Condition monitoring of gears using transmission error

Author:
Chin, Jacky
Publication Date:
2022
DOI:
https://doi.org/10.26190/unsworks/24007
License:
https://creativecommons.org/licenses/by/4.0/
Link to license to see what you are allowed to do with this resource.

Downloaded from http://hdl.handle.net/1959.4/100300 in https://

unsworks.unsw.edu.au on 2024-05-27
Condition monitoring of gears using transmission error

Zhan Yie Chin

A thesis in fulfilment of the requirements for the degree of

Doctor of Philosophy

School of Mechanical and Manufacturing Engineering

Faculty of Engineering

University of New South Wales, Australia

December 2021
Thesis Title and Abstract

i
Originality, Copyright and Authenticity
Statements

ii
Inclusion of Publications Statement

iii
Acknowledgements
This research was partly supported by the Australian Government through the Australian
Research Council’s Discovery Projects funding scheme (project DP160103501 and
DP190103231).

I would not have made it through my PhD without the support of many people. I would therefore
like to take this opportunity to acknowledge those who have accompanied me throughout this
challenging but delightful journey.

First and foremost, to my supervisor Associate Professor Pietro Borghesani, who believes in me
more than I do myself, I would like to extend my immense gratitude for your insightful mentorship
and boundless patience. Thank you for inspiring me with your enthusiasm in research, and for
your generosity in sharing your knowledge.

To my supervisor Professor Zhongxiao Peng, who has guided me since day one, I would like to
express my sincere appreciation for your kind advice and invaluable insights. Thank you for
giving me the opportunity to be part of this research group and for being a constant source of
encouragement throughout this journey.

To my supervisor Dr. Wade Smith, I am deeply grateful for your unwavering support and
constructive ideas. Thank you for your practical advice in the development of my project. It has
been a pleasure growing under your guidance.

To Emeritus Professor Robert Randall, thank you for your unparalleled wisdom and willingness
to share your expertise.

To all my colleagues from the research group, thank you for the productive discussions, for
helping in the experiments, and for the memorable gatherings.

Many thanks to our laboratory technical officers for your assistance in my experiments.

I would also like to thank UNSW for the financial support through the UIPA scholarship.

Finally, to my parents, my sister, and my partner, thank you very much for your unconditional
love and support.

iv
Abstract
Condition monitoring is crucial for safe and economic machine operations. Although
vibration dominates gear condition monitoring and has been undeniably proven successful in
early fault detection, its effectiveness in fault severity assessment is still debated. Disadvantages
of vibration in such applications include remoteness from the gears, a complex transfer path,
insensitivity to average wear, masking by other sources, and strong dependence on operating
conditions. Solutions are available, but usually require complex gear models and/or large datasets
for the training of data-driven methods.

An underappreciated alternative to vibration is gear transmission error (TE), based on

angular encoder measurements of the gear shafts. The more direct connection of the sensors to
the gears suggests a much easier conversion of TE into a micrometre measurement of gear wear
and tooth deflections. However, the lack of scientific analysis of TE leaves three key challenges.
Firstly, like vibration, current TE measurements lack an absolute reference and are therefore
insensitive to average wear. Secondly, despite being much simpler than that of vibration, a transfer
path still exists between the gear mesh and measurements, affecting accuracy at higher operating
speeds. Finally, profile-error and tooth-deflection components must be separated from within TE
to allow for their correlation with wear and cracks, respectively.

This thesis aims to address these gaps to enable the effective use of TE in gear condition
monitoring through several innovations. Firstly, a procedure was proposed to obtain “absolute
TE”, which includes information on average wear depth. Then, a technique was developed for the
removal of transfer-path effects, extending the applicability of TE to higher speeds. Finally, the
latter was integrated within a broader method, able to separate wear- and crack-related
components and automatically estimate crack severity. To support this work, a vast set of unique
experimental wear and crack tests were conducted, providing new insights on these faults and
their impact on TE.

These developments are not separate, and form a coherent strategy for the use of TE in
fault severity assessment. Its accurate, reliable and physically justified results constitute a crucial
resource for future work in the prognostics and health management of gears.

v
List of Publications

• * Zhan Yie Chin, Pietro Borghesani, Yuanning Mao, Wade A. Smith, Robert B. Randall,
Use of transmission error for a quantitative estimation of root-crack severity in gears,
Mech. Syst. Signal Process. 171 (2022) 108957.
https://doi.org/10.1016/j.ymssp.2022.108957.
• * Zhan Yie Chin, Wade A. Smith, Pietro Borghesani, Robert B. Randall, and Zhongxiao
Peng, “Absolute transmission error: A simple new tool for assessing gear wear,” Mech.
Syst. Signal Process., vol. 146, p. 107070, Jan. 2021, doi: 10.1016/j.ymssp.2020.107070.
• Robert B. Randall, Zhan Yie Chin, Wade A. Smith, and Pietro Borghesani,
“Measurement and use of transmission error for diagnostics of gears,” in Survishno, 2019,
no. July.
• Ke Feng, Pietro Borghesani, Wade A. Smith, Robert B. Randall, Zhan Yie Chin, Jinzhao
Ren, Zhongxiao Peng, “Vibration-based updating of wear prediction for spur gears,”
Wear, vol. 426–427, 2019, doi: 10.1016/j.wear.2019.01.017.
• 1
H. André, Q. Leclère, D. Anastasio, Y. Benaïcha, K. Billon, M. Birem, F. Bonnardot,
Z.Y. Chin, F. Combet, P.J. Daems, A.P. Daga, R. De Geest, B. Elyousfi, J. Griffaton, K.
Gryllias, Y. Hawwari, J. Helsen, F. Lacaze, L. Laroche, X. Li, C. Liu, A. Mauricio, A.
Melot, A. Ompusunggu, G. Paillot, S. Passos, C. Peeters, M. Perez, J. Qi, E.F. Sierra-
Alonso, W.A. Smith, X. Thomas, “Using a smartphone camera to analyse rotating and
vibrating systems: Feedback on the SURVISHNO 2019 contest,” Mech. Syst. Signal
Process., vol. 154, p. 107553, Jun. 2021, doi: 10.1016/J.YMSSP.2020.107553.

* Publications included in this thesis.

1
This publication is prepared by the organisers of a contest which I have participated in.

vi
Abbreviations
AE Acoustic emission
CBM Condition-based maintenance
DTE Dynamic transmission error
EMP Electromagnetic particle
FE Finite element
FRF Frequency response function
GM Gearmesh
GTE Geometric transmission error
HCF High-cycle fatigue
HTF Hunting tooth frequency
HTP Hunting tooth period
IAS Instantaneous angular speed
LCF Low-cycle fatigue
OMA Operational modal analysis
PE Potential energy
RMS Root mean square
RUL Remaining useful life
SA Synchronous average
STE Static transmission error
TE Transmission error
TVMS Time-varying mesh stiffness
VFD Variable frequency drive

vii
List of Figures
Figure 2-1. Deflected teeth due to load vs ideal rigid gear (top); teeth with geometrical errors vs
ideal involute gear (bottom). The scale of the effects is exaggerated for clarity. ....................... 19
Figure 2-2. Worn tooth vs healthy gear (left); cracked tooth vs healthy gear (right). The scale of
the effects is exaggerated for clarity. .......................................................................................... 19
Figure 2-3. Expected STE variation in the case of healthy gear (qualitative). ........................... 21
Figure 2-4. Effects of local and uniform material removal on TE. ............................................. 24
Figure 2-5. Hunting tooth period in the case of hunting tooth design, example with 12 and 17 gear
teeth. ............................................................................................................................................ 26
Figure 3-1. Overall methodology of thesis. ................................................................................ 30
Figure 3-2. Relationships between profile error, tooth flexibility and transfer function. ........... 32
Figure 3-3. Methodology to remove transfer function and separate profile error and tooth
flexibility. .................................................................................................................................... 33
Figure 3-4. Crack detection and model-based severity assessment. ........................................... 34
Figure 3-5. Obtaining absolute TE and its application to identify average wear depth. ............. 35
Figure 3-6. Spur gear test rig. ..................................................................................................... 36
Figure 3-7. Gearbox components. ............................................................................................... 36
Figure 3-8. Schematic diagram of spur gear test rig. .................................................................. 37
Figure 3-9. Gear tooth with small (left), medium (middle) and large (right) root crack. ........... 38
Figure 3-10. Tooth surfaces – (a, b) driving gear; (c, d) driven gear; (a, c) healthy; (b, d) worn.
.................................................................................................................................................... 40
Figure 3-11. Surface images of tooth at different wear stage (as described in [106]). ............... 41
Figure 4-1. Relationship between GTE, STE and Δ𝑇𝑇𝑇𝑇 in (a) angular domain, (b) order-amplitude
domain. ....................................................................................................................................... 44
Figure 4-2. Relationship between STE, DTE and 𝐻𝐻 (𝐹𝐹𝐹𝐹𝐹𝐹) in the order domain at input speed of
5 Hz (left column) and 10 Hz (right column). ............................................................................ 47
Figure 4-3. Summary of the proposed methodology. ................................................................. 47
Figure 4-4. DTE signals at different speeds and loads after synchronous averaging w.r.t input
shaft and band-pass filtering for input speed of (a) 2 Hz, (b) 5 Hz, and (c) 10 Hz. .................... 52
Figure 4-5. Order domain of DTE signals at different speeds and loads for (a) 2 Hz, (b) 5 Hz, and
(c) 10 Hz. .................................................................................................................................... 53
Figure 4-6. An example of the removal of transfer function from the DTE signal at 10 Hz and 9.9
Nm, (a) DTE, fitted transfer function and equivalent STE in the order spectrum and (b) DTE and
equivalent STE in angular domain (one shaft revolution). ......................................................... 54

viii
Figure 4-7. Original DTE (top) and estimated equivalent STE (bottom) for all speeds at ~10 Nm.
.................................................................................................................................................... 55
Figure 4-8. Order-domain regression analysis to estimate compliance and GTE using equivalent
STE signals. ................................................................................................................................ 56
Figure 4-9. Angular-domain regression analysis equivalent to the order-domain approach. ..... 56
Figure 4-10. (a) Estimated compliance and (b) Estimated GTE. ................................................ 57
Figure 4-11. Order spectra of DTE for operating speeds of 10, 15 and 20 Hz. .......................... 58
Figure 4-12. An example of the removal of transfer function from the DTE signal at 20 Hz and
9.9 Nm, (a) DTE, fitted FRF and equivalent STE in the order spectrum and (b) DTE and
equivalent STE in angular domain (one shaft revolution). ......................................................... 59
Figure 4-13. (a) Estimated compliance and (b) Estimated GTE considering higher-speed data
(including speeds of 2, 5, 10, 15 and 20 Hz). ............................................................................. 60
Figure 4-14. Comparison between GTE estimated from the proposed methodology and GTE
measured directly at low speed and low load, for all crack depths. ............................................ 61
Figure 5-1. Procedure for detection and severity assessment of crack. ...................................... 64
Figure 5-2. Theoretical tooth-meshing compliance (healthy) before filtering (top); after filtering
(bottom). ..................................................................................................................................... 65
Figure 5-3. Theoretical meshing compliance at different crack levels, before band-pass filtering.
.................................................................................................................................................... 67
Figure 5-4. Estimated compliance, with the same starting phase as theoretical compliance. ..... 68
Figure 5-5. (a) Estimated compliance vs healthy reference (GM-synchronous average); (b) error
between the two. ......................................................................................................................... 69
Figure 5-6. Comparison between the estimated compliance from the large-crack experimental
signals and the theoretical compliances simulated by the model for different crack sizes. The best-
matching simulated compliance will indicate the most likely crack depth. ................................ 70
Figure 5-7. Crack severity assessment for large crack (nominal depth = 50%). RMS of the
difference between estimated compliance and modelled results with different crack sizes. ...... 71
Figure 5-8. Crack severity assessment for large crack (nominal depth = 50%). RMS of the
difference between estimated compliance (considering higher-speed data) and modelled results
with different crack sizes. ........................................................................................................... 72
Figure 5-9. (a) Estimated compliance and (b) Estimated GTE for the medium crack. ............... 73
Figure 5-10. Crack severity assessment for medium crack (nominal depth = 36%). RMS of the
difference between estimated compliance and modelled results with different crack sizes. ...... 73
Figure 5-11. (a) Estimated compliance vs healthy reference (GM-synchronous average); (b) error
between the two for the small crack............................................................................................ 74

ix
Figure 5-12. Crack severity assessment for small crack (nominal depth = 19%). RMS of the
difference between estimated compliance and modelled results with different crack sizes, with a
zoom of the same plot in the red box. ......................................................................................... 75
Figure 5-13. Estimated compliance vs healthy reference, i.e., GM-synchronous average (top);
error between the two (bottom), for the medium crack. ............................................................. 77
Figure 5-14. Comparison between the estimated compliance from the medium-crack experimental
signals and the theoretical compliances simulated by the model for different crack sizes. The best-
matching simulated compliance will indicate the most likely crack depth. ................................ 78
Figure 5-15. (a) Estimated compliance and (b) Estimated GTE, for the small crack. ................ 78
Figure 5-16. Comparison between the estimated compliance from the small-crack experimental
signals and the theoretical compliances simulated by the model for different crack sizes. The best-
matching simulated compliance will indicate the most likely crack depth. ................................ 79
Figure 5-17. Synchronous average of vibration (acceleration) measurements with respect to input
shaft, for small, medium and large cracks (from top to bottom). Left column: measurements at 10
Hz and 20 Nm; right column: measurements at 20 Hz and 20 Nm. ........................................... 80
Figure 5-18. Relative contribution of Hertzian compliance towards the total tooth-pair compliance
(healthy teeth with parameters as per the previous sections of this chapter). ............................. 81
Figure 6-1. Synchronous averages of conventional (zero mean) TE for one input shaft revolution
on a spur gearbox; (a) mild wear, (b) severe wear. ..................................................................... 86
Figure 6-2. Illustration of wear depth as absolute geometric transmission error. ....................... 87
Figure 6-3. Graphical representation of hunting tooth period rephasing methodology. ............. 88
Figure 6-4. Processing method to calculate absolute TE. ........................................................... 91
Figure 6-5. Absolute TE (low speed, low load) at different wear stages with respect to healthy
gear condition.............................................................................................................................. 93
Figure 6-6. Final worn tooth images showing wear depths measured along line of action of (a)
driving gear with average wear depth of 544 𝜇𝜇𝜇𝜇; (b) driven gear with average wear depth of 218
𝜇𝜇𝜇𝜇. .............................................................................................................................................. 94
Figure 6-7. Comparison of combined average wear depth calculated using mean of absolute TE
from GTE, STE and DTE signals, against equivalent wear depth calculated from particle mass
measurements (top); Error of wear depth based on GTE, STE and DTE vs wear depth based on
particle mass (bottom). Note that a zero-error value at the beginning is only representing the fact
that Test Number 0 is used as reference. .................................................................................... 96
Figure 6-8. Comparison of wear depth calculated from tacho signals at different operating
conditions. ................................................................................................................................... 98

x
Figure 6-9. Changes in TE (GTE) components throughout wear evolution. An angle of 360°
represents a full gearmesh period in (b) and (e), a full pinion shaft rotation in (c) and (f), and a
full driven gear shaft rotation in (d) and (g).............................................................................. 100
Figure 6-10. Spectrum of vibration measurement at Test Number 4 (after approximately 12000
pinion cycles). ........................................................................................................................... 102
Figure 6-11. RMS of velocity measured using accelerometers at different locations throughout
wear evolution........................................................................................................................... 103
Figure 7-1. GM signals of GTE obtained before and after stops for inspections and/or surface
measurements............................................................................................................................ 106
Figure 7-2. GM signals of STE, obtained before and after stops for inspections and/or surface
measurements............................................................................................................................ 107
Figure 7-3. Evolution of GTE (left column) and STE (right column) gearmesh harmonics as wear
progresses.................................................................................................................................. 108
Figure 7-4. GM signals of GTE and STE from lubricated test (first half). ............................... 110
Figure 7-5. GM signals of GTE and STE from lubricated test (last half). ................................ 111
Figure 7-6. Evolution of GTE and STE GM harmonics as wear progresses for lubricated test.
.................................................................................................................................................. 112
Figure 7-7. Comparison of RMS values of GTE and STE with pitted area.............................. 113
Figure 7-8. Vibration signals of initial healthy and final wear stages at different speeds and loads.
.................................................................................................................................................. 115
Figure 7-9. Comparison of RMS values of vibration at 10 Hz and 20 Nm with pitted area..... 116
Figure 7-10. Comparison of RMS values of vibration at 20 Hz and 20 Nm with pitted area... 117

xi
List of Tables
Table 2-1. Summary of the effect of inputs and transfer function on the different forms of TE. 21
Table 3-1. Summary of gear parameters for the crack tests and to be used in the TVMS model.
.................................................................................................................................................... 39
Table 3-2. Summary of experimental details. ............................................................................. 41
Table 5-1. FRF identification based on different tests. ............................................................... 76
Table 5-2. Estimated crack sizes obtained excluding a single speed from the list of TE
experimental signals.................................................................................................................... 76
Table 5-3. Estimated crack sizes obtained including the 1 Hz TE signals.................................. 77
Table 5-4. Crack severity assessment using a TVMS model with 10% error. ............................ 83

xii
Table of Contents
Thesis Title and Abstract ............................................................................................................ i

Originality, Copyright and Authenticity Statements............................................................... ii

Inclusion of Publications Statement ......................................................................................... iii

Acknowledgements .................................................................................................................... iv

Abstract........................................................................................................................................ v

List of Publications .................................................................................................................... vi

Abbreviations ............................................................................................................................ vii

List of Figures........................................................................................................................... viii

List of Tables ............................................................................................................................. xii

Chapter 1 Introduction............................................................................................................... 1

1.1 Background......................................................................................................................... 1

1.2 Objectives of research ........................................................................................................ 3

1.3 Thesis structure ................................................................................................................... 4

Chapter 2 Literature review ...................................................................................................... 7

2.1 Gear dynamics and failure mechanisms ............................................................................. 7

2.1.1 Gear dynamics and models ......................................................................................... 8

2.1.2 Gear cracks .................................................................................................................. 9

2.1.3 Gear wear .................................................................................................................. 11

2.1.4 Overview of gear condition monitoring strategies .................................................... 13

2.2 Vibration- and AE-based gear condition monitoring ....................................................... 14

2.3 Encoder measurements and transmission error ................................................................ 17

2.3.1 Terminology and definitions ..................................................................................... 20

2.3.2 Historical developments of TE .................................................................................. 22

2.3.3 TE-based gear condition monitoring ......................................................................... 22

2.4 Summary and gaps............................................................................................................ 26

Chapter 3 Methodology and experimentation........................................................................ 29

xiii
 3.1 Overall methodology of thesis .......................................................................................... 29

3.2 Removal of transfer-path effects and separation of profile error and tooth flexibility
(Objective 1) ........................................................................................................................... 31

3.3 Assessing crack severity using TE (Objective 2) ............................................................. 33

3.4 Assess wear severity using TE (Objective 3) ................................................................... 34

3.5 Experimental facilities and data........................................................................................ 35

3.5.1 Test rig ...................................................................................................................... 35

3.5.2 Crack tests ................................................................................................................. 37

3.5.3 Dry wear test ............................................................................................................. 39

3.5.4 Lubricated wear test .................................................................................................. 40

3.5.5 Summary table of experiments .................................................................................. 41

Chapter 4 Removal of transfer-path effects in TE and separation of profile error and tooth
flexibility .................................................................................................................................... 42

4.1 Conceptual understanding of transmission error .............................................................. 42

4.2 Methodology..................................................................................................................... 47

4.2.1 Calculation and pre-processing of DTE signals ........................................................ 48

4.2.2 Removal of transfer function and regression to estimate GTE and compliance ....... 48

4.3 Results and Discussions.................................................................................................... 51

4.3.1 Detailed results on data up to 10 Hz.......................................................................... 51

4.3.2 Extension of the methodology to higher speeds ........................................................ 57

4.4 Additional results.............................................................................................................. 61

Chapter 5 Use of TE in gear crack severity assessment ........................................................ 62

5.1 Methodology..................................................................................................................... 62

5.1.1 Separation of cracked tooth from healthy teeth......................................................... 64

5.1.2 Model-based estimation of crack size ....................................................................... 66

5.2 Results and discussion ...................................................................................................... 68

5.2.1 Application to large crack (up to 10 Hz) ................................................................... 68

5.2.2 Application to higher speed data ............................................................................... 71

xiv
 5.2.3 Application to smaller cracks .................................................................................... 72

5.2.4 Further validation of the proposed methodology ...................................................... 76

5.3 Additional results and discussions .................................................................................... 77

Chapter 6 Absolute TE: a simple new tool for assessing gear wear ..................................... 84

6.1 Problem statement: mean component of TE overlooked .................................................. 85

6.1.1 Calculation of conventional (zero mean) TE ............................................................ 85

6.1.2 Limitations of conventional TE for wear analysis .................................................... 85

6.2 Proposed methodology: hunting-tooth-period rephasing for absolute transmission error 87

6.2.1 Conceptual methodology........................................................................................... 87

6.2.2 Summary of processing method for absolute transmission error .............................. 90

6.3 Application to gear wear assessment ................................................................................ 91

6.3.1 Quantification of absolute transmission error ........................................................... 91

6.3.2 GTE, STE, DTE and comparison with mass measurements ..................................... 95

6.3.3 Sensitivity to angular resolution: encoders vs tachos ................................................ 97

6.3.4 Analysis of the evolution of TE components throughout the wear process .............. 98

6.3.5 Discussion: important considerations of the proposed method ............................... 100

6.3.6 Comparison with vibration measurements .............................................................. 101

Chapter 7 Comparative study on the use of TE in wear monitoring: dry vs lubricated

conditions ................................................................................................................................. 104

7.1 Pre-processing of TE signals for all tests ....................................................................... 104

7.2 Dry test ........................................................................................................................... 105

7.3 Lubricated test ................................................................................................................ 109

7.3.1 Sensitivity of GTE and STE to gear wear ............................................................... 109

7.3.2 Comparison of TE with vibration............................................................................ 114

Chapter 8 Discussion .............................................................................................................. 118

8.1 Discussion on TE-based crack severity assessment ....................................................... 118

8.2 Discussion on TE-based wear severity assessment ........................................................ 120

Chapter 9 Conclusions and future work ............................................................................... 122

xv
 9.1 Summary of outcomes .................................................................................................... 122

9.2 Future work..................................................................................................................... 124

References ................................................................................................................................ 127

xvi
Chapter 1 Introduction

1.1 Background

Gears are widely used components in rotating machines, found in many industries that
require power transmission. However, more than many other mechanical components, they are
prone to failure which could lead to catastrophic losses, both in economic terms and safety. Gears
can fail due to a variety of failure mechanisms, mostly captured within the families of tooth cracks
and different types of wear. An efficient maintenance strategy of gear systems is therefore crucial
to ensure their safe operation, while balancing the direct cost of maintenance and downtime. This
can be optimally achieved by condition monitoring, the fundamental component enabling
condition-based maintenance (CBM) [1]. CBM relies on condition monitoring to provide
information on the health status of the gears. This allows informed and optimal decisions on when
it is necessary to perform maintenance, avoiding both costly failures and over-maintenance.

Condition monitoring can be further categorised into fault detection, diagnostics, severity
assessment and prognostics. Gear detection and diagnostics can be considered as well-established
in most cases, while severity assessment and prognostics are still fundamental research topics.
The more established measurements used for condition monitoring are oil/wear-debris analysis
[2] and vibration [3].

Wear-debris or lubricant analysis involves the collection of wear particles from the
lubricant of the gearbox. In the case of gear wear, the resulting particles released in the oil are
therefore captured by the diagnostic system. Despite the presence of on-line systems capable of

1
broad detection of mechanical issues in the drivetrain, most applications require time-consuming
manual analysis of the particles to provide a precise diagnostic conclusion. An automated process
for such analysis is still the object of fundamental research, would still rely on extensive datasets
of historical particles, and would only cover gear wear (not cracks).

Vibration has dominated the field of condition monitoring for decades, and relies on the
measurements of accelerometers, usually installed on the gearbox casing, to extract information
on the gear condition. The basic principle for vibration-based condition monitoring is the fact that
both wear and cracks introduce time-varying internal forces at the gear-meshing point, which in
turn generate vibrations in the whole gearbox. The main advantages of vibration sensors are the
ease of installation, the possibility of on-line monitoring and the fact that they are already used
for the detection of faults in other components such as shafts and bearings, thus offering synergies
with other applications. The main drawbacks of vibrations are the sensitivity to sources unrelated
to gears and the physical distance between the meshing point and the sensor. This distance results
in a complex transfer path, which renders the measured vibration signals significantly different
from the original time-varying gear-meshing forces where the symptoms of gear faults manifest.
Moreover, the effect of this transfer path on the vibration signal is dependent on operating
conditions, and the transfer path includes an overall unknown scaling factor between force and
vibration. Therefore, while vibration has been proven highly effective in terms of early detection
and diagnostics, its use in severity assessment and prognostics is challenging, unless large datasets
for training data-driven models or complex gear dynamic models are available.

More recent and mostly research-oriented works have also proposed alternative
measurements with potential for this application, including instantaneous angular speed [4],
acoustic emission [5], and transmission error [6].

Transmission error (TE) is a parameter that is highly relevant to the kinematics of gear
meshing. In the modelling of gearboxes [7], it is considered as the “source” of internal excitation
that leads to the generation of vibration. TE, as suggested literally by its name, is the kinematic
error in the transmission of motion between the two gears, with respect to the ideal case of perfect
and rigid involute-profile teeth. Such deviations are mainly due to profile errors and tooth
deflections, and can be amplified by the system dynamics. Typical gear faults directly generate
TE components: for instance, wear affects the tooth profiles, while cracks reduce tooth stiffness
and thus increase the deflection of the damaged teeth under load. The source-separation and
transfer path problems affecting vibration are therefore highly reduced with TE, making it a
perfect candidate for gear condition monitoring. Although requiring a more sophisticated
measurement apparatus, TE-based condition monitoring is becoming increasingly feasible, thanks

2
to the availability of accurate shaft encoders, and the spread of the Industry 4.0 concept of
integrated sensors.

However, the work on TE-based techniques is still much limited compared to vibration.

The main unresolved questions regarding TE are the following:

Gap 1. From TE to tooth profile and deflection. Despite being much closer to the gear, TE is
still potentially subject to distortions from the transfer path between the gear-meshing point and
the encoders. This transfer path does not include the complex dynamics of the casing, but it is
potentially affected by the torsion and bending of the shafts, especially when measured at higher
speeds. Moreover, even neglecting those effects, TE shows combined effects of both tooth profile
and deflections, each carrying different diagnostic information (mostly about wear and cracks,
respectively). No work available in the literature (see Chapter 2 for details) has so far shown if it
is possible to remove transfer path effects and identify tooth profile and deflections separately.

Gap 2. Crack severity assessment. While there are well-established techniques to detect and
diagnose cracks using different sensor technologies, the work on severity assessment of cracks is
still largely incomplete. Although some techniques in the literature demonstrated the ability to
assess crack severity, they require either a large amount of data to train data-driven algorithms or
a sophisticated dynamic model. No signal-based approach has been developed to effectively
tackle crack severity problems (see Chapter 2 for details).

Gap 3. Wear severity assessment. Wear depth is the most crucial information that describes the
severity of worn teeth in gears, but it is currently not entirely captured by available TE
measurements. In particular, only the variations of wear depth along the flank are currently
incorporated into TE signals, while the average wear depth is discarded. This is because the
current methodology extracting TE from encoder measurements does not include a DC value, i.e.,
in the absence of a reference, the average of the TE signal is arbitrarily set to zero. Despite the
key value of the average wear-depth information, no work available in the literature (see Chapter
2) has so far even discussed this clear drawback, and a solution has not been proposed.

1.2 Objectives of research

The previous section briefly highlighted the remaining obstacles to the implementation
of TE-based gear condition monitoring. Therefore, the overall aim of this research is to develop
innovative solutions to enable TE-based gear condition monitoring. In order to achieve this,
three main research objectives are listed as follows, matching the three gaps identified in the
previous subsection.

3
 Objective 1: Develop a technique to obtain estimates of profile and tooth deflections
based on TE measurements. It is necessary to address the two challenges highlighted in Gap 1
before TE can be reliably used to study gear faults. The expected outcome is the ability to obtain
information of tooth flexibility (for cracks) and profile error (for wear).

Objective 2: Develop a reliable approach to assess tooth crack severity based on TE.
With the understanding that a crack reduces the stiffness of a tooth, it is possible to separate a
cracked tooth from other healthy teeth. An algorithm that performs this identification
automatically is to be developed. The flexibility of the cracked tooth must then be assessed and
compared to theoretical tooth-compliance models to identify the crack size. The expected
outcome is the ability to quantify the severity of the crack, thus providing quantitative information
for risk assessment and maintenance planning.

Objective 3: Develop a reliable approach to assess wear depth based on TE. This
objective aims to develop a methodology that is capable of retaining information on the average-
wear depth in TE signals. The expected outcome is the ability to quantify overall wear depth in
gears directly using TE signals. Since this is physically impossible with vibration, a successful
achievement of this objective would form a significant advance in gear wear monitoring, and a
valuable tool for risk assessment and maintenance planning.

1.3 Thesis structure

This section highlights the structure of the thesis, with a brief introduction to the contents
of each chapter. There are a total of nine chapters in this thesis, three of which are adapted from
publications. Chapters 4 to 5 consist of material from a published journal paper titled “Use of
transmission error for a quantitative estimation of root-crack severity in gears” [8]. The contents
of Chapter 6 are instead adapted from a published journal paper titled “Absolute transmission
error: a simple new tool for assessing gear wear” [9]. The brief discussion and organisation of
each of the following eight chapters is presented below.

Chapter 2 reviews in detail the existing literature relevant to the field, further justifying
the formulation of the research objectives and the need for this TE study. The literature review
begins with the fundamentals of gear dynamics and the common gear failure mechanisms,
followed by an overview of gear condition monitoring strategies. The next sections provide a
survey on gear condition monitoring for specific sensors, including the most popular, vibration,
acoustic emission, instantaneous angular speed and of course transmission error. An in-depth
discussion of transmission error including its origin and historical developments is provided since

4
this measurement type forms the focus of the thesis. Finally, a summary detailing the research
gaps and linking them to the literature is included.

Chapter 3 lays out the methodology of this thesis, designed to address the research
objectives described above. The chapter first outlines the overall methodological framework of
the thesis, followed by three sections that briefly describe the key research tasks. The three tasks
correspond to the three objectives mentioned in Chapter 1, and the execution and results of each
are reported in Chapters 4 to 6. The last section of chapter 3 provides the description of the
experimental facilities and tests that have supported the innovations developed in this thesis.

Chapter 4 details the innovative development of a technique that first removes transfer-
path effects in TE measurements, and then separates the tooth profile component from the tooth
deflection component. The technique employs measurements obtained at multiple speeds and
loads. The different speeds aid in the estimation and removal of the transfer path, while the
different loads allow separating load-dependent deflections from load-independent profile
components. The former can be utilised for the study of crack severity, while the latter are useful
for wear monitoring.

Chapter 5 proposes a new technique for the severity assessment of gear cracks using the
deflections extracted from TE in Chapter 4. The automated procedure first detects the cracked
tooth in the signal to separate it from the healthy teeth. Severity assessment is then achieved by a
model-based approach that compares the experimental meshing-compliance signal to an
equivalent signal derived from an existing theoretical time-varying mesh stiffness model.

Chapter 6 reports the development of the innovative “absolute transmission error”. The
main innovation is the rephasing of TE measurements based on a reference TE signal obtained
from unworn gears. This allows measuring the DC shift of TE signals due to wear, thus finally
providing an accurate estimate of average wear depth. For the first time, and without any
modelling, TE-based direct measurement of wear depth in gears is achieved.

Chapter 7 consists of a further observational study on the monitoring of two different

types of wear using TE signals. In particular, the chapter compares the results obtained from dry
and lubricated wear tests, the first mainly leading to abrasive wear and the latter to pitting. The
chapter mainly analyses the evolution of TE harmonics due to wear and highlights the observed
differences in the two wear mechanisms.

Chapter 8 discusses the overall achievements and limitations of TE-based gear condition
monitoring in this thesis. The feasibility of the use of TE in industrial settings is also mentioned,

5
while highlighting its advantages and constraints when compared to the currently dominant
vibration-based techniques.

Chapter 9 summarises the key conclusions of this research project, providing an account
of the contribution towards the field of gear condition monitoring and recommendations for future
work on TE.

6
Chapter 2 Literature review

This chapter provides a review of the existing literature to identify the gaps that form the
basis of this research project. Section 2.1 provides an overview of gear dynamics, the most
common gear failure mechanisms, and the fundamentals of gear condition monitoring. The next
sections then provide a more in-depth survey on gear condition monitoring based on specific
sensor technologies: the most popular vibration, including extensions to AE in section 2.2, as well
as instantaneous angular speed and finally transmission error in section 2.3. Given that the last is
the main focus of this thesis, section 2.3 will also introduce the fundamental physical concepts
behind transmission error, including its origin, historical developments, and its relationship to
common gear faults. Finally, a summary of the literature review and the research gaps are
mentioned in section 2.4. Some of the material presented in this chapter consists of adapted
extracts from the papers 2 titled “Use of transmission error for a quantitative estimation of root-
crack severity in gears” [8], and “Absolute transmission error: a simple new tool for assessing
gear wear” [9].

2.1 Gear dynamics and failure mechanisms

In order to provide a strong physical background to the gear condition monitoring

developments of this thesis, it is essential to introduce the fundamentals of gear dynamics and the

2
Permission has been granted by co-authors.

7
most common gear failure mechanisms. To address the first, subsection 2.1.1 provides a brief
overview of the extensive gear modelling literature. Regarding the second, failure of gears can
occur due to various reasons such as operation beyond design limits, end of service life,
inappropriate assembly, misalignment, manufacturing imperfections, intrusive particles etc [10].
However, most of these root-causes trigger two common failure mechanisms: tooth root-cracks
and various types of wear, which will be discussed in sections 2.1.2 and 2.1.3 respectively. Finally,
subsection 2.1.4 provides the fundamental concepts behind their diagnostics and prognostics, thus
introducing the more detailed condition-monitoring sections which follow.

2.1.1 Gear dynamics and models

Conceptually, gear models are built to study and simulate the dynamic response of gear
systems to pre-defined inputs, which usually consist of the operating conditions (speeds and
torques), and tooth-meshing parameters representing the health-state of the gears.

The key input parameters affected by faults usually are the time-varying mesh stiffness
and tooth profile error [11]. Time-varying mesh stiffness (TVMS) is an important component of
the gear model, because this is one of the main sources of internal excitation that lead to vibration
of gearboxes, even in healthy conditions [12]. It is time-varying due to the change in number
(single- vs double-pair contact) and position (root vs tip) of the contact points throughout the
meshing cycle. The mesh stiffness of a tooth pair is therefore usually modelled as a periodically
variable-stiffness spring between the two gears. In healthy ideal gears, the TVMS has a
fundamental frequency equal to the gearmesh frequency

𝑓𝑓𝐺𝐺𝐺𝐺 = 𝑍𝑍1 𝑓𝑓1 = 𝑍𝑍2 𝑓𝑓2 (2-1)

where 𝑍𝑍1 and 𝑍𝑍2 represent the number of teeth of the two gears and 𝑓𝑓1 and 𝑓𝑓2 the angular speed
of their shafts.

It is sometimes convenient to consider the reciprocal of TVMS in certain applications,

which can be called mesh compliance. Two main families of approaches are commonly used in
literature to model the TVMS: analytical potential energy (PE) methods [13]–[15] and finite
element (FE) methods [16], [17]. FE methods are usually considered the most accurate, but
require a careful selection of meshing parameters and are generally considered more
labour-intensive. PE methods are based on approximated analytical expressions, and are therefore
simpler to implement and allow a more direct interpretation of the key physical phenomena
resulting in the TVMS. In the simplest PE methods, the gear tooth is considered as a cantilevered
beam with varying cross section, and the TVMS is obtained using Timoshenko’s beam theory [18].

8
In more sophisticated models, the TVMS also considers tooth axial compression, Hertzian contact
(often linearised), and fillet-foundation effects.

Profile errors are present even in healthy gears due to manufacturing tolerances,
misalignments and even design choices such as tip-relief. Mark [19] provided the most established
model for the calculation of the combined profile error input, accounting for the geometric
alteration of all the teeth flanks in mesh at the same time. In the proposed approach, geometric
errors of two flanks meshing with each other are simply added, while the geometric error of two
tooth-pairs in contact at the same time are combined with an average, weighted by the
corresponding tooth-pair stiffness. Assuming that most profile errors due to manufacturing are
similar across all teeth, healthy gears are expected to show fairly gearmesh-periodic profile errors.

Model outputs correspond to displacements and rotations of the gearbox elements, and
often include virtual sensors installed on the gearbox. These can be in the form of lateral
displacement (typically at the position of the bearings in the model), torsional displacement of a
shaft (which can be differentiated to obtain instantaneous angular speed or IAS), or even
transmission error by considering the torsional displacements of two shafts.

The conversion of the inputs into outputs is performed by dynamic models based on
fundamental mechanics. Two main families of models are found, both simplifying the continuous
system to a finite number of degrees of freedom: lumped parameter models and FE models. A
lumped parameter model is obtained by treating the different elements (e.g., shafts, couplings,
bearings) as rigid bodies, springs and dampers [20], [21]. These models can be as simple as a
single-DOF system [22], or as complicated as a 26-DOF [23], depending on the assumptions and
simplifications for other gearbox elements such as the consideration of the flexibility of shafts
and bearings. On the other hand, FE models usually consider a larger number of degrees of
freedom by decomposing each element into smaller portions [24], [25]. Both approaches result in
MDOF models whose transfer path significantly distorts the inputs of TVMS and profile error
(the main excitation components). Consequently, the resulting dynamic responses still carry
information on the inputs (and therefore faults) but their relationship with stiffness and profile
changes is not obvious and biased by operating conditions as the inputs pass through the transfer
path. In the next chapters of this thesis the effect of this transfer path on actual measurements will
be investigated and techniques will be proposed to revert to estimates of the original inputs.

2.1.2 Gear cracks

Gear cracks are usually caused by bending fatigue of a gear tooth, which can eventually
lead to the complete loss of a tooth if undetected. Bending fatigue occurs in gears when teeth are

9
repeatedly subject to bending stress exceeding the local fatigue strength. This phenomenon is
mostly affecting the tooth root on the meshing flank [26], and can be exacerbated by overloading,
inappropriate assembly, misalignment, poor design and stress raisers due to unexpected
indentations.

The crack initiates at the high stress point at the root and propagates towards a zero-stress
point near the tooth centreline below the root circle [27]. As the crack propagates beyond the
tooth centreline, the zero-stress point shifts transversely until it reaches the root of the other side
of the tooth. Fatigue is categorised into high-cycle fatigue (HCF) and low-cycle fatigue (LCF).
HCF occurs at low stress and the deformation is elastic, leading to operating cycles of more than
104. LCF, on the other hand, has lower operating cycles and the deformation is primarily plastic.
Both HCF and LCF leads to crack growth.

In terms of fatigue, crack growth consists of three progressive stages – crack initiation,
crack propagation and fracture, often referred to as Stage 1 to 3 respectively [28]. Stage 1 occurs
when local plastic deformation causing cracks to nucleate at a microscopic level of the material,
even though the material has not yielded, i.e., the maximum bending stress is less than material
yield strength. This often happens due to structural discontinuities of the material such as grain
boundaries. Most of the cycles of the fatigue life of a component are spent in Stage 1 to initiate
the crack. Stage 2 is crack propagation, where the crack grows in the direction normal to the
maximum tensile stress. This stage is characterised by a steady growth of the crack, which can be
governed by established crack growth equations such as Paris’ law [29], applied in crack-growth
studies by most researchers of the field. Finally, Stage 3 is where fracture occurs, and it could be
ductile, brittle or mixed mode depending on the stress level and material toughness.

The work on gear crack propagation has been studied extensively by Lewicki et al. [30]–
[34]. It is worth mentioning that research in the conditioning monitoring of gearboxes has been
extensively conducted for decades but most of the experiments are carried out using seeded faults
instead of naturally generated cracks. In [35], [36], a project was conducted to seed a gear tooth
root crack that closely resembles a natural crack instead of one that was seeded with a cutting tool.
This was achieved by using a tooth-bending rig that could induce bending failure on the tooth.

It is interesting to note that backup ratio, which is the ratio of the rim thickness to tooth
height, affects the propagation path of a gear crack, as studied in [30]. Lewicki [30] studied the
effect of rim thickness on the propagation path of the gear crack through simulation using FRANC
2D and experimental studies. It was found that if the backup ratio is sufficiently small, the crack
could propagate along the rim. Also demonstrated in [37], the crack propagation path of a spur
gear is dependent on its backup ratio and the initial point of the crack.

10
 In general, cracks result in two main effects: a reduction of the load-bearing section and
plastic deformations leading to cracks opening even under no load. These result in two main
changes in the inputs discussed in section 2.1.1: (i) a localised reduction in the portion of TVMS
corresponding to the cracked tooth in mesh, and (ii) a localised tooth profile error due to plastic
deformations. The first has always been considered as the dominant phenomenon, and second
remains to be studied from a theoretical point of view, as it is difficult to predict. Both effects
result in faulty-gear-shaft periodic components in the TVMS and profile.

In [38], Mark et al. developed a transmission-error model that considers the effect of
linear-profile plastic deformation on a single tooth. On the other hand, a series of established and
reliable TVMS models are available in the literature for the simulation of the effects of cracks in
gears.

The TVMS with a cracked tooth is usually calculated based on a pre-defined crack
propagation path: usually straight lines [39], and sometimes parabolic curves [15]. In the case of
a cracked tooth, the bending and shear stiffnesses are most significantly affected. Changes in
Hertzian and axial stiffness are negligible because the strength of the tooth in the direction of the
tooth width remains the same. There are also more sophisticated TVMS models based on PE
method. Chen et al. evaluated the TVMS of a spur gear pair with tooth profile modification and
root crack as well as establishing the effects of the mentioned tooth errors and applied torque on
TVMS, load sharing and static transmission error [40]. Ma et al. developed an improved TVMS
model for cracked spur gears [41] by considering the effects of reduction of fillet-foundation
stiffness and extended tooth contact of cracked gears. FE modelling of cracks is also studied by
some researchers. For instance, in [16], a 2D FE model considering linear elastic fracture
mechanics is used to study the crack propagation path for gears of different contact ratio.
Calculation of TVMS using 3D FE model is also possible, such as the one established in [14] to
validate the PE method.

Later in this thesis, one such TVMS PE model will be used as a reference for crack
severity estimation (see Chapter 5).

2.1.3 Gear wear

Other than bending fatigue, gears can suffer from wear, leading to the removal of material
from the tooth surfaces. The rate of wear is highly associated with the operating conditions,
including load, temperature and lubricant. The wear process is often divided into three stages,
each with different wear rate at nominal operating conditions. The first stage is running-in, where
high spots or asperities on the surface are smoothened so that the two mating surfaces adapt to

11
each other, with a decreasing wear rate as this stage progresses. The second stage is where steady
wear occurs due to constant tribological characteristics. The third stage has a rapidly increasing
wear rate as the gears are approaching the end of their service life. The common types of wear
include abrasive wear and pitting [42], which are also the focus of this research project (see
Chapters 6 and 7).

Gear kinematics is characterised by the combination of rolling and sliding contacts

between the teeth of the two gears in mesh [43], [44]. Pure rolling occurs at the pitch line, whereas
sliding takes place at both the addendum and dedendum, with maximum sliding at the tip and the
root of the tooth during the start or end of engagement. Since abrasive wear occurs primarily due
to the relative sliding of the mating surfaces [45], the tip and the root of the teeth are the most
affected by this type of wear. Abrasive wear is often significant when there is insufficient or
contaminated lubricant, whose role is to provide an adequate oil film separating the mating
surfaces.

Pitting, also known as contact fatigue or surface fatigue, occurs when the Hertzian contact
stresses are higher than the material fatigue limit, gradually forming cavities on the tooth surfaces
as micro-cracking occurs due to the high stresses [46]. This failure mechanism is often observed
at the pitch line, where there is single-tooth-pair contact and load sharing across the surface is
limited due to low oil film thickness. Pitting can be categorised into micro-pitting and macro-
pitting (sometimes also referred to as spalling) based on the average size of the pits. Pitting can
lead to destructive incidents, as shown in the case study of Ref. [47], where a helical gear used in
the gearbox of a bus failed due to misalignment, causing destructive pitting and spalling, which
ultimately led to breakage of several teeth.

In general, it is understood that wear introduces geometric modifications to the tooth since
it changes the tooth profile, and also potentially reduces the tooth stiffness due to thinning of the
gear teeth (only in the most severe cases). Under the assumption that wear would be fairly evenly
distributed across the different teeth, it is expected to have a main effect on the gearmesh-periodic
components of TVMS and profile.

The most applied model by researchers for wear is Archard wear equation [48], which is
based on the assumption that the volume of material removal is proportional to the work done by
friction forces. In [49], the authors adopted the generalised Archard wear equation to develop a
wear prediction numerical model. While most studies focused on the effect of cracks on TVMS,
there are also some studies considering the effect of wear. In [50], TVMS is modelled based on
PE method with the consideration of a worn tooth profile.

12
2.1.4 Overview of gear condition monitoring strategies
Condition monitoring of gears can be categorised in four stages, which are detection,
diagnostics, severity assessment and prognostics. Detection is primarily identifying and detecting
the presence of a fault, while diagnostics involves determining the type and the location of the
fault. After the nature of the fault is identified, severity assessment is then performed to examine
how severe the fault is at the current stage. The last stage of condition monitoring is prognostics,
which involves the prediction of future fault severity and the remaining useful life (RUL) of gears,
based on industry standards or pre-defined safety criteria. It is important to realise that prognostics
is highly dependent on accurate severity assessment. Effective prognostics is only achievable with
a reliable severity-assessment technique or indicator, unless there is a large number of data
available for data-driven methods.

For the condition monitoring of gears, detection and diagnostics are considered well-
established, at least for the common faults such as cracks, wear and spalls. However, severity
assessment and prognostics remain as fundamental research topics which require further
investigation. While the early detection of faults remains an important area in condition
monitoring, severity assessment is at least as important to balance cost and the risk of failure, in
particular in the later stages of fault development. In many applications, it is desired to run a
machine even after an early fault has been detected, to maximise the possible RUL of the faulty
component, so as to minimise cost. This is why severity assessment is also crucial to ensure
minimal costs, while not compromising the risk management of possible failures.

Various types of measurement technologies can be used for gear condition monitoring,
with two main approaches possible.

The first is oil/wear-debris analysis [2]. In the case of gear wear, particles are deposited
from the tooth surface into the lubricant of the gearbox. Wear-debris or lubricant analysis is
performed to collect these particles that contain hidden information regarding the nature of the
wear mechanism [51]. Although some on-line systems are available for detection, most
applications of this measurement require labour-intensive and time-consuming manual analysis
of the wear particles to reveal the diagnostic information. Wear-debris analysis is also limited to
only wear but not applicable to cracks, since a crack does not release significant particles into the
lubricant. For these reasons this approach will not be discussed further in this thesis.

The other family of approaches instead relies on signals captured by sensors permanently
installed on the machine. These signals are acquired with analogue-to-digital converters and
undergo a series of signal-processing steps to extract clear diagnostic information and provide

13
continuous condition monitoring. The most popular sensor technology is vibration [3], usually
based on measurements by accelerometers installed on the gearbox casing in proximity of the
bearing housings. Other sensor technologies which have gained popularity (albeit mostly in
research application so far) are acoustic emission sensors, which will be covered together with
vibration in section 2.2, and encoder measurements. The latter measure angular displacements of
the gear shafts and can be used independently to obtain instantaneous angular speed, or jointly to
compute transmission error. Both will be discussed in section 2.3.

2.2 Vibration- and AE-based gear condition monitoring

Vibration has undoubtedly dominated the field of condition monitoring of gears, with the
vast majority of traditional condition monitoring approaches using accelerometers to measure
vibration on the gearbox casing [3],[52]–[58]. Given the importance of this measurement
technology, this chapter will provide a short review of some key aspects of vibration-based gear
condition monitoring, despite the focus of this thesis on the alternative technology of TE. An
exhaustive survey of the body of works regarding vibration is out of the scope of this study and
would not contribute to clarity and conciseness. On the contrary, this section will concentrate on
the characteristics that make vibration-based condition monitoring so popular, and on those
limitations that are still not resolved after decades of research.

Under sufficient speed and load, fault-induced TVMS and profile changes result in
specific forces at the meshing point. Such forces generate torsional and radial motion of the shafts,
which in turn compress the bearing rollers, exciting the gearbox casing and causing vibration on
its surface. Therefore, the periodic components of TVMS and profile due to cracks and wear are
transferred to accelerometer measurements. Vibration-based condition monitoring is therefore
often focussed on the enhancement and identification of such components in real-case situations.

First of all, the periodicity of gear-vibration signals is compromised in practice by

inevitable small speed fluctuations. In particular, signal features are phase-locked to the shaft
angle rather than being periodic in the time domain. Tachos and encoders have been used in
conjunction with accelerometers for decades in order to address this issue and perform a time-
angle domain transformation usually referred to as ‘order-tracking’ [59]. The traditional method
to perform order tracking is known as computed order tracking, which resamples the signals in
constant increments of shaft rotation angle (rather than time). To achieve that, the construction of
a phase-time map is firstly obtained based on zero-crossing of a reference tacho or encoder signal,
and then an interpolation is performed. A cubic spline interpolation has been shown to produce
the best results in the generation of both phase-time maps and resampled signals [59]. Another

14
way to obtain the phase-time relationship compared to the zero-crossing approach is using phase
demodulation of the main encoder signal harmonics to produce a continuous relationship between
angle and time [60]. It was also shown that order tracking can be performed without a reference
tacho/encoder signal [60], using the acceleration signal itself, band-pass filtered around a specific
gearmesh harmonic. A demodulation band is required and the method to choose the demodulation
band was mentioned in [61]. Recently, modified order tracking algorithms such as Vold-Kalman
Filter Order Tracking [62], [63] and Order Tracking Transforms [64] have been introduced. The
former can extract smoother order waveform compared to traditional computed order tracking
while the latter combines order tracking and Fourier Transform into a single step.

In the study of gear signals, synchronous averaging (SA), often follows the order-tracking
step. SA is a well-known technique implemented to further enhance the periodic vibration
components symptomatic of gear condition, and attenuate noise and signals not synchronous with
the shaft under analysis [65]. Since gear signals are angular-periodic, this is a traditional way to
separate a gear signal from signals of other gears or components of the gearbox. To obtain the
signals for different gears of the same gearbox, SA can be performed on the signal separately with
different synchronous periods. SA is performed by taking the average of the signal segments
where each of them corresponds to one period of a synchronising signal. In the frequency domain
(or order domain), this will give a comb filter characteristic with the transfer function which
effectively passes the harmonics of the synchronising signal while rejecting other frequency
components [66]. It is suggested to perform order tracking prior to SA as this could reduce the
small speed fluctuations that cause the signal to be not periodic. This can also produce integer
number of samples per revolution and a defined initial point for the phase reference [67].

A common problem with acceleration measurements is the strong effect of the transfer
path between the fault-symptomatic forces and the measurement point. This transfer function is
complex, as it includes modes of the internal gearbox elements, the casing and the support
structure where the gearbox is installed. Due to these problems, it is difficult to identify and
remove the transfer function and convert the acceleration measurement into an estimate of tooth
deflection or profile error [68]. In addition to this, the overall scaling of this transfer function is
usually unknown, and it highly depends on the gearbox installation layout. This is the main reason
for the absence of a reliable and repeatable technique for crack severity assessment based on
accelerometer measurements.

It is worth mentioning one approach to measurements with accelerometers on the casing,

which in limited circumstances removes the dynamic effects of the transfer function. It is the ALR
(average log ratio) method developed by Mark et al. [69]–[71], of which a particularly interesting

15
example is presented in [70]. Assuming that the transfer path is approximately unaffected by gear
faults, it takes the log ratio of (growing) harmonics of the gear shaft from the point where a change
is starting to occur (as reference). Measurements are always taken at the same speed, so that the
harmonics are affected in the same way by the transfer path. The harmonics taken depend on the
type of fault expected, but always exclude the gearmesh harmonics and adjacent sidebands and a
small number of low harmonics. For the case of a growing tooth root crack, as in [70], the
concentration is on a band of harmonics centred on the midpoint between the gearmesh harmonics
(including zero) as this is where local faults dominate. In this study, it is shown that by taking
measurements at different loads it is also possible to differentiate between the load-dependent and
non-load-dependent components, thus separating profile and TVMS effects. However, the
changes are only in terms of changes in harmonic orders, and cannot be expressed as waveforms,
or calibrated in absolute terms.

The same influence of transfer paths that hinders accurate severity assessment are used
as an advantage for accelerometer-based early fault detection. Resonant responses to the high-
frequency events caused by gear faults in fact increase the sensitivity of vibration to incipient gear
faults. In [72], Wang successfully applied resonance demodulation to detect an incipient tooth
crack, which caused very limited changes in the tooth stiffness, but resulted in impacts exciting
high-frequency resonances. After order-tracking and SA, the technique removed the gearmesh
harmonics to obtain a “residual signal” 3 that was then bandpass-filtered around a resonance, and
demodulated via envelope to obtain a clear one-per-rev event representative of the fault. Another
recent development by Yang et al. [73] focused on the same phenomenology, but deconvolved
rather than demodulating the “residual signal” to obtain an estimate of the cracked-induced
impulsive event itself. This was done by fitting a transfer function to the vibration signal using
the matrix pencil method. Metrics from the resulting signals were then fed to a linear-
discriminant-analysis classifier, which was previously trained using a portion of the same dataset.
This served as a proof of concept for a methodology for severity assessment using vibration
signals.

Summarising, given their strong sensitivity to impulsive and abrupt events,

accelerometers are remarkable in early fault detection, but their application in severity assessment
is strongly limited by the complexity of the transfer path and the unknown overall scale factor
between measurement and input (TVMS and profile). Two potential avenues to enable vibration-
based fault severity assessment were proposed: one (e.g., Mark’s ALR) uses a ratio of spectra

3
It is noted that in the field of gear diagnostics, often the residual signal is not the random part obtained by
removing the SA from the signal, but rather the SA component without the gearmesh harmonics.

16
obtained in faulty and healthy conditions at the same speed and load, the other relies on data-
driven methods (e.g., Yang et al.) to classify signal features against a reference database. The
main drawback of both strategies is that they generally require large amounts of historical data
[74], [75], comprising a significant number of observational faults on a specific machine with a
fixed transfer function.

Signal-processing tools similar to vibration are commonly applied to acoustic emission

(AE) signals, which are obtained with piezoelectric sensors similar to accelerometers. They are
more commonly used in civil applications [76], but are also applicable to gearboxes, as shown in
[5], [77]–[79]. Tan et al. showed that the source of AE in gears comes from the asperity contact
due to rolling and sliding between the gears in mesh [77]. A comparative experimental study has
been conducted on the diagnostic and prognostic capabilities of acoustic emission, vibration and
spectrometric oil analysis for spur gears [78]. Feng et al. proposed a generalised Gaussian
cyclostationary model and established a correlation between gear wear and statistical properties
of AE signals, which can be applied to monitor surface degradation due to wear [5]. In [79], the
authors extracted the theoretical relationship between surface roughness and AE generation
mechanisms for the first time in mechanical systems, and estimated surface roughness using a
model-based technique. Overall, AE signals are mainly studied for their sensitivity to early crack
propagation and micro-scale surface conditions. As such they are only effective for early fault
detection and trending, but probably not particularly promising for accurate fault-severity
estimation.

2.3 Encoder measurements and transmission error

Tachos and encoders also represent an alternative to accelerometers through their direct
use as torsional vibration sensors. The use of encoders for the detection of phase-modulation
effects induced by cracked teeth has been proposed already, with the use of instantaneous angular
speed (IAS) methods, such as the one shown in [4], [80]–[82]. Despite having a less complex
transfer function compared to vibration, the scaling issue is still unresolved using a single encoder
and IAS, and no IAS-based method has been able to reliably quantify crack depth.

An alternative to IAS is obtained by combining two encoder measurements from input

and output shaft to estimate gear transmission error. Transmission error (TE) is the difference
between the angular displacement of gears that are in mesh, once the gear ratio is taken into
account. Strictly speaking, TE is formally defined as the deviation of the angular displacement
𝜃𝜃2 (𝑡𝑡) of the driven gear with respect to that of the driving gear 𝜃𝜃1 (𝑡𝑡), with 𝜃𝜃2 (𝑡𝑡) properly scaled
by the gear ratio [83], such that any material removal would result in a negative value of TE [84]:

17
 𝑅𝑅𝑏𝑏1
TE(𝑡𝑡) = 𝜃𝜃2 (𝑡𝑡) − 𝜃𝜃1 (𝑡𝑡) (2-2)
𝑅𝑅𝑏𝑏2
where 𝑅𝑅𝑏𝑏1 and 𝑅𝑅𝑏𝑏2 are the base radii of the input and output gears, respectively.

However, TE is often found in literature with different units and sign conventions,
depending on the application and choice of the authors. In terms of units, TE is often expressed
directly in microns, by multiplying the above equation by 𝑅𝑅𝑏𝑏2 , and often with an opposite sign
convention, so that material removal and deflections result in a positive TE:

TE(𝑡𝑡) = ±[𝜃𝜃1 (𝑡𝑡)𝑅𝑅𝑏𝑏1 − 𝜃𝜃2 (𝑡𝑡)𝑅𝑅𝑏𝑏2 ] (2-3)

Under the assumption of ideal and healthy gears, TE is constant and equal to zero, since
the fundamental theorem of gearing dictates that the products of angular displacement and base
radius on the two shafts must be identical (𝜃𝜃1 𝑅𝑅𝑏𝑏1 = 𝜃𝜃2 𝑅𝑅𝑏𝑏2) [44]. This behaviour is however never
observed in practice, since it requires the gears to be perfectly rigid and have perfect involute-
profile geometry. In reality, this is impossible to achieve due to tooth deflection under load (see
top diagram in Figure 2-1) and inevitable geometrical imperfections (see bottom diagram in
Figure 2-1) such as manufacturing errors, contributing to fluctuations in the output speed even
when input speed is constant. Thus, TE indicates any deviation from the ideal gear meshing
behaviour.

18
Figure 2-1. Deflected teeth due to load vs ideal rigid gear (top); teeth with geometrical errors vs
ideal involute gear (bottom). The scale of the effects is exaggerated for clarity.

Gear faults such as wear (see left diagram in Figure 2-2) and cracks (see right diagram in
Figure 2-2) would also affect the profile and stiffness of gear teeth, introducing alternations to TE.
A thorough understanding of TE therefore forms a crucial basis for this research project. This
section provides the terminology and definitions of TE, as well as its origin and historical
developments.

Figure 2-2. Worn tooth vs healthy gear (left); cracked tooth vs healthy gear (right). The scale of
the effects is exaggerated for clarity.

19
2.3.1 Terminology and definitions
When analysing TE, it is common to follow a three-stage decomposition [3].

The simplest measurement of TE is obtained at very low load and very low speed, at
which teeth can be considered rigid and the behaviour of the system quasi-static. This TE
measurement is called Geometric Transmission Error (GTE), because it is directly related to the
geometric profile of the teeth, i.e., it is zero only if the teeth are perfect involutes. Therefore, the
profile error discussed in the previous sections directly corresponds to GTE, showing a first
obvious advantage of this measurement approach.

At non-negligible loads, but still very low speed, tooth deflections add to GTE giving the
so-called Static Transmission Error (STE). STE can in fact be considered as the additive
combination of the two inputs seen in the previous section: profile errors (GTE) and TVMS, albeit
here present as compliance 𝑐𝑐(𝑡𝑡) = 1/TVMS multiplied by total contact force 𝐹𝐹(𝑡𝑡), considered as
the sum of all contact forces at a single instant (one or two depending on the number of tooth-
pairs in contact):

𝑆𝑆𝑆𝑆𝑆𝑆(𝑡𝑡) = 𝐺𝐺𝐺𝐺𝐺𝐺(𝑡𝑡) + 𝑐𝑐(𝑡𝑡) ∙ 𝐹𝐹(𝑡𝑡) (2-4)

This further supports the strong connection between TE measurements and fault-
symptomatic changes in profile and TVMS. In healthy conditions, the effects of the TVMS are
visible on STE and, since TVMS is expected to be periodic with the gearmesh frequency, the
same periodicity is expected to dominate STE. In fact, in most spur gears the main observed
pattern is dominated by the one- to two-pair contact cyclic repetition, which results in a
compliance changing approximately by a factor of 2: higher when only one tooth-pair is in mesh
(contact around the pitchline) and lower when two pairs are in contact (root-tip contact). This
translates proportionally to STE, which sees the tooth deflections approximately doubled in the
single tooth-pair contact zone. A qualitative representation of the expected STE from a healthy
gear set with zero GTE is shown in Figure 2-3.

20
 Figure 2-3. Expected STE variation in the case of healthy gear (qualitative).

Finally, under more general operating conditions, with higher speed and non-negligible
load, time-varying tooth deflections and profile errors interact with the resonances of the gearbox
drivetrain and result in the so-called Dynamic Transmission Error (DTE) [85]. DTE is a distorted
form of STE, which still contains the information on profile and stiffness, yet biased by a transfer
function, also known as a frequency response function (FRF), between STE and DTE. Despite
this being seemingly similar to the problem of vibration measurements, there are two key
advantages of TE, even in its DTE form. Firstly, the transfer function effects are significantly
simpler than vibration, due to the closeness of the sensor to the source, and the torsional nature of
the measurement. Secondly, the overall scale factor of the measurement (DC gain) is known and
equal to 1, allowing for a direct conversion of DTE in microns of wear or tooth deflection.

Table 2-1. Summary of the effect of inputs and transfer function on the different forms of TE.

GTE STE DTE

Profile errors Yes Yes Yes
Tooth deflections No Yes Yes
Transfer function No No Yes

21
 More details on the relationships between inputs, GTE, STE, DTE and the system’s
transfer function will be given in Chapter 4, with a more rigorous mathematical description, which
forms the basis for the newly proposed methodology aimed at extracting profile and deflection
from TE. The next subsections will instead provide further details on TE, including its history and
use for condition monitoring.

2.3.2 Historical developments of TE

One of the first comprehensive studies of TE was published in 1958 by Harris [86], who
demonstrated the behaviour of spur gears under low speed with static transmission error plots,
now commonly known as Harris maps. Even twenty years earlier, however, Walker [87], [88]
investigated geometric imperfections and tooth deflection in gears, developing a method to
determine the profile modification required to compensate for tooth deflections under heavy load.
Walker’s work was later followed up by Harris [86] with the introduction of the TE concept, with
Harris, Gregory and Munro making a number of important early contributions to the field [89]–
[91]. In [89], a detailed experimental investigation was presented to show the dynamic behaviour
of spur gears using TE measurements. The measurement of TE was actually a challenge until an
optical method [92] was developed by the National Engineering Laboratory, which was one of
the pioneers in the development of TE measurement techniques. When TE was first defined, it
was well understood that it consists of both alternating and mean (DC) components, with the mean
component considered to be affected solely by the mean deflection of teeth under load. It was
shown that both components could be measured, but only for a gear pair of 1:1 ratio [90]. The DC
component of TE did not receive much attention until the late 1980s, when the advantages of
plotting TE curves with their DC components were stated [91]. It was mentioned that the DC
component could help in gear quality assessment during design, such as determining the
likelihood of tooth separation and the accuracy of bi-directional positioning, i.e., backlash.

2.3.3 TE-based gear condition monitoring

With the more prevalent use of shaft encoders in large-scale industrial gearboxes found
in wind turbines and production factories, TE has gained popularity as a diagnostic tool for
gearboxes. Compared to the mid-20th century, when the TE concept was originally developed, the
measurement of TE has become very convenient with the availability of relatively inexpensive
yet sophisticated encoders, usually mounted on the free (unloaded) ends of the gear shafts.

Endo et al. [93] were the first to use STE for the diagnostics of spalls and cracks. In [93],
a differential diagnosis using a cepstral technique was performed to differentiate spalls and cracks
based on how they vary the TE signal. A processing procedure for TE diagnostics was
demonstrated in [3] to extract the features of a simulated tooth root crack. Park et al. [6] applied

22
ensemble empirical mode decomposition to TE to extract fault features due to spalls and cracks
and a pattern recognition algorithm was used to distinguish between these two types of faults.
Recent developments in the diagnostics of localised faults in planetary gearboxes include the
combined use of TE and mesh phasing to perform differential diagnosis of cracks and spalls, as
well as to identify the faulty planet gear [94], [95].

In contrast to the case of localised faults, TE has rarely been used for the detection of
quantification of uniform wear. Studies in [96], [97] showed that wear leads to the increase of
peak-to-peak values of the gearmesh components in TE signals. It was also recently demonstrated
in [98], [99] that the three types of TE (GTE, STE and DTE) could be measured using encoders
under different operating conditions, and this has broadened the potential of TE as a reliable gear
diagnostic tool compared to conventional lateral acceleration or torsional vibration measurements.
Ref. [98] also demonstrated the use of GTE and STE signals from a lubricated gear wear
experiment to diagnose the progression of wear and the effects of wear (particularly mild and
severe pitting) on GTE and STE. However, despite a partially confirmed correlation between wear
severity and TE harmonics, none of the previous studies were able to quantify wear depth directly.

Compared to the transfer function affecting accelerometer measurements, the STE-to-

DTE transfer function is much simpler, as it is only affected by the torsional and radial modes of
the internal elements of the gearbox (shafts, bearings, couplings and gears) rather than also by
those of the casing and support structure. Often, this means that very few modes of the STE-to-
DTE transfer function lie in the frequency range of interest. Below the gearmesh frequency, DTE
is likely to be very similar to STE, since the presence of resonances at such low frequency would
be an undesirable design issue. Therefore, for a coarse analysis of the crack-induced tooth
deflections, as shown in [99], DTE can be used directly, without transfer function removal. This
however only allows for the reconstruction of the low-frequency effect of the crack, with potential
inaccuracies in the identification of the exact severity.

If the analysis is extended to a higher frequency range, encompassing a few harmonics of

the gearmesh frequency, the removal of a simple transfer function might be required. This was
recently done by Lu et al. [100], who used a cepstrum-based operational modal analysis (OMA)
to estimate STE from DTE measurements conducted with healthy gears. The paper fitted a single-
pole transfer function based on the random part of the TE signal (the noise floor in between shaft
harmonics), which is expected to follow the trends of the transfer function. This allowed
effectively reconstructing the gearmesh waveform, thus focussing on the mid frequency range of
0.5-5 times the gearmesh frequency. Another approach to account for the transfer function,
proposed in [101], is to actually implement a dynamic model able to simulate DTE based on an

23
input stiffness profile, and then identify the current crack severity by changing the input until the
simulated DTE matches the experimental results. Compared to the model-based approach, which
is affected by the quality of the dynamic model and requires onerous simulations in time domain,
the cepstrum-OMA method is significantly more efficient, but relies on the fitting of random TE
components which could in other practical cases be strongly corrupted by other torsional vibration
sources. Finally, none of the few works in this area have however fully exploited the combination
of multiple loads and speeds, often available in geared transmissions, in order to estimate and
then remove the transfer function and, at the same time, separate GTE from tooth deflections,
without the need for a dynamic model.

It is understood from section 2.3.2 that any TE measurement (including GTE, STE and
DTE) should include not only a time-varying component, but also a DC (average) component,
which has been neglected by researchers. This component is however carrying a crucial
information for worn gears: the average wear depth. This is better understood in the limit case of
perfectly uniform wear, where the tooth keeps an involute profile, but shifted inwards towards the
centreline. In this case (see Figure 2-4, right) the DC value in GTE would be the only component
able to represent wear severity, and in general the only parameter directly measuring average wear
depth. Despite the shifted involute case being unrealistic, the example in Figure 2-4 shows a limit
case which is often a large component of real wear profiles, which can be decomposed into a
shifted involute component and a zero-mean irregular profile.

Figure 2-4. Effects of local and uniform material removal on TE.

As seen in section 2.3.1, STE, is the sum of instantaneous GTE and tooth deflection,
which is in turn proportional to compliance 𝑐𝑐(𝑡𝑡) and total load 𝐹𝐹(𝑡𝑡). Under low-speed conditions,
the total load, which is directly proportional to torque, can be considered constant. Under such an

24
assumption, the difference between STE and GTE is directly proportional to the instantaneous
compliance. While the average compliance is a result of the combination of shaft, bearing and
tooth flexibility, shaft and bearing stiffness is almost constant. Therefore, the time-varying part
of compliance is directly proportional to the time-varying meshing compliance. Consequently, as
shown in [99], the combination of GTE and STE measurements allows separating the profile error
contribution from that of tooth deflections, estimating instantaneous compliance, and detecting
and quantifying root cracks in teeth.

Literature studies have in fact focused on the analysis of TE gearmesh harmonics and
their dependence on different faults. These studies have discussed changes in the amplitudes of
the gearmesh harmonics as well as the appearance of additive or multiplicative effects
(modulations). On the other hand, the DC shift of TE signals has not been given much attention
in the literature, with the only exception being gear quality control applications and DC shifts due
to mean deflection under load, and only in the very special case of a 1:1 gear ratio [102]. The
reason for this lack of attention is the fact that, except for the 1:1 gear ratio case, the DC shift of
TE is not directly available in conventional TE measurements. In a typical hunting tooth design,
traditional TE signals have an arbitrary DC value as only one of the two encoders/tachometers is
used to provide a zero-phase reference, with the other gear being arbitrarily oriented, depending
on the selection of the first tacho pulse of the reference shaft. Thus, in conventional TE, the DC
shift (signal mean) has no meaning and is set to zero, leaving only the variations around the mean
for analysis. This problem is not present in a 1:1 gear ratio, where the same tooth pair will come
into contact every revolution of a shaft and a fixed reference on one shaft results also in a fixed
phase on the other shaft, explaining why the change in DC shift due to load could easily be
identified in [102]. Except for this very unusual gear arrangement, there is currently no way of
identifying the DC shift of TE signals, and therefore its potential for TE-based diagnostics has so
far been neglected.

In chapter 6, a technique using the concept of hunting tooth frequency (HTF) will be
developed to allow retaining the DC component of TE. Therefore, it is beneficial to revisit this
concept. This is formally defined as the frequency at which a particular tooth of a gear meshes
with a particular tooth on the other gear [83]. Taking the reciprocal of the HTF produces the
hunting tooth period (HTP), which is the fundamental period of the complete mesh of a pair of
gears (i.e., the joint period of both gears). In the case of hunting-tooth design, i.e., when the
number of teeth of the two gears have no common factor, the length of a hunting tooth period is
equal to the period of one of the two shafts times the number of teeth of the other.

25
 Figure 2-5. Hunting tooth period in the case of hunting tooth design, example with 12 and 17
gear teeth.

It is interesting that the author of [103] utilised a signal with a length of HTP to construct
a three dimensional map known as the local meshing plane, which could assess the gear fault
severity in terms of the meshing quality of different tooth pairs. Other than that, the HTP has not
been given much attention in the field of condition monitoring, with practitioners instead using
for diagnostics the synchronously averaged signals over each shaft-period to identify changes in
the gearmesh harmonics and sidebands.

Only Li et al. have attempted prognosis using TE [101], achieved by first obtaining the
current crack severity based on their previous work [104], and then further utilising Paris law to
predict crack propagation and determine the RUL of the gear. The main drawback of this study is
the reliance on a dynamic model able to simulate the effect of the crack on the final TE
measurement, including the transfer function. These two papers, despite innovative, have the
drawback of using a model-in-the-loop approach for the identification of TVMS directly. This
means that multiple time-domain simulations are run for many candidates of TVMS until
agreement between simulated and measured results are obtained. This approach, despite effective
on the data used in the papers, is very inefficient due to the time-domain simulations. Moreover,
it neglects profile errors and their effect on TE, often dominant over deflections, especially for
large facewidths.

2.4 Summary and gaps

Based on the surveyed literature, the majority of the developments on the condition
monitoring of gears are based on vibration measurements. While vibration has been well-
established in detection and diagnostics of gears, its use for severity assessment remains a
challenge due to a complex transfer path from the source of gearmesh to the measurement point

26
on the gearbox casing. Without large training data or a complex dynamic model, it is hard to
obtain an accurate and scaled measurement of the fault using vibration.

On the other hand, TE is more closely related to the gearmesh and has been considered
as a key quantity for the study of sound- and vibration-emission in gearboxes for decades, but its
use in condition monitoring is very limited. Further work is also required before TE can be fully
utilised as an effective tool in the field.

The analysis of the literature on TE-based gear diagnostics allows adding some additional
consideration to the gaps outlined in the Introduction to this thesis (section 1.1).

Gap 1. From TE to tooth profile and deflection. Measured TE can be generally considered a
DTE, i.e., with the effects of profile errors and deflection distorted by the transfer function from
the gear mesh to the encoder measurements. This transfer function can be identified and removed
by OMA or accounted for in simulation. The simulation path is however inefficient and relies on
accurate models, which in real complex machines could become onerous and difficult to obtain.
The OMA path has only been explored using the cepstral analysis of the random part of the TE
signal, which however requires the measurements to be unaffected by any other source. Even after
the removal of the transfer function, a rigorous approach to separate compliance and profile
effects in STE, where they are linearly combined, was not reported. What was completely missed
in the few proposed approaches is the possibility to actually exploit the availability of multiple
speeds and loads. The former have predictable effects on the interaction of STE components with
the transfer function, while the latter acts as a weighting parameter on the linear combination of
compliance and profile in STE.

Gap 2. Crack severity assessment. As for Gap 1, the only algorithm so far attempting crack-
severity estimation using TE has not exploited the common case of multiple loads available.
Combining information from multiple loads in the estimation of crack severity is crucial to render
it more robust and insensitive to profile error.

Gap 3. Wear severity assessment. The DC component of TE remains the crucial missing
information for the achievement of reliable wear-severity estimation. Vibration-based methods
are in this case not able to capture average wear depth, due to the insensitivity of accelerometers
to DC values. TE is conceptually able to capture DC shifts of the involute profiles, but current
signal processing approaches for its calculation do not include an absolute reference and therefore
lose this crucial information. The relative angle between input and output tacho pulses is in fact
meaningless unless a fixed hunting-tooth-period reference is used to compare a healthy gear and
a worn gear. If it is ensured that the same position within a hunting tooth period is attained for

27
both cases, then the change in the relative angle between input and output tacho would indicate
wear depth. The relationship between HTP and the DC component of TE has been so far
completely neglected and needs to be investigated in order to solve this issue.

28
Chapter 3 Methodology and experimentation

To address the research gaps mentioned in the previous chapter and thus achieve the
objectives of this research, a methodology has been designed. In this chapter, an overview of the
methodology that formulates the complete condition-monitoring framework to assess crack- and
wear- severity is presented, together with an outline of experimental investigations that support
this research. An introduction of the overall methodology of this thesis is first described in section
3.1 of the chapter, followed by three sections (3.2, 3.3. and 3.4) that split the thesis in three
corresponding tasks. Each of the three sections addresses one of the three research objectives
discussed in the introduction. Lastly, section 3.5 outlines the experimental facilities, such as the
test rig, gear specimens and sensors, as well as the details of the collected data. Some of the
material presented in this chapter consists of adapted extracts from the papers4 titled “Use of
transmission error for a quantitative estimation of root-crack severity in gears” [8], and “Absolute
transmission error: a simple new tool for assessing gear wear” [9].

3.1 Overall methodology of thesis

To tackle the gear-condition-monitoring challenges and gaps discussed in the previous

chapters, a complete condition-monitoring framework is required. This section provides an

4
Permission has been granted by co-authors.

29
overview of the overall methodology of this thesis, as illustrated in Figure 3-1. The methodology
includes the study and severity assessment of the two gear faults in question, i.e., crack and wear.

Figure 3-1. Overall methodology of thesis.

Theoretically, a TE signal can be separated in its average value and its time-varying
component. The average value carries important information on the average wear depth of the
teeth, while the time-varying component is critical in the detection of individual-tooth problems
such as root cracks. However, the time-varying component of TE still combines both tooth
deflections and profile alterations, the latter resulting from manufacturing and the uneven
component of wear. Moreover, the dynamics of the gearbox internal elements (e.g., flexibility of
shafts, bearings and couplings, and rotational inertias of the drivetrain) result in the time-varying
part of TE being distorted, especially when operating at high angular speed.

The removal of the transfer path effects on the time-varying part and the separation of
tooth-profile and tooth-deflection effects (Objective 1) are tackled by the first task presented in
section 3.2. The tasks result in separate and unbiased estimates of time-varying mesh-compliance
and profile alterations. The first forms the starting point for the second task of this thesis,
described in section 3.3, which compares it with theoretical models to estimate crack severity
(Objective 2), while the second is retained as supporting information on wear. The profile-related
part of the time-varying TE component is however of secondary importance to the DC value of
TE, which is directly linked to the average wear depth. Section 3.4 describes the methodology
adopted to accurately estimate the DC value of TE, and therefore quantify average wear depth,
the key component of wear severity (Objective 3). Moreover, an approach to combine this
information with the time-varying component is discussed, in order to provide further information
on the most likely wear mechanism.

Throughout the thesis, a set of general assumptions are used:

30
 1. The encoder measurements at the free ends of the shafts are not affected by static
deflection of the shafts, and they follow the motion of the gears up to a very high
frequency [99].
2. Deflection and geometric error combine linearly into a quasistatic component, which
passes through the same transfer function to give rise to a TE measurement [100].
3. Transfer functions are not affected by load significantly, and any nonlinear effects (e.g.,
Hertzian contact) is neglected and linearised. This assumption will be discussed more in
detail in Chapter 4.

3.2 Removal of transfer-path effects and separation of profile error and tooth
flexibility (Objective 1)

TE measurements consist of effects from profile error and tooth deflection, both distorted
by the transfer path. To achieve Objective 1, this thesis developed a methodology based on the
physical understanding of these three effects contributing to TE, and the different way in which
they are in turn affected by operating speed and load.

The component of TE corresponding to tooth profile alterations is scarcely affected by

load. On the other hand, tooth deflections, which are obtained as a product of tooth compliance
and load, are strongly load-dependent. With this knowledge and the availability of TE signals
obtained at different loads, it is possible to separate the information regarding profile from that of
tooth flexibility.

It is known that TE measurements would appear different when collected at different

operating speeds. While this is true, the actual profile alterations and tooth compliance behind the
generation of TE should in theory be identical across all speeds, i.e., speed does not change the
actual condition of the tooth but rather distorts the way in which this input reflects into TE. These
distortions are due to the fact that the input (profile and compliance) passes through different
regions of the transfer function in the frequency spectrum when TE is measured at different speeds.
The relationships between profile error, tooth deflection and transfer function are illustrated in
Figure 3-2.

31
 Figure 3-2. Relationships between profile error, tooth flexibility and transfer function.

Following the understanding of these important relationships, a methodology with two

main steps is developed. The first step is the removal of transfer-function effects from TE using
an OMA-based approach. In order to achieve this, TE signals recorded at multiple speeds are
required in order to observe the trend of the frequency response function (FRF) in the TE spectrum
within the frequency range of interest. Once the FRF is identified, it can be deconvolved from the
TE measurement to obtain an “equivalent static TE”, i.e., a TE unbiased by the gear dynamics.
The second step is the implementation of a regression procedure to separate the profile-error and
tooth-flexibility components from the equivalent static TE. This instead requires the signals to be
recorded at multiple loads to observe the load-induced tooth deflections. Therefore, this
methodology, as depicted in Figure 3-3, exploits the combination of multiple speeds and loads,
allowing the reconstruction of profile error and meshing compliance.

32
 Figure 3-3. Methodology to remove transfer function and separate profile error and tooth
flexibility.

To validate the methodology in this section as well as the crack-estimation technique in

the following section, encoder signals have been acquired at multiple operating speeds and loads
from experiments with different crack sizes. Further details of the experiments are available in
later sections of this chapter.

The full details of this approach and corresponding results obtained on experimental TE
measurements are provided in Chapter 4.

3.3 Assessing crack severity using TE (Objective 2)

A crack reduces the stiffness of a tooth, and leads to plastic deformation that causes
profile alterations. The two effects can be separated using the methodology described in the
previous section. As such, a meshing compliance that consists of information about the reduction
of stiffness can be obtained to assess gear crack severity. The approach used in this thesis to tackle
this objective is based on a two-step methodology. Firstly, the tooth showing the strongest
deviation from the average tooth behaviour is identified as the most likely candidate for a crack.
Then, the compliance that represents the cracked tooth is compared with the theoretical
compliance model of different crack sizes, and the crack depth is identified based on the best fit.
While the development of the signal pre-processing and overall model-based methodology is a
novel contribution of this thesis, the implementation of the theoretical compliance model of ref.
[15] for the UNSW test-rig was a courtesy of PhD student Y. Mao, conducting a parallel thesis in

33
our group, and mainly focussing on gear modelling. The methodology described above is
illustrated in Figure 3-4.

The full details of this approach and corresponding results obtained with tooth cracks of
different severities are provided in Chapter 5.

Figure 3-4. Crack detection and model-based severity assessment.

3.4 Assess wear severity using TE (Objective 3)

The procedure described in section 3.2 is significant only for the time-varying component
of TE, which actually coincides with the traditional way of measuring TE. The DC value of TE,
not even captured in traditional TE measurements, is the focus of this task of the thesis, addressing
Objective 3. In order to estimate the average value of TE, a rephasing technique is developed
based on a reference TE measurement, taken in healthy conditions. Using this signal as a reference,
the average component of the TE signal is estimated and used as a direct measurement for average
wear depth.

The methodology is illustrated in Figure 3-5.

34
 Figure 3-5. Obtaining absolute TE and its application to identify average wear depth.

The proposed technique is discussed in detail in Chapter 6, and validated using encoder
and tachometer measurements from a dry gear wear experiment.

The DC value of TE constitutes the most important component in wear severity

assessment, especially for medium and severe cases. However, the information available in the
time-varying part of TE is also considered, providing information on the shape of the profile and,
consequently on the characteristic patterns of different types of wear mechanisms. Chapter 7
discusses this additional component and the results obtained from the case of abrasive wear and
fatigue pitting, thus completing the analysis of all components of TE.

3.5 Experimental facilities and data

None of the developments in this thesis would have been possible without the extensive
experimental campaigns which form an integral part of the work of this thesis, with the help of
other students. This section describes the experimental facilities, the data collection and testing
procedures. Three main types of tests have been conducted using a spur gear test rig – crack tests,
dry wear tests and lubricated wear tests.

3.5.1 Test rig

All experimental data analysed in this research was collected from the spur gear test rig
of the UNSW Tribology and Machine Health Monitoring Group. The test rig was designed by
previous research students of the group, but was further improved and adapted for the
requirements of precise TE measurements. Figure 3-6 and Figure 3-7 show the entire test rig and
a closer view of the gearbox components respectively. A schematic diagram of the test rig is also
provided in Figure 3-8.

35
 The test rig features a 4-kW induction motor connected to the input shaft, controlled by
a variable frequency drive (VFD). At the other end of the drivetrain, an electromagnetic particle
(EMP) brake is connected to the output shaft to apply load.

Figure 3-6. Spur gear test rig.

Figure 3-7. Gearbox components.

36
 Figure 3-8. Schematic diagram of spur gear test rig.

The gear pair itself was different in each campaign, as different dimensions, materials
and surface treatments were chosen to accelerate or enhance a certain type of fault. All gears were
however spur gears, with a centre distance of 71 mm. Specifications of each gear pair is discussed
in the following sections for each campaign.

For the acquisition of TE signals, encoders were mounted on the unloaded (free) end of
both shafts to record angular motion as rigidly as possible with respect to the gears, at least up to
a certain frequency. Different models of the encoders were used on the different tests, as the test-
rig was improved during the duration of this thesis. Details on the different types of encoders are
provided in the following sections. In addition to encoders, vibration signals were recorded from
accelerometers mounted on the gearbox casing, whose location, number and model varied in the
three tests. The measurement setup was finally completed by a torque meter measuring the input
torque. All signals were sampled at 200 kHz for the crack tests and 100 kHz for the wear tests.

3.5.2 Crack tests

To validate the first two tasks briefly introduced in sections 3.2 and 3.3, crack tests were
conducted using a hunting-tooth-designed steel gearset with a 27-tooth driving gear and a 44-
tooth driven gear, with a gear module of 2 mm, a facewidth of 5 mm, a bore radius of 10 mm and
a standard contact angle of 20°. The gears have induction-hardened teeth to avoid significant
development of wear during the tests. In these campaigns, the encoders were the Heidenhain ROD
426, which generate a once-per-rev tacho signal and 3600 pulses per shaft revolution.

A series of tests were conducted at different operating loads and speeds, with pinion
cracks of three different sizes (small, medium and large), as shown in Figure 3-9. These gear
cracks are artificially generated slots that have a width of 0.35 mm and are at approximately 45°
from the fillet to the tooth centreline across the whole facewidth. Differently from a real crack,
these slots are expected to show a ‘closing crack’ effect on GTE, i.e., a situation where GTE

37
shows an opposite sign with respect to load-induced deflections, as observed in [99]. This will be
further discussed in Chapter 4. Measurements were collected at nominal input speeds of 1, 2, 5,
10, 15 and 20 Hz and nominal loads of 0, 5, 10, 15 and 20 Nm with respect to the input shaft. The
nominally zero-load cases are actually characterised by about 2 Nm at the torque meter, even if
the brake was turned off.

Cracks are located on the active tooth flank, and therefore load-induced deflections result
in an opening of the cracks. The selection of the severity levels was done considering similar
previous studies in the literature. Even the most severe case (50%) would in many applications
result in a non-negligible remaining useful life of the gear, as shown for instance in [105], where
a 50% cracked tooth required 21 days of operation to reach failure (run-to-failure test with initial
artificially seeded 50% slot).

Small (19 %) Medium (36 %) Large (50 %)

Figure 3-9. Gear tooth with small (left), medium (middle) and large (right) root crack.

A summary of the variables used for the modelling of healthy and cracked tooth
compliance in the potential energy method (to be applied in Chapter 5) of ref. [15] is reported in
Table 3-1.

38
 Table 3-1. Summary of gear parameters for the crack tests and to be used in the TVMS model.

Number of teeth – Pinion 27

Number of teeth – Driven gear 44
Module 2 mm
Tooth width (facewidth) 5 mm
Bore radius (shaft radius) 10 mm
Pressure angle 20°
Young’s modulus 206 GPa
Poisson’s ratio 0.3
Shear modulus 79.23 GPa
Addendum 1 × module
Tip clearance 0.25 × module
Crack initial position (angle 𝜓𝜓 in ref.
35°
[15])
Crack propagation angle 45°

3.5.3 Dry wear test

To validate the methodology introduced in section 3.4 and for the comparative study in
Chapter 7, a dry gear wear experiment was conducted. The module 2 mm gearset used in this
campaign consists of a pinion with 19 teeth and a driven gear with 52 teeth, and thus a hunting
tooth design to limit uneven wear. The gears are made of mild steel JIS S45C and were surface-
treated with a layer of black oxide. The tooth surfaces were not hardened, giving low surface
durability and allowing the rapid development of abrasive wear long before bending fatigue
occurs. The encoders used in these tests were Heidenhain ROD 426, which produce 1000 pulses
per shaft revolution and a once-per-rev tacho signal that serves as a phase reference.

To wear the gears, the rig was run nominally at an input shaft speed and load of 10 Hz
and 5 Nm, respectively. Each wear period lasted about 5 minutes, or approximately 3000 cycles
of the pinion, with 14 wear periods (≅ 42000 pinion cycles) completed in total. A set of TE
measurements was recorded at different operating conditions so as to obtain GTE (2 Hz, 0 Nm),
STE (2 Hz, 5 Nm) and DTE (10 Hz, 5 Nm) measurements before and after each wear period. The
duration of each record was at least the length of one HTP (rounded up to 30 s and 10 s for the 2
Hz and 10 Hz cases, respectively), so that the signals were sufficiently long for the application of
the proposed method. A total of 15 sets of records were acquired, which are named Test 0 to 14

39
in the results in later sections. Test 0 represents the unworn case, and is used as the reference
healthy signal. In between the wear periods, the rig was stopped and an additional procedure
undertaken – the removal, using a small brush, of any wear particles retained on the gear tooth
surfaces. This was performed to investigate the differences in GTE with and without the presence
of these particles. Other than TE measurements, every two wear periods (10 minutes), the wear
particles removed from the gears (both with the brush and the majority that had fallen naturally)
were collected using a piece of adhesive paper placed carefully at the bottom of the gearbox casing.
These paper sheets were weighed before and after collecting the particles to obtain an approximate
measurement of the amount of material removed from the gears at different stages throughout the
test. These particle mass measurements were complemented by mass measurements of the gears
themselves taken at the start and end of the entire test. The total mass of the particles was weighed
to be 12.57 g whereas the actual change in total mass of both gears was measured to be 13.35 g,
showing that the particle measurements were reliable with an error margin of 6%.

Figure 3-10. Tooth surfaces – (a, b) driving gear; (c, d) driven gear; (a, c) healthy; (b, d) worn.

Figure 3-10 shows images of both the driving and driven gears to provide an indication
of the surface roughness of the tooth surface before and after the wear test.

3.5.4 Lubricated wear test

On top of the dry wear experiment, a lubricated gear wear test was used in the
comparative study of Chapter 7. Similarly to the previous test, the gearset is a speed-reduction
setup with a 19-tooth driving gear and a 52-tooth driven gear, both made of mild steel JIS S45C.
The gear tooth also had low surface durability so as to allow surface degradation before bending
fatigue could occur. For the acquisition of TE signals, encoders which give a once-per-rev tacho
signal and a 512-pulse-per-rev signal were mounted on the free ends of the input and output shafts.

To develop wear on the gears, the rig was run in a lubricated condition for about 3.25
million (mil.) input shaft cycles. TE measurements were taken at different operating conditions
at 2 Hz and 0 Nm for GTE, as well as 2 Hz and 20 Nm for STE (measured at the input shaft).
Unfortunately, DTE measurements were not available in that experiment due to a resonance of
the encoders which affected the TE signals at higher frequency. Vibration measurements depicted

40
in later sections were instead taken at 20 Hz and 20 Nm, because they were not affected by the
encoder resonance. There were also stop intervals throughout the experiment to take moulds of
the tooth surfaces to monitor the progression of surface pits. Some images of the moulds are
shown in Figure 2 to illustrate the surface conditions as pitting progresses. The surface images in
Figure 2 are of the driving gear, which had a higher rate of pit development as compared to the
driven gear. It can be seen that pitting started at around 0.224 million cycles. Throughout the test,
pits continued to develop with increasing size and severe material loss towards the end. A more
detailed study of the moulds with image analysis can be found in [106]. The same study also
shows how to obtain the estimates of pitted area used in Chapter 7.

Figure 3-11. Surface images of tooth at different wear stage (as described in [106]).

3.5.5 Summary table of experiments

The following table summarises the features of each experimental campaign.

Table 3-2. Summary of experimental details.

Experiments Crack Dry wear Lubricated wear

Driving gear tooth 27 19 19
number
Driven gear tooth 44 52 52
number
Tooth facewidth 5 20 20
(mm)
Tooth surface Induction-hardened Black oxide coating Black oxide coating
treatment
Sampling frequency 200 100 100
(kHz)
Encoders 3600 1000 512
(pulses per rev)
Accelerometers 4 3 2
(all B&K 4396) (2 x B&K 4396 and (B&K4396 and
1 x B&K 4394) B&K4394)

41
Chapter 4 Removal of transfer-path effects in TE
and separation of profile error and tooth flexibility

When measured at higher operating speeds, TE is subjected to dynamic effects due to the
transfer path. Moreover, even if these dynamic effects were removed (or if TE is measured at low
speed) a static transmission error (STE) always contains the combined effects of profile alterations
(with respect to the involute) and tooth flexibility. These need to be separated before crack/wear
severity can be investigated. This chapter presents the development and validation of the
methodology introduced in section 3.2 to address the two aforementioned challenges, achieving
Objective 1 of this thesis. The material presented in this chapter is an adapted version of the work
published in the journal paper 5 titled “Use of transmission error for a quantitative estimation of
root-crack severity in gears” [8]. A few parts were also moved to previous chapters.

4.1 Conceptual understanding of transmission error

In this section, the underlying relationships and interactions between GTE, tooth
compliance, STE and DTE will be recapped and discussed in more detail than in Chapter 2, and
with mathematical expression, in order to support the methodological choices discussed in later
sections.

5
Permission has been granted by co-authors.

42
 As mentioned in Chapter 2, GTE corresponds to profile error, while STE includes both
GTE and tooth deflections, which in turn are obtained as a product of meshing compliance 𝑐𝑐(𝑡𝑡)
and the total contact force 𝐹𝐹, which can be considered constant at low speed, when the transfer
function has a quasi-static behaviour. In an ideal healthy case, where all teeth are assumed to be
identical, the meshing compliance 𝑐𝑐(𝑡𝑡) is periodic with the gearmesh frequency. Assuming
instead that one of the gears has a cracked tooth, the meshing compliance becomes synchronous
with the corresponding shaft. Therefore, in the general case the Fourier Series of compliance
consists of harmonics 𝐶𝐶[𝑘𝑘] at frequencies 𝑓𝑓𝑘𝑘 , multiples of the cracked-gear shaft speed 𝑠𝑠 (𝑓𝑓𝑘𝑘 =
𝑘𝑘 ⋅ 𝑠𝑠, and 𝑠𝑠 = shaft speed). The shaft-synchronous spectral components of 𝑆𝑆𝑆𝑆𝑆𝑆(𝑡𝑡) will therefore
show the same relationship with the spectral components of GTE and compliance:

𝑆𝑆𝑆𝑆𝑆𝑆[𝑘𝑘] = 𝐺𝐺𝐺𝐺𝐺𝐺[𝑘𝑘] + 𝐶𝐶[𝑘𝑘] ∙ 𝐹𝐹 (4-1)

where 𝑆𝑆𝑆𝑆𝑆𝑆[𝑘𝑘] and 𝐺𝐺𝐺𝐺𝐺𝐺[𝑘𝑘] are the Fourier coefficients of 𝑆𝑆𝑆𝑆𝑆𝑆(𝑡𝑡)and 𝐺𝐺𝐺𝐺𝐺𝐺(𝑡𝑡), corresponding to
a frequency 𝑓𝑓𝑘𝑘 , or, more precisely, the equivalent order-spectrum components.

These relationships are shown in the angular and order domains in Figure 4-1, which
qualitatively shows a plausible STE for a single-stage gear transmission with a 27-tooth pinion
with one cracked tooth (70% crack severity), a contact ratio of 1.68, and a constant total contact
force 𝐹𝐹 = 1 kN. Figure 4-1(a) shows the waveforms of GTE, tooth deflections Δ𝑇𝑇𝑇𝑇(𝑡𝑡) = 𝑐𝑐(𝑡𝑡) ⋅
𝐹𝐹, and the resulting STE over a single pinion shaft revolution. In this example, GTE is assumed
to be unaffected by the crack, thus showing a regular 27-per-rev pattern, whereas tooth deflection
is characterised by a superposition of the healthy 27-per-rev pattern and the 1-per-rev crack
symptom. The former is simply due to the regular transition between single- and double-tooth-
pair contact, whereas the crack symptoms result in an additional compliance over the whole tooth-
contact duration (1.68 gearmesh periods). As is intuitively understandable, the additional crack
deflection is maximum during single-tooth-pair contact, since all the load is concentrated on the
cracked tooth. Despite being a 1-per-rev event, the short duration of the crack symptom in time
(1.68 gearmesh periods) necessarily results in a spread over a large frequency range, i.e., the
amplitude increase of a large number of shaft harmonics, as shown in Figure 4-1(b).

43
 Figure 4-1. Relationship between GTE, STE and Δ𝑇𝑇𝑇𝑇 in (a) angular domain, (b) order-
amplitude domain.

When TE measurements are taken at higher speeds, dynamic effects are no longer
negligible; the contact force is no longer constant, but modified by the resonances of the internals,
as shown in [100]. DTE is measured, rather than STE, and the transfer function between them
corresponds to the change in force. Incidentally, the changed force at the mesh acts through the
corresponding forces at the four bearings to change vibration responses at external measurement
points through four further transfer functions. In this work we assume that STE does not only
represent an actual TE measurement taken at low speed, but also the general linear combination
of the two inputs of compliance and profile error, with the first weighted by the time-averaged
load. This input is therefore considered as an equivalent STE which is the source of the DTE
signals measured at higher speed. This interpretation is fairly accepted in the literature, as
evidenced by refs. [107]–[109]. For conciseness, the term STE will be used to refer to this
combination of inputs rather than using “equivalent STE”. As such, a single transfer function
between the two components of STE (GTE and deformations) and the measured DTE is expected,
as supported by the simplified model and experimental results shown in ref. [100]. Conceptually,
the best way to understand the meaning of this transfer function is the following: STE is a
combination of tooth profile errors and deflection, serving as an input to the system dynamics;
this input is amplified/attenuated/delayed by the resonances and antiresonances of the system

44
(transfer function 𝐻𝐻(𝑓𝑓)), and it is finally measured as torsional vibrations at the encoder positions
(combined in the DTE measurement). The inputs to the system (profile error and TVMS) are
obviously unaffected by speed as they are physical and geometrical properties of the gears. As
such, their combination into an equivalent STE is also unaffected by speed if sampled in the
angular domain of the reference shaft. This means that 𝑆𝑆𝑆𝑆𝑆𝑆[𝑘𝑘], i.e., the 𝑘𝑘th harmonic of STE in
the order domain, remains constant at all speeds. The transfer function between STE and DTE
also remains constant independent of speed, but in the raw frequency domain (Hz) rather than in
orders. Such a transfer function 𝐻𝐻(𝑓𝑓) is also expected to be relatively unaffected by the operating
speed and load of the system, as long as non-linearities, mainly due to bearings and couplings,
are not significant in the frequency range of interest.

Under these assumptions, it is possible to define different measurements of DTE,

obtained at different speeds 𝑠𝑠𝑚𝑚 as:

𝐷𝐷𝐷𝐷𝐸𝐸𝑚𝑚 [𝑘𝑘] = 𝐻𝐻(𝑘𝑘 ⋅ 𝑠𝑠𝑚𝑚 ) ⋅ 𝑆𝑆𝑆𝑆𝑆𝑆[𝑘𝑘] (4-2)

This means that, since DTE is a combination of an order-domain invariant quantity
𝑆𝑆𝑆𝑆𝑆𝑆[𝑘𝑘] and a frequency-domain invariant quantity 𝐻𝐻(𝑓𝑓), 𝐷𝐷𝐷𝐷𝐸𝐸𝑚𝑚 [𝑘𝑘] will vary with speed as the
harmonics of STE move higher or lower in frequency following 𝑠𝑠𝑚𝑚 , and interact with a different
portion of the transfer function 𝐻𝐻(𝑘𝑘 ∙ 𝑠𝑠𝑚𝑚 ). By denoting with 𝐹𝐹𝑚𝑚 the average load of the 𝑚𝑚-th test
(i.e., the corresponding constant force that would be observed if the test was conducted at low
speed, thus obtaining STE), we can further expand this relationship into:

𝐷𝐷𝐷𝐷𝐸𝐸𝑚𝑚 [𝑘𝑘] = 𝐻𝐻(𝑘𝑘 ⋅ 𝑠𝑠𝑚𝑚 ) ⋅ {𝐺𝐺𝐺𝐺𝐺𝐺[𝑘𝑘] + 𝐶𝐶[𝑘𝑘] ∙ 𝐹𝐹𝑚𝑚 } (4-3)

45
 An example of this relationship, applied to the simulated case of Figure 4-1, is shown in

Figure 4-2, for the two cases of pinion shaft speed 𝑠𝑠1 = 5 Hz (left column) and 𝑠𝑠2 = 10
Hz (right column). The top row shows two identical STE spectral diagrams, where the only effect
of speed is the dilation of the frequency axis. The transfer function is identical in frequency, and
thus compressed along the order-axis as shaft speed grows (mid row). This phenomenon is
illustrated with the aid of a double axis in the mid row of Figure 4-2, in which the axis at the top
of the plots represents the order axis, and the axis at the bottom represents the frequency axis.
This relative dilation of order and frequency axes, and therefore the change in the location of each
shaft harmonic along the transfer function, results in the different spectra of DTE for the two
speeds (bottom row).

46
 Figure 4-2. Relationship between STE, DTE and 𝐻𝐻 (𝐹𝐹𝐹𝐹𝐹𝐹) in the order domain at input speed
of 5 Hz (left column) and 10 Hz (right column).

4.2 Methodology

The methodology proposed in this section aims at inverting the chain of relationships
shown in section 4.1, and therefore works backwards from DTE at different loads and speeds to
obtain a robust estimate of compliance and GTE. A summary of the processing procedure of the
methodology is presented in Figure 4-3.

Figure 4-3. Summary of the proposed methodology.

47
4.2.1 Calculation and pre-processing of DTE signals
The methodology relies on a series of 𝑚𝑚 = 1, … 𝑀𝑀 input- and output-shaft encoder
measurements, obtained at known and constant input shaft speed 𝑠𝑠𝑚𝑚 in Hertz and gear load 𝐹𝐹𝑚𝑚 in
Newtons (torque/gear base radius). The extraction of angular signals 𝜃𝜃𝑚𝑚,𝑖𝑖𝑖𝑖 (𝑡𝑡) and 𝜃𝜃𝑚𝑚,𝑜𝑜𝑜𝑜𝑜𝑜 (𝑡𝑡) from
raw encoder measurements was performed by applying both phase demodulation-based (PDOT)
[110] and zero-crossing methods [59], with both methods giving similar results (PDOT was then
chosen).

DTE is computed in the time domain as:

𝐷𝐷𝐷𝐷𝐷𝐷𝑚𝑚 (𝑡𝑡) = 𝜃𝜃𝑚𝑚,𝑖𝑖𝑖𝑖 (𝑡𝑡) ⋅ 𝑅𝑅𝑖𝑖𝑖𝑖 − 𝜃𝜃𝑚𝑚,𝑜𝑜𝑜𝑜𝑜𝑜 (𝑡𝑡)𝑅𝑅𝑜𝑜𝑜𝑜𝑜𝑜 (4-4)

where 𝑅𝑅𝑖𝑖𝑖𝑖 and 𝑅𝑅𝑜𝑜𝑜𝑜𝑜𝑜 are the base radii of input and output gear, respectively.

Then, DTE is resampled into 𝐷𝐷𝐷𝐷𝐸𝐸𝑚𝑚 (𝑛𝑛Δ𝜃𝜃) using constant input-shaft angular increments.
When resampling to the angular domain, an integer number of samples 𝑁𝑁 per gearmesh period is
chosen, so that 𝑁𝑁 ⋅ 𝑍𝑍 ⋅ Δ𝜃𝜃 = 2𝜋𝜋 represents one complete input shaft rotation, 𝑍𝑍 is the number of
teeth on the input-shaft gear, and the resulting 𝐷𝐷𝐷𝐷𝐷𝐷𝑚𝑚 (𝑛𝑛Δ𝜃𝜃) is defined for 𝑛𝑛 = 0, 1, … , 𝑁𝑁 ⋅ 𝑍𝑍 − 1.

𝐷𝐷𝐷𝐷𝐷𝐷𝑚𝑚 (𝑛𝑛Δ𝜃𝜃) is also band-pass filtered to retain only harmonics in the frequency range of
interest. The lower limit of the pass-band is chosen so that the average DTE is removed, together
with the first two shaft harmonics. The latter are often compromised by eccentricities and
misalignments. The higher bound of the same band is instead set just above a few multiples of the
gearmesh frequency, in order to retain sufficient information to well-approximate the gearmesh
waveform, but avoid extending the OMA to too many modes.

Using the same symbol as in section 2, the order spectrum of each signal 𝐷𝐷𝐷𝐷𝐷𝐷𝑚𝑚 (𝑛𝑛Δ𝜃𝜃) is
denoted as 𝐷𝐷𝐷𝐷𝐷𝐷𝑚𝑚 [𝑘𝑘], where 𝑘𝑘 represents the order of the harmonics of the reference shaft, limited
by the pass-band to the range of integer values between 𝑘𝑘𝑚𝑚𝑚𝑚𝑚𝑚 and 𝑘𝑘𝑚𝑚𝑚𝑚𝑚𝑚 . 𝐷𝐷𝐷𝐷𝐷𝐷𝑚𝑚 [𝑘𝑘] can be
obtained by synchronous averaging of the signal with respect to the reference shaft. To avoid
biased effects (such as encoder-accuracy errors) coming from one of the two shafts, the average
should be performed over an integer number of hunting-tooth periods, so that the signal includes
integer numbers of both input and output shaft revolutions.

4.2.2 Removal of transfer function and regression to estimate GTE and compliance
A variety of approaches are possible for the identification and removal of the transfer
function, including the methodology recently published in [100]. In this chapter, a new
methodology is proposed which exploits the availability of multiple speeds. The proposed
approach exploits the fact that input shaft harmonics of STE would be distorted differently at

48
different operating speeds, due to their interaction with a different part of the transfer function
𝐻𝐻(𝑓𝑓) of Equation (4-2). The proposed OMA approach therefore attempts to estimate the STE-to-
DTE transfer function based only on the relative amplitude and phase of shaft harmonics at
different operating speeds. Differently from [100], this approach therefore relies only on the
deterministic signal, which is expected to be less affected by other non-gear-related sources and
background noise. In order to estimate and remove the transfer function 𝐻𝐻(𝑓𝑓) between STE and
DTE, the following pole-zero parameterisation is chosen:

𝑓𝑓 2 𝑓𝑓
∏𝐵𝐵𝑏𝑏=1 �− � � + 2𝑗𝑗𝜁𝜁𝑧𝑧 [𝑏𝑏] � � + 1�
𝑓𝑓𝑧𝑧 [𝑏𝑏] 𝑓𝑓𝑧𝑧 [𝑏𝑏]
𝐻𝐻(𝑓𝑓; 𝛈𝛈) = (4-5)
𝑓𝑓 2 𝑓𝑓
∏𝐴𝐴𝑎𝑎=1 �− � � + 2𝑗𝑗𝜁𝜁𝑝𝑝 [𝑎𝑎] � � + 1�
𝑓𝑓𝑝𝑝 [𝑎𝑎] 𝑓𝑓𝑝𝑝 [𝑎𝑎]
with 𝐴𝐴 pole-pairs and 𝐵𝐵 zero-pairs, described by the parameter array

𝛈𝛈 = �𝑓𝑓𝑝𝑝 [1], … , 𝑓𝑓𝑝𝑝 [𝐴𝐴], 𝜁𝜁𝑝𝑝 [1], … , 𝜁𝜁𝑝𝑝 [𝐴𝐴], 𝑓𝑓𝑧𝑧 [1], … , 𝑓𝑓𝑧𝑧 [𝐵𝐵], 𝜁𝜁𝑧𝑧 [1], … , 𝜁𝜁𝑧𝑧 [𝐵𝐵]� (4-6)

which contains the corresponding frequencies and damping ratios of the poles and zeros of the
transfer function. It must be noted that, thanks to the properties of TE, the zero-frequency gain of
𝐻𝐻(𝑓𝑓) is known to be 1.

Using this parameterisation in Equation (4-3), the following procedure can be applied.
For any possible set of poles and zeros 𝛈𝛈 it is possible to estimate STE at each load and speed:

𝐷𝐷𝐷𝐷𝐷𝐷𝑚𝑚 [𝑘𝑘]
� 𝑚𝑚 [𝑘𝑘; 𝛈𝛈] =
𝑆𝑆𝑆𝑆𝐸𝐸 (4-7)
𝐻𝐻(𝑘𝑘𝑠𝑠𝑚𝑚 ; 𝛈𝛈)
� 𝑚𝑚 [𝑘𝑘; 𝛈𝛈] is estimated for each 𝑚𝑚 = 1, … 𝑀𝑀 , a linear regression problem can be
Once 𝑆𝑆𝑆𝑆𝑆𝑆
established to find:

𝑀𝑀 −1 𝑀𝑀

𝐶𝐶̂ [𝑘𝑘; 𝛈𝛈] = � � (𝐹𝐹𝑚𝑚 − 𝐹𝐹� )2 � � 𝑚𝑚 [𝑘𝑘; 𝛈𝛈] − 𝑆𝑆𝑆𝑆𝑆𝑆

� �𝑆𝑆𝑆𝑆𝑆𝑆 ����� [𝑘𝑘; 𝛈𝛈] � ⋅ (𝐹𝐹𝑚𝑚 − 𝐹𝐹� ) (4-8)
𝑚𝑚=1 𝑚𝑚=1

and

� [𝑘𝑘; 𝛈𝛈] = �����

𝐺𝐺𝐺𝐺𝐺𝐺 𝑆𝑆𝑆𝑆𝑆𝑆 [𝑘𝑘; 𝛈𝛈] − 𝐶𝐶̂ [𝑘𝑘; 𝛈𝛈] ∙ 𝐹𝐹� (4-9)
with:

𝑀𝑀 𝑀𝑀
1 1
𝑆𝑆𝑆𝑆𝑆𝑆 � 𝑚𝑚 [𝑘𝑘; 𝛈𝛈] and 𝐹𝐹� = � 𝐹𝐹𝑚𝑚
����� [𝑘𝑘; 𝛈𝛈] = � 𝑆𝑆𝑆𝑆𝑆𝑆 (4-10)
𝑀𝑀 𝑀𝑀
𝑚𝑚=1 𝑚𝑚=1

The optimal set of parameters 𝛈𝛈𝑜𝑜𝑜𝑜𝑜𝑜 can be found as a least-square-error problem

49
 𝑀𝑀 𝑘𝑘𝑚𝑚𝑚𝑚𝑚𝑚 −1

𝛈𝛈𝑜𝑜𝑜𝑜𝑜𝑜 = argmin � � � �𝐺𝐺𝐺𝐺𝐺𝐺 � 𝑚𝑚 [𝑘𝑘; 𝛈𝛈]�2 �

� [𝑘𝑘; 𝛈𝛈] + 𝐶𝐶̂ [𝑘𝑘; 𝛈𝛈] ∙ 𝐹𝐹𝑚𝑚 − 𝑆𝑆𝑆𝑆𝑆𝑆 (4-11)
𝛈𝛈
𝑚𝑚=1 𝑘𝑘=𝑘𝑘𝑚𝑚𝑚𝑚𝑚𝑚

� 𝑚𝑚 [𝑘𝑘; 𝛈𝛈] calculated

Equation (4-11) calculates 𝛈𝛈𝑜𝑜𝑜𝑜𝑜𝑜 by minimising the error between 𝑆𝑆𝑆𝑆𝑆𝑆
� [𝑘𝑘; 𝛈𝛈] obtained
from Equation (4-7), and the equivalent STE calculated based on 𝐶𝐶̂ [𝑘𝑘; 𝛈𝛈] and 𝐺𝐺𝐺𝐺𝐺𝐺
� 𝑚𝑚 , 𝐶𝐶̂ and 𝐺𝐺𝐺𝐺𝐺𝐺
from Equations (4-8) to (4-10). It is important to highlight that 𝑆𝑆𝑆𝑆𝑆𝑆 � are all estimates
based on measured 𝐷𝐷𝐷𝐷𝐷𝐷𝑚𝑚 [𝑘𝑘] and the transfer function 𝐻𝐻(𝑘𝑘𝑠𝑠𝑚𝑚 ; 𝛈𝛈), which is dependent on the
parameter 𝛈𝛈. Therefore, while actual STE, compliance and GTE are independent of 𝐻𝐻, all their
estimates have a strong dependence on the transfer function and its parameter 𝛈𝛈 as per Equations
(4-7) to (4-10).

Finally, by substituting 𝛈𝛈𝑜𝑜𝑜𝑜𝑜𝑜 into Equations (4-8) to (4-10), the most likely compliance
𝐶𝐶̂ [𝑘𝑘] = 𝐶𝐶̂ �𝑘𝑘; 𝛈𝛈𝑜𝑜𝑜𝑜𝑜𝑜 � and profile error 𝐺𝐺𝐺𝐺𝐺𝐺 � �𝑘𝑘; 𝛈𝛈𝑜𝑜𝑜𝑜𝑜𝑜 � can be obtained. Transforming
� [𝑘𝑘] = 𝐺𝐺𝐺𝐺𝐺𝐺

𝐶𝐶̂ [𝑘𝑘] and 𝐺𝐺𝐺𝐺𝐺𝐺

� [𝑘𝑘] back into the angular domain through an inverse discrete Fourier Transform
� (𝑛𝑛Δ𝜃𝜃)
(DFT) will result in an estimate of the compliance and GTE waveforms 𝑐𝑐̂ (𝑛𝑛Δ𝜃𝜃) and 𝐺𝐺𝐺𝐺𝐺𝐺
over one shaft revolution. In order to eliminate small numerical errors, it is recommended to force
conjugate symmetry in 𝐶𝐶̂ [𝑘𝑘] and 𝐺𝐺𝐺𝐺𝐺𝐺
� [𝑘𝑘] before the inverse DFT, and therefore obtain real
angular domain signals.

It is worth highlighting that the order-domain regression performed in Equations (4-8)

� (𝑛𝑛Δ𝜃𝜃) , is
� [𝑘𝑘] → 𝐺𝐺𝐺𝐺𝐺𝐺
and (4-9), combined with the inverse DFTs 𝐶𝐶̂ [𝑘𝑘] → 𝑐𝑐̂ (𝑛𝑛Δ𝜃𝜃) and 𝐺𝐺𝐺𝐺𝐺𝐺
equivalent to a series of linear regressions performed directly in the angular domain (one for each
instant 𝑛𝑛Δ𝜃𝜃). This is because of the linearity property of the DFT and its inverse, combined with
the fact that regression only includes linear operations.

It is also worth mentioning that if the methodology is applied to an original DTE signal
with an integer number of hunting-tooth periods, the encoder-accuracy errors would be averaged
out in the estimated compliance, leaving these errors only in the estimated GTE. In terms of crack
severity assessment, these errors do not interfere with the results, which only take into account
the estimated compliance but not GTE. If an estimated GTE that is unbiased by these errors is to
be obtained, absolute TE [9] (also discussed in Chapter 6) should be exploited instead of the
conventional TE used in this Chapter.

50
4.3 Results and Discussions

The results obtained using the data at lower speed (up to 10 Hz and for a large crack) are
presented first, and discussed in detail to show the application of the methodology described in
section 4.2. The following subsection discusses the extension of the technique to higher speeds.
All results of estimated compliance in this section will be further applied in Chapter 5 for the
investigation of crack severity. While this chapter focuses on the removal of transfer function and
the separation of profile error and tooth flexibility, some of the final results of crack-severity
estimation is briefly mentioned to discuss the accuracy of the transfer-function removal procedure.

4.3.1 Detailed results on data up to 10 Hz

DTE signals were obtained at different speeds and loads (2-10 Hz, 5-20 Nm for this
subsection), using the pre-processing methodology described in subsection 4.2.1. Since the crack
was present on the pinion, the input shaft was used as reference shaft for order-tracking and
analysis. The order-tracking of the shaft rotation signals was done using number of samples per
gearmesh 𝑁𝑁=100, thus resulting in DTE signals with 2,700 samples per pinion revolution. The
pass-band for DTE was set to 3-149 orders (limits included) of the input shaft so as to include the
first five gearmesh harmonics. Figure 4-4 shows the DTE signals at different speeds and loads for
the large crack.

51
 Figure 4-4. DTE signals at different speeds and loads after synchronous averaging w.r.t input
shaft and band-pass filtering for input speed of (a) 2 Hz, (b) 5 Hz, and (c) 10 Hz.

The next stage of processing involves the removal of transfer function effects in the DTE
signals to obtain equivalent STE signals, as described in subsection 4.2.2. The DTE signals were
first plotted in the order domain (Figure 4-5) for a preliminary analysis to qualitatively initialise
the number and rough location of poles and zeros. Note that Figure 4-5 shows DTE signals before
bandpass filtering, to allow a clearer analysis of the order domain. No clear zero was identified,
and only one pole seemed to dominate the transfer function in this range at about 600 Hz
(approximately the 300th, 120th and 60th input shaft order for input speeds of 2 Hz, 5 Hz and 10
Hz, respectively). A second pole is possibly affecting the very last part of the 10 Hz order

52
spectrum, just above the 300th harmonic of the shaft, but is neglected for simplicity in this
subsection. The resonance initialisation was qualitatively done based on the general trend of the
spectrum including both the base noise level and shaft harmonics protruding from it.

Figure 4-5. Order domain of DTE signals at different speeds and loads for (a) 2 Hz, (b) 5 Hz,
and (c) 10 Hz.

The OMA approach of subection 4.2.2 was then applied to accurately estimate the
location (in Hz) and damping ratio of the pole. Starting from an observed initial estimate of natural
frequency (600 Hz) and damping ratio (1%), a gradient descent optimisation was applied to solve
the least-squares problem of Equation (4-11) and refine the estimation of the parameters. The pole
was correctly estimated at a frequency of 573 Hz with a damping ratio of 1.35%, both almost
perfectly aligned with the results obtained in [100], which used a different OMA technique for
the analysis of TE in healthy gears on the same test rig.

With the identification of the parameters of the transfer function, equivalent STE was
obtained for all tests shown in Figure 4-5. An example of the reconstruction of STE is shown in
Figure 4-6 for the test at 10 Hz and 9.9 Nm. It can be seen that high-frequency fluctuations of
DTE due to the resonance at 573 Hz have been successfully removed.

53
 Figure 4-6. An example of the removal of transfer function from the DTE signal at 10 Hz and
9.9 Nm, (a) DTE, fitted transfer function and equivalent STE in the order spectrum and (b) DTE
and equivalent STE in angular domain (one shaft revolution).

The effectiveness of transfer function removal was further evaluated by comparing

equivalent STE signals obtained at different speeds but under similar loads. The example of
Figure 4-7 shows how all equivalent STE curves reconstructed at about 10 Nm are very similar
in both shape and amplitude. The fact that the same input (equivalent STE) is reconstructed almost
identically from different responses (DTE at different speeds) is proof of the successful
application of the OMA-based removal of the transfer function 𝐻𝐻(𝑓𝑓), effectively inverting the
relationship of Equation (4-2).

54
 Figure 4-7. Original DTE (top) and estimated equivalent STE (bottom) for all speeds at ~10
Nm.

The STEs reconstructed in Figure 4-7 however still do not show clear symptoms of the
crack, due to the fact that profile errors are superimposed on the deflections. In order to facilitate
the identification of the cracked tooth, and, most importantly, the estimation of its severity, profile
and deflection effects must be separated. After obtaining equivalent STE at different loads, this
separation of meshing compliance and GTE can be achieved based on the linear regression
described in subsection 4.2.2. The linear regression is performed in the order domain to improve
computational efficiency within the optimisation loop (see example in Figure 4-8), but it is in all
respects equivalent to an angular-domain regression (Figure 4-9) as mentioned in section 4.2.2.

55
Figure 4-8. Order-domain regression analysis to estimate compliance and GTE using equivalent
STE signals.

Figure 4-9. Angular-domain regression analysis equivalent to the order-domain approach.

56
 The estimated compliance and GTE are shown in Figure 4-10. The once-per-rev
phenomenon due to the cracked tooth can be clearly observed from the estimated compliance in
Figure 4-10 (a), as well as from GTE (b), which shows the plastic deformation effect. The plastic
deformation appears to be a reduction in tooth spacing – a ‘closing crack’– as shown by its
opposite direction vs crack-compliance. This has been explained for similar tests in ref. [99] as a
possible result of the release of residual stresses from hardening when the slot was machined. It
should be noted that in the case of a real crack, the opposite effect is to be expected, where the
crack is permanently open when unloaded, due to the plastic deformation at the crack tip as part
of its development [38].

Figure 4-10. (a) Estimated compliance and (b) Estimated GTE.

4.3.2 Extension of the methodology to higher speeds

The extension of the result to operating speeds of 15 and 20 Hz requires the
consideration of a larger frequency range of interest.

Figure 4-11 shows how the second pole at about 3,250 Hz becomes important when
considering speeds above 10 Hz, and this resonance must be included in the estimation of STE.
A zero is not clearly visible, and the transfer function fitting will therefore only include two poles.

57
 Figure 4-11. Order spectra of DTE for operating speeds of 10, 15 and 20 Hz.

The optimisation is performed as in the previous subsection, with an initialisation of

natural frequencies at 600 and 3,250 Hz and damping ratios of 1%. Figure 4-12 shows the result
of the fitting and the removal of the FRF for the 20 Hz – 9.9 Nm case. The two resonances were
identified at 575 Hz and 3,250 Hz, with damping ratios of 0.81% and 1.81%, respectively. The
position of the first resonance confirms the results obtained in this work and [100] with a single-
resonance transfer function, whereas the second resonance was required only in this further
analysis given the extended frequency range of interest (from 10 to 20 Hz max. input shaft speed).

By changing the initialisation and obtaining different fitted frequencies for the second
pole, it was found that its position has a significantly smaller impact on the overall STE
reconstruction, due to its much smaller residue and the fact that it is outside the frequency range
of interest (i.e., above 5.5 times the gearmesh) even at the highest speed. That is probably why it
was observed that, as long as the initialisation frequency of the pole was within the 2,700-3,700
Hz range, the optimisation algorithm did not update its position significantly while keeping high
estimation accuracy, and damping was rather tuned in order to achieve the correct effect on the
frequency range of interest.

58
Figure 4-12. An example of the removal of transfer function from the DTE signal at 20 Hz and
9.9 Nm, (a) DTE, fitted FRF and equivalent STE in the order spectrum and (b) DTE and
equivalent STE in angular domain (one shaft revolution).

Applying the same steps used in the previous section, we obtain the estimated compliance
and GTE shown in Figure 4-13. The GTE and compliance curves are very similar to those
estimated using only low-speed tests (see Figure 4-10), and it is presented later in Chapter 5 that
the final result shows an even superior estimation of the crack depth (to be presented in Figure
5-8 later), consistent with the use of more data points. The balance between using more data points
and requiring a more complex transfer function depends on each gearbox configuration, and
specifically on the ratio between operating speed range and location of the first resonances.

59
 Figure 4-13. (a) Estimated compliance and (b) Estimated GTE considering higher-speed data
(including speeds of 2, 5, 10, 15 and 20 Hz).

A trial with a single pole (as in section 4.3.1) resulted in a slightly lower accuracy of
crack estimation (53% crack depth vs 52% in Figure 5-8), suggesting dominant poles have the
most significant impact on the procedure. Nonetheless, this is very case-dependent, and it is
suggested to include at least one out-of-band pole if its presence is clear from the trend in the
spectrum.

A further test was done adding to the two poles initialised at 600 and 3250 Hz a third pole
initialised at 820 Hz and a zero at 750 Hz. This was done following a small spectral perturbance
observed around those frequencies which could be interpreted as a possible pole-zero combination.
The position of the previous poles did not change significantly (i.e., 569 and 3249 Hz), the
additional pole was updated to 806 Hz and the zero to 758 Hz. The accuracy of the estimation
improved slightly, with similar GTE and compliance waveforms and an estimated crack depth of
50%. This confirms that the identification of dominant poles and the correct slope of the transfer
function (difference in number of poles and zeros) is sufficient for a good estimation of
compliance and GTE. Non-dominant poles will in fact only affect the transfer function locally,
and thus bias few harmonics in DTE. These biases are likely reduced in the averaging procedure
embedded in the regression-based estimation of compliance and GTE.

60
4.4 Additional results

This section shows additional results that might be of interest to readers.

Figure 4-14 shows the comparison of GTE estimated with the proposed methodology and
that measured directly based on low speed/low load tests. This comparison is a further validation
of the proposed methodology, since it shows the effectiveness of the GTE estimation.

Figure 4-14. Comparison between GTE estimated from the proposed methodology and GTE
measured directly at low speed and low load, for all crack depths.

61
Chapter 5 Use of TE in gear crack severity
assessment

With the successful estimation of meshing compliance in Chapter 4, it is then possible to

assess crack severity. This chapter aims to achieve Objective 3, by proposing a methodology that
reliably detects a cracked tooth, and estimates the crack depth by comparing the experimental TE
signal to theoretical models. The materials presented in this chapter is also part of the work
published in the journal paper 6 titled “Use of transmission error for a quantitative estimation of
root-crack severity in gears” [8].

5.1 Methodology

As discussed in Chapter 2, the main and most predictable effects of a crack are shaft-
synchronous changes in stiffness. Cracks have been shown to affect GTE as well, but this effect
is less understood, and GTE also includes the effects of wear and manufacturing deviations from
the involute. Hence, crack severity assessment should focus on the time-varying meshing stiffness
(TVMS) changes. In this study the time-varying meshing compliance (the reciprocal of TVMS)
will be used for simplicity.

The potential energy method can be used to model meshing compliance, both in healthy
condition and in the presence of a crack. A detailed discussion of the potential energy method and

6
Permission has been granted by co-authors.

62
implementation choices is beyond the scope of this thesis, and the proposed methodology is
expected to yield similar results with other models, e.g., FE method. A series of different
implementations are in fact available in the literature, and the methodology used in this chapter
follows the procedure described in ref. [15], interpreted and implemented for the experimental
setup described in section 3.5.2. This model undoubtedly includes a series of approximations, but
it has been selected because it is established and relatively simple to implement, considering that
the aim of this chapter is not the development or validation of a sophisticated model, but rather
the proposal of a TE-based crack severity estimation approach. It should be kept in mind that the
primary result of the method is to find the change in compliance of the cracked tooth, and
ascribing this to a “crack depth” is just a commonly used way of assessing the severity of such a
compliance reduction. In practical applications, the actual details of the crack causing a
compliance change will not be known, and will almost certainly not be the same as the model
used to assign the “crack depth”. Most (but not all) analytical crack simulation models assume
development of the crack along a straight line, and at a particular angle, and also that they can be
represented in two dimensions (i.e., exactly the same angle and depth across the whole width of
the tooth). The trajectory of the crack has a dominant influence on factors such as the effect of
the crack on the gear body stiffness, but that cannot be known in a practical case, and the only use
of “crack depth” in this chapter is as a nominal “equivalent crack depth” as an indicator of severity.
A strong argument can be made that compliance change is a more relevant indicator of severity
than crack depth itself. In the rest of the chapter, “crack depth” is taken to mean the equivalent
depth of the simulated idealised crack as modelled using the method described in the chapter. In
a practical case, the change in compliance might for example be calibrated against an FE model,
based on previous experience with crack development in particular gears, or with specific
geometric features such as rim thickness.

In the potential energy method, the overall compliance is obtained as a contribution of

tooth-bending, tooth-shear, tooth-axial-compressive, fillet-foundation and Hertzian-contact
compliances (see section 5.3 for an investigation of nonlinear Hertzian effect on total tooth-pair
compliance). Each term is time-varying, and more specifically dependent on the instantaneous
rotation angle of the gears, which influences the number of tooth-pairs in contact (one or two in
most spur gear applications) and the location of the contact point(s). As per the modelling
assumptions of ref. [15], the presence of a crack increases the bending and shear components of
compliance for the faulty tooth.

This methodology first automatically detects the crack and separates it from the healthy
teeth in the meshing-compliance signal estimated from the procedure in Chapter 4. Then, the

63
estimated compliance is compared with theoretical compliance calculated using the potential
energy method, in order to assess crack severity (i.e., crack depth). The procedure is illustrated in
Figure 5-1.

Figure 5-1. Procedure for detection and severity assessment of crack.

5.1.1 Separation of cracked tooth from healthy teeth

This step aims to automatically distinguish the cracked tooth from other healthy teeth. In
order to do this, the portion of the estimated compliance corresponding to each tooth needs to be
extracted from 𝑐𝑐̂ [𝑛𝑛Δ𝜃𝜃]. Each tooth starts its contact at the beginning of a double-tooth-pair contact.
To identify the location of such a point within a gearmesh period of 𝑐𝑐̂ [𝑛𝑛Δ𝜃𝜃], the synchronous
average of the estimated compliance (with respect to the gearmesh period) is first obtained as:

𝑍𝑍−1
1
𝑐𝑐̅(𝑛𝑛Δ𝜃𝜃) = � 𝑐𝑐̂ (𝑛𝑛Δ𝜃𝜃 + 𝑧𝑧𝑧𝑧) (5-1)
𝑍𝑍
𝑧𝑧=0

with 𝑛𝑛 = 0, … , 𝑁𝑁 − 1 and 𝑍𝑍 equal to the number of teeth on the cracked gear. This estimated
compliance is then compared to the theoretical value 𝑐𝑐0 (𝑛𝑛Δ𝜃𝜃) obtained from the potential energy
method [15] (as described at the beginning of section 5.1) with healthy teeth and initial angle
corresponding to the starting point of the double-tooth-pair contact. It is worth mentioning that
since the theoretical compliance is calculated from a quasistatic model, this comparison must only
be performed after an equivalent STE has been obtained from DTE by removing the dynamic
(transfer function) influences using the OMA procedure described in the previous section. The
theoretical compliance 𝑐𝑐0 (𝑛𝑛Δ𝜃𝜃) is band-pass filtered in the same shaft harmonic range
[𝑘𝑘𝑚𝑚𝑚𝑚𝑚𝑚 , 𝑘𝑘𝑚𝑚𝑚𝑚𝑚𝑚 ] before use. Figure 5-2 shows an example of theoretical healthy meshing compliance
for the entire duration of a tooth contact period, based on the parameters of the experimental setup
described in section 3.5.1. The reason for this band-pass filter is to accommodate the fact that a
real compliance signal would have a smoother transition when the contact between the two gears
changes from single-tooth-pair to double-tooth-pair, or vice versa.

64
Figure 5-2. Theoretical tooth-meshing compliance (healthy) before filtering (top); after filtering
(bottom).

Assuming that a single cracked tooth would not affect the synchronous average 𝑐𝑐̅(𝑛𝑛Δ𝜃𝜃)
significantly, the maximum of a circular cross-correlation between 𝑐𝑐̅(𝑛𝑛Δ𝜃𝜃) and the corresponding
theoretical healthy compliance 𝑐𝑐0 (𝑛𝑛Δ𝜃𝜃) allows identification of the location of the starting point
𝑛𝑛𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 of the double-tooth-pair contact:

𝑁𝑁−1
1 𝑛𝑛 − 𝑛𝑛′
𝑛𝑛𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 = argmax � � 𝑐𝑐0 (𝑛𝑛Δ𝜃𝜃) ⋅ 𝑐𝑐̅ ��𝑛𝑛 − 𝑛𝑛′ − 𝑁𝑁 ⋅ � �� Δ𝜃𝜃�� (5-2)
𝑛𝑛′ =0,1,…,𝑁𝑁−1 𝑁𝑁 𝑁𝑁
𝑛𝑛=0

where ⌊ ⌋ represents the round-down operation required by the circular-indexing. Since 𝑐𝑐̅(𝑛𝑛Δ𝜃𝜃)
and 𝑐𝑐̂ (𝑛𝑛Δ𝜃𝜃) have the same initial phase, a circular rephasing of 𝑐𝑐̂ (𝑛𝑛Δ𝜃𝜃) is possible, so that its first
sample is aligned with the starting point of the double-tooth-pair contact:

𝑛𝑛 − 𝑛𝑛𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖
𝑐𝑐̂𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 (𝑛𝑛Δ𝜃𝜃) = 𝑐𝑐̂ ��𝑛𝑛 − 𝑛𝑛𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 − 𝑁𝑁𝑁𝑁 ⋅ � �� Δ𝜃𝜃� (5-3)
𝑁𝑁𝑁𝑁
where the term 𝑁𝑁𝑁𝑁 ⋅ ⌊(𝑛𝑛 − 𝑛𝑛𝑖𝑖𝑛𝑛𝑛𝑛𝑛𝑛 )⁄𝑁𝑁𝑁𝑁⌋ is again inserted just to ensure a circular rephasing. After
this initial rephasing, the tooth-specific portions of the estimated compliance 𝑐𝑐̂𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 (𝑛𝑛Δ𝜃𝜃) are
extracted into a series 𝑧𝑧 = 0, … , 𝑍𝑍 − 1 of shorter signals

𝑛𝑛Δ𝜃𝜃 + 𝑧𝑧𝑧𝑧
𝑐𝑐̂𝑧𝑧 (𝑛𝑛Δ𝜃𝜃) = 𝑐𝑐̂𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑎𝑎𝑠𝑠𝑠𝑠𝑠𝑠 �𝑛𝑛Δ𝜃𝜃 + 𝑧𝑧𝑧𝑧 − 𝑁𝑁𝑁𝑁 ⋅ � �� with 𝑛𝑛 = 0, … , ⌊𝜀𝜀𝜀𝜀⌋ − 1 (5-4)
𝑁𝑁𝑁𝑁

65
where 𝜀𝜀 is the gear contact ratio (i.e., the ratio between the duration of a tooth contact and a
gearmesh period), and the term 𝑁𝑁𝑁𝑁 ⋅ ⌊(𝑛𝑛Δ𝜃𝜃 + 𝑧𝑧𝑧𝑧)⁄𝑁𝑁𝑁𝑁⌋ is added for circular sampling of the last
tooth.

Then, the following measure 𝛿𝛿𝑧𝑧 of deviation from the “average tooth compliance” is
computed for each tooth:

⌊𝜀𝜀𝜀𝜀⌋−1 𝑍𝑍−1 2
1
𝛿𝛿𝑧𝑧 = � �𝑐𝑐̂𝑧𝑧 (𝑛𝑛Δ𝜃𝜃) − � 𝑐𝑐̂𝑧𝑧′ (𝑛𝑛Δ𝜃𝜃)� (5-5)
𝑍𝑍 ′
𝑛𝑛=0 𝑧𝑧 =0

The maximum value of 𝛿𝛿𝑧𝑧 allows the identification of 𝑧𝑧𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 (cracked tooth number), and
the corresponding compliance. The average of the identified cracked-tooth compliance is then
removed from 𝑐𝑐̂𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 (𝑛𝑛Δ𝜃𝜃) before further processing, since the mean is often affected by the
remaining low-shaft harmonics even after the band-pass filtering of DTE:

⌊𝜀𝜀𝜀𝜀⌋−1
1
0-mean (𝑛𝑛Δ𝜃𝜃)
𝑐𝑐̂𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 = 𝑐𝑐̂𝑧𝑧𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 (𝑛𝑛Δ𝜃𝜃) − � 𝑐𝑐̂𝑧𝑧𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 (𝑛𝑛′Δ𝜃𝜃) (5-6)
⌊𝜀𝜀𝜀𝜀⌋ ′
𝑛𝑛 =0

5.1.2 Model-based estimation of crack size

The potential energy method [15] (as described at the beginning of section 5.1) was used
to calculate the theoretical tooth-meshing compliance 𝑐𝑐𝑑𝑑 (𝑛𝑛Δ𝜃𝜃) for different crack scenarios 𝑑𝑑 =
0,1, … , 𝐷𝐷 − 1, with the desired crack-depth resolution. An example of the effect of a crack on the
theoretical compliance is shown in Figure 5-3, obtained using the actual parameters of the
experimental setup discussed in section 3.5.1.

66
Figure 5-3. Theoretical meshing compliance at different crack levels, before band-pass filtering.

Theoretical compliance curves were band-passed filtered in the input-shaft-harmonic

range of [𝑘𝑘𝑚𝑚𝑚𝑚𝑚𝑚 , 𝑘𝑘𝑚𝑚𝑚𝑚𝑚𝑚 ]. Then, a tooth-specific portion for each crack severity is extracted and its
0-mean (𝑛𝑛Δ𝜃𝜃):
mean is removed, as done for the estimated compliance 𝑐𝑐̂𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐

⌊𝜀𝜀𝜀𝜀⌋−1
1
𝑐𝑐𝑑𝑑0-mean (𝑛𝑛Δ𝜃𝜃) = 𝑐𝑐𝑑𝑑 (𝑛𝑛Δ𝜃𝜃) − � 𝑐𝑐𝑑𝑑 (𝑛𝑛′Δ𝜃𝜃) (5-7)
⌊𝜀𝜀𝜀𝜀⌋ ′
𝑛𝑛 =0

The crack size, 𝑑𝑑𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 , is estimated based on the best match (minimum RMS error)
between 𝑐𝑐𝑑𝑑0-mean (𝑛𝑛Δ𝜃𝜃) and 𝑐𝑐̂𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐
0−mean (𝑛𝑛Δ𝜃𝜃),
over the entire tooth

⎧ ⌊𝜀𝜀𝜀𝜀⌋−1 ⎫
⎪ 1 2⎪
𝑑𝑑𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 = argmin � 0-mean 0-mean
� �𝑐𝑐̂𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 (𝑛𝑛Δ𝜃𝜃) − 𝑐𝑐𝑑𝑑 (𝑛𝑛Δ𝜃𝜃)� (5-8)
𝑑𝑑 ⎨ ⌊𝜀𝜀𝜀𝜀⌋ 𝑛𝑛=0 ⎬
⎪ ⎪
⎩ ⎭

67
5.2 Results and discussion

5.2.1 Application to large crack (up to 10 Hz)

Using the meshing compliance estimated in Figure 4-10, the automated identification of
the cracked tooth starts with the rephasing and segmentation of compliance into single-tooth
records, as described in subsection 5.1.1. Figure 5-4 shows the rephased estimated compliance
signal together with the theoretical reference. The theoretical compliance was bandpass-filtered
in the same way (3rd to 149th input shaft harmonics, inclusive) as the TE-based compliance to
enable an unbiased comparison. It is important to highlight that no prior knowledge of the
theoretical gearmesh stiffness was used in the calculation of the experimental compliance, so the
comparison supports the effectiveness of each step of the procedure applied so far. The
comparable values of theoretical and estimated compliances for healthy teeth across all tests
suggests a sufficiently accurate model implementation and TE signal processing.

Figure 5-4. Estimated compliance, with the same starting phase as theoretical compliance.

The cracked tooth is identified as that showing the highest deviation from the “average”
tooth behaviour as described in subsection 5.1.1. Figure 5-5 (top) shows the comparison of the
estimated compliance with its gearmesh-synchronous average on the top, clearly highlighting the
part affected by the crack. Figure 5-5 (bottom) shows how the tooth-deviation measure 𝛿𝛿𝑧𝑧 allows
the automated detection of the 7th tooth as that with the crack. Tooth-contact periods 6 and 8 are
also partially affected, as expected given the contact ratio of 1.68. It is interesting to note that this

68
cross-tooth effect is stronger on tooth 8, when the tip of the cracked tooth is engaged, as expected
from the potential energy method predictions shown in Figure 5-2. The portion of the estimated
compliance corresponding to the cracked-tooth-contact period is then extracted for further
processing.

Figure 5-5. (a) Estimated compliance vs healthy reference (GM-synchronous average); (b) error
between the two.

The next phase of processing involves the comparison between the estimated compliance
with the theoretical cracked-tooth compliance profiles calculated from the potential energy
method, as shown in subsection 5.1.2. Figure 5-6 shows a comparison between the estimated
compliance (cracked part) and the theoretical compliance for different crack sizes. Note that the
average of the theoretical compliance signals was removed to allow for an unbiased comparison
with the zero-mean estimated compliance. The estimated and theoretical compliances follow a
similar shape, but the estimated compliance seems to have a broader single-contact than expected,
possibly explained as a “smearing” effect due to biases in the FRF removal or due to unmodelled
contact deformation. The latter could include the unmodelled crack-induced effects of plastic
deformations and increased compliance on the duration of the cracked-tooth contact (effective
contact ratio), which would also likely change with different load cases.

69
 Figure 5-6. Comparison between the estimated compliance from the large-crack experimental
signals and the theoretical compliances simulated by the model for different crack sizes. The
best-matching simulated compliance will indicate the most likely crack depth.

The RMS of the difference between estimated and theoretical compliance for different
theoretical crack depths is shown in Figure 5-7. As per Equation (5-8), the minimum RMS-error
indicates the most likely crack severity, which in this case is 54% vs the nominal 50%. It is
important to mention that, given the large thickness of the artificial slot used to simulate the crack
and its rounded tip, the reference 50% is to be taken as an indication, rather than a precise measure
of the equivalent crack depth.

70
 Figure 5-7. Crack severity assessment for large crack (nominal depth = 50%). RMS of the
difference between estimated compliance and modelled results with different crack sizes.

5.2.2 Application to higher speed data

The previous section presented the results for the detection and severity assessment of
crack in detail, with data up to operating speed of 10 Hz. This section shows crack estimation
results from the compliance in Figure 4-13, which is obtained considering higher-speed data. With
a more accurate removal of transfer function compared to only data up to 10 Hz, an even superior
estimation of crack depth (52%) is achieved.

71
 Figure 5-8. Crack severity assessment for large crack (nominal depth = 50%). RMS of
the difference between estimated compliance (considering higher-speed data) and modelled
results with different crack sizes.

5.2.3 Application to smaller cracks

This section demonstrates the application of the methodology to the small and medium
cracks shown in Figure 3-9. In this section all tests with speed 2-20 Hz and load 5-20 Nm were
used and corrected with a two-pole FRF, as in subsection 4.3.2. Some processed results are shown
in this section for relevant discussions, while additional results have been included in section 5.3.
A trial with 3 poles and 1 zero yielded overall comparable results and is therefore not included
for the sake of brevity.

For the medium crack, a clear once-per-rev phenomenon can still be observed in the
estimated compliance (due to reduction in stiffness) and estimated GTE (plastic deformation from
a ‘closing crack’), as shown in Figure 5-9, albeit with lower magnitude than the large crack. As
shown in Figure 5-10, the crack size was estimated at 41%, with an error of only 5% of the root
section. This confirms the robustness of the methodology in tackling the assessment of gear crack
severity, even if the error seems to increase with smaller cracks. This is likely due to the lower
“signal-to-noise ratio”, understood as the effect of the crack on compliance vs other sources of its
variability (e.g., noise, gear eccentricity, misalignment and errors of the compliance estimation
procedure).

72
 Figure 5-9. (a) Estimated compliance and (b) Estimated GTE for the medium crack.

Figure 5-10. Crack severity assessment for medium crack (nominal depth = 36%). RMS of the
difference between estimated compliance and modelled results with different crack sizes.

When the methodology was applied to the small crack, it correctly identified the broken
tooth, but with a very small margin vs the other teeth, as shown in Figure 5-11.

73
 Figure 5-11. (a) Estimated compliance vs healthy reference (GM-synchronous average); (b)
error between the two for the small crack.

The severity assessment component surprisingly provides a good estimate of the crack
severity: 18% vs the nominal 19%. However, given the flatness of the RMS curve of Figure 5-12
around the minimum, the confidence in this estimate is much lower than for the previous cracks.
It is expected that general results for small cracks would often be less accurate than this
particularly favourable example. The presence of a discontinuity in the zoom of the RMS curve
near 0% is caused by a small model discrepancy, which gives a small increase in stiffness for very
small cracks.

74
 Figure 5-12. Crack severity assessment for small crack (nominal depth = 19%). RMS of the
difference between estimated compliance and modelled results with different crack sizes, with a
zoom of the same plot in the red box.

The lower confidence provided by the technique for small faults is however not
particularly problematic, since in these cases early-detection-focused techniques are preferable to
severity assessment, which becomes more important only when the crack is closer to failure and
thus also likely to develop faster. The progressive flattening (loss of curvature around the
minimum) of the RMS-error curve for smaller cracks could be used as a measure of confidence
in the result, which conversely is seen to greatly improve as the crack approaches failure.

It is interesting to note that the estimation of the transfer function across different crack-
severity tests was fairly consistent, as shown in Table 5-1, confirming that the fitting could be
performed only once at the beginning of the crack evolution. If there is an expectation that the
FRF might change over time due to ageing of the system, it is reasonable to think that at least the
manual analysis of the FRF and the initialisation of poles and zeros (the only manual component
of this procedure) can be executed only once at installation, whereas the fine-tuning and updating
can be left to the automated optimisation procedure which can be run at regular intervals in time.

75
 Table 5-1. FRF identification based on different tests.

Large crack Medium crack Small crack

1st natural frequency (Hz) 575 578 580
2nd natural frequency (Hz) 3250 3250 3250
1st mode damping ratio (%) 0.82 0.58 0.88
2nd mode damping ratio (%) 1.64 1.87 1.92

5.2.4 Further validation of the proposed methodology

A further validation of the overall procedure was conducted by excluding one speed test
from the data used in sections 4.3.2 and 5.2.3 (i.e., 2-20 Hz and 5-20 Nm). The results obtained
using the two-pole FRF shown in Table 5-2 confirm the observations of the previous sections,
with the small crack showing the highest variability in the severity estimation results. Of particular
interest is the case excluding the 2 Hz tests, which are the least affected by the transfer function
and could be considered as almost identical to STE. The fact that similar accuracy was obtained
confirms the quality of the OMA-based removal of the transfer function.

Table 5-2. Estimated crack sizes obtained excluding a single speed from the list of TE
experimental signals.

Excluded Small crack Medium Large crack

speed (19%) crack (36%) (50%)
2 Hz 21% 44% 51%
5 Hz 18% 41% 53%
10 Hz 12% 39% 52%
15 Hz 0% 38% 50%
20 Hz 27% 44% 55%

The 0%-estimate obtained excluding the 15-Hz test is mostly caused by the small model
discrepancy already observed in Figure 5-12.

A final validation was then conducted by adding to the data used in sections 5.2.2 and
5.2.3 the 1-Hz speed tests, affected by a low-frequency resonance (at about 30 Hz) whose removal
was challenging since it only affected this particular speed. The results shown in Table 5-3 show
that the estimation is still acceptable, especially for medium and large cracks despite the
“corrupted” additional TE signals. This robustness is likely due to the regression procedure which
does reduce the impact of outliers on the overall result.

76
 Table 5-3. Estimated crack sizes obtained including the 1 Hz TE signals.

Small crack Medium Large crack FRF used

(19%) crack (50%)
(36%)
All speeds including 1 Hz 25% 44% 53% 2-pole
Speeds in the range 1-10 Hz 31% 46% 55% 1-pole

5.3 Additional results and discussions

This section shows additional results and discussions that might be of interest to readers.

In the body of the chapter, the complete processed results have been demonstrated for the
large crack, but not for small and medium cracks. These additional results are shown in Figure
5-13 and Figure 5-14 (medium crack) and Figure 5-15 and Figure 5-16 (small crack).

Medium crack

Figure 5-13. Estimated compliance vs healthy reference, i.e., GM-synchronous average (top);
error between the two (bottom), for the medium crack.

77
 Figure 5-14. Comparison between the estimated compliance from the medium-crack
experimental signals and the theoretical compliances simulated by the model for different crack
sizes. The best-matching simulated compliance will indicate the most likely crack depth.

Small crack

Figure 5-15. (a) Estimated compliance and (b) Estimated GTE, for the small crack.

78
Figure 5-16. Comparison between the estimated compliance from the small-crack experimental
signals and the theoretical compliances simulated by the model for different crack sizes. The
best-matching simulated compliance will indicate the most likely crack depth.

Comparison with vibration measurements

Figure 5-17 shows the synchronous average of vibration (acceleration) measurements

obtained from the same crack tests used in this chapter, i.e., for small, medium and large cracks
(from top to bottom). Measurements obtained at input speed of 10 Hz and input load of 20 Nm
are presented on the left column, whereas those obtained at input speed of 20 Hz are on the right
column. It can be observed from Figure 5-17 that vibration suffers from a more complex transfer
path compared to TE signals, leading to the masking of the components of interest such as the
gearmesh and the symptoms of crack. It is still possible to observe the effect of the large crack at
20 Hz and 20 Nm, but not at 10 Hz and 20 Nm. For other crack sizes, the signals do not seem to
show clear indication of the crack. Further processing of these signals could reveal the crack, but
this would require a more complicated OMA-removal procedure or sophisticated dynamic models,
due to the complexity of the transfer path.

79
Figure 5-17. Synchronous average of vibration (acceleration) measurements with respect to
input shaft, for small, medium and large cracks (from top to bottom). Left column:
measurements at 10 Hz and 20 Nm; right column: measurements at 20 Hz and 20 Nm.

80
Investigation of nonlinear Hertzian effect on total tooth-pair compliance

In this thesis, a constant value for the Hertzian stiffness is used, according to the same
choice adopted in [15], and based on a common linearisation of the more complex and non-linear
Hertzian effects. According to established literature [41], the latter is approximately following a
power law where stiffness is proportional to 𝐹𝐹 0.1. This short section briefly justifies the choice of
approximating this mildly load-dependent relationship with a constant.

In the specific experimental case study of this thesis, the maximum load variation is 5 to
20 Nm. Therefore, for the same contact point, forces observed in the highest load case would be
4 times higher than those of the lowest load case. According to the aforementioned nonlinear
relationship, the resulting Hertzian stiffness calculated with the nonlinear formula will be 40.1 =
1.15 times higher in the 20 Nm case, i.e., +15% vs the 5 Nm case. In terms of compliance, this
means that considering a constant value for Hertzian compliance (for instance evaluated at 5 Nm),
1
the error committed at maximum load (20 Nm) against the fully nonlinear case is 1 − = 13%.
1.15

This percentage error is however only affecting the relatively minor Hertzian term of the TVMS,
which accounts for about 6% of the total tooth-pair compliance (see Figure 5-18 obtained with
the same model and parameters used in the previous sections of this chapter). Therefore, the
overall compliance error expected from the linearisation of Hertzian effects (constant Hertzian
stiffness) are in this case limited to about 1% of the total compliance (13% × 6% ≈ 1%).

Figure 5-18. Relative contribution of Hertzian compliance towards the total tooth-pair
compliance (healthy teeth with parameters as per the previous sections of this chapter).

81
 Despite this analysis being specific of this thesis, the use of standard gears (module design)
and the considerations over the typical load ranges are applicable to a vast series of practical cases,
thus rendering these considerations quite general. Even in cases where load ranges could change
significantly more, the nonlinear component of Hertzian stiffness can still be neglected if the most
extreme load cases are simply discarded.

82
Investigation on the effect of modelling inaccuracies on crack severity assessment

To investigate the effect of modelling inaccuracies on the robustness of the technique for
crack severity assessment, a sensitivity study is demonstrated here by deliberately applying an
error of 10% (increase) to the TVMS modelled using the potential energy method. This results in
1
a theoretical meshing compliance that is reduced by a factor of 1 − = 9.1%. The same severity
1.1

assessment procedure is conducted using this theoretical meshing compliance, and the results are
shown in Table 5-4.

While the results in Table 5-4 are less accurate compared to previous sections, the
estimated crack sizes are still reasonably acceptable, and the technique is able to clearly
differentiate the different crack sizes. It also seems that the large crack is less susceptible to
modelling inaccuracies, as the estimated crack size only shows little change when compared to
the more accurate result (52%) in Figure 5-8.

Table 5-4. Crack severity assessment using a TVMS model with 10% error.

Actual Crack Size Estimated Crack Size

Small (19%) 27%
Medium (36%) 47%
Large (50%) 56%

83
Chapter 6 Absolute TE: a simple new tool for
assessing gear wear

In this chapter, the concept of ‘absolute transmission error’ – including both average and
time-varying component – is proposed as an important diagnostic tool, and for the first time a
general methodology for its estimation is presented. The method uses an HTP phase reference,
obtained from tachometers or encoders on both the input and output shafts. The importance of the
DC component of TE is highlighted and shown to be a very powerful tool for assessing gear wear.
The chapter thus has two main aims. The first one is to introduce a new signal processing
methodology that can restore an HTP phase reference for encoder signals and enables the
calculation of absolute TE. The second aim is to show the application of this new methodology
in quantifying the DC component of TE signals, which indicates the combined average wear depth
of a tooth pair relative to a reference measurement, ideally from the gears in their unworn state.
This chapter describes the application of TE in a new perspective and the term “absolute
transmission error” is introduced. The significance and ability of absolute TE in gear wear severity
assessment are demonstrated with evidence from a dry gear wear experiment.

The chapter is organised as follows. Section 6.1 presents the procedure to calculate
conventional TE and its limitations for gear wear diagnostics. Section 6.2 conceptually explains
the proposed methodology in detail and provides a summary of processing steps to calculate
absolute TE. Section 6.3 applies the proposed methodology to calculate wear depth using encoder
measurements recorded at different speeds and loads, and the results are compared with mass

84
measurements of the collected wear particles as a validation. Section 6.3 also demonstrates the
possibility of using tachometer measurements to quantify wear depth, some important
considerations of the proposed methodology are discussed and the effectiveness of absolute TE
as a wear severity diagnostic tool is compared with vibration.

The work presented in this chapter is adapted from a published journal paper 7 titled
“Absolute transmission error: a simple new tool for assessing gear wear” [9].

6.1 Problem statement: mean component of TE overlooked

6.1.1 Calculation of conventional (zero mean) TE

To fully understand the value and details of the new procedure proposed in this chapter,
it is necessary to first discuss in detail how TE is conventionally calculated. The first step of the
conventional TE calculation is to obtain the phase-time maps for both shafts, i.e., 𝜃𝜃1 (𝑡𝑡) and 𝜃𝜃2 (𝑡𝑡),
which can be extracted using phase demodulation from the encoders mounted on the shafts [110].
The phase-time maps can also be obtained using a simple time-domain approach, e.g., ‘zero-
crossing method’ [59] by detecting the rising edge of the pulses in the encoder signals and using
cubic spline interpolation to calculate the phase at each time instant. It is often desired to have a
phase reference for both shafts, and thus a one-pulse-per-revolution tacho signal is required to
zero the phase time maps at a time instant when a tacho pulse rises. These phase-time maps are
essentially the torsional vibration of the shafts. Subtracting 𝜃𝜃1 (𝑡𝑡) (scaled with the gear ratio
𝑍𝑍1 ⁄𝑍𝑍2 ) from 𝜃𝜃2 (𝑡𝑡) results in a TE signal in terms of the phase of the driven shaft which includes
an arbitrary mean value. TE can also be scaled along the line of action by multiplying the phase
signals by the base radius of their respective gear before the subtraction. The mean component
(DC shift) is then set to zero because it does not represent anything, since the phase-time maps of
the driving and driven shafts, which are obtained from encoder signals, have arbitrary initial phase.
This results in a signal which represents only the fluctuations of TE around an unknown mean
value.

6.1.2 Limitations of conventional TE for wear analysis

Gear wear monitoring has mostly focused on the analysis of gearmesh harmonics since it
has been suggested that uniform wear across all teeth can alter the gearmesh harmonics, both in
vibration [111] and TE [97] signals. While the analysis of the variation of gearmesh harmonics in
TE could provide some indication of wear severity, a much more direct information is missing in

7
Permission has been granted by co-authors.

85
conventional TE measurements – the average wear depth, strictly related to the DC value of TE.
The analysis of gearmesh harmonics could reveal the characteristics of the tooth profile and shape,
but it is difficult to extract information about the average wear depth. Figure 6-1 shows the
synchronous averages of conventional (zero mean) TE signals with respect to the input shaft of a
spur gearbox with a 19-tooth pinion at two different wear stages – mild and severe cases. It can
be seen that there are significant changes in both the amplitude and phase of the signals as wear
progresses. However, due to the shape of the signal being different, it is difficult to identify any
clear phase shift.

Figure 6-1. Synchronous averages of conventional (zero mean) TE for one input shaft
revolution on a spur gearbox; (a) mild wear, (b) severe wear.

Figure 6-2 illustrates the presence of a combined wear depth resulting in a non-zero DC
value for TE (represented by the red line along the line of action). This component could not be
measured using the conventional TE measurements in which the mean of the signal was removed
before analysis, i.e., the TE signal varies around zero mean. A zero-mean TE signal thus neglects
the clearest information about wear depth. The absolute TE introduced in this chapter can thus be
thought of as a complete transmission error that includes the DC component, directly representing
the average loss of material on the tooth flank.

86
 Figure 6-2. Illustration of wear depth as absolute geometric transmission error.

However, without a phase reference in the fundamental period of the complete mesh of
two gears (HTP), this DC value cannot generally be measured since the worn and healthy TE
signals would show different initial phases. In particular, each TE signal would start with a
different combination of driving and driven gear teeth in contact, depending on the arbitrary
choice of the first tacho pulse on the reference shaft. The next section introduces a simple, yet
novel, methodology that allows restoring the HTP phase reference of TE signals measured at
different times (worn gears) with respect to an initial reference TE signal (healthy gears), enabling
the identification of the absolute TE.

6.2 Proposed methodology: hunting-tooth-period rephasing for absolute

transmission error

6.2.1 Conceptual methodology

This section describes the methodology to obtain absolute transmission error. As
explained in section 6.1.1, when encoder (or tacho) signals are arbitrarily recorded in time, the
starting phase of the phase-time maps of the two shafts is arbitrary, and so by convention is usually
set to zero when calculating TE. In order to overcome this limitation, some sort of additional
phase reference must be included. This can be achieved using the concept of a hunting tooth
period and a reference measurement, the latter typically from an unworn gearset. The
methodology consists of three main steps – order tracking, locating the HTP reference and
rephasing the signals according to the HTP reference, as illustrated in Figure 6-3. In the figure,
the algorithm is explained with a once-per-rev tacho example, and an unrealistic (but explicative)
gear ratio of 3:5 (i.e., 𝑍𝑍1 = 3, 𝑍𝑍2 = 5). The blue and red lines represent the tacho signals of the
driving and driven shafts, respectively. The algorithm aims to rephase the worn gear signals (a
pair of input and output tacho signals) with respect to equivalent healthy gear signals so that they
can be compared.

87
Figure 6-3. Graphical representation of hunting tooth period rephasing methodology.

88
 As a first step (Figure 6-3 (a)), order tracking is performed to fix an initial phase reference
for a shaft revolution, and to truncate the signal to a length of one HTP (important for the later
circular-shift operation).

Even though order tracking with respect to the driving gear ensures that the signals are
phased consistently with respect to the same gear, the angle of the driven gear at the beginning of
the signal is still arbritrarily shifted in the HTP. To obtain the DC value of transmission error,
both healthy and worn cases must start at the same point in the hunting tooth period so that the
signals are phased consistently with respect to both shafts. In order to achieve this, a shifting
procedure is required, which is demonstrated in Figure 6-3 (b). The worn gear signals are shifted
by increments corresponding to multiples of one driving shaft period (Δ1 ). Note that Δ2 and Δ𝐻𝐻𝐻𝐻𝑃𝑃
are the angles corresponding to the driven shaft period and the HTP. The shifting procedure is
performed (𝑍𝑍2 − 1) times, producing 𝑍𝑍2 differently rephased replicas of the signals, including
the original (unshifted) signals. One of these replicas must be very close in phase to the unworn
𝑘𝑘
reference signal, with a delay 𝑑𝑑+ which is the smallest among those of all the possible replicas
(marked with a green box in Figure 6-3 (b)), indicating the necessary rephasing of the signal.

Once the number of shifts (𝑘𝑘𝑠𝑠ℎ𝑖𝑖𝑖𝑖𝑖𝑖 ) required is identified, the final step is to rephase the
signals with a circular shift by a length of (𝑘𝑘𝑠𝑠ℎ𝑖𝑖𝑖𝑖𝑖𝑖 × Δ1 ) as shown in Figure 6-3 (c). In MATLAB,
this is done using the circshift function. After rephasing the signals, the starting phases of
the driven gear phase-time maps are directly comparable. If using only once-per-rev tacho signals
(used in Figure 6-3 for illustrative purposes), the delay in the 𝑍𝑍1 driven gear periods of the HTP
(applies for a true hunting tooth design, otherwise the HTP is < 𝑍𝑍1 periods) can be averaged to
give an approximation of the DC component of transmission error (which must be significantly
smaller than a tooth spacing). Using only 1xRev tachos could however mean that a phase-time
map with poor resolution is used to order track the signals, which could lead to less accurate
results. If encoder signals are available, the signals from the worn gearset can be rephased by the
same shift identified above, and the phase-time maps generated, now with meaningful starting
phase. The TE for the reference and worn gears can then be calculated with the starting phase
retained, and the difference between these two TE signals produces the ‘absolute transmission
error’, which describes the changes in the gear profile due to wear, with respect to the reference
gears. Note that this absolute transmission error includes variations around the mean as well as
the DC shift due to wear, the latter being directly proportional to the mean wear depth of the gears.
Note also that the rephasing procedure can be applied directly to the phase angle signals, rather
than to the tacho or encoder signals.

89
6.2.2 Summary of processing method for absolute transmission error
The previous section provided an intuitive explanation of the concept behind the
calculation of absolute TE. This section instead summarises the complete processing procedure
for its efficient and accurate calculation, as shown in Figure 6-4. Starting from raw encoder signals,
the first step is always to obtain the phase-time maps from both the input and output shafts of both
healthy and worn gear signals. The signals are then order-tracked with respect to the input shaft.
The order tracking must be carried out immediately after the phase-time maps are extracted, in
contrast to conventional TE in which order tracking can be performed after the TE signal is
obtained. This is because order tracking removes any speed fluctuation effects in the phase signal,
leaving a phase signal that represents only profile characteristics. This is crucial for the rephasing
procedure where a very small offset (𝑑𝑑+𝑘𝑘 ) is to be detected, in turn for the identification of the
optimal shift required to rephase the signals. After order tracking, the output shaft phase signal of
the worn gears is rephased with respect to that of the healthy gears directly in terms of radians by
using the following equations to calculate 𝑘𝑘𝑠𝑠ℎ𝑖𝑖𝑖𝑖𝑖𝑖 :

𝑑𝑑+0 = wrap(< [𝜃𝜃2 (𝜃𝜃1 )]worn − [𝜃𝜃2 (𝜃𝜃1 )]ref >)

(6-1)
𝑑𝑑+𝑘𝑘 = wrap(𝑑𝑑+0 − 𝑘𝑘 ∙ 𝐺𝐺𝐺𝐺 ∙ 2𝜋𝜋) with 𝑘𝑘 = 1, 2, … , 𝑍𝑍2 − 1

𝑘𝑘𝑠𝑠ℎ𝑖𝑖𝑖𝑖𝑖𝑖 = argmin��𝑑𝑑+𝑘𝑘 �� with 𝑘𝑘 = 0, 1, … , 𝑍𝑍2 − 1 (6-2)

𝑘𝑘

where < > represents the mean over 𝜃𝜃1 , wrap( ) brings an angle into the range of [-𝜋𝜋, 𝜋𝜋] and 𝐺𝐺𝐺𝐺 =
𝑍𝑍1
𝑍𝑍2
is the gear ratio.

For illustration purposes, the gap 𝑑𝑑𝑚𝑚𝑚𝑚𝑚𝑚 in Figure 6-3 has been exaggerated. Note that the
real wear depth must be a fraction of a gear tooth spacing, which is hardly visible in raw signals.
The maximum practical limit for the gap (the tooth thickness itself) also ensures that the choice
of the minimum 𝑑𝑑+𝑘𝑘 as the actual gap is the only one physically possible, given that all others will
be greater than a tooth spacing. The equations provided in this section account for both mean
phase leads and lags, although their use in determining phase leads might not be of interest since
the phase change due to wear (material loss) would always be a lag. These equations implement
efficiently the conceptual method illustrated in Figure 6-3 (b).

90
 Figure 6-4. Processing method to calculate absolute TE.

The TE signals scaled along the line of action for both healthy and worn gears can then
be calculated, with consistent phasing, using the equation 𝑇𝑇𝑇𝑇 = 𝑅𝑅2 𝜃𝜃2 − 𝑅𝑅1 𝜃𝜃1 , where 𝑅𝑅1 and 𝑅𝑅2
are the base radii of the driving and driven gears respectively. The difference of these two TE
signals produces the absolute TE, a signal that describes the changes of gear profile due to wear,
with a healthy gearset as the reference. This absolute TE, with the length of a hunting tooth period,
is a signal that fully describes the meshing condition of every tooth pair of a gearset.

6.3 Application to gear wear assessment

This section demonstrates the application of the proposed methodology of section 6.2
with data from the gear wear experiment of section 3.5.3. A healthy gear was chosen as the
reference signal and wear depths were calculated using the proposed method.

6.3.1 Quantification of absolute transmission error

Since a differential comparison between TE signals at different wear stages is required,
it was favourable to first resample the signals so that they possessed a fixed (and identical) number
of samples per revolution of both shafts and per gearmesh. This could be easily done by defining

91
a number of samples per gearmesh cycle, and multiplying this value by the number of teeth of the
gear mounted on the relevant shaft to obtain the samples per revolution of each shaft.

To extract the phase signals from the encoders, both time-based and phase demodulation-
based (PDOT) methods [110] were applied, giving similar results. The output shaft phase signals
were order tracked [61] with respect to the corresponding input shaft phase signals to remove
speed fluctuations and to serve as a precursor for the rephasing procedure. The rephasing
algorithm was demonstrated in section 6.2 with once-per-rev tacho signals as an example for
simple graphical illustration. To improve accuracy, this can be performed with high precision
encoder signals which in this case have 1000 pulses-per-rev. The tacho signals were still essential
to serve as phase reference markers on both shafts.

The equations provided in section 6.2 can then be utilised to determine 𝑘𝑘𝑠𝑠ℎ𝑖𝑖𝑖𝑖𝑖𝑖 and thus to
locate the HTP reference in the worn cases with respect to the reference healthy case. To rephase
the signals as shown in Figure 6-3 (c), note that the signals must correspond to exactly one HTP
because the process requires circular shifting (i.e., as if the two ends of the signals were joined).
This means that any phase angle signals that are to be shifted must be wrapped.

Following this process, the TEs at different wear stages throughout the test were obtained
and scaled to represent TEs along the line of action. To observe the evolution of wear, a reference
TE representing the healthy condition of the gears was chosen, and it was subtracted from the rest
of the TE curves to produce what might be called absolute TEs (i.e., differential with respect to
an unworn reference), as shown in Figure 6-5. Note that the absolute TEs in Figure 6-5 are scaled
along the line of action (to be comparable to wear depths), and are synchronous averages of the
output shaft. It can be observed that there were significant DC shifts as wear progressed (Test 0
to Test 14) while there were minimal changes in the gearmesh components, especially after the
first few measurements. This highlights two strong advantages of this approach against the zero-
mean conventional TE. First, even if a potential correlation between wear depth and gearmesh
harmonic could exist for mild wear, it is definitely not the case for medium and severe wear,
where gearmesh components seem almost unaltered. Second, even for mild wear, the use of this
correlation for severity assessment would require a tedious calibration, whereas absolute TE
provides a direct quantification of wear depth.

92
 Figure 6-5. Absolute TE (low speed, low load) at different wear stages with respect to healthy
gear condition.
Note that the sign convention used in Figure 6-5 is intentionally set to the opposite of the
one originally defined in [84], so that here the lag due to material loss produces a positive
transmission error, to show the increasing trend of the average wear depth.

It was discovered in preliminary experiments that there had been slippage of the encoder
stators between successive measurements. If this occurs in practice, it is usually observable as a
sudden, unusual increase in the mean TE level. In such cases, despite losing a precise historical
reference, the relative wear depth could still be measured by establishing a new (and ‘known’)
reference, i.e., the change in wear depth could still be monitored even if the first healthy reference
was lost. This is one advantage of this technique as it monitors the changes (phase shifts) due to
wear, with a healthy reference being ideal but not a necessity.

The reference GTE used here was actually not the first GTE data point collected during
the experiment; in fact, it was a measurement that was taken after 20 minutes of running. At the
very beginning of the wear test, the gear was experiencing unstable run-in wear which included
the removal of the black oxide on the gear tooth surface and any non-uniform characteristics
across the teeth. Moreover, it was determined from a previous experiment that an initial running-
in was required to take up any backlash in the system components. Ensuring that all components
were tightened was a crucial step because these backlashes would lead to “slippage” in the
encoders, thus losing the original phase references of the shafts and compromising the wear depth

93
measurements. This was aggravated in this laboratory setting by the unorthodox mounting of the
encoders, clamped onto retort stands. In industrial applications, where encoder stators are rigidly
and securely mounted to the gearbox casing, this problem will not be encountered.

Figure 6-6. Final worn tooth images showing wear depths measured along line of action of (a)
driving gear with average wear depth of 544 𝜇𝜇𝜇𝜇; (b) driven gear with average wear depth of
218 𝜇𝜇𝜇𝜇.

Figure 6-6 shows the images of two worn teeth from the end of the experiment – one from
each of the driving and driven gears – to provide an indication of the profile, with estimated
measurements of wear depth at different positions along the arc of contact. These wear depths
were measured along the line of action to allow comparison with the absolute TE. It is difficult to
identify exactly the pair of positions where two teeth come into contact. However, it seems that
the position where the wear depth of the driving gear tooth is maximum (660 𝜇𝜇𝜇𝜇) has a trough,
whereas the position with the minimum wear depth (110 𝜇𝜇𝜇𝜇) in the driven gear tooth appears to
have a crest. It is likely that these two points come into contact. Taking the sum of these values
gives a combined wear depth of 770 𝜇𝜇𝜇𝜇, which is within the range of the measured absolute TE
at the final wear stage (Figure 6-5). Moreover, the sum of the average wear depth (calculated
using the measurements in Figure 6-6) of both teeth gives a value of 762 𝜇𝜇𝜇𝜇, which is also very
close to the final combined average wear depth estimated from absolute TE. This is the first step
in the validation of the proposed method, and the next section further validates the method with
the mass measurements of the wear particles.

It should be noted that comparable measurements for the identification of wear depth
must be performed under the same load because load variation would cause an additional DC shift
due to shaft, bearing and tooth deflections. In the case illustrated in Figure 6-6, excessive wear

94
would give a change in tooth stiffness, but this would normally not be permitted in industrial
settings. The change in meshing stiffness is disguised in the experiments because of the high
flexibility of the shafts relative to the teeth, and the limitation of torque compared to the rated
load of the gears. The absolute TE can be calculated in the same way even if there is lubricant
since oil film thicknesses (ranging from 0.1 to 10 𝜇𝜇𝜇𝜇 depending on a variety of factors [112]) are
typically much smaller than wear depths.

6.3.2 GTE, STE, DTE and comparison with mass measurements

Figure 6-5 showed only the absolute TEs calculated using the GTE measurements. This
section analyses the wear depths calculated using the mean of the absolute TE measurements
under different operating speeds and loads. These wear depth measurements are also compared
to the equivalent values calculated from the mass measurements of the collected wear particles,
where 1 𝜇𝜇𝜇𝜇 of wear depth along the line of action corresponds to approximately 15 𝑚𝑚𝑚𝑚 of
particles. It is important to realise that this wear depth is the combined average wear depth of a
tooth pair, but not of a single tooth. Therefore, in obtaining these wear depth values from mass
measurements, several assumptions have been made. The first assumption is how the wear is
distributed between the gears. Since the two gears used in the experiment here were made of the
same material, scaling the distribution of wear according to the gear ratio seemed a reasonable
assumption. This assumption might be invalid if gears of different materials or hardnesses are run
together. Other than that, wear is assumed to be uniform across the face width of a tooth, and
uniform for all teeth of a gear.

95
Figure 6-7. Comparison of combined average wear depth calculated using mean of absolute TE
from GTE, STE and DTE signals, against equivalent wear depth calculated from particle mass
measurements (top); Error of wear depth based on GTE, STE and DTE vs wear depth based on
particle mass (bottom). Note that a zero-error value at the beginning is only representing the fact
that Test Number 0 is used as reference.

Figure 6-7 shows the wear depths calculated using GTE, STE and DTE measurements
and their comparison to the equivalent wear depths calculated from the particle mass
measurements. Each test number represents 5 minutes of running (see section 3.5.3). As
mentioned earlier, the particle mass measurements were taken every 10 minutes, hence occurring
only at even numbers in Figure 6-7. Since the first GTE measurement was 20 minutes after
wearing the gears, a correction was made for the mass of the particles collected before this point
to allow for valid comparison. Also, since there was a 6% error between the actual changes in
masses of the gears and particle masses, the “mass-based” wear depth was corrected by scaling
up by 6% the particle mass measurements so as to represent the actual evolution of the masses of
the gears, assuming that the error was uniform across all points. From Figure 6-7, it can be
observed that the trends of the wear depth measurements from GTE, STE and DTE are very
similar and monotonic, indicating that the increasing DC shift caused by wear is also almost
independent of operating conditions. The errors of the final GTE, STE and DTE measurements
compared to mass-based wear depth calculations are 13.2%, 11.8% and 14.5%, respectively,
showing good correspondence and validation of the methodology. The errors at earlier wear
stages such as Test Number 2 are higher, but that is for very small wear levels (approximately 50

96
𝜇𝜇𝜇𝜇). A mean TE lower than the corresponding measured mass loss is possibly explained by the
fact that the worn surfaces are not perfectly flat across the tooth face, an observation which is
qualitatively confirmed by visual inspection. It is interesting that even the DTE measurement
gives a very good indication of wear levels, despite being affected by system resonances; the
slightly increased discrepancy is because the phase references are affected at higher speed due to
a shaft harmonic being close to a torsional resonance.

6.3.3 Sensitivity to angular resolution: encoders vs tachos

In cases where encoders are not available, the proposed technique could still be applied
with once-per-rev tachos to obtain average wear depth, but at the cost of accuracy (though small).
This section demonstrates the quantification of wear depth using only tacho signals 8 from input
and output shafts. Figure 6-8 shows the wear depths calculated at different operating conditions
using only tacho signals, and their comparison to the ones from the particle mass and the encoder
signals at low speed and load (i.e., GTE measurement).

Perhaps surprisingly, observing the trends in Figure 6-8, even measurements based on
tacho signals alone give good indications of the wear depth, which further shows the robustness
and effectiveness of the technique. It was expected that there would be more fluctuations in the
results, especially for the low-speed (2 Hz) and high-load (5 Nm) case. The test rig was unable to
maintain a constant speed when applying load under low speed, thus causing more uncertainties
in the phase because of the poor resolution of the phase-time maps. This could be further
explained by looking at the tacho measurements at higher speed (10 Hz) and higher load (5 Nm),
in which the trend is smoother. However, despite having the smoothest trend, the measurement at
10 Hz and 5 Nm has the largest error, whereas the final measurement point at 2 Hz and 5 Nm
almost completely intersects the mass-based measurement, even though it has greater fluctuations
throughout the duration of the test. Nevertheless, all measurements show good correspondence,
with acceptable error margins.

8
For the simplicity of processing, the wear depth values from the tacho signals were calculated using
threshold-crossing of a trigger level, to avoid the necessity to remove end effects and ripples due to the
phase demodulation of once-per-rev tacho signals.

97
 Figure 6-8. Comparison of wear depth calculated from tacho signals at different operating
conditions.

6.3.4 Analysis of the evolution of TE components throughout the wear process

This section analyses the different TE components throughout the wear process. All the
results in Figure 6-9 were extracted from the GTE measurements (2 Hz, 0 Nm, before stops) using
the encoders. Figure 6-9 (a) shows again the DC value of each GTE measurement corresponding
to a different stage of wear, in comparison with mass-based estimates of wear depth. Figure 6-9
(b-d) are ‘conventional’ (no reference, thus without DC level) GTE signals after synchronously
averaging with respect to gearmesh (b), input shaft (c) and output shaft (d). In order to study the
changes of the input and output shaft components of GTE, the gearmesh harmonics have been
removed from the signals shown in Figure 6-9 (c, d), so that the shaft-specific components can be
revealed and observed clearly without being masked by the gearmesh components. Figure 6-9 (e-
g) are the same synchronous averages as Figure 6-9 (b-d) respectively but with the healthy
reference subtracted, thus showing only the variation of TE due to wear.

Observing Figure 6-9 (b) and (e), it can be seen that, from healthy to mild wear, the
gearmesh harmonic content of TE generally increases, as expected from previous findings
discussed in the literature. However, the same magnitude seems to stabilise soon after the early
wear stage, indicating that wear rates along the tooth flank become more and more even as wear
progresses. This makes gearmesh harmonics insensitive to wear in medium to severe cases.

From Figure 6-9 (c, d), it is observed that there is an eccentricity in both gears, causing a
1xRev harmonic in the raw GTE signals. This seems to cause a non-uniform development of wear

98
around the pinion, as shown by the evolution of the first pinion-synchronous harmonic in Figure
6-9 (f), which highlights changes with respect to the healthy case. The fact that this phenomenon
is not observed in Figure 6-9 (g) for the output shaft could be explained by the fact that most wear
happens on the pinion teeth due to their higher number of wear cycles. This phenomenon however
requires further investigation in future work. It is worth noting that while these TE waveform
components exhibit wear-related developments, none of them give anywhere near as good an
indication of wear severity as the figures extracted from the mean of the absolute TE, shown in
Figure 6-9 (a). Moreover, their physical relationship with wear is likely dependent on other factors
such as eccentricity, which makes it difficult to generalise their use.

99
 Figure 6-9. Changes in TE (GTE) components throughout wear evolution. An angle of 360°
represents a full gearmesh period in (b) and (e), a full pinion shaft rotation in (c) and (f), and a
full driven gear shaft rotation in (d) and (g).

6.3.5 Discussion: important considerations of the proposed method

There are some requirements and limitations that need to be considered in order for the
proposed method to work properly. A first important constraint is on the length of the acquired
signals, which must be at least equal to a HTP to allow rephasing (≤ 𝑍𝑍2 periods of the input shaft).

Secondly, to obtain a meaningful absolute TE that accurately describes the changes due
to wear, signals from a reference gearset are required. Ideally an unworn gearset is chosen to be
the reference, but the method shown here can still quantify the changes in profile if signals of
unworn gears are not available, which is still useful as long as the state or condition of the
reference gears is known. Measurements should ideally be taken at low load and low speed to
obtain absolute TE from GTE, which describes the TE component due to gear tooth profile only.
It was however shown (section 6.3.2) that even under load (STE) and higher speed (DTE), the

100
estimation of the combined average wear depth was accurate, since constant load and speed result
in little changes in static and dynamic deflections. Therefore, the comparison between TE signals
should be performed under the same speed and load conditions. Future studies will investigate
the effects of speed and load on the quality of the results. The data used in this chapter could not
be recorded at higher speed and load conditions because there was no lubricant; the dry wear
experiment was performed with the purpose of accelerating the wear process to prove the
application of the method. As mentioned before, even in lubricated conditions, the calculation of
the average wear depth will not be affected because oil film thickness is usually not more than a
few microns, negligible compared to meaningful wear-depth estimates.

The mounting of the tachometers or encoders plays a crucial role for the accuracy of the
method as it depends on phase references of the shafts. As observed in preliminary experiments,
slippage can occur between the encoder rotor and the shaft and between the encoder stator and
mounting, causing an error in the absolute TE estimation. Thus, all backlash in the system
(components between the gear and where the tachometers/encoders are mounted), such as
clearances between keyways and couplings, must be taken up right at the beginning, before
reference signals are recorded in the healthy state. From an industrial application point of view,
gears with large numbers of teeth might require encoders with proportionally higher numbers of
pulses per rev to have the same resolution.

6.3.6 Comparison with vibration measurements

It is worth comparing absolute TE with vibration measurements to further demonstrate
its effectiveness as a wear severity indicator. Figure 6-10 shows the velocity spectrum in
logarithmic scale after integrating one of the accelerometer measurements, recorded after
approximately 12000 pinion cycles (Test Number 4), and at 10 Hz and 5 Nm. As expected, the
vibration signals are dominated by the components of the gearmesh frequency and its harmonics.
Any changes, then, in RMS of the vibration signals would only be sensitive to variations of the
profile around its mean involute curve, and they would not react to uniform material removal
(shifted involute).

101
Figure 6-10. Spectrum of vibration measurement at Test Number 4 (after approximately 12000
pinion cycles).

This is confirmed in Figure 6-11, which shows the RMS of the velocity measurements
from the three accelerometers mounted at different locations (and orientations) on the gearbox
casing, over the wear process. From Figure 6-11 (a) and (b), it can be observed that the RMS
increases monotonically from the beginning until Test Number 4, and moderately stablilises after
that. This phenomenon is consistent with the changes of the gearmesh components of TE (see
Figure 6-9 (b) and (e)). Therefore, as for the gearmesh component of TE, they can only provide
an indication of wear severity until a certain mild level of wear, but are completely insensitive to
severe wear. This is because in a worn gear, where the “shifted involute” profile gives only
changes in the DC values of TE, vibration signals would not exhibit any changes. Also note that
vibration is only generated by dynamic tooth deflection, itself requiring either inertial or spring
forces to vary with the relative angular position.

102
 Figure 6-11. RMS of velocity measured using accelerometers at different locations throughout
wear evolution.

Moreover, for the quantification of wear depth, vibration would require the identification
and calibration of a scale parameter. This often relies on complex models to take into account
different operating conditions as well as the implicit relationships between average wear and the
scale of specific wear patterns such as a double-scalloped profile. Further, a vibration signal also
depends on the sensor position and related structural transfer path. This is shown in Figure 6-11
(c), where the RMS did not show clear changes due to wear even at the beginning of the test, as
compared to Figure 6-11 (a) and (b).

TE is less affected by the aforementioned challenges suffered by vibration because its

measurement is directly at the source, highlighting the potential advantages of the application of
TE in gear diagnostics. Further, the proposal of absolute TE in this chapter has allowed the
quantification of wear, which has not been achieved with vibration according to the literature (and
in principle cannot be).

103
Chapter 7 Comparative study on the use of TE in
wear monitoring: dry vs lubricated conditions

This chapter is included in this thesis as a compendium of the previous chapter, and
provides a summary of additional experimental observations on TE signals measured in dry and
lubricated wear tests, respectively dominated by abrasive wear and fatigue pitting. This chapter
does not aim at offering new analysis techniques, but rather new insights on the effect of different
wear mechanisms on the spectrum of TE signals. Despite probably constituting a lesser
contribution with respect to the innovations of the previous chapters, the information provided in
this chapter is deemed useful for future developments of wear-mechanism-specific prognostic
techniques.

7.1 Pre-processing of TE signals for all tests

The data used in this chapter was measured at low speed (2 Hz), and therefore does not
require the preprocessing steps discussed in Chapter 4 for the removal of the transfer function.
Moreover, the availability of low-load tests and the dominance of wear over tooth-bending
deflections, even at high load, also allow skipping the separation of profile and flexibility effects.
The limited effects of compliance in these two campaigns result from the larger facewidth of the
gears (20 mm), compared to the thin teeth of the cracked case (5 mm). Finally, the application of
the Chapter 4 approach to this data is not only unnecessary, but also hindered by the very limited

104
variety of recorded operating conditions (especially speeds), which is due to the fact that these
campaigns were run before the full development of Chapter 4. The findings of this chapter are
however general, in the sense that, in high speed cases, the same analysis would apply to
equivalent STE and GTE reconstructed after removing the transfer function and separating profile
and compliance effects.

The TE signals are first obtained from the encoder signals using a procedure similar to
that outlined in Chapter 6. To extract information related to uniform wear (across all teeth) only,
the synchronous averages of a gearmesh (GM) period at different wear stages were then calculated,
i.e., only GM harmonics are retained in the order spectrum. To allow a reasonable and valid
comparison between the two wear tests, the pre-processing procedure of the signals is performed
in the same way for all data in this chapter to extract the GM signals of TE.

7.2 Dry test

The dry test results presented in this section include four measurements for each wear
condition. These arise from the combination of two loading conditions (nominally 0 and 5 Nm,
leading to GTE and STE) and two successive measurement records obtained before and after each
stop (see section 3.5.3 for details). It should be noted that all results for dry test presented in this
section are absolute TE signals, with respect to the reference at time “01h06m”. While the main
aim of this section remains the analysis of general trends of profile-induced TE in the case of dry
wear, allowing a later comparison with lubricated tests, the availability of these four measurement
conditions allows further insights on the capability of TE to measure profile. In particular, in the
comparison of GTE vs STE and before vs after stops, the aim is to highlight the effect of load and
other parameters on TE-measurements. While the trend of the average wear depth (DC values)
has already been thoroughly discussed in section 6.3, this section will focus on the analysis of the
waveform representing profile-shape changes.

Unfortunately, the comparison of STE and GTE, even at the earlier stages of wear, cannot
in this case provide a reference angle to locate the pitchline, because the two-to-one tooth-pair-
contact transition for 20 mm facewidth gears results in less than 1 𝜇𝜇m difference in deflection,
which is below other differences ascribed to Hertzian contact deformation (discussed later). This
prevents the drawing of fully proven conclusions on the nature of some observed TE patterns, and
considerations will have to rely on assumptions on the location of the pitchline within a TE
gearmesh period.

Figure 7-1 shows the GTE signals obtained before and after stops for surface inspections.
Except for a small brushing of the surfaces to remove some small deposits, there is no expected

105
difference between the results before and after stops. The only physical explanation which sounds
plausible with no further information, is a localised temperature effect, with the gears cooling
down during the stops. The sharper profile features, observed on severely worn gears in the centre
of the x-axis would be “softer” at higher temperature. The evolution of GTE after stops (likely a
more reliable measure of “true” profile) shows the evolution of a double-scalloped profile, with
a resulting minimum-wear valley located in the centre, clearly visible especially after about 3
hours of testing. Assuming that the pitchline is indeed located in the middle of the x-axis, this
would confirm the surface observations of a sharp ridge located in the single-pair contact area.

Figure 7-1. GM signals of GTE obtained before and after stops for inspections and/or surface
measurements.

106
 From an analysis of STE (Figure 7-2) it appears that the differences before and after
stopping are reduced, probably due to the fact that local surface deformations are less temperature
dependent under higher loads. This is reasonable if it is considered that sharper ridges would
completely deflect under sufficient load, independently of the temperature.

Figure 7-2. GM signals of STE, obtained before and after stops for inspections and/or surface
measurements.

A more reliable analysis of the trends of those profiles can be performed by analysing the
evolution of the harmonics of TE, as shown in Figure 7-3. The figure presents the GTE results in
the left column, with the STE results on the right. Within each plot, the results obtained from data

107
before and after stops are shown in the left and right columns, respectively. Each row of the figure
shows the changes of the corresponding GM harmonics, except for the last, which shows the
overall RMS of the signal. Test #1 to #14 corresponds to the 14 data points in Figure 7-1 and
Figure 7-2 in the same order, e.g., for GTE 01h16m to 03h46m (before stops) or 01h17m to
03h47m (after stops).

Figure 7-3. Evolution of GTE (left column) and STE (right column) gearmesh harmonics as
wear progresses.

By observing the GTE and STE harmonics in Figure 7-3, most of the GM harmonics do
not seem to be a good trending parameter for wear in dry conditions, except for the 2nd, 3rd and 4th
GM harmonics of GTE obtained after stops, likely corresponding to the sharp event due to the
pitchline ridge. These three harmonics do not however offer robust diagnostics, as shown by the

108
fact that they are only visible at low load and after stops. Some GM harmonics did increase at the
beginning of the wear test (e.g., Test #1 to Test #6 of 2nd GM harmonic of GTE before stops), but
did not consistently show a monotonic increase as wear progresses to more severe levels. These
observations can be explained by the fact the type of wear in dry condition is dominantly abrasive
wear, quickly leading to significant loss of materials from the tooth surfaces. In medium and
severe wear, the changes in tooth profile are no longer dominant symptoms of wear, and the
overall material loss continues as a “shifted involute” of the tooth profile, making average wear
depth the most sensitive measure. In such cases, it is more suitable to monitor the DC value of
absolute TE as a trending parameter, as detailed in Chapter 6.

7.3 Lubricated test

7.3.1 Sensitivity of GTE and STE to gear wear

This section presents the results obtained from the lubricated test. The TE signals were
processed in the same way with the procedure described in section 7.1.

There are more data points collected in the lubricated test than in the dry test, since wear
had been significantly slowed down by the presence of lubricant. Figure 7-4 and Figure 7-5 show
the first half and last half of the data points of the experiment, respectively. The application of
absoute TE for this campaign was not possible due to a slippage observed in the encoder mounting.
These were less secure than those used in the crack tests and resulted in random slip of the stator
of the encoder for long campaign such as this one.

Observing the results in Figure 7-4 and Figure 7-5, it can be seen that the peak-to-peak
value of both GTE and STE signals consistently increases as wear progresses. Unlike in the dry
test, wear appears to be mostly concentrated in a specific area of the tooth, and does not present
a clear double-scalloped profile. Moreover, at any wear stage the GTE and STE signals seem to
be very similar, and their differences tend to decrease, rather than grow, with wear severity.

Figure 7-6 further shows the evolution of the harmonics of TE due to wear in the
lubricated condition. Similarly to the dry test results, the first four rows show the changes of the
first four GM harmonics, while the fifth row shows the overall RMS. Test #1 to #40 corresponds
to the 40 data points in Figure 7-4 and Figure 7-5 in the same order, i.e., 00h00m to 67h06m.

The fact that wear is distributed over a single area is confirmed by the fact that once wear
commences the first GM harmonic dominates the signal, and contributes to most of the RMS. All
other harmonics show non-monotonic patterns and a smaller sensitivity to wear severity.

109
Figure 7-4. GM signals of GTE and STE from lubricated test (first half).

110
Figure 7-5. GM signals of GTE and STE from lubricated test (last half).

111
 Figure 7-6. Evolution of GTE and STE GM harmonics as wear progresses for lubricated test.

The overall RMS for the lubricated test monotonically increases, differently to the
observation in dry test, dominated by the first gearmesh harmonic. This was further compared
with the pitted area on the pinion (the most affected by this type of wear), to check if the growth
of RMS could be used as a trending indicator

112
 Figure 7-7 shows a comparison of the RMS values of GTE and STE throughout the
experiment with the pitted area. By plotting the RMS values and pitted area with two different
scales, it can be seen that there is a reasonably good correspondence in their trends. For a more
accurate interpretation of wear rate when comparing with the pitted area, the horizontal axis is
plotted in units of million (mil.) cycles, where Test #1 and Test #40 correspond to the 0th and
3.25th mil. cycles respectively. For all the trends, the highest rate of wear seems to be from 0.5
mil. cycles to 1.1 mil cycles. Both GTE and STE closely follow as the trend of pitted area
increases, at least until around 2.4 million cycles. After 2.4 million cycles, the pitted area starts
to stabilise with slight variation while both GTE and STE exhibit an increasing trend from the
start to the end. The phenomenon after 2.4 million cycles is possibly because there are still
changes of the tooth profile occurring that affect GTE as wear progressed in depth rather than
across the surface. It is understood that GTE is mainly caused by the change in tooth profile due
to wear, instead of the pitting directly. This is however still one step forward in establishing TE
as a wear indicator. Note that the results for GTE and STE presented here differ slightly from
those presented in [98] and [99], partly because of an improvement in consistency in the technique
used to extract the TE, but mainly because the particular result for the GTE near the end of the
test [98] was artificially low, because of error.

Figure 7-7. Comparison of RMS values of GTE and STE with pitted area.

113
7.3.2 Comparison of TE with vibration
To compare the ability of TE and vibration to monitor gear wear, vibration signals
measured at different speeds (2 Hz, 6 Hz, 10 Hz, 16 Hz and 20 Hz) and loads (0 Nm and 20 Nm)
were analysed. The vibration signals at different wear stages were first order-tracked with respect
to the input shaft. Synchronous averages of a gearmesh period were then calculated for all wear
stages. Figure 7-8 shows the averaged vibration signals of only two wear stages, i.e., the initial
healthy stage and final wear stage, at different speed and load conditions. It can be observed that
at a speed of 6 Hz or higher, vibration shows some change in its amplitude of the gearmesh content
due to wear. The change is however not as obvious as TE, where the amplitude of the final stage
is close to four times that of the initial stage. Moreover, vibration shows virtually no changes due
to wear at all at a low speed of 2 Hz because it depends on the deflection of teeth as well as
dynamic forces to exhibit the wear-induced changes. On the other hand, TE should work equally
well for both low speed and high speed, as well as different load conditions, since even a DTE
signal should reveal changes in GTE due to wear.

114
 Figure 7-8. Vibration signals of initial healthy and final wear stages at different speeds and
loads.

Figure 7-9 and Figure 7-10 both show a comparison of the RMS values of the amplitudes
of the gearmesh content of vibration with pitted area, at speeds of 10 Hz and 20 Hz respectively.
From Figure 7-9, the RMS values of vibration at 10 Hz seem to follow the trend of pitted area,
but not as closely as TE. However, at 20 Hz, the RMS values have quite different trends and do
not show good correspondence with pitted area. At other speeds, there were no clear trends
observed. This further emphasises the influence of speed on vibration in gear condition
monitoring.

115
 It is understood that with further processing of vibration signals, it is possible to extract
more signal features that are related to pitting. For example, different frequency ranges of
vibration would show very different features. The intention here is however to demonstrate the
close relationship between TE with changes on the gear tooth surface, highlighting the “closeness”
of TE to the fault and the very limited requirements in terms of signal processing.

Figure 7-9. Comparison of RMS values of vibration at 10 Hz and 20 Nm with pitted area.

116
Figure 7-10. Comparison of RMS values of vibration at 20 Hz and 20 Nm with pitted area.

117
Chapter 8 Discussion

This thesis has developed several TE-based techniques which aim to demonstrate and
establish TE as a powerful tool for gear condition monitoring. Considering that gear cracks and
wear are the most common gear faults, the tools developed in this thesis to accurately estimate
crack depth in millimeters and average wear depth in microns render TE a valuable alternative to
vibration. The limited amount of pre-processing required in the proposed crack and wear severity
assessment techniques makes this new gear condition monitoring strategy appealing for industry
applications, especially when compared to many sophisticated and multi-parameter methods
proposed for vibration measurements. Vibration has instead an excellent early fault detection
performance, while TE has shown outstanding severity assessment capabilities, especially for
medium and severe wear, thus suggesting strong opportunities to combine the two in a full-life
cycle condition monitoring strategy.

In this chapter, the performance of TE is discussed in detail for crack and wear severity
assessment (sections 8.1 and 8.2 respectively), summarising and analysing the findings of the
previous chapters.

8.1 Discussion on TE-based crack severity assessment

In chapter 4, new TE-based methods have been proposed to remove transfer-path effects
and separate components of STE, thus allowing for the estimation of profile errors and meshing
compliance. The latter is then applied in chapter 5 to estimate gear root crack severity. The

118
methodology is able to accurately estimate crack depth as a percentage of tooth base thickness,
with a maximum error of about 5% of the root thickness. This is achieved by combining dynamic
transmission error measurements acquired at different speeds and loads to improve robustness,
applying a simple OMA step to remove the transfer function, and comparing the estimated
compliance with theoretical values obtained using the potential energy method. Compared to
similar attempts with vibration signals (e.g., [73]), the transfer functions of TE are much simpler
and therefore require a significantly lower number of parameters to be fitted, increasing the
robustness and simplicity of the algorithms employed.

The crack severity estimation is then performed based on a comparison with a relatively
simple quasistatic gearmesh stiffness model, and in particular searching for the simulated crack
giving the minimum RMS error against the estimated experimental compliance. It was found that
the curvature of this RMS-error (see Figure 5-8, Figure 5-10 and Figure 5-12) around the
minimum point (the estimated crack depth) grows with fault severity. This means that the
confidence on the severity assessment grows with the severity of the fault. This is mostly due to
the fact that the compliance loss resulting from the same increase in crack severity (e.g., +1%) is
more pronounced in the case of a severe crack (e.g., from 50% to 51%) than a small crack (e.g.,
from 20% to 21%), as obvious even from the stiffness model results (e.g., Figure 5-3). This
behaviour is advantageous (and commonly sought after) in practical condition monitoring terms,
because the diagnostic procedure becomes more accurate as the system approaches failure. A
lower confidence in the estimation of small cracks is not considered a serious problem, since at
such an early stage fault-severity estimation is less important than detection, which could be
achieved by a number of detection-focused techniques, possibly integrating TE with vibration
measurements.

The proposed procedure relies on a relatively simple quasistatic tooth stiffness model,
whose accuracy is expected to influence the quality of the results. However, even if slight
differences in the shape of theoretical and estimated compliance were observed, they did not
compromise the overall quality and repeatability of the results. Different phenomena could be the
cause of this mismatch, e.g., unmodelled contact deformations and/or “low-pass” filtering effects
of the estimation procedure resulting in a “smeared” compliance curve. The adoption of more
sophisticated tooth-compliance models is likely to provide some answers and solutions to this
issue, and in general improve the effectiveness of model-based crack severity estimation methods.
Examples of additional phenomena that have not been considered in the model used in this thesis
are: the coupling effect of the fillet foundation [113], the impact of cracks on fillet foundation
stiffness [114], and nonlinear Hertzian stiffness [115]. That said, the fact that successful fault

119
severity estimation was possible with a relatively simple model shows that this procedure is
relatively robust to modelling inaccuracies.

Before attempting to include these additional effects into the model, their relative
importance should be taken into consideration and weighed against the increased complexity.
This is particularly relevant if attempting to include a nonlinear Hertzian stiffness model, which
would prevent the use of compliance as a metric for fault severity assessment. In that case, the
methodology described in chapter 5 would require an adaptation to the nonlinear case. Simple
considerations of the typical contribution of Hertzian stiffness to the overall tooth deflections,
combined with the mild nonlinearity of this component, show that in most cases the benefits of a
linear model largely outweigh the small errors resulting from such an approximation.

8.2 Discussion on TE-based wear severity assessment

In chapter 6, a new perspective on gear transmission error has been introduced. The DC
component in TE signals, which has hitherto received very little attention and only in limited
design-based studies, is discussed there in detail and applied for the first time to gear wear
assessment. Because of the arbitrary initial phase of recorded encoder signals, this DC component
has not been calculated before, except for the 1:1 gear ratio case. A new rephasing method has
thus been proposed to restore the hunting tooth period phase reference, so that two TE signals
with common phase reference can be compared. The subtraction of two TE signals, one being a
reference healthy case and the other from a worn gearset, will produce a complete representation
of gear profile changes, termed the ‘absolute transmission error’. For the first time, this has
allowed including the DC component of TE in the quantification of the combined wear depth.
This is a significant advance compared to previous studies, where TE-based gear diagnostics has
focused on changes in amplitude and phase of the gearmesh components, themselves unable to
provide a reliable and quantitative assessment of total wear depth (only variations around the
mean).

The capability of TE to measure this all-important average wear depth directly in microns
greatly advances the prospects of TE with respect to acceleration measurements. The latter have
no chance of capturing this information directly, and therefore require significant processing to
extract this information indirectly from the time-varying components of the signal. Even then, the
wear tests of Chapter 6 showed that the correlation between average wear depth and time-varying
components is only present for very mild wear, when the profile changes shape rather than simply
shifting towards the tooth centreline. As with gear cracks, rather than supplanting vibration-based

120
approaches, TE shows a strong complementarity to vibration. The latter is very effective for early
detection and trending of mild wear, thanks to its high-frequency content and responsiveness to
localised profile changes and micro-scale surface alterations [116], while TE is particularly
valuable in quantifying and tracking medium to severe wear progression. This complementarity
offers significant potential for multi-sensor integrated gear-wear monitoring.

The innovation demonstrates the robustness of the proposed methodology in obtaining

the absolute TE measurements and the effectiveness of the DC component of TE as a wear
indicator, with validation from encoder measurements taken from a dry gear wear test and under
different operating conditions. The estimates of average wear depth were almost unaffected by
speed and load and, somehow interestingly, it was found that even 1×Rev tachometer
measurements can give a reasonable indication of the wear depth, after rephasing using the
proposed method. This means that the proposed technique is effective in wear monitoring even
with very simple sensors, thus rendering TE-based wear monitoring accessible to a wide range of
applications.

Despite the most relevant innovation and significant contribution being the DC
components, the study of gearmesh harmonics of TE for dry and lubricated wear tests (Chapter 7)
has demonstrated interesting trends, especially in the mild lubricated wear case. Abrasive wear in
dry tests seem to trend well only with the DC value of TE (average wear depth), and only for the
very mild stages there is a growing first gearmesh harmonic. Higher orders (2nd, 3rd and 4th) are
growing in severe dry wear, but easily masked by load or temperature (the latter assumed to be
the cause of pre- and post-stop differences). Lubricated tests, on the other hand, showed a strong
monotonic 1x and 2x gearmesh increase throughout most of the campaign. These trends seem to
be consistent with the shape of the resulting profiles, which shows a steep ridge at the pitchline
in dry wear, which could cause an abrupt (higher-frequency) event in the signal, and a more
smoothly distributed profile change for pitting. Despite these observations being preliminary and
based only on two test campaigns, they seem to confirm that the DC value of TE is the most
important quantity for moderate and severe wear, and that time-varying components of TE offer
valuable diagnostic information in the earlier stages. The detailed analysis of profile changes in
the dry wear test campaign also highlighted the seemingly relevant effect of local surface
deformation, which can affect the estimation of profile. In particular, STE and GTE measurements
(i.e., with and without significant load) resulted in different profile estimations, with STE showing
less pronounced ridges, possibly due to Hertzian contact effects.

121
Chapter 9 Conclusions and future work

This chapter highlights the conclusions and recommendations for future work in TE-
based gear condition monitoring. Section 9.1 summarises the key outcomes of the thesis, by
providing an account for each research objective. Section 9.2 outlines the limitations of the
innovations developed in this thesis, and then suggests potential future research directions to
further improve the current work.

9.1 Summary of outcomes

The thesis aimed at developing innovative solutions to enable TE-based gear condition
monitoring. The developments focus on two main types of gear faults, i.e., root crack and gear
wear, and aimed at obtaining their accurate and reliable severity assessment. In order to achieve
this, the relationships between profile errors, tooth flexibility, and transfer path in TE signals have
been investigated thoroughly. This study has led to the ability to separate these effects from each
other, as demonstrated in Chapter 4. This then allowed the developments of Chapter 5, which
utilised the estimated tooth flexibility to conduct crack severity assessment. In Chapter 6, another
new innovation is developed to address the lacking DC component in conventional TE signals. A
new rephasing method (absolute TE) was developed and enabled the direct measurement of
average wear depth. Although Chapter 7 has not offered novel algorithms, the comparative and
in-depth analysis of fault symptoms induced by dry and lubricated wear produced new insights
that help in distinguishing between different wear mechanisms (namely abrasive wear and fatigue

122
pitting) by trending the evolution of the TE gearmesh harmonics. With the aforementioned
achievements, the following three objectives have been achieved.

Objective 1: Develop a technique to obtain estimates of profile and tooth deflections based
on TE measurements.

• Utilising TE signals from different operating conditions, a multiple-load and multiple-

speed approach has been developed to separate the effects of profile errors, tooth
flexibility, and transfer path effects.
• The OMA-based technique to remove the transfer path effect can be implemented for any
operating condition, including higher speed cases. The multiple-speed signals enable the
identification and fitting of the resonances, so that the transfer function can be removed
to reconstruct an equivalent STE, a combination of the two inputs of profile error and
tooth deflection.
• A regression analysis then exploits the multiple loads to separate profile errors from tooth
flexibility (i.e., compliance), based on the linear relationship between deflections and
load.

Objective 2: Develop a reliable approach to assess tooth crack severity based on TE.

• To assess crack severity, a technique is proposed to compare the estimated compliance

(see Objective 1) with a theoretical meshing compliance reference 9 to quantify crack
depth.
• It is shown that for medium and large cracks, the technique was able to accurately identify
crack depth, while the accuracy and confidence tend to decrease for small cracks.
• Even down to about 20% of the tooth thickness, the technique is still able to detect the
cracked tooth, despite the overall meshing stiffness changing by only about 6%.
• The increasing accuracy and confidence with more severe faults suit the typical
requirements of a prognostic system, whose importance grows as the machine approaches
failure.

Objective 3: Develop a reliable approach to assess wear depth based on TE.

• A novel innovation has been developed to allow for the inclusion of the DC component
in TE, thus enabling the direct quantification of wear depth using in-situ measurements
for the first time.

9
Although the theoretical meshing compliance is computed based on the established potential-energy
method, any method to compute the TVMS would be suitable for this technique.

123
 • This innovation involves a new rephasing algorithm to obtain the DC component, leading
to “absolute transmission error” when combined with the conventional method to
calculate TE signals.
• The DC component of absolute TE is a direct measurement of average wear depth, which
is proven to be a much more sensitive wear monitoring tool than any harmonics of TE,
especially for mild to severe wear.
• From a practical point of view, the technique requires minimal computation effort and
time, and can be applied even in large industrial gearboxes. It has been shown to work
well both in low and high speed/load, rendering the approach reliable and robust to
operating conditions.
• Though simple, this technique not only provides a new perspective on TE-based
diagnostics, but is also a powerful and practical diagnostic tool to quantify wear severity.

9.2 Future work

This thesis focused on developing signal processing techniques for TE-based gear
condition monitoring, validated using experimental data. It was the aim of this thesis to establish
TE as a robust tool to monitor gear faults, alternative to existing measurements such as vibration,
AE and IAS. Despite this aim being achieved, the thesis also highlighted limitations of TE which
offer interesting opportunities for future work. This section discusses these limitations and
provides recommendations to address these issues for further advances in TE.

1. The main obstacle for TE to be prevalent in condition monitoring is the ability to obtain TE
measurements. The additional installation of rotary encoders on both gear shafts is required.
However, this is becoming less of a problem because encoders (at least one) are already often
installed to use in conjunction with accelerometers to perform order-tracking. Moreover,
aligned with the goals of the Industry 4.0 paradigm in implementing smart systems,
sophisticated sensors such as rotary encoders are expected to be more widespread in future
industries. The estimation of average wear depth (through absolute TE) only requires two
1xRev tacho signals, which can be obtained with very simple and non-intrusive
measurements at exposed ends of the shafts. A detailed analysis of the effect of encoder
properties on the accuracy of the other diagnostic information was beyond the scope of this
thesis and offers a potential avenue for further applied research.

124
2. An OMA-based approach using TE signals of different speeds was proposed to remove
transfer-path effects. The approach requires a manual inspection of the TE frequency
spectrum for a preliminary analysis for the initialisation of the poles and zeros, and this could
be automated by further developments. Another limitation of the proposed method is indeed
the requirement of measurements at multiple conditions of speed and load. However, most
gearboxes in practical applications run (or at least able to run) at different speed and load
conditions. Moreover, it is also possible to remove transfer-path effects without the need for
variable speed conditions, for instance in combination with other OMA-based approaches
(e.g., cepstrum-based OMA [100]) and/or model-based methods, thus improving the OMA
component of the proposed technique. A more sophisticated OMA, or even better, a model-
based approach, would also provide a more reliable indication of the number of poles and
zeros that need to be considered, a choice that can impact the diagnostic results.

3. The crack experiments and their corresponding data were obtained for artificially seeded
slots, which have plastic deformations (closing crack effect), unlike naturally growing cracks.
Tests with a real, naturally developed crack (as in [105]), rather than a slot, would be another
interesting future work, as it would offer new insights on the effectiveness of the technique
and possibly some answers to explain some of the discrepancies observed with respect to
theoretical compliance. These tests could also show how the technique performs when real
cracks develop non-uniformly across the facewidth, if an equivalent stiffness reduction can
be estimated in the form of an “equivalent average crack depth” across the facewidth, and
lead naturally into prognostics.

4. The development of absolute TE was validated with an accelerated-wear experiment under

dry conditions, thus restricting the upper limits of speed and load. Future work should
establish the limitations of the proposed method in terms of the ranges of speed and load
conditions. Moreover, a test should be repeated under lubricated conditions with an encoder
configuration similar to that used in this thesis for dry tests. This will enable the calculation
of absolute TE and its validation under different wear mechanisms such as pitting.

5. The TE measurements used in this thesis are obtained from encoders mounted on the free
ends of the shafts. The free end might not be accessible for some machines. It would be
useful to study the feasibility and limitations of mounting encoders on loaded ends, such as
using through-shaft encoders.

125
6. In this thesis, the innovations are developed and validated based on a single-stage spur
gearbox. There are however various types of complicated gearboxes in the industry. It is
therefore necessary to extend the use of TE to other types of gears or gearboxes, such as
helical gears and multistage gearboxes. The main problem is likely to be the difficulty of
obtaining angular measurements on intermediate shafts especially in compact gearbox
configurations and on moving shafts (planet-gear shafts). However, it is possible that even
established source-separation techniques (e.g., synchronous-averaging) can be used to
separate the contribution of each stage to an overall TE obtained from two encoders at the
two ends of the drivetrain. This type of source separation would unlikely solve entirely the
issue especially for the estimation of average wear depth (DC value common to all stages).
However, even a total combined wear depth of all stages could be used as a fault detection,
if not diagnostic, tool. Section 7 has also shown how at the onset of wear the average wear
depth is well correlated with the amplitude of the first gearmesh harmonics. This correlation
could be exploited to attribute average wear depth to the stage showing anomalous variations
in its gearmesh synchronous components.

126
References
[1] A. K. S. Jardine, D. Lin, and D. Banjevic, “A review on machinery diagnostics and
prognostics implementing condition-based maintenance,” Mechanical Systems and Signal
Processing, vol. 20, no. 7. pp. 1483–1510, 2006, doi: 10.1016/j.ymssp.2005.09.012.

[2] Z. Peng and N. Kessissoglou, “An integrated approach to fault diagnosis of machinery
using wear debris and vibration analysis,” Wear, vol. 255, no. 7, pp. 1221–1232, 2003,
doi: https://doi.org/10.1016/S0043-1648(03)00098-X.

[3] R. B. Randall, Vibration-based condition monitoring: industrial, aerospace and

automotive applications. Chichester, UK: Wiley, 2011.

[4] M. Zhao, X. Jia, J. Lin, Y. Lei, and J. Lee, “Instantaneous speed jitter detection via encoder
signal and its application for the diagnosis of planetary gearbox,” Mech. Syst. Signal
Process., vol. 98, pp. 16–31, 2018, doi: https://doi.org/10.1016/j.ymssp.2017.04.033.

[5] P. Feng, P. Borghesani, H. Chang, W. A. Smith, R. B. Randall, and Z. Peng, “Monitoring

gear surface degradation using cyclostationarity of acoustic emission,” Mech. Syst. Signal
Process., vol. 131, pp. 199–221, 2019, doi: https://doi.org/10.1016/j.ymssp.2019.05.055.

[6] S. Park, S. Kim, and J.-H. Choi, “Gear fault diagnosis using transmission error and
ensemble empirical mode decomposition,” Mech. Syst. Signal Process., vol. 108, pp. 262–
275, 2018, doi: https://doi.org/10.1016/j.ymssp.2018.02.028.

[7] X. Liang, M. J. Zuo, and Z. Feng, “Dynamic modeling of gearbox faults: A review,” Mech.
Syst. Signal Process., vol. 98, pp. 852–876, Jan. 2018, doi:
10.1016/J.YMSSP.2017.05.024.

[8] Z. Y. Chin, P. Borghesani, Y. Mao, W. A. Smith, and R. B. Randall, “Use of transmission

error for a quantitative estimation of root-crack severity in gears,” Mech. Syst. Signal
Process., vol. 171, p. 108957, May 2022, doi: 10.1016/j.ymssp.2022.108957.

[9] Z. Y. Chin, W. A. Smith, P. Borghesani, R. B. Randall, and Z. Peng, “Absolute

transmission error: A simple new tool for assessing gear wear,” Mech. Syst. Signal
Process., vol. 146, p. 107070, Jan. 2021, doi: 10.1016/j.ymssp.2020.107070.

[10] “Appearance of Gear Teeth - Terminology of Wear and Failure,” ANSI/AGMA 1010-F14,
2014.

[11] H. Nevzat Özgüven and D. R. Houser, “Mathematical models used in gear dynamics—A

127
 review,” J. Sound Vib., vol. 121, no. 3, pp. 383–411, 1988, doi:
https://doi.org/10.1016/S0022-460X(88)80365-1.

[12] H. Ma, J. Zeng, R. Feng, X. Pang, Q. Wang, and B. Wen, “Review on dynamics of cracked
gear systems,” Engineering Failure Analysis, vol. 55. Pergamon, pp. 224–245, Sep. 01,
2015, doi: 10.1016/j.engfailanal.2015.06.004.

[13] D. C. H. Yang and J. Y. Lin, “Hertzian damping, tooth friction and bending elasticity in
gear impact dynamics,” J. Mech. Des. Trans. ASME, vol. 109, no. 2, pp. 189–196, Jun.
1987, doi: 10.1115/1.3267437.

[14] F. Chaari, T. Fakhfakh, and M. Haddar, “Analytical modelling of spur gear tooth crack
and influence on gearmesh stiffness,” Eur. J. Mech. - A/Solids, vol. 28, no. 3, pp. 461–468,
May 2009, doi: 10.1016/J.EUROMECHSOL.2008.07.007.

[15] H. Ma, R. Song, X. Pang, and B. Wen, “Time-varying mesh stiffness calculation of
cracked spur gears,” Eng. Fail. Anal., vol. 44, pp. 179–194, 2014, doi:
https://doi.org/10.1016/j.engfailanal.2014.05.018.

[16] Y. Pandya and A. Parey, “Failure path based modified gear mesh stiffness for spur gear
pair with tooth root crack,” Eng. Fail. Anal., vol. 27, pp. 286–296, Jan. 2013, doi:
10.1016/J.ENGFAILANAL.2012.08.015.

[17] O. D. Mohammed, M. Rantatalo, and J.-O. Aidanpää, “Improving mesh stiffness

calculation of cracked gears for the purpose of vibration-based fault analysis,” Eng. Fail.
Anal., vol. 34, pp. 235–251, Dec. 2013, doi: 10.1016/J.ENGFAILANAL.2013.08.008.

[18] X. Liang, H. Zhang, M. J. Zuo, and Y. Qin, “Three new models for evaluation of standard
involute spur gear mesh stiffness,” Mech. Syst. Signal Process., vol. 101, pp. 424–434,
2018, doi: https://doi.org/10.1016/j.ymssp.2017.09.005.

[19] W. D. Mark, Performance-based gear metrology: kinematic-transmission-error

computation and diagnosis. John Wiley \& Sons, 2012.

[20] Z. Chen and Y. Shao, “Dynamic simulation of spur gear with tooth root crack propagating
along tooth width and crack depth,” Eng. Fail. Anal., vol. 18, no. 8, pp. 2149–2164, Dec.
2011, doi: 10.1016/j.engfailanal.2011.07.006.

[21] O. D. Mohammed and M. Rantatalo, “Dynamic response and time-frequency analysis for
gear tooth crack detection,” Mech. Syst. Signal Process., vol. 66–67, pp. 612–624, 2016,
doi: https://doi.org/10.1016/j.ymssp.2015.05.015.

128
[22] G. Litak and M. I. Friswell, “Dynamics of a Gear System with Faults in Meshing Stiffness,”
Nonlinear Dyn., vol. 41, no. 4, pp. 415–421, 2005, doi: 10.1007/s11071-005-1398-y.

[23] S. Jia and I. Howard, “Comparison of localised spalling and crack damage from dynamic
modelling of spur gear vibrations,” Mech. Syst. Signal Process., vol. 20, no. 2, pp. 332–
349, Feb. 2006, doi: 10.1016/J.YMSSP.2005.02.009.

[24] Z. Wan, H. Cao, Y. Zi, W. He, and Z. He, “An improved time-varying mesh stiffness
algorithm and dynamic modeling of gear-rotor system with tooth root crack,” Eng. Fail.
Anal., vol. 42, pp. 157–177, 2014, doi: https://doi.org/10.1016/j.engfailanal.2014.04.005.

[25] H. Ma, X. Pang, R. Feng, R. Song, and B. Wen, “Fault features analysis of cracked gear
considering the effects of the extended tooth contact,” Eng. Fail. Anal., vol. 48, pp. 105–
120, 2015, doi: https://doi.org/10.1016/j.engfailanal.2014.11.018.

[26] R. Errichello, “Gear Bending Fatigue Failure and Bending Life Analysis BT -
Encyclopedia of Tribology,” Q. J. Wang and Y.-W. Chung, Eds. Boston, MA: Springer
US, 2013, pp. 1467–1468.

[27] P. J. L. Fernandes, “Tooth bending fatigue failures in gears,” Eng. Fail. Anal., 1996, doi:
10.1016/1350-6307(96)00008-8.

[28] N. Perez, “Fatigue Crack Growth BT - Fracture Mechanics,” N. Perez, Ed. Cham:
Springer International Publishing, 2017, pp. 327–372.

[29] P. Paris and F. Erdogan, “A critical analysis of crack propagation laws,” J. Basic Eng., vol.
85, no. 4, pp. 528–533, Dec. 1963, [Online]. Available:
http://dx.doi.org/10.1115/1.3656900.

[30] D. G. Lewicki and R. Ballarini, “Effect of Rim Thickness on Gear Crack Propagation Path,”
J. Mech. Des., vol. 119, no. 1, pp. 88–95, 1997, doi: 10.1115/1.2828793.

[31] D. G. Lewicki and R. Ballarini, “Rim thickness effects on gear crack propagation life,”
Int. J. Fract., vol. 87, no. 1, pp. 59–86, 1997, doi: 10.1023/A:1007368801853.

[32] D. G. Lewicki and R. Ballarini, “Gear crack propagation investigations,” TriboTest, vol.
5, no. 2, pp. 157–172, 1998, doi: 10.1002/tt.3020050206.

[33] D. G. Lewicki, R. F. Handschuh, L. E. Spievak, P. A. Wawrzynek, and A. R. Ingraffea,

“Consideration of Moving Tooth Load in Gear Crack Propagation Predictions,” J. Mech.
Des., 2001, doi: 10.1115/1.1338118.

129
[34] D. G. Lewicki, “Gear Crack Propagation Path Studies-Guidelines for Ultra-Safe Design,”
J. Am. Helicopter Soc., vol. 47, no. 1, pp. 64–72, 2002, doi: 10.4050/JAHS.47.64.

[35] D. B. Stringer, B. D. Dykas, K. E. LaBerge, A. J. Zakrajsek, and R. F. Handschuh, “A new

high-speed, high-cycle, gear-tooth bending fatigue test capability,” 2011.

[36] B. Stringer, K. E. Laberge, C. J. Burdick, and B. A. Fields, “Natural fatigue crack initiation
and detection in high quality spur gears,” New York, 2012.

[37] J. G. Verma, S. Kumar, and P. K. Kankar, “Crack growth modeling in spur gear tooth and
its effect on mesh stiffness using extended finite element method,” Eng. Fail. Anal., vol.
94, pp. 109–120, 2018, doi: https://doi.org/10.1016/j.engfailanal.2018.07.032.

[38] W. D. Mark and C. P. Reagor, “Static-transmission-error vibratory-excitation

contributions from plastically deformed gear teeth caused by tooth bending-fatigue
damage,” Mech. Syst. Signal Process., vol. 21, no. 2, pp. 885–905, 2007, doi:
10.1016/j.ymssp.2006.05.002.

[39] S. Wu, M. J. Zuo, and A. Parey, “Simulation of spur gear dynamics and estimation of fault
growth,” J. Sound Vib., vol. 317, no. 3–5, pp. 608–624, Nov. 2008, doi:
10.1016/J.JSV.2008.03.038.

[40] Z. Chen and Y. Shao, “Mesh stiffness calculation of a spur gear pair with tooth profile
modification and tooth root crack,” Mech. Mach. Theory, vol. 62, pp. 63–74, Apr. 2013,
doi: 10.1016/j.mechmachtheory.2012.10.012.

[41] H. Ma, X. Pang, R. Feng, J. Zeng, and B. Wen, “Improved time-varying mesh stiffness
model of cracked spur gears,” Eng. Fail. Anal., vol. 55, pp. 271–287, 2015, doi:
https://doi.org/10.1016/j.engfailanal.2015.06.007.

[42] P. J. L. Fernandes and C. McDuling, “Surface contact fatigue failures in gears,” Eng. Fail.
Anal., vol. 4, no. 2, pp. 99–107, 1997, doi: https://doi.org/10.1016/S1350-6307(97)00006-
X.

[43] B. Errichello and J. Muller, “How to analyze gear failures.”

https://www.machinerylubrication.com/Read/150/gear-failures (accessed Nov. 15, 2021).

[44] R. G. Budynas and J. K. Nisbett, Shigley’s Mechanical Engineering Design, vol. New
York,. 2010.

[45] A. Kahraman and H. Ding, “Wear in Gears BT - Encyclopedia of Tribology,” Q. J. Wang

and Y.-W. Chung, Eds. Boston, MA: Springer US, 2013, pp. 3993–4001.

130
[46] V. Vullo, “Surface Durability (Pitting) of Spur and Helical Gears BT - Gears: Volume 2:
Analysis of Load Carrying Capacity and Strength Design,” V. Vullo, Ed. Cham: Springer
International Publishing, 2020, pp. 73–147.

[47] O. Asi, “Fatigue failure of a helical gear in a gearbox,” Eng. Fail. Anal., vol. 13, no. 7, pp.
1116–1125, 2006, doi: https://doi.org/10.1016/j.engfailanal.2005.07.020.

[48] J. F. Archard, W. Hirst, and T. E. Allibone, “The wear of metals under unlubricated
conditions,” Proc. R. Soc. London. Ser. A. Math. Phys. Sci., vol. 236, no. 1206, pp. 397–
410, Aug. 1956, doi: 10.1098/rspa.1956.0144.

[49] A. Flodin and S. Andersson, “Simulation of mild wear in spur gears,” Wear, vol. 207, no.
1–2, pp. 16–23, Jun. 1997, doi: 10.1016/S0043-1648(96)07467-4.

[50] W. Chen, Y. Lei, Y. Fu, and L. Hou, “A study of effects of tooth surface wear on time-
varying mesh stiffness of external spur gear considering wear evolution process,” Mech.
Mach. Theory, vol. 155, p. 104055, Jan. 2021, doi:
10.1016/J.MECHMACHTHEORY.2020.104055.

[51] M. J. B. T.-T. H. (Second E. NEALE, Ed., “D14 - Wear debris analysis,” Oxford:
Butterworth-Heinemann, 1995, p. D14.1-D14.9.

[52] W. J. Staszewski and G. R. Tomlinson, “Application of the wavelet transform to fault

detection in a spur gear,” Mechanical Systems and Signal Processing, vol. 8, no. 3. pp.
289–307, 1994, doi: 10.1006/mssp.1994.1022.

[53] F. Combet, L. Gelman, and G. Lapayne, “Novel detection of local tooth damage in gears
by the wavelet bicoherence,” Mech. Syst. Signal Process., vol. 26, no. 1, pp. 218–228, Jan.
2012, doi: 10.1016/j.ymssp.2011.07.002.

[54] P. D. McFadden and J. D. Smith, “A signal processing technique for detecting local defects
in a gear from the signal average of the vibration,” Proc. Inst. Mech. Eng. Part C J. Mech.
Eng. Sci., 1985, doi: 10.1243/PIME_PROC_1985_199_125_02.

[55] P. D. McFadden, “Detecting Fatigue Cracks in Gears by Amplitude and Phase

Demodulation of the Meshing Vibration,” J. Vib. Acoust. Stress. Reliab. Des., vol. 108,
no. 2, pp. 165–170, 1986, doi: 10.1115/1.3269317.

[56] D. Brie, M. Tomczak, H. Oehlmann, and A. Richard, “Gear crack detection by adaptive
amplitude and phase demodulation,” Mech. Syst. Signal Process., vol. 11, no. 1, pp. 149–
167, 1997, doi: https://doi.org/10.1006/mssp.1996.0068.

131
[57] A. Belsak and J. Flasker, “Detecting cracks in the tooth root of gears,” Eng. Fail. Anal.,
vol. 14, no. 8, pp. 1466–1475, 2007, doi:
https://doi.org/10.1016/j.engfailanal.2007.01.013.

[58] A. Belsak and J. Flasker, “Method for detecting fatigue crack in gears,” Theor. Appl. Fract.
Mech., vol. 46, no. 2, pp. 105–113, 2006, doi:
https://doi.org/10.1016/j.tafmec.2006.07.002.

[59] K. R. Fyfe and E. D. S. Munck, “Analysis of computed order tracking,” Mech. Syst. Signal
Process., vol. 11, no. 2, pp. 187–205, 1997, doi: 10.1006/mssp.1996.0056.

[60] F. Bonnardot, M. El Badaoui, R. B. Randall, J. Danière, and F. Guillet, “Use of the

acceleration signal of a gearbox in order to perform angular resampling (with limited speed
fluctuation),” Mech. Syst. Signal Process., vol. 19, no. 4, pp. 766–785, Jul. 2005, doi:
10.1016/j.ymssp.2004.05.001.

[61] M. D. Coats and R. B. Randall, “Single and multi-stage phase demodulation based order-
tracking,” Mech. Syst. Signal Process., vol. 44, no. 1–2, pp. 86–117, 2014, doi:
10.1016/j.ymssp.2013.09.016.

[62] K. S. Wang, D. Guo, and P. S. Heyns, “The application of order tracking for vibration
analysis of a varying speed rotor with a propagating transverse crack,” Eng. Fail. Anal.,
vol. 21, pp. 91–101, Apr. 2012, doi: 10.1016/j.engfailanal.2011.11.020.

[63] S. Schmidt, P. S. Heyns, and J. P. de Villiers, “A tacholess order tracking methodology

based on a probabilistic approach to incorporate angular acceleration information into the
maxima tracking process,” Mech. Syst. Signal Process., vol. 100, pp. 630–646, Feb. 2018,
doi: 10.1016/j.ymssp.2017.07.053.

[64] P. Borghesani, P. Pennacchi, S. Chatterton, and R. Ricci, “The velocity synchronous

discrete Fourier transform for order tracking in the field of rotating machinery,” Mech.
Syst. Signal Process., vol. 44, no. 1–2, pp. 118–133, Feb. 2014, doi:
10.1016/J.YMSSP.2013.03.026.

[65] P. D. Mcfadden, “A revised model for the extraction of periodic waveforms by time
domain averaging,” pp. 83–95, 1987.

[66] S. Braun, “The synchronous (time domain) average revisited,” Mech. Syst. Signal Process.,
vol. 25, no. 4, pp. 1087–1102, 2011, doi: 10.1016/j.ymssp.2010.07.016.

[67] E. B. Halim, M. A. A. Shoukat Choudhury, S. L. Shah, and M. J. Zuo, “Time domain

132
 averaging across all scales: A novel method for detection of gearbox faults,” Mech. Syst.
Signal Process., vol. 22, no. 2, pp. 261–278, Feb. 2008, doi: 10.1016/j.ymssp.2007.08.006.

[68] P. D. McFadden and J. D. Smith, “Effect of Transmission Path on Measured Gear

Vibration,” J. Vib. Acoust. Stress. Reliab. Des., vol. 108, no. March 1986, 1987.

[69] W. D. Mark, H. Lee, R. Patrick, and J. D. Coker, “A simple frequency-domain algorithm

for early detection of damaged gear teeth,” Mech. Syst. Signal Process., vol. 24, no. 8, pp.
2807–2823, 2010, doi: 10.1016/j.ymssp.2010.04.004.

[70] J. A. Hines and W. D. Mark, “Bending-fatigue damage-detection on notched-tooth spiral-

bevel gears using the average-log-ratio, ALR, algorithm,” Mech. Syst. Signal Process., vol.
43, no. 1, pp. 44–56, 2014, doi: https://doi.org/10.1016/j.ymssp.2013.09.013.

[71] M. E. Wagner, W. D. Mark, and A. C. Isaacson, “Implementation of the Average-Log-

Ratio ALR gear-damage detection algorithm on gear-fatigue-test recordings,” Mech. Syst.
Signal Process., vol. 154, 2021, doi: 10.1016/j.ymssp.2020.107590.

[72] W. WANG, “EARLY DETECTION OF GEAR TOOTH CRACKING USING THE

RESONANCE DEMODULATION TECHNIQUE,” Mech. Syst. Signal Process., vol. 15,
no. 5, pp. 887–903, 2001, doi: https://doi.org/10.1006/mssp.2001.1416.

[73] X. Yang, M. J. Zuo, and Z. Tian, “Development of crack induced impulse-based condition
indicators for early tooth crack severity assessment,” Mech. Syst. Signal Process., vol. 165,
p. 108327, Feb. 2022, doi: 10.1016/J.YMSSP.2021.108327.

[74] D. Wang, “K-nearest neighbors based methods for identification of different gear crack
levels under different motor speeds and loads: Revisited,” Mