Condition monitoring of rolling

element bearings during low and

variable speed conditions
 Andreas Klausen

Condition monitoring of rolling element bearings

during low and variable speed conditions

Doctoral Dissertation for the Degree Philosophiae Doctor (PhD) at

the Faculty of Engineering and Science, Specialisation in Mechatronics

University of Agder
Faculty of Engineering and Science
2019
Doctoral Dissertation at the University of Agder 227
ISSN: 1504-9272
ISBN: 978-82-7117-926-7

c Andreas Klausen, 2019

Printed by Media 07
Oslo
Acknowledgments

This project has been carried out in the period between October 2014 and January 2019
as part of an integrated PhD program that merged the last year of my master’s education
with the first year as a Research Fellow. The work towards this doctoral thesis has been
carried out at the University of Agder, campus Grimstad, and the project is funded by
the Ministry of Education in Norway. During this project I have learned a lot, mostly
from moving to a new field of research which is not covered of my Mechatronics education.
During this time, Prof. Kjell G. Robbersmyr has been my main supervisor. In addition,
I have had the supervision of two co-supervisors: Prof. Hamid R. Karimi until December
2016 and Dr. Huynh V. Khang from January 2017.
I want to thank my main supervisor Prof. Kjell G. Robbersmyr for encouraging me in
perusing this field of research and for his supervision. I also want to thank Prof. Hamid
R. Karimi for his great help with writing papers and setting the theoretical framework of
the plan. Finally, I would like to thank my co-supervisor Dr. Huynh V. Khang for his
supervision and teaching me a great deal about publishing papers.
Incorporating both theoretical and experimental work has taught me that experiments
doesn’t always work as theorized, and problems occur. I am grateful for the help from
the lab staff at the Mechatronics lab, especially Roy W. Folgerø for aiding in designing
the test rig and building it afterwards.
The time spent at the office would not be as great if it weren’t for my good colleagues.
I have had the pleasure sharing an office with my friends Sondre Tørdal, Rune Husveg,
Sondre Nordås and Philipp Pasolli who have always brought an enjoyable atmosphere for
work and discussions. I want to thank Sondre Tørdal and Morten Ottestad for taking
their time with discussions when I asked for help. I have also had great collaborations
with my friends studying condition monitoring of rotating machinery, Surya, Martin and
Jagath. Thank you for sharing your knowledge and working with me.
To my parents Håkon, Marianne, and my brother and sister Kristian and Solveig:
Thank you for your unconditional love and support. Last, but not least, to my fiancée
Veronica, thank you for mending my doubts, enduring my stress and late nights, and for
your continuous love and support through my tough times.

v
Summary

Rotating machinery is common in industry, and the reliability is critical for continued op-
eration and safety, as undetected component faults may force emergency shutdowns. With
the increased complexity of machines, spare parts and maintenance crew are often not
immediately available for a quick overhaul, which increases downtime. This is especially
true when repairing an off-shore wind mill because suitable weather condition is required
to board it. Rolling element bearings carry the shaft load using elements that rotate with
shaft between two raceways. Despite being a mature technology, bearings are the most
common component in rotating machinery to fail because of highly dynamic loading. The
expected bearing lifetime is challenging to determine due to the statistical nature of man-
ufacturing quality, and difficulties in modeling the actual operating conditions. Condition
monitoring is therefore vital to increase reliability on critical machines. Vibration mea-
surements have been used for several decades to estimate the bearing health state, and
the currently available methods can solve most condition monitoring problems. However,
low and variable speed conditions can present a problem for these existing techniques,
because of low impact energy and non-stationary vibration signals. Such conditions are
typical in wind mills which operate with low rotor speed and are subjected to variable
wind speeds. A bearing is normally fully functional for a limited amount of time after the
initial fault. Estimating this time frame is challenging as the degradation trend must be
accurately predicted. However, an estimate of the remaining useful life is beneficial for
maintenance planning. Despite being a mature research field, most of the up-to-date tech-
niques require historic failure data to set parameters of prediction models. Such failure
data is often not available when installing a condition monitoring system, which render a
challenge for estimating the remaining useful life on new machines.
In this project, new techniques are proposed to alleviate the aforementioned challenges
in fault diagnosis and prognosis. Two new algorithms are developed for fault detection of
bearings operating under low and variable speed conditions. Additionally, a new algorithm
is proposed for estimating the remaining useful life on new machines with no historic
failure data. The developed methods are described in this thesis, while details are provided
in the appended papers.

vi
Publications

The following listed articles have been published or submitted for publication in peer
reviewed conference proceedings and journals. The version presented in this thesis differs
only in formatting and minor errata.

Paper A A. Klausen, R. W. Folgerø, K. G. Robbersmyr and H. R. Karimi. Accelerated

Bearing Life-time Test Rig Development for Low Speed Data Acquisition. Journal of
Modeling, Identification and Control, 38(3):143-156, 2017. doi: 10.4173/mic.2017.3.4.

Paper B A. Klausen, K. G. Robbersmyr and H. R. Karimi. Autonomous Bearing Fault

Diagnosis Method based on Envelope Spectrum. IFAC-PapersOnLine, 50(1):13378-
13383, 2017. doi: 10.1016/j.ifacol.2017.08.2262.

Paper C A. Klausen and K. G. Robbersmyr. Cross-correlation of Whitened Vibration

Signals for Low-Speed Bearing Diagnostics. Mechanical Systems and Signal Pro-
cessing, 118:226-244, 2019. doi: 10.1016/j.ymssp.2018.08.048.

Paper D A. Klausen, H. V. Khang and K. G. Robbersmyr. Multi-band Identification

for Enhancing Bearing Fault Detection in Variable Speed Conditions. Under review
at Mechanical Systems and Signal Processing.

Paper E A. Klausen, H. V. Khang and K. G. Robbersmyr. Novel Threshold Calcula-

tions for Remaining Useful Lifetime Estimation of Rolling Element Bearings. IEEE
International Conference on Electrical Machines (ICEM), Greece, 1912-1918, 2018.
doi: 10.1109/ICELMACH.2018.8507056

Paper F A. Klausen, H. V. Khang and K. G. Robbersmyr. Novel RMS Based Health

Indicator used for Remaining Useful Lifetime Estimation of Bearings. Under review
at Mechanical Systems and Signal Processing.

vii
 viii

The following articles were written and published during the time of this project but are
not included in the thesis.

Paper G S. T. Kandukuri, A. Klausen, H. R. Karimi and K. G. Robbersmyr. A review

of diagnostics and prognostics of low-speed machinery towards wind turbine farm-
level health management. Renewable and Sustainable Energy Reviews, 53:697-708.
doi: 10.1016/j.rser.2015.08.061.

Paper H S. T. Kandukuri, A. Klausen and K. G. Robbersmyr. Fault diagnostics of wind

turbine electric pitch systems using sensor fusion approach. Journal of Physics:
Conference Series, 1037(3):032036. doi: 10.1088/1742-6596/1037/3/032036.

Paper I M. A. A. Ismail and A. Klausen. Multiple Defect Size Estimation of Rolling

Bearings using Autonomous Diagnosis and Vibrational Jerk. 7th World Conference
on Structural Control and Monitoring (7WCSCM), China, July 2018.
Contents

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Fault diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Prognostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Motivation and Problem Statement . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 Experimental test data 15

2.1 In-house test rig - Paper A . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Downloaded datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.1 Variable speed datasets . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.2 Accelerated lifetime datasets . . . . . . . . . . . . . . . . . . . . . . 18
2.2.3 Constant speed datasets . . . . . . . . . . . . . . . . . . . . . . . . 18

3 Fault diagnostics 19
3.1 Theoretical background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1.1 Order tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1.2 Signal whitening . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.1.3 Envelope spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1.4 Fault detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 Automatic fault detection - Paper B . . . . . . . . . . . . . . . . . . . . . 22
3.3 Whitened cross-correlation spectrum - Paper C . . . . . . . . . . . . . . . 25
3.4 Resonance frequency identification - Paper D . . . . . . . . . . . . . . . . . 27

4 Prognostics 33
4.1 RMS health indicator - Paper E . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Filter bank RMS - Paper F . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.3 Remaining useful life estimation - Paper F . . . . . . . . . . . . . . . . . . 39

ix
CONTENTS x

5 Concluding Remarks 47
5.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.3 Further Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

References 51

Appended Papers 57

A Accelerated Bearing Life-time Test Rig Development for Low Speed

Data Acquisition 59
A.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
A.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
A.2.1 Accelerated life-time testing . . . . . . . . . . . . . . . . . . . . . . 63
A.2.2 Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
A.2.3 Concept Development . . . . . . . . . . . . . . . . . . . . . . . . . 64
A.3 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
A.3.1 The chosen solution . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
A.3.2 Test rig design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
A.3.3 Axial and Radial Loads . . . . . . . . . . . . . . . . . . . . . . . . 71
A.3.4 Complete test rig . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
A.3.5 Controller hardware . . . . . . . . . . . . . . . . . . . . . . . . . . 77
A.3.6 Load controller design . . . . . . . . . . . . . . . . . . . . . . . . . 78
A.3.7 Accelerated life-time test software design . . . . . . . . . . . . . . . 79
A.3.8 Case study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
A.3.9 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
A.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

B Autonomous Bearing Fault Diagnosis Method based on Envelope Spec-

trum 91
B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
B.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
B.2.1 Experimental Test Setup . . . . . . . . . . . . . . . . . . . . . . . . 95
B.2.2 Pre-process Vibration Signal . . . . . . . . . . . . . . . . . . . . . . 96
B.2.3 Narrow-band Identifier Algorithm . . . . . . . . . . . . . . . . . . . 96
B.2.4 Envelope Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
B.2.5 Autonomous Fault Diagnosis . . . . . . . . . . . . . . . . . . . . . . 98
B.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
CONTENTS xi

B.3.1 Vibration Mode Score . . . . . . . . . . . . . . . . . . . . . . . . . 102

B.3.2 Vibration Trend Score . . . . . . . . . . . . . . . . . . . . . . . . . 104
B.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

C Cross-correlation of Whitened Vibration Signals for Low-Speed Bearing

Diagnostics 109
C.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
C.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
C.2.1 Experimental test setup . . . . . . . . . . . . . . . . . . . . . . . . 114
C.2.2 Order tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
C.2.3 Angle synchronous average . . . . . . . . . . . . . . . . . . . . . . . 118
C.2.4 Autoregressive whitening filter . . . . . . . . . . . . . . . . . . . . . 119
C.2.5 Hilbert transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
C.2.6 Cross-correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
C.2.7 Order spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
C.2.8 Theoretical Discussion on WCCS . . . . . . . . . . . . . . . . . . . 122
C.2.9 Complete algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 124
C.2.9.1 Whitened Cross-correlation Spectrum . . . . . . . . . . . 124
C.2.9.2 Comparison methods . . . . . . . . . . . . . . . . . . . . . 125
C.2.9.3 Performance metrics . . . . . . . . . . . . . . . . . . . . . 127
C.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
C.3.1 Kurtosis trend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
C.3.2 Whitened Cross-correlation Spectrum . . . . . . . . . . . . . . . . . 128
C.3.3 Comparison to other methods . . . . . . . . . . . . . . . . . . . . . 133
C.3.3.1 Performance metrics . . . . . . . . . . . . . . . . . . . . . 135
C.3.4 Supplementary benchmark case study . . . . . . . . . . . . . . . . . 137
C.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

D Multi-band Identification for Enhancing Bearing Fault Detection in Vari-

able Speed Conditions 145
D.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
D.2 Bearing fault diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
D.3 Proposed method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
D.3.1 Complete algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
D.3.2 Order tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
D.3.3 Cepstrum pre-whitening . . . . . . . . . . . . . . . . . . . . . . . . 152
CONTENTS xii

D.3.4 Order tracking and cepstrum pre-whitening . . . . . . . . . . . . . 153

D.4 Comparison methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
D.4.1 Spectral kurtosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
D.4.2 Cepstrum pre-whitening . . . . . . . . . . . . . . . . . . . . . . . . 154
D.4.3 Diagnosis score . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
D.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
D.5.1 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
D.5.2 Test rig 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
D.5.3 Test rig 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
D.5.4 Test rig 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
D.6 Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
D.6.1 Spectral kurtosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
D.6.2 Cepstrum pre-whitening . . . . . . . . . . . . . . . . . . . . . . . . 167
D.6.3 Diagnosis score . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
D.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

E Novel Threshold Calculations for Remaining Useful Lifetime Estimation

of Rolling Element Bearings 173
E.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
E.2 ISO standard threshold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
E.3 Proposed threshold transformation . . . . . . . . . . . . . . . . . . . . . . 179
E.4 Practical test of threshold setting . . . . . . . . . . . . . . . . . . . . . . . 180
E.5 Threshold setting for multiple physical health indicators . . . . . . . . . . 182
E.6 Practical test of MD threshold . . . . . . . . . . . . . . . . . . . . . . . . . 183
E.6.1 IMS dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
E.6.2 PHM Challenge dataset . . . . . . . . . . . . . . . . . . . . . . . . 184
E.6.3 Low-speed test rig . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
E.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190

F Novel RMS Based Health Indicators used for Remaining Useful Lifetime
Estimation of Bearings 193
F.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
F.2 Velocity vs acceleration RMS . . . . . . . . . . . . . . . . . . . . . . . . . 197
F.3 Proposed RMS health indicator . . . . . . . . . . . . . . . . . . . . . . . . 199
F.3.1 Filter bank RMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
F.3.2 Filterbank RMS failure thresholds . . . . . . . . . . . . . . . . . . . 201
CONTENTS xiii

F.4 Remaining useful life estimation . . . . . . . . . . . . . . . . . . . . . . . . 203

F.4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
F.4.2 Spearman coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . 203
F.4.3 RMS trend selection . . . . . . . . . . . . . . . . . . . . . . . . . . 204
F.4.4 Degradation model . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
F.4.5 Initial parameter setting . . . . . . . . . . . . . . . . . . . . . . . . 205
F.4.6 Particle filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
F.4.7 Weighted RUL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
F.5 Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
F.5.1 Test rig 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
F.5.2 Test rig 2 — in-house setup . . . . . . . . . . . . . . . . . . . . . . 211
F.5.3 Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
F.6 Further work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
F.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
List of Figures

1.1 Simplified schematic of a bearing condition monitoring system. . . . . . . . 3

1.2 Bearing dimensions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 Simplified description of bearing prognostics procedure. . . . . . . . . . . . 7

2.1 Schematic of the accelerated bearing life-time test rig. . . . . . . . . . . . . 16

2.2 The finished test rig in the laboratory at the University of Agder. . . . . . 17

2.3 Damaged bearings after two accelerated life time tests. (a) Rollers; (b)
Outer race; (c) Inner race. . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.1 Envelope in time and frequency domain for the three fault types. . . . . . . 22

3.2 The vibration spectrum shown with the nine identified resonance mode
bands. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.3 The first harmonic and its side-bands, and the prominence threshold given
by 3 times the noise floor. The search width is given by the stapled lines. . 24

3.4 The three prominent harmonics and side-bands identified in the third fre-
quency band. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.5 The fault scores over time for the last 100 datasets. The root mean square
(RMS) is displayed to compare the fault score with the vibration energy. . 25

3.6 The WCCS used to diagnose a roller fault. Harmonics and side-bands
linked to the roller fault are marked in all three subfigures. (a) Spectrum
of the whitened vibration signal xw ; (b) Spectrum of the envelope xenv . (c)
Proposed WCCS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.7 Resonance band identification using the proposed method on a simulated

signal—Part 1. (a) shaft vibration; (b) bearing vibration; (c) shaft speed;
(d) measured vibration signal; (e) frequency spectrum of raw signal; (f)
order spectrum. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

xiv
LIST OF FIGURES xv

3.8 Resonance band identification using the proposed method on a simulated

signal—Part 2. (a) order spectrum after CPW; (b) frequency spectrum
after inverse order tracking with a red line showing the amplitude-filtered
signal; (c) amplitude-filtered frequency spectrum raised to power of 5, with
a suitable band-pass filter region marked in red; (d) envelope order spec-
trum after band-pass filtration of original signal. . . . . . . . . . . . . . . . 29
3.9 Results from diagnosing an outer race fault on a variable speed dataset.
(a) shaft speed during measurement; (b) frequency spectrum of the raw
vibration signal; (c) frequency spectrum after using the proposed method,
where three suitable band-pass filter areas are marked; (d)–(f) envelope
order spectra after band-pass filtration at the three marked areas, respec-
tively. Red triangles show identified harmonics related to the fault. . . . . 30

4.1 Difference between RMS values calculated using velocity and acceleration
units. (a) RMS of the velocity signal; (b) RMS of the acceleration signal.
The mean R̄ and standard deviation σ are estimated using the first 50
hours of data files. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Comparison of the RMS threshold using velocity and acceleration units.
(a) Velocity-based RMS; (b) Acceleration-based RMS. . . . . . . . . . . . . 36
4.3 Ra compared with the mean degradation trend to highlight oscillations. . . 36
4.4 A collection of RMS trends with failure FTs for the vibration dataset. (a)
R15 ; (b) R11 ; (c) R6 ; (d) R3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.5 Identified RMS trends with high Spearman coefficient, and output of the
corresponding PFs. Rows 1-5 indicate i = [15, 14, 12, 6, 3]. (column 1) Ri ,
FT R̂i , and median and 95% CI of initial PF output; (column 2) µα over
time for the initiated PF. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.6 PF prediction of each identified RMS trend at t = 130 hours. . . . . . . . . 43
4.7 Weighted RUL decision. (a) weighted mean, 95% CI and true RUL of the
dataset; (b) weights for each PF output. . . . . . . . . . . . . . . . . . . . 44
4.8 Weighted RUL PDF over time shown with the weighted mean RUL and
true RUL. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

A.1 Test rig pre-design function tree. *An Open-Loop force feedback is a feed-
forward calculation of the assumed forces based on a particular input. . . 65
A.2 Chosen shaft design and bearing placements. (a) Counteracting forces on
all the bearings. (b) Position of the required bearings. . . . . . . . . . . . . 68
A.3 Bearing housings as placed on the shaft. . . . . . . . . . . . . . . . . . . . 68
LIST OF FIGURES xvi

A.4 The sensor layout on the test bearing housing: (a) Front of the housing
with the bearing in the middle and sensors around it. (b) Side of the test
bearing housing with the proximity sensor location. . . . . . . . . . . . . . 69
A.5 Concentric unit housing the thrust bearing. . . . . . . . . . . . . . . . . . 70
A.6 Axial load setup. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
A.7 Closeup details of the axial load setup. . . . . . . . . . . . . . . . . . . . . 71
A.8 Radial load setup. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
A.9 Closeup details of the radial load setup. . . . . . . . . . . . . . . . . . . . . 72
A.10 Bracket with linear rolling element bearings connected to the test bearing
housing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
A.11 Free Body Diagram of the axial load setup. . . . . . . . . . . . . . . . . . . 73
A.12 Free Body Diagram of radial load setup. . . . . . . . . . . . . . . . . . . . 74
A.13 Complete 3D model with numbered components. The width of the test rig
is 350 mm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
A.14 Electronic hardware connections. . . . . . . . . . . . . . . . . . . . . . . . 77
A.15 Control scheme for both linear actuators. . . . . . . . . . . . . . . . . . . . 79
A.16 Control scheme for the accelerated life-test. . . . . . . . . . . . . . . . . . . 80
A.17 Resulting kurtogram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
A.18 Squared Envelope Spectrum after bandpass filtering. . . . . . . . . . . . . 82
A.19 Shaft rotational speed at a reference speed of 20 rpm. . . . . . . . . . . . . 82

B.1 The frequency spectrum of the vibration signal in addition to its envelope. 97
B.2 The frequency spectrum of the vibration signal combined with the nine
narrow-bands that are identified for this dataset. . . . . . . . . . . . . . . . 97
B.3 For H = 1, the frequency bands for identifying the harmonic value and the
noise are shown. In addition, the harmonic is marked as a green dot and
the noise level is the red-stapled line. . . . . . . . . . . . . . . . . . . . . . 100
B.4 For H = 1, the frequency bands for identifying the sub-bands are shown.
In addition the identified sub-bands are marked with a red and blue dot. . 100
B.5 For H = 2, the harmonic and two sub-bands are identified. . . . . . . . . 101
B.6 Complete algorithm flowchart. . . . . . . . . . . . . . . . . . . . . . . . . . 102
B.7 For the third narrow-band, M = 3, a total of 3 harmonics related to the
inner-race fault are identified. . . . . . . . . . . . . . . . . . . . . . . . . . 103
B.8 For the sixth narrow-band, M = 6, a total of 3 harmonics related to the
inner-race fault is identified. . . . . . . . . . . . . . . . . . . . . . . . . . . 103
B.9 For the last 100 datasets, the characteristic fault scores are calculated using
(B.15) and shown in a). The RMS trend calculated using (B.16) is shown
in b) for comparison. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
LIST OF FIGURES xvii

B.10 The de-assembled test bearing after the accelerated life test. a) shows two
damaged balls, b) shows a small pit in the outer-ring, and c) shows a large
area of pitting on the inner-ring. . . . . . . . . . . . . . . . . . . . . . . . . 106

C.1 Simplified view of the accelerated bearing life-time test bench. . . . . . . . 115
C.2 The damaged components in the disassembled bearing. (a) Three damaged
rollers. (b) Damaged area on the outer-race. (c) Large damaged area on
the inner-race. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
C.3 Example scenario where an impulse train is revealed after applying the
autoregressive whitening filter. (a) The example input signal. (b) The
residual after applying the whitening filter. (c) The AICC criteria as a
function of the filter order together with the optimal value. (d) The resid-
ual power as a function of the filter order. . . . . . . . . . . . . . . . . . . 121
C.4 Flowchart of WCCS used to diagnose the low-speed bearing for faults. . . . 124
C.5 Three additional methods used to compare performance with the WCCS.
(a) Method A involves diagnosing the bearing using the Envelope Spec-
trum. (b) Method B improves on the Envelope Spectrum by bandpass
filtering the whitened vibration signal around the narrow-band yielding
maximum kurtosis. (c) Method C involves the fast calculation of the Spec-
tral Correlation which is a tool for analyzing cyclostationary signals. . . . . 126
C.6 Kurtosis of the vibration datasets measured at 20 rpm shaft speed. (a)
The kurtosis at an early stage of damage. (b) The kurtosis at an advanced
stage of damage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
C.7 The WCCS used to diagnose a roller fault at kurtosis point 1 in Figure
C.6. Harmonics and side-bands linked to the roller fault are marked in all
three subfigures. (a) The spectrum of the whitened vibration signal Vw .
(b) The spectrum of the envelope Venv . (c) The proposed WCCS and a
red stapled line which corresponds to the mean spectrum value. . . . . . . 130
C.8 The WCCS computed 10,000 revolutions prior to kurtosis point 1 in Figure
C.6. The roller damage is not visible. . . . . . . . . . . . . . . . . . . . . . 131
C.9 Baseline WCCS at the start of the accelerated life-test. A peak at the char-
acteristic frequency for an outer-race fault is visible, however the bearing
is undamaged at this stage. . . . . . . . . . . . . . . . . . . . . . . . . . . 131
C.10 The WCCS used to diagnose an outer-race fault at kurtosis point 2 in
Figure C.6. Harmonics linked to the outer-race fault are marked in all
three subfigures. (a) The spectrum of the whitened vibration signal Vw .
(b) The spectrum of the envelope Venv . (c) The proposed WCCS. . . . . . 132
LIST OF FIGURES xviii

C.11 The WCCS used to diagnose an inner-race fault at kurtosis point 3 in Figure
C.6. Harmonics and side-bands linked to an inner-race fault are marked
in all three subfigures. (a) The spectrum of the whitened vibration signal
Vw . (b) The spectrum of the envelope Venv . (c) The proposed WCCS. . . 134
C.12 The resulting Envelope Spectrum from using method A. The roller fault
harmonics are barely visible, while shaft and motor vibration dominate
the spectrum. The mean value of the spectrum is marked for comparison
purposes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
C.13 Resulting Kurtogram in method B indicating a maximum of kurtosis at a
central frequency of 24,800 Hz and a bandwidth of 1600 Hz. The kurtosis
value at that narrow-band is 18.2. . . . . . . . . . . . . . . . . . . . . . . . 136
C.14 Squared envelope spectrum of the band-pass filtered vibration signal in
Method B. The roller fault harmonics and side-bands are visible, and the
mean value of the spectrum is marked for comparison purposes. . . . . . . 136
C.15 Resulting Enhanced Envelope Spectrum from the Fast Spectral Correlation
(Method C). The roller fault harmonics and side-bands are barely visible,
and the mean value of the spectrum is marked for comparison purposes. . . 137
C.16 The WCCS on the CWRU healthy dataset 100DE recorded on the drive-end.138
C.17 The WCCS used to diagnose an inner-race fault located at the fan-end of
the CWRU dataset. The vibration was measured at the drive-end. Har-
monics and side-bands linked to an inner-race fault are marked in all three
subfigures. (a) The spectrum of the whitened vibration signal Vw . (b) The
spectrum of the envelope Venv . (c) The proposed WCCS. . . . . . . . . . . 139

D.1 Simplified bearing condition monitoring setup. . . . . . . . . . . . . . . . . 150

D.2 Flow diagram of the proposed method. x(t) and θ(t) are the vibration signal
and encoder signal, respectively. fc and fbw are the central frequency and
bandwidth of the identified resonance mode. . . . . . . . . . . . . . . . . . 151
D.3 Simulation showing the resonance frequency spreading effect, and determin-
istic component de-spreading. (a) shaft vibration; (b) bearing vibration;
(c) shaft vibration frequency spectrum; (d) bearing vibration frequency
spectrum; (e) shaft vibration order spectrum; (f) bearing vibration order
spectrum. One order is equal 10 Hz in this example. . . . . . . . . . . . . . 152
D.4 Visualization of the ratio between resonance mode amplitude versus the
most prominent amplitude of other components in the frequency spectrum
(ROR) after each step in the proposed method. . . . . . . . . . . . . . . . 154
LIST OF FIGURES xix

D.5 Resonance band identification using the proposed method on a simulated

signal. (a) shaft vibration; (b) bearing vibration; (c) shaft speed; (d) mea-
sured vibration signal; (e) frequency spectrum of raw signal; (f) order spec-
trum; (g) order spectrum after CPW; (h) frequency spectrum after inverse
order tracking with a red line showing the amplitude-filtered signal; (i)
amplitude-filtered frequency spectrum raised to power of 5, with a suitable
band-pass filter region marked in red; (j) envelope order spectrum after
band-pass filtration of original signal. . . . . . . . . . . . . . . . . . . . . . 157
D.6 Simplified schematic of the test setup in . . . . . . . . . . . . . . . . . . . 157
D.7 Results from diagnosing an outer race fault on test rig 1. (a) shaft speed
during measurement; (b) frequency spectrum of the raw vibration signal;
(c) frequency spectrum after using the proposed method, where three suit-
able band-pass filter areas are marked; (d)–(f) envelope order spectra after
band-pass filtration at the three marked areas, respectively. Red triangles
show identified harmonics related to the fault. . . . . . . . . . . . . . . . . 158
D.8 Results from diagnosing a roller fault on test rig 1. (a) shaft speed dur-
ing measurement; (b) frequency spectrum of the raw vibration signal; (c)
frequency spectrum after using the proposed method, where three suit-
able band-pass filter areas are marked; (d)–(f) envelope order spectra after
band-pass filtration at the three marked areas, respectively. Red triangles
show identified harmonics related to the fault. . . . . . . . . . . . . . . . . 160
D.9 Results from diagnosing an inner race fault on test rig 1. (a) shaft speed
during measurement; (b) frequency spectrum of the raw vibration signal;
(c) frequency spectrum after using the proposed method, where three suit-
able band-pass filter areas are marked; (d f) envelope order spectra after
band-pass filtration at the three marked areas, respectively. Red and blue
triangles show identified harmonics and side-bands related to the fault. . . 161
D.10 Simplified schematic of the test rig used at CWRU. . . . . . . . . . . . . . 161
D.11 Time-frequency diagram of the enveloped vibration data. An identified
ridge is marked as a red-stapled line. This ridge is identified as the third
order of shaft vibration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
D.12 Results from diagnosing an inner race fault on test rig 2. (a) shaft speed
during measurement; (b) frequency spectrum of the raw vibration signal;
(c) frequency spectrum after using the proposed method, where 5 suit-
able band-pass filter areas are marked; (d)–(h) envelope order spectra after
band-pass filtration at each identified mode. Red triangles show identified
harmonics elated to the fault. . . . . . . . . . . . . . . . . . . . . . . . . . 163
LIST OF FIGURES xx

D.13 Results from diagnosing an outer race fault on test rig 3. (a) shaft speed
during measurement; (b) frequency spectrum of the raw vibration signal;
(c) frequency spectrum after using the proposed method, where 3 suitable
band-pass filter areas are marked; (d)–(f) envelope order spectra after band-
pass filtration. Red triangles show identified harmonics related to the fault. 164
D.14 Simplified schematic of the in-house accelerated life-time test rig . . . . . . 164
D.15 Results from using the fast Kurtogram detailed in Section D.4.1 on all the
experimental datasets. Pairs consisting of a Kurtogram and the resulting
ES are given for each dataset. (a)–(b) test rig 1, dataset 1; (c)–(d) test
rig 1, dataset 2; (e)–(f) test rig 1, dataset 3; (g)–(h) test rig 2, dataset
1; (i)–(j) test rig 3, dataset 1. In a Kurtogram, red lines indicate central
frequencies identified using the proposed method. In envelope spectra, red
and blue triangles indicate prominent harmonics and side-bands related to
the fault, respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
D.16 Results from using the cepstrum pre-whitening method detailed in Section
D.4.2 on all the experimental datasets. Envelope spectra with identified
harmonics and side-bands are shown for each dataset. (a)–(c) test rig 1,
dataset 1 through 3, respectively; (d) test rig 2, dataset 1; (e) test rig 3,
dataset 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

E.1 Difference between RMS values calculated using vibration in velocity and
acceleration units. (a) The RMS of the velocity signal. (b) The RMS of the
acceleration signal. Both plots show a red-stapled line which is the mean
RMS value of the first 300 data files. . . . . . . . . . . . . . . . . . . . . . 178
E.2 Comparison of the RMS threshold using velocity and acceleration units
for dataset 1. (a) The RMS of the velocity signal. (b) The RMS of the
acceleration signal. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
E.3 The HI plotted with the HIt threshold for dataset 1. . . . . . . . . . . . . . 185
E.4 Comparison of the RMS threshold using velocity and acceleration units
for dataset 2. (a) The RMS of the velocity signal. (b) The RMS of the
acceleration signal. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
E.5 The HI plotted with the HIt threshold for dataset 2. . . . . . . . . . . . . . 187
E.6 Comparison of the RMS threshold using velocity and acceleration units
for dataset 3. (a) The RMS of the velocity signal. (b) The RMS of the
acceleration signal. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
E.7 The HI plotted with the HIt threshold for dataset 3. . . . . . . . . . . . . . 189
LIST OF FIGURES xxi

F.1 Comparison between velocity- and acceleration-based RMS for the IMS
dataset. (a) velocity-based RMS and its FT; (b) acceleration-based RMS
and its FT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
F.2 Ra compared to an artificially created mean trend to highlight oscillations. 199
F.3 Collection of RMS trends with FTs for the IMS dataset. (a) R15 ; (b) R11 ;
(c) R6 ; (d) R3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
F.4 Flowchart of the proposed method. . . . . . . . . . . . . . . . . . . . . . . 204
F.5 Identified RMS trends with high Spearman coefficient, and output of the
corresponding PFs—Part 1. Rows 1-3 indicate i = [15, 14, 12]. (column 1)
Ri , FT R̂i , and median and 95% CI of initial PF output; (column 2) µα
over time for the initiated PF; (column 3) predicted PF trend at t = 130
hours. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
F.6 Identified RMS trends with high Spearman coefficient, and output of the
corresponding PFs—Part 2. Rows 1-2 indicate i = [6, 3]. (column 1) Ri ,
FT R̂i , and median and 95% CI of initial PF output; (column 2) µα over
time for the initiated PF; (column 3) predicted PF trend at t = 130 hours. 210
F.7 Weighted RUL of the IMS dataset. (a) weighted mean, 95% CI and true
RUL of the IMS dataset; (b) weights for each PF output. . . . . . . . . . . 212
F.8 Weighted RUL PDF of the IMS dataset over time. The weighted mean and
true RUL are shown for reference. . . . . . . . . . . . . . . . . . . . . . . . 213
F.9 Comparison of velocity- and acceleration-based RMS of the in-house test
rig dataset. (a) velocity-based RMS; (b) acceleration-based RMS. . . . . . 214
F.10 Collection of RMS trends with FTs for the in-house test rig dataset. (a)
R40 ; (b) R19 ; (c) R11 ; (d) R1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
F.11 R1 and initial median and 95% CI output of the corresponding PF. (a) R1
with marked part as optimization input, blue line for FT, and red lines for
median and 95% CI of PF; (b) α over time for the initiated PF. . . . . . . 214
F.12 Weighed RUL of the in-house test rig dataset. The weighted mean, 95%
CI and true RUL are shown. . . . . . . . . . . . . . . . . . . . . . . . . . . 215
F.13 Weighted RUL PDF of the in-house test rig dataset over time. The weighted
mean and true RUL are shown for reference. . . . . . . . . . . . . . . . . . 215
F.14 Health stage division from energy cycles of a high-frequency band on the
IMS dataset. (a) High-frequency band R15 and identified HS transitions;
(b) low-frequency band R3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
F.15 Health stage division from energy cycles of a high-frequency band on the
in-house test rig dataset. (a) High-frequency band R40 and identified HS
transitions; (b) low-frequency band R1 . . . . . . . . . . . . . . . . . . . . . 218
List of Tables

2.1 Run to failure experiments using the in-house test rig. . . . . . . . . . . . . 16

A.1 Operator settings for the test (nc = 6). . . . . . . . . . . . . . . . . . . . . 83

A.2 Characteristic bearing fault frequencies in shaft orders. . . . . . . . . . . . 83

B.1 Inner ring diagnostic score for the nine narrow-bands in the example data. 104
B.2 Characteristic fault frequencies for the test bearing in orders. . . . . . . . . 104

C.1 Characteristic fault orders for the test bearing given in shaft orders assum-
ing zero contact angle and no roller slip. The expected harmonic order and
the accompanying side-band order (due to non-homogeneous radial load)
for the given fault type are given. . . . . . . . . . . . . . . . . . . . . . . . 116
C.2 Specifications of the datasets at the three kurtosis points. . . . . . . . . . . 129
C.3 Performance metric for each method. The WCCS scores the highest in
both categories. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
C.4 Specifications for the tested CWRU datasets. DE=Drive-end, FE=Fan-end 138

D.1 Diagnosis score per band for each dataset using the proposed method. The
best mode for each dataset is written in bold. “DS X.Y” means “Test rig
X - dataset Y”. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
D.2 Diagnosis score per dataset for each method. For the proposed method, the
best scoring ES is the basis. The best method for each dataset is written
in bold. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

E.1 Threshold value calculations. . . . . . . . . . . . . . . . . . . . . . . . . . . 184

xxii
Nomenclature

Abbreviations IMS intelligent maintenance systems

ADC analog-to-digital converter LSC low speed conditions

AE acoustic emission PCA principal component analysis

BPFI ball pass frequency inner race PDF probability density function

BPFO ball pass frequency outer race PF particle filter

PHI physical health indicator

BSF ball spin frequency
RMS root mean square
CBM condition based maintancne
rpm rounds per minute
CM condition monitoring
RUL remaining useful life
CPW cepstrum pre-whitening
RVM relevance vector machine
CS2 second order cyclostationary
SOM self-organizing map
CWR case western reserve
STD standard deviation
DE drive-end
SVM support vector machine
DFR discrete Fourier transform
TSA time synchronous average
DRS discrete/random separator
VHI virtual health indicator
FE fan-end
VSC variable speed condition
FIR finite impulse response Variables
FTF fundamental train frequency a crack size

FT failure threshold α modified Paris-Law parameter

HI health indicator β modified Paris-Law parameter

hp horse power C bearing dynamic capacity

xxiii
LIST OF TABLES xxiv

D bearing pitch diameter p number of autoregressive model

parameters
d roller diameter
φ radial plane load angle
∆f frequency step
P bearing load
∆nc loading cycles
qj j’th autoregressive model param-
∆t discrete time step
eter
∆θ shaft position difference
Ra vibration RMS
E signal energy R̂a acceleration-based RMS threshold

F discrete Fourier transform R̄ Mean initial RMS value

fot (θ) interpolated vibration signal ρL lower Spearman threshold

Fp positive-sided frequency spectrum R RMS of frequency band i

i

fr shaft frequency Rv velocity-based vibration RMS

R̂i FT of Ri R̂v velocity-based RMS threshold

i, j, k indexes σa STD of acceleration-based RMS

L10 bearing nominal life σα process noise STD

m R̂i FT scaling factor σh measurement noise STD

µα mean α σi STD of Ri during baseline

µi mean of Ri during baseline σv STD of velocity-based RMS

N normal distribution θ(t) shaft position

N vibration signal samples W PF weight for RUL estimation

Nb number of frequency bands w particle weight

nb samples in each frequency band X discrete frequncy spectrum

nc number of cycles x(t) vibration signal

να process noise xa (t) analytic vibration signal

νh measurement noise xenv (t) vibration envelope

LIST OF TABLES xxv

xi (t) Hilbert transform of vibration sig- xw (t) whitened vibration signal

nal
z particle
xot (t) order-tracked vibration signal

xv (t) velocity-based vibration signal Zb number of rollers

Chapter 1

Introduction

1.1 Background
Rotating machines have bearings which keep the shaft in place while providing as little
friction and noise as possible. Various types of bearings are available, each with their own
strengths and weaknesses. Rolling element bearings, plain bearings, magnetic bearings,
and fluid film bearings are some commercially available types. Rolling element bearings
are frequently used in gearboxes, motors and pumps, because of their ability to withstand
high loads at high speeds. There are four main parts in a rolling element bearing. An
inner race is fastened to the rotating shaft, while an outer race sits stationary inside a
housing. Rolling elements move around the shaft between the raceways, and a cage keeps
the elements at a near-fixed distance from each other to better distribute the shaft load
and hinder them from crashing into each other.
Different rolling element types are available, and the optimal choice depends on the
application. Roller elements are cylinder-shaped and provide a long surface line to dis-
tribute the load. Therefore, they can withstand high radial loads, but are not suitable for
axial loads due to the regular cylindrical design. Tapered and spherical rollers are more
suited for axial loads and misaligned systems but have a more complicated design and are
therefore more expensive to manufacture. Additionally, more specialized designs, such
as gear bearings and needle bearings, can be found in industry. In this thesis, the focus
is on ball element bearings which can commonly be found in electrical machines. These
bearings can be used at a high rotational speed, and the ball design permits both axial
and radial loads. In addition, ball bearings are one of the cheapest to manufacture given
the simplicity of a round element shape.
Oil or grease lubricant must be provided in the bearing to prevent metal-to-metal
contact and fill gaps made by imperfections on the surface of rollers and raceways. With-
out adding a lubricant, the metal-to-metal friction is significant, which results in early

1
failure. The oil pressure between roller and raceway causes small elastic deformation of
the elements. This cyclic deformation will eventually lead to fatigue, and the bearing
may fail even if it is perfectly lubricated. The bearing life can be estimated by using
the load, speed, temperature, lubricant viscosity, and load capacity of the bearing. The
estimated life of a bearing is generally known as L10 , which denotes how many million
revolutions 10 out of 100 bearings should last if installed correctly and operated under
the specified conditions. If a higher confidence is desired, e.g. L5 , the estimated life is
drastically reduced Similarly, L50 can be almost a magnitude greater than L10 . For these
reasons, it is difficult to know accurately when a certain bearing will fail. It is similar
to throwing a dice, as the outcome of a single throw is improbable to guess, while it is
easier to estimate the probability distribution of each outcome if the dice is thrown thou-
sands of times. This statistical nature of bearing life may arise from small differences in
material and production quality between batches. If the actual load, temperature, and
speed characteristics is dissimilar to the modeled conditions, the life estimation may get
quite inaccurate. Lifetime calculators are often available via the bearing manufacturers
web portal, and can be used to calculate L10 .

Without knowing the health indication of a machine, two maintenance practices can
be employed. With a reactive maintenance strategy, the components are replaced after
failure, which is ideal for non-critical appliances. A preventive maintenance strategy
aims to replace components after a certain amount of time in use. Such a strategy brings
increased maintenance cost, and high probabilities of early failure due to frequent machine
overhaul and run-in of new components.

A condition-based maintenance (CBM) strategy aims at identifying the health of com-

ponents, and replacing the ones that are worn out. This gives an economical advantage
as fewer maintenance stops are necessary, and fewer parts are replaced before the end
of their useful life. Additionally, knowledge of component health also improves system
reliability. Although visual inspection at regular intervals is a normal procedure to as-
sess the component health, this is difficult to perform on a rolling element bearing for
a couple of reasons. First, the machine must be stopped and disassembled, and second,
microscopic cracks have to be assessed on all rollers and raceways. Detecting tiny cracks
can be difficult even in a laboratory setting, which makes this procedure infeasible. In-
stead, the bearing health can be estimated by measuring the machine states while it is
running. Sensors are used to acquire physical states of the system, and signal processing
algorithms can be used to detect faults or estimate the level of wear. Different types of
sensors and algorithms can be used for this purpose, and the next section elaborates on
the state-of-the-art within condition monitoring (CM) systems for bearings.
1.2 State of the Art

1.2.1 Fault diagnosis

Initial bearing faults are typically small surface defects on either the raceway or roller.
When a roller passes the defect, an impulse of vibration is produced. This impulse excites
the resonance modes of the bearing and supporting structure. To detect this phenomenon,
a piezo-electric vibration accelerometer is fastened to the bearing housing, and the output
voltage is digitized with an analog-to-digital converter (ADC). The measured voltage
signal is proportional to the acceleration, e.g. with a rate of 100 mV/g. A bearing with
an outer race fault is shown in Fig. 1.1, and a CM system monitors the vibration signal.
The bearing has a surface fault on the outer race that generates vibration when rollers
pass, and an accelerometer responds to this vibration. The signal is afterwards processed
with a computer to generate results which an operator can analyze to determine the health
state.

Figure 1.1: Simplified schematic of a bearing condition monitoring system.

The bearing resonance frequency is typically in the thousands of Hz, while the cyclic
impact frequency is lower. Analyzing the kinematics of a bearing under the assumption
of no slip reveals the cyclic impact frequency. Four characteristic fault frequencies are
important for diagnosing a bearing:
 
fr d
Fundamental train frequency, FTF = 1 − cos φ (1.1)
2 D
Ball pass frequency outer race, BPFO = Zb FTF (1.2)
 
Zb fr d
Ball pass frequency inner race, BPFI = 1 + cos φ (1.3)
2 D
 2 !
fr D d
Ball spin frequency, BSF = 1− cos φ (1.4)
2d D
where fr is the shaft frequency, d is the roller diameter, D is the pitch diameter, Zb is the
number of rollers, and φ is the radial plane load angle. Cyclic impacts at either of these
frequencies indicate a damaged bearing. The bearing dimensions are shown in Fig. 1.2.

Figure 1.2: Bearing dimensions.

The radial load is spread un-evenly to the bottom half of the rollers, while the top
half gets larger clearance in the raceway. Due to the radial load zone, the load angle for
the rolling elements may differ, as the bottom half gets an angular contact point, while
the top half may slide on the edges due to axial load. In a ball bearing, the roller speed
(BSF) is dependent on the load angle, and thus the speed differs between rollers. While
the cage is moving at the mean speed of all rollers, some of them must slip on the raceway
if moving too fast or too slow. Therefore, there is slight variation in cyclic impact period,
and the bearing impact vibration can be modeled as second order cyclostationary (CS2)
[1]. On the other hand, deterministic vibration components originating from shafts and
gearboxes are phase-locked to the shaft. The bearing vibration can be separated from the
deterministic components by using a discrete/random separator (DRS). Time synchronous
average (TSA) [2], linear prediction filter [3], and self-adaptive noise cancellation [4] are
examples of DRSs. After separating the random components, the signal is denoted as
whitened.
The vibration amplitude is often demodulated to extract the cyclic frequencies, and
envelope analysis is one of the most successful methods for this purpose [5]. In the
early days of bearing vibration analysis, analog rectifiers and low-pass filters were used to
achieve the vibration envelope before digitizing the signal [6]. With faster digital technol-
ogy, the Hilbert envelope has emerged as a new alternative with several advantages over
analog rectification [4, 7]. To achieve the Hilbert envelope, the positive-sided frequency
spectrum is first inversely transformed to the time domain, yielding the analytic signal.
The analytic signal is the original signal plus an imaginary part, which is the Hilbert
transform of the signal. Then, the absolute value of the analytic signal is denoted the
Hilbert envelope. Finally, the envelope spectrum is analyzed to detect prominent peaks
at integer multiples of the bearing characteristic frequencies, i.e. harmonics.
The vibration signal should be band-pass filtered at the bearing resonance frequency
before enveloping to separate the signal of interest from other sources. An inherit advan-
tage of the Hilbert envelope is that the signal can be bandpass filtered in the positive-side
of the spectrum, which gives the analytic signal directly. There are several ways of identi-
fying the bearing resonance region. One approach is to track changes in the raw vibration
spectrum over time and place the filter over emerging resonance modes [4], but that re-
quires historic data of the system. More recent methods are based on the fourth statistical
moment, kurtosis, which has a high value for impulsive signals such as bearing vibration.
Spectral kurtosis is a tool for identifying frequency bands with high values of kurtosis
[8, 9, 10], which is assumed the optimal frequency band for Hilbert envelope analysis.
However, in case of non-Gaussian noise and large single impulses, spectral kurtosis may
fail to identify the optimal frequency region. Hence, other methods such as Protrugram
[11] and harmonic-to-noise ratio [12] have been proposed to mitigate those issues.
As stated in [4], the aforementioned methods are largely enough to solve most bearing
diagnostics cases. However, there are two operating conditions that can pose difficulties
when diagnosing a bearing, namely variable and low speed conditions. During variable
speed conditions (VSCs), fault impacts no longer occur at a fixed time interval, therefore
the cyclic frequencies are spread in the envelope spectrum. Computed order tracking [13]
largely resolves the problem by transforming the vibration signal from the time domain to
the shaft angle domain. This, however, requires knowledge of the shaft position acquired
either by a tachometer or an encoder. Alternatively, the vibration signal can be used to
estimate the shaft position by tracking frequency ridges in the time-frequency spectrum
[14, 15, 16, 17]. Cepstrum pre-whitening (CPW) is proposed in [18] to whiten the vibration
signal during VSC. Deterministic vibration components from shafts and gearboxes are
periodic, but not sinusoidal, and therefore form multiple harmonics in the frequency
spectrum. In the cepstrum, these harmonics form a single peak at the quefrency of the
deterministic component, while cyclostationary signals form no significant peaks at all.
By performing a series of liftering operations, the deterministic components are removed
by setting the whole real cepstrum to zero, except for at the zeroth quefrency. Afterwards,
the envelope spectrum of the order tracked whitened signal should reveal the bearing fault.
However, CPW causes spectrum normalization that increases the noise floor, which could
mask fault related harmonics. Methods for diagnosing bearings operating under VSC
should be further investigated.
Low speed conditions (LSC) pose a second challenge for bearing diagnosis. The impact
energy is depending on the fault size, bearing load, and shaft frequency. With a slower
shaft frequency, the impact energy is lower, which makes the impacts harder to detect
when the signal is contaminated by noise. To obtain a high-resolution envelope spectrum,
the signal must be measured for the duration of several shaft revolutions. At low speed,
the data collection time increases drastically, which makes a high impact on memory
footprint and required computational power. Some types of signal processing algorithms
are not feasible to use if the signal contains too many samples, such as iterative algorithms.
To overcome the issues of low impact energy, several studies have successfully applied
acoustic emission (AE) sensors, which are sensitive in the frequency range between 100
kHz and 1 MHz [19]. Such sensors require a high sampling rate ADC to properly sample
the signal, which increases the price of the CM system. Multiple studies have shown
that signal features such as AE energy [20], AE amplitude [21], AE counts [22], linear
prediction filter coefficient values [23], root mean square (RMS) and kurtosis [24] are
sensitive to the bearing fault at low rotational speed. Classification of faults using such
features is demonstrated in [25] with support vector machine (SVM) and relevance vector
machine (RVM). To reduce the amount of data required for fault diagnosis using envelope
spectrum, a peak-hold down-sampling technique is proposed in [26], and heterodyne signal
enveloping is demonstrated in [27]. A few studies are dedicated to using the vibration
signal for diagnosing faults at low speed as well [28, 29]. Most of the referred papers
conclude that AE is more suited to detect incipient bearing faults during LSC, compared
to regular accelerometers. However, high cost and complexity make AE infeasible for some
machines. Therefore, new methods for diagnosing a bearing under LSC with vibration
signal should be developed.
Detecting initial faults in a bearing is an important first step towards predictive main-
tenance. However, in many cases the machine should not stop immediately after detection
due to safety reasons or maintenance planning. Bearings can be in operation after the
initial fault, and the time until a complete failure is commonly referred to as the remain-
ing useful life (RUL). Prognostic techniques have been developed to estimate the RUL,
and the state-of-the-art within prognostics is elaborated on in the next section.

1.2.2 Prognostics

The bearing prognostics procedure can be divided into four major steps: Health indicator
(HI) estimation, initial fault detection, model prediction, and failure threshold (FT). A
simplified example of this procedure is shown in Fig. 1.3, and the steps are detailed
hereafter.
 4.0

3.5
Health indicator

3.0
Health indicator
2.5 Failure threshold
Model output
2.0
Future values Initial fault
1.5
Time left?
1.0

136 137 138 139 140 141 142 143 144

Days in operation

Figure 1.3: Simplified description of bearing prognostics procedure.

Health indicator estimation

The actual health state of a bearing can be assessed by quantifying the level of wear
in the bearing. This would require disassembly of the machine and examination of the
dismantled bearing, typically by a microscope. This is not feasible for most machines,
therefore the actual state is estimated by an HI calculated from sensor data, such as
vibration or AE. Generally, HIs can be categorized into physical HIs (PHIs) and virtual
HIs (VHIs) [30]. PHIs are generated from signals closely related to the physics of failure,
for example the vibration or AE signal. One of the most common PHIs is the RMS
[31, 32, 33], as it is very closely linked to the vibration energy. Other PHIs are the
kurtosis [34], and the energy ratio between the original signal and the linear prediction
residual [35].
VHIs, on the other hand, have no direct correlation with physics of failure. For exam-
ple, the combination of multiple PHIs using statistical methods or principle component
analysis (PCA) is a VHI. In [36], the Mahalanobis distance is used to fuse 14 PHIs into
a VHI, and in [37] a self-organizing map (SOM) is used to combine multiple PHIs. More
examples of HIs are given in [38].

Initial fault detection

Estimating the bearing RUL can be done after the first fault has developed, and the
HI starts to increase. In [31], a threshold on the kurtosis value is used to detect initial
degradation. A statistical alarm threshold on the HI itself is also used in some research.
For example, the mean and standard deviation (STD) of the HI is obtained with baseline
data, and the initial degradation starts when the HI passes its mean plus 3 times the
standard deviation [36, 39].

Model prediction

After passing initial degradation, the future trend of the HI can be predicted to estimate
the time until the FT is reached. Such algorithms can be divided into physics-based and
data-based methods [40]. Physics-based methods use models that are closely linked to the
failure mechanics of the component. Bearing vibration energy is often exponential as an
increasing number of cracks results in a greater amount of impacts. Therefore, variants
of the Paris-Erdogan crack law [37, 41, 42] have been used to track bearing degradation.
Data-based or statistical methods predict the future trend based on historic data alone,
without relying on physics of the fault. Some examples are the linear prediction model
[43] and the wiener process model [36].

Failure threshold

The prognosis objective is to estimate the time until the chosen HI reaches the FT, and
therefore the accuracy depends greatly on the FT value. Depending on the choice of HI,
it may be difficult to determine a suitable FT, especially for the more abstract VHIs.
In [40] it is stated that a challenge for rotating machinery prognosis is to define FTs for
newly designed HIs. Two ISO standards can be used for two PHIs, RMS and peak-to-
peak [40, 44], but are not applicable to other HIs. Consequently, most of the recently
developed algorithms require historic failure data to determine a suitable FT. This is often
not possible to achieve in practice due to lack of proper historic data when installing CM
equipment on a new machine. Therefore, new methods for RUL estimation that does not
involve historic failure data should be developed.

1.3 Motivation and Problem Statement

This project focuses on three topics under active research. The first topic is LSC bearing
fault diagnosis. Most of the identified works have focused on using AE to detect low-
energy impacts in the bearing. However, AE equipment is expensive, and accordingly
used mostly on large, critical machines. Developing new methods using vibration signal
to detect incipient faults in a bearing is desirable for LSC bearing diagnosis.
The second topic is bearing fault diagnosis under VSC in modern drive trains. While
computed order tracking and cepstrum pre-whitening have proven to be effective for
diagnosing bearings operating under VSC, there is considerable noise floor in the envelope
spectrum due to normalization of the entire vibration signal. Therefore, other methods
that can isolate the bearing vibration better should to be investigated.
Finally, the last topic is remaining useful life (RUL) estimation of bearings. The need
for historic failure data is common in all the referred literature. Once a CM system is
installed on a new machine, it is impractical to require historical failure data before RUL
can be estimated. Therefore, the last part of the thesis focuses on developing prognosis
algorithms that do not require historic failure data.

1.4 Contributions
This thesis is based on 6 papers that have been published or submitted for publication
during the project period.

1.4.1 Paper A: Accelerated Bearing Life-time Test Rig Devel-

opment for Low Speed Data Acquisition
Summary: Experimental data is required to verify the performance of new diagnostics
and prognostics algorithms. Some online repositories have vibration data captured during
VSC or constant speed operation. However, vibration data captured during LSC was not
available at the time of project initiation. Therefore, we designed and built an in-house
test rig to generate our own vibration data. The test rig can apply both axial and radial
loads to a small 40 mm bore-diameter 6008-type ball element bearing. The rig is struc-
turally designed to handle long periods of operation to wear out the bearing from healthy
to faulty stage. A permanent magnet synchronous motor with a planetary gearbox can
run the bearing at low speed down to 20 rounds per minute (rpm) and variable speed up
to 500 rpm. The test rig controller manages the entire test during the bearing life, and
handles regular sensor data logging, motor control, and bearing health check. In other
words, the operator does not have to be present and the accelerated life tests can run 24/7.

Contributions: The new test rig design allows for both axial and radial loads at the
same time, while being able to gather sensor data at low and variable speed settings. The
acquired data is used in all the appended papers in this thesis.

This paper has been published as:

A. Klausen, R. W. Folgerø, K. G. Robbersmyr and H. R. Karimi. Accelerated Bearing
Life-time Test Rig Development for Low Speed Data Acquisition. Journal of Modeling,
Identification and Control, 38(3):143-156, 2017. doi: 10.4173/mic.2017.3.4.
1.4.2 Paper B: Autonomous Bearing Fault Diagnosis Method
based on Envelope Spectrum
Summary: Analyzing vibration spectra for resonance regions and characteristic fault
harmonics is time consuming, and false alarms or erroneous conclusions may be drawn.
This paper presents an automatic method for bearing fault diagnosis. First, the vibration
spectrum is split into all resonance regions by use of bandpass filters. For each resonance
region, an automatic peak detection algorithm searches for harmonics of characteristic
fault frequencies in the envelope spectrum. For every identified harmonic, a fault score
level increases, and a higher score signifies greater confidence in fault detection. Experi-
mental results show how all three major fault types are detected during a single accelerated
lifetime test.

Contributions: In the direction of autonomous fault detection, the paper presents a

method for automatic detection of faults in a bearing.

This paper has been published as:

A. Klausen, K. G. Robbersmyr and H. R. Karimi. Autonomous Bearing Fault Diagno-
sis Method based on Envelope Spectrum. IFAC-PapersOnLine, 50(1):13378-13383, 2017.
doi: 10.1016/j.ifacol.2017.08.2262.

1.4.3 Paper C: Cross-correlation of Whitened Vibration Signals

for Low-Speed Bearing Diagnostics
Summary: Bearing fault diagnosis using vibration signal during LSC is a challenge due to
low impact energy. This paper presents a new method for extracting more useful informa-
tion of the vibration signal. First, the bearing vibration is isolated with a discrete/random
separator. Afterwards, the envelope spectrum is correlated with the whitened vibration
spectrum to combine useful information in both spectra. The reasoning is that fault
related harmonics may be identifiable directly in the raw vibration spectrum, and the
cross-correlation may improve strength of fault related harmonics. Experimental results
at 20 rpm verifies that the presented method significantly improves the diagnosis capa-
bilities compared to other methods reported in the literature.

Contributions: A new computationally efficient method for extracting more fault re-
lated harmonics of a vibration signal.

This paper has been published as:

A. Klausen and K. G. Robbersmyr. Cross-correlation of Whitened Vibration Signals for
Low-Speed Bearing Diagnostics. Mechanical Systems and Signal Processing, 118:226-244,
2019. doi: 10.1016/j.ymssp.2018.08.048.

1.4.4 Paper D: Multi-band Identification for Enhancing Bearing

Fault Detection in Variable Speed Conditions
Summary: This paper presents a new method for identifying bearing resonance regions
for bearings operating in VSC. It is known that order tracking makes shaft-speed de-
pendent vibration components de-spread in the frequency spectrum. At the same time,
bearing resonance is time dependent, and order tracking therefore causes a spread of res-
onance energy. The presented method exploits these phenomena by first normalizing the
vibration spectrum after order tracking by means of CPW. Afterwards, the inverse order
tracking is applied to de-spread and highlight resonance regions. Experimental results
from three test rigs show that the proposed method can highlight multiple resonance
bands, and that the envelope spectra contain numerous harmonics related to the fault.
Compared to two other methods reported in literature, the fault diagnosis score is higher.

Contributions: A new method to identify all resonance regions in a vibration signal

when operating under VSC.

This paper has been submitted as:

A. Klausen, H. V. Khang and K. G. Robbersmyr. Multi-band Identification for Enhanc-
ing Bearing Fault Detection in Variable Speed Conditions. Under review at Mechanical
Systems and Signal Processing.

1.4.5 Paper E: Novel Threshold Calculations for Remaining Use-

ful Lifetime Estimation of Rolling Element Bearings
Summary: ISO standard 10816-3 defines allowable RMS vibration levels in the position
and velocity domains which may be used as an FT for RUL estimation. However, initial
bearing faults cause high-frequency resonance vibration, and integration from acceleration
to velocity reduces the effect of this component. Therefore, an early degradation trend
is difficult to predict using the velocity- or position-based RMS. This paper presents a
method for transforming the RMS threshold to the acceleration domain. Experimental
results show that the acceleration-based RMS allows for longer prediction time, and that
the transformed failure threshold is valid compared to final values.
Contributions: Two new failure thresholds for vibration signals that does not require
any historic failure data.

This paper has been published as:

A. Klausen, H. V. Khang and K. G. Robbersmyr. Novel Threshold Calculations for
Remaining Useful Lifetime Estimation of Rolling Element Bearings. IEEE International
Conference on Electrical Machines (ICEM), Greece, 1912-1918, 2018. doi: 10.1109/ICEL-
MACH.2018.8507056

1.4.6 Paper F: Novel RMS Based Health Indicator used for Re-
maining Useful Lifetime Estimation of Bearings
Summary: The vibration RMS is a good health indicator due to its connection to the
signal energy. However, the trend is often non-stationary, which makes it difficult to pre-
dict future values. This paper presents a novel method for splitting the vibration signal
into multiple frequency bands for RMS calculations. Without using digital filters, the
proposed method splits the signal energy into multiple frequency bands using a single dis-
crete Fourier transform. The RMS is calculated for each band, and suitable RMS trends
for RUL estimation are identified online. Failure thresholds are calculated by extending
the method presented in Paper E. A particle filter combined with the Paris-Erdogan law
is applied to predict the RUL using suitable RMS trends. The method is verified with
experimental results from two test rigs.

Contributions: A novel method for subdividing the vibration signal into multiple fre-
quency bands for RMS calculations. A method for predicting remaining useful life without
using historic failure data.

This paper has been submitted as:

A. Klausen, H. V. Khang and K. G. Robbersmyr. Novel RMS Based Health Indicator
used for Remaining Useful Lifetime Estimation of Bearings. Under review at Mechanical
Systems and Signal Processing.

1.5 Outline
The rest of this thesis is divided into four main chapters. Chapter 2 contains brief de-
scriptions of the datasets used in this thesis, which includes the in-house test rig and
downloaded datasets from open repositories. Next, Chapter 3 is devoted to the contri-
butions within fault diagnosis during LSC and VSC. The chapter starts with a short
background on existing techniques before elaborating on the contributions made from
three papers written during the project. Chapter 4 involve the estimation of remaining
useful life for bearings. In particular, contributions from the two last papers are pre-
sented. Finally, Chapter 5 concludes the thesis and presents the reader with limitations
and possibilities for further work.
Chapter 2

Experimental test data

This chapter briefly describes the bearing vibration datasets used to generate experimental
results. Section 2.1 presents an in-house test rig that was built during the project period,
and the section is based on Paper A. Datasets from other sources have also been used
to verify the performance of the developed methods, and these datasets are described in
Section 2.2.

2.1 In-house test rig - Paper A

The test rig was designed and built at the University of Agder during the project period
to conduct accelerated lifetime tests of bearings from healthy to faulty conditions. A brief
description of the test rig is presented here, and more details are given in Paper A. Fig.
2.1 shows a schematic of the test rig, and Fig. 2.2 shows a picture of the finished test rig
in the laboratory.
A permanent magnet motor drives the shaft with 4 bearings. The shaft can be op-
erated with stable speeds between 20 and 500 rpm using a variable frequency drive and
a planetary gearbox. Continuous variable speed operation is also permitted using this
setup. A sealed 6008 type test bearing is installed in a custom housing on the left end of
the shaft. Two electric linear actuators apply axial and radial loads to the test bearing,
and Each actuator is connected to a lever arm to amplify the load, and load cells are used
as rotational points to measure the applied loads. To the right in Fig. 2.1, the red lever
arm is pushing on an axial bearing to generate up to 12 kN axial load. To the left, the
second red lever pushes the test bearing housing upwards, producing up to 12 kN radial
load. The two larger bearings are installed in the center to stabilize the shaft and bear
the radial loads. A piezo-electric accelerometer is mounted on the test bearing housing to
measure the vibrations. In addition, shaft movement is measured with an eddy current
proximity sensor, and the bearing temperature is determined with a PT100 element. The

15
 Figure 2.1: Schematic of the accelerated bearing life-time test rig.

proximity sensor data is acquired to verify the vibration analysis, as both sensors are
recorded at the same time. In this project, datasets from three run to failure tests have
been used. Table 2.1 shows the test conditions and which papers the datasets have been
used in.

Table 2.1: Run to failure experiments using the in-house test rig.
Paper Speed [rpm] Radial load [kN] Axial load [kN] Lifetime [106 rounds]

B 250 9 4 42.7
A 50 9 7 6
C 20 9 7 6
E 100 9 7 6
D 50±35 9 5 35.7
F 100 9 5 35.7

Experimental results used in Paper B originates from the first successful attempt to
damage the bearing, and this particular test was conducted before the planetary gearbox
was installed. It was, however, observed that the measurements acquired during LSC
below 100 rpm suffered from too large speed deviations due to variable friction and speed
controller overshoots. Therefore, the planetary gearbox was installed for the remaining
experiments to stabilize low speed operation. The second dataset used in Papers A, C
and E included low speed measurements down to 20 rpm. Variable speed operation was
also included in the last test, and datasets from this experiment were primarily used in
 Figure 2.2: The finished test rig in the laboratory at the University of Agder.

Papers D and F. The surface faults on two of the disassembled bearings are shown in Fig.
2.3.

Figure 2.3: Damaged bearings after two accelerated life time tests. (a) Rollers; (b) Outer
race; (c) Inner race.
2.2 Downloaded datasets
The datasets from three other sources have been used extensively in this project as a
second comparison with in-house datasets.

2.2.1 Variable speed datasets

The datasets collected by Mishra et al. [45] have been applied for variable speed fault
diagnosis. The test rig features a variable speed motor, bearings with artificially created
faults, and a small radial load of 5 kg. A laser vibrometer measured the vibration at
a sample rate of 50 kHz for ≈ 20 seconds for each dataset. The shaft encoder provided
readings of the shaft position during the measurements. Three types of faults were tested,
namely inner race, outer race and roller fault. During the measurements, the shaft speed
was adjusted manually between 600 and 900 rpm. These datasets are extensively used in
Paper D.

2.2.2 Accelerated lifetime datasets

The datasets generated by NSF I/UCR Center for Intelligent Maintenance Systems (IMS)
have been used for testing prognostics algorithms. These datasets were first used in [46]
and consists of three run-to-failure tests. The test rig has an induction motor running a
shaft with 4 bearings at a constant speed of 2000 rpm. Radial load of ≈ 26700 N is applied
to wear out the test bearings. For each test, the vibration at each bearing housing was
recorded for 1 second at a rate of 20480 Hz, every 10 minutes. An oil circulation system
lubricates the bearings, and a magnetic plug triggers a stop signal when enough metal
debris is accumulated. These datasets are used extensively in Paper E and F.

2.2.3 Constant speed datasets

Case Western Reserve (CWR) University has an open database with bearing fault test
data [47]. The test rig has an induction motor connected to a dynamometer as a load
torque. Bearings with pre-seeded faults are installed on either the drive-end (DE) or
fan-end (FE) for each test. For each fault and location combination, the vibration signal
was measured at DE, FE and the motor base (BA). Vibration was sampled at a rate of
12 kHz for FE faults, while measured at both 12 and 48 kHz for DE faults. There was
no radial load, but the torque load varied between 0, 1, 2 and 3 horse power (hp). While
unloaded, the shaft speed was 1797 rpm, and 3 hp load caused the induction motor to
run at a lower speed of 1730 rpm. These datasets are used extensively in Papers C and
E.
Chapter 3

Fault diagnostics

This chapter presents three new methods for fault diagnosis developed during the project
period. First, some relevant background used in the developed methods is given in Section
3.1. Afterwards, the three developed methods with results are described. Section 3.2 is
based on Paper B and introduce a method for automatic bearing fault detection. A
new method is developed in Paper C for extracting more fault-related information from
the vibration signal during LSC, and the method is described in Section 3.3. Paper D
demonstrates a new method for identifying resonance frequencies in the vibration signal
while operating under VSC, and Section 3.4 covers this method briefly. The papers are
appended to provide more detailed information about the proposed methods.

3.1 Theoretical background

The developed methods take advantage of background theory that is given in this section.
Section 3.1.1 explain order tracking which is used to transform the vibration signal from
time domain to shaft angle domain. In Section 3.1.2, three methods for separating random
components from the vibration signal are disclosed. Next, Section 3.1.3 shows how the
envelope spectrum is calculated. Finally, Section 3.1.4 explains how the envelope spectrum
can be analyzed to detect and classify bearing faults.

3.1.1 Order tracking

The vibration signal is normally sampled at a constant rate, for example 51200 samples
per second (Hz). A fixed number of samples between bearing fault impacts is recorded if
the shaft is rotating at a constant speed. However, if there are changes in the shaft speed
during the measurement, the impact frequency is no longer constant. Fourier techniques
assume stationary frequency components in the signal, and this assumption fails under
VSC. Therefore, the energy of cyclic components is spread over a larger frequency area,

19
which makes diagnosis difficult. Order tracking transforms the vibration signal from the
time domain to the shaft angle domain [13], which results in a fixed number of samples
between impacts. The shaft position or speed must be recorded along with the vibration
measurements to perform order tracking. An encoder or tachometer is normally used to
acquire the shaft position.
Let x(t) be the vibration signal sampled in the time domain, and xn is the same signal
discretized at time index n. The shaft position signal is given by θ(t) and assumed to be
sampled at the same rate as the vibration signal. A cubic spline is used to describe the
vibration signal between the discrete samples. This is given as

fot (θ) = interpolate(θ(t), x(t)) , (3.1)

where interpolate(θ(t), x(t)) is the desired interpolation function (cubic spline) that is used
to approximate the function fot (θ) = x(t). To order track the signal, vibration samples at
a fixed ∆θ period are calculated from fot (θ). The resulting order-tracked vibration signal
is xot (θ). The x-axis of an order tracked vibration spectrum is given in orders, where the
1st order is the same as the shaft frequency. More details on this method are given in the
appended Paper C, Section C.2.2.

3.1.2 Signal whitening

Bearing vibration is slightly random and is therefore modeled like a second order cy-
clostationary (CS2) signal as described in Section 1.2.1. Shaft and gearbox vibration
components are phase-locked to the shaft and are therefore deterministic. Random and
deterministic components can be separated using a DRS. For bearing fault diagnosis,
the random signal is retained for further analysis. A whitened vibration signal contains
mostly random, uncorrelated components such as white noise. Therefore, the random
signal output of a DRS is a whitened signal. In this project, three types of whitening
approaches are utilized.
The time synchronous average (TSA) is used to remove vibration components
that are phase-locked to the shaft. The vibration signal is split into segments of equal
lengths, each containing the vibration of one shaft rotation. Then, the mean of these
segments is the average vibration occurring per shaft revolution, i.e. the TSA. By sub-
tracting this average from of each shaft revolution of the original signal, the result is a
slightly whitened vibration signal xw (t). If the signal is order tracked, the angle syn-
chronous average is determined instead. More details are given in the appended Paper C,
Section C.2.3.
An autoregressive model is trained to predict the next sample based on a linear
combination of p previous samples. This model is formulated as
p
X
xn+1 = xw,n+1 − qj xn−j−1 , (3.2)
j=1

where xw,n is the n’th sample of the whitened vibration signal (model residual), and qj
is the j’th model parameter. Deterministic signal components can be predicted based
on previous samples, while random components are left in the residual. The model pa-
rameters are identified using the Yule-Walker equations which is a least-squares approach
involving the autocorrelation of the vibration signal [48, 49]. The model parameters
are determined for each signal, and the whitened vibration signal is extracted using the
trained model. More details are given in the appended Paper C, Section C.2.4.
Cepstrum pre-whitening [18] is an effective method that utilizes the cepstral do-
main to whiten the signal. Vibration components from shafts and gearboxes are periodic,
but not sinusoidal. Therefore, the components make up multiple harmonics in the fre-
quency spectrum. In the cepstral domain, these harmonics form a single peak at the
quefrency equal to the period of the deterministic component. By performing a series of
liftering operations, the deterministic components are mitigated by setting the whole real
cepstrum to zero, except for the zeroth quefrency [50, 51]. This is simplified by [18]
 
−1 F (x(t))
xw = F , (3.3)
|F (x(t))|

where F and F −1 are the forward and inverse discrete Fourier transform, respectively.
Detailed information are given in the appended Paper D, Section D.6.2 and [18, 50, 51].

3.1.3 Envelope spectrum

The cyclic frequencies are extracted by demodulating the whitened vibration signal with
the Hilbert envelope. The analytic signal is acquired by taking the inverse Fourier trans-
form of the positive-sided frequency spectrum. If Fp (x(t)) is the positive-sided frequency
spectrum, then the analytic signal is

xa (t) = x(t) + ixi (t) = F −1 (Fp (x(t))) , (3.4)

where xi (t) is the Hilbert transform of x(t). Afterwards, the envelope xenv (t) is calculated
with the absolute value as
xenv (t) = |xa (t)| . (3.5)

If, however, the signal is order-tracked, xenv (θ) is determined instead. Taking the Fourier
transform of xenv returns the envelope spectrum.
3.1.4 Fault detection
The envelope spectrum can be analyzed to detect faults in the bearing as a faulty bearing
will produce amplitude-modulated vibration signal. Fig. 3.1 shows the envelope in the
time and frequency domain for three different fault types. The envelope of resonance

Figure 3.1: Envelope in time and frequency domain for the three fault types.

vibration impulses is periodic, but not sinusoidal, resulting in several harmonics in the
spectrum. The side-bands appear due to amplitude modulation when the fault moves in
and out of the radial load zone. The envelope spectrum is examined for the frequency
domain characteristics in Fig. 3.1 to diagnose the bearing. If the spectrum does not
contain any prominent harmonics or side-bands related to the fault, the bearing is either
healthy, or the vibration signal contains too much noise.

3.2 Automatic fault detection - Paper B

Each fault impact causes resonance vibration in the bearing itself and the surrounding
structure. It is beneficial to isolate the resonance frequency vibration before making
the envelope spectrum to improve fault detection. However, identifying the resonance
frequency requires a detailed model of the system, which is not suitable to make for
all systems. Paper B presents a method for isolating resonance modes in the vibration
 signal and make an envelope spectrum for each of them. Manual analysis of several
envelope spectra is time consuming, therefore an automatic fault detection algorithm is
also presented in this paper.
A dataset from the in-house test rig is used to demonstrate the automatic fault de-
tection algorithm. During recording of this dataset, the bearing had an inner race fault,
and the motor ran at 250 rpm. The vibration signal is first order-tracked, and the TSA
is removed. The next step is to divide the vibration spectrum into resonance modes. A
low-pass filter is applied to the vibration spectrum amplitude, and the local minima define
points for dividing the spectrum into resonance bands. The vibration spectrum is shown
in Fig. 3.2, where the y-axis is the vibration amplitude, and the x-axis is the frequency
in Hz. In this example, 9 identified bands are investigated further.

1 Spectrum
0.08
Envelope
Amplitude [m/s2 ]

0.06 Band

0.04 2

0.02
3 4 5 6 7 8 9
0.00
0 2000 4000 6000 8000 10000
Frequency [Hz]

Figure 3.2: The vibration spectrum shown with the nine identified resonance mode bands.

The vibration signal is bandpass filtered at each identified band, and the envelope
spectrum is afterwards calculated. For each envelope spectrum, a harmonic search al-
gorithm is applied to identify prominent harmonics and side-bands related to bearing
faults.
The search algorithm focuses a single fault case at a time, for example an inner race
fault. Side-bands appear with 1 order spacing from the harmonic due to radial load
modulation. Fig. 3.3 shows the search for the first harmonic and side-bands in the third
frequency band.
The load angle φ normally makes up to 2 % difference in characteristic fault frequen-
cies. Therefore, prominent harmonics are searched for within a band of ±2 % width of the
investigated characteristic fault frequency. The maximum values within these bands are
chosen, and if the harmonic and one of the side-bands are greater than the threshold of
three times the noise floor (3N), the harmonic is prominent. A score value is increasing for
each prominent harmonic. The score is the ratio between the prominent harmonic and the
 0.04
Harmonic
Amplitude [m/s2 ]

0.03 Neg. Side

Pos. Side
0.02 3N

0.01

0.00
5 6 7 8 9
Orders

Figure 3.3: The first harmonic and its side-bands, and the prominence threshold given by
3 times the noise floor. The search width is given by the stapled lines.

threshold, multiplied with the harmonic number squared. The squaring causes the score
to increase quickly if multiple harmonics are identified. This value can be interpreted as
the probability of fault detection. In the third frequency band, the algorithm can identify
three prominent harmonics as shown in Fig. 3.4. This procedure is continued with the 8

0.05
Neg. Side
0.04
Amplitude [m/s2 ]

Harmonic
Pos. Side
0.03

0.02

0.01

0.00
0 5 10 15 20 25 30
Orders

Figure 3.4: The three prominent harmonics and side-bands identified in the third fre-
quency band.

other frequency bands shown in Fig. 3.2, and the score is summed to a single value for
this dataset.
The scores for each fault type can be monitored over time for easier visualization of
fault propagation. Fig. 3.5 shows the normalized fault scores (divided by their respective
maximum value) for the last 100 datasets captured at 250 rpm using the in-house test rig.
The normalized scores are small for the first 70 datasets, but the green trend increases to
its maximum value at dataset 71. A fault score of less than 10 is typically observed for
 1.0
RMS / max(RMS)
0.8
Normalized score

Inner ring–max = 923.58

0.6 Outer ring–max = 623.21
Rollers–max = 126.40
0.4

0.2

0.0
0 10 20 30 40 50 60 70 80 90 100
Dataset number

Figure 3.5: The fault scores over time for the last 100 datasets. The root mean square
(RMS) is displayed to compare the fault score with the vibration energy.

noise, therefore 126.40 indicates a high probability of a roller fault. The outer ring is also
damaged afterwards as seen by the increase of the blue trend during the next 20 datasets.
Finally, the inner race was also damaged near the end of the lifetime.
The fault scores are compared with the RMS of the vibration signal, which is a measure
of the mean vibration energy in the signal, given by
s
1 T
Z
RMS = x(t)2 dt (3.6)
T 0

At dataset 71, the RMS value is unchanged from previous values, while the roller fault
score increased to 126. Therefore, the proposed method can give earlier fault detection
compared to the RMS, and the fault is automatically classified.

3.3 Whitened cross-correlation spectrum - Paper C

The automatic fault score method presented in the previous section can be used for fault
diagnosis in many cases. During LSC, however, the bandpass filtered envelope spectrum
may not provide correct diagnosis information. For this reason, AE sensors have been used
in many applications to improve signal sensitivity and catch the low energy vibration.
Paper C presents a new method for extracting more information from the vibration
signal for fault diagnosis during LSC. First, the vibration signal is order-tracked, and
the TSA is removed. Afterwards, an autoregressive model with a limited number of
coefficients is trained to further whiten the signal. The result is a vibration signal that is
partially whitened as some deterministic components are still left in the signal. Bearing
faults can in some cases be detected in the vibration spectrum directly, without using
the envelope. Before the envelope spectrum was invented, this was the common method
for fault diagnosis. However, the bearing must be significantly damaged for prominent
harmonics to appear in the vibration spectrum. Additionally, the resonance frequency
needs to be low enough for identifying deterministic components [1].

The proposed whitened cross-correlation spectrum (WCCS) aims to combine the ad-
vantages of the vibration and envelope spectrum by correlating the useful information.
The envelope spectrum is more likely to contain fault related information but may contain
a high noise floor during low speed operation. The vibration spectrum may also show
small signs of the bearing fault, only slightly higher than the noise floor. When combining
these two signals, the information of both are fused in a single spectrum. To perform the
fusing, the cross-correlation of the vibration and envelope signal is calculated and used
for diagnosis analysis.

Vibration datasets from the in-house test rig captured during shaft speed of 20 rpm
are used to verify the proposed method. The first fault to develop during the accelerated
lifetime test is a roller fault, and Fig. 3.6 shows the result diagnosing the bearing using
the proposed method.

Fig. 3.6 (a) shows the spectrum of the partially whitened vibration signal. The yellow
lines show harmonic locations of 2 times BSF, and green stapled lines show the FTF
side-bands. The first 5 harmonics are not prominent compared to the noise floor, but
multiple harmonics of the side-bands are visible. On the contrary, the envelope spectrum
in Fig. 3.6 (b) has a low noise floor and a few prominent harmonic amplitude values.
The fault-related harmonics are only slightly higher than the noise floor, and therefore it
is difficult to diagnose the bearing based on this spectrum alone. The proposed method
combines these two spectra into the WCCS which is shown in Fig. 3.6 (c). The perceived
noise floor is reduced compared to the spectrum of the whitened vibration signal, and
the first five harmonics are now prominent compared to the noise floor. In addition, a
prominent peak at FTF is showing, which further indicates a roller fault.

A limitation with the WCCS is that the deterministic bearing fault components are
the result of a low-pass filter applied to the high-frequency resonance signal [1]. If the
resonance frequency is too high, the whitened vibration spectrum may not show signs of
the bearing fault at all. In this case, the envelope spectrum will be correlated with mostly
white noise, which should neither improve or decrease the diagnosis capability of WCCS.
Therefore, this method can be used even if the resonance frequency is high.
 ×10−5
6
Amplitude [m/s2 ]

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders
(a)
−3
×10
Amplitude [m/s2 ]

Harmonics of 2xBSF Side-bands

0.5

0.0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders
(b)
×10−2
1.5
Amplitude [m/s2 ]

1.0 FTF

0.5

0.0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders
(c)

Figure 3.6: The WCCS used to diagnose a roller fault. Harmonics and side-bands linked to
the roller fault are marked in all three subfigures. (a) Spectrum of the whitened vibration
signal xw ; (b) Spectrum of the envelope xenv . (c) Proposed WCCS.

3.4 Resonance frequency identification - Paper D

Whitening methods can significantly improve the bearing diagnosis by isolating random
signal components. However, white measurement noise is also random, and is therefore not
removed in a whitening process. CPW is a whitening method proposed in [18] for bearing
diagnosis under VSC. A brief description of CPW is given in Section 3.1.2. This method
normalizes the vibration spectrum to reduce the power of deterministic components. The
negative side effect is an increase in white noise energy, which could mask fault-related
 ×10−2
1
1
Vibration [m/s2 ]

Vibration [m/s2 ]
0
0

0.000 0.025 0.050 0.075 0.100 0.000 0.025 0.050 0.075 0.100
Time Time
×101 (a) (b)
5
Random
Shaft Speed [Hz]

Vibration [m/s2 ]
impacts
1.5
0

1.0
−5
0 5 10 15 20 0 5 10 15 20
Time Time
×104 (c) ×105 (d)
Amplitude [m/s2 ]

Amplitude [m/s2 ]

0.75
2
0.50
1
0.25

0 0.00
0.0 2.5 5.0 7.5 10.0 0 250 500 750 1000
Frequency [kHz] Orders
(e) (f)

Figure 3.7: Resonance band identification using the proposed method on a simulated
signal—Part 1. (a) shaft vibration; (b) bearing vibration; (c) shaft speed; (d) measured
vibration signal; (e) frequency spectrum of raw signal; (f) order spectrum.

components. Although this whitening procedure is convenient, the diagnosis can be more
accurate by isolating the resonance frequency instead.
Paper D presents a new method for identifying bearing resonance frequency when
operating under VSC. The method utilizes CPW and order tracking to highlight shaft-
speed invariant signal components. The procedure is shown as an example in Figs. 3.7
and 3.8 and explained hereafter. Deterministic shaft vibration harmonics are modeled as
27 sine wave components, and an underdamped second order model represents the bearing
resonance vibration. The shaft and bearing vibration are shown in Figs. 3.7 (a) and (b),
respectively. The shaft frequency ranges between 10 and 16 Hz as shown in Fig. 3.7 (c).
The complete vibration signal is given in Fig. 3.7 (d), where two extra random impacts
 2.0
Amplitude [m/s2 ]

Amplitude [m/s2 ]
Amplitude-filtered
1.5 4

1.0
2
0.5

0.0 0
0 250 500 750 1000 0.0 2.5 5.0 7.5 10.0
Orders Frequency [kHz]
×101 (a) ×10−3 (b)
Amplitude [m/s2 ]

Amplitude [m/s2 ]
Harmonic
1.0 1.0

0.5 0.5

0.0 0.0
0.0 2.5 5.0 7.5 10.0 0 10 20 30
Frequency [kHz] Orders
(c) (d)

Figure 3.8: Resonance band identification using the proposed method on a simulated
signal—Part 2. (a) order spectrum after CPW; (b) frequency spectrum after inverse
order tracking with a red line showing the amplitude-filtered signal; (c) amplitude-filtered
frequency spectrum raised to power of 5, with a suitable band-pass filter region marked
in red; (d) envelope order spectrum after band-pass filtration of original signal.

are added to verify the proposed methods ability to ignore noise. First, the signal is order-
tracked, which transforms the frequency spectrum as shown in Fig. 3.7 (e), into the order
spectrum in Fig. 3.7 (f). The maximum value of the low-frequency peaks is amplified
due to de-spreading of deterministic shaft-dependent components, and spreading of time-
dependent vibration components. Afterwards, CPW is applied to normalize the signal
as shown in Fig. 3.8 (a). Time-dependent signal components, such as bearing resonance
vibration, are afterwards de-spread when inverse order tracking is applied, as shown in
Fig. 3.8 (b).
A low-pass filter is applied on the spectrum amplitude to cluster single peaks into
resonance bands. By raising the low-pass filter output to a power of 5, the resonance
mode at 6000 Hz is isolated as shown in Fig 3.8 (c). The original signal is afterwards
band-pass filtered at this frequency, given by the red-stapled square, and the envelope
spectrum is shown in Fig 3.8 (d). In this spectrum, there are multiple harmonics related
to the fault, which shows that the proposed method can be used to identify prominent
resonance modes.
 ×104

Amplitude [m/s2 ]
1.0
Shaft Speed [Hz]

15
0.5
10
0.0
0 10 20 0 5 10 15
Time [s] Frequency [kHz]
(a) ×103 (b)
1.5
Amplitude [m/s2 ]

Amplitude [m/s2 ]
1 2 3
1.0
1
0.5

0
0.0 2.5 5.0 7.5 10.0 0 10 20 30 40
Frequency [kHz] Orders
×102 (c) ×101 (d)
1.0
Amplitude [m/s2 ]

Amplitude [m/s2 ]

Harmonic

0.5 1

0.0 0
0 10 20 30 40 0 10 20 30 40
Orders Orders
(e) (f)

Figure 3.9: Results from diagnosing an outer race fault on a variable speed dataset. (a)
shaft speed during measurement; (b) frequency spectrum of the raw vibration signal; (c)
frequency spectrum after using the proposed method, where three suitable band-pass filter
areas are marked; (d)–(f) envelope order spectra after band-pass filtration at the three
marked areas, respectively. Red triangles show identified harmonics related to the fault.

As an example, a variable speed dataset is used to detect an outer race fault using
the proposed method. A description of the dataset is given in Section 2.2.1. Fig. 3.9 (a)
shows the shaft frequency, which ranges between 10 and 15 Hz. The vibration spectrum
is shown in Fig. 3.9 (b), and it is dominated by low-frequency deterministic components.
By applying the proposed method, resonance modes are highlighted as seen in Fig. 3.9
(c). Three suitable frequency bands are chosen manually to diagnose the bearing. The
envelope spectrum is calculated for each band, and shown in Figs. 3.9 (d)–(f). In all
three envelope spectra, the automatic diagnosis method described in Paper B is used to
detect prominent harmonics. The results make it easy to diagnose the bearing correctly,
as there are multiple harmonics in each envelope spectrum. Selection of resonance bands
can be automated using the automatic diagnosis algorithm presented in Section 3.2.
Chapter 4

Prognostics

This chapter describes a new method for estimating the RUL of a bearing without utilizing
historical failure data. Section 4.1 is based on Paper E and briefly explains how the RMS
vibration FT can be chosen based on ISO standard 10816-3 [44]. Next, Section 4.2,
based on Paper F, describes how the vibration signal can be split into multiple frequency
bands for RMS calculation to achieve more suitable trends for RUL estimation. Finally,
the results of estimating RUL using a particle filter (PF) and the Paris-Erdogan law are
shown in Section 4.3.

4.1 RMS health indicator - Paper E

Estimating the bearing RUL is challenging as the degradation trend must be predicted.
More so, without historical failure data of the machine, it is difficult to determine a
suitable HI and corresponding FT that correlates well with the actual degradation. To
tackle this problem, the proposed methodology is to rely on industrial standards that can
be generalized for many applications. The RMS is closely related to the vibration energy
and can be used to assess the degradation of bearings because a larger number of defects
results in higher vibration energy. ISO standard 10816-3 [44], “Mechanical vibration—
Evaluation of machine vibration by measurements on non-rotating parts—Part 3”, defines
allowable vibration RMS values in the velocity domain, which can be used to set the FT.
The RMS of a signal x can be calculated with
v
u
u1 X N
RMS(x) = t x2 , (4.1)
N k=1 k

where N is the number of samples in the signal and xk is the discrete value at time index
k. Four classes are defined in ISO 10816-3: A–“new machine condition”, B–“Unlimited
long-term operation allowed”, C–“Short-term operation allowed” and D–“vibration causes

33
damage”. Vibration level D can be used as an FT because the vibration from continued
operation can damage other components such as pump seals and gears. Vibration is
commonly measured in the acceleration domain using accelerometers, because the piezo-
electric crystal responds directly to pressure changes, and generates a voltage signal. To
get velocity-based vibration, the signal must be integrated, either digitally or using analog
circuits. The latter is preferred for accurate integration in standalone devices that checks
only for high velocity-based RMS values. However, integrating the acceleration-based vi-
bration signal to velocity domain presents drawbacks for bearing RUL estimation: high-
frequency components are attenuated such that the effective bandwidth is reduced. Initial
bearing wear generate high-frequency vibration above 1000 Hz, and therefore contribute
insignificantly to the velocity-based RMS. The accelerated lifetime dataset introduced in
Section 2.2.2 is used to show the difference between velocity- and acceleration based RMS.
Let Ra = RMS(x) be the acceleration-based RMS and Rv = RMS(xv ) be the velocity-
based RMS, where xv is the vibration signal in the velocity domain. Fig. 4.1 (a) shows
Rv and the mean initial value R̄v .
An initial degradation alarm is set as the mean value R̄v plus 5 times the STD σv to
achieve a low probability of false alarm. In Fig. 4.1 (a), the alarm is triggered when there
is only a few hours left of the useful life, which gives little time to plan maintenance. Ra
shown in Fig. 4.1 (b), is different, and the initial degradation alarm is triggered much
sooner compared to Rv . In this case, there is approximately 70 hours left of the actual
life. However, the FT must be defined for Ra to make it useful for estimating the RUL.
Paper E presents a new method for making an FT for Ra using ISO 10816-3 [44].
An analytic transformation that applies for any machine was not feasible to determine,
because a detailed model of the vibration signal is required for that E. The transformation
can instead be determined experimentally for each machine using baseline data. The ratio
between Ra and Rv is determined to transform the velocity-based threshold R̂v to the
acceleration-based threshold R̂a . The ratio is given by

Rr = R̄a /R̄v , (4.2)

where the mean RMS values during baseline measurements are used to determine the
ratio. Afterwards, the ratio is used to transform the threshold from velocity-domain to
acceleration-domain using
R̂a = Rr R̂v . (4.3)

This transformation basically dictates that Ra must increase by a scale of R̄v /Rv to reach
the FT.
ISO standard 10816-3 [44] defines the vibration levels for machines that are 15 kW
or larger. For 15 kW, the standard defines that vibration causes damage when Rv = 4.5
 5
Rv R̄v
4
RMS [mm/s]

3
σv = 0.147
R̄v + 5σv crossed
2 R̄v = 0.720

0
0 20 40 60 80 100 120 140 160
Time [hours]
(a)
8
Ra R̄a
6
RMS [m/s2 ]

4
σa = 0.014
2 R̄a = 0.756 R̄a + 5σa crossed

0
0 20 40 60 80 100 120 140 160
Time [hours]
(b)

Figure 4.1: Difference between RMS values calculated using velocity and acceleration
units. (a) RMS of the velocity signal; (b) RMS of the acceleration signal. The mean R̄
and standard deviation σ are estimated using the first 50 hours of data files.

mm/s. As the motor used in the test rig may not be as powerful, the threshold is slightly
reduced to R̂v = 4 mm/s. Using the baseline data in Fig. 4.1, the threshold is determined
using (4.3) as R̂a = 4.34 m/s2 . The RMS trends are re-drawn in Fig. 4.2 together with
the respective thresholds. R̂v is reached by the velocity-based RMS as seen in Fig. 4.2
(a), showing that the ISO standard threshold is useful for stopping the machine. The
acceleration-based threshold R̂a shown in Fig. 4.2 (b) is also reached near the end of the
useful life, which validates the transformation for this dataset. Paper E gives more details
on the method and also more practical examples using other datasets.
While Ra can be used as an HI for RUL estimation, the trend is not ideal for predicting
the future degradation level. The main reason is the oscillating behavior, which makes Ra
a non-monotonic trend. Ra is cropped in Fig. 4.3 to better show the oscillations. The red-
stapled line shows the mean degradation path, which is ideal to extract. Non-monotonic
trends render a challenge for determining optimal model parameters to predict future
 Rv R̂v = 4.0 mm/s
4
RMS [mm/s]

σv = 0.147 R̄v + 5σv crossed

2 R̄v = 0.720 0.7 hours left

0
0 20 40 60 80 100 120 140 160
Time [hours]
(a)
8
Ra R̂a = 4.34 m/s2
6
RMS [m/s2 ]

4
σa = 0.014 R̄a + 5σa crossed
2 R̄a = 0.756 73.5 hours left

0
0 20 40 60 80 100 120 140 160
Time [hours]
(b)

Figure 4.2: Comparison of the RMS threshold using velocity and acceleration units. (a)
Velocity-based RMS; (b) Acceleration-based RMS.

3.0
Ra
2.5
Mean trend
RMS [m/s2 ]

2.0

1.5

1.0

0.5
80 90 100 110 120 130 140 150 160
Time [hours]

Figure 4.3: Ra compared with the mean degradation trend to highlight oscillations.

values. Therefore, the mean degradation path should be extracted from the vibration
signal. The next section describes a new method for extracting more useful information
from the vibration signal by splitting the signal into multiple frequency bands.

4.2 Filter bank RMS - Paper F

The non-stationary oscillations of Ra can be mitigated by considering the RMS of the
components within a certain frequency band rather than the entire signal. The vibration
signal can be split into multiple components using a filter bank, where each component
contains a frequency sub-band of the input signal. This is achieved by iteratively passing
the signal through low- and high-pass finite impulse response (FIR) filters and decimating
each output signal with a factor of 2. The number of levels in the filter bank decides how
narrow each frequency sub band is. The frequency content for each band can afterwards
be used to calculate RMS. Digital FIR filters use convolution to filter the signal, and only
the overlapping part between the filter kernel and the input signal should be preserved
to avoid adding artifacts. Therefore, there is information loss in the filtering. A larger
filter kernel yields sharper frequency cutoff, but the information loss is greater per level
of filtering compared to a shorter filter.
However, the time domain filter bank not necessary to acquire to calculate the RMS
within different frequency bands. Paper F presents a novel method for splitting the
frequency spectrum into suitable bands for RMS calculations, without losing signal infor-
mation/energy in the process. This is achieved with a single discrete Fourier transform
(DFT), and the method is described in the following.
The RMS can be calculated based on the energy E of the signal, such as
r
E(x)
RMS(x) = . (4.4)
T
The signal energy can be calculated in both time and frequency domain with
n
X n
X
2
E(x) = |xi | ∆t = |Xi |2 ∆f , (4.5)
i=1 i=1

where ∆t is the time step between samples, Xi is the i’th DFT bin, and ∆f is the
frequency step. Let XL be the frequency bins between negative and positive 50% Nyquist
frequency, with the rest set to zero. Similarly, let XH contain the other high-frequency
bins, and the rest are zero. This arrangement represents a single level filter bank where
the frequency content is split in half. It is demonstrated in Section F.3.1 of Paper F that
the energy of the entire signal can afterwards be calculated as

E(X) = E(XL ) + E(XH ) . (4.6)

Eq. (4.6) shows that the total signal energy can be calculated from the spectrum bins
directly. Further, if the spectrum is split evenly into Nb bands, the signal energy and
RMS are calculated with
Nb
X
E(X) = E(X(i−1)nb +1:inb ) (4.7)
i=1
v
u Nb
uX
RMS(X) = t RMS(X 2
(i−1)nb +1:inb ) , (4.8)
i=1

where nb = n/Nb is the number of samples in the frequency band. For brevity, the RMS
of a frequency band i is calculated with

Ri = RMS(X(i−1)nb +1:inb ) . (4.9)

To acquire an FT for each Ri , the conservation of RMS must be conserved as indicated

in (4.8). By substituting the RMS values for the respective thresholds, (4.8) becomes
v
u Nb
uX
R̂a = t R̂i2 , (4.10)
i=1

where R̂i is the FT for Ri . A simple approach for solving (4.10) is to let R̂i ∀ i ∈ [1, Nb ]
be equal. However, some energy bands may have larger mean values than others, and
therefore this FT may not serve the purpose very well for estimating RUL. Instead, the
assumption for solving (4.10) is that white noise is spread across the entire DFT, and that
the final value of each Ri is proportional to the noise standard deviation. In this case,
the FT is set as
R̂i = µi + mσi , (4.11)

where µi and σi are the mean and standard deviation of Ri during baseline measurements.
After combining (4.10) and (4.11), and squaring the equation, m can be determined by
solving the quadratic formula
Nb
X Nb
X Nb
X
R̂a2 =m 2 2
(σi ) + m 2µi σi + (µi )2 , (4.12)
i=0 i=0 i=0

for the maximum valued m. With m and baseline vibration measurements, the thresholds
for all Ri can be determined with (4.11)
Fig. 4.4 shows four RMS trends generated by frequency sub-bands of the previously
described dataset. R15 in Fig. 4.4 (a) is based on frequency content within [4480, 4800] Hz.
The trend is not monotonously increasing and is therefore not suitable for RUL estimation.
On the other hand, R3 in Fig. 4.4 (d), is very suitable for RUL estimation since the trend
is monotonic. Another difference between R15 and R3 is the time of triggering the initial
alarm, as R15 triggers it almost 25 hours before R3 . These observations suggest that
 1.5
R15 R11
1.5 R̂15 R̂11
RMS [m/s2 ]

RMS [m/s2 ]
1.0
1.0

µ15 + 5σ15 0.5 µ11 + 5σ11

0.5

0.0 0.0
0 40 80 120 160 0 40 80 120 160
Time [hours] Time [hours]
(a) (b)
1.00
R6 0.8 R3
0.75 R̂6 R̂3
RMS [m/s2 ]

RMS [m/s2 ]
0.6
0.50
0.4
µ6 + 5σ6 µ3 + 5σ3
0.25 0.2

0.00 0.0
0 40 80 120 160 0 40 80 120 160
Time [hours] Time [hours]
(c) (d)

Figure 4.4: A collection of RMS trends with failure FTs for the vibration dataset. (a)
R15 ; (b) R11 ; (c) R6 ; (d) R3 .

the energy of high-frequency components increase first, but are more volatile, while the
energy in low-frequency components changes last but rises more steadily. R11 and R6 in
Figs. 4.4 (b) and (c) are calculated from frequency bands between the two others. As is
observed here, R6 is more monotonic compared to R11 . In addition to these observations,
the FT is reached for all the trends in this dataset. The next section shows how suitable
RMS trends can be identified and used to estimate the bearing RUL.

4.3 Remaining useful life estimation - Paper F

Since it is unknown which of the RMS trends are suitable for RUL estimation, the most
monotonously increasing ones needs to be identified online. The Spearman rank correla-
tion coefficient ρ [52] assesses how well two signals are correlated in terms of their rank.
Here, the RMS trend and time are used as input, which returns a measure of how mono-
tonic the RMS signal is. By iteratively calculating the Spearman coefficient for all trends
with each new measurement, the most suitable trends for RUL estimation are determined
online. Bearing degradation is estimated using the modified Paris-Erdogan law [41] given
by [37]
da
= αaβ , (4.13)
dnc
where a is the crack size, nc is the number of cycles, and α and β are the model parameters.
This model is utilized to replicate the exponential degradation of a bearing [53].
A particle filter (PF) is applied to filter noise and predict the future Ri trend using the
modified Paris-Erdogan law. The model is re-formulated to a state-space representation
given by

a = ak−1 + αk−1 aβk−1 ∆nc
 k



αk = αk−1 (4.14)



h
k = ak + ν h ,

where αk−1 ∼ N (µα , σα2 ) is a random variable given by a normal distribution N , β is

constant, νh ∼ N (0, σh2 ) is measurement noise, and ∆nc is the number of cycles since last
update. Initial states of Np particles zkj ∀ j ∈ [1, Np ] at time index k are given with
" # " #!
Ri,k σh2 0
zkj ∼N , , (4.15)
µα 0 σα2

where Ri,k is the measured Ri value at time index k, and the weight of each particle is set
to wkj = 1/Np . Afterwards, the PF predicts new values for its states using the state-space
model, and particles with high probabilities of matching new measurements are given
a high weight. The weights describe a probability density function (PDF) that dictate
which parameter values make the state-space model match the measured samples best.
The RUL estimation of each particle is achieved by iteratively running the state-space
model in (4.14) until the FT is reached. Combining the particle RUL with their weight
creates the PDF of the RUL estimation from trend Ri .
When initiating a new PF, the parameters are first initialized to suitable values. Non-
linear least squares (NLS) is applied to determine the initial mean and constant param-
eters, i.e. Θ1 = (a1 , µα , β), where a1 is the initial crack size. The cost function is the
squared error between simulated values using (4.14) and Ri measurement samples. This
cost function is minimized by optimizing values in Θ1 . The measurement noise variance
is afterwards determined by calculating the variance of the baseline data, i.e. σh2 = σi2 .
There is also a process noise on α given by να ∼ N (0, σα2 ), which is the uncertainty of
the model. α is a function of the cyclic load on the bearing, which may change during
the degradation period, and therefore this process noise is added when predicting new
particle states. Therefore, the performance of the PF is dependent on the process and
measurement noise. The process noise is set to σα2 = µ2α to give the PF some margins
when filtering the RMS trend.
A new PF (P Fi ) is initiated for all Ri trends that achieve a Spearman coefficient
greater than ρ̂. In theory, all Nb RMS trends can therefore be utilized to estimate the
RUL, and the available information is weighed to make a single RUL decision. The weights
are based on the current Spearman coefficient value for the trend, and a higher value gives
a higher weight. The weight for P Fi at time index k is given by

Wi,k = ρ3i,k − ρ3L 1 − ρ3L ,



(4.16)

where ρi,k is the Spearman coefficient for trend Ri at time index k, and ρL is the lower
Spearman threshold. Non-monotonic trends get a zero weight as they cannot be predicted
by the Paris-Erdogan law. These weights are multiplied to the corresponding PFs particle
weights, and afterwards, the RUL PDF of all PFs are combined to a weighted PDF. The
weighted mean and 95% CI of this PDF is the estimate of the bearing RUL. More details
on this implementation is given in Section F.4 in Paper F.
The IMS dataset introduced in Section 2.2.2 is used to validate the performance of
this proposed method. The dataset is split into Nb = 32 bands, giving a bandwidth of
400 Hz. In addition, 30 measurement samples are used to optimize the initial PF values,
and the Spearman coefficient must be greater than ρ̂ = 0.9 to initiate a new PF. The
lower Spearman threshold is set to ρL = 0.7 to neglect non-monotonic trends. Fig. 4.5
shows the identified trends and predicted output of the initialized PFs. Each subplot row
is allocated for a single RMS trend number, given by the index i in the upper left corner.
The columns of subplots contain the following. Column 1 shows the identified Ri with
its FT R̂i , in addition to the median and 95% confidence interval (CI) prediction of the
initialized PF. Column 2 shows the median and 95% CI of the µα parameter in the model.
In addition, the median and 95 % CI of the predicted PF output are given in Fig. 4.6
when t = 130 to show convergence of PFs.
At t = 101 hours, the first monotonic RMS trend is identified, which is R15 shown
in Fig. 4.5 (a). The trend starts to increase linearly at the beginning, but at t ≈ 115
hours, the value increases quickly before gradually decreasing. The Spearman coefficient
ρ15 gets below 0.7 when t ≈ 130 hours due to the fluctuations in R15 , and P F15 will not
contribute to the weighted RUL after this point. µα is shown in Fig. 4.5 (b), and the
parameter eventually converges to the median value, signifying that the predicted trend
is not updated on new samples. The predicted trend is shown in Fig. 4.6 (a) for t = 130
hours. The 95% CI is much smaller compared to the initial PF prediction due to the
convergence of µα .
R14 and R12 are later identified as monotonic trends at t ≈ 108 hours, and the initial
PF outputs are shown in Figs. 4.5 (c) and (e) for the two trends, respectively. R14 and R12
 R̂i Ri Median 95% CI

×10−5
i = 15 ρ15 ≤ 0.7
1
RMS [m/s2 ]

1.0

µα
0
0.0
100 110 120 130 140 150 160 110 120 130 140 150
(a) ×10−5 (b)
i = 14 ρ14 ≤ 0.7
2.0 1
RMS [m/s2 ]

µα
1.0
0

0.0
110 120 130 140 150 160 110 120 130 140 150
(c) ×10−5 (d)
i = 12 ρ12 ≤ 0.7
RMS [m/s2 ]

1.0 2.5
µα

0.0
0.0
110 120 130 140 150 160 110 120 130 140 150
(e) −5 (f)
×10
1.0
i=6 1
RMS [m/s2 ]

µα

0.5
0
0.0
110 120 130 140 150 160 120 130 140 150
(g) ×10−5 (h)
i=3
RMS [m/s2 ]

2
0.5
µα

−2
0.0
120 130 140 150 160 130 140 150
Time [hours] Time [hours]
(i) (j)

Figure 4.5: Identified RMS trends with high Spearman coefficient, and output of the
corresponding PFs. Rows 1-5 indicate i = [15, 14, 12, 6, 3]. (column 1) Ri , FT R̂i , and
median and 95% CI of initial PF output; (column 2) µα over time for the initiated PF.
 R̂i Ri Median 95% CI

i = 15 i = 14
2
RMS [m/s2 ]

RMS [m/s2 ]
1
1

0 0
100 110 120 130 140 150 160 110 120 130 140 150 160
Time [hours] Time [hours]
(a) (b)
i = 12 1.0 i=6
RMS [m/s2 ]

RMS [m/s2 ]
1
0.5

0 0.0
110 120 130 140 150 160 110 120 130 140 150 160
Time [hours] Time [hours]
(c) (d)
i=3
RMS [m/s2 ]

0.5

0.0
120 130 140 150 160
Time [hours]
(e)

Figure 4.6: PF prediction of each identified RMS trend at t = 130 hours.

are similar to R15 , due to the RMS fluctuations. Additionally, the Spearman coefficient
for these trends go below 0.7 at the points shown by the vertical stapled lines. The future
PF predictions at t = 130 hours are shown in Figs. 4.6 (b), (c) for these two trends.
The median of both PFs pass the FT near end of life due to good parameter estimation
at the start. However, since the Spearman coefficient is lowered towards 0.7, the RUL
estimation won’t count in the weighted RUL decision.
A more promising trend R6 is later identified at t = 113 hours, as shown in Fig. 4.5
(g). The trend is more monotonic compared to the previous three ones, and the Spearman
coefficient never gets below 0.7. However, the trend is not increasing at a similar rate all
the time. When µα converges, as seen in Fig. 4.6 (h), the predicted trend increases faster,
as µα has increased from initial median value. The predicted PF output at t = 130 hours
shown in Fig. 4.6 (d) increases faster than the new measured samples. Therefore, the
 100
95% CI
90
Weighted mean
80
Remaining useful life [hours]

True RUL
70
60
50
40
30
20
10
0
100 105 110 115 120 125 130 135 140 145 150 155 160
Time [hours]
(a)
1.0

0.8
Trend weights

0.6 W15
W14
0.4 W12
W6
0.2
W3
0.0
100 105 110 115 120 125 130 135 140 145 150 155 160
Time [hours]
(b)

Figure 4.7: Weighted RUL decision. (a) weighted mean, 95% CI and true RUL of the
dataset; (b) weights for each PF output.

RUL is underestimated by this trend.

The last identified RMS trend is R3 shown in Fig. 4.6 (i). This trend is monotonic
and is very useful for RUL estimation. The initial PF prediction matches the future
samples well, which signifies a good RUL estimation by this trend. µα slowly converges
to a median value in Fig. 4.5 (j) which imply that the initial value was chosen close to
the final converged value. The predicted PF output at t = 130 is directed towards the
true RUL, and therefore the RUL estimation is accurate at this point.
The weighted mean and true RUL are shown in Fig. 4.7 (a), while the weights for
each PF are shown in Fig. 4.7 (b). The weighted RUL is fluctuating around the true
RUL throughout the estimation period, and changes for each new PF that is initialized.
At t = 130 hours, the estimated RUL is the most accurate compared to the true RUL.
Here, only W12 , W6 and W3 have significant weights. In Fig. 4.6, the predictions of
P F12 overestimates the RUL, P F6 underestimates the RUL, and P F3 matches the RUL.
Therefore, the weighted mean value falls between these three predictions, i.e. matching
the true RUL.
Afterwards, R12 get a low Spearman coefficient within the next 5 hours. Therefore, the
weights W15 , W14 and W12 are all turned to zero as seen in Fig. 4.7 (b) by t = 135 hours.
After that, the RUL is estimated by P F6 and P F3 only, and the estimation matches the
true RUL well.
A 3D plot in Fig. 4.8 shows how the weighed RUL PDF is changing over time. This
plot makes it easier to visualize how the weighted mean is determined. At the beginning,
the weighted mean RUL fluctuates due to several PFs being active, each with a high
variance in µα . After t = 135 hours, only P F6 and P F3 are left, and the weighted mean
RUL stabilizes between two peaks.

Figure 4.8: Weighted RUL PDF over time shown with the weighted mean RUL and true
RUL.
Chapter 5

Concluding Remarks

5.1 Conclusions
The focus area of this project is fault detection and remaining useful lifetime (RUL)
estimation of rolling element bearings. Several fault diagnosis and prognosis algorithms
are proposed in this thesis to tackle different challenges in bearing health monitoring:
automatic fault classification, low speed fault detection, variable speed fault detection,
and remaining useful lifetime estimation on new machines. A test rig was first designed
and built to make bearing failure sensor data at different operating conditions, such as
low and variable speed. With the in-house datasets and downloaded datasets from open
repositories, five algorithms have been proposed to tackle the aforementioned challenges.
With many machines in a plant to monitor, manual data analysis is cumbersome
due to a large amount of data. As a solution, an automatic diagnosis algorithm has been
developed to detect the most common bearing faults using the vibration signal. Vibration
data captured at 250 rpm was successfully diagnosed for three common fault types, namely
the rollers and two raceways.
During low speed conditions, a more robust algorithm is necessary compared to the
envelope spectrum. A new algorithm has been developed for extracting more useful
information from the vibration signal, by utilizing the deterministic components as well
as the envelope. This made it possible to detect bearing faults during low-speed operation
using a vibration accelerometer. With dataset from the in-house test rig, a roller fault
was diagnosed at 20 rpm using this new method. Compared to three other methods
reported in the literature, the proposed method resulted in stronger harmonics and more
side-bands related to the fault.
For variable speed conditions, a new method was developed for estimating the bearing
resonance frequency regions. The signal is bandpass filtered around each region, so the
envelope spectrum is easier to analyze for faults as there is less noise in the signal. In

47
most of the investigated vibration datasets, several resonance modes were identified using
this method. The envelope spectrum of each bandpass filtered resonance mode contained
multiple harmonics and side-bands related to the fault. Compared to two other methods
reported in the literature, the proposed method yielded the highest fault score using the
automatic diagnosis algorithm.
Estimating the remaining useful lifetime of a bearing on new machines without his-
toric failure data is the last challenge tackled in this thesis. Most prognosis algorithms
reported in the literature require historic failure data for parameter tuning and setting
the failure threshold. Such data may not be available in the industry, and therefore new
algorithms needs to be developed for these situations. A new method has been developed
for producing a general failure threshold for vibration root mean square (RMS) based
on ISO 10816-3. This method only requires knowledge of the nominal machine power
output, which is normally available in most plants. Experimental results from three test
rigs show that the proposed method generated failure thresholds that are surpassed close
to the actual time of failure.
The vibration RMS is, however, mostly non-stationary, and therefore fluctuates around
a mean trend. These fluctuations make it difficult to predict the RUL using a mathe-
matical model. A new approach for splitting the vibration signal into multiple frequency
bands is proposed to remove signal non-stationarity. The RMS is calculated for each
band, and monotonic RMS trends are identified using the Spearman coefficient. A par-
ticle filter algorithm is applied on each trend to predict future states by assuming the
degradation follows the Paris-Erdogan law. Experimental results from the in-house setup
and an online repository validate the performance of the proposed method.
In addition to the results presented in Chapters 3 and 4, several other datasets are
tested with the proposed methods in the appended papers. In particular, Papers D, E
and F include results using datasets acquired on the in-house test rig during low speed
operation. The presented results show that the proposed methods are also applicable
during low speed conditions.

5.2 Limitations
The automatic failure diagnosis algorithm proposed in Paper B requires a minimum vi-
bration signal length to acquire a fine frequency resolution. If the resolution is poor,
the algorithm has a trouble in distinguishing fault related peaks from surrounding noise.
Measuring the vibration signal for at least 30 rounds should be enough for the algorithm
to function properly.
The whitened cross-correlation spectrum (WCCS) proposed in paper C for low speed
bearing diagnosis is limited by machine resonance frequencies. An analytic derivation of
the vibration spectrum shows that the deterministic components related to bearing faults
are the result of a low-pass filter applied to the resonance vibration. If the resonance
frequency is too high, the deterministic components may not be distinguishable from
noise, and the proposed method may not improve the fault diagnosis.
In Paper D, a new approach for identifying resonance frequency bands is proposed. A
clear requirement is the vibration signal must be recorded during variable speed operation,
and that the shaft position must be acquired using an encoder. In addition, some of
the resonance bands may contain components from other sources than the bearing, and
therefore, several bands should be analyzed for bearing faults.
The failure threshold created in Paper E is based on the vibration energy, and therefore
other failure criteria, such as maximum vibration, will require a different failure thresh-
old. Further, the transformation from velocity-based to acceleration-based RMS failure
threshold is based on experimental data acquired on each machine and may therefore not
be optimal in all cases.
The particle filter implementation for remaining useful lifetime estimation in Paper F
is computationally heavy, and may not be usable on small, embedded computers.

5.3 Further Work

Vibration data from real machines in the industry should be used to verify the perfor-
mance of the proposed methods. In this thesis, vibration data has been acquired from
either accelerated life-time tests or bearings with artificial faults. Bearing damages in real
machines may have different characteristics compared to laboratory tests. Accelerated
lifetime tests make somewhat comparable vibration data, but the high load may make
the bearing break in different way than through normal fatigue over several years.
A vibration sensor has been used in this thesis as a cheaper alternative compared to
acoustic emission sensors. To further reduce the cost of the diagnosis system, the shaft
frequency may be estimated based on the vibration signal directly rather than using an
encoder. Algorithms for estimating this has been reported in the literature, but more
tests should be performed to verify the performance at low speed operation.
The method developed for low speed fault diagnosis should be validated on low speed
vibration data from another test rig, as a single source of vibration data is not enough
to validate the algorithm. In addition, it would be interesting to see if there is a limit to
how low the shaft speed can be before the fault is indistinguishable from noise.
The methods developed in Paper F for splitting the vibration signal into frequency
bands for RMS calculations can be used for estimating the RUL on new machines. How-
ever, the presented RUL estimation algorithm is computationally heavy, and more efficient
methods for utilizing the RMS bands should be investigated.
References

[1] R. B. Randall, J. Antoni, and S. Chobsaard. The relationship between spectral cor-
relation and envelope analysis in the diagnostics of bearing faults and other cyclosta-
tionary machine signals. Mechanical Systems and Signal Processing, 15(5):945–962,
2001. doi:10.1006/mssp.2001.1415.

[2] S. Braun. The extraction of periodic waveforms by time domain averaging. Acta
Acustica United with Acustica, 32(2):69–77, 1975.

[3] W. Wang and A. K. Wong. Autoregressive Model-Based Gear Fault Diagnosis. Journal
of Vibration and Acoustics, 124(2):172, 2002. doi:10.1115/1.1456905.

[4] R. B. Randall and J. Antoni. Rolling element bearing diagnostics—a tutorial. Me-
chanical Systems and Signal Processing, 25(2):485–520, 2011. doi:10.1016/j.ymssp.
2010.07.017.

[5] P. D. McFadden and J. D. Smith. Model for the vibration produced by a single point
defect in a rolling element bearing. Journal of Sound and Vibration, 96(1):69–82, 1984.
doi:10.1016/0022-460X(84)90595-9.

[6] M. S. Darlow, R. H. Badgley, and G. W. Hogg. Application of High-Frequency Reso-

nance Techniques for Bearing Diagnostics in Helicopter Gearboxes. Technical report,
US Army Air Mobility Research and Development Laboratory, 1974.

[7] M. Feldman. Hilbert transform in vibration analysis. Mechanical Systems and Signal
Processing, 25(3):735–802, 2011. doi:10.1016/j.ymssp.2010.07.018.

[8] R. Dwyer. Detection of non-Gaussian signals by frequency domain kurtosis estima-

tion. In IEEE International Conference on Acoustics, Speech, and Signal Processing,
ICASSP, volume 8, pages 607–610. IEEE, 1983.

[9] J. Antoni. The spectral kurtosis: a useful tool for characterising non-stationary signals.
Mechanical Systems and Signal Processing, 20(2):282–307, 2006. doi:10.1016/j.
ymssp.2004.09.001.

51
[10] J. Antoni. Fast computation of the kurtogram for the detection of transient faults.
Mechanical Systems and Signal Processing, 21(1):108–124, 2007. doi:10.1016/j.
ymssp.2005.12.002.

[11] T. Barszcz and A. JabLoński. A novel method for the optimal band selection for vi-
bration signal demodulation and comparison with the Kurtogram. Mechanical Systems
and Signal Processing, 25(1):431–451, 2011. doi:10.1016/j.ymssp.2010.05.018.

[12] X. Xu, M. Zhao, J. Lin, and Y. Lei. Envelope harmonic-to-noise ratio for periodic
impulses detection and its application to bearing diagnosis. Measurement, 91:385–397,
2016. doi:10.1016/j.measurement.2016.05.073.

[13] K. R. Fyfe and E. D. S. Munck. Analysis of computed order tracking. Mechanical

Systems and Signal Processing, 11(2):187–205, 1997. doi:10.1006/mssp.1996.0056.

[14] J. Urbanek, T. Barszcz, and J. Antoni. A two-step procedure for estimation of

instantaneous rotational speed with large fluctuations. Mechanical Systems and Signal
Processing, 38(1):96–102, 2013. doi:10.1016/j.ymssp.2012.05.009.

[15] T. Wang, M. Liang, J. Li, and W. Cheng. Rolling element bearing fault diagnosis via
fault characteristic order (FCO) analysis. Mechanical Systems and Signal Processing,
45(1):139–153, 2014. doi:10.1016/j.ymssp.2013.11.011.

[16] Y. Wang, G. Xu, Q. Zhang, D. Liu, and K. Jiang. Rotating speed isolation and
its application to rolling element bearing fault diagnosis under large speed variation
conditions. Journal of Sound and Vibration, 348:381–396, 2015. doi:10.1016/j.jsv.
2015.03.018.

[17] J. Shi, M. Liang, and Y. Guan. Bearing fault diagnosis under variable rotational speed
via the joint application of windowed fractal dimension transform and generalized
demodulation: a method free from prefiltering and resampling. Mechanical Systems
and Signal Processing, 68:15–33, 2016. doi:10.1016/j.ymssp.2015.08.019.

[18] P. Borghesani, P. Pennacchi, R. B. Randall, N. Sawalhi, and R. Ricci. Application

of cepstrum pre-whitening for the diagnosis of bearing faults under variable speed
conditions. Mechanical Systems and Signal Processing, 36(2):370–384, 2013. doi:
10.1016/j.ymssp.2012.11.001.

[19] (ISO). 22096 condition monitoring and diagnostics of machines - acoustic emission,
2007.
[20] M. Elforjani and D. Mba. Accelerated natural fault diagnosis in slow speed bearings
with acoustic emission. Engineering Fracture Mechanics, 77(1):112–127, 2010. doi:
10.1016/j.engfracmech.2009.09.016.

[21] N. Jamaludin, D. Mba, and R. H. Bannister. Condition monitoring of slow-speed

rolling element bearings using stress waves. Proceedings of the Institution of Mechan-
ical Engineers, Part E: Journal of Process Mechanical Engineering, 215(4):245–271,
2001. doi:10.1177/095440890121500401.

[22] M. Elforjani and D. Mba. Detecting natural crack initiation and growth in slow
speed shafts with the Acoustic Emission technology. Engineering Failure Analysis,
16(7):2121–2129, 2009. doi:10.1016/j.engfailanal.2009.02.005.

[23] N. Jamaludin and D. Mba. Monitoring extremely slow rolling element bearings: Part
I. NDT and E International, 35(6):359–366, 2002. doi:10.1016/S0963-8695(02)
00006-3.

[24] E. Y. Kim, A. C. C. Tan, J. Mathew, and B. S. Yang. Condition monitoring of low

speed bearings: A comparative study of the ultrasound technique versus vibration
measurements. Australian Journal of Mechanical Engineering, 5(2):177–189, 2008.
doi:10.1007/978-1-84628-814-2_21.

[25] A. Widodo, E. Y. Kim, J.-D. Son, B.-S. Yang, A. C. C. Tan, D.-S. Gu, B.-K. Choi,
and J. Mathew. Fault diagnosis of low speed bearing based on relevance vector machine
and support vector machine. Expert Systems with Applications, 36(3):7252–7261, 2009.
doi:10.1016/j.eswa.2008.09.033.

[26] T. R. Lin, E. Kim, and A. C. C. Tan. A practical signal processing approach

for condition monitoring of low speed machinery using Peak-Hold-Down-Sample al-
gorithm. Mechanical Systems and Signal Processing, 36(2):256–270, 2013. doi:
10.1016/j.ymssp.2012.11.003.

[27] B. Van Hecke, J. Yoon, and D. He. Low speed bearing fault diagnosis using acoustic
emission sensors. Applied Acoustics, 105:35–44, 2016. doi:10.1016/j.apacoust.
2015.10.028.

[28] Q. Xiong, Y. Xu, Y. Peng, W. Zhang, Y. Li, and L. Tang. Low-speed rolling bearing
fault diagnosis based on EMD denoising and parameter estimate with alpha stable
distribution. Journal of Mechanical Science and Technology, 31(4):1587–1601, 2017.
doi:10.1007/s12206-017-0306-y.
[29] C. Mishra, A. K. Samantaray, and G. Chakraborty. Rolling element bearing fault
diagnosis under slow speed operation using wavelet de-noising. Measurement, 103:77–
86, 2017. doi:10.1016/j.measurement.2017.02.033.

[30] C. Hu, B. D. Youn, P. Wang, and J. T. Yoon. Ensemble of data-driven prognostic

algorithms for robust prediction of remaining useful life. Reliability Engineering &
System Safety, 103:120–135, 2012. doi:10.1016/j.ress.2012.03.008.

[31] N. Li, Y. Lei, J. Lin, and S. X. Ding. An improved exponential model for predicting
remaining useful life of rolling element bearings. IEEE Transactions on Industrial
Electronics, 62(12):7762–7773, 2015. doi:10.1109/TIE.2015.2455055.

[32] Y. Lei, N. Li, and J. Lin. A new method based on stochastic process models for
machine remaining useful life prediction. IEEE Transactions on Instrumentation and
Measurement, 65(12):2671–2684, 2016. doi:10.1109/TIM.2016.2601004.

[33] Z. Huang, Z. Xu, X. Ke, W. Wang, and Y. Sun. Remaining useful life prediction for
an adaptive skew-Wiener process model. Mechanical Systems and Signal Processing,
87:294–306, 2017. doi:10.1016/j.ymssp.2016.10.027.

[34] Z.-X. Zhang, X.-S. Si, and C.-H. Hu. An Age-and State-Dependent Nonlinear Prog-
nostic Model for Degrading Systems. IEEE Transactions on Reliability, 64(4):1214–
1228, 2015. doi:10.1109/tr.2015.2419220.

[35] R. Li, P. Sopon, and D. He. Fault features extraction for bearing prognos-
tics. Journal of Intelligent Manufacturing, 23(2):313–321, 2012. doi:10.1007/
s10845-009-0353-z.

[36] Y. Wang, Y. Peng, Y. Zi, X. Jin, and K.-L. Tsui. A two-stage data-driven-based prog-
nostic approach for bearing degradation problem. IEEE Transactions on Industrial
Informatics, 12(3):924–932, 2016. doi:10.1109/TII.2016.2535368.

[37] Y. Lei, N. Li, S. Gontarz, J. Lin, S. Radkowski, and J. Dybala. A model-based method
for remaining useful life prediction of machinery. IEEE Transactions on Reliability,
65(3):1314–1326, 2016. doi:10.1109/TR.2016.2570568.

[38] J. Zhu, T. Nostrand, C. Spiegel, and B. Morton. Survey of condition indicators

for condition monitoring systems. In Annual conference of the prognostics and health
management society, volume 5, pages 1–13, 2014.

[39] X. Jin, Y. Sun, Z. Que, Y. Wang, and T. W. S. Chow. Anomaly detection and fault
prognosis for bearings. IEEE Transactions on Instrumentation and Measurement,
65(9):2046–2054, 2016. doi:10.1109/TIM.2016.2570398.
[40] Y. Lei, N. Li, L. Guo, N. Li, T. Yan, and J. Lin. Machinery health prognostics: A
systematic review from data acquisition to RUL prediction. Mechanical Systems and
Signal Processing, 104:799–834, 2018. doi:10.1016/j.ymssp.2017.11.016.

[41] P. Paris and F. Erdogan. A critical analysis of crack propagation laws. Journal of
basic engineering, 85(4):528–533, 1963. doi:10.1115/1.3656900.

[42] D. Xu, Q. Zhu, X. Chen, Y. Xu, S. Wang, et al. Residual fatigue life prediction of
ball bearings based on paris law and rms. Chinese Journal of Mechanical Engineering,
25(2):320–327, 2012. doi:10.3901/cjme.2012.02.320.

[43] Y. Qian, R. Yan, and S. Hu. Bearing degradation evaluation using recurrence quan-
tification analysis and kalman filter. IEEE Transactions on Instrumentation and Mea-
surement, 63(11):2599–2610, 2014. doi:10.1109/tim.2014.2313034.

[44] ISO. Mechanical vibration—Evaluation of machine vibration by measurements on

non-rotating parts—Part 3 (Standard No. 10816-3). ISO, 1998.

[45] C. Mishra, A. K. Samantaray, and G. Chakraborty. Rolling element bearing de-

fect diagnosis under variable speed operation through angle synchronous averaging of
wavelet de-noised estimate. Mechanical Systems and Signal Processing, 72:206–222,
2016. doi:10.1016/j.ymssp.2015.10.019.

[46] H. Qiu, J. Lee, J. Lin, and G. Yu. Wavelet filter-based weak signature detection
method and its application on rolling element bearing prognostics. Journal of Sound
and Vibration, 289(4-5):1066–1090, 2006. doi:10.1016/j.jsv.2005.03.007.

[47] Case Western Reserve University Bearing Data Center Website. 2015. URL: http:
//csegroups.case.edu/bearingdatacenter/home.

[48] G. U. Yule. On a method of investigating periodicities in disturbed series, with special

reference to Wolfer’s sunspot numbers. Philosophical Transactions of the Royal Society
of London, 226:267–298, 1927. doi:10.1098/rsta.1927.0007.

[49] G. Walker. On periodicity in series of related terms. Proceedings of the Royal Society
of London, 131(818):518–532, 1931. doi:10.1175/1520-0493(1931)59<277:OPISOR>
2.0.CO;2.

[50] N. Sawalhi and R. B. Randall. Signal pre-whitening using cepstrum editing (liftering)
to enhance fault detection in rolling element bearings. In COMADEM, Stavanger,
Norway, pages 330–336, May 2011.
[51] R. B. Randall, N. Sawalhi, and M. Coats. A comparison of methods for separation
of deterministic and random signals. International Journal of Condition Monitoring,
1(1):11–19, 2011. doi:10.1784/204764211798089048.

[52] C. Spearman. The proof and measurement of association between two things. The
American journal of psychology, 15(1):72–101, 1904. doi:10.1093/ije/dyq191.

[53] P. Rycerz, A. Olver, and A. Kadiric. Propagation of surface initiated rolling contact
fatigue cracks in bearing steel. International Journal of Fatigue, 97:29–38, 2017. doi:
10.1016/j.ijfatigue.2016.12.004.
Appendices

57
Paper A

Accelerated Bearing Life-time Test

Rig Development for Low Speed
Data Acquisition

Andreas Klausen, Roy Werner Folgerø, Kjell G. Robbersmyr and Hamid

Reza Karimi

59
This paper has been published as:

A. Klausen, R. W. Folgerø, K. G. Robbersmyr and H. R. Karimi. Accelerated

Bearing Life-time Test Rig Development for Low Speed Data Acquisition. Journal of
Modeling, Identification and Control, 38(3):143-156, 2017. doi: 10.4173/mic.2017.3.4.
 Accelerated Bearing Life-time Test Rig
Development for Low Speed Data
Acquisition
Andreas Klausen*, Roy Werner Folgerø*, Kjell G. Robbersmyr*, Hamid
Reza Karimi**

*University of Agder
Department of Engineering Sciences
Jon Lilletunsvei 9, 4879 Grimstad, Norway

**Politecnico di Milano
Department of Mechanical Engineering
Via La Masa 1, 20156 Milan, Italy

Abstract – Condition monitoring plays an important role in rotating ma-

chinery to ensure reliability of the equipment, and to detect fault conditions
at an early stage. Although health monitoring methodologies have been thor-
oughly developed for rotating machinery, low-speed conditions often pose a
challenge due to the low signal-to-noise ratio. To this aim, sophisticated al-
gorithms that reduce noise and highlight the bearing faults are necessary to
accurately diagnose machines undergoing this condition. In the development
phase, sensor data from a healthy and damaged bearing rotating at low-speed
is required to verify the performance of such algorithms. A test rig for per-
forming accelerated life-time testing of small rolling element bearings is de-
signed to collect necessary sensor data. Heavy loads at high-speed conditions
are applied to the test bearing to wear it out fast. Sensor data is collected
in intervals during the test to capture the degeneration features. The main
objective of this paper is to provide a detailed overview for the development
and analysis of this test rig. A case study with experimental vibration data is
also presented to illustrate the efficacy of the developed test rig.

A.1 Introduction
Rolling element bearings, or bearings for short, are necessary in rotating machinery to
reduce the degree of freedom of moving parts. A typical bearing is made of an inner-race
fastened to the shaft, a stationary outer-race, and rollers in between that transfers the
shaft load. The relative distance between each roller is kept constant by a cage. Bearings

61
are precisely manufactured to withstand the dynamic loads acting on the shaft. Metal-to-
metal contact is reduced to a minimum by lubricating the bearing with either oil or grease.
However, wear will always be present, even in perfectly lubricated bearings. As a roller
moves in and out of the radial load zone, the local lubrication is pressurized and causes
stress to the rollers and the raceways. After millions of rotations, this cyclic load wears
out the bearing components, and cause single or multiple faults which can occur in four
different locations: the rollers, the inner-race, the outer-race, and the cage; although cage
failures are uncommon. A worn bearing has increased friction, which in turn increases
the machine temperature, noise, and vibration levels. A completely worn bearing could
cause total system breakdown, injuries, and costly downtime.

It is important to monitor the condition of critical components, such as bearings, to

schedule maintenance when necessary and avoid breakdowns. This is especially important
for machines operating offshore, like windmill farms [1], as they are difficult to access and
typically has a low profit margin. For bearing condition monitoring, the vibration signal
is broadly accepted as a reliable data source. When a roller hits a fault, the impact energy
causes the bearing to vibrate at its resonance frequency. This vibration is measured using
an accelerometer, and is analyzed to determine the bearing condition. However, bearings
faults of low-speed machines are difficult to diagnose as the fault impact energy is lowered,
resulting in a low signal-to-noise ratio (SNR). Algorithms that reduce noise and highlight
low-energy impacts are important to correctly diagnose a machine operating at low speed.
Noise reduction examples in literature are autoregressive model filters [2], adaptive filters
[3], and neural networks filters [4], for instance. Highlighting low-energy impacts can
also be accomplished with a bandpass filter at the estimated bearing resonance frequency
using the fast Kurtogram [5], or blindly deconvolve the raw signal to preserve parts with
high kurtosis [6]. Analyzing proximity data to diagnose a faulty bearing has also been
explored in [7] where the authors state that a proximity sensor is more sensitive than
accelerometers during low-speed conditions. Acoustic emission sensors have also gained
prominence in the past decade, promising better sensitivity compared to accelerometers
during low-speed conditions. Apart from the sensor type selection, low-speed bearing
sensor data is necessary to assess the diagnostic accuracy of a newly developed algorithm.
Sensors can be used to acquire test data, and they can be mounted on industrial machinery
in use, or on specialized test rigs in research facilities. The latter setup type is the most
ideal for research as it grants full control of the test and the environment of the bearings.
However, for verification purposes, it may be advantageous to test algorithms on industrial
machinery as well.

In the literature, different bearing test rigs have been used to generate signal data.
Some designs [8, 9, 10, 11, 12] combine a torque source and a test bearing that is monitored
for faults. These rigs are unable to apply heavy loads to the bearing, and are thus unable
to naturally wear out the bearing within a feasible amount of time. Other test benches
also include heavy load capabilities [13, 14], but due to a low rotating speed, the bearing
did not wear out naturally. In these cases, the test bearing is installed with an artificially
introduced, pre-seed fault, often shaped like a hole or a line. Comparing sensor data from
a healthy and a damaged bearing is a typical scenario for testing a diagnostic algorithm.
One missing feature is the ability to follow the development of faults as it would happen
in a real scenario. Remaining useful life prognostics of a machine is often dependent on
the changes in sensor data as the amount of wear increases. To capture this trend, the
bearing may be run to failure on an accelerated life-time test rig. Their designs include
equipment for applying loads targeting the test bearing to accelerate its lifetime. The
test rigs in [15, 16, 17, 18, 19, 20, 21, 22] combine high speed and heavy load to achieve
faults in a reasonable amount of time. Unfortunately, their equipment does not handle
low-speed scenarios for logging data. Some other designs have the loading capacity to
wear out the bearing even during low-speed scenarios [23, 24], using only axial load.
The presented research gives a detailed overview of the development of a new test rig
used to acquire low-speed sensor data. From the literature review, the adopted method-
ology for this rig is to accelerate the bearing life-time using heavy axial and radial loads
at medium-to-high shaft speed. Sensor data is acquired in intervals during the test at
lower speeds, reaching down to 20 revolutions per minute (rpm). Required hardware and
software design for performing the test is also presented.
The rest of this paper is organized as follows. Section A.2 describes all specifications
for the test rig, and the different solutions that fulfill these requirements. A combination
of solutions are chosen and merged into a final design in Section A.3. Additionally, a case
study from an accelerated life-time test is also presented. Finally, conclusions are drawn
in Section A.4.

A.2 Methods
In this section, a short description of accelerated life-time testing is first given. Then,
specifications for the test rig are provided, and possible solutions to fulfill them are dis-
cussed.

A.2.1 Accelerated life-time testing

Accelerated life-time testing (ALT) is the process of speeding up the degradation of a
component by overloading it to uncover faults in a short amount of time. The test rig
described in this paper is designed to drastically reduce the lifetime of a bearing from
several years to a few weeks. Subjecting the test bearing with large forces reduces the
number of revolutions a bearing will hold. This comes from the following:
 3
C
L10 = , (A.1)
P

where L10 denotes the bearing life-time in million revolutions with a 90% confidence (hence
the lower case 10), C is the dynamic capacity of the bearing, and P is the weighted sum
of radial and axial loads in Newtons. By running the bearing with a heavy load at a
high speed, the life-time is accelerated to uncover faults relatively fast.

A.2.2 Specifications
Before designing the test rig, it is important to set up specifications describing the main
features. Further development of concepts is based on these specifications:

1. The driving motor must be able to rotate the shaft at medium-to-high speed to
accelerate the life-time of the bearing, and at a low speed (about 20 rpm).

2. The accessible radial and axial loads must be at least 10 kN. The load magnitude
must also be easy to change.

3. It must be possible to mount sensors measuring bearing data including: tempera-

ture, vibration, shock pulses, acoustic emission, and shaft radial movement.

4. The test bearing housing must be big enough to house a 6008 size rolling element
bearing with a 68mm outer diameter size.

5. The test rig equipment, software, and controller hardware must be designed for
unsupervised 24/7 accelerated life-time testing.

The reason for designing a test rig around a small 6008 size bearing is to minimize the
cost, space and complexity, as a larger bearing require heavier loads in addition to stronger
support structure. In the next subsection, solutions that fulfill the specifications are
presented.

A.2.3 Concept Development

Based on the specifications there are various solutions capable of achieving the desired
functions. A function tree, illustrated in Figure A.1, visualizes feasible options to achieve
the main function which is to accelerate the life-time of a bearing. The main function
is split into four sub functions that must be achieved, and for each there are several
Figure A.1: Test rig pre-design function tree. *An Open-Loop force feedback is a feed-
forward calculation of the assumed forces based on a particular input.

feasible solutions. The viability of each are discussed, with respect to the specifications,
to determine which should be kept in the design phase:
Torque source: It is important that the torque source can control the test bear-
ing shaft at low speed. If the shaft experiences too much unintentional speed fluctua-
tions around the setpoint, the sensor measurements may be hard to analyze. A high-
performance motor and frequency drive combination is necessary to fulfill this require-
ment. In addition, variable speed must also be realizable to simulate some machines
undergoing variable conditions, such as winches or windmills. Most motor types have the
possibility to be driven at both high and low speed. A hydraulic motor requires an exter-
nal pressure source to be operational during the entire accelerated life-time test. Using a
centralized hydraulic power unit (HPU) in a laboratory for 24/7 operation is not ideal due
to power consumption and equipment safety. Installing an HPU for the test rig increase
the complexity and cost of parts. Using an electric motor is more ideal since electricity
is almost always readily available. The different electric motor types are: brushed DC
motor, brush-less DC motor, induction motor, and permanent magnet motor. Brushed
DC motors may require more maintenance due to wear of the brushes, and therefore
these are not the most suitable. Brush-less DC motors and permanent magnet motors
are quite similar in function, as they are synchronous machines. Low-speed operation is
necessary, which require an encoder or resolver in the motor. Most induction motors are
controlled using sensor-less speed control to reduce cost, but a position-feedback is im-
portant for low-speed scenarios to properly control the motor, and log the shaft position
during measurements.
Transmit Rotation: The lowest speed requirement is 20 rpm, and it is improbable
that a motor coupled to the shaft directly will be able to rotate steadily at that low speed.
This is because the shaft may be slightly unbalanced or bent, resulting in a variable load
torque. Without a reasonably high inertia, the motor controller will most likely not react
fast enough to the changes in the load torque. Therefore, a gearing ratio of at least 5
is probably necessary to increase the load inertia, and allow the motor to operate at a
higher speed. This can be achieved with a belt drive, but a planetary gearbox is more
compact to install.

Apply Loads: Using a hydraulic linear actuator (Act.) to load the bearing allows for
a high amount of variable force, but require an HPU stationed nearby. Pneumatic linear
actuators have some of the same disadvantages, but cannot subject the bearing with high
forces due to low permissible air pressure. Hydraulic and pneumatic valves often leak over
time, which require the control system to reapply pressure consistently. A scissor jack may
create a high force, but varying the load requires manual operation and is not convenient.
Electric linear actuators have the advantage that they can self-lock after the desired force
is applied, therefore power is only necessary when the load is changing. However, they are
often limited by their maximum force. Considering this, a mechanical lever can be used
to amplify the load if necessary. It is important to allow for 24/7 operation, and as such
it is inconvenient to keep a hydraulic or pneumatic power unit continuously powered on.

Measure Loads: Actuators have losses from friction and power conversions, and
these losses may be non-linear, which makes them hard to estimate. Therefore open-loop
force calculations will include errors and are thus not ideal to use. Strain gauges can be
added to the structure to measure loads with good accuracy. However, such measurements
require high-resolution analog-to-digital converters, a model of the system, and a way to
calibrate the strain gauge signal based on a known reference. In addition, it may be
difficult to suppress unwanted noise due to temperature changes, or to decouple load
sources. Off the shelf load cells are pre-calibrated and temperature compensated by the
manufacturer, and are thus much easier to use and less error prone.

A.3 Results and discussion

In this section, solutions for the sub functions are chosen and designed. Later, all parts
are assembled on a suitable steel rig. The software and electronic hardware designs for
performing the accelerated life-time test are elaborated afterwards. Finally, a case-study
from an accelerated life-time test using the assembled test rig is presented.
A.3.1 The chosen solution
Based on the discussion of available components in Section A.2.3, the following choices
are made:
The available electric motor types should all be capable of controlling the shaft at
low-speed, assuming that some gearbox is installed, and the controller utilizes a motor
position feedback sensor. A permanent magnet motor (PMM) with a high resolution
built-in encoder, and a variable frequency drive with a high-performance controller was
chosen to drive the shaft. This combination can drive the motor up to 4000 rpm, and due
to the high efficiency of PMM, it is simples to keep cool during low-speed conditions as
squirrel cage induction motor fans typically produce little airflow during this condition.
A planetary gearbox is installed to transmit the rotation torque from the motor to the
shaft. It has a gearing ratio of 1:7, and is designed to be attached directly to the chosen
PMM.
Two electric linear actuators were chosen to produce the radial and axial loads. Each
have a loading capacity of 2.5 kN, which is too low considering the specifications. To
amplify the low force, two levers are designed to amplify them.
Pre-calibrated load cells are used to measure the loads due to difficulties of applying
strain-gauges. The chosen cells are shaped as bolts, and can measure the reaction force
in a rotational joint up to 20 kN. These will be used as hinges for the two levers, and
measure the reaction forces at the same time.

A.3.2 Test rig design

The design of the test rig starts with the shaft. The test bearing is installed on one end
of the shaft, and the gearbox is connected via a coupling on the other end. Radial and
axial loads are applied using linear actuators and levers. Aside from the test bearing, a
minimum of two support bearings are also needed to counteract the added radial forces.
To avoid wearing out the support bearings, durable ones with higher dynamic capacities
compared to the test bearing are necessary. Also, it is important that the support bearings
ignore the axial load on the shaft since it must propagate to the test bearing. Radial
bearings within split plummer block housings are suitable for these specifications, as this
combination give no axial support without installing extra equipment. The axial load
must originate from a stationary source, hence a thrust (axial) bearing is installed on the
shaft that allow for this load transfer. Four bearings are therefore installed on the shaft:
a test bearing, two radial support bearings, and a thrust bearing. Figure A.2 (a) shows
the acting forces on the shaft, and Figure A.2 (b) shows the bearing positions.
Stationary houses are required to transfer load to the rig structure, and to mount
Figure A.2: Chosen shaft design and bearing placements. (a) Counteracting forces on all
the bearings. (b) Position of the required bearings.

Figure A.3: Bearing housings as placed on the shaft.

sensors. Two suitable split plummer bearing housings are chosen to accommodate the two
larger radial bearings. Inside these housings, the two bearings are free to slide in the axial
direction. This is critical to transfer the axial load to the test bearing. The test bearing
housing is processed in a CNC machine from a steel block. Figure A.3 shows the bearing
houses on the shaft. The outer area of the test bearing housing can accommodate up to
three vibration accelerometers, or shock pulse sensors, via stud mounts. A temperature
sensor reaches the outer ring of the bearing via a hole in the housing. An eddy current
proximity sensor is stud mounted in a drilled hole in the split plummer bearing housing
close to the test bearing. Acoustic emission sensors can be placed using glue on any flat
surface on the housing. Figure A.4 shows the possible sensor layout on the test bearing
housing.
The axial load is added to the shaft via a concentric unit that is placed on the thrust
Figure A.4: The sensor layout on the test bearing housing: (a) Front of the housing with
the bearing in the middle and sensors around it. (b) Side of the test bearing housing with
the proximity sensor location.

bearing. Its curved area makes the load transfer from the lever easier. Figure A.5 shows
the concentric unit combined with the thrust bearing. The concentric unit is stationary
and acts as a bridge between the lever and the shaft.

From the specifications, at least 10 kN radial and axial loads must be available. Two
electric linear actuators, each with a maximum force output of 2.5 kN, are used to apply
the loads. Additionally, two levers are designed to amplify each linear actuator force to 17
kN. The high amplification was chosen to avoid loading each actuator 100% to achieve the
required force, and to establish some flexibility for testing stronger bearings in the future.
Figure A.6 shows the lever used to transfer axial load to the shaft. The two lengths L1
and L2 are determined in such a way that L1 /L2 = 17/2.5 = 6.8. The concentric unit
is located at the upper end of the lever, and the linear actuator at the bottom. The
lever features a curved surface that ensures proper transfer of the force to the concentric
unit. The revolution point A is a load cell bolt that can measure the reaction force in the
hinge. The axial force is calculated based on the geometry of the lever and the reaction
force feedback from the load cell. Section A.3.3 includes more details. The load cell is
 Figure A.5: Concentric unit housing the thrust bearing.

Figure A.6: Axial load setup.

supported by brackets bolted to the main steel structure. Figure A.7 shows the load cell
connected to the lever on the axial load setup.
A similar setup is used to produce the radial load which is applied to the test bear-
ing housing from underneath. A second linear actuator is connected to a lever which
amplifies the radial load. Figure A.8 shows the radial load setup. The lever follows the
aforementioned geometry ratio to amplify the linear actuator force to 17 kN by making
L3 /L4 = 6.8. A load cell bolt is used as the revolute joint to measure the reaction force
in the hinge. Point B on Figure A.8 pushes the bottom of the test bearing housing when
the actuator is retracting. The load cell is supported by two brackets as shown in Figure
A.9. As shown, the lever pushes directly on the bottom of the housing, and the force
propagates through the test bearing, to the shaft.
The added axial force propagates through the test bearing, to the test rig structure.
The radial force is transferred through the test bearing, to the support bearing housings.
 Figure A.7: Closeup details of the axial load setup.

Figure A.8: Radial load setup.

Considering this, the test bearing housing require structural support in the axial direction,
without interfering in the radial direction. Linear rolling element bearings on rails take
care of this behavior. Using four suitable linear bearings between the housing and a
support bracket, it is free to move in the vertical direction while supported in the axial
direction. Figure A.10 shows the housing connected to the linear bearings.

A.3.3 Axial and Radial Loads

In the previous subsection, the linear actuator and lever combinations producing axial
and radial loads were presented. Load cell bolts are used as lever hinges, and they
measure the reaction force in a single direction, as indicated by an arrow on the bolt. The
optimal measurement direction must be determined to measure the complete reaction
force. During installation, it is important that the load cell orientations are correct.
To this aim, the force equilibrium equations are established from a Free Body Diagram
 Figure A.9: Closeup details of the radial load setup.

Figure A.10: Bracket with linear rolling element bearings connected to the test bearing
housing.

(FBD). The FBD of the axial load setup is shown in Figure A.11. Here, three forces are
acting on the lever, disregarding gravity: FA is the axial force, R1 is the load cell reaction
force, and F1 is the linear actuator force. θ3 is the angular orientation of the load cell
necessary to measure the entire reaction force R1 , i.e. the direction of this force vector.
The force and moment equilibrium are given in (A.2) and (A.3) respectively.
       
cos θ1 cos θ3 FA 0
X        
F = F1   sin θ1  + R1  sin θ3  +  0  = 0 .
       (A.2)
0 0 0 0

   
− cos θ2 cos θ1
X    
M = L1  − sin θ2
 × F1  sin θ1 
   
0 0
     
cos θ2 FA 0
     
+ L2  sin θ2  ×  0  = 0
    
. (A.3)
0 0 0
P
where × represent the cross product in a right-hand coordinate system, M is the sum
P
of moments, and F is the sum of forces, and the three dimensions are x, y, and z
 Figure A.11: Free Body Diagram of the axial load setup.

respectively. Evaluating (A.3) yields the following equilibrium in the z-direction

F1 L1 (sin θ2 cos θ1 − cos θ2 sin θ1 ) − L2 FA sin θ2 = 0 . (A.4)

Afterwards, (A.4) is rearranged to isolate the axial force

 
F1 L1 sin θ1
FA = cos θ1 − . (A.5)
L2 tan θ2

The linear actuator force is isolated from the y-dimension of (A.2)

sin θ3
F1 = −R1 . (A.6)
sin θ1

Inserting (A.6) into (A.5) yields

 
R1 L1 sin θ3 1 1
FA = − . (A.7)
L2 tan θ2 tan θ1

The x-dimension in (A.2) is used to determine the hinge reaction force angle

F1 cos θ1 + R1 cos θ3 + FA = 0 . (A.8)

Inserting (A.6) and (A.7) into (A.8) yields

cos θ1
− R1 sin θ3 + R1 cos θ3
sin θ1
 
L1 1 1
+ R1 sin θ3 − = 0. (A.9)
L2 tan θ2 tan θ1
 Figure A.12: Free Body Diagram of radial load setup.

Afterwards, (A.9) is rearranged by dividing all terms by R1 cos θ3 , and isolating θ3

 
1
θ3 = arctan    . (A.10)
1 L1 1 1
tan θ1
+ L2 tan θ1
− tan θ2

θ3 is a constant angle independent of the axial load magnitude, as it is only dependent

on the geometry and orientation of the lever. During installation on the test rig, it is
important that the rotational orientation of the load cell matches the angle θ3 , otherwise
the measured reaction force is smaller than in reality. With the load cell in place, R1 can
be measured and recalculated into FA using (A.7).
The FBD of the radial load setup is shown in Figure A.12. Here, FR is the radial load,
R2 is the load cell reaction force, F2 is the linear actuator force, and θ6 is the angular
orientation of the load cell. The force and moment equilibrium for the FBD are given in
(A.11) and (A.12) respectively.
       
cos θ5 cos θ6 0 0
X        
F = F2 
 sin θ5  + R2  sin θ6  + −FR  = 0 .
       (A.11)
0 0 0 0

   
− cos θ4 cos θ5
X    
M = L3  − sin θ4
 × F2  sin θ5 
   
0 0
     
L4 0 0
     
+ 0  × −FR  = 0 . (A.12)
    
0 0 0
Evaluating (A.12) yields the following equilibrium in the z-direction,

L3 F2 (sin θ4 cos θ5 − cos θ4 sin θ5 ) − L4 FR = 0 . (A.13)

Afterwards, (A.13) is rearranged to isolate FR :

F 2 L3
FR = (sin θ4 cos θ5 − cos θ4 sin θ5 ) . (A.14)
L4
The linear actuator force is isolated from the x-dimension in (A.11) as

cos θ6
F2 = −R2 . (A.15)
cos θ5
Inserting (A.15) into (A.14) yields

R2 cos θ6 L3
FR = (cos θ4 tan θ5 − sin θ4 ) . (A.16)
L4
The y-dimension in (A.11),

F2 sin θ5 + R2 sin θ6 − FR = 0 , (A.17)

is used to determine the hinge reaction force angle. Inserting (A.15) and (A.16) into
(A.17) yields

R2 cos θ6
− sin θ5 + R2 sin θ6
cos θ5
R2 cos θ6 L3
+ (sin θ4 − cos θ4 tan θ5 ) = 0 . (A.18)
L4
θ6 is isolated in (A.18) by dividing all terms by R2 cos(θ6 ), and rearranging the result as
 
L3
θ6 = arctan (cos θ4 tan θ5 − sin θ4 ) − tan θ5 . (A.19)
L4

Here, θ6 is a constant angle independent of the radial load, and is only dependent on
the geometry and orientation of the lever. After installing the load cell bolt with the
orientation given by θ6 , the radial force FR is calculated using the measured hinge reaction
force in (A.16).

A.3.4 Complete test rig

Based on the sub function solutions in the previous sections, a complete design of the
test rig is created. The chosen solutions achieve the different sub functions; however,
the final design shows how the components are placed on a suitable steel structure. The
complete test rig is shown in Figure A.13. Details of all numbered components are given
in the following: To drive the shaft, a permanent magnet synchronous AC motor (1) is
Figure A.13: Complete 3D model with numbered components. The width of the test rig
is 350 mm.

chosen. It is powered by a variable frequency drive to realize variable-speed conditions. A

1:7 ratio planetary gearbox (2) is installed to aid the motor during low-speed conditions.
This combination enables the shaft to reach speeds up to 500 rpm and down to 20 rpm.
A flexible claw coupling (3) that does not transfer axial load is installed to provide
shaft rotation. A thrust bearing (4) is installed to transfer axial load to the shaft.
Two bearings (5) are used to counteract the radial load subjected to the test bearing,
and they are free to slide in the axial direction inside their housings. On the left end of
the shaft at position (6), a proximity sensor is mounted to measure the radial movement
of the shaft close to the test bearing.
The test bearing housing (7) is made of steel and is currently manufactured to house
a 6008-type ball bearing. Other similarly sized bearings with the same bore diameter
may be used, provided that a new housing is created if the outer diameter is different.
An accelerometer is placed on the side of the test bearing housing, and the outer ring
temperature is measured with a tip sensitive PT100 resistance temperature detector.
Other sensors are also available for installation as illustrated in Figure A.4.
The support bracket (8) is attached to the test bearing housing via linear ball element
bearings (9). These bearings allow for vertical load transfer to the test bearing, while
 Figure A.14: Electronic hardware connections.

they support the shaft in the axial direction. The bracket itself is bolted to the structure.
(10) and (11) are electric linear actuators. They can produce loads up to 2.5 kN in
their axial direction. Coupled with the levers (12) and (13), the system can theoretically
subject the bearing with 17 kN of load in both axial and radial direction.
Two load cells, (14) and (15), act as hinges for the levers while measuring the respective
reaction force up to 20 kN.
Machine damping feet (16) are used to reduce the vibration coming from external
sources, and their height can also be adjusted to level out the rig on slightly uneven
surfaces.

A.3.5 Controller hardware

Electronic hardware is used to control the active components, convert between analog
and digital signals, and store sensor data for further processing. A detailed overview of
the hardware is shown in Figure A.14. A computer with internet access is connected to
a dedicated controller with a built-in ARM CPU and Field Programmable Gate Array
(FPGA). The main task of the computer is to provide a local and remote interface to
the test setup software, and to store data during measurement intervals. A program on
the ARM CPU controls all the active components based on the feedback signals and test
configuration. The FPGA is responsible for acquiring and sending the analog signals
provided by the various IO via a databus.
The permanent magnet motor (PMM) is controlled by a motor drive with a built-in
regulator, and requires only a speed reference from the controller via a real-time EtherCat
connection. The shaft angular position is acquired from a quadrature encoder in the
motor. The electric linear actuators are powered by a 12 V DC motor drive, and the
reference signal to the drive is between 0-5 V. An input of 0 V corresponds to full retraction
force, 5 V to full push force, and 2.5 V for no force. The test-bearing temperature and
load cell reaction forces are acquired via an analog-in module. A 24-bit, high-frequency
±30 V IEPE-compatible analog input module measures the vibration in the accelerometer
attached to the test-bearing housing. The accelerometer produces a signal of 100 mV/g
in the linear range up to 10 kHz. In addition, a proximity sensor measures the radial
shaft movement in one direction. It has a measurement range of 1.1 mm and a resolution
of 18.5 nm, which makes it capable of detecting small changes in shaft position when the
bearing is damaged.

A.3.6 Load controller design

The electric linear actuators are controlled to apply the correct radial and axial loads to
the test bearing. The choice of control system is not trivial due to several factors: 1) The
linear actuator consists of a DC motor connected to a screw joint which moves in and out
when the motor rotates. The friction between the screw and DC motor is considerable,
and the actuator will therefore self-lock when not active. 2) The recommended duty
cycle is less than 10% to avoid overheating. Specifically, the manufacturer recommends a
maximum of 2 minute operation per 20 minutes. 3) Experimental testing on the finished
test rig revealed that the measured axial and radial loads varies with the shaft angular
position. The variation resembles a sine wave with a certain amplitude if the shaft speed
is constant. This variation is most likely caused by asymmetries on the shaft. It may be
slightly bent from the manufacturing process, or perhaps not perfectly circular. Due to
this disturbance, the actuators are difficult to control in a closed loop.
In any case, care must be taken not to overload the bearing or overheat the actuators.
A PID controller would attempt to cancel out the cyclic disturbance from the shaft, risking
overshoots and overheating. Using small gains would be a possibility, however that will
result in a poor bandwidth.
The self-locking ability is instead exploited to create a simple control system similar
to Sliding Mode Control. For each actuator setup, the error e between the reference load
F (ref ) and the actual load F is continuously checked. If this error is outside a threshold
T , the actuator is activated with maximum power (u = ±1) to reduce the error. Once
 Figure A.15: Control scheme for both linear actuators.

the error is within the threshold, the actuator is deactivated (u = 0). By configuring
this threshold slightly larger than the amplitude of the shaft disturbance, the actuator
is only powered in a short period after load reference change. The control scheme for
each actuator load setup is shown in Figure A.15. Here, R1 and R2 denotes the measured
reaction force in the axial and radial load cell, respectively. The block “To F ” determines
the axial or radial load using Equation (A.7) or (A.16), respectively.
It should be noted that this control scheme is not suitable for following a continuously
changing reference signal, i.e. a sine wave. For such an input, the controller will actuate
the system in steps when the error exceeds the threshold, resulting in a step-wise load
change.

A.3.7 Accelerated life-time test software design

An accelerated bearing life-time test may last for several weeks, or even months. There-
fore, it is impractical to manually make measurements at a defined interval, or to monitor
the test in person. To overcome this, an automated lab test program is developed. This
program controls all the active components such as the motor and the linear actuators,
and initiates sensor data logging at pre-defined intervals. The flowchart of the test pro-
gram is shown in Figure A.16. Initially, the operator starts by preparing nc configurations
(configs for short) that the program runs through. Each config contains specifications for:
the shaft speed, bearing load, and the duration. The duration is specified in either num-
ber of revolutions (revs) if sensor data is logged, or number of minutes if not. Afterwards,
the test may start from the first config, and after it is finished, the next config is loaded
and performed. Once all configs are finished, the program resets to the first one.
The test is stopped automatically if the vibration root mean square (RMS) exceeds a
 Figure A.16: Control scheme for the accelerated life-test.

preset threshold TRM S . The RMS is calculated using

v
u n
u1 X
VRM S = t V2, (A.20)
n i=1 i

where n is the number of samples in the dataset, and Vi denotes the i ’th sample in the
dataset V . Once the RMS reaches the threshold, the bearing is considered completely
worn out, and the test is automatically stopped. It is also possible to use different sources
for stopping the test such as motor torque or bearing temperature. Additionally, trends
of the RMS, motor torque, and bearing temperature is automatically uploaded to a secure
folder on the internet, and may be observed remotely by the operator. Further, the test
computer may also be accessed via a secure remote desktop connection in case the test
must be stopped/changed outside of working hours. Using this setup, the test rig is safe
to operate 24/7 without supervision.

A.3.8 Case study

As described in Section A.2.1 the test rig is used to perform accelerated life-time tests on
bearings. In addition to being able to run at high-speed to degenerate the bearing faster,
it also operates at slower speeds to obtain low-speed vibration data. For this accelerated
life-time test, the following options are set:

• The shaft is set to run at 500 rpm to wear out the test bearing.
 fb-kurt.2 - K max =40.6 @ level 7, Bw= 200Hz, f c =25100Hz
40
0
1
35
1.6
2 30
2.6
3 25
level k

3.6
20
4
4.6
15
5
5.6 10
6
6.6 5

7
0
0 0.5 1 1.5 2 2.5
frequency [Hz] 104

Figure A.17: Resulting kurtogram.

• The radial load is set to 9 kN, and the axial load to 7 kN. This results in a nominal
lifetime of L10 = 6 million revolutions according to the bearing manufacturer’s
online calculator [25].

• The bearing vibration is recorded at a sample rate of 51.2 kHz, while the bearing
temperature, motor torque, and shaft angular position are recorded at a sample rate
of 512 Hz.

• Sensor datasets are obtained every 10 minutes at 500, 250, 100, 50, and 20 rpm,
successively. The measurement duration is set to 100 revolutions.

• The automatic stop was set to an RMS threshold of 3 g (1 g = 9.81 m/s2 ).

A summary of the configuration loaded into the test software is given in Table A.1.
Two weeks and approximately 6 million revolutions later, the bearing is sufficiently
damaged that the RMS threshold is triggered. To diagnose the bearing for faults, the
datasets acquired at 50 rpm are used on this occasion. To determine the fault condition
in the bearing, a state-of-the-art method named the Fast Kurtogram [5] is employed. The
method decomposes the vibration data into frequency narrow-bands at different central
frequencies and widths. The kurtosis of each decomposed signal is calculated using

µ4
Kurt{x} = , (A.21)
σ4
 ×10−7
3.0
Harmonics of 2xBSF Side-bands
2.5

2.0
Amplitude

1.5

1.0

0.5

0.0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36
Shaft Orders

Figure A.18: Squared Envelope Spectrum after bandpass filtering.

22
Shaft speed [rpm]

20

18

0.0 0.5 1.0 1.5 2.0 2.5 3.0

Time [s]

Figure A.19: Shaft rotational speed at a reference speed of 20 rpm.

where µ4 is the fourth central moment, and σ is the standard deviation. A high kurtosis
value imply an impulsive signal which resembles bearing fault impact vibration. The
frequency narrow-band with the highest kurtosis is therefore chosen as the optimal filter,
and this band should contain the bearing fault impact vibration. Using this method,
the kurtogram is employed at every dataset acquired at 50 rpm shaft speed from the
accelerated life-time test. After the signal is filtered using the optimal filter, it should
contain the high-frequency resonance vibration from the bearing. To identify the cyclic
frequencies in the signal, the signal is demodulated. Using the Hilbert transform, the
complex-valued analytic signal is obtained. By computing the absolute value of this
analytic signal, the envelope is obtained. The envelope spectrum (ES) is the Fourier
transform of this envelope, which contain all cyclic impact frequencies. It has been shown
that the envelope should be squared to remove extra peaks in the spectrum [26], hence
resulting in the squared ES (SES). The resulting SES after bandpass filtration is manually
analyzed to identify prominent peaks at the characteristic fault frequencies for the present
bearing, which are given in Table A.2. In this table, BPFI is the ball pass frequency inner
 Table A.1: Operator settings for the test (nc = 6).
Conf. Speed FA FR Rec.? Duration

1 500 rpm 7 kN 9 kN No 10 min

2 500 rpm 7 kN 9 kN Yes 100 revs
3 250 rpm 7 kN 9 kN Yes 100 revs
4 100 rpm 7 kN 9 kN Yes 100 revs
5 50 rpm 7 kN 9 kN Yes 100 revs
6 20 rpm 7 kN 9 kN Yes 100 revs

Table A.2: Characteristic bearing fault frequencies in shaft orders.

BPFI BPFO FTF BSF

6.88 5.12 0.43 3.33

race, BPFO is the ball pass frequency outer race, FTF is the fundamental train frequency,
and BSF is the ball spin frequency. One shaft order is defined as the shaft speed, and ball
faults are shown at 2xBSF as there is an impact at the inner and outer race successively
during one spin. Equations for calculating these characteristic frequencies are given in
[27]. These characteristic bearing frequencies are of interest because they describe how
many times an incipient fault on a certain location is passed on each shaft revolution. The
earliest sign of bearing fault is identified after 5.37 million revolutions, and in this case, a
ball fault is progressing. The kurtogram is shown in Figure A.17, where the width of each
frequency band is given as levels in the y-axis, and the central frequency is given on the
x-axis. The kurtosis values are displayed as colors, and the maximum kurtosis of 40.6 is
identified at the central frequency 25,100 Hz with a bandwidth of 200 Hz (level 7). The
resulting SES after band-pass filtering is shown in Figure A.18. Here, integer multiples of
2xBSF (harmonics) are marked as yellow lines together with side-bands marked as green
stapled lines. The side-bands spaced apart by the FTF are shown due to the amplitude
modulation from the non-homogeneous radial load [27]. As there are multiple harmonics
with accompanying side-bands close to the theoretical 2xBSF, it is likely that the vibration
is caused by a ball fault.
The fast Kurtogram filtered SES from a dataset recorded at 20 rom was also analyzed,
however it did not show prominent peaks at this low speed. More advanced processing
methods are required for diagnosing the bearing during this low-speed working condition.
To check if the problem of diagnosing the bearing at 20 rpm is due to speed fluctuations,
the shaft speed for the 10 first revolutions is derived from encoder data, and shown in
Figure A.19. Accordingly, the speed fluctuations are within ±10%, which may deteriorate
the diagnostic capabilities of the fast Kurtogram slightly. Angular re-sampling should be
applied to reduce the blurring effect of the speed fluctuation [27].

A.3.9 Future work

In the case study in Section A.3.8, it was shown that early fault detection is possible
using the fast Kurtogram [5] at 50 rpm, however not at 20 rpm. The main purpose of
this case study, however, is to demonstrate that the test rig is capable of accelerating the
life-time of the bearing, and that the acquired signals may be used for fault diagnosis.
Future research will focus on new methods and algorithms for fault detection at the low
speed of 20 rpm.
After determining the initial fault, it is advisable to predict the remaining useful
life-time (RUL) using prognostics methods. These models are important to help schedule
maintenance and avoid unexpected breakdowns. Typically, extracted features from sensor
data is used to create a prognostic model, and a variety of models and methods have
already been suggested in the literature [28, 29]. However, there are still challenges to
overcome in designing a standardized scheme that can be applied on any machine [30].
Additionally, one apparent challenge in rotating machinery prognostics is to make accurate
estimations during variable speed or load conditions [31].
On the presented test rig, the bearing temperature and vibration, motor torque, and
shaft movement is measured for the whole life-span of the bearing during the accelerated
life-time test. Further, the working speed of the bearing is also changed between each
measurement, and may even be variable during the measurement period. In comparison to
bearing test rigs relying on pre-seeded faults, accelerated life-time tests allows for tracking
the changes in sensor data features as time/wear progresses, making RUL estimation
possible. The acquired sensor data will be used in future research to develop algorithms
which estimate RUL for machines operating under variable conditions.

A.4 Conclusion
An accelerated life-time test rig for rolling element bearings have been developed in this
paper. Using electric actuators and levers, the test bearing can be subjected with heavy
radial and axial loads. Additionally, utilizing a variable frequency drive, a motor, and a
planetary gearbox, both high and low speed working conditions are realizable. Sensors
are installed to measure physical data on the test bearing including vibration, shaft radial
movement, and temperature. Tests show that the test rig can be used to generate low-
speed vibration data during all bearing states ranging from healthy to completely worn.
Therefore, all the design criteria from Section A.2.2 are fulfilled. The list below shows
 √ √
conclusive remarks of the design specifications, where ( ) means fulfilled, (≈ ) means
partially fulfilled, and (×) means not fulfilled:
√
1. ( ) - Results show that the low speed of 20 rpm is achievable using a permanent
magnet motor and a planetary gearbox. There is a small fluctuation of ±2 rpm,
which may be corrected using digital angular re-sampling.
√
2. ( ) - Theoretically, the load setups can apply up to 17 kN of load in the radial and
axial direction.
√
3. ( ) - The flat edges of the test bearing housing allow for using all the requested
sensor types.
√
4. ( ) - The test bearing housing can house the 6008-type bearing with 68mm outer
diameter.
√
5. ( ) - Using suitable hardware and software, the test rig may be operated safely
24/7 without supervision.
REFERENCES

[1] S. T. Kandukuri, A. Klausen, H. R. Karimi, and K. G. Robbersmyr. A review of

diagnostics and prognostics of low-speed machinery towards wind turbine farm-level
health management. Renewable and Sustainable Energy Reviews, 53:697–708, 2016.
doi:10.1016/j.rser.2015.08.061.

[2] C. Junsheng, Y. Dejie, and Y. Yu. A fault diagnosis approach for roller bearings
based on EMD method and AR model. Mechanical Systems and Signal Processing,
20(2):350–362, 2006. doi:10.1016/j.ymssp.2004.11.002.

[3] J. Antoni and R. B. Randall. Unsupervised noise cancellation for vibration signals:
part I-evaluation of adaptive algorithms. Mechanical Systems and Signal Processing,
18(1):89–101, 2004. doi:10.1016/S0888-3270(03)00012-8.

[4] J. Zarei, M. A. Tajeddini, and H. R. Karimi. Vibration analysis for bearing fault
detection and classification using an intelligent filter. Mechatronics, 24(2):151–157,
2014. doi:10.1016/j.mechatronics.2014.01.003.

[5] J. Antoni. Fast computation of the kurtogram for the detection of transient faults.
Mechanical Systems and Signal Processing, 21(1):108–124, 2007. doi:10.1016/j.
ymssp.2005.12.002.

[6] R. Peled, S. Braun, and M. Zacksenhouse. A blind deconvolution separation of

multiple sources, with application to bearing diagnostics. Mechanical Systems and
Signal Processing, 19(6):1181–1195, 2005. doi:10.1016/j.ymssp.2005.08.019.

[7] P. Shakya, A. K. Darpe, and M. S. Kulkarni. Bearing diagnosis using proximity probe
and accelerometer. Measurement, 80:190–200, 2016. doi:10.1016/j.measurement.
2015.11.029.

[8] N. Sawalhi and R. B. Randall. Simulating gear and bearing interactions in the pres-
ence of faults: Part I. The combined gear bearing dynamic model and the simulation
of localised bearing faults. Mechanical Systems and Signal Processing, 22(8):1924–
1951, 2008. doi:10.1016/j.ymssp.2007.12.001.

87
 [9] R. Kumar and M. Singh. Outer race defect width measurement in taper roller bearing
using discrete wavelet transform of vibration signal. Measurement, 46(1):537–545,
2013. doi:10.1016/j.measurement.2012.08.012.

[10] D. Siegel, H. Al-Atat, V. Shauche, L. Liao, J. Snyder, and J. Lee. Novel method
for rolling element bearing health assessment—A tachometer-less synchronously av-
eraged envelope feature extraction technique. Mechanical Systems and Signal Pro-
cessing, 29:362–376, 2012. doi:10.1016/j.ymssp.2012.01.003.

[11] N. Sawalhi and R. B. Randall. Vibration response of spalled rolling element bearings:
Observations, simulations and signal processing techniques to track the spall size.
Mechanical Systems and Signal Processing, 25(3):846–870, 2011. doi:10.1016/j.
ymssp.2010.09.009.

[12] S. A. Niknam, V. Songmene, and Y. H. J. Au. The use of acoustic emission informa-
tion to distinguish between dry and lubricated rolling element bearings in low-speed
rotating machines. International Journal of Advanced Manufacturing Technology,
69(9-12):2679–2689, 2013. doi:10.1007/s00170-013-5222-4.

[13] T. R. Lin, E. Kim, and A. C. C. Tan. A practical signal processing approach

for condition monitoring of low speed machinery using Peak-Hold-Down-Sample al-
gorithm. Mechanical Systems and Signal Processing, 36(2):256–270, 2013. doi:
10.1016/j.ymssp.2012.11.003.

[14] N. Jamaludin, D. Mba, and R. H. Bannister. Condition monitoring of slow-speed

rolling element bearings using stress waves. Proceedings of the Institution of Mechan-
ical Engineers, Part E: Journal of Process Mechanical Engineering, 215(4):245–271,
2001. doi:10.1177/095440890121500401.

[15] B.-Q. Fan, K.-M. Lee, X.-P. Ouyang, and H.-Y. Yang. Soft-switchable dual-pi con-
trolled axial loading system for high-speed emu axle-box bearing test rig. IEEE
Transactions on Industrial Electronics, 62(12):7370–7381, 2015. doi:10.1109/TIE.
2015.2458303.

[16] B. Muruganatham, M. A. Sanjith, B. Krishnakumar, and S. A. V. S. Murty. Roller

element bearing fault diagnosis using singular spectrum analysis. Mechanical Systems
and Signal Processing, 35(1):150–166, 2013. doi:10.1016/j.ymssp.2012.08.019.

[17] J. B. Ali, N. Fnaiech, L. Saidi, B. Chebel-Morello, and F. Fnaiech. Application of

empirical mode decomposition and artificial neural network for automatic bearing
fault diagnosis based on vibration signals. Applied Acoustics, 89:16–27, 2015. doi:
10.1016/j.apacoust.2014.08.016.
[18] B. Zhang, C. Sconyers, C. Byington, R. Patrick, M. E. Orchard, and G. Vachtsevanos.
A probabilistic fault detection approach: Application to bearing fault detection.
IEEE Transactions on Industrial Electronics, 58(5), 2011. doi:10.1109/TIE.2010.
2058072.

[19] J. Yu. Bearing performance degradation assessment using locality preserving pro-
jections and Gaussian mixture models. Mechanical Systems and Signal Processing,
25(7):2573–2588, 2011. doi:10.1016/j.ymssp.2011.02.006.

[20] R. Li, P. Sopon, and D. He. Fault features extraction for bearing prognos-
tics. Journal of Intelligent Manufacturing, 23(2):313–321, 2012. doi:10.1007/
s10845-009-0353-z.

[21] H. K. Danielsen, F. G. Guzmán, K. V. Dahl, Y. J. Li, J. Wu, G. Jacobs, G. Burghardt,

S. Fæster, H. Alimadadi, S. Goto, et al. Multiscale characterization of White Etching
Cracks (WEC) in a 100Cr6 bearing from a thrust bearing test rig. Wear, 370:73–82,
2017. doi:10.1016/j.wear.2016.11.016.

[22] H. Qiu, J. Lee, J. Lin, and G. Yu. Wavelet filter-based weak signature detection
method and its application on rolling element bearing prognostics. Journal of Sound
and Vibration, 289(4-5):1066–1090, 2006. doi:10.1016/j.jsv.2005.03.007.

[23] W. Caesarendra, P. B. Kosasih, A. K. Tieu, C. A. S. Moodie, and B.-K. Choi. Con-

dition monitoring of naturally damaged slow speed slewing bearing based on ensem-
ble empirical mode decomposition. Journal of Mechanical Science and Technology,
27(8):2253–2262, 2013. doi:10.1007/s12206-013-0608-7.

[24] M. Elforjani and D. Mba. Accelerated natural fault diagnosis in slow speed bearings
with acoustic emission. Engineering Fracture Mechanics, 77(1):112–127, 2010. doi:
10.1016/j.engfracmech.2009.09.016.

[25] SKF. Bearing lifetime calculator, (Link accessible October 2017). URL: http://
webtools3.skf.com/BearingCalc/.

[26] D. Ho and R. B. Randall. Optimisation of bearing diagnostic techniques using sim-

ulated and actual bearing fault signals. Mechanical Systems and Signal Processing,
14(5):763–788, 2000. doi:10.1006/mssp.2000.1304.

[27] R. B. Randall and J. Antoni. Rolling element bearing diagnostics—a tutorial.

Mechanical Systems and Signal Processing, 25(2):485–520, 2011. doi:10.1016/j.
ymssp.2010.07.017.
[28] Y. Peng, M. Dong, and M. J. Zuo. Current status of machine prognostics in condition-
based maintenance: a review. International Journal of Advanced Manufacturing
Technology, 50(1):297–313, 2010. doi:10.1007/s00170-009-2482-0.

[29] F. Ahmadzadeh and J. Lundberg. Remaining useful life estimation: review. Inter-
national Journal of System Assurance Engineering and Management, 5(4):461–474,
2014. doi:10.1007/s13198-013-0195-0.

[30] J. Lee, F. Wu, W. Zhao, M. Ghaffari, L. Liao, and D. Siegel. Prognostics and
health management design for rotary machinery systems-Reviews, methodology and
applications. Mechanical Systems and Signal Processing, 42(1):314–334, 2014. doi:
10.1016/j.ymssp.2013.06.004.

[31] A. Heng, S. Zhang, A. C. C. Tan, and J. Mathew. Rotating machinery prognos-

tics: State of the art, challenges and opportunities. Mechanical Systems and Signal
Processing, 23(3):724–739, 2009. doi:10.1016/j.ymssp.2008.06.009.
Paper B

Autonomous Bearing Fault

Diagnosis Method based on
Envelope Spectrum

Andreas Klausen, Kjell G. Robbersmyr and Hamid Reza Karimi

91
This paper has been published as:

A. Klausen, K. G. Robbersmyr and H. R. Karimi. Autonomous Bearing Fault Di-

agnosis Method based on Envelope Spectrum. IFAC-PapersOnLine, 50(1):13378-13383,
2017. doi: 10.1016/j.ifacol.2017.08.2262.
 Autonomous Bearing Fault Diagnosis
Method based on Envelope Spectrum
Andreas Klausen*, Kjell G. Robbersmyr* and Hamid Reza Karimi**

*University of Agder
Department of Engineering Sciences
Jon Lilletunsvei 9, 4879 Grimstad, Norway

**Politecnico di Milano
Department of Mechanical Engineering
Via La Masa 1, 20156 Milan, Italy

Abstract – Rolling element bearings are one of the fundamental compo-

nents of a machine, and their failure is the most frequent cause of machine
breakdown. Monitoring the bearing condition is vital to preventing unex-
pected shutdowns and improving their maintenance planning. Specifically, the
bearing vibration can be measured and analyzed to diagnose bearing faults.
Accurate fault diagnosis can be achieved by analyzing the envelope spectrum
of a narrow-band filtered vibration signal. The optimal narrow-band is cen-
tered at the resonance frequency of the bearing. However, how to determine
the optimal narrow-band is a challenge. Several methods aim to identify the
optimal narrow-band, but they are not always precise. The bearing fault vi-
bration components are lost if the narrow-band is incorrectly chosen, thus
leading to an incorrect fault diagnosis. For on-line systems, it is critical that
bearing faults are diagnosed with a high degree of confidence. In this ar-
ticle, a method for analyzing multiple narrow bands is presented. Bearing
faults are detected autonomously by a narrow-band envelope spectrum-based
algorithm. This algorithm removes the need for manual spectrum analysis,
allowing operators to focus on more important tasks. Bearing fault vibra-
tion data from an accelerated life-test is used to verify the performance of
the proposed method. The proposed method accurately diagnoses the worn-
out bearing for three characteristic defect types and shows when one fault
propagates to a second one.

93
B.1 Introduction
Rolling Element Bearings (REBs, or bearings for short) are used in all kinds of rotating
machinery. They are designed to restrict the shaft rotational motion for transferring
loads to stationary housings. Bearings are worn out after exceeding a certain life-time
based on the rotational speed, the combined loads, and the bearing design. When a
bearing suddenly fails, the rapidly increasing vibration may damage other components,
causing a complete breakdown of the machine. Replacing a bearing before it fails is of
great importance to avoid breakdowns and expensive overhauls. The exact life-time of a
bearing cannot be perfectly predicted because the failure rate of a particular bearing is
based on statistics. In addition, changes in bearing vibration is only detectable during the
final stage of the life-time. Therefore it is useful to continually monitor the condition of
critical components such as bearings to prevent sudden machine breakdowns. Sensors can
be placed on the bearing housing to measure a physical quantity containing information
about the condition. The bearing vibration signal is strongly linked to the amount of
wear and is a suitable signal to use for bearing condition monitoring.
Roller impacts in a damaged bearing produce amplitude modulated ringing with the
carrier frequency equal to the resonance frequency of the bearing. The frequency of the
modulation waveform reveals the fault location in the bearing. As such, the vibration
signal must be demodulated to diagnose the fault location and severity. After demodu-
lation, the frequency spectrum can be analyzed to diagnose bearing faults. [1] used the
Hilbert-Huang transform to demodulate the vibration signal and diagnosed bearing faults
using the frequency spectrum. The bearing characteristic frequencies are derived from
the bearing design and kinematics under the no-slip assumption. [2] calculated these
characteristic frequencies and diagnosed the damaged bearing by observing the envelope
spectrum for peaks at these frequencies. The vibration signal contains components that
are not linked to impacts in the bearing, but from other sources such as mass unbalance,
misalignment, and others. These noisy components can be removed by applying a band-
pass filter on a narrow-band around the resonance frequency of the bearing. The result
is a much clearer frequency diagram to analyze. However, it is difficult to determine the
optimal narrow-band analytically, but several attempts in the literature have been made
to determine it experimentally. [3] introduced the Fast Kurtogram which uses the kurto-
sis to determine the optimal narrow-band. The vibration signal is band-pass filtered at
various narrow-bands, and the kurtosis value determines the extent of impulsiveness in
the filtered data. The filter specification returning the highest kurtosis can be selected
as the optimal narrow-band for further envelope spectrum analysis. However, while there
are cases when the kurtogram detects the optimal narrow-band, there are also cases when
it fails. [4] presented a new method inspired by the kurtogram that attempts to overcome
some of the drawbacks of the fast kurtogram. It is named the Protrugram and it calcu-
lates the kurtosis of the envelope spectrum of the band-pass filtered signal rather than the
time-signal. The advantage is the ability to detect transients with a small signal-to-noise
ratio. [5] presented a new feature named envelope harmonic-to-noise ratio (EHNR) that
can replace the kurtosis calculation in the fast kurtogram. Its advantage over the kurtosis
is that it is not sensitive to random single impulses and is therefore more robust.
The methods existing in literature aim at identifying the optimal narrow-band. How-
ever, incorrect selection leads to filtering out components related to the bearing fault.
Also, the bearing fault impacts may excite more than one resonance frequency in the
system, and thus multiple narrow-bands should be investigated on a single vibration
dataset. In this article, we present a new method for diagnosing faults in a bearing, based
on narrow-band envelope spectrum analysis. Initially, the entire vibration frequency spec-
trum is divided into multiple narrow-bands and band-pass filtered. Next, the envelope
spectrum of each filtered signal is realized by applying the Hilbert-Huang transform and
the Fast Fourier transform. Each spectrum is afterwards analyzed to diagnose the bear-
ing for faults. Since there are multiple spectra to analyze, the task becomes impractical
to perform manually. Therefore, a method for autonomous bearing fault diagnosis is
presented in this article. The method is based on automatic envelope spectrum analy-
sis to diagnose the bearing for faults. Only the shaft speed and the four characteristic
bearing fault frequencies derived from the bearing design, are required prior knowledge.
An algorithm searches the envelope spectra for harmonics linked to a characteristic fault
frequency and scores the dataset based on the number of harmonics identified and their
prominence compared to surrounding noise. This score is monitored over time to detect
condition changes in the bearing and determine the fault location. In this article, the
proposed method is presented along with results from diagnosing faults in a damaged
bearing. The test bearing is worn naturally over time during an accelerated life-test,
and three characteristic faults are identified in the de-assembled bearing. The presented
method accurately identifies all three fault types, and score trends also show when one
fault progresses to a second one.

B.2 Methods

B.2.1 Experimental Test Setup

The vibration signal is collected on a bearing test bench. The test bearing is a 6008
type with a dynamic and static load rating of 17.8kN and 11kN , respectively. The test
bearing is naturally worn over time during an accelerated life-time test. To reduce the
bearing life-time, radial and axial loads of, respectively, 9kN and 4kN are applied to the
test bearing. During the test, the vibration signal V (t) and encoder position signal θ(t)
is measured every 30 minutes at a reference speed of θ̇(ref ) = 250 revolutions per minute
(rpm). The sampling frequency is set to 51.2kHz and the measurement duration is 24s
(100 revolutions). The test bearing was worn out after surpassing 42 million revolutions.

B.2.2 Pre-process Vibration Signal

The measured vibration signal contains frequency components that are smeared across
the frequency spectrum due to small variations in motor speed. The vibration signal
is re-sampled to obtain an equal number of samples per round and spaced at a fixed
angle interval to reduce frequency smearing. The number of samples retained per round
is calculated using the reference shaft speed θ̇(ref ) and the sampling rate. The encoder
signal θ(t) is used as the reference angular shaft position, and the vibration signal is
re-sampled using linear interpolation. This method is named Order-Tracking.
After applying the order-tracking algorithm, the re-sampled vibration signal Vot (t) is
obtained. The next step is to filter the stationary vibration components not linked to
bearing fault impacts. Most of these components are synchronous with the shaft position
and are generated by e.g. misalignments or mass unbalance. A Time Synchronous Average
(TSA) algorithm is used to remove these components. The vibration signal is first split
up into multiple parts each with a time length of 1 round. The average of these parts is
the synchronous vibration signal Vs (t). Vs (t) is afterwards subtracted from the vibration
signal Vot (t) to remove the synchronous components. This subtraction results in the
asynchronous vibration signal Vas (t). More information on order tracking and TSA can
be found in [6].

B.2.3 Narrow-band Identifier Algorithm

The narrow-bands are identified in the vibration signal frequency spectrum. A Fast
Fourier Transform (FFT) algorithm is used to obtain the frequency spectrum. It re-
turns the absolute amplitude Y and the frequency f of each sine wave component. The
frequency spectrum is obtained using (B.1).

[Y, f ] = FFT(V (t)) (B.1)

Y (i) denotes the amplitude at the i’th frequency bin, and similarly f (i) denotes the
frequency at the i’th bin. After obtaining the spectrum, an envelope is determined and
placed on top of it. The envelope is obtained by modifying the amplitude of the spectrum
to the maximum value within ±15Hz from each frequency bin. Afterwards, a low-pass
filter is applied to smoothen the transitions. The resulting envelope is seen on top of the
spectrum in Fig. B.1.

0.0010
Vibration Spectrum
Amplitude Y (V )

0.0008 Spectrum Envelope

0.0006

0.0004

0.0002

0.0000
0 2000 4000 6000 8000 10000
Frequency f (Hz)

Figure B.1: The frequency spectrum of the vibration signal in addition to its envelope.

To determine each narrow-band in the spectrum, each local minimum of the envelope
determines the separation frequency. These separation frequencies are shown in Fig. B.2
and each narrow-band is given a number ranging from 1 to 9 in this example.

0.0010
1 Vibration Spectrum
Amplitude Y (V )

0.0008 Narrow-band separation

0.0006

0.0004 2

0.0002
3 4 5 6 7 8 9
0.0000
0 2000 4000 6000 8000 10000
Frequency f (Hz)

Figure B.2: The frequency spectrum of the vibration signal combined with the nine
narrow-bands that are identified for this dataset.

The first three narrow-bands contain the most energy and should be analyzed for bearing
fault impacts. However, the remaining six are also investigated in case the bearing natural
frequency is within these narrow-bands.

B.2.4 Envelope Spectrum

Impacts in the bearing cause it to ring (vibrate) at its resonance frequency. The narrow-
band may contain frequency components related to this ringing in the bearing. The
resonance frequency of the bearing is not directly linked to any characteristic bearing
fault frequencies, but acts as an information carrier in an amplitude modulated signal.
The information we seek lies in the frequency of the modulation waveform. Therefore it is
necessary to de-modulate the signal to diagnose faults in the bearing. Each narrow-band
is analyzed by performing three tasks. First, the vibration signal is band-pass filtered
around the narrow-band frequency. Second, a Hilbert-Huang transform is applied to the
bandpass-filtered signal to demodulate the signal. The result is the envelope of the sig-
nal, i.e. the modulation waveform. Finally, the envelope is transformed to the frequency
domain via the FFT algorithm in (B.1). The resulting spectrum is named the Envelope
Spectrum. It can be analyzed directly to find prominent peaks at the characteristic fault
frequencies derived from the bearing kinematics. However, recall that the vibration sig-
nal is split up into multiple narrow-bands that should be analyzed using the Envelope
Spectrum method. Analyzing multiple spectra manually is impractical due to time lim-
itations. Therefore, there is a need for an autonomous bearing fault diagnostic system
that can analyze the spectra. The next subsection presents such an autonomous bearing
fault diagnosis system.

B.2.5 Autonomous Fault Diagnosis

The following method aims to mimic the way a diagnostician would analyze an envelope
spectrum. Typically, one can observe the spectrum for prominent peaks corresponding to
the characteristic bearing fault frequencies. Also, there may be several harmonics and also
sub-bands if roller or inner-ring faults are present. Due to slip or a non-zero contact angle
in the bearing, the spectrum peaks may not coincide perfectly with the characteristic fault
frequencies. The diagnostician also has to account for this possible error. These tasks are
built into the autonomous bearing fault diagnosis algorithm.
To illustrate the method, the envelope spectrum of the third narrow-band (M = 3)
shown in Figure B.2 is used as an example. The test bearing has an inner-race fault
that will be characterized by harmonics at the inner-race ball pass frequency, and sub-
bands equal to the shaft frequency. The characteristic fault frequency for the inner race
is fc = 6.88 orders, and the expected sub-band frequency is fsb = 1 order. One order is
the same as the shaft speed as shown in (B.2)

1 order ≡ fs (B.2)

where fs is the shaft speed. Converting the envelope spectrum frequency to orders
removes the need for scaling the fault frequencies based on the shaft speed. The steps
in the autonomous fault diagnosis algorithm are detailed in the following. Suppose the
envelope spectrum is given by the amplitude Y and the frequency f . Let the harmonic
number H = 1.
Step 1: Identify the maximum amplitude within a small frequency band around a
characteristic fault frequency. This maximum value is the harmonic value and is calculated
as follows:

Mh , I = max(Y (i)) : fh1 ≤ f (i) ≤ fh2 (B.3)

where fh1 = cα fc · (H − w), fh2 = cα fc · (H + w), and I is the index i where Y (i) is maxi-
mum. The frequency band is determined by the slip tolerance w and the fault frequency
correction factor cα . These values are calculated using (B.4) and (B.5) respectively.

(
0.02, for H = 1
w= (B.4)
0.01, else
(
1, for H = 1
cα = (B.5)
α, else

The correction factor α is calculated in (B.6) to adjust the characteristic fault fre-
quency to account for roller slip and contact angle.

α = f (I)/(fc · H) (B.6)

where f (I) is the frequency where the most prominent harmonic peak is observed.
Step 2: A “noise level” within a small band around the corrected fault frequency is
calculated using (B.7) to determine whether the maximum peak is prominent enough.

1 hX i 
N= Y (i) : fn1 ≤ f (i) ≤ fn2 − Y (I) (B.7)
n−1
where fn1 = αfc · (H − 0.02), fn2 = αfc · (H + 0.02), and n is the number of elements in
the summation. N is the mean of all spectrum values within the frequency band, except
for the value of the harmonic. The frequency bands used to determine the maximum
harmonic value and the noise are shown as an example for H = 1 in Fig. B.3.
Step 3: If the fault frequency is amplitude modulated by a sub-band frequency fsb ,
there should be at least one prominent peak ±fsb away from αfc . If it is not amplitude
modulated, skip this step. Two prominent sub-bands are searched for: one negative
sub-band and one positive sub-band. The negative sub-band is calculated using (B.8).

Mns = max(Y (i)) : fns1 ≤ f (i) ≤ fns2 (B.8)

 0.00035
Harmonic
0.00030
Amplitude Y (V )
Noise
0.00025 fh1 fh2

0.00020 fn1 fn2

0.00015
0.00010
0.00005
0.00000
6.0 6.5 7.0 7.5 8.0
Orders f /fs

Figure B.3: For H = 1, the frequency bands for identifying the harmonic value and the
noise are shown. In addition, the harmonic is marked as a green dot and the noise level
is the red-stapled line.

where fns1 = Hαfc − fsb − 0.05fs , and fns2 = Hαfc − fsb + 0.05fs . Here a 5% tolerance
away from the shaft speed is used to generate the frequency band where the negative
sub-band should be. The positive sub-band is calculated in a similar manner in (B.9).

Mps = max(Y (i)) : fps1 ≤ f (i) ≤ fps2 (B.9)

where fps1 = Hαfc + fsb − 0.05fs , and fps2 = Hαfc + fsb + 0.05fs . The frequency bands
are shown together with the maximum sub-band values for H = 1 in Fig. B.4.

0.00035
fns1 fps2
0.00030 fns2 fps1
Amplitude Y (V )

0.00025
Neg. Sub.
0.00020 Pos. Sub.
0.00015
0.00010
0.00005
0.00000
6.0 6.5 7.0 7.5 8.0
Orders f /fs

Figure B.4: For H = 1, the frequency bands for identifying the sub-bands are shown. In
addition the identified sub-bands are marked with a red and blue dot.

Step 4: If the harmonic and possible sub-bands are prominent enough, a score is cal-
culated based on the harmonic value. However, the harmonic value must pass a threshold
if a score should be computed. The threshold is calculated using (B.10).
 T = 3N (B.10)

If there should be sub-bands present around the fault frequency, the following relation
must be true to calculate a score:

Mh > T and (Mns > T or Mps > T ) (B.11)

Similarly if fsb = 0, the following relation must be true to calculate a score:

Mh > T (B.12)

If the relation above is true, the harmonic score is calculated in Step 5. Otherwise,
skip to Step 6.
Step 5: A scoring system is developed to quantify the prominence of the harmonic
peaks. The score for each identified harmonic is calculated using (B.13).

S(H) = (Mh /3N ) · H 2 (B.13)

where S(H) is the score for the H’th harmonic. The score is unit-less and rewards
narrow-bands with multiple harmonics. A large score implies a high probability of damage
being present. Afterward, more harmonics are searched for by increasing the harmonic
number H = H + 1 and returning to Step 1. Figure B.5 shows an example where the
second harmonic and sub-bands are identified in mode 3. In this example, the negative
sub-band and the harmonic is greater than the threshold, and therefore the score of 10.21
is calculated using (B.13).

0.00030

0.00025
Amplitude Y (V )

Harmonic
0.00020 Neg. Sub.
Pos. Sub.
0.00015

0.00010

0.00005

0.00000
12.5 13.0 13.5 14.0 14.5
Orders f /fs

Figure B.5: For H = 2, the harmonic and two sub-bands are identified.

Step 6: Once all the prominent harmonics are identified, a combined score for the
narrow-band can be calculated using (B.14).
 nH
X
Sm (M ) = S(i) (B.14)
i=1
where nH is the number of harmonics identified, and M is the narrow-band number.
The score summation can be extended by also summing over all the narrow-bands in the
vibration dataset. This combined score is calculated using (B.15).

nM
X
Str = Sm (i) (B.15)
i=1
where nM is the number of narrow-bands in the vibration dataset. Since Str is a single
scalar value for a given dataset and a given fault type, it can be monitored over time like
RMS and Kurtosis. This provides an excellent opportunity to visualize the changes in
the bearing condition and to set up automatic warnings or alarms for the operator. The
complete flow diagram for the algorithm described in this section is shown in Fig. B.6.

Figure B.6: Complete algorithm flowchart.

B.3 Results and Discussion

B.3.1 Vibration Mode Score

The vibration dataset whose frequency spectrum is shown in Fig. B.2, contains at least
two narrow-bands that include the characteristic fault components for an inner-ring fault.
The first one is narrow-band three, M = 3. The harmonics observed in the envelope
spectrum of this narrow-band are shown in Fig. B.7.
 0.0014
Neg. Sub.
0.0012
Amplitude Y (V )
Harmonic
0.0010 Pos. Sub.
0.0008
0.0006
0.0004
0.0002
0.0000
0 5 10 15 20 25 30
Orders f /fs

Figure B.7: For the third narrow-band, M = 3, a total of 3 harmonics related to the
inner-race fault are identified.

The score calculated using (B.14) for the third narrow-band is Sm (3) = 39.33. Whether
this value is large or small is difficult to determine. But normally, analyzing an undamaged
bearing should return a score quite close to zero. In addition, analysis of the sixth narrow-
band (M = 6) also reveals signs of a fault on the inner-ring. The harmonics observed in
the analysis of this narrow-band are shown in Fig. B.8.

0.00030
Neg. Sub.
0.00025
Amplitude Y (V )

Harmonic
0.00020 Pos. Sub.

0.00015

0.00010

0.00005

0.00000
0 5 10 15 20 25 30
Orders f /fs

Figure B.8: For the sixth narrow-band, M = 6, a total of 3 harmonics related to the
inner-race fault is identified.

The score calculated using (B.14) for the sixth narrow-band is Sm (6) = 36.14, which
is almost as much as the third narrow-band score. Therefore, there are reasons to ana-
lyze more than one narrow-band on the vibration signal. For reference and comparison
purposes, the score for the nine narrow-bands are given in Table B.1.
 Table B.1: Inner ring diagnostic score for the nine narrow-bands in the example data.
M 1 2 3 4 5 6 7 8 9
Sm (M ) 3.35 0 39.33 0 0 36.14 1.36 1.75 0

B.3.2 Vibration Trend Score

A great usage scenario for the proposed method is to monitor the score for the three
characteristic bearing fault types over time. The algorithm in Fig B.6 is repeated for
each characteristic fault by altering fc and fsb to the combinations given in Table B.2.
The combined vibration dataset score is calculated using (B.15) for the three character-
istic bearing faults and plotted in a graph. The proposed method will be compared to
monitoring the RMS value of the vibration data. The RMS is calculated using (B.16).
v
u nt
u1 X
VRM S = t V (i)2 (B.16)
nt i=1

where nt is the number of samples in the vibration dataset. The last 100 datasets
acquired from the accelerated life test detailed in Section B.2.1 are used to monitor the
progressing bearing faults. The scores calculated using (B.15) for the three fault types
are shown in Fig. B.9 a), and, for comparison, the RMS value as calculated using (B.16)
is shown in Fig. B.9 b).

Table B.2: Characteristic fault frequencies for the test bearing in orders.
Inner Race Outer Race Roller
fc 6.88 5.12 6.66
fsb 1 0 0.43

In Fig. B.9 a) the scores are normalized to fit all three graphs in a single figure, but the
respective maximum values are given in the legend. Analyzing Fig. B.9 a) it is seen that
until the 71st dataset the condition of the bearing is unchanged and in a healthy state.
At this point, however, the proposed algorithm returns a high score for the roller fault,
while it has previously been close to zero. This finding corresponds well with the increase
in RMS value as shown in Fig. B.9 b). However, the roller fault does not appear in the
vibration datasets for long. Instead, an outer-ring damage propagates during the next 20
datasets. Finally, the inner-ring is also damaged, as is seen during the final 10 datasets in
Fig. B.9 a). The score values calculated using the proposed method correspond well with
the findings in the de-assembled bearing. The internal damage in the bearing is shown in
Fig. B.10.
 1.0
a) Inner-Ring Str
max = 938.44
0.8
Normalized Score

Rollers Str
max = 126.31
0.6 Outer-Ring Str
max = 623.16

0.4

0.2

0.0
0 10 20 30 40 50 60 70 80 90 100
Dataset Number

0.025
b) Vibration RMS
0.020
Vibration RMS

0.015

0.010

0.005

0.000
0 10 20 30 40 50 60 70 80 90 100
Dataset Number

Figure B.9: For the last 100 datasets, the characteristic fault scores are calculated using
(B.15) and shown in a). The RMS trend calculated using (B.16) is shown in b) for
comparison.
Figure B.10: The de-assembled test bearing after the accelerated life test. a) shows two
damaged balls, b) shows a small pit in the outer-ring, and c) shows a large area of pitting
on the inner-ring.

Fig. B.10 shows signs of damage in two rollers, the outer-ring, and the inner-ring. The
locations of the faults correspond very well with the results from the proposed method.
This comparison shows that the proposed algorithm is very efficient in diagnosing the
bearing for the characteristic faults.

B.4 Conclusion
Raw vibration data is often bandpass-filtered around the bearing fundamental frequency
only to preserve frequency components related to fault impacts. However, determining the
resonance frequency of the bearing is a challenging task. Instead of identifying the optimal
narrow-band using Kurtogram, the vibration data is divided into multiple narrow-bands
and each band is analyzed to diagnose the bearing. However, this is a tremendous task to
perform manually and thus an automatic bearing fault diagnosis algorithm is proposed in
this article. It is capable of identifying the three most common faults in a bearing with
good accuracy. The proposed algorithm has been explained in detail and validated using
experimental data.
REFERENCES

[1] L. Guo, J. Chen, and X. Li. Rolling bearing fault classification based on envelope
spectrum and support vector machine. Journal of Vibration and Control, 15(9):1349–
1363, 2009. doi:10.1177/1077546308095224.

[2] C. Mishra, A. K. Samantaray, and G. Chakraborty. Rolling element bearing defect

diagnosis under variable speed operation through angle synchronous averaging of
wavelet de-noised estimate. Mechanical Systems and Signal Processing, 72:206–222,
2016. doi:10.1016/j.ymssp.2015.10.019.

[3] J. Antoni. Fast computation of the kurtogram for the detection of transient faults.
Mechanical Systems and Signal Processing, 21(1):108–124, 2007. doi:10.1016/j.
ymssp.2005.12.002.

[4] T. Barszcz and A. JabLoński. A novel method for the optimal band selection for
vibration signal demodulation and comparison with the Kurtogram. Mechanical
Systems and Signal Processing, 25(1):431–451, 2011. doi:10.1016/j.ymssp.2010.
05.018.

[5] X. Xu, M. Zhao, J. Lin, and Y. Lei. Envelope harmonic-to-noise ratio for periodic
impulses detection and its application to bearing diagnosis. Measurement, 91:385–
397, 2016. doi:10.1016/j.measurement.2016.05.073.

[6] R. B. Randall and J. Antoni. Rolling element bearing diagnostics—a tutorial.

Mechanical Systems and Signal Processing, 25(2):485–520, 2011. doi:10.1016/j.
ymssp.2010.07.017.

107
Paper C

Cross-correlation of Whitened
Vibration Signals for Low-Speed
Bearing Diagnostics

Andreas Klausen and Kjell G. Robbersmyr

109
This paper has been published as:

A. Klausen and K. G. Robbersmyr. Cross-correlation of Whitened Vibration Sig-

nals for Low-Speed Bearing Diagnostics. Mechanical Systems and Signal Processing,
118:226-244, 2019. doi: 10.1016/j.ymssp.2018.08.048.
 Cross-correlation of Whitened Vibration
Signals for Low-Speed Bearing Diagnostics
Andreas Klausen and Kjell G. Robbersmyr

University of Agder
Department of Engineering Sciences
Jon Lilletunsvei 9, 4879 Grimstad, Norway

Abstract – Rolling-element bearings are crucial components in all rotating

machinery, and their failure will initially degrade the machine performance,
and later cause complete shutdown. The period between an initial crack and
complete failure is short due to crack propagation. Therefore, early fault
detection is important to avoid unexpected machine shutdown and to aid in
maintenance scheduling. Bearing condition monitoring has been applied for
several decades to detect incipient faults at an early stage. However, low-
speed conditions pose a challenge for bearing fault diagnosis due to low fault
impact energy. To reliably detect bearing faults at an early stage, a new
method termed Whitened Cross-correlation Spectrum (WCCS) is proposed.
The method computes the cross-correlation between the whitened vibration
signal and its envelope. In this paper, it is detailed how this correlation
can improve the fault diagnosis compared to analyzing the envelope spectrum
alone. Compared to other methods reported in the literature, the WCCS pro-
vides accurate fault detection without involving experimentally tuned settings
or bandpass-filtering. Vibration data at 20 revolutions per minute rotational
speed from an accelerated life-time test of a 40 mm bore size bearing is used
to verify the performance of the proposed method. An additional case study
using the WCCS on a difficult dataset from the Case Western Reserve Uni-
versity database is also presented to verify the performance.

C.1 Introduction
Rolling-element bearings, or bearings for short, are crucial components in all rotating
machinery. Their failure is one of the most common cause of machine breakdown. A
worn bearing is characterized by increased vibration levels, internal looseness, and higher
friction. The increase of vibration can damage nearby components, and lead to a full
stop of the machine. If worn bearings are not replaced in time, costly downtime or
personnel injuries may occur. Condition monitoring techniques can be applied to estimate

111
the bearing health and remaining useful life-time. Data from sensors that measure a
physical quantity, like the vibration, are used as input to such a system. The data is
further analyzed using signal processing algorithms, before the results are presented to an
operator. Based on the results, the operator can decide whether the bearing is in a healthy
state, or if it is worn and should be replaced. Such condition monitoring systems have
been used for several decades to monitor the health of all kinds of rotating machinery
components. The most common sensor type to use is vibration accelerometers, as the
bearing vibration is closely linked to the amount of internal wear. An incipient fault, on
either bearing race-way or a roller, causes an impulse of vibration every time it is struck.
The spectral frequency of the resulting vibration is based on the resonance frequency of
the system and is normally in the thousands of Hertz. Further, the resonance frequency
of a bearing system is normally not known as it is difficult to determine analytically or
experimentally. However, the spectral frequency is not directly of interest when diagnosing
a bearing. The cyclic frequency between each impact impulse may reveal its fault. By
analyzing the kinematics of a bearing under no-slip conditions, the characteristic cyclic
frequencies for the different fault types are determined. If the cyclic vibration frequency
match any of the characteristic frequencies, the bearing is likely to be damaged at that
certain location. To determine the cyclic impact frequencies of the measured vibration
signal, the Fourier transform of its demodulated envelope may be analyzed to identify
the bearing cyclic vibration. During low-speed conditions, there may be some challenges
in diagnosing bearing faults using a vibration accelerometer. The fault impact energy is
dependent on the shaft speed and is decreased if the shaft speed decreases [1]. Background
noise, however, is the same no matter what the shaft speed is, and therefore the signal-
to-noise-ratio (SNR) decreases with decreasing shaft speed. The envelope spectrum may
not reveal the fault if the characteristic cyclic frequencies are masked in background noise
during low-speed conditions. In this case, more advanced signal processing methods are
necessary to highlight the low energy impacts during low-speed conditions.

A new method was proposed in [2] for diagnosing a low-speed bearing using extracted
discrete wavelet packets that contains the bearing fault vibrations. The results show that
the multiple band-pass filtered autoregressive envelope spectrum provided clear indication
of faults at 60 revolutions per minute (rpm) rotational speed. However, one disadvantage
is that the Adaptive Network-based Fuzzy Inference System trained to choose the bands,
requires pre-labeled training data, and such historic data may not be available for every
system. In recent years, research on acoustic emission (AE) technology has shown that
it is sensitive to early sub-surface cracks in the bearing. While vibration sensors are only
sensitive to impacts on the bearing surface, AE sensors may detect changes in the bearing
sub-surface, which should aid in early fault diagnosis. The biggest disadvantage is the cost
of the equipment, and the high required sample rate, as the signal of interest is typically
in the range of 100 kHz and 1 MHz [3]. The disadvantages of the high sampling rate
is however reduced significantly by an efficient down-sampling technique that does not
affect signal quality [4]. The results indicate that even when down-sampling by a ratio
of 500, the bearing fault signature could be captured by the AE sensor. In [5], a thrust
bearing is run to failure at 72 rpm, and the resulting AE was measured at four locations
on the outer ring. The results indicate that the AE energy increased with increasing
fault size, and that the fault type is detectable in certain pre-processed spectra. In [6],
a bearing with a pre-seeded fault at an extremely low speed of 1.12 rpm is diagnosed
using AE. It was concluded that parameters such as AE amplitude and energy provided
valuable information on the condition of a low-speed rotating bearing. In addition, a
method for detecting a fault in the bearing using AE signal was also presented. The
method consists of grouping multiple stress-wave signals by the centroid of autoregressive
model coefficients. If two distinct groups are formed, the bearing is considered faulty.
If no distinct groups were formed, the stress wave signals are considered to only contain
noise, and the bearing is assumed healthy. In [7], a low-speed bearing rotating at 10 rpm
was diagnosed using Support Vector Machine and Relevance Vector Machine. The input
features were generated from both vibration and AE sources, and it was concluded that
the classifier trained on AE data had the best accuracy. In [8], the performance of AE
and vibration was compared on a large slew bearing rotating at 8 rpm. Both signals
were pre-processed using a combination of multivariate Principal Component Analysis
and Ensemble Empirical Mode Decomposition, which adaptively decomposed the signal
into different time scales. It was shown that both the vibration and AE signal contained
enough information to diagnose the bearing for a seeded inner ring fault after proper
pre-processing. Other research has also shown success in using AE for low-speed bearing
fault diagnosis [9, 10, 11, 12].

Most of the referenced works state that AE is superior to vibration signals for bearing
fault diagnosis under low-speed conditions. While a vibration sensor can capture the fault
impact vibration, the signal is often masked by vibration from other machine components,
and background noise. In addition, the bearing fault impact energy gets lower as the
shaft speed is reduced. Therefore, a solid algorithm is necessary to extract the weak
impulses generated in the bearing during low-speed conditions to properly diagnose its
condition. In this article, vibration signals are used to diagnose a 6008 ball-type bearing
for faults. The bearing is run to failure on an accelerated life-time test bench at 500
rpm. The vibration signal is measured every 10 minutes at 20 rpm for the duration of
50 revolutions to acquire low-speed vibration signals. Initially, the vibration signal is re-
sampled using encoder data to achieve a constant angle increment rather than time. Next,
shaft-synchronous vibration components are removed by subtracting the average vibration
per revolution. To highlight the slightly random vibration impulses in the bearing, an
Autoregressive Model (ARM) is afterwards trained to predict deterministic components
that are not related to the bearing fault. The components predicted by the ARM are
removed to retain random components such as the bearing vibration. After this process,
the vibration signal is whitened as it contains mostly random components. Finally, the
cross-correlation between the whitened vibration signal and its envelope is computed.
This cross-correlation results in an element-wise multiplication of the frequency spectrum
of the whitened vibration signal and its envelope. The justification of computing the cross-
correlation is that the bearing characteristic fault vibration may be visible directly in the
spectrum of the raw vibration signal [13], as well in the spectrum of the demodulated
high-frequency vibration signal. Therefore, to exploit both signals, the cross-correlation
is calculated to fuse the information in both spectra. Frequency amplitude peaks that are
present in both spectra will be amplified, while peaks that are only present in either of
the two are attenuated. Using the Fourier transform, the resulting spectrum of the cross-
correlation signal, termed the Whitened Cross-correlation Spectrum (WCCS), is analyzed
for faults related to the bearing. The results indicate that, even at the low-speed of 20 rpm,
early identification of fault is possible. For verification purposes, a difficult dataset from
the Case Western Reserve University bearing fault database is also successfully diagnosed
using the WCCS. The rest of the paper is organized as follows: the experimental setup
and the algorithms used in the proposed method and comparison methods are presented
in Section C.2. Afterwards, results of using the proposed method are presented, and the
performance is compared to other methods reported in the literature, like the Envelope
Spectrum, in Section C.3. Finally, conclusions are drawn in Section C.4.

C.2 Methods

C.2.1 Experimental test setup

The vibration data is collected on an accelerated bearing life-time test rig that is shown
in Figure C.1. A permanent magnet motor combined with a 1:7 reduction planetary
gearbox is used to drive the test bearing shaft, and a variable-frequency drive allows for
variable speed conditions. The test bearing at the left end of the shaft is a 6008 type
bearing with 40 mm bore size, and its characteristic fault orders are given in Table C.1.
These frequencies are given in magnitudes of shaft orders, where the shaft speed equals
1 order, i.e. OS = 1. The rated dynamic load of the bearing is C = 17.8 kN, and the
static load rating is C0 = 11.6 kN. Radial and axial loads are applied by two electric
 Figure C.1: Simplified view of the accelerated bearing life-time test bench.

linear actuators geared with mechanical levers. The two support bearings at the middle
of the shaft aid to counteract the radial force on the test bearing. The axial bearing
is installed to transfer the stationary axial load to the rotating shaft. A unidirectional
vibration accelerometer is stud-mounted to the side of the test bearing housing with the
direction and placement as shown in Figure C.1. Its linear range is between 2 Hz and
10 kHz, and the nominal sensitivity is 100 mV/g with a maximum peak acceleration
of 60 g. The vibration is sampled at 51.2 kHz using a 24-bit ±30 V A/D converter. A
quadrature incremental encoder is located inside the motor. With 1024 pulses per channel,
a total of 4096 pulse edges can be identified per motor shaft revolution. Additionally, the
motor rotates 7 times faster than the test bearing shaft due to the gearbox, resulting in
a resolution of 28,672 pulse edges per revolution on the test bearing shaft. To record the
shaft position, the number of pulse edges passed is sampled at a frequency of 512 Hz.
The test bearing is subjected to a radial and axial load of 9 kN and 7 kN respectively
to reduce the rotational life-time. The test procedure is as follows: The test bearing is
generally driven at a speed of 500 rpm. Every 10 minutes, a measurement cycle is initiated.
During this cycle, the bearing vibration and motor encoder position is measured for the
duration of 50 revolutions at shaft speeds of 500, 250, 100, 50, and 20 rpm successively.
After each cycle, the shaft speed returns to 500 rpm. In this article, the vibration data
collected at 20 rpm will be used since low-speed bearing fault diagnostics is the primary
objective. The bearing life-time was roughly 6 million revolutions (14 days of continuous
operation) before complete stop. The bearing was afterwards disassembled to identify
what components were damaged. Faults were identified on three of the rollers, the outer-
race, and the inner-race. Pictures of the disassembled bearing are shown in Figure C.2.
Figure C.2: The damaged components in the disassembled bearing. (a) Three damaged
rollers. (b) Damaged area on the outer-race. (c) Large damaged area on the inner-race.

Table C.1: Characteristic fault orders for the test bearing given in shaft orders assuming
zero contact angle and no roller slip. The expected harmonic order and the accompanying
side-band order (due to non-homogeneous radial load) for the given fault type are given.

Characteristic fault orders Acronym Value Side-band

Fundamental train OF T 0.43 -

Ball pass outer race OBP O 5.12 -
Ball pass inner race OBP I 6.88 OS
2x Ball spin O2BS 6.66 OF T
C.2.2 Order tracking

The motor speed controller is incapable of rotating the shaft at a completely constant
speed. Consequently, the time period between impacts in a bearing will vary based on
the instantaneous speed. The A/D converter that measures the vibration signal is storing
data at a fixed time interval given by the sampling rate. Because of the instantaneous
speed changes, this will result in a blurred frequency spectrum as the Fourier transform
assumes a stationary process. To mitigate the symptoms, the vibration signal is digitally
re-sampled with respect to the shaft encoder pulses. The re-sampled signal has a fixed
shaft angle interval rather than a time interval. This makes sure that there is, theoretically,
a fixed number of samples between each bearing fault impact. The following algorithm is
inspired by the work done by Fyfe and Munck in [14].
On the presented test rig, the rotor angular position is measured at a sampling fre-
quency of 512 Hz. Due to the fast sampling, it is assumed that the motor speed is constant
between each measurement. Under this assumption, the shaft angle at any time t can
be obtained using linear interpolation techniques. The shaft position is initially interpo-
lated to match the digital sampling times of the vibration signal V (t). This results in a
vibration signal that is given with respect to the shaft angle, but the shaft angle interval
is not fixed. To fix the interval, the vibration signal is interpolated to achieve a fixed
angle interval. Here, a second-order cubic spline interpolation is chosen as it provides
good interpolation performance at a low computational cost. The spline interpolation is
initialized as:
fot (θ) = interp{θ, V } (C.1)

where interp{·} is the cubic spline interpolation function, and fot (θ) is a continuous ap-
proximation of V describing points between the discrete samples of V using cubic spline
interpolation. Using fot (θ), the vibration data is re-sampled to achieve a fixed shaft angle
interval. The size of this angular interval can be arbitrarily chosen, but some guideline
should be followed to avoid over- or under-sampling. Here, the angular interval is chosen
to preserve the original sampling rate and total number of samples as close as possible
using:
  −1
Fs
∆θd = round , (C.2)
θ̇(ref)

where θ̇(ref) is the reference shaft speed in Hz, Fs is the vibration data sampling rate in Hz,
and round{·} rounds the number to the nearest integer. The reason for using the round{·}
function and inverse, is to make sure that an integer number of samples contain a complete
shaft revolution. This is crucial when performing Angle Synchronous Averaging to avoid
overlap between revolutions. Finally, the vibration signal is re-sampled to achieve the
constant shaft angle interval using:

Vot (j · ∆θd ) = fot (j · ∆θd ). (C.3)

where Vot is the order tracked vibration signal, and j is a positive integer. In the next
Subsection, the average vibration per shaft revolution is calculated to remove the shaft-
synchronous vibration components.

C.2.3 Angle synchronous average

The vibration signal includes components from sources that are synchronous to the shaft
rotation. Typical components include vibration from shaft misalignment, shaft unbal-
ance, and motor vibration. In addition, the meshing vibration from fixed-axis gearboxes
are also integer multiples of the shaft speed. It should be noted, however, that the test
rig in Section C.2.1 does not include a fixed-axis gearbox. These shaft-synchronous vi-
bration components may mask the bearing fault and should therefore be removed before
further processing of the signal. The bearing characteristic fault orders in Table C.1 are
non-integer orders, and are therefore not shaft-synchronous. Therefore, by estimating the
shaft-synchronous vibration components and removing that from the vibration signal, the
bearing fault vibration is not affected. Using a deterministic/random separator (DRS)
such as the time synchronous average in [15], the shaft-synchronous components are esti-
mated and removed from the signal. In this research, the signal was first order-tracked by
means of computed order tracking, and therefore a modified Angle Synchronous Average
(ASA) is performed instead. An estimation of the shaft-synchronous components is found
by calculating the mean value of the vibration signal per revolution using:
nr −1
1 X
Vs = Vot [j · Nr : (j + 1) · Nr ] , (C.4)
nr j=0

where nr is the number of revolutions the shaft has turned during the measurement,
Nr = 1/∆θd is the number of samples measured per revolution, Vot [a : b] imply all discrete
samples of Vot including Vot [a] up to and including Vot [b − 1], and the first sample is at
index 0. This results in the ASA, which should contain the shaft-synchronous vibration
components. The ASA is afterwards removed from each revolution in the order tracked
vibration signal using:

Vas [j · Nr : (j + 1) · Nr ] = Vot [j · Nr : (j + 1) · Nr ] − Vs : j ∈ [0, nr − 1], (C.5)

where Vas is the asynchronous vibration data containing the bearing fault. More infor-
mation on the general time synchronous average is found in [15]. In the next subsection,
other deterministic signal components are separated and removed using an autoregressive
whitening filter.
C.2.4 Autoregressive whitening filter
The ASA algorithm attenuated most of the shaft-synchronous vibration originating from
the shaft and the motor. However, there are still deterministic vibration components
not related to the bearing fault present in the dataset. One such source is the planetary
gearbox used in the test rig. Planetary gearboxes emit a complex vibration pattern due
to multiple moving components and non-integer orders forcing frequencies [16]. These
components are, however, deterministic and therefore predictable. The characteristic
bearing orders from Table C.1 assumes a zero contact angle and no slip. However, as the
local radial and axial load relationship for each roller may be different from one another,
each roller moves at slightly different speeds. The bearing cage moves at the mean speed of
all rollers, forcing some of them to slip on the raceway. Consequently, angular deviations
of up to 2% between impacts may occur [15], and the bearing vibration can therefore be
modeled as a cyclostationary process. This allows the bearing vibration to be separated
from the planetary gearbox vibration using a DRS. A linear prediction model as described
in [15] is used in this research to model the deterministic components. To this end, an
Autoregressive Model (ARM) is estimated to predict the future values of the vibration
signal based on the p previous values. The ARM acts as a model of the deterministic part
of the signal. The ARM is afterwards subtracted from the original signal, separating the
random components. Separating the deterministic components whitens the signal, and
hence this process is described as a whitening filter. The ARM is given by:
p−1
X
Vas [k] = − a[j] · Vas [k − j − 1] + Vw [k], (C.6)
j=0

where a[j] is the j’th parameter of the model, Vw is the whitened vibration signal (or ARM
residual), and p is the ARM order. Random signal components, including the bearing
vibration, are not perfectly predictable and will therefore be a part of the residual Vw . The
unknown parameters are identified by solving the Yule-Walker equations [17, 18] for ARM
training. Wang and Wong [19] describes a method for efficiently training the ARM using
Levinson-Durbin Recursion (LDR). This training method allows for the identification of
all ARM up to an order of pmax . Theoretically, the ARM can have as many parameters
as the input signal length. C.7red However, to avoid overfitting, i.e. the ARM starts
to predict random data, the maximum order must be less than the number of samples
between bearing fault impulses [20]. This limit also imposes a selective whitening as the
deterministic components related to the bearing fault are preserved in the residual. The
maximum number of samples is calculated using:
 
1
pmax = floor , (C.7)
max{Obcf }∆θd
where Obcf is a list containing the characteristic bearing fault orders for the present
bearing, floor{·} returns the nearest integer rounding in the negative direction, and max{·}
returns the maximum number of a list. Since the type of fault is unknown during analysis,
the maximum characteristic fault order is used.
The training method in [19] uses the Akaike Information Criterion (AIC) to determine
the optimal model order up to pmax . One modification is however implemented in this
research by using the Corrected Akaike Information Criterion (AICC ) [21] rather than the
standard AIC. This change is made for its advantage on finite size datasets. The optimal
ARM order popt is identified by the minimal value of AICC . Using the optimal ARM, the
residual of the ARM is calculated using:
popt −1
X
Vw [k − popt ] = Vas [k] + aopt [j] · Vas [k − j − 1], (C.8)
j=0

where aopt is the optimal ARM. The rest of the subsection is devoted to an example using
the autoregressive whitening filter. Let y be a simulated signal containing a sum of sines,
impulses, and white Gaussian noise given by:
10
X
y = It + W + sin {j · 0.5 · 2π · t + φ[j]} , (C.9)
j=1

where It is an impulse train of strength 1.5 with frequency 1 Hz, W is white Gaussian
noise with a standard deviation σ = 1/5, and φ[j] is a random phase ∈ [0, 2π]. The
signal is simulated at a sample rate of 100Hz for 10 seconds, and is shown in Figure C.3
(a). Here, it is difficult to spot the impulses occurring every second. An ARM is trained
to predict the deterministic components to separate the random components. Since the
impulses occur at a rate of 1 Hz, the maximum ARM order is set to pmax = 99 to avoid
prediction of the impulses. After model training, the residual is given by (C.8) and shown
in Figure C.3 (b). The residual only contains the impulses and Gaussian noise and is
therefore suitable for further processing. In addition, the AICC for each ARM order up
to pmax is shown in Figure C.3 (c), and the residual power σ 2 is shown in Figure C.3
(d). The residual power decreases continuously with increasing filter order, but the AICC
prevents overfitting the model. The optimal filter order is popt = 19 for this example.

C.2.5 Hilbert transform

The ARM residual retains the high-frequency amplitude-modulated bearing fault vibra-
tion impacts as well as low-frequency deterministic components related to the fault. The
bearing characteristic frequency for a specific fault is the frequency of the modulation
waveform. This modulation waveform is obtained by determining the envelope of the
 ×103
−0.8
3 (a) (c)
Noisy signal y Optimal
−0.9
Signal

AICC
0 −1.0

−1.1

−3 −1.2
0 1 2 3 4 5 6 7 8 9 10 0 20 40 60 80 100
Time [s] Filter order

×10−1
1.6
(b) (d)
Residual Optimal

Resdial power
1.5 1.4
Residual

0.0 1.2

−1.5 1.0
0 1 2 3 4 5 6 7 8 9 10 0 20 40 60 80 100
Time [s] Filter order

Figure C.3: Example scenario where an impulse train is revealed after applying the autore-
gressive whitening filter. (a) The example input signal. (b) The residual after applying
the whitening filter. (c) The AICC criteria as a function of the filter order together with
the optimal value. (d) The residual power as a function of the filter order.

signal. The envelope is obtained by computing the absolute value of the analytic signal.
The analytic signal is acquired using:

Va (θ) = Vas + H{Vas } = Vas + iVi , (C.10)

where H{·} is the Hilbert transform, i is the imaginary unit, and Vi is the complex part of
the analytic signal. The zero-mean envelope is obtained by computing the Pythagorean
distance in the complex plane and removing the mean value using:
q q
Venv = |Va | = Vas + Vi − mean{ Vas2 + Vi2 }.
2 2
(C.11)

where mean{·} returns the mean value of an array. In the next subsection, bearing
fault vibration components are amplified by computing the cross-correlation between the
whitened vibration signal and its envelope.

C.2.6 Cross-correlation
The enveloped vibration signal contains the demodulated cyclic frequency of the bear-
ing fault impacts. In addition, the selectively whitened vibration signal also contains
deterministic components at the bearing cyclic frequency. Instead of analyzing either of
the two signals, the signals are fused by means of the cross-correlation operation. The
resulting signal contain amplified frequency components related to the fault. A more
in-depth discussion on the effects of this cross-correlation is given in Section C.2.8. The
cross-correlation is calculated using:

Vcc (α) = Vw (θ) ? Venv (θ) = Vw∗ (−θ) ∗ Venv (θ), (C.12)
∗
where α is the angular lag, ? is the notation for cross-correlation, superscript indicates
the complex conjugate, and ∗ is convolution. Vw∗ is reversed to match the definition
of cross-correlation, hence noting the input with (−θ). The cross-correlated signal is
transformed to the order domain in the next subsection to be analyzed for bearing faults.

C.2.7 Order spectrum

The cross-correlated signal Vcc contains frequency components directly related to the
bearing fault. Harmonics and side-bands of the bearing characteristic frequencies can be
observed in the order spectrum of Vcc . To avoid spectral leakage, a Hann window is used
in the Fourier transform. The amplitude of the windowed Fourier transform is calculated
using:
Y = |F {w · Vcc }| , (C.13)

where w is the Hann-window, and w·Vcc must be treated as an element-wise multiplication.

The absolute value of the spectrum is used because the phase of each component is not
of interest, only the amplitude. The frequency bins of the resulting spectrum are given
in orders using:
j
O[j] = , (C.14)
nw · ∆θd
where nw is the size of the Hann window. The final order spectrum Y (O) is termed
the Whitened Cross-correlation Spectrum (WCCS) and represents the last step in the
proposed algorithm.

C.2.8 Theoretical Discussion on WCCS

The effects of the cross-correlation in (C.12) on the WCCS is not readily apparent in the
time domain. The reason for why this cross-correlation amplifies the bearing vibration is
difficult to visualize in the time domain. Advantageously, the time domain convolution
may be represented by an element-wise multiplication in the frequency domain thanks
to the convolution theorem. Considering this, a different way of calculating the cross-
correlated vibration is

Vcc (θ) = F −1 {F {Vw∗ (−θ)} · F {Venv (θ)}}, (C.15)

where F −1 is the inverse Fourier transform, and Vw∗ and Venv must be padded with zeros
to match the length of the original Vcc in (C.12). Inserting (C.15) into (C.13) shows that
the WCCS can be calculated with

Y = |F {w · F −1 {F {Vw∗ (−θ)} · F {Venv (θ)}}}|, (C.16)

which contain numerous Fourier transforms. However, if the Hann window w is replaced
by a rectangular window (i.e. no window), (C.16) may be written as

Y = |F {Vw∗ (−θ)} · F {Venv (θ)}|. (C.17)

(C.17) shows that the WCCS is actually a similarity measure between the spectrum of
the whitened vibration signal, and the spectrum of the envelope. If both spectra have
high values at certain frequency bins, the WCCS results in a magnified amplitude at
these bins. At bins with a high amplitude on only one of the spectra, its amplitude is
attenuated. The bearing characteristic fault frequency can be observed directly in the
raw/deterministic vibration spectrum [13]. Additionally, it is determined in [22] that
the demodulated vibration signal (envelope) also contains components at the bearing
characteristic frequencies. As the characteristic bearing fault vibration may be observed in
both the selectively whitened vibration signal and its envelope, the cross-correlation may
therefore amplify the bearing vibration and attenuate unwanted noise and uncorrelated
peaks. This also shows that it is only the low-frequency part of the two signals that
are used. Therefore to rapidly calculate the cross-correlation in (C.12), the two signals
should be low-pass filtered and decimated to reach a suitable sampling frequency. For
example, as most harmonics of a bearing fault are visible up to 50 orders, the signals
should be decimated to this amount before applying the cross-correlation to reduce the
computational load.
However, there are some possible limitations to the presented WCCS that must be
addressed. The presented method performs best when low-frequency bearing fault compo-
nents are discernible from background noise in both the selectively whitened deterministic
signal and its envelope. A theoretical model of the bearing vibration signal is thoroughly
analyzed in [23]. In this reference, it is shown that the deterministic bearing signal com-
ponents are the result of a low-pass filter applied to the high-frequency bearing resonance
vibration signal. Consequently, the deterministic components related to the bearing fault
are only discernible from background noise if the bearing resonance frequency overlaps
this low-pass filter. Typically, small bearings have a high resonance frequency, and there-
fore this condition may not be met in all scenarios. A recent publication [24] demonstrate
the limitation of using the deterministic torque vibration signal for bearing fault diagno-
sis. It is also shown that the demodulated high-frequency vibration signal is more suited
for bearing fault diagnosis. However, it should be noted that even if the low-frequency
bearing fault components in the deterministic signal are only marginally stronger than
 Figure C.4: Flowchart of WCCS used to diagnose the low-speed bearing for faults.

the background noise, the high-frequency signal envelope will still be amplified by the
cross-correlation step in (C.12), although not by much. In the case that the deterministic
signal peaks are completely drowned in noise, the amplification will be almost unity, i.e.
no improvement. Such a scenario is shown and discussed in Section C.3.2.

C.2.9 Complete algorithms

In this subsection, the proposed method is summarized with an accompanying flowchart
illustrating the implementation. In addition, other established methods in literature are
presented to establish a comparison with the proposed method.

C.2.9.1 Whitened Cross-correlation Spectrum

The flowchart for producing the WCCS is shown in Figure C.4. Initially, experimen-
tal data consisting of the speed reference θ̇(ref ) , vibration signal V (t), and encoder data
θ(t), are obtained in an accelerated life-test as described in Section C.2.1. This data
is used as input to the Order Track algorithm elaborated in Section C.2.2. After this
re-sampling process, a constant angle increment between each sample is realized. Next,
the synchronous components are estimated with the Angle Synchronous Average (ASA)
algorithm from Section C.2.3 and subtracted from the order tracked signal Vot to re-
tain asynchronous components, Vas . Afterwards, a linear Autoregressive Model (ARM)
is used as a DRS in Section C.2.4 to separate the bearing fault vibration from other de-
terministic components. This process acts as a whitening filter and returns the whitened
vibration signal. The envelope is obtained by using the Hilbert transform and calculat-
ing the absolute value of the analytic signal in Section C.2.5. This envelope contains
the demodulated high-frequency bearing fault vibration, random peaks, and white noise.
The cross-correlation between the whitened vibration signal and its envelope is calculated
in Section C.2.6 to amplify the bearing fault vibration, and attenuate white noise and
random peaks. Vcc is finally processed through a Hann-windowed Fourier transform as
elaborated in Section C.2.7, which results in the WCCS that is plotted for analysis. The
performance of the proposed method is compared against three other methods reported
in the literature, and they are explained in the next Subsection.

C.2.9.2 Comparison methods

To showcase the performance of the WCCS, the method is compared to three other
methods reported in the literature. These three are the Envelope Spectrum [22], Fast
Kurtogram [25], and the Fast Spectral Correlation [26]. These methods are similarly
computationally heavy, and do not involve experimentally tuned settings. To generate a
fair comparison, the whitened vibration signal Vw (θ) is used as input to all three methods.

Method A: Envelope Spectrum

The Envelope Spectrum [22] is a well-known method used for bearing fault diagnosis, and
its flowchart is shown in Figure C.5 (a). The method involves demodulating the vibration
signal via an envelope function to identify the characteristic bearing fault frequencies in
the frequency spectrum. In this application, the absolute valued analytic signal is applied
to get the envelope.

Method B: Kurtogram

To further improve the envelope spectrum, a narrow-band filter centered on the resonance
frequency of the bearing vibration can be applied beforehand to increase the signal-to-
noise ratio. However, the resonance frequency is normally unknown, and is difficult to
obtain analytically. To estimate this resonance frequency, the Fast Kurtogram [25] is
employed. The flow chart of this method is shown in Figure C.5 b). In summary, the
Fast Kurtogram passes the whitened vibration signal through a complex, narrow-band
filter-bank with varying central frequencies and spectral width. The kurtosis is calculated
for each filtered signal, and the narrow-band that maximizes the kurtosis is assumed to
Figure C.5: Three additional methods used to compare performance with the WCCS. (a)
Method A involves diagnosing the bearing using the Envelope Spectrum. (b) Method B
improves on the Envelope Spectrum by bandpass filtering the whitened vibration signal
around the narrow-band yielding maximum kurtosis. (c) Method C involves the fast
calculation of the Spectral Correlation which is a tool for analyzing cyclostationary signals.

contain the bearing resonance frequency. The squared envelope spectrum of the narrow-
band filtered signal Vbp (θ) is analyzed for the bearing characteristic fault frequencies. In
obtaining the Kurtogram, the excess kurtosis is calculated using:
µ4
Kurt{x} − E = − E, (C.18)
σ4

where µ4 is the fourth central moment, σ is the standard deviation, and E is the kurtosis
value for Gaussian white noise. For a real signal E = 3, and for a complex signal E =
2. Henceforth, the excessive kurtosis is named kurtosis to avoid repeating the word
“excessive”.

Method C: Fast Spectral Correlation

The Spectral Correlation (SC) is a tool for analyzing cyclostationary signals (signals
with hidden periodicities or repetitive patterns), like the high-frequency vibration signal
exhibited from bearing fault impacts. The resonance vibration is amplitude-modulated
with a comparably lower-frequency modulation waveform, making it ideal to analyze
using such cyclostationary tools. The SC is a two-dimensional spectrum showing the
cyclic-spectral frequency relationship. The usage of the SC has been limited in condition
monitoring applications due to its high computational cost. A fast version of the SC, Fast
SC, was developed in [26] which computes an estimate of the SC several magnitudes faster
than the original algorithm. From the estimated SC, the Enhanced Envelope Spectrum
[26] is identified by calculating the mean, absolute value of the SC in the direction of the
spectral frequency. The Enhanced Envelope Spectrum is used in this method, and the
flowchart is shown in Figure C.5 (c).

C.2.9.3 Performance metrics

Two performance metrics are created to quantify the performance of the WCCS and
the three other methods. The first metric evaluates the method’s ability to discern the
bearing characteristic fault frequency harmonics from the noise floor (mean value of the
spectrum), and is calculated using:
nh
!
1 X
P1 = Yh,j /mean{Y } , (C.19)
nh j=1

where nh is the number of harmonics in the spectrum, and Yh,j is the amplitude of
harmonic j. P 1 represents the mean ratio of the harmonic values to the noise floor. This
dimensionless ratio is comparable between each method, as the amplitude-scale of each
spectrum is widely different from each other.
The second metric evaluates the method’s ability to discern the side-bands from the
noise floor, and is calculated using:
nh
1 X
P2 = ns,j , (C.20)
nh j=1

where ns,j is the number of side-bands linked to harmonic j in the spectrum. Visible
side-bands are marked in the spectra and are expected to be prominent compared to the
noise floor. P 2 represents the mean number of visible side-bands per harmonic in the
spectrum.

C.3 Results
Vibration data is collected from an accelerated life-time test as described in Section C.2.1.
In this section, the measurements recorded at 20 rpm will be used to show the performance
of the WCCS, and how it compares to the three other methods.
 3.0 105 (b)
(a)
2.5 104
3
2.0 1 2 103
Kurtosis

Kurtosis
1.5
102
1.0
101
0.5
100
0.0
4.5 4.6 4.7 4.8 4.9 5.0 5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 5.9 5.6 5.7 5.8 5.9 6.0 6.1
Million Revolutions passed Million Revolutions passed

Figure C.6: Kurtosis of the vibration datasets measured at 20 rpm shaft speed. (a) The
kurtosis at an early stage of damage. (b) The kurtosis at an advanced stage of damage.

C.3.1 Kurtosis trend

The kurtosis is a time domain feature that can be used to indicate damage in the bearing.
This is because a high kurtosis value implies an impulsive signal containing shocks from,
i.e., bearing fault impacts. To show the damage propagation of the bearing during the
accelerated life-time test, the kurtosis is calculated using (C.18) for the 20 rpm datasets,
and shown in Figure C.6. In this figure, there are two graphs. (a) and (b) both shows the
kurtosis of the vibration data, but at different y- and x-axis scales. The x-axis indicates
how many million revolutions the bearing has rotated before the data is recorded. Initially,
the kurtosis value is low, but it rises to very high values when the bearing is sufficiently
damaged. In Figure C.6 (a) there are three points that are marked. These will be referred
to as kurtosis points in the following, and they indicate changes in the bearing health.
Specifications for the datasets linked to these points are given in Table C.2. In the next
Subsection, the WCCS is used to diagnose the bearing for faults at these three points.

C.3.2 Whitened Cross-correlation Spectrum

The proposed method from Section C.2.9.1 generates the WCCS which is used to ana-
lyze the vibration data at the various kurtosis points. At kurtosis point 1, there is no
notable change in the kurtosis value. However, the WCCS generated from the proposed
method shows a significant change compared to earlier measurements. The spectrum
of the whitened vibration signal, the envelope spectrum, and the WCCS are shown in
Figures C.7 (a)-(c), respectively. Here, the order spectra up to 50 orders are displayed.
In addition, lines that indicate harmonics of O2BS and side-bands are also shown. From
Table C.1, the characteristic fault frequency for a roller fault is 6.66 orders, and the side-
bands are spaced apart by the fundamental train at 0.43 orders (OF T ). Due to roller slip
and non-zero contact angle, the actual frequency is observed at 6.71 orders (deviation
 Table C.2: Specifications of the datasets at the three kurtosis points.
Kurtosis point → Point 1 Point 2 Point 3

Sampling frequency Fs (kHz) 51.2 51.2 51.2

Duration (s) 153 153 153
Million revolutions passed 5.37 5.46 5.76
Number of samples 7,833,600 7,833,600 7,833,600
Kurtosis of raw data kurt{V } 1.14 1.10 1.41
Shaft speed θ̇(ref ) (Hz) 0.33 0.33 0.33
Fault type Roller Outer-race Inner-race
Max ARM order pmax 22,391 22,391 22,391
Optimal ARM order popt 22,325 22,160 22,094
Kurtosis of whitened data kurt{Vw } 0.69 0.65 0.61

of +0.75%). Examining Figure C.7 (a) reveals that the low-frequency vibration signal
contains multiple side-bands around harmonics of the characteristic frequency. However,
the first five harmonics are not readily visible. On the contrary, the envelope spectrum
in Figure C.7 (b) shows small signs of the harmonics, and a few side-bands. Incorporat-
ing the cross-correlation fuses these two spectra, and amplifies the vibration components
that are visible in both. Figure C.7 (c) show the WCCS. Here, the perceived noise-floor
is smaller than in Figure C.7 (a), and the harmonics from Figure C.7 (b) are shown as
discrete peaks. In addition, a prominent peak at the OF T is also visible due to the strong
vibration occurring when the damaged roller is in the radial load zone. From these re-
sults, it is shown that even if a characteristic peak is only readily visible on one of the two
spectra, it is preserved after the cross-correlation step. This phenomenon is best observed
at the second harmonic (2x 6.71 orders), where the envelope spectrum in Figure C.7 (b)
shows a small peak, while the whitened vibration spectrum in Figure C.7 (a) does not
indicate a significant peak. Even with these signals, the WCCS in Figure C.7 (c) still
preserves the second harmonic peak from the envelope spectrum. Therefore, it should be
unproblematic to use the proposed method even if the deterministic signal shows small
to no signs of the bearing fault.
By visual inspection between the three spectra in Figure C.7, it is evident that the
WCCS is easier to analyze than either of the two other spectra. From these results, it is
clear that the bearing suffers from a roller fault at this stage. For comparison purposes, the
WCCS from a dataset captured ≈10,000 revolutions earlier is obtained, and the spectrum
 ×10−5
6
(a)
Harmonics of 2xBS Side-bands
5

4
Amplitude

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

×10−4
8
(b)
Harmonics of 2xBS Side-bands

6
Amplitude

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

×10−2
1.50
(c)
Harmonics of 2xBS Side-bands Mean
1.25

1.00 OF T
Amplitude

0.75

0.50

0.25

0.00
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

Figure C.7: The WCCS used to diagnose a roller fault at kurtosis point 1 in Figure C.6.
Harmonics and side-bands linked to the roller fault are marked in all three subfigures.
(a) The spectrum of the whitened vibration signal Vw . (b) The spectrum of the envelope
Venv . (c) The proposed WCCS and a red stapled line which corresponds to the mean
spectrum value.

is shown in Figure C.8. Here, there are no prominent amplitudes related to a roller fault,
and the bearing seems undamaged. There are, however, four other prominent peaks in
the spectrum, and these are marked. 14OS is equal to two times the motor speed due
to the gearbox. This vibration component was not attenuated in the whitening process
and is hence noise in the spectrum. The second peak is marked as OBP O which is the
characteristic frequency for an outer race fault at roughly 5.12 orders. It should normally
 ×10−2
1.50
14OS
1.25
OBP O
1.00
Amplitude

OBP O −1
0.75 OBP O +1

0.50

0.25

0.00
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

Figure C.8: The WCCS computed 10,000 revolutions prior to kurtosis point 1 in Figure
C.6. The roller damage is not visible.

×10−3
5

4 OBP O
14OS
Amplitude

3
OBP O −1 OBP O +1
2

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

Figure C.9: Baseline WCCS at the start of the accelerated life-test. A peak at the char-
acteristic frequency for an outer-race fault is visible, however the bearing is undamaged
at this stage.

be visible if there is an outer-race fault. However, the authors believe that the outer race
is undamaged at this point. That is because vibration at this cyclic frequency can also
be spotted at the beginning of the accelerated life-time test. A baseline WCCS after the
initial run-in of the healthy bearing is shown in Figure C.9. In this spectrum, a peak at
the characteristic frequency for an outer-race fault is still visible. As the bearing isn’t
damaged at this point, the authors conclude that there is vibration at the OBP O frequency
regardless of the bearing condition. The authors suspect that rollers passing the radial
load zone may cause the vibration at the OBP O frequency. This passing may cause the
shaft to move slightly up and down, which would explain the vibration. It could also be
from stressing and de-stressing a roller once it passes the radial load zone. Additionally,
the 1 order side-bands in Figures C.8 and C.9 are effects of amplitude-modulation of the
OBP O vibration. This could be caused by shaft mass-unbalance, or if the shaft is slightly
bent. Nevertheless, to accurately diagnose an outer-race fault using the proposed method,
there must be several stronger harmonics in the spectrum, as a single peak appears to
 ×10−4
1.0
(a)
Harmonics of BPO
0.8
Amplitude

0.6

0.4

0.2

0.0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

×10−4
8
(b)
Harmonics of BPO

6
Amplitude

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

×10−2
4
(c)
Harmonics of BPO

3
Amplitude

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

Figure C.10: The WCCS used to diagnose an outer-race fault at kurtosis point 2 in Figure
C.6. Harmonics linked to the outer-race fault are marked in all three subfigures. (a) The
spectrum of the whitened vibration signal Vw . (b) The spectrum of the envelope Venv .
(c) The proposed WCCS.

always be visible.
Moving on, kurtosis point 2 from Figure C.6 marks a new health state for the bearing.
The spectrum of the whitened vibration signal, the envelope spectrum, and the WCCS
are shown in Figures C.10 (a)-(c), respectively. The roller fault is no longer visible in
the WCCS in Figure C.10 (c). Instead, harmonics linked to an outer race fault are very
prominent in the spectrum. The characteristic fault frequency for outer-race faults is
5.12 orders with no side-bands according to Table C.1. The actual observed harmonic
frequency is 5.18 orders (deviation of +1.17%). Compared to a healthy state, there are
now multiple strong harmonics in the spectrum, which strengthen the probability that
there is an outer-race fault. Examining Figure C.10 (a) and (b) reveals that the outer-race
vibration peaks in the WCCS mostly arrive from the spectrum of the whitened vibration
signal. This is a second case showing that even if only one of the two input spectra shows
significant signs of bearing vibration, the resulting WCCS does not suffer significantly
from this. The envelope spectrum in Figure C.10 doesn’t show any prominent harmonics
at the outer-race characteristic frequency. The outer-race fault stays visible for quite
some time until kurtosis point 3 which marks yet a change of state in the bearing. The
spectrum of the whitened vibration signal, the envelope spectrum, and the WCCS are
shown in Figures C.11 (a)-(c), respectively. The WCCS in Figure C.11 (c) shows multiple
harmonics and side-bands linked to an inner-race fault. The fault is characterized by
strong harmonics at multiples of 6.88 orders accompanied by side-bands spaced 1 order
away from the harmonic according to Table C.1. In the observed spectrum, the actual
harmonic frequency is 6.82 (deviation of -0.87%). The peaks are very prominent, and there
are multiple side-bands, making it easy to conclude the type of fault. Harmonics from the
outer-race fault may also be observed in the spectrum, however they are not marked to
avoid clutter in the plot. Examining the whitened vibration spectrum in Figure C.11 (a)
shows prominent peaks at the third and higher harmonics. Meanwhile the first and second
harmonic are visible in the envelope spectrum in Figure C.11 (b). The cross-correlation of
the two signals make a clearer spectrum as shown in Figure C.11. Shortly after kurtosis
point 3, the kurtosis value increases drastically as seen on the chart in Figure C.6 (b), and
the bearing is approaching the end of its useful life. Luckily, the WCCS highlighted roller
faults nearly 500,000 rounds prior to kurtosis increasing drastically. With an expected
life-time of ≈ 6 million revolutions, this corresponds to 0.5/6 = 8.33% of remaining useful
life after the first fault was diagnosed. In the next subsection, the performance of the
WCCS is compared to the three other methods.

C.3.3 Comparison to other methods

For the comparisons shown in this subsection, the dataset resulting in kurtosis point 1
from Figure C.6 is used, because early fault detection is the most critical. The WCCS
shown in Figure C.7 (c) indicates that there is a roller fault in the bearing. This spectrum
is compared to the three other processing methods (A, B, and C), whose algorithms are
elaborated in Section C.2.9.2.
 ×10−4
1.0
(a)
Harmonics of BPI Side-bands
0.8
Amplitude

0.6

0.4

0.2

0.0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

×10−4
8
(b)
Harmonics of BPI Side-bands

6
Amplitude

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

×10−1
1.0
(c)
Harmonics of BPI Side-bands
0.8
Amplitude

0.6

0.4

0.2

0.0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

Figure C.11: The WCCS used to diagnose an inner-race fault at kurtosis point 3 in Figure
C.6. Harmonics and side-bands linked to an inner-race fault are marked in all three
subfigures. (a) The spectrum of the whitened vibration signal Vw . (b) The spectrum of
the envelope Venv . (c) The proposed WCCS.

Method A: Envelope Spectrum

The resulting envelope spectrum from method A is shown in Figure C.12. The spectrum
barely highlights the roller fault harmonics and its side-bands. They could easily be
ignored due to the first harmonic’s low amplitude. Shaft and motor vibrations, marked
as OS and 14OS , are the most prominent peaks in the spectrum. Therefore, method A is
not suitable for diagnosing faults on the presented dataset.
 ×10−4
7
14OS Harmonics of 2xBS Side-bands Mean
6

5
Amplitude

3
OS
2

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

Figure C.12: The resulting Envelope Spectrum from using method A. The roller fault
harmonics are barely visible, while shaft and motor vibration dominate the spectrum.
The mean value of the spectrum is marked for comparison purposes.

Method B: Kurtogram

The resulting Kurtogram from method B is shown in Figure C.13. From this result, the
chosen narrow-band is centered at 24,800 Hz with a bandwidth of 1600 Hz, because it
represents the highest kurtosis. The squared envelope spectrum of the band-pass filtered
signal is returned from this method and shown in Figure C.14. In this spectrum, it is
possible to discern seven harmonics related to a roller fault. However, motor-synchronous
vibration at 14OS and 35OS dominates the spectrum. The harmonics and side-bands of
the roller fault frequency are more prominent compared to the non-filtered envelope spec-
trum. However, compared to the WCCS shown in Figure C.7 (c), this squared envelope
spectrum has a higher noise-floor.

Method C: Fast Spectral Correlation

The resulting Enhanced Envelope Spectrum (EES) from method C is shown in Figure
C.15. In the EES, the harmonics linked to the roller fault are small compared to the
vibration from the shaft and motor. The spectrum is dominated by peaks at 1OS , 14OS ,
and 35OS . The harmonics linked to the roller fault are partially visible, however they are
hard to discern from the noise-floor. Compared to the WCCS, it is harder to detect the
harmonic peaks, and there are fewer side-bands.

C.3.3.1 Performance metrics

The performance of each method is quantified using the performance metrics from Section
C.2.9.3. The spectra indicating a roller fault at kurtosis point 1 in Figure C.6 is used to
evaluate the performance, and the values are given in Table C.3. The metrics show that
the WCCS scores the best among the four different spectra. The Fast Kurtogram (B) is
 fb-kurt.2 - K max =18.2 @ level 4, Bw= 1600Hz, fc =24800Hz
18
0
1 16
1.6
14
2
2.6 12
3
level k 10
3.6
4
8
4.6
5 6

5.6
4
6
6.6 2

7
0 0.5 1 1.5 2 2.5
frequency [Hz] 104

Figure C.13: Resulting Kurtogram in method B indicating a maximum of kurtosis at a

central frequency of 24,800 Hz and a bandwidth of 1600 Hz. The kurtosis value at that
narrow-band is 18.2.

×10−6
3.0
14OS Harmonics of 2xBS Side-bands Mean
2.5

2.0
Amplitude

1.5
35OS
1.0

0.5

0.0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

Figure C.14: Squared envelope spectrum of the band-pass filtered vibration signal in
Method B. The roller fault harmonics and side-bands are visible, and the mean value of
the spectrum is marked for comparison purposes.

second best, and the Envelope Spectrum (A) and the Fast Spectral Correlation (C) are
tied for the third place. Further, the resulting spectrum from methods A, B, and C also
contained strong peaks from the shaft and motor vibration, which reduces the visibility
of the peaks linked to bearing faults. Therefore, based on these criteria, the WCCS is
the best spectrum to perform fault diagnosis on among the four. It should be noted
that the minimum value of the Enhanced Envelope spectrum was subtracted before the
criteria was calculated, as the spectrum shown in Figure C.15 appears to have an offset
in amplitude.
 ×10−10
5
14OS Harmonics of 2xBS Side-bands Mean
4
Amplitude

2
35OS
OS
1

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

Figure C.15: Resulting Enhanced Envelope Spectrum from the Fast Spectral Correlation
(Method C). The roller fault harmonics and side-bands are barely visible, and the mean
value of the spectrum is marked for comparison purposes.

Table C.3: Performance metric for each method. The WCCS scores the highest in both
categories.
Method P1 P2

WCCS 5.49 6.57

A 2.95 1.42
B 4.42 3.85
C 2.65 2.00

C.3.4 Supplementary benchmark case study

The bearing fault database from the Case Western Reserve University (CWRU) [27] is
a standard benchmark for diagnostics algorithm performance. Smith and Randall [28]
published an extensive benchmark study where the performance of three established di-
agnostic techniques were evaluated on all available datasets. The three methods involved
the squared envelope spectrum (SES); pre-whitening before SES; and pre-whitened Spec-
tral kurtosis [25] and SES.
A dataset recorded during healthy conditions is examined first. This is to investigate
the phenomenon of a visible vibration at the frequency of the outer-race OBP O . The
dataset used is named 100DE in which the vibration is measured at the drive-end, and
there is no damage on either fan-end or drive-end bearing. Relevant specifications are
given in Table C.4. The WCCS of this dataset is shown in Figure C.16.
The resulting WCCS show multiple peaks at integer multiples of the shaft speed, in
addition to other non-integer peaks. The most interesting part is the inclusion of a peak
at OBP O and two side-bands spaced 1 order away from it. This is the same phenomenon
 Table C.4: Specifications for the tested CWRU datasets. DE=Drive-end, FE=Fan-end
Dataset number → 100DE 275DE

Sampling frequency Fs (kHz) 48 12

Duration (s) 10.11 10.11
Accelerometer location DE DE
Number of samples N 485,335 121,351
Kurtosis of raw data -0.0428 1.25
Shaft speed θ̇(ref ) (Hz) 28.83 29.58
Fault type Healthy FE Inner-race
DE ball pass order outer-race OBP O 3.5961 -
FE ball pass order inner-race OBP I - 4.974
Max ARM order pmax 308 82
Optimal ARM order popt 308 82
Kurtosis of whitened data 0.0708 1.22

×10−3
3.0

2.5

2.0
Amplitude

OBP O +1
1.5
OBP O −1 OBP O
1.0

0.5

0.0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Orders

Figure C.16: The WCCS on the CWRU healthy dataset 100DE recorded on the drive-end.

shown in the healthy case of the first database in Figure C.8. It was suggested that the
vibration at OBP O comes from the passing of rollers over the radial load zone. A bent
shaft would cause this OBP O vibration to be modulated by the shaft frequency, which
would result in the side-bands spaced apart by 1 order.
The performance of the WCCS is also tested on a dataset recorded with a known
fault. A particular dataset (275DE) with an inner-race fault located at the fan-end (FE),
and the vibration recorded at the drive-end (DE), proved to be difficult to diagnose using
the aforementioned benchmark methods ([28], Table B4). The results are indicated as
“partially successful” for the first two methods, and “not successful” for the last one as
 ×10−3
8
(a)
Harmonics of BPI Side-bands

6
Amplitude

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

×10−2
1.0
(b)
Harmonics of BPI Side-bands
0.8
Amplitude

0.6

0.4

0.2

0.0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

×10−1
7
(c)
Harmonics of BPI Side-bands
6

5
Amplitude

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50
Orders

Figure C.17: The WCCS used to diagnose an inner-race fault located at the fan-end of the
CWRU dataset. The vibration was measured at the drive-end. Harmonics and side-bands
linked to an inner-race fault are marked in all three subfigures. (a) The spectrum of the
whitened vibration signal Vw . (b) The spectrum of the envelope Venv . (c) The proposed
WCCS.

shown in Table B4 in [28]. Performing diagnosis on this dataset is difficult as the bearing
vibration must pass through the noisy induction motor. Relevant specifications for this
dataset are given in Table C.4. Note that the ball-pass inner race order is very close to 5
shaft orders, which may cause problems during spectrum analysis.

The spectrum of the whitened vibration signal, the spectrum of the envelope, and the
WCCS are shown in Figures C.17 (a)-(c), respectively. The vibration spectrum in Figure
C.17 (a) shows 10 harmonics of the OBP I , and multiple side-bands spaced one order away
from the harmonic. However, there also multiple other peaks present which makes the
analysis more difficult. The envelope spectrum in Figure C.17 (b) doesn’t contain as many
peaks, but most are related to the bearing fault. However, from the third harmonic, the
amplitude of the OBP I harmonics diminish in value. The WCCS in Figure C.17 (c) is
capable of combining the harmonics and side-bands of the whitened vibration spectrum
and the envelope spectrum. This spectrum has a lower perceived noise-floor, and fewer
peaks that are not related to the bearing fault. From these results it is clear that the
WCCS is useful to successfully diagnose the fan-end bearing, and that it is easier to
analyze than either the whitened vibration spectrum or the envelope spectrum.

C.4 Conclusions
The Whitened Cross-correlation Spectrum (WCCS) is proposed in this paper for di-
agnosing bearing faults. The method consists of whitening the vibration signal using
synchronous average and an autoregressive model. The key feature of the method is
the cross-correlation between the selectively whitened vibration signal and its envelope.
This correlation results in an element-wise multiplication of the frequency spectra of the
two signals. It has been shown in the literature that both the raw deterministic vibra-
tion signal and the envelope may contain low-frequency signal components related to the
bearing fault. Therefore, this correlation causes frequency components that are visible
in both to be amplified, while components that do not share correlation are attenuated.
The computational complexity of the time-domain convolution can also be minimized
by down-sampling the whitened vibration signal and the envelope prior to computing the
cross-correlation. Vibration data from a bearing accelerated life-time test has been used to
verify the diagnostic performance of WCCS. The method aids to accurately diagnose the
bearing for all three fault types before the bearing is seriously worn out. The early fault
was identified at ≈ 8.33% of remaining useful life, which should be long enough to sched-
ule a machine overhaul. Compared to other methods in the literature that are similarly
computationally heavy, the WCCS provided more prominent harmonics and side-bands
compared to the other methods. Additionally, a difficult dataset from the Case Western
Reserve University (CWRU) has been successfully diagnosed using the WCCS.

Funding
This work has been funded by the Ministry of Education and Research in Norway.
REFERENCES

[1] C. Mishra, A. K. Samantaray, and G. Chakraborty. Rolling element bearing de-

fect diagnosis under variable speed operation through angle synchronous averaging of
wavelet de-noised estimate. Mechanical Systems and Signal Processing, 72:206–222,
2016. doi:10.1016/j.ymssp.2015.10.019.

[2] J. Altmann and J. Mathew. Multiple Band-Pass Autoregressive Demodulation for

Rolling-Element Bearing Fault Diagnosis. Mechanical Systems and Signal Processing,
15(5):963–977, 2001. doi:10.1006/mssp.2001.1410.

[3] (ISO). 22096 condition monitoring and diagnostics of machines - acoustic emission,
2007.

[4] T. R. Lin, E. Kim, and A. C. C. Tan. A practical signal processing approach

for condition monitoring of low speed machinery using Peak-Hold-Down-Sample al-
gorithm. Mechanical Systems and Signal Processing, 36(2):256–270, 2013. doi:
10.1016/j.ymssp.2012.11.003.

[5] M. Elforjani and D. Mba. Accelerated natural fault diagnosis in slow speed bearings
with acoustic emission. Engineering Fracture Mechanics, 77(1):112–127, 2010. doi:
10.1016/j.engfracmech.2009.09.016.

[6] N. Jamaludin, D. Mba, and R. H. Bannister. Condition monitoring of slow-speed

rolling element bearings using stress waves. Proceedings of the Institution of Mechan-
ical Engineers, Part E: Journal of Process Mechanical Engineering, 215(4):245–271,
2001. doi:10.1177/095440890121500401.

[7] A. Widodo, E. Y. Kim, J.-D. Son, B.-S. Yang, A. C. C. Tan, D.-S. Gu, B.-K. Choi, and
J. Mathew. Fault diagnosis of low speed bearing based on relevance vector machine
and support vector machine. Expert Systems with Applications, 36(3):7252–7261, 2009.
doi:10.1016/j.eswa.2008.09.033.

[8] M. Žvokelj, S. Zupan, and I. Prebil. Multivariate and multiscale monitoring of large-
size low-speed bearings using Ensemble Empirical Mode Decomposition method com-

141
 bined with Principal Component Analysis. Mech. Syst. Signal Pr., 24(4):1049–1067,
2010. doi:10.1016/j.ymssp.2009.09.002.

[9] N. Jamaludin and D. Mba. Monitoring extremely slow rolling element bearings: Part
I. NDT and E International, 35(6):359–366, 2002. doi:10.1016/S0963-8695(02)
00006-3.

[10] E. Y. Kim, A. C. C. Tan, J. Mathew, and B. S. Yang. Condition monitoring of low

speed bearings: A comparative study of the ultrasound technique versus vibration
measurements. Australian Journal of Mechanical Engineering, 5(2):177–189, 2008.
doi:10.1007/978-1-84628-814-2_21.

[11] M. Elforjani and D. Mba. Detecting natural crack initiation and growth in slow
speed shafts with the Acoustic Emission technology. Engineering Failure Analysis,
16(7):2121–2129, 2009. doi:10.1016/j.engfailanal.2009.02.005.

[12] M. Elforjani and D. Mba. Condition Monitoring of Slow-Speed Shafts and Bear-
ings with Acoustic Emission. Strain, 47(SUPPL. 2):350–363, 2011. doi:10.1111/j.
1475-1305.2010.00776.x.

[13] N. Tandon and A. Choudhury. A review of vibration and acoustic measurement meth-
ods for the detection of defects in rolling element bearings. Tribology international,
32(8):469–480, 1999. doi:10.1016/S0301-679X(99)00077-8.

[14] K. R. Fyfe and E. D. S. Munck. Analysis of computed order tracking. Mechanical

Systems and Signal Processing, 11(2):187–205, 1997. doi:10.1006/mssp.1996.0056.

[15] R. B. Randall and J. Antoni. Rolling element bearing diagnostics—a tutorial. Me-
chanical Systems and Signal Processing, 25(2):485–520, 2011. doi:10.1016/j.ymssp.
2010.07.017.

[16] Z. Feng and M. J. Zuo. Vibration signal models for fault diagnosis of planetary
gearboxes. Journal of Sound and Vibration, 331(22):4919–4939, 2012. doi:10.1016/
j.jsv.2012.05.039.

[17] G. U. Yule. On a method of investigating periodicities in disturbed series, with special

reference to Wolfer’s sunspot numbers. Philosophical Transactions of the Royal Society
of London, 226:267–298, 1927. doi:10.1098/rsta.1927.0007.

[18] G. Walker. On periodicity in series of related terms. Proceedings of the Royal Society
of London, 131(818):518–532, 1931. doi:10.1175/1520-0493(1931)59<277:OPISOR>
2.0.CO;2.
[19] W. Wang and A. K. Wong. Autoregressive Model-Based Gear Fault Diagnosis. Jour-
nal of Vibration and Acoustics, 124(2):172, 2002. doi:10.1115/1.1456905.

[20] V. C. M. N. Leite, J. G. B. da Silva, G. F. C. Veloso, L. E. B. da Silva, G. Lambert-

Torres, E. L. Bonaldi, and L. E. d. L. de Oliveira. Detection of localized bearing faults
in induction machines by spectral kurtosis and envelope analysis of stator current.
IEEE Transactions on Industrial Electronics, 62(3):1855–1865, 2015. doi:10.1109/
TIE.2014.2345330.

[21] N. Sugiura. Further analysts of the data by akaike’s information criterion and
the finite corrections: Further analysts of the data by akaike’s. Communications in
Statistics-Theory and Methods, 7(1):13–26, 1978. doi:10.1080/03610927808827599.

[22] P. D. McFadden and J. D. Smith. Model for the vibration produced by a single point
defect in a rolling element bearing. Journal of Sound and Vibration, 96(1):69–82, 1984.
doi:10.1016/0022-460X(84)90595-9.

[23] R. B. Randall, J. Antoni, and S. Chobsaard. The relationship between spectral

correlation and envelope analysis in the diagnostics of bearing faults and other cyclo-
stationary machine signals. Mechanical Systems and Signal Processing, 15(5):945–962,
2001. doi:10.1006/mssp.2001.1415.

[24] M. R. Shahriar, P. Borghesani, and A. C. C. Tan. Electrical Signature Analysis-

Based Detection of External Bearing Faults in Electromechanical Drivetrains. IEEE
Transactions on Industrial Electronics, 65(7):5941–5950, 2018. doi:10.1109/tie.
2017.2782240.

[25] J. Antoni. Fast computation of the kurtogram for the detection of transient faults.
Mechanical Systems and Signal Processing, 21(1):108–124, 2007. doi:10.1016/j.
ymssp.2005.12.002.

[26] J. Antoni, G. Xin, and N. Hamzaoui. Fast computation of the spectral correlation.
Mechanical Systems and Signal Processing, 92:248–277, 2017. doi:10.1016/j.ymssp.
2017.01.011.

[27] Case Western Reserve University Bearing Data Center Website. 2015. URL: http:
//csegroups.case.edu/bearingdatacenter/home.

[28] W. A. Smith and R. B. Randall. Rolling element bearing diagnostics using the Case
Western Reserve University data: A benchmark study. Mechanical Systems and Signal
Processing, 64-65:100–131, 2015. doi:10.1016/j.ymssp.2015.04.021.
Paper D

Multi-band Identification for

Enhancing Bearing Fault Detection
in Variable Speed Conditions

Andreas Klausen, Huynh Van Khang and Kjell G. Robbersmyr

145
This paper has been submitted as:

A. Klausen, H. V. Khang and K. G. Robbersmyr. Multi-band Identification for

Enhancing Bearing Fault Detection in Variable Speed Conditions. Under review at
Mechanical Systems and Signal Processing.
 Multi-band Identification for Enhancing
Bearing Fault Detection in Variable Speed
Conditions
Andreas Klausen, Huynh Van Khang and Kjell G. Robbersmyr

University of Agder
Department of Engineering Sciences
Jon Lilletunsvei 9, 4879 Grimstad, Norway

Abstract – Rolling element bearings are crucial components in rotating ma-

chinery, and avoiding unexpected breakdowns using fault detection methods
is an increased demand in industry today. Variable speed conditions render a
challenge for vibration-based fault diagnosis due to the non-stationary impact
frequency. Computed order tracking transforms the vibration signal from
time domain to the shaft-angle domain, allowing order analysis with the enve-
lope spectrum. To enhance fault detection, the bearing resonance frequency
region is isolated in the raw signal prior to order tracking. Identification of
this region is not trivial but may be estimated using kurtosis-based methods
reported in the literature. However, such methods may fail in the presence of
relatively strong non-Gaussian noise. Cepstrum pre-whitening has also been
proposed for this diagnosis challenge, however the noise floor may increase
significantly from the normalization of the entire spectrum. In this paper,
a new approach for identifying multiple resonance regions is proposed. The
proposed method highlights all resonance frequencies in the signal by com-
bining computed order tracking and cepstrum pre-whitening in a new way.
Simulations and experimental results prove the validity of the method, and
comparisons with two existing methods show the increase in effectiveness of
the proposed method.

D.1 Introduction
Rolling-element bearings, or bearings for short, are critical components in rotating ma-
chinery. Unexpected bearing failure may cause machine breakdown and unplanned stops,
followed by human safety risks and economic loss [1, 2, 3]. To avoid this scenario, crit-
ical machine components, such as bearings, should be monitored to detect irregularities
early. Condition monitoring systems with sensor data input have been used for decades
to diagnose rotating machinery for various faults. Commonly, the vibration measured

147
with an accelerometer is analyzed to detect bearing faults [4, 5, 6]. Electric motor stator
current [7, 8, 9] and acoustic emission [10, 11] are also reported as viable sensor sources
for bearing fault diagnosis. However, in this research, vibration signals are used towards
drive-train applications, where accelerometers are always in place.
Bearing fault diagnosis methods based on the theory of cyclostationary (CS) signals
have received substantial attention in both industry and academia [12].
Bearing vibration signals are weakly categorized as second order CS (CS2) signals due
to internal slip and varying radial/axial load ratios for each roller [13].
The envelope spectrum (ES) [14] is a high-frequency demodulation technique that is
often used for bearing diagnosis owing to its simplicity for identifying the cyclic frequency
of CS2 signals.
Such existing techniques are commonly used in industrial production, in which the
machinery mainly operates in a steady state, i.e. the motor shaft rotates at a near-fixed
speed. This is not the case for electric drive-trains that often operate under variable speed
conditions (VSC) based on driver’s command.
Under VSC, bearing fault impacts no longer occur at a fixed time interval as the rela-
tion between time and shaft angle is not linear. The time domain vibration signal can be
transformed to the shaft-angle domain by applying computed order tracking [15]. This
transformation requires knowledge of the shaft position acquired from a tachometer or
an encoder. Alternatively, the shaft phase can be extracted from high-energy harmonics
in the vibration signal [16], or the electric motor current [17]. The bearing resonance
vibration should be isolated using a band-pass filter prior to order tracking to detect
bearing faults effectively. Determining the resonance frequency may be achieved by a
hammer tap test [18], or from a thin-shell vibration computer program [19]. However,
these operations are complex and hard to generalize for multiple different bearings and
machines. An alternative is to estimate an optimal band for band-pass filtration using
the vibration signal itself. Kurtosis-based methods, such as spectral kurtosis [20], may be
used to identify optimal frequency regions. A high kurtosis value implies that the signal is
impulsive, such as the quickly damped impact vibration from bearing fault impacts. Con-
sequently, a frequency band with a high kurtosis value should contain bearing vibration.
Methods such as the fast Kurtogram [21] and the Protrugram [22] both use the kurtosis
to determine optimal frequency bands for filtering. Applications of these methods [23, 24]
show their applicability in determining the optimal filter band. However, in the case of
strong non-Gaussian noise or single impacts, kurtosis-based methods may fail to identify
the optimal frequency band. Cepstrum pre-whitening (CPW) has also been proposed for
fault diagnosis during VSC [25].
However, the spectrum normalization of CPW increases the noise floor which makes
it more difficult to detect fault-related peaks in the ES.
To address the mentioned challenges, a new method for identifying multiple resonance
frequency regions is proposed in this paper. Instead of relying on kurtosis, the proposed
method exploits the effect of resonance frequency spread due to order tracking, and the
spectrum normalization effect of CPW [26, 27]. Advantageously, the method is robust
against non-Gaussian noise and random impacts, and the noise floor is kept to a minimum
compared to CPW as the signal is band-pass filtered at the bearing resonance region(s).
After applying order tracking and CPW, the normalized spectrum is transformed back
to the time domain, and time variant resonance modes rises from normalized spectrum.
Raising this spectrum to a higher power allows separation resonance frequency modes,
which can be chosen for bandpass filtration. Simulations and experimental results from
three test rigs validate the proposed method. The rest of the paper is organized as follows.
An introduction to bearing fault diagnosis using a vibration accelerometer is given in
Section D.2. The algorithms used in the proposed method are elaborated in Section D.3.
Methods used for comparison purposes are detailed in Section D.4. Experimental results
from simulations and vibration datasets from three test rigs are shown in Section D.5.
Comparisons between the proposed method versus spectral kurtosis and CPW are given
in Sections D.6. Finally, conclusions are drawn in Section D.7.

D.2 Bearing fault diagnosis

A roller element bearing is made of four main components: an inner race fastened to the
rotating shaft; an outer race stationary inside a housing; rollers/balls rotating in between
the raceways; and a cage which keeps the distance between rollers constant. An exemplary
bearing with an outer race fault is shown to the left in Fig. D.1. Fault impacts of passing
rollers cause quickly damped vibration pulses at the resonance frequency of the bearing
as seen in Fig. D.1. The impact frequency, referred as a characteristic bearing frequency,
is determined by the kinematics of the bearing. The expected characteristic frequencies
associated with each fault type (inner race, outer race and roller) can be determined
with bearing dimensions, such as the number of rollers, the roller diameter, and the pitch
diameter [18]. A resonance frequency mode should be isolated using a band-pass filter
to remove frequency components, which are not related to the bearing impacts. Usually,
multiple frequency bands contain bearing resonance vibration as indicated in [19, 28].
The proposed method allows for identifying several resonance frequency bands that can
be used for band-pass filtration. To perform diagnosis, the Hilbert ES is analyzed for
prominent peaks at the characteristic frequencies of the bearing.
 Figure D.1: Simplified bearing condition monitoring setup.

D.3 Proposed method

D.3.1 Complete algorithm

The flow diagram of the proposed method is shown in Fig. D.2. At first, the vibration
signal x(t) is acquired using an accelerometer, and the shaft position signal θ(t) is captured
using an encoder.
A suitable frequency band [fc , fbw ] is first identified using the proposed method before
computing a band-pass filtered Hilbert envelope order spectrum [14].

D.3.2 Order tracking

The vibration signal is usually measured at a fixed sample rate in the time domain.
During constant speed operation, this corresponds to a near-fixed number of vibration
samples between bearing fault impacts. Under VSC, however, this correspondence is no
longer true. To acquire a fixed number of samples between impacts, the vibration signal
is re-sampled from the time domain to the shaft-angle domain (also known as the order
domain). The vibration signal is re-sampled using any interpolation-method of choice
with
fot (θ) = Interpolate{x = θ(t), y = x(t)} , (D.1)

where fot (θ) is a continuous description of the vibration signal in the order domain, where
values between the discrete samples are described using the desired interpolation method.
In this research, cubic spline interpolation is chosen to compromise between compu-
tational burden and accuracy [15]. Using fot (θ), the vibration signal is re-sampled at
Figure D.2: Flow diagram of the proposed method. x(t) and θ(t) are the vibration signal
and encoder signal, respectively. fc and fbw are the central frequency and bandwidth of
the identified resonance mode.

intervals of ∆θ with
xot [j] = fot (j · ∆θ) , (D.2)

where j ≥ 0 is an integer, xot [j] is the j’th sample of the order tracked vibration signal,
and (j · ∆θ) must be within limits of the measured θ(t).
It should be noted that the frequency spectrum of an order tracked signal is shown as
a function of orders, where the 1st order is equal the shaft speed. In addition to order
tracking, it is also possible to perform the inverse, i.e. time tracking, which transforms
the signal back to the time domain. Deterministic vibration components from shafts are
de-spread in the order spectrum, while the bearing resonance vibration is spread across a
larger area as the resonance vibration is a function of time. Numerical examples of order
tracking shaft and bearing vibration are shown in Fig. D.3. Figs. D.3 (a) and (b) show
the simulated shaft vibration and bearing resonance vibration, respectively. The shaft
vibration is the sum of 27 sine waves described in the shaft angle domain. A bearing with
an outer race fault is included in the simulation. Each roller impact causes a response
given by an underdamped second order system with a resonance frequency of 6000 Hz.
Each signal is simulated for 20 seconds with a shaft speed that ranges between 10 and
16 Hz. The frequency spectra of the shaft and bearing vibration are shown in Figs. D.3
 Vibration [m/s2 ]

Vibration [m/s2 ]
×10−2
1 1

0 0

−1
0.00 0.02 0.04 0.06 0.08 0.10 0.00 0.02 0.04 0.06 0.08 0.10
Time Time
(a) (b)
Vibration [m/s2 ]

Vibration [m/s2 ]
×105 ×101
0.75 5.0
0.50
0.25 2.5
0.00 0.0
0 100 200 300 400 0 2000 4000 6000 8000 10000
Frequency [Hz] Frequency [Hz]
(c) (d)
Vibration [m/s2 ]

Vibration [m/s2 ]
×105 ×101
0.75 5.0
0.50
0.25 2.5
0.00 0.0
0 10 20 30 40 0 200 400 600 800 1000
Orders Orders
(e) (f )

Figure D.3: Simulation showing the resonance frequency spreading effect, and determinis-
tic component de-spreading. (a) shaft vibration; (b) bearing vibration; (c) shaft vibration
frequency spectrum; (d) bearing vibration frequency spectrum; (e) shaft vibration order
spectrum; (f) bearing vibration order spectrum. One order is equal 10 Hz in this example.

(c) and (d), respectively. As seen in Fig. D.3 (c) shaft vibration components are spread
across the low-frequency region, while the bearing resonance is centered at 6000 Hz in
Fig. D.3 (d). After applying order tracking, the order spectrum of the shaft and bearing
vibration are shown in Figs. D.3 (e) and (f), respectively. In D.3 (e), the shaft vibration
components are sharpened, while the bearing vibration order spectrum in Fig. D.3 (f)
shows that the resonance vibration energy is spread across a larger frequency area. The
proposed method exploits this phenomenon to identify the bearing resonance frequency
region.

D.3.3 Cepstrum pre-whitening

As elaborated in [25], deterministic vibration components originating from shafts and gear-
boxes are periodic, but not sinusoidal, and thus form multiple harmonics in the frequency
domain. In the cepstral domain, these harmonics form a single peak at the quefrency
equal to the period of the deterministic signal component. However, bearing vibration
CS2 components do not form significant peaks in the cepstrum, as they are not exactly
periodic. Therefore, it is possible to eliminate deterministic components by performing a
series of liftering operations around the quefrencies of deterministic components [26, 27].
By setting the whole real cepstrum to a zero value, except at zero quefrency, all deter-
ministic components are attenuated. This method is referred as cepstrum pre-whitening
(CPW) and is simplified using the following [25]:
F {V }
 
−1
xcpw = F , (D.3)
|F {V }|
where F and F −1 are the forward and inverse Fourier transform, respectively.
Inspecting (D.3) reveals that CPW normalizes the frequency spectrum amplitude, i.e.
turns the spectrum flat.

D.3.4 Order tracking and cepstrum pre-whitening

As previously elaborated, resonance vibration is spread in the order domain, while deter-
ministic components are de-spread, and CPW normalizes the amplitude of all frequency
components. By combining CPW and order tracking, it is possible to highlight resonance
frequency modes in the frequency spectrum. This is visualized in Fig. D.4 and explained
hereafter: First, the resonance-to-other ratio (ROR) is defined as the ratio between res-
onance mode amplitude and the most prominent amplitude of other components in the
frequency spectrum.
The raw vibration spectrum is mostly dominated by deterministic low-frequency com-
ponents from shaft and gearbox vibration, therefore the ROR is lower than 1. By im-
plementing the order tracking algorithm, the ROR is even more reduced because of the
spread of resonance frequency and de-spread of deterministic components. Using CPW
afterwards causes all frequency components to get the same amplitude, and thus the ROR
is equal to 1. The inverse order tracking is used to spread deterministic components and
de-spread resonance frequency modes, thus achieving a ROR greater than 1. As the ROR
is greater than 1, raising the amplitude to a power increases the ROR even further. To
avoid singular dominant peaks in the spectrum, the amplitude is filtered using a forward-
backward first order IIR low pass filter, and afterwards raised to a power of 5. The final
spectrum is analyzed to identify the most prominent resonance frequency mode, and the
original signal is band-pass filtered around this mode to enhance bearing fault diagnosis.
While prominent resonance modes are manually chosen in this paper, it is possible to
automate the mode splitting using an automatic diagnosis system described in [28].

D.4 Comparison methods

D.4.1 Spectral kurtosis

Spectral kurtosis is a well-known method for identifying a suitable frequency band as a
high kurtosis value signifies impulsive signal components such as bearing impact vibration.
The fast Kurtogram [21] makes a good estimate of the spectral kurtosis at a fraction of
Figure D.4: Visualization of the ratio between resonance mode amplitude versus the most
prominent amplitude of other components in the frequency spectrum (ROR) after each
step in the proposed method.

the computational cost by calculating the kurtosis at multiple levels of low- and high-pass
filtration in an iterative manner. The frequency band with the highest kurtosis should
be the most suitable frequency band. The Kurtogram shows kurtosis values at different
central frequencies (x-axis) and bandwidth (y-axis), where an increase in level signifies a
decrease in bandwidth. Spectral kurtosis may fail if the signal is contaminated by strong
non-Gaussian noise or random impacts [22]. The vibration signal is filtered with the band
that has the highest kurtosis value, and the envelope order spectrum is calculated.

D.4.2 Cepstrum pre-whitening

The proposed method is also compared to the CPW algorithm in [25]. In that algorithm,
the signal is first order tracked, and discrete components are removed by means of CPW.
Afterwards, the envelope order spectrum of the resulting signal is computed.

D.4.3 Diagnosis score

When diagnosing a fault with an ES, there are two factors that determine the confidence
of the diagnosis: the number of characteristic bearing frequency harmonics, and their
prominence compared to the noise floor. With several harmonics it is easier to verify
the actual fundamental harmonic frequency, and prominence compared to the noise floor
decreases the chance of giving a false diagnosis. To compare the quality of the envelope
spectra obtained with the proposed method and the comparison methods, an automatic
bearing diagnosis system [28] is applied. This system can identify faults in the bearing
using an envelope order spectrum, and a diagnosis score is calculated based on the number
of harmonics and their prominence compared to the noise floor. In a simple form, the
score is calculated with
Nh
X Yh,i 2
S= i , (D.4)
i=1
3µn,i
where Nh is the number of prominent harmonics identified, Yh,i is the amplitude of the
i’th harmonic, and µn,i is the mean value of the noise floor around harmonic i. A high
number of harmonics is rewarded with a high score due to i2 , and prominence to the noise
floor increases the score gained per identified harmonic. Typically, a score lower than 10
can be caused by noise, while a score larger than 100 gives a high confidence of the fault
[28]. For a harmonic i to be identified as prominent,

Yh,i > 3µn,i and (Yns,i > 3µn,i or Yps,i > 3µn,i ) , (D.5)

where Yns,i and Yps,i are the negative and positive side-bands associated with a fault
type, respectively. The diagnosis score for each algorithm is directly compared with
experimental results to verify the performance of the proposed method.

D.5 Experimental Results

Simulations and experimental datasets from three test rigs are used to validate the effec-
tiveness of the proposed method. The first source [29] features a variable-speed capable
test rig installed with bearings that are pre-seeded with three different faults. The second
test rig is located at the Case Western Reserve University [30], and the chosen dataset is
recorded during VSC with an inner race fault. The final experimental dataset is from an
in-house test rig [31] where the bearing has been naturally worn through an accelerated
life-time test.

D.5.1 Simulation
A simulation of a bearing fault vibration signal is used to verify the applicability of the
proposed method to identify resonance frequency regions. A simulated signal consisting
of bearing and shaft vibration, two random impacts, and white Gaussian noise is given
by
n
X
xsim (t) = w(t) + Ab (θ̇)g ∗ δ(i/OO − θ(t))
i=1
m
X
+ Aj (θ̇) sin(2πjθ(t) + φj ) + I(t), (D.6)
j=1

where w(t) is white Gaussian noise, n is the number of outer race fault impulses during the
simulation, Ab is the bearing impact vibration amplitude as a function of shaft speed, g is
the impulse response of an underdamped second order system with a resonance frequency
of 6000 Hz, δ is the dirac-delta function, OO is the characteristic fault order for an outer
race fault, θ is the shaft position in rounds, m = 27 is the number of shaft vibration
harmonics, Aj is the amplitude of the j’th harmonic as a function of shaft speed, φj is
a random phase for the j’th harmonic, and I(t) is the vibration from two impacts with
high amplitude and a response following an underdamped second order system with a
resonance frequency of 3500 Hz.
The bearing and shaft vibration amplitudes are linearly dependent on the instanta-
neous shaft speed, and the random impacts are added to validate the proposed methods’
effectiveness to avoid highlighting frequency regions containing random impacts. Fig. D.5
shows results from using the proposed method on the simulated signal. Figs. D.5 (a) and
(b) show a 100 ms long snapshot of the shaft vibration and bearing vibration, respec-
tively, and Fig. D.5 (c) shows the shaft speed during the simulation. The full simulated
vibration xsim is shown in Fig. D.5 (d), and its raw frequency spectrum is shown in Fig.
D.5 (e). As observed, most of the energy is stored in the low-frequency deterministic shaft
components. After order tracking the signal, all shaft deterministic peaks are amplified
and sharpened as shown in Fig. D.5 (f). CPW is applied to normalize the spectrum as
shown in Fig. D.5 (g). The inverse order tracking causes the resonance frequency region
at 6000 Hz to de-spread, and the ROR is therefore increased above unity as shown in Fig.
D.5 (h). This allows raising the spectrum to a power to further separate the resonance
regions from the rest as shown in Fig. D.5 (i). Here, the amplitude-filtered spectrum
is raised to a power of 5 to increase ROR. As observed, there are no high peaks around
3500 Hz, indicating that the proposed method is robust against random impacts. Finally,
the most visible area in this spectrum (marked as a red square) is chosen for band-pass
filtration of the original vibration signal. A filter bandwidth of 50 orders (500 Hz in this
example) is used to retain several harmonics in the final ES. After applying the ES algo-
rithm shown in Fig. D.2, multiple harmonics of the outer race fault order OO = 3.54 are
detected using the automatic bearing diagnosis system, as shown in Fig. D.5 (j).

D.5.2 Test rig 1

The first test rig is installed with bearings that are pre-seeded with different faults, and
the motor is operated under VSC. The datasets used are provided as supplementary
materials in [29], and a simplified schematic of the test setup is shown in Fig. D.6.
Both the vibration signal and the shaft angular velocity is available in the repository.
The characteristic fault frequencies for the test bearing are: ball-pass inner race order
OI = 5.41, ball-pass outer race order OO = 3.59, two times ball (roller) spin order
OB = 4.75, and the fundamental cage order OC = 0.40.
 Vibration [m/s2 ]

Vibration [m/s2 ]
×10−2
1 1

0 0

−1
0.00 0.02 0.04 0.06 0.08 0.10 0.00 0.02 0.04 0.06 0.08 0.10
Time Time
(a) (b)

Vibration [m/s2 ]
Shaft Speed [Hz]

×101
5 Random
1.5 impacts
0
1.0
−5
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0
Time Time
(c) (d)
×104 ×105
Amplitude

Amplitude
0.75
2
0.50
1 0.25
0 0.00
0 2000 4000 6000 8000 10000 0 200 400 600 800 1000
Frequency [Hz] Orders
(e) (f )

2 5.0
Amplitude

Amplitude Amplitude-filtered
1 2.5

0 0.0
0 200 400 600 800 1000 0 2000 4000 6000 8000 10000
Orders Frequency [Hz]
(g) (h)
×101 ×10−3
Amplitude

Amplitude

1 1 Harmonic

0 0
0 2000 4000 6000 8000 10000 0 5 10 15 20 25 30
Frequency [Hz] Orders
(i) (j)

Figure D.5: Resonance band identification using the proposed method on a simulated
signal. (a) shaft vibration; (b) bearing vibration; (c) shaft speed; (d) measured vibration
signal; (e) frequency spectrum of raw signal; (f) order spectrum; (g) order spectrum after
CPW; (h) frequency spectrum after inverse order tracking with a red line showing the
amplitude-filtered signal; (i) amplitude-filtered frequency spectrum raised to power of 5,
with a suitable band-pass filter region marked in red; (j) envelope order spectrum after
band-pass filtration of original signal.

Figure D.6: Simplified schematic of the test setup in

Two times ball order spin is utilized as the roller hits the inner and outer race succes-
sively during one spin.
 ×104
Shaft Speed [Hz]

1.0

Amplitude
15
0.5
10
0.0
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5
Time [s] Frequency [kHz]
(a) 3 (b)
×10
1.5
Amplitude

Amplitude
1 2 3
1.0
1
0.5
0
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 0 5 10 15 20 25 30 35 40
Frequency [kHz] Orders
2 (c) 1 (d)
×10 ×10
1.0
Amplitude

Amplitude
Harmonic
0.5 1

0.0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Orders Orders
(e) (f )

Figure D.7: Results from diagnosing an outer race fault on test rig 1. (a) shaft speed
during measurement; (b) frequency spectrum of the raw vibration signal; (c) frequency
spectrum after using the proposed method, where three suitable band-pass filter areas
are marked; (d)–(f) envelope order spectra after band-pass filtration at the three marked
areas, respectively. Red triangles show identified harmonics related to the fault.

Dataset 1—outer race fault

The first dataset in the repository is recorded while a bearing with an outer race fault is
installed on the test rig, and the results are shown in Fig. D.7. Fig. D.7 (a) shows the shaft
speed during the measurement and Fig. D.7 (b) shows the frequency spectrum of the raw
vibration signal. As shown in Fig. D.7 (b), the low frequency deterministic components
dominate the spectrum. After applying the proposed method, the resulting spectrum is
shown in Fig. D.7 (c), in which several resonance frequency regions can be chosen due
to multiple peaks with high value. Therefore, three marked frequency regions are used
for obtaining the ES. Band-pass filtered envelope spectra of the three marked areas are
shown in Figs. D.7 (d), (e), and (f), respectively. In these spectra, harmonics of the
outer race fault order OO are shown as red triangles, and multiple prominent harmonics
are identified in all three spectra. The diagnosis score for each band is shown in Table
D.1, under “DS 1.1”. With the high scores on all three bands, the fault diagnosis of an
outer race fault is confirmed. These results demonstrate that multiple bearing resonance
frequencies may be captured in the vibration signal.

Dataset 2—roller fault

The second dataset from test rig 1 is recorded when the test bearing has a roller fault,
and the results are shown in Fig. D.8. The shaft speed varies between 10 and 15 Hz
Table D.1: Diagnosis score per band for each dataset using the proposed method. The
best mode for each dataset is written in bold. “DS X.Y” means “Test rig X - dataset Y”.
Band \ Dataset DS 1.1 DS 1.2 DS 1.3 DS 2 DS 3

Band 1 1615 1733 448 41 0

Band 2 1581 287 202 49 0
Band 3 895 328 212 154 166
Band 4 - - - 73 -
Band 5 - - - 26 -

as seen in Fig. D.8 (a), and the deterministic low-frequency components dominate the
raw vibration spectrum as indicated in Fig. D.8 (b). Applying the proposed method
reveals three suitable band-pass filter areas, which are marked in Fig. D.8 (c). With a
roller fault, the characteristic vibration patterns are harmonics of two-times roller spin OB
with side-bands given by the fundamental cage OC . Due to the radial load on bearings,
there may be little vibrations when the damaged roller is outside the radial load zone.
Therefore, the most visible vibration can be from impacts in the radial load zone, which
the damaged roller is passing once per fundamental cage revolution [18]. In case of
this dataset, only harmonics of OC are identified when analyzing the band-pass filtered
envelope order spectra in Figs. D.8 (d), (e), and (f). All of the envelope spectra show
multiple prominent harmonics at the cage frequency, which strongly verifies the diagnosis.
The high diagnosis score given in Table. D.1 under “DS 1.2” also verifies the effectiveness
of the proposed method to detect bearing faults.

Dataset 3—inner race fault

The final test is performed with an inner race fault in the bearing, and the results are
shown in Fig. D.9. Figs. D.9 (a) and (b) show the shaft speed and the frequency spec-
trum of the raw vibration signal, respectively. After applying the proposed method, the
highlighted resonance modes are shown in Fig. D.9 (c). The three areas marked with
red-stapled lines are chosen for band-pass filtered ES analysis due to their prominence.
The envelope order spectra for these three areas are shown in Figs. D.9 (d)–(f), respec-
tively. An inner race fault is characterized by harmonics at the inner race frequency OI
with side-bands located 1 order away due to the damaged part rotating in and out of the
radial load zone [18]. Therefore, the automatic diagnosis method searches for prominent
harmonics and side-bands. With this method, multiple harmonics and side-bands are
identified, which can be seen in the envelope spectra. The score for each band is given
in Table D.1 under “DS 1.3”. Due to low prominence of the higher order harmonics, the
 ×104
Shaft Speed [Hz]

Amplitude
1.0
15
0.5
10
0.0
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5
Time [s] Frequency [kHz]
(a) 2 (b)
×10
Amplitude

Amplitude
2 1 2 3 4

1 2

0
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 0 1 2 3 4 5 6
Frequency [kHz] Orders
1 (c) (d)
×10
4
Amplitude

Amplitude
1.0 Harmonic

0.5 2

0.0 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Orders Orders
(e) (f )

Figure D.8: Results from diagnosing a roller fault on test rig 1. (a) shaft speed during
measurement; (b) frequency spectrum of the raw vibration signal; (c) frequency spectrum
after using the proposed method, where three suitable band-pass filter areas are marked;
(d)–(f) envelope order spectra after band-pass filtration at the three marked areas, re-
spectively. Red triangles show identified harmonics related to the fault.

maximum score is not as great when compared to the two first datasets. However, with
a score greater than 100, the diagnosis confidence is high.

D.5.3 Test rig 2

Data from a second test rig is utilized to further quantify the performance of the proposed
method. This test rig is located at the Case Western Reserve University (CWRU), and
multiple datasets are available for download [30]. A simplified schematic of the test rig
is shown in Fig. D.10. Both fan-end (FE) and drive-end (DE) bearing faults at different
severity levels are available in the database, with accelerometer measurements taken at
the FE, DE and base (BA).
A study [32] thoroughly analyzed all datasets in the repository and benchmarked
multiple fault diagnosis techniques to determine their performance in detection of faults
using data from three accelerometers. Dataset 174 proved difficult to diagnose due to
speed variation, and consequently the three methods utilized resulted in a mis-diagnosis
as reported in Table B3 [32]. A bearing with an inner race fault (OI = 5.401) is installed
at the DE when dataset 174 is recorded, and in this research, the DE accelerometer data
is used to diagnose the machine.
The time-frequency spectrum of the enveloped vibration signal, shown in Fig. D.11,
reveals that the machine is operated under VSC during recording of this dataset. That
 ×104
Shaft Speed [Hz]

Amplitude
1.5
15
1.0
0.5
10
0.0
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5
Time [s] Frequency [kHz]
(a) 2 (b)
×10
Amplitude

Amplitude
1.5 1 2 3
1.0
1.0
0.5
0.5
0.0
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 0 5 10 15 20 25 30 35 40
Frequency [kHz] Orders
1 (c) 1 (d)
×10 ×10
2
Amplitude

Amplitude
Harmonic
2 Side-band
1

0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Orders Orders
(e) (f )

Figure D.9: Results from diagnosing an inner race fault on test rig 1. (a) shaft speed
during measurement; (b) frequency spectrum of the raw vibration signal; (c) frequency
spectrum after using the proposed method, where three suitable band-pass filter areas are
marked; (d f) envelope order spectra after band-pass filtration at the three marked areas,
respectively. Red and blue triangles show identified harmonics and side-bands related to
the fault.

Figure D.10: Simplified schematic of the test rig used at CWRU.

is because visible amplitude ridges show a reduction in frequency over time. The red-
stapled line has been identified as the third shaft order harmonic vibration. Integrating
this frequency ridge over time and dividing by three gives the relative position of the
shaft. Order tracking the raw vibration signal using this position gave high peaks at
the assumed shaft order and its second and third harmonics. Additionally, the identified
shaft speed is slightly below the supposed fixed-speed of 1797 rpm (29.95 Hz) [30], further
indicating that the identified shaft speed is correct.
Results from using the proposed method are shown in Fig. D.12. Fig. D.12 (a) shows
the identified shaft speed over time, and the initial speed of about 28.2 Hz is just below
the specified shaft speed at 29.95 Hz. Fig. D.12 (b) shows the frequency spectrum of
Figure D.11: Time-frequency diagram of the enveloped vibration data. An identified
ridge is marked as a red-stapled line. This ridge is identified as the third order of shaft
vibration.

the raw vibration signal. After using the proposed method, multiple resonance modes
are identified as shown in Fig. D.12 (c). Five bands are chosen for bandpass filtration to
properly investigate most of the resonance bands. The band-pass filtered envelope spectra
are shown in Fig. D.12 (d)-(h). As there is an inner race on the bearing, side-bands of 1
order should be visible together with the harmonics. However, no such side-bands were
identifiable on either of the envelope spectra. This may be the cause of improper order
tracking or speed estimation. Therefore, the diagnosis score is based on the harmonics of
OI alone. All the envelope spectra show harmonics of the fault, and one of the bands have
a higher score than 100 as indicated in Table D.1 under “DS 2.1”. These results show
that order tracking is necessary to diagnose the fault in this dataset, and the proposed
method aids in identifying suitable band-pass filter specifications.

D.5.4 Test rig 3

Data from an in-house test rig is also used to demonstrate the performance of the proposed
method. A 6008 type, 40 mm bore diameter roller element bearing is worn naturally
through an accelerated life-time test by applying radial and axial load. Details of the test
rig are given in [31], and a schematic is shown in Fig. D.14. The bearing has a dynamic
loading capacity of C = 17.8 kN and a static load capacity C = 11 kN. It was loaded with
a radial load of 9 kN and an axial load of 5 kN constantly during the test. The bearing
vibration was measured with an accelerometer, the radial shaft movement was measured
with an eddy current proximity sensor, and the shaft angular position is acquired using
an encoder. The position data from the proximity sensor is used for verification purposes,
 ×103
Shaft Speed [Hz]

Amplitude
28 0.75
0.50
26
0.25
24 0.00
0.0 0.2 0.4 0.6 0.8 1.0 1.2 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5
Time [s] Frequency [kHz]
(a) ×10 2 (b)
4
Amplitude

Amplitude
1 2 3 4 5
2
2
1

0
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 0 5 10 15 20 25 30 35 40
Frequency [kHz] Orders
(c) ×10 1 (d)
Amplitude

Amplitude
1.0
2
0.5

0 0.0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Orders Orders
(e) (f )
1.0
Amplitude

Amplitude
5.0 Harmonic
0.5
2.5

0.0 0.0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Orders Orders
(g) (h)

Figure D.12: Results from diagnosing an inner race fault on test rig 2. (a) shaft speed
during measurement; (b) frequency spectrum of the raw vibration signal; (c) frequency
spectrum after using the proposed method, where 5 suitable band-pass filter areas are
marked; (d)–(h) envelope order spectra after band-pass filtration at each identified mode.
Red triangles show identified harmonics elated to the fault.

while the vibration signal is used in this research. During the test, the bearing is run at 500
rpm to accelerate the lifetime. However, every 10 minutes the test rig controller enters
a measurement cycle. Within this cycle, vibration and proximity data was measured
during various configurations of constant speed and variable speed. After approximately
34.6 million revolutions, an outer-race fault is detected after analyzing data from the
eddy current proximity sensor. The characteristic outer race frequency for this bearing is
OO = 5.12 orders. In this research, a vibration measurement file captured during variable
speed operation right after detecting the outer race fault is used. The shaft speed reference
was set to mimic a wave generated using a Pierson–Moskowitz wave spectrum with a mean
speed at 50 rpm, significant wave height of 67 rpm, and a significant wave period of 10 s.
Using the proposed method yields the results shown in Fig. D.13. The recorded shaft
speed is shown in Fig. D.13 (a), and the raw vibration spectrum is shown in Fig. D.13 (b).
The raw spectrum contains peaks mostly at lower frequencies, but there are two single
peaks at about 10 kHz and 13 kHz. Using the proposed method yields the spectrum
shown in Fig. D.13 (c). In this spectrum, there are three areas of interest for bandpass
filtration. The envelope spectra are shown in Figs. D.13 (d)-(f). The spectrum from the
 ×104
Shaft Speed [Hz]

Amplitude
1.0
1
0.5

0 0.0
0 5 10 15 20 25 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0
Time [s] Frequency [kHz]
(a) 1 (b)
×10
Amplitude

Amplitude
1 2 3
2
2
1

0 0
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 0 5 10 15 20 25 30 35 40
Frequency [kHz] Orders
1 (c) (d)
×10
Amplitude

Amplitude
2 Harmonic
0.5
1

0.0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Orders Orders
(e) (f )

Figure D.13: Results from diagnosing an outer race fault on test rig 3. (a) shaft speed
during measurement; (b) frequency spectrum of the raw vibration signal; (c) frequency
spectrum after using the proposed method, where 3 suitable band-pass filter areas are
marked; (d)–(f) envelope order spectra after band-pass filtration. Red triangles show
identified harmonics related to the fault.

Figure D.14: Simplified schematic of the in-house accelerated life-time test rig

first mode shows no sign of any bearing fault, and this mode may just contain shaft speed
invariant noise. There are no signs of bearing fault in the second ES shown in Fig. D.13
(e) either. The only prominent peak is at 35 orders, which most likely sources from the
electric motor driving the test rig. The final mode at a high frequency beyond 16 kHz
contain some bearing fault related vibrations. The ES in Fig. D.13 (f) shows multiple
harmonics of OO , which verifies the diagnosed fault. Table D.1 shows the diagnosis score,
and only the third band has a score greater than 0. However, with a value larger than
 Max Kurtosis=1815.2 @ level 1.6, fc=12499Hz, fbw=8333Hz Max Kurtosis=26.1 @ level 7, fc=1464Hz, fbw=195Hz
0 1750 0 25
1 1
1.6 1500 1.6
2 2 20
2.6 1250 2.6

Kurtosis

Kurtosis
3 3
Level

Level
3.6 1000 3.6 15
4 4
4.6 750 4.6 10
5 5
5.6 500 5.6
6 6 5
6.6 250 6.6 Proposed center
7 7 0
0 5 10 15 20 25 0 5 10 15 20 25
Frequency [kHz] Frequency [kHz]
(a) (c)
Amplitude

Amplitude
Harmonic
100 500

0 0
0 5 10 15 20 25 30 35 40 0 1 2 3 4 5 6
Orders Orders
(b) (d)
Max Kurtosis=7.3 @ level 6.6, fc=2213Hz, fbw=260Hz Max Kurtosis=8.9 @ level 1, fc=6000Hz, fbw=12000Hz
0 7 0
1 1 8
1.6 6 1.6 7
2 2
2.6 5 2.6 6
Kurtosis

Kurtosis
3 3
Level

Level
3.6 4 3.6 5
4 4 4
4.6 3 4.6
5 5 3
5.6 2 5.6 2
6 6
6.6 1 6.6 1
7 0 7 0
0 5 10 15 20 25 0 5 10 15 20
Frequency [kHz] Frequency [kHz]
(e) (g)
Amplitude

Amplitude

Harmonic
Side-band 2000
500

0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Orders Orders
(f ) (h)
Max Kurtosis=238.7 @ level 2.6, fc=14933Hz, fbw=4266Hz
0
1
1.6 200
2
2.6
Kurtosis

3 150
Level

3.6
4
4.6 100
5
5.6 50
6
6.6
7 0
0 5 10 15 20 25
Frequency [kHz]
(i)
Amplitude

2000

0
0 5 10 15 20 25 30 35 40
Orders
(j)

Figure D.15: Results from using the fast Kurtogram detailed in Section D.4.1 on all the
experimental datasets. Pairs consisting of a Kurtogram and the resulting ES are given
for each dataset. (a)–(b) test rig 1, dataset 1; (c)–(d) test rig 1, dataset 2; (e)–(f) test rig
1, dataset 3; (g)–(h) test rig 2, dataset 1; (i)–(j) test rig 3, dataset 1. In a Kurtogram,
red lines indicate central frequencies identified using the proposed method. In envelope
spectra, red and blue triangles indicate prominent harmonics and side-bands related to
the fault, respectively.
100, there is high confidence with the diagnosis.

D.6 Comparisons

D.6.1 Spectral kurtosis

The fast Kurtogram detailed in Section D.4.1 is tested for all the experimental datasets to
check the performance of the proposed method. Results from applying this algorithm are
shown in Fig. D.15. Here, a pair consisting of a Kurtogram and an ES is shown for each
dataset. Starting with dataset 1 from test rig 1, the Kurtogram is shown in Fig. D.15
(a), while the resulting ES is shown in Fig. D.15 (b). In the Kurtogram, kurtosis values
for many combinations of frequency bands are displayed, and the band with the highest
kurtosis is deemed the most optimal band for band-pass filtration. Additionally, red lines
show central frequencies identified using the proposed method. As seen in Fig. D.15 (a),
the highest kurtosis value is found in a high frequency band with a central frequency of
12500 Hz and bandwidth of 8333 Hz. In comparison to the proposed method, this is a
much higher frequency band. The corresponding ES is shown in Fig. D.15 (b). Due to
the un-optimal filter specifications, there are only two prominent outer race harmonics
visible in the spectrum.
For the second dataset with test rig 1, the Kurtogram is shown in Fig. D.15 (c). Here,
the highest kurtosis value is identified at the same central frequency as the second mode
from the proposed method. The ES in Fig. D.15 (d) hence shows multiple harmonics of
the fundamental cage frequency.
Similarly, for dataset 3, the Kurtogram is shown in Fig. D.15 (e). Again, the optimal
frequency band with the highest kurtosis value is very close to the ones given by the
proposed methods. Therefore, the ES in Fig. D.15 (f) show multiple prominent harmonics
and side-bands.
For test rig 2, the Kurtogram is shown in Fig. D.15 (g). Here, a large band centered
at 6000 Hz with bandwidth of 12000 Hz is chosen for filtration. In comparison with the
red lines, this band encompasses several of the modes given by the proposed method.
Therefore, the resulting ES in Fig. D.15 (h) should contain multiple harmonics of the
characteristic fault. However, only two prominent harmonics were identified, which is less
than the spectra attained with the proposed method.
The final dataset with test rig 3 is also tested, and the Kurtogram is shown in Fig.
D.15 (i). An optimal band is identified at a central frequency around 15000 Hz, which
is different than the ones given by the proposed method. In this case, the Kurtogram
performs better than the proposed method, as the ES in Fig. D.15 (j) indicate multiple
 ×101 ×101
4
Amplitude

Amplitude
4 Harmonic

2 2

0 0
0 5 10 15 20 25 30 35 40 0 1 2 3 4 5 6
Orders Orders
×101 (a) ×101 (b)
Amplitude

Amplitude
2 Harmonic
Side-band 2
1

0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Orders Orders
(c) ×101 (d)
Amplitude

0
0 5 10 15 20 25 30 35 40
Orders
(e)

Figure D.16: Results from using the cepstrum pre-whitening method detailed in Section
D.4.2 on all the experimental datasets. Envelope spectra with identified harmonics and
side-bands are shown for each dataset. (a)–(c) test rig 1, dataset 1 through 3, respectively;
(d) test rig 2, dataset 1; (e) test rig 3, dataset 1.

Table D.2: Diagnosis score per dataset for each method. For the proposed method, the
best scoring ES is the basis. The best method for each dataset is written in bold.
Proposed Kurtogram CPW

Test rig 1 - outer race 1615 26 692

Test rig 1 - roller 1733 245 268
Test rig 1 - inner race 448 38 49
Test rig 2 - inner race 154 41 116
Test rig 3 - outer race 166 240 76

harmonics of the outer race fault.

D.6.2 Cepstrum pre-whitening

The cepstrum pre-whitening method elaborated on in Section D.4.2 is also used compare
performance with the proposed method. Fig. D.16 shows the resulting envelope spectra
from this method. The first ES in Fig. D.16 is from test rig 1, dataset 1 with an outer
race fault. With this method, multiple harmonics of the fault are identified, however
the harmonics are almost buried in the noise floor. Compared to band-pass filtration,
the cepstrum pre-whitening approach is bound to return an increased noise floor, as the
white noise covering the entire spectrum is brought to the same amplitude level as other
signals due to spectrum normalization.
The second ES in Fig. D.16 (b) is acquired from the second dataset using test rig
1. Here, multiple harmonics of the cage frequency is detected. Similarly, the ES from
dataset 3 in Fig. D.16 (c) also presents multiple prominent harmonics, but not as many
when compared to the proposed method.
For test rig 2, the ES in Fig. D.16 (d) shows up to four prominent harmonics, which
is about the same as the proposed method. Finally, the ES from test rig 3 is shown in
Fig. D.16 (e), and there up to four harmonics are displayed.

D.6.3 Diagnosis score

To quantify the performance of each method, the diagnosis score for all experimental
datasets are shown in Table D.2 for the three methods. For the proposed method, the
frequency mode resulting in the highest diagnosis score is used. The method with the
highest diagnosis score is highlighted with a bold font. As seen, the proposed method
scores the highest on the first four experimental dataset, while the Kurtogram method
is better for the final dataset. This shows that the proposed method shows an excellent
ability to discern resonance frequency vibration from the vibration signal.

D.7 Conclusions
In this paper, a new method for identifying bearing resonance frequency modes in a vibra-
tion signal is proposed. The method exploits the resonance vibration distortion caused by
order tracking and the spectrum normalization feature of cepstrum pre-whitening (CPW).
The algorithm first transforms the signal to the order domain before applying CPW and
returning to the time-domain via inverse order tracking. The inverse order tracking causes
a de-spread in resonance frequency modes, and since the spectrum is already normalized
by means of CPW, resonance mode amplitudes rise above the normalized noise floor. Fre-
quency areas containing prominent peaks in this spectrum are chosen as band-pass filter
regions, and the envelope order spectrum of each band-pass filtered signal is analyzed to
identify the bearing fault. Simulations and experimental results from three test rigs oper-
ating under variable speed conditions validate the proposed method. Comparisons with
the fast Kurtogram and CPW with the same datasets show that the proposed method per-
forms better or similarly well on all cases. In the experimental cases, multiple resonance
frequency modes were identified using the proposed method, and most of the obtained
envelope spectra contained multiple harmonics and side-band related to the fault. The
proposed method may also be extended to facilitate automatic resonance mode detection
and fault diagnosis.
REFERENCES

[1] W. Zhou, T. G. Habetler, and R. G. Harley. Bearing condition monitoring methods

for electric machines: A general review. In Proc. IEEE Int. SDEMPED, pages 3–6.
IEEE, 2007. doi:10.1109/demped.2007.4393062.

[2] Report of large motor reliability survey of industrial and commercial installations,
Part I. IEEE Transactions on Industry Applications, 21(4):853–864, 1985. doi:
10.1109/tia.1985.349532.

[3] N. Li, Y. Lei, J. Lin, and S. X. Ding. An improved exponential model for predicting
remaining useful life of rolling element bearings. IEEE Transactions on Industrial
Electronics, 62(12):7762–7773, 2015. doi:10.1109/TIE.2015.2455055.

[4] F. Immovilli, A. Bellini, R. Rubini, and C. Tassoni. Diagnosis of bearing faults in

induction machines by vibration or current signals: A critical comparison. IEEE
Transactions on Industry Applications, 46(4):1350–1359, 2010. doi:10.1109/TIA.
2010.2049623.

[5] F. Immovilli, M. Cocconcelli, A. Bellini, and R. Rubini. Detection of generalized-

roughness bearing fault by spectral-kurtosis energy of vibration or current signals.
IEEE Transactions on Industrial Electronics, 56(11):4710–4717, 2009. doi:10.1109/
TIE.2009.2025288.

[6] B. Yang, R. Liu, and X. Chen. Fault diagnosis for a wind turbine generator bearing
via sparse representation and shift-invariant K-SVD. IEEE Transactions on Indus-
trial Informatics, 13(3):1321–1331, 2017. doi:10.1109/TII.2017.2662215.

[7] X. Gong and W. Qiao. Bearing fault diagnosis for direct-drive wind turbines
via current-demodulated signals. IEEE Transactions on Industrial Electronics,
60(8):3419–3428, 2013. doi:10.1109/TIE.2013.2238871.

[8] S. Singh and N. Kumar. Detection of Bearing Faults in Mechanical Systems

Using Stator Current Monitoring. IEEE Transactions on Industrial Informatics,
13(3):1341–1349, 2017. doi:10.1109/TII.2016.2641470.

169
 [9] E. Elbouchikhi, V. Choqueuse, Y. Amirat, M. E. H. Benbouzid, and S. Turri.
An Efficient Hilbert–Huang Transform-Based Bearing Faults Detection in Induc-
tion Machines. IEEE Transactions on Energy Conversion, 32(2):401–413, 2017.
doi:10.1109/TEC.2017.2661541.

[10] L. Zhou, F. Duan, D. Mba, and E. Faris. A comparative study of helicopter

planetary bearing diagnosis with vibration and acoustic emission data. In IEEE
Int C Prognostics and Health Management (ICPHM), pages 246–251. IEEE, 2017.
doi:10.1109/ICPHM.2017.7998336.

[11] M. Kang, J. Kim, and J.-M. Kim. An FPGA-based multicore system for real-time
bearing fault diagnosis using ultrasampling rate AE signals. IEEE Transactions on
Industrial Electronics, 62(4):2319–2329, 2015. doi:10.1109/TIE.2014.2361317.

[12] D. Abboud, J. Antoni, S. Sieg-Zieba, and M. Eltabach. Envelope analysis of rotating

machine vibrations in variable speed conditions: a comprehensive treatment. Me-
chanical Systems and Signal Processing, 84:200–226, 2017. doi:10.1016/j.ymssp.
2016.06.033.

[13] R. B. Randall, J. Antoni, and S. Chobsaard. The relationship between spectral cor-
relation and envelope analysis in the diagnostics of bearing faults and other cyclosta-
tionary machine signals. Mechanical Systems and Signal Processing, 15(5):945–962,
2001. doi:10.1006/mssp.2001.1415.

[14] P. D. McFadden and J. D. Smith. Model for the vibration produced by a single point
defect in a rolling element bearing. Journal of Sound and Vibration, 96(1):69–82,
1984. doi:10.1016/0022-460X(84)90595-9.

[15] K. R. Fyfe and E. D. S. Munck. Analysis of computed order tracking. Mechanical

Systems and Signal Processing, 11(2):187–205, 1997. doi:10.1006/mssp.1996.0056.

[16] J. Shi, C. Shen, X. Jiang, W. Huang, and Z. Zhu. An Auto Instantaneous Frequency
Order Extraction Method for Bearing Fault Diagnosis under Time-Varying Speed
Operation. In Int. Conf. on Sensing, Diagnostics, Prognostics, and Control (SDPC),
pages 623–627, 2017. doi:10.1109/SDPC.2017.122.

[17] J. Wang, Y. Peng, and W. Qiao. Current-aided order tracking of vibration signals for
bearing fault diagnosis of direct-drive wind turbines. IEEE Transactions on Industrial
Electronics, 63(10):6336–6346, 2016. doi:10.1109/TIE.2016.2571258.
[18] R. B. Randall and J. Antoni. Rolling element bearing diagnostics—a tutorial.
Mechanical Systems and Signal Processing, 25(2):485–520, 2011. doi:10.1016/j.
ymssp.2010.07.017.

[19] M. S. Darlow, R. H. Badgley, and G. W. Hogg. Application of High-Frequency

Resonance Techniques for Bearing Diagnostics in Helicopter Gearboxes. Technical
report, US Army Air Mobility Research and Development Laboratory, 1974.

[20] J. Antoni. The spectral kurtosis: a useful tool for characterising non-stationary
signals. Mechanical Systems and Signal Processing, 20(2):282–307, 2006. doi:10.
1016/j.ymssp.2004.09.001.

[21] J. Antoni. Fast computation of the kurtogram for the detection of transient faults.
Mechanical Systems and Signal Processing, 21(1):108–124, 2007. doi:10.1016/j.
ymssp.2005.12.002.

[22] T. Barszcz and A. JabLoński. A novel method for the optimal band selection for
vibration signal demodulation and comparison with the Kurtogram. Mechanical
Systems and Signal Processing, 25(1):431–451, 2011. doi:10.1016/j.ymssp.2010.
05.018.

[23] R.-B. Sun, Z.-B. Yang, X.-F. Chen, and J.-W. Xiang. Sparse representation based
on spectral kurtosis for incipient bearing fault diagnosis. In Prognostics and System
Health Management Conference (PHM-Harbin), pages 1–6. IEEE, 2017. doi:10.
1109/PHM.2017.8079185.

[24] V. C. M. N. Leite, J. G. B. da Silva, G. F. C. Veloso, L. E. B. da Silva, G. Lambert-

Torres, E. L. Bonaldi, and L. E. d. L. de Oliveira. Detection of localized bearing
faults in induction machines by spectral kurtosis and envelope analysis of stator
current. IEEE Transactions on Industrial Electronics, 62(3):1855–1865, 2015. doi:
10.1109/TIE.2014.2345330.

[25] P. Borghesani, P. Pennacchi, R. B. Randall, N. Sawalhi, and R. Ricci. Application

of cepstrum pre-whitening for the diagnosis of bearing faults under variable speed
conditions. Mechanical Systems and Signal Processing, 36(2):370–384, 2013. doi:
10.1016/j.ymssp.2012.11.001.

[26] N. Sawalhi and R. B. Randall. Signal pre-whitening using cepstrum editing (liftering)
to enhance fault detection in rolling element bearings. In COMADEM, Stavanger,
Norway, pages 330–336, May 2011.
[27] R. B. Randall, N. Sawalhi, and M. Coats. A comparison of methods for separation
of deterministic and random signals. International Journal of Condition Monitoring,
1(1):11–19, 2011. doi:10.1784/204764211798089048.

[28] A. Klausen, K. G. Robbersmyr, and H. R. Karimi. Autonomous Bearing Fault

Diagnosis Method based on Envelope Spectrum. IFAC-PapersOnLine, 50(1):13378–
13383, 2017. doi:10.1016/j.ifacol.2017.08.2262.

[29] C. Mishra, A. K. Samantaray, and G. Chakraborty. Rolling element bearing defect

diagnosis under variable speed operation through angle synchronous averaging of
wavelet de-noised estimate. Mechanical Systems and Signal Processing, 72:206–222,
2016. doi:10.1016/j.ymssp.2015.10.019.

[30] Case Western Reserve University Bearing Data Center Website. 2015. URL: http:
//csegroups.case.edu/bearingdatacenter/home.

[31] A. Klausen, R. W. Folgerø, K. G. Robbersmyr, and H. R. Karimi. Accelerated

Bearing Life-time Test Rig Development for Low Speed Data Acquisition. Modeling,
Identification and Control, 38(3):143–156, 2017. doi:10.4173/mic.2017.3.4.

[32] W. A. Smith and R. B. Randall. Rolling element bearing diagnostics using the Case
Western Reserve University data: A benchmark study. Mechanical Systems and
Signal Processing, 64-65:100–131, 2015. doi:10.1016/j.ymssp.2015.04.021.
Paper E

Novel Threshold Calculations for

Remaining Useful Lifetime
Estimation of Rolling Element
Bearings

Andreas Klausen, Huynh Van Khang and Kjell G. Robbersmyr

173
This paper has been published as:

A. Klausen, H. V. Khang and K. G. Robbersmyr. Novel Threshold Calculations

for Remaining Useful Lifetime Estimation of Rolling Element Bearings. IEEE Inter-
national Conference on Electrical Machines (ICEM), Greece, 1912-1918, 2018. doi:
10.1109/ICELMACH.2018.8507056
Novel Threshold Calculations for Remaining
Useful Lifetime Estimation of Rolling
Element Bearings
Andreas Klausen, Huynh Van Khang and Kjell G. Robbersmyr

University of Agder
Department of Engineering Sciences
Jon Lilletunsvei 9, 4879 Grimstad, Norway

Abstract – The prognostics objective is to avoid sudden machinery break-

downs and to estimate the remaining useful life after initial degradation. Typ-
ically, physical health indicators are derived from available sensor data, and
a mathematical model is tuned to fit them. The time it takes for the model
to reach a failure threshold is the estimated remaining useful life. The fail-
ure threshold may be determined from historical failure data, but that is not
always readily available. ISO standard 10816-3 defines permissible velocity vi-
bration levels for machines that may be used as a failure threshold. However,
velocity vibration is not suitable for bearing prognostics due to the effect of in-
tegration from acceleration. In this paper, the drawbacks of velocity vibration
are explained, and two new failure thresholds using acceleration vibration data
are proposed. Results from three run-to-failure tests are provided to show the
performance of the proposed failure thresholds.

E.1 Introduction
Rolling element bearings are used in most rotating machinery. Their purpose is to transfer
the shaft load to a stationary housing, and make sure that the shaft rotates smoothly.
However, bearing damage is the most common cause of breakdown in rotating machinery,
and an unexpected breakdown may result in costly downtime and/or personal harm[1, 2].
It is therefore wise to monitor the machine health using sensor data from i.e. a vibra-
tion accelerometer and condition monitoring (CM) techniques. Using signal processing
methods, incipient faults in a bearing can be detected before complete breakdown[3].
However, it is not always feasible to perform maintenance as soon as a fault is detected.
For remote locations, such as offshore wind farms, weather conditions and pre-planned
maintenance trips contribute significantly to feasibility of performing a machine overhaul.
It is common that a faulted bearing can be in operation for a certain amount of time after
the initial fault. This amount of time is referred to as the remaining useful life (RUL).

175
For maintenance planning, it is beneficial to estimate the RUL based on the sensor data
acquired by the CM system, i.e. make a prediction on how long the machine may operate
before an emergency stop is necessary. The general approach is to acquire one or sev-
eral physical health indicators (PHIs) from available sensor data. A virtual health index
(VHI) can be generated from fusion or dimension reduction methods of the chosen PHIs.
Most approaches reported in literature use a certain data-driven model[4] to estimate the
RUL, where the future trend of a PHI or VHI is estimated using a certain model, and
the time/cycles it takes for this model to reach a failure threshold (FT), is the RUL.
Some examples of mathematical models are the exponential model, linear model, and the
Paris-Erdogan crack-law model. Often, bearing degradation is unstable, and therefore an
exponential model may fit best. One common PHI for rotating machinery is the root-
mean-square (RMS) of the vibration signal, as it is closely related to the total vibration
energy. RMS is also robust against random impacts, as short-lived events contribute lit-
tle to the mean of squared vibration. References[5, 6, 7] used the RMS for predicting
the RUL with different mathematical models, such as the improved exponential model,
a stochastic process model, and a Skew-Wiener process. In[8], 14 PHIs are fused into a
VHI using the logarithmic Mahalanobis distance (MD), and the RUL is estimated using a
Brownian motion (BM) statistical model. While the methods presented in the four papers
show excellent performance to predict the RUL, one important issue is not thoroughly
discussed: The FT setting. It appears that the FT used to predict the RUL is just set
to whatever the PHI/VHI is at the end of the vibration dataset that is utilized, i.e. the
FT is reverse-engineered to prove the performance of the model and parameter tuning.
In a real CM case where a new machine with no historical failure data is monitored, it is
impossible to use the methods described in[5, 6, 7, 8] without setting the FT in advance.
One attempt of defining the FT is given in[9], where the FT is said to be set dynamically
based on each machine. However, the procedure requires historical failure data sets that
must be used for training, which is not always readily available for every machine. There-
fore, there is a need for a general procedure to set the FT without relying on historical
failure data.

To deal with this unknown FT, this paper proposes methods for defining new FTs for
the RMS and the logarithmic MD. These FTs are based on the ISO 10816-3 standard[10]
that proposes viable vibration levels for machines operating at speeds faster than 120
rpm. The proposed FTs can be used on any machine given the availability of some
vibration data captured during a known healthy state. The rest of the paper is organized
as follows. In Section E.2 the ISO 10816-3 standard is explained, and a reason for why
the velocity RMS (VRMS) is not optimal to be used directly for RUL estimation is given.
Next, Section E.3 provides a method for transforming the VRMS threshold to acceleration
unit. Section E.4 gives an example of this transformation on an experimental dataset. In
Section E.5, a method for determining a threshold for the logarithmic MS is presented.
Section E.6 provides experimental results for three different datasets, and shows how well
the proposed FTs compare to the actual RUL. Finally, Section E.7 concludes the paper.

E.2 ISO standard threshold

ISO standard 10816-3[10] lists acceptable VRMS and displacement RMS (DRMS) levels
for rotating machines within the power range between 15 kW and 50 MW. The standard
divides vibration levels in four groups: A, B, C and D, where vibration levels in group A
belong to new machines, and group D contain vibration levels that cause damage to the
machine itself and/or connected components. The RMS value of a signal x is calculated
using
s
RT
0
x(t)2 dt
RMS{x} = , (E.1)
T
where T is the period of the signal.
For machinery prognostics purposes, it is of interest to estimate the RUL after the
initial damage, and vibration group D in the ISO standard may be used as an FT for the
end of serviceable life. However, the vibration level thresholds in the ISO standard are
only given for VRMS and DRMS signals for vibration signal components between 10 and
1000 Hz. Common accelerometers typically measure the vibration in acceleration units,
and therefore the signal must be integrated to achieve the vibration in a velocity unit.
The main issue with using the VRMS FT is the effect of integration itself, as damaged
bearings tend to give increased vibration levels at high frequencies, typically higher than
1000 Hz. Integration causes these high-frequency components to be mitigated as it acts
like a low-pass filter. Low-frequency vibration components only rise in value once the
bearing is critically damaged, and therefore changes in the bearing health may not be
detected early if the vibration is monitored in velocity units. An example of this can
be seen in Fig. E.1 where the VRMS and ARMS are compared in a real case. The
vibration measurements used in this example are collected by the NSF I/UCR Center for
Intelligent Maintenance Systems (IMS)[11]. The second run-to-failure test consisting of
982 data files of vibration measurements is utilized. Fig. E.1 (a) shows the VRMS, the
initial mean value µVRMS , and the initial standard deviation (STD) σVRMS . Initial mean
and STD are calculated based on mt = 300 (50 hours) data files in the healthy state as
µVRMS = 0.147, σVRMS = 0.720, µARMS = 0.014, and σARMS = 0.756. For the purpose of
detecting initial degradation, a statistical alarm is set at a window of 5 times STD, i.e.
5σ. As can be seen, the VRMS value does not surpass this alarm limit before reaching the
 5
VRMS µV RM S

4
RMS [mm/s]

3
σV RM S = 0.147
5σ crossed
µV RM S = 0.720
2

0
0 20 40 60 80 100 120 140 160
Time [hours]
(a)
8
ARMS µARM S

6
RMS [m/s2 ]

4
σARM S = 0.014
µARM S = 0.756 5σ crossed
2

0
0 20 40 60 80 100 120 140 160
Time [hours]
(b)

Figure E.1: Difference between RMS values calculated using vibration in velocity and
acceleration units. (a) The RMS of the velocity signal. (b) The RMS of the acceleration
signal. Both plots show a red-stapled line which is the mean RMS value of the first 300
data files.

last few hours of RUL, and, at this stage, the value increases quickly. From a prognostics
point of view, it may be too late to schedule maintenance as there is limited time left
before a complete failure. Fig. E.1 (b) shows the ARMS trend. After approximately 90
hours, the ARMS value surpasses the alarm limit, which signifies a trend of degradation
in the bearing. The alarm is triggered much sooner when using the ARMS compared to
when using the VRMS, and permits a much larger window for scheduling maintenance.
Therefore, it is advantageous to transform the VRMS ISO threshold value to acceleration
unit to estimate the RUL. The proposed threshold transformation is given in the next
subsection.
E.3 Proposed threshold transformation
Using a vibration accelerometer and an analog/digital converter, the signal is acquired in
m/s2 at a sample rate Fs , for a period T , and saved as xraw . The raw vibration signal
is first filtered using a 2nd order IIR high-pass filter with a cutoff-frequency at 10 Hz to
remove signal drift and comply with the ISO 10816-3 standard[10]. The transfer function
of this filter is
Xacc (s) s2
Gacc (s) = = 2 , (E.2)
Xraw (s) s + as + b
where Xraw and Xacc are the frequency transformed raw vibration and filtered vibration,
respectively, s = iω is the complex variable, ω is the frequency in rad/s, and a and b are
the filter constants. The integral of the vibration acceleration is the vibration velocity
xvel , and the transfer function is given by
Xvel (s) 1 s
Gvel (ω) = = Gacc (s) = 2 , (E.3)
Xraw (s) s s + as + b
where Xvel is the frequency transformed velocity vibration. From ISO 10816-3, the VRMS
values are only valid for vibration components up to 1000 Hz, and therefore an extra
low-pass filter should be applied at 1000 Hz. However, the authors argue that it is
unnecessary to apply a low-pass filter at 1000 Hz because 1/s reduces the signal power
by approximately 38 dB at 1000 Hz, and therefore an extra low-pass filter should be
unnecessary. The RMS is the square root of the mean energy of a signal which can be
calculated in the time-domain or in the frequency domain using

ZT F
Zs /2
2
E{x} = |x(t)| dt = |X(s = iω)|2 dω, (E.4)
0 −Fs /2

where T is the time signal period, and Fs is the sampling frequency in rad/s. Using the
energy definition, the RMS of a signal x is calculated using
r
E{x}
RMS{x} = . (E.5)
T
To transform the VRMS threshold to an acceleration unit, the RMS ratio (RMSr ) between
the ARMS and the VRMS is calculated using
s
RMS{xacc } E{xacc }
RMSr = = . (E.6)
RMS{xvel } E{xvel }

To calculate this ratio analytically, the energies of the two signals are needed. For conve-
nience, the energies are resolved in the frequency domain. One assumption for calculating
the energy is that xraw (t) is real-valued only, and therefore the negative frequency com-
ponents are the complex conjugates of the positive frequency components. The energy
can therefore be calculated by integrating over the positive frequencies and multiply by
2. The energy of the vibration acceleration signal is calculated using
Zω2
E{xacc } = 2 |Xacc (iω)|2 dω (E.7)
ω1
Zω2 2
−ω 2 Xraw (iω)
=2 dω (E.8)
−ω 2 + aiω + b
ω1

where the integration bounds ω1 and ω2 are given by

ω1 = 0 rad/s, ω2 = Fs /2. (E.9)

Similarly, the energy of the velocity vibration is calculated using

Zω2
E{xvel } = 2 |Xvel (iω)|2 dω (E.10)
ω1
Zω2 2
iωXraw (iω)
=2 dω. (E.11)
−ω 2 + aiω + b
ω1

Ideally, the RMSr would be calculated analytically for all cases using the acceleration
and velocity energy given in (E.7) and (E.10), respectively. However, the raw vibration
signal, Xraw (s), is an unavoidable term in the integration of both energy quantities. Xraw
is the true vibration signal passing through the dynamic system of the machine itself, the
dynamics of the accelerometer, and the characteristics of the analog/digital converter.
Therefore, Xraw can be significantly different between two machines, even during similar
shaft speeds and power outputs. It is not feasible to accurately model Xraw for each
case, hence RMSr should be determined experimentally on every machine and after every
maintenance overhaul. This is done by calculating the mean VRMS and ARMS values
for a machine during a known healthy state, and get the ratio using (E.6). Once the ratio
is determined, the acceleration threshold is finally calculated using

ARMSt = RMSr · VRMSt , (E.12)

where ARMSt is the ARMS threshold and VRMSt is the VRMS threshold. In the next
section, experimental vibration signal is used to determine the ARMSt using the methods
described in this section.

E.4 Practical test of threshold setting

To test the proposed ARMSt , the same IMS dataset presented in Section E.2 is used.
ISO standard 10816-3 suggests a VRMSt = 4.5 mm/s for machines larger than 15 kW.
 5
VRMS VRMSt = 4.0 mm/s
4
RMS [mm/s]

3
σV RM S = 0.147 5σ crossed
µV RM S = 0.720 0.7 hours left
2

0
0 20 40 60 80 100 120 140 160
Time [hours]
(a)
8
ARMS ARMSt = 4.34 m/s2

6
RMS [m/s2 ]

4
σARM S = 0.014
µARM S = 0.756 5σ crossed
2 73.5 hours left

0
0 20 40 60 80 100 120 140 160
Time [hours]
(b)

Figure E.2: Comparison of the RMS threshold using velocity and acceleration units for
dataset 1. (a) The RMS of the velocity signal. (b) The RMS of the acceleration signal.

Since the machine used in[11] is probably not powerful, the VRMSt is lowered to 4.0
mm/s. Using the mean and STD values presented in Section E.2, RMSr is determined
as 1085.22 1/s using (E.6). The new threshold is computed by (E.12) as ARMSt =
1085.22 1/s · 4 mm/s = 4.34 m/s2 . Fig. E.2 shows a comparison of the thresholds where
(a) shows the VRMS, and (b) shows the ARMS. As seen on the two graphs, the VRMS and
the ARMS reaches their respective thresholds at approximately the same time (163 hours).
This result show that the proposed transformation can be used to get a proper threshold
for the ARMS with a physical meaning. The alarm-triggered point (5σ) shows how long
time is left before the threshold is reached. As can be seen, the maintenance window is
much larger using the ARMS when compared to the VRMS: 73.5 hours compared to 0.7
hours.
The presented ARMS threshold may be used directly for RUL estimation using algo-
rithms in[5, 6, 7]. However, some prognostic algorithms combine multiple PHIs to create
a VHI for the system. The advantage is that each PHI responds differently to different
degradation trends. Therefore, by fusing multiple PHIs, the initial degradation trend may
be easier to detect. In the next section, multiple PHIs are fused using the Mahalanobis
distance (MD), and a proposed threshold equation for the MD is outlined.

E.5 Threshold setting for multiple physical health in-

dicators
Some prognostics models use multiple PHIs including the ARMS to determine the RUL.
Such a prognostics model using multiple PHIs is proposed in[8], where 14 time-domain
PHIs are calculated from the vibration signal. The PHIs are fused using the MD to
create a single VHI that represent the data file. According to[8], the MD is linearized by
calculating the logarithm of the MD. For brevity, a health index (HI) is defined here as
the logarithmic MD. Parameters for a Brownian motion (BM) model are estimated in an
on-line fashion using a Kalman filter (KF) to best fit the HI. Since the HI is assumed to
be linearly increasing with time once the initial degradation of the bearing has started,
this is a valid model to use. The RUL is calculated by using the estimated BM model
and predict how long time it would take for the model to reach a certain failure threshold
(FT) value. However, in[8], the FT used for estimating the RUL is not discussed. It is
impossible to calculate the RUL in a real situation if the FT is not known. Therefore,
the rest of this section is dedicated to a method for making a suitable threshold for the
HI proposed in[8].
First, the MD is calculated using
p
MD = (x − µ)T S −1 (x − µ), (E.13)

where x = [x1 , x2 , . . . , xn ] is a vector of the n chosen PHIs, S is the covariance matrix of

a set of observations, and µ = [µ1 , µ2 , . . . , µn ] is the mean value of these observations. To
determine S and µ, a set of mt observations during healthy machine condition are used.
Finally, the HI is determined using

HI = log{MD}. (E.14)

The FT for the HI is determined using the ARMS and a single-dimension MD (MD1 )
x−µ
MD1 = , (E.15)
σ
where σ is the standard deviation of the first mt observations of x, and µ is the mean value
of these observations. Each PHI result in different MD1 trends, however the assumption
for the proposed threshold is that the average MD1 characteristics for all vibration based
PHIs are similar to the MD1 of the ARMS. Specifically, that the average maximum value
for all MD1 ’s is the same as the maximum value of the MD1 of the ARMS. Following this
assumption, it should be possible to upscale the ARMSt to account for multiple PHIs in
the MD. To this end, the MD can be estimated by using the ARMS alone, as in
v
u n
uX (ARMS − µARMS )2
MD ≈ t 2
, (E.16)
i=1
σARM S

√ ARMS − µARMS
≈ n· , (E.17)
σARMS
2
where S is replaced with σARM S because, as a covariance matrix, it is diagonal for PHIs
that are equal. Given this estimated MD, it is possible to create a threshold for the MD
and the HI using the ARMSt as
√ ARMSt − µARMS
MDt = n· + µMD , and (E.18)
σARMS
HIt = log{MDt } + µHI , (E.19)

where MDt and HIt are the proposed MD and HI thresholds, respectively, and µMD and
µHI are the mean value of the MD and HI for the first mt samples, respectively. The mean
value is added to the threshold to account for the initial value of the MD and HI. Note
that (E.13) and (E.14) are used for computing the MD and HI, respectively, while (E.18)
and (E.19) are only used to determine their respective thresholds. In the next section,
the proposed HIt is calculated for multiple test cases.

E.6 Practical test of MD threshold

E.6.1 IMS dataset

The proposed HIt threshold is here calculated for the IMS bearing dataset introduced
in Section E.4. From Section E.4, the ARMSt = 4.34 m/s2 , µARMS = 0.756 m/s2 , and
σARMS = 0.0141 m/s2 . A total of n = 14 PHIs from[8] are utilized to make the HI using
(E.14) where the covariance and mean values are calculated from the mt = 300 (50 hours)
first data files. The mean value of HI is µHI = 0.905 for this dataset. Given these values,
the threshold is calculated using (E.19) as
√ 4.34 − 0.756
HIt = log{ 14 · } + 0.905 = 7.7655.
0.0141
Table E.1 shows these calculated values under column “Dataset 1”. The HI and the HIt
are both shown in Fig. E.3 for the IMS dataset, and as shown, the HIt is reached by the HI
close to the final data file. This finding is consistent with the ARMS and ARMSt shown
 Table E.1: Threshold value calculations.
Value ↓ \ Dataset → 1 2 3

VRMSt [m/s] 4.0 3.0 4.0

mt [files] 300 300 125
RMSr [1/s] 1085 2803 368
ARMSt [m/s2 ] 4.34 8.41 1.47
2
σARMS [m/s ] 0.0141 0.0235 0.0049
µARMS [m/s2 ] 0.756 0.378 0.447
µHI [-] 0.905 1.053 2.404
HIt [-] 7.7655 8.208 9.068

in Fig. E.2 (b) in that the computed threshold is reached near the end. The possible
advantage of the HI compared to the ARMS in terms of prognostics, is that the general
trend of the HI appears to increase linearly over time from first degradation point, while
the ARMS exhibits a more exponentially increasing trend near the end. The linear HI
trend has an advantage of allowing simpler models for estimating the RUL, such as the
BM used in[8]. The choice of RUL estimation model depends on which trend should be
used, as the ARMS and HI shows different characteristics.

E.6.2 PHM Challenge dataset

The second dataset used for testing the ARMSt and HIt is a part of the Pronostia
database[12]. This database was part of the IEEE PHM 2012 Prognostics Challenge that
encouraged scientists and engineers to make accurate RUL estimations based on vibration
and temperature data. The database consists of 17 accelerated life-time tests conducted
during three different operating conditions. The ARMSt and HIt are determined using
the vibration data in dataset “Bearing1 3”, i.e. the third accelerated life-time test during
operating condition 1. A total of 2375 vibration data files was collected during this test,
and the time interval between two files is 10 seconds. The vibration data in each file was
collected for a duration of 0.1 s at a rate of Fs = 25.6 kHz.
The VRMSt = 3m/s due to the small 250 W motor used on the Pronostia test rig.
The first mt = 300 data files (50 minutes) are assumed to be from a healthy machine state.
Important values and calculated thresholds are given in Table E.1 under column “Dataset
2”. The VRMS and ARMS are shown in Figs. E.4 (a) and (b), respectively, together with
their respective thresholds. The VRMS reaches and surpasses the proposed threshold of
3 mm/s at the end of the accelerated life-time test, indicating that the bearing is severely
damaged. The ARMS in Fig. E.4 (b) shows very similar characteristics as the VRMS,
 10 HI HIt = 7.77

6
HI

0
0 20 40 60 80 100 120 140 160
Time [hours]

Figure E.3: The HI plotted with the HIt threshold for dataset 1.

and the alarm at 5σ is triggered almost at the same time for both signals. The main
difference is that the ARMS value is less noisy than the VRMS, which should make it
easier to predict RUL using the ARMS. In addition, the threshold ARMSt is reached at the
end of the test. The HI is also computed for this dataset using the methods elaborated
in Section E.5 and its threshold is given in Table E.1 under column “2”. The HI and
the HIt are both shown in Fig. E.5 for this test. As seen, there is a clear advantage of
calculating the logarithm of the MD, as the HI is increasing almost linearly after the initial
degradation in this test. The HIt threshold is also very close to the actual final value of
the HI at the end of the test, which shows the performance of the proposed threshold
calculations. A filtered HI (HIf ) is overlaid in Fig. E.5 to make the degradation trend
easier to identify. The HI and HIt can be used in conjunction with the BM algorithm
in[8] to estimate the bearing RUL.

E.6.3 Low-speed test rig

The final database used originate from an in-house low-speed bearing test rig. A detailed
description of the test rig and the procedures for acquiring vibration sensor data are given
in[13]. The dataset is acquired during an accelerated life-time test of a 6008 type bearing,
and the vibration signal is measured while the shaft speed is as low as 100 rpm. Each
vibration data file was recorded every 18 minutes at a sample rate of 51.2 kHz for a
duration of 10 seconds. The test in total lasted 14 full days, i.e. 14 · 24 = 336 hours,
before a torque limit alarm was triggered and forced a stop of the test. The last 507 data
files (152 hours) of the accelerated life-time test is used to make the proposed thresholds.
In ISO 10816-3, the VRMS threshold is set for components between 10 and 1000 Hz for
 8
VRMS VRMSt = 3.0 mm/s

6
RMS [mm/s]

σV RM S = 0.023
4 µV RM S = 0.138

2
5σ crossed
191.0 minutes left
0
0 50 100 150 200 250 300 350 400
Time [minutes]
(a)
12
ARMS ARMSt = 8.41 m/s2
10
RMS [m/s2 ]

6
σARM S = 0.023
4 µARM S = 0.378

2 5σ crossed
181.8 minutes left
0
0 50 100 150 200 250 300 350 400
Time [minutes]
(b)

Figure E.4: Comparison of the RMS threshold using velocity and acceleration units for
dataset 2. (a) The RMS of the velocity signal. (b) The RMS of the acceleration signal.

machines operating at speeds higher than 600 rpm. For slower machines, between 120
and 600 rpm, the frequency band is set from 2 to 1000 Hz. As the shaft speed in this test
is 100 rpm, the 2nd order IIR high-pass filter in (E.2) is configured to a cutoff frequency
at 80% · 100/60 = 1.33 Hz to preserve the vibration energy at the shaft frequency and
higher. The motor power is 1.1 kW, and the VRMSt is therefore set to 4 mm/s as a
conservative value. The mt = 125 (40 hours) first data files are considered to be recorded
during a healthy state, and are therefore used for calculating the thresholds. Table E.1
column “Dataset 3” shows calculated values and thresholds for this dataset. The VRMS,
ARMS, and their respective thresholds are shown in Fig. E.6. The VRMS exceeds its
threshold near the end of the test, and the 5σ alarm is triggered only 2.1 hours before
the threshold is reached. In a real application, this will translate to a very short time
for RUL estimation and maintenance planning. The ARMS also exceed its threshold
at the end, and the initial 5σ alarm is triggered much sooner, at 31.5 hours left before
reaching the threshold. It should be noted, however, that the ARMS shows an extremely
 10
HI HIt = 8.21 HIf

6
HI

0
0 50 100 150 200 250 300 350 400
Time [minutes]

Figure E.5: The HI plotted with the HIt threshold for dataset 2.

sudden change in degradation ratio near the end, and therefore it might be difficult to fit
a degradation model to this signal.
Using the 14 PHIs in[8] the HI for this dataset is calculated using (E.14). The HI
is shown in Fig. E.7. As shown, the HI does not indicate a linear trend, but rather
an exponentially increasing trend. However, the increase near the end (140 hours) is
less sudden when compared to the ARMS in Fig. E.6 (b). Therefore, it may be easier
to make a model that follows the trend of the HI compared to the ARMS. In addition,
the threshold value is reached near the end, which suggests the validity of the proposed
threshold.

E.7 Conclusions
Two failure thresholds (FTs) are proposed in this paper for remaining useful life-time
(RUL) estimation of rotating machinery. The FTs are based on the ISO 10816-3 stan-
dard for acceptable vibration levels, and historic failure data for each specific machine
is not required. The ISO standard provides acceptable levels for velocity RMS (VRMS)
vibration which are not suitable for RUL estimation. That is because the initial bear-
ing wear induce increased vibration levels at high frequencies, typically higher than 1000
Hz. Velocity is the integral of acceleration, and this integration causes high-frequency
vibration components to be attenuated. The vibration energy at low frequencies only
rises significantly once the bearing is critically damaged, and therefore it may be too late
to schedule maintenance. To improve the RUL estimation capability of the ISO VRMS
threshold, the first presented FT transforms it to acceleration unit. Experimental results
show that: 1) the transformed acceleration based RMS (ARMS) threshold is reached by
 8
VRMS VRMSt = 4.0 mm/s

6
RMS [mm/s]

4
σV RM S = 0.099 5σ crossed
µV RM S = 1.224 2.1 hours left
2

0
0 20 40 60 80 100 120 140 160
Time [hours]
(a)
2.00
ARMS ARMSt = 1.47 m/s2
1.75

1.50
RMS [m/s2 ]

1.25

1.00
σARM S = 0.005
0.75 µARM S = 0.447 5σ crossed
31.5 hours left
0.50

0.25
0 20 40 60 80 100 120 140 160
Time [hours]
(b)

Figure E.6: Comparison of the RMS threshold using velocity and acceleration units for
dataset 3. (a) The RMS of the velocity signal. (b) The RMS of the acceleration signal.

the ARMS at approximately the same time as the VRMS reaches its threshold, and 2)
that the ARMS show signs of early degradation earlier when compared to the VRMS. The
second FT presented is for a health index (HI) based on multiple vibration physical health
indicators (PHIs) fused into a single virtual health indicator (VHI) using the Mahalanobis
distance (MD). Experimental results show that the HI threshold is suitable, and that the
HI presents a different degradation trend than the ARMS. In practice, these FTs can be
employed for RUL estimation on any rotating machine in which a tolerable VRMS level is
given either by the ISO 10816-3 standard, or from experience by an operator. For future
work it may be beneficial to replace some PHIs in the MD, and also to scale up the ARMS
threshold for other VHIs than the MD.
 14
HI HIt = 9.07 HIf
12

10

8
HI

0
0 20 40 60 80 100 120 140 160
Time [hours]

Figure E.7: The HI plotted with the HIt threshold for dataset 3.
REFERENCES

[1] W. Zhou, T. G. Habetler, and R. G. Harley. Bearing condition monitoring methods

for electric machines: A general review. In Proc. IEEE Int. SDEMPED, pages 3–6.
IEEE, 2007. doi:10.1109/demped.2007.4393062.

[2] Report of large motor reliability survey of industrial and commercial installations,
Part I. IEEE Transactions on Industry Applications, 21(4):853–864, 1985. doi:
10.1109/tia.1985.349532.

[3] R. B. Randall and J. Antoni. Rolling element bearing diagnostics—a tutorial.

Mechanical Systems and Signal Processing, 25(2):485–520, 2011. doi:10.1016/j.
ymssp.2010.07.017.

[4] Y. Lei, N. Li, L. Guo, N. Li, T. Yan, and J. Lin. Machinery health prognostics: A
systematic review from data acquisition to RUL prediction. Mechanical Systems and
Signal Processing, 104:799–834, 2018. doi:10.1016/j.ymssp.2017.11.016.

[5] N. Li, Y. Lei, J. Lin, and S. X. Ding. An improved exponential model for predicting
remaining useful life of rolling element bearings. IEEE Transactions on Industrial
Electronics, 62(12):7762–7773, 2015. doi:10.1109/TIE.2015.2455055.

[6] Y. Lei, N. Li, and J. Lin. A new method based on stochastic process models for
machine remaining useful life prediction. IEEE Transactions on Instrumentation
and Measurement, 65(12):2671–2684, 2016. doi:10.1109/TIM.2016.2601004.

[7] Z. Huang, Z. Xu, X. Ke, W. Wang, and Y. Sun. Remaining useful life prediction for
an adaptive skew-Wiener process model. Mechanical Systems and Signal Processing,
87:294–306, 2017. doi:10.1016/j.ymssp.2016.10.027.

[8] Y. Wang, Y. Peng, Y. Zi, X. Jin, and K.-L. Tsui. A two-stage data-driven-based prog-
nostic approach for bearing degradation problem. IEEE Transactions on Industrial
Informatics, 12(3):924–932, 2016. doi:10.1109/TII.2016.2535368.

191
 [9] K. Javed, R. Gouriveau, and N. Zerhouni. A new multivariate approach for prognos-
tics based on extreme learning machine and fuzzy clustering. IEEE Transactions on
Cybernetics, 45(12):2626–2639, 2015. doi:10.1109/TCYB.2014.2378056.

[10] ISO. Mechanical vibration—Evaluation of machine vibration by measurements on

non-rotating parts—Part 3 (Standard No. 10816-3). ISO, 1998.

[11] H. Qiu, J. Lee, J. Lin, and G. Yu. Wavelet filter-based weak signature detection
method and its application on rolling element bearing prognostics. Journal of Sound
and Vibration, 289(4-5):1066–1090, 2006. doi:10.1016/j.jsv.2005.03.007.

[12] P. Nectoux, R. Gouriveau, K. Medjaher, E. Ramasso, B. Chebel-Morello, N. Zer-

houni, and C. Varnier. PRONOSTIA: An experimental platform for bearings acceler-
ated degradation tests. In IEEE International Conference on Prognostics and Health
Management, PHM’12., pages 1–8. IEEE Catalog Number: CPF12PHM-CDR, 2012.

[13] A. Klausen, R. W. Folgerø, K. G. Robbersmyr, and H. R. Karimi. Accelerated

Bearing Life-time Test Rig Development for Low Speed Data Acquisition. Modeling,
Identification and Control, 38(3):143–156, 2017. doi:10.4173/mic.2017.3.4.
Paper F

Novel RMS Based Health Indicators

used for Remaining Useful Lifetime
Estimation of Bearings

Andreas Klausen, Huynh Van Khang and Kjell G. Robbersmyr

193
This paper has been submitted as:

A. Klausen, H. V. Khang and K. G. Robbersmyr. Novel RMS Based Health Indi-

cator used for Remaining Useful Lifetime Estimation of Bearings. Under review at
Mechanical Systems and Signal Processing.
 Novel RMS Based Health Indicators used
for Remaining Useful Lifetime Estimation of
Bearings
Andreas Klausen, Huynh Van Khang and Kjell G. Robbersmyr

*University of Agder
Department of Engineering Sciences
Jon Lilletunsvei 9, 4879 Grimstad, Norway

Abstract – Estimating the remaining useful life (RUL) of bearings from

healthy to faulty is important for predictive maintenance. The bearing fault
severity can be estimated based on the energy or root mean square (RMS) of
vibration signals. However, the vibration RMS is often not monotonic, which
renders a challenge for predicting the RUL. To tackle the problem, a new
method is proposed for splitting the vibration signal into multiple frequency
bands for RMS calculations. Compared to a digital filter bank, the proposed
method is lossless, and the computational burden is much smaller. Some of
the resulting RMS trends are more relevant for RUL estimation due to their
monotonicity. Suitable RMS trends are identified using the Spearman coeffi-
cient, and the RUL is afterwards estimated with the Paris-Erdogan law and
particle filters. Historical failure data is not required to set any parameters,
which is a clear advantage of the proposed method. Experimental results from
two test rigs validate the performance of the proposed method.

F.1 Introduction
Bearings are common in rotating machines, and bearing defects result in increased vibra-
tion, temperature, and friction. Up to 44% failures in the most common motors, namely
induction motors, are due to bearing faults [1]. The vibration from severe bearing faults
may cause damage to other machine components, such as gears, stators and pump seals,
and should therefore be detected. Unscheduled stops can cause long downtime and huge
expenses due to maintenance and productivity losses. Therefore, monitoring the bearing
health condition is important to avoid emergency shutdowns and plan maintenance.
A condition monitoring (CM) system allows for detecting faults, and therefore en-
hances machine reliability. On rotating machines, bearing CM systems can be used to
detect faults using vibration signals, typically with the envelope spectrum [2]. Continued
operation of a machine after detecting the initial bearing fault is beneficial for planning

195
maintenance, or even necessary if the machine should be shut down in a regulated man-
ner rather than immediately. As such, an estimation of the bearing remaining useful life
(RUL) is important to select between regulated or emergency stop.
A bearing health indicator (HI) can be compared to a failure threshold (FT) to es-
timate the RUL by assuming that a machine should stop if the HI reaches the FT. The
future HI trend can be predicted using a mathematical model, and the time until FT is
reached is the estimated RUL. The bearing HI can be assessed by examining the level of
wear on the bearing rollers and raceways. However, the actual bearing HI is impractical
to determine, as it would require an offline inspection after dissembling the bearing [3],
resulting in productivity loss. Instead, the HI can be estimated using sensor signals that
are related to the amount of bearing wear. Vibration signals [4] can be used for this pur-
pose, as defects in the raceways or rollers increases the vibration energy. The vibration
signal is often reduced to features which can afterwards be used as the estimated HI.
HIs can be categorized into physical HIs (PHIs) and virtual HIs (VHIs) [5]. PHIs
are generated from primarily physical signals and are directly related to the physics of
failure. Examples are the root mean square (RMS) [6], kurtosis [7], and characteristic
bearing fault frequency amplitudes [8]. VHIs do not correlate directly with the physics of
failure [5] and can be calculated by combining multiple PHIs. The Mahalanobis distance
was applied in [9] to combine 14 PHIs into a single VHI, and principal component analysis
(PCA) was used in [10] to estimate the principal component of multiple PHIs.
The bearing degradation can be divided into two or multiple health stages (HSs) [3].
During the first HS, there is no apparent degradation, while the second HS often show
linear increase, and the third could be unstable growth. The RUL is normally estimated
after transitioning from the first to the second HS. This transition can be detected using
baseline measurements of the kurtosis value [6] and the RMS signal [11].
After transitioning to the second HS, the future degradation trend can be estimated
by a mathematical model which closely resembles the physics of failure. The exponential
model [6, 12], Brownian Motion [9], and Paris-Erdogan law [13] have been used to predict
the future bearing degradation trend. The Kalman filter [14] and particle filter [15] have
been used to update parameters for mathematical models and predict the future degra-
dation trend. Alternatively, data-driven methods allow for tracking the trend without
knowing the physics of failure. A Gaussian process model [16] and least squares support
vector machine (LSSVM) [10] are examples of data-driven models.
Setting a proper FT is the final prerequisite for predicting RUL using the aforemen-
tioned methods. For complex VHIs, historical failure data from a similar setup is often
necessary to create a suitable FT [17, 18, 13]. However, for machines with no histor-
ical failure data, many of the methods reported in the literature may not work. The
implemented solutions in the referred papers require use of historic failure data to set
parameters, rendering a challenge for estimating RUL on new machines. In [19], the FT
is set based on the vibration RMS as guided by the ISO standard 10816-3 [20], without
involving historic failure data. However, estimating the RUL using RMS is difficult as
vibration energy is often not monotonic.
To address the existing challenges, a new approach for extracting the mean degradation
trend of the vibration RMS signal is proposed. The vibration signal is split into multiple
frequency bands to separate monotonic components from non-monotonic ones. Instead of
using a digital filter bank, the new approach utilizes only the discrete Fourier transform of
the vibration signal. Compared to digital filters, the new approach yields no information
loss, and is less computationally expensive. The RMS is calculated for each frequency
band components, and the Spearman coefficient is used to determine monotonic RMS
trends suitable for RUL estimation. FTs of the generated RMS trends are determined by
extending the FT calculations in [19]. The degradation trend is predicted by updating
parameters of the Paris-Erdogan law using a particle filter (PF). Historic failure data is
not required, because the FTs are based on general guidelines provided by ISO 10816-
3 [20]. The performance of the proposed method is demonstrated on two experimental
datasets.
The rest of the paper is organized as follows. Practical differences between velocity-
and acceleration-based RMS are discussed in Section F.2. Next, the proposed approach
for subdividing the vibration signal into multiple RMS trends is detailed in Section F.3.
Afterwards, the algorithms used for RUL estimation are elaborated in Section F.4, and
the experimental results are given in Section F.5. Finally, suggestions for further work
are given in Section F.6, and conclusions are drawn in Section F.7.

F.2 Velocity vs acceleration RMS

ISO Standard 10816-3 [20] defines levels of velocity-based vibration RMS values for ro-
tating machines with over 15 kW power output. Four levels are given, ranging between
A - D: A=“standard acceptance for new machines”; B=“unlimited operation is possible”;
C=“short term operation allowed”; D=“vibration causes damage”. A machine should
stop before reaching level D, and therefore the boundary between level C and D can be
used as an FT. The RMS of a vibration signal x(t) can be calculated with
s
RT
0
x(t)2 dt
RMS(x(t)) = , (F.1)
T
where T is the measurement period. The vibration signal x(t) is in this study measured
with an accelerometer, which gives the vibration in acceleration units. Through integra-
tion, the velocity signal xv (t) is acquired, which is used for checking vibration severity
according to ISO 10816-3 [20]. For brevity, Ra = RMS(x(t)) and Rv = RMS(xv (t)). How-
ever, Rv is not suitable for RUL estimation, because early defects generate high frequency
resonance vibration, and integration reduces the effect of high frequency components as
verified in [19]. To alleviate this limitation, the FT is transformed to the acceleration do-
main before estimating bearing RUL. This procedure was proposed in [19] and is briefly
described in the following.
To experimentally transform the FT from velocity to acceleration units, the mean
Ra and Rv values at a known steady state of the machine (baseline) are calculated.
Afterwards, the ratio between these is determined with [19]

R̄a
Rr = , (F.2)
R̄v
where R̄a and R̄v are the mean Ra and Rv during baseline measurements, respectively.
Using this ratio, the acceleration-based FT R̂a is calculated using [19]

R̂a = Rr R̂v , (F.3)

where R̂v is the velocity-based FT from the ISO standard.

Vibration data collected by the NSF I/UCR Center for Intelligent Maintenance Sys-
tems (IMS) [21] is used to validate this transformation. Accelerometer data from the
second run-to-failure test, consisting of 982 vibration measurements, are utilized. Every
10 minutes, 1 second of vibration data was measured at a sample rate of 20480 Hz. At the
end of this test, an outer race fault was observed in the dismantled bearing. The motor
size is assumed smaller than 15 kW, and therefore the FT is set slightly lower than level
D in the ISO standard, at R̂v = 4 mm/s.
The first 30 hours of vibration data (180 datasets with 10 minutes interval) is set as
a baseline, and the mean RMS values are R̄a = 0.757 m/s2 and R̄v = 0.755 mm/s. Using
(F.2) and (F.3) results in R̂a = 4.01 m/s2 . Fig. F.1 (a) shows Rv and R̂v , while Fig. F.1
(b) shows Ra and the resulting FT R̂a . As observed in Fig. F.1 (a), Rv starts increasing
from the mean value near the end of useful life (after 150 hours), giving a very short
time to shut the machine down and plan maintenance. However, Ra starts increasing
much sooner (t ≈ 80 hours), which allows for predicting RUL and plan maintenance in
advance. The two trends reach their respective FTs near the end of life, which shows that
the transformation can be used for this purpose.
Ra in Fig. F.1 (b) is, however, not suitable for RUL estimation due to oscillations
around the mean degradation. Fig. F.2 shows Ra and an artificially created mean trend
to highlight the cyclic behavior. For example, the sudden increase of Ra at t ≈ 115 hours,
and the subsequent decrease, render a challenge for determining the actual degradation
 5 8
Ra
4 R̂a
6
RMS [mm/s]

RMS [m/s2 ]
3 Rv
4
R̂v
2
2
1

0 0
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160
Time [hours] Time [hours]
(a) (b)

Figure F.1: Comparison between velocity- and acceleration-based RMS for the IMS
dataset. (a) velocity-based RMS and its FT; (b) acceleration-based RMS and its FT.

3.0
Ra
2.5 Mean trend
RMS [m/s2 ]

2.0

1.5

1.0

0.5
80 90 100 110 120 130 140 150 160
Time [hours]

Figure F.2: Ra compared to an artificially created mean trend to highlight oscillations.

trend. A more stable trend should be extracted from the vibration signal to achieve better
RUL estimation. In the next subsection, a new approach for extracting more useful RMS
data from the vibration signal is proposed.

F.3 Proposed RMS health indicator

F.3.1 Filter bank RMS

A digital filter bank can be used to subdivide a signal into multiple components, where
each component contains a frequency sub-band of the signal. This can be achieved by
iteratively passing the signal through a low- and high-pass finite impulse response (FIR)
filter and decimating each output signal to half the frequency. Such a procedure has
some limitations: Digital FIR filters use convolution to filter the signal, and only the
overlapping part between the filter kernel and the signal should be preserved to avoid
adding artifacts. This means that a high order FIR filter kernel with a sharp frequency
response will remove much of the signal energy. In addition, the energy loss is exponential
for each level of filtering. If, however, the FIR filter kernel is small, there can be large
frequency overlap between the signal components. Computing a digital filter bank can also
be computationally taxing if the input signal is long, and if multiple frequency levels are
required. Alternatively, the vibration signal can be split into sub-bands using the discrete
wavelet transform (DWT) [22]. A time-domain filter bank is, however, not necessary to
acquire the RMS in different frequency bands. The following explains how the spectrum
bins from a single discrete Fourier transform (DFT) can be used to directly calculate the
vibration RMS in a certain frequency band.
An alternative representation of RMS is given by the energy E of the signal, such as
r
E(x)
RMS(x) = . (F.4)
T
The signal energy can be calculated in both time and frequency domain with
Xn Xn
2
E(x) = |xi | ∆t = |Xi |2 ∆f , (F.5)
i=1 i=1

where n is the length of the sampled signal, x = (x1 , . . . , xn ) is the vibration signal given
in discrete time samples, X = (X1 , . . . , Xn ) is the frequency spectrum, ∆t is the time
interval between samples, and ∆f is the frequency step between each spectrum bin.
To explain the procedure, the frequency spectrum is assumed ordered from 0 Hz to
Nyquist frequency, and from negative Nyquist up to 0 Hz. Let XL = (XLP , Zn/2 , XLN ),
where XLP contains the spectrum bins of the lower positive frequencies (i.e. 0 Hz to half
the Nyquist frequency), Zn/2 is n/2 zeros, and XLN contains the spectrum bins of the
lower negative frequencies. Similarly, let XH = (Zn/4 , XHP , XHN , Zn/4 ), where XHP and
XHN are the spectrum bins of higher positive and negative frequencies, respectively. This
arrangement represents a single level filter bank that splits X at half Nyquist frequency
completely. The total energy of the signal can afterwards be calculated as

E(X) = E(XL + XH )
Xn
= |XL,i + XH,i |2 ∆f (F.6)
i=1

where XL,i and XH,i are the i’th bin of XL and XH , respectively. Let the spectrum bins
be given by their complex values, i.e. XL,i = aL,i + jbL,i , XH,i = aH,i + jbH,i , then (F.6)
becomes

n q
X 2
= 2
(aL,i + aH,i ) + (bL,i + bH,i )2 ∆f
i=1
Xn
a2L,i + 2aL,i aH,i + a2H,i + b2L,i + 2bL,i bH,i + b2H,i ∆f .


= (F.7)
i=1
 Given that XL and XH represent a complete signal separation at half Nyquist fre-
quency, the overlap between XL and XH is zero. Therefore, 2aL,i aH,i = 2bL,i bH,i = 0, and
(F.7) is reduced to
n
X
E(X) = (a2L,i + a2H,i + b2L,i + b2H,i )∆f
i=1
n
X n
X
= (|aL,i + jbL,i |2 )∆f + (|aH,i + jbH,i |2 )∆f
i=1 i=1

= E(XL ) + E(XH ) . (F.8)

Eq. (F.8) shows that the energy of the entire signal can be calculated by the energy of
separate spectrum bins. It can also be shown as an extension of (F.4) that
r
E(XL ) + E(XH )
RMS(X) =
T
p
= RMS(XL )2 + RMS(XH )2 . (F.9)

Eq. (F.8) also indicates that the signal can be split into multiple equally-sized frequency
bands such that
Nb
X
E(X) = E(X(i−1)nb +1:inb ) (F.10)
i=1
v
u Nb
uX
RMS(X) = t RMS(X 2
(i−1)nb +1:inb ) , (F.11)
i=1

where Nb is the number of frequency bands the spectrum is split into, and nb = n/Nb .
For brevity, the RMS of a frequency band i is defined as

Ri = RMS(X(i−1)nb +1:inb ) , (F.12)

and the RMS value for band i at time index k is given by Ri,k . In summary, the spectrum
X is first obtained for the entire vibration signal x, and afterwards the energy of bins
belonging to frequency band i is used to calculate Ri using (F.4). The next step is to
obtain FTs for all Ri .

F.3.2 Filterbank RMS failure thresholds

To get an FT for each Ri , (F.11) is reconsidered such that
v
u Nb
uX
R̂a = t R̂i2 , (F.13)
i=1
where R̂i is the FT for Ri . To solve (F.13) for any number of bands Nb , it is necessary
to make some assumptions about the signal x. It is assumed that the vibration signal
is corrupted with normally distributed noise. Such noise is evenly distributed in the
frequency domain, and therefore the energy in all frequency bands should be affected by
this noise. It is also assumed that the energy increase from bearing wear is proportional
to the baseline noise variance in each frequency band. With these assumptions, the FTs
are defined as
R̂i = µi + mσi , (F.14)

where µi and σi are the mean and standard deviation (STD) of Ri during baseline measure-
ments, respectively, and m is a constant scaling factor. To determine m, the conservation
of energy given by (F.13) is considered, such that
v
u Nb
uX
R̂a = t (µi + mσi )2 (F.15)
i=0
Nb
X
R̂a2 = (µi + mσi )2 (F.16)
i=0
Nb
X Nb
X Nb
X
2
=m σi2 +m 2µi σi + (µi )2 . (F.17)
i=0 i=0 i=0

Solving (F.17) for m using the quadratic formula yields two possible solutions, m1 and
m2 . Due to squaring of the equation in (F.16), both a positive and negative solution for
m are possible. The highest valued solution should be chosen so that the FT is above the
mean value, i.e.
m = max(m1 , m2 ) . (F.18)

The rest of this section contains an example of the proposed RMS filter bank. The
IMS dataset introduced in Section F.2 is subdivided into Nb = 32 frequency bands (320
Hz bandwidth), and Ri is calculated for each band using (F.12). The mean µi and STD
σi of each Ri are determined with the baseline measurements. With these values, the
FTs for each Ri are calculated using (F.14) after solving (F.17) for m. Additionally, a
degradation alarm is triggered at time index kai when Ri,kai > µi + 5σi .
Fig. F.3 shows four RMS trends at different frequency ranges. Fig. F.3 (a) shows
R15 from frequency band [4480, 4800] Hz, and the oscillations in the trend are similar
to Ra shown in Fig. F.1 (b). On the other hand, R3 ([640, 920] Hz) in Fig. F.3 (d) is
monotonic after the initial alarm is triggered, and is very suitable for RUL estimation.
The two other RMS trends in Figs. F.3 (b) and (c) are calculated from frequency bands
between the two others, and it is seen that RMS trends from lower frequency bands are
more monotonic. The initial degradation alarm is also triggered at different times, and
 1.5 1.0
R15 R11 R6 R3
0.8
1.5 R̂15 R̂11 0.8 R̂6 R̂3
RMS [m/s2 ]

RMS [m/s2 ]

RMS [m/s2 ]

RMS [m/s2 ]
1.0 0.6
0.6
1.0
0.4
0.4
µ15 + 5σ15 0.5 µ11 + 5σ11 µ6 + 5σ6 µ3 + 5σ3
0.5
0.2 0.2

0.0 0.0 0.0 0.0

0 40 80 120 160 0 40 80 120 160 0 40 80 120 160 0 40 80 120 160
Time [hours] Time [hours] Time [hours] Time [hours]
(a) (b) (c) (d)

Figure F.3: Collection of RMS trends with FTs for the IMS dataset. (a) R15 ; (b) R11 ; (c)
R6 ; (d) R3 .

RMS trends from high frequency bands appear to trigger it first. In addition, the FTs
R̂i are also shown as blue lines in Fig. F.3. The FTs are reached near the end of useful
life and are therefore considered useful for RUL estimation. The next section details how
suitable RMS trends can be selected online and used to estimate the RUL.

F.4 Remaining useful life estimation

F.4.1 Overview
Suitable RMS trends for RUL estimation are identified by using the Spearman coefficient
[23] in an online manner, which determines the monotonicity of each Ri . A particle
filter (PF) is initialized for each Ri trend with a high Spearman coefficient, and the
Paris-Erdogan law [24] is used to predict the degradation trend. The model parameters
are initialized using a non-linear least squares (NLS) algorithm. For each new sample,
the PFs are updated, and the RUL probability density function (PDF) for each PF is
estimated. Finally, a weighted PDF combining the RUL estimation of all PFs is used to
estimate bearing RUL. A flowchart of the proposed method is shown in Fig. F.4.

F.4.2 Spearman coefficient

The Spearman coefficient [23] is a rank-based correlation between two signals. It can be
used to describe how monotonic a signal is, even when the signal is not linearly increasing.
The Spearman coefficient between a signal x and time t is calculated as

cov(rank(x), rank(t))
Spearman(x, t) = , (F.19)
STD(rank(x))STD(rank(t))

where cov(·, ·) is the covariance of two trends, and rank(·) is the rank of a signal. Recall
that Ri,k is a sample at index k of Ri , and kai is the index for when Ri,kai ≥ µi + 5σi for
 Figure F.4: Flowchart of the proposed method.

the first time. Then, the running Spearman coefficient for Ri is defined as

Spearman(Ri,ka :k , t) if k ≥ kai
i
ρi,k = (F.20)
0 else.

Eq. (F.20) is used to continually check whether Ri increases monotonously over time.

F.4.3 RMS trend selection

The Spearman coefficient is used to numerically describe how monotonic the RMS trends
are. Three criteria must be met by a certain Ri at time index k for it to be used for RUL
estimation:

1. k ≥ kai + ks, where ks is the minimum number of samples used to calculate Spear-
man coefficient.

2. ρi,k ≥ ρ̂, where ρ̂ is the Spearman threshold.

3. i ≤ min(Ip ), where Ip is a set of indices i belonging to R trends previously accepted

for RUL estimation.

The reasons of the criteria are as follows. 1) To calculate a stable Spearman coefficient, it
is necessary with several samples. Therefore, ks is set as a minimum number of samples.
2) The Spearman coefficient must be higher than the threshold to avoid non-monotonic
trends. 3) It is expected that RMS trends of low-frequency bands have less oscillations
compared to higher frequency ones. Therefore, the index i of new trends must be smaller
than all the other selected ones to avoid estimating RUL on unnecessary many trends. The
degradation model chosen to predict each RMS trend is detailed in the next subsection.
F.4.4 Degradation model
The bearing degradation level is assumed to be monotonously increasing and never self-
healing. In addition, as an increasing number of defects develop in the bearing, the
vibration level increases, which results in an exponential degradation rate. A study [25]
shows that bearing crack propagation may be modeled with the Paris-Erdogan law [24]
This model describes the crack propagation rate in materials under cyclic load, and is
given by
da √
= c(∆k)m , ∆k = ∆σγ πa , (F.21)
dnc
where a is the crack size, nc is the cycle number, c, m and γ are material constants, and
∆σ is the cyclic load amplitude. The material constants can be estimated via experimen-
tal testing with a known cyclic load, while measuring the crack length. However, it is
impractical to measure the crack length within a bearing during operation. Therefore, a
is instead estimated with an HI based on the vibration measurements [26].
√
Let α = γc∆σ π and β = m/2, then the modified Paris-Erdogan law is [13]

da
= αaβ . (F.22)
dnc

Eq. F.22 is afterwards re-written in the form of a state-space model as [13]


a = ak−1 + αk−1 aβk−1 ∆nc
 k



αk = αk−1 (F.23)



h
k = ak + ν h ,

where αk−1 ∼ N (µα , σα2 ) is a random variable given by a normal distribution N , β is a

constant, νh ∼ N (0, σh2 ) is measurement noise, and ∆nc is the number of cycles since last
update.
A PF is applied to update the model parameters based on new samples, but reasonable
initial values should be set first. The modified Paris-Erdogan law in (F.23) has five
unknown parameters (a1 , µα , β, σα2 , σh2 ) which must be identified. The mean values Θ1 =
(a1 , µα , β) are estimated using a non-linear least squares (NLS) algorithm, and afterwards
the variances Θ2 = (σα2 , σh2 ) are set based on other criteria

F.4.5 Initial parameter setting

Three model parameters Θ1 = (a1 , µα , β) are identified by minimizing an objective func-
tion using NLS. The objective function is the mean square error between measurements
and model prediction. Let the measurement HI be h = (Ri,k−ko:k )T , where ko is the num-
ber of samples used in the optimization. To obtain the objective function, the ordinary
differential equation in (F.22) is solved for a as

a(nc , Θ1 ) = (C1 (1 − β) + αnc (1 − β))1/(1−β) , β 6= 1 , (F.24)

where C1 is based on initial conditions. The process noise and measurement noise are
assumed zero in the minimization algorithm. C1 is determined by solving (F.24) with
initial values a = a1 and nc = 0. After substitution of C1 , (F.24) becomes
 1/(1−β)
a(nc , Θ1 ) = αnc (1 − β) + a11−β . (F.25)

The unknown parameters Θ1 are identified using an NLS minimization routine described
as

arg min F (h, Θ1 ) , (F.26)

Θ1



 min(h) ≤ a1 ≤ max(h)

subject to 0 < µα < ∞ (F.27)



0.5 < β < 1.25 ,
ko
X
where F = 0.5 (hi − a(nc = i − 1, Θ1 ))2 . (F.28)
i=1

A trust region reflective algorithm [27] is used in this research to minimize (F.26). The
constraints for β are set so that the exponential rate does not get unstable.
The variances in Θ2 = (σα2 , σh2 ) are set based on baseline data and optimized values in
Θ1 . The measurement noise variance σh2 is simply identified as the variance of Ri during
the baseline measurements, i.e. σh2 = σi2 . A particle filter is quite dependent on the choice
of process noise, i.e. σα2 , and a larger variance gives more headroom for the filter in case
the chosen model does not fit very well with measured data. To give enough headroom
for the filter, the mean value µα is used as the standard deviation, such that σα2 = µ2α .

F.4.6 Particle filter

After initializing the model parameters, a sequential importance sampling (SIS) PF is
applied to further update the parameters in the degradation model. General PF theory
is given in [13, 26, 28] for interested readers. A set of initial particles zkj ∀ j ∈ [1, Np ] are
initialized at time index k with
" # " #!
Ri,k σh2 0
zkj ∼N , , (F.29)
µα 0 σα2
where Np is the number of particles. The particle weights are initialized as wkj = 1/Np .
The next model state is predicted using the particles with
" # " #
j j j j j β
a a + (µ + ν α,k ) (a ) ∆t
zkj = kj = k−1 α
j
k−1
, (F.30)
µα µα
j
where να,k ∼ N (0, σα2 ) is the process noise which is added due to the uncertainty of bearing
load, and ∆t is the time between cycles. When a new measurement Ri,k is available, the
particle weights are updated and normalized with

w̃kj = wk−1
j
p(Ri,k | zkj ) (F.31)
, Np
X j
wkj = w̃kj w̃k , (F.32)
j=1

where w̃kj is the un-normalized particle weight and

(Ri,k − ak )2
 
j 1
p(Ri,k | zk ) = p exp − . (F.33)
2πσh2 2σh2

To deal with particle degeneracy, a re-sampling step is utilized if necessary. When

, Np
X j
Nef f = 1.0 wk < Np /2 , (F.34)
i=1

particles are re-sampled according to a systematic re-sampling [29] approach, and particle
weights are re-initialized as wkj = 1/Np .
The RUL estimated by particle j at time index k is given by

lkj = inf{lkj : aj (lk + tk ) ≥ R̂i | aj1:k } , (F.35)

where aj (lk + tk ) is the state value for particle j at time tk + lk , aj1:k is the estimated state
value at 1, . . . , k and R̂i is the FT for Ri . To solve (F.35), the state of each particle j are
simulated using the state transition function given by (F.30) up to the time aj (lk + tk ) ≥
R̂i . With the estimated RUL and weight for each particle, the probability density function
(PDF) for RUL lk is approximated by
Np
X
p(lk | Ri,1:k ) = wkj δ(lk − lkj ) , (F.36)
j=1

where δ() is the delta-dirac function.

F.4.7 Weighted RUL

The vibration signal is split into Nb frequency bands, where each band is used to calculate
RMS Ri . The running Spearman coefficient is used to detect monotonously increasing
RMS trends, and whenever a new trend has a high Spearman coefficient, a new PF is
initialized and updated for that trend. In theory, up to Nb PFs can be initialized, and the
RUL estimated from all are used in a weighted RUL decision. Let P Fi denote the PF for
j j
Ri , and let wi,k and li,k denote the weight and RUL, respectively, for particle j at time
index k in P Fi . When Ri stops being monotonic, the PF output associated with it should
count less in a weighted RUL estimation. Additionally, once the Spearman coefficient of
a trend is less than ρL , that trend is neglected in the weighted RUL estimation. Then,
let the weight of each P Fi be

Wi,k = ρ3i,k − ρ3L 1 − ρ3L ,



(F.37)

where the Spearman coefficient is cubed to prioritize monotonic trends. Afterwards, the
weighted RUL PDF at time index k is defined as
Nb Np

p(¯lk | R1:Nb ,1:k ) = δ(¯lk − li,k

X X j j
Wi,k wi,k ). (F.38)
i=1 j=1

Instead of using the median of the PDF in (F.38) as the estimated RUL, the weighted
mean is instead chosen. The reason is that the median of an even number of PFs will
most likely fall under either one of them, even if the medians are far from each other.
Instead, the weighted mean of the PDF is chosen as the estimated RUL. This weighted
mean is calculated with
PNb PNp j j
i=1 Wi,k j=1 wi,k li,k
Weighted mean(k) = PNb PNp j . (F.39)
i=1 Wi,k j=1 wi,k

The weighted mean and 95% confidence interval (CI) of the weighted RUL PDF are
determined after each measurement update to track the estimated RUL over time.

F.5 Experimental results

F.5.1 Test rig 1

To validate the proposed method, the IMS bearing dataset [21] is utilized. ks = 30 is
the minimum number of samples for Spearman calculation, ko = 30 samples are used to
initialize PFs, and ρ̂ = 0.9 is the Spearman threshold. In addition, Np = 1000 particles are
used in each PF, and ρL = 0.7 is the lower Spearman threshold in (4.16). Figs. F.5 and
F.6 shows the results of using the proposed method on this dataset. Subplots in the first
column show the identified RMS trends with high Spearman coefficients, and the output
of the initialized PF. Ri ’s are shown by black lines, and the median and 95% confidence
interval of the initialized PF output are shown as red and red-stapled lines, respectively.
 R̂i Ri Median 95% CI

×10−5
i = 15
ρ15 ≤ 0.7
1.5 1.0 1.5
RMS [m/s2 ]

RMS [m/s2 ]
0.5

µα
1.0 1.0

0.5 0.0 0.5

0.0 −0.5 0.0

100 110 120 130 140 150 160 110 120 130 140 150 100 110 120 130 140 150 160
Time [hours] Time [hours] Time [hours]
(a) ×10−5 (b) (c)
i = 14
ρ14 ≤ 0.7
2.0 1.0 2.0
RMS [m/s2 ]

RMS [m/s2 ]
1.5 1.5
0.5
µα

1.0 1.0
0.0
0.5 0.5
−0.5
0.0 0.0
100 110 120 130 140 150 160 110 120 130 140 150 100 110 120 130 140 150 160
Time [hours] Time [hours] Time [hours]
(d) (e) (f )
1.5 i = 12 1.5
ρ12 ≤ 0.7 0.0
RMS [m/s2 ]

RMS [m/s2 ]

1.0 1.0
0.0
µα

0.5 0.5
0.0

0.0 0.0
110 120 130 140 150 160 110 120 130 140 150 110 120 130 140 150 160
Time [hours] Time [hours] Time [hours]
(g) (h) (i)

Figure F.5: Identified RMS trends with high Spearman coefficient, and output of the
corresponding PFs—Part 1. Rows 1-3 indicate i = [15, 14, 12]. (column 1) Ri , FT R̂i ,
and median and 95% CI of initial PF output; (column 2) µα over time for the initiated
PF; (column 3) predicted PF trend at t = 130 hours.

In addition, blue lines indicate the FT. Each row in Figs. F.5 and F.6 corresponds to a
single index i, which is given in the upper left corner of the leftmost subplot. The second
column show the median and 95% confidence interval (CI) of the µα parameter. The
predicted PF output at t = 130 hours is given in column 3 to show how the particle filters
are converging over time.
At t = 101 hours (k = 608), a new PF is initialized for R15 , indicating a change
in bearing health. This RMS trend, shown in Fig. F.5 (a), is from a high-frequency
 R̂i Ri Median 95% CI

×10−5
1.0 1.0
i=6
1.0
0.8 0.8
RMS [m/s2 ]

RMS [m/s2 ]
0.6 0.5 0.6

µα
0.4 0.4
0.0
0.2 0.2

0.0 −0.5 0.0

110 120 130 140 150 160 120 130 140 150 110 120 130 140 150 160
Time [hours] Time [hours] Time [hours]
(a) ×10−5 (b) (c)
i=3 3
0.8 0.8
2
RMS [m/s2 ]

RMS [m/s2 ]
0.6 0.6
1
µα

0.4 0 0.4

−1
0.2 0.2
−2
0.0 0.0
120 130 140 150 160 130 140 150 120 130 140 150 160
Time [hours] Time [hours] Time [hours]
(d) (e) (f )

Figure F.6: Identified RMS trends with high Spearman coefficient, and output of the
corresponding PFs—Part 2. Rows 1-2 indicate i = [6, 3]. (column 1) Ri , FT R̂i , and
median and 95% CI of initial PF output; (column 2) µα over time for the initiated PF;
(column 3) predicted PF trend at t = 130 hours.

band within [4480, 4800] Hz. Initially, R15 increases monotonously, but at t = 117 hours
the trend starts to oscillate around a mean value. This behavior is similar to the cyclic
behavior observed in the full RMS Ra . µα has converged before t = 110 hours as seen in
Fig. F.5 (b), and therefore the estimated trend will continue until the end of RUL. Fig.
F.5 (c) shows the median and 95% CI output of P F15 at t = 130 hours. A vertical black
line in Fig. F.5 (a) shows when the running Spearman value is less than the requirement
of 0.7. At this time, the trend is determined not suitable for RUL estimation after all,
and its weight W15 is 0 as given by (F.37).
R14 and R12 passes the criteria for RUL estimation at t ≈ 108 hours, and the initial
output of the PFs are shown in Figs. F.5 (d) and (g), respectively. R14 and R12 are
both similar to R15 , and gets a low Spearman coefficient eventually as indicated by black-
stapled vertical lines due to the non-monotonic behavior of the trend. The parameter
µα in P F14 and P F12 also converge as indicated in Figs. F.5 (b) and (e), respectively.
Therefore, the converged PF output trends will continue until the Spearman coefficient
values go below 0.7. The converged output of P F14 and P F12 at t = 130 hours are shown
in Fig. F.5 (f) and (i), respectively.
At t = 113 hours, P F6 is initialized, and the initial PF output is shown in Fig. F.6
(a). R6 has less oscillations compared to the three previously identified RMS trends,
and therefore the Spearman coefficient never gets below 0.7. The parameter µα quickly
converges as shown in Fig. F.6 (b). This is because the sudden increase in value of R6 at
t ≈ 117 hours gives particles with high µα value a large weight. The predicted output is
shown in Fig. F.6 (c) at t = 130 hours. The converged output should reach the FT early
at t ≈ 150, but the PF is continually updated on new samples. Given the uncertainty of
bearing load, the PF may follow the future measurement samples.
When t = 124 hours, P F3 is initialized, and R3 together with the PF output are shown
in Fig. F.6 (d). R3 is the most monotonic trend of the five, and the initial PF output
median matches the future samples well. At t = 130 hours, µα starts to converge as seen
in Fig. F.6 (e), and the PF output is shown in Fig. F.6 (f). The model prediction at this
point is directed at the end of lifetime, and therefore the RUL is accurately estimated at
this point.
The weighed RUL mean and 95% CI for this dataset are shown in Fig. F.7 (a). Here,
the true RUL is shown as a black solid line, while the weighted mean and 95% CI are the
red and red-stapled lines, respectively. The weighted mean oscillates around the true RUL
until the end of life. At t = 125 hours, all five PFs have a weight as shown in Fig. F.7 (b).
From t = 130 hours, the estimated RUL is close to the true RUL. The reason for this is as
follows. The median P F12 estimate in Fig. F.5 (i) over-estimates the RUL. In addition,
the median of P F6 underestimates the RUL while P F3 matches the RUL, as indicated
in Figs. F.6 (c) and (f). The weighted mean will therefore fall somewhere between these
three outcomes. Afterwards, the weights W15 , W14 and W12 decreases towards 0 due to a
low Spearman coefficient, and R6 and R3 are the only two trends left near the end of life.
During this time, the estimated RUL is very close to the true RUL.
To visualize how the weighted RUL PDF changes over time, Fig. F.8 shows a 3D
plot where the z-axis is the smoothened PDF. This plot shows why the weighted mean is
calculated in comparison to the median. The true RUL is mainly situated between large
peaks, while the discretely calculated median would have been situated under either peak
with the highest weight. Therefore, it is more natural to estimate the RUL as the center
between PFs.

F.5.2 Test rig 2 — in-house setup

Data from the second test rig is used to further validate the performance of the proposed
method. A 6008-type roller element bearing is worn naturally in an accelerated life test
 100
95% CI
90
Weighted mean
80 True RUL
Remaining useful life [hours]

70

60

50

40

30

20

10

100 105 110 115 120 125 130 135 140 145 150 155 160
Time [hours]
(a)
1.0

0.8
Trend weights

0.6 W15
W14
0.4
W12
0.2 W6
W3
0.0
100 105 110 115 120 125 130 135 140 145 150 155 160
Time [hours]
(b)

Figure F.7: Weighted RUL of the IMS dataset. (a) weighted mean, 95% CI and true RUL
of the IMS dataset; (b) weights for each PF output.

by applying radial and axial loads. The dynamic capacity of the bearing is 17.8 kN, and
the static capacity is 11 kN. With constant radial and axial loads of 9 kN and 5 kN,
respectively, the bearing lasted approximately 34.6 million revolutions before failing due
to an outer race fault. At a shaft speed of 100 rpm, vibration data was sampled every 15
minutes at a rate of 51200 Hz for 6 seconds. The 154 last hours of operation are used to
verify the performance of the proposed method. More details of the test rig are given in
[30].
The velocity- and acceleration-based RMS are shown in Figs. F.9 (a) and (b), respec-
tively. A 1.1 kW motor is used, and therefore the velocity-based FT is set to R̂v = 4.0
mm/s. The FT R̂v is not reached entirely at the end of useful life, as shown in Fig. F.9
(a). On the other hand, the transformed acceleration-based FT R̂a is far from reached in
Fig. F.9 (b). The transformation from R̂v to R̂a is not analytic and may therefore not be
accurate. The test was stopped due to a high rate of change in Ra , and the machine could
possibly have been run for more cycles. If the test was run for a few more measurement
Figure F.8: Weighted RUL PDF of the IMS dataset over time. The weighted mean and
true RUL are shown for reference.

cycles, R̂a might have been reached.

The proposed method is used to split the vibration signal into Nb = 64 frequency
bands, resulting in a frequency bandwidth of 400 Hz. A few of the RMS trends are
shown in Fig. F.10. R40 in Fig. F.10 (a) is oscillating, and hence not suitable for RUL
estimation. R19 and R11 in Figs. F.10 (b) and (c) have less oscillations, but do not
increase steadily. R1 in Fig. F.10 (d) on the other hand, increases almost linearly after
the alarm is triggered, which makes the trend suitable for RUL estimation.

Using the proposed RUL estimation algorithm, the first and only identified RMS band
is R1 at t = 120 hours. The initial PF output and R1 are shown in Fig. F.11 (a), and
µα is shown in Fig. F.11 (b). The trend is re-drawn in Fig. F.11 (c) to show the PF
output at t = 130 hours, which follows the future samples well. Since the predicted trend
is similar to new samples, the estimated RUL is accurate. The weighted RUL is shown
in Fig. F.12, and it follows the true RUL well. In this case, the estimated FT is reached
very close to the actual end of life, and therefore the RUL estimation is accurate. The
3D plot in Fig. F.13 also shows how the weighted RUL PDF changes over time.
 5 2.0

4
1.5
RMS [mm/s]

RMS [m/s2 ]
3 Rv Ra
1.0
R̂v R̂a
2
0.5
1

0 0.0
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160
Time [hours] Time [hours]
(a) (b)

Figure F.9: Comparison of velocity- and acceleration-based RMS of the in-house test rig
dataset. (a) velocity-based RMS; (b) acceleration-based RMS.

0.125 0.04
R40 R19 R11 R1
0.05 0.3
R̂40 0.100 R̂19 R̂11 R̂1
0.03
RMS [m/s2 ]

RMS [m/s2 ]

RMS [m/s2 ]

RMS [m/s2 ]
0.04
0.075 0.2
0.03 0.02
0.050
0.02 µ11 + 5σ11 µ1 + 5σ1
µ40 + 5σ40 µ19 + 5σ19 0.1
0.01
0.01 0.025

0.00 0.000 0.00 0.0

0 40 80 120 160 0 40 80 120 160 0 40 80 120 160 0 40 80 120 160
Time [hours] Time [hours] Time [hours] Time [hours]
(a) (b) (c) (d)

Figure F.10: Collection of RMS trends with FTs for the in-house test rig dataset. (a)
R40 ; (b) R19 ; (c) R11 ; (d) R1 .

R̂i Ri Median 95% CI

×10−4
i=1
0.75
0.3 0.3
0.50
RMS [m/s2 ]

RMS [m/s2 ]

0.2 0.25 0.2

µα

0.00
0.1 0.1
−0.25
0.0 0.0
110 120 130 140 150 120 130 140 150 110 120 130 140 150
Time [hours] Time [hours] Time [hours]

Figure F.11: R1 and initial median and 95% CI output of the corresponding PF. (a) R1
with marked part as optimization input, blue line for FT, and red lines for median and
95% CI of PF; (b) α over time for the initiated PF.
 70
95% CI
60 Weighted mean
True RUL
Remaining useful life [hours]

50

40

30

20

10

110 115 120 125 130 135 140 145 150

Time [hours]

Figure F.12: Weighed RUL of the in-house test rig dataset. The weighted mean, 95% CI
and true RUL are shown.

Figure F.13: Weighted RUL PDF of the in-house test rig dataset over time. The weighted
mean and true RUL are shown for reference.

F.5.3 Comparisons
In this section, other research using the same IMS dataset are compared with the proposed
method. In [31], the RUL was estimated using a feedforward artificial neural network
(FFNN). However, training such an FFNN requires historic failure data, and therefore
the results cannot be compared to the proposed method. In [32], the proportional hazard
model and logistic regression model were used, but historic failure data of the machine
was required as well. Relevance vector machine (RVM) was used in [33], and the results
show good estimation of the RUL. But, historical failure data was required to train the
RVM. An enhanced phase space warping (PSW) method is proposed in [34] to combine
the advantages of physics-based and data-driven techniques. The estimated RUL is close
to the actual RUL, but historical failure data from a different test on the same machine
was used to determine a few parameters. Soualhi et al. [35] proposed a method using
an artificial ant clustering (AAC) technique for classifying faults, a hidden Markov model
(HMM) to detect changes in degradation stage, and an adaptive neuro-fuzzy inference
system (ANFIS) for RUL estimation. However, historical failure data was necessary to
train the AAC. Ahmad et al. [36] presented a hybrid technique for RUL estimation that
rectifies RMS fluctuations, and uses least squares minimization to fit a quadratic model
to the rectified RMS. The gradient of the quadratic model is used as FT, and it was set
based on available historical failure data. The resulting RUL estimation is quite accurate.
However, due to usage of historic failure data to set the FT, the results cannot be compared
to the proposed method. The authors could not identify publications showcasing RUL
estimation on the IMS dataset without use of historic failure data. Therefore, direct RUL
estimation comparison is not performed.

F.6 Further work

This section presents opportunities for further work to improve the proposed method.
High-frequency RMS bands over 2 kHz may be used for health stage (HS) division. Energy
is required to form cracks in materials, and that energy is dissipated during fracture, i.e.
strain energy release. It has been observed that the RMS of a high frequency band seem
to mimic this physical phenomenon. This information may be used to divide the life of the
bearing into multiple health stages (HSs). In the IMS dataset, R15 is a high-frequency
RMS band that shows signs of energy oscillations, seemingly due to crack formation.
Fig. F.14 (a) shows the result of HS division due to strain energy release. The initial
degradation alarm is first triggered for R15 , which marks the transition to HS 2. The RMS
is increasing slowly over time, until R15 jumps to a high value ≈ 1 m/s2 . Now the RMS is
seemingly high enough for more damage to occur, and during the fracture, R15 starts to
decrease, which marks the change to a new HS. With the increased amount of damage,
bearing impacts also increase vibration energy at lower frequency RMS bands. R3 in Fig.
F.14 (b) is a low-frequency band, and the RMS starts to increase linearly during HS 3.
At t = 130 hours, R15 has reached a local minimum before it starts to increase again and
 1.4
R15
1.2 Linear trend
1.0 HS Change
RMS [m/s2 ]

0.8

0.6

0.4

0.2

0.0
(a)
0.8
R3

0.6
RMS [m/s2 ]

0.4

0.2

HS 1 HS 2 HS 3 HS 4 HS 5+
0.0
70 75 80 85 90 95 100 105 110 115 120 125 130 135 140 145 150 155 160
Time [hours]
(b)

Figure F.14: Health stage division from energy cycles of a high-frequency band on the
IMS dataset. (a) High-frequency band R15 and identified HS transitions; (b) low-frequency
band R3 .

reaches another maximum at t = 143 hours. The rate of R3 increases again afterwards,
as there are seemingly more bearing impacts. In HS 4, R15 decreases and increases within
10 hours, and the bearing enters yet another HS. From HS 5+, the energy cycles are very
short, just a few hours, indicating that critical failure is imminent.
For the second test rig, the RMS in a high-frequency band could indicate HS changes.
Fig. F.15 (a) shows R40 and the identified peaks at the vertical stapled lines. At first, R40
increases until it reaches a top value, after which point it decreases again. This signifies
an HS change, and the low-frequency R1 in Fig. F.15 (b) increases linearly from that
point. When then next RMS top point is reached at t = 135 hours, the rate of change in
R1 does not increase. Instead, the linear trend stops to increase for 7 hours, until it starts
again at t = 142 hours. At this point, R40 reaches a small peak point. This phenomenon
may be investigated in future work, as this information could be used to make a better
RUL estimation. For example, the energy oscillation in a high-frequency band can be
predicted to assume a higher rate of lower-frequency RMS trends ahead of time.
The RMS filter bank has been used to estimate RUL by identifying monotonic RMS
trends, and updating model parameters using a particle filter. However, the presented
 0.10
R40
0.08 Linear trend
HS Change
RMS [m/s2 ]

0.06

0.04

0.02

0.00
(a)
0.30
R1
0.25

0.20
RMS [m/s2 ]

0.15

0.10

0.05
HS 1 HS 2 HS 3 HS 4 HS 5+
0.00
100 105 110 115 120 125 130 135 140 145 150
Time [hours]
(b)

Figure F.15: Health stage division from energy cycles of a high-frequency band on the
in-house test rig dataset. (a) High-frequency band R40 and identified HS transitions; (b)
low-frequency band R1 .

method may be improved in future work. The frequency bands were evenly split with
equal frequency bandwidth, and it should be investigated if there are more optimal ways
to subdivide the energy bands. Secondly, it should be checked if the presented filter bank
could be used for other purposes than calculating the RMS. Finally, there may be better
ways to use all the RMS bands for RUL estimation rather than filtering each trend with
a PF.

F.7 Conclusions
In this paper, a new method for subdividing the vibration signal into multiple frequency
bands for root mean square (RMS) calculations is proposed. The method utilizes a single
discrete Fourier transform (DFT) per signal, and individual bins are used to acquire
the signal energy within a frequency band. The Spearman coefficient is used to identify
monotonic RMS trends that are suitable for remaining useful life (RUL) estimation. It
is observed that low-frequency RMS bands are most monotonic, while higher frequency
RMS bands show earlier sign of degradation. The failure threshold (FT) for vibration
RMS, developed in earlier research, has been extended for the RMS frequency bands. A
particle filter (PF) is applied to estimate parameters for the Paris-Erdogan law, and a
single instance is used for every monotonic RMS trend. The RUL is afterwards estimated
by weighing the output RUL of all initiated PFs. Experimental results show that the
proposed method produces good RUL estimations without the use of historic failure data.
The resulting RUL estimation in this research is, however, less accurate compared to
referenced work that uses historical failure data for model training. Therefore, more
investigation on the RUL estimation algorithm is necessary to improve performance.
REFERENCES

[1] P. Zhang, Y. Du, T. G. Habetler, and B. Lu. A survey of condition monitoring and
protection methods for medium-voltage induction motors. IEEE Transactions on
Industry Applications, 47(1):34–46, 2011. doi:10.1109/tia.2010.2090839.

[2] P. D. McFadden and J. D. Smith. Model for the vibration produced by a single point
defect in a rolling element bearing. Journal of Sound and Vibration, 96(1):69–82,
1984. doi:10.1016/0022-460X(84)90595-9.

[3] Y. Lei, N. Li, L. Guo, N. Li, T. Yan, and J. Lin. Machinery health prognostics: A
systematic review from data acquisition to RUL prediction. Mechanical Systems and
Signal Processing, 104:799–834, 2018. doi:10.1016/j.ymssp.2017.11.016.

[4] R. K. Singleton, E. G. Strangas, and S. Aviyente. The use of bearing currents

and vibrations in lifetime estimation of bearings. IEEE Transactions on Industrial
Informatics, 13(3):1301–1309, 2017. doi:10.1109/tii.2016.2643693.

[5] C. Hu, B. D. Youn, P. Wang, and J. T. Yoon. Ensemble of data-driven prognostic

algorithms for robust prediction of remaining useful life. Reliability Engineering &
System Safety, 103:120–135, 2012. doi:10.1016/j.ress.2012.03.008.

[6] N. Li, Y. Lei, J. Lin, and S. X. Ding. An improved exponential model for predicting
remaining useful life of rolling element bearings. IEEE Transactions on Industrial
Electronics, 62(12):7762–7773, 2015. doi:10.1109/TIE.2015.2455055.

[7] Y. Lei, N. Li, and J. Lin. A new method based on stochastic process models for
machine remaining useful life prediction. IEEE Transactions on Instrumentation
and Measurement, 65(12):2671–2684, 2016. doi:10.1109/TIM.2016.2601004.

[8] N. Gebraeel, M. Lawley, R. Liu, and V. Parmeshwaran. Residual life predictions from
vibration-based degradation signals: a neural network approach. IEEE Transactions
on industrial electronics, 51(3):694–700, 2004. doi:10.1109/TIE.2004.824875.

221
 [9] Y. Wang, Y. Peng, Y. Zi, X. Jin, and K.-L. Tsui. A two-stage data-driven-based prog-
nostic approach for bearing degradation problem. IEEE Transactions on Industrial
Informatics, 12(3):924–932, 2016. doi:10.1109/TII.2016.2535368.

[10] C. Lu, J. Chen, R. Hong, Y. Feng, and Y. Li. Degradation trend estimation of
slewing bearing based on LSSVM model. Mechanical Systems and Signal Processing,
76:353–366, 2016. doi:10.1016/j.ymssp.2016.02.031.

[11] W. Wang. A model to predict the residual life of rolling element bearings given
monitored condition information to date. IMA Journal of Management Mathematics,
13(1):3–16, 2002. doi:10.1093/imaman/13.1.3.

[12] D. Wang and K.-L. Tsui. Statistical modeling of bearing degradation signals. IEEE
Transactions on Reliability, 66(4):1331–1344, 2017. doi:10.1109/tr.2017.2739126.

[13] Y. Lei, N. Li, S. Gontarz, J. Lin, S. Radkowski, and J. Dybala. A model-based

method for remaining useful life prediction of machinery. IEEE Transactions on
Reliability, 65(3):1314–1326, 2016. doi:10.1109/TR.2016.2570568.

[14] R. K. Singleton, E. G. Strangas, and S. Aviyente. Extended Kalman filtering for

remaining-useful-life estimation of bearings. IEEE Transactions on Industrial Elec-
tronics, 62(3):1781–1790, 2015. doi:10.1109/tie.2014.2336616.

[15] Y. Qian and R. Yan. Remaining useful life prediction of rolling bearings using an
enhanced particle filter. IEEE Transactions on Instrumentation and Measurement,
64(10):2696–2707, 2015. doi:10.1109/tim.2015.2427891.

[16] P. Boškoski, M. Gašperin, D. Petelin, and D̄. Juričić. Bearing fault prognostics using
Rényi entropy based features and Gaussian process models. Mechanical Systems and
Signal Processing, 52:327–337, 2015. doi:10.1016/j.ymssp.2014.07.011.

[17] S. A. Khan, A. E. Prosvirin, and J.-M. Kim. Towards bearing health prognosis using
generative adversarial networks: Modeling bearing degradation. In International
Conference on Advancements in Computational Sciences (ICACS), pages 1–6. IEEE,
2018. doi:10.1109/ICACS.2018.8333495.

[18] J. Zhu, N. Chen, and W. Peng. Estimation of Bearing Remaining Useful Life based
on Multiscale Convolutional Neural Network. IEEE Transactions on Industrial Elec-
tronics, 66(4):3208–3216, 2018. doi:10.1109/tie.2018.2844856.

[19] A. Klausen, H. Van Khang, and K. G. Robbersmyr. Novel Threshold Calculations

for Remaining Useful Lifetime Estimation of Rolling Element Bearings. In 2018 XIII
 International Conference on Electrical Machines (ICEM), pages 1912–1918. IEEE,
2018. doi:10.1109/ICELMACH.2018.8507056.

[20] ISO. Mechanical vibration—Evaluation of machine vibration by measurements on

non-rotating parts—Part 3 (Standard No. 10816-3). ISO, 1998.

[21] H. Qiu, J. Lee, J. Lin, and G. Yu. Wavelet filter-based weak signature detection
method and its application on rolling element bearing prognostics. Journal of Sound
and Vibration, 289(4-5):1066–1090, 2006. doi:10.1016/j.jsv.2005.03.007.

[22] B. Duong, S. Khan, D. Shon, K. Im, J. Park, D.-S. Lim, B. Jang, and J.-M. Kim.
A Reliable Health Indicator for Fault Prognosis of Bearings. Sensors, 18(11):3740,
2018. doi:10.3390/s18113740.

[23] C. Spearman. The proof and measurement of association between two things. The
American journal of psychology, 15(1):72–101, 1904. doi:10.1093/ije/dyq191.

[24] P. Paris and F. Erdogan. A critical analysis of crack propagation laws. Journal of
basic engineering, 85(4):528–533, 1963. doi:10.1115/1.3656900.

[25] P. Rycerz, A. Olver, and A. Kadiric. Propagation of surface initiated rolling contact
fatigue cracks in bearing steel. International Journal of Fatigue, 97:29–38, 2017.
doi:10.1016/j.ijfatigue.2016.12.004.

[26] D. An, J.-H. Choi, and N. H. Kim. Prognostics 101: A tutorial for particle filter-
based prognostics algorithm using Matlab. Reliability Engineering & System Safety,
115:161–169, 2013. doi:10.1016/j.ress.2013.02.019.

[27] M. A. Branch, T. F. Coleman, and Y. Li. A subspace, interior, and conjugate gradient
method for large-scale bound-constrained minimization problems. SIAM Journal on
Scientific Computing, 21(1):1–23, 1999. doi:10.1137/S1064827595289108.

[28] M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp. A tutorial on particle

filters for online nonlinear/non-Gaussian Bayesian tracking. IEEE Transactions on
signal processing, 50(2):174–188, 2002. doi:10.1109/78.978374.

[29] J. D. Hol, T. B. Schon, and F. Gustafsson. On resampling algorithms for particle

filters. In IEEE Nonlinear Statistical Signal Processing Workshop, pages 79–82. IEEE,
2006. doi:10.1109/nsspw.2006.4378824.

[30] A. Klausen, R. W. Folgerø, K. G. Robbersmyr, and H. R. Karimi. Accelerated

Bearing Life-time Test Rig Development for Low Speed Data Acquisition. Modeling,
Identification and Control, 38(3):143–156, 2017. doi:10.4173/mic.2017.3.4.
[31] A. K. Mahamad, S. Saon, and T. Hiyama. Predicting remaining useful life of ro-
tating machinery based artificial neural network. Computers & Mathematics with
Applications, 60(4):1078–1087, 2010. doi:10.1016/j.camwa.2010.03.065.

[32] H. Liao, W. Zhao, and H. Guo. Predicting remaining useful life of an individual
unit using proportional hazards model and logistic regression model. In Annual
Reliability and Maintainability Symposium, RAMS’06, pages 127–132. IEEE, 2006.
doi:10.1109/RAMS.2006.1677362.

[33] A. Widodo and B.-S. Yang. Application of relevance vector machine and survival
probability to machine degradation assessment. Expert Systems with Applications,
38(3):2592–2599, 2011. doi:10.1016/j.eswa.2010.08.049.

[34] Y. Qian, R. Yan, and R. X. Gao. A multi-time scale approach to remaining useful life
prediction in rolling bearing. Mechanical Systems and Signal Processing, 83:549–567,
2017. doi:10.1016/j.ymssp.2016.06.031.

[35] A. Soualhi, H. Razik, G. Clerc, and D. D. Doan. Prognosis of bearing failures

using hidden Markov models and the adaptive neuro-fuzzy inference system. IEEE
Transactions on Industrial Electronics, 61(6):2864–2874, 2014. doi:10.1109/tie.
2013.2274415.

[36] W. Ahmad, S. A. Khan, and J.-M. Kim. A hybrid prognostics technique for rolling
element bearings using adaptive predictive models. IEEE Transactions on Industrial
Electronics, 65(2):1577–1584, 2018. doi:10.1109/TIE.2017.2733487.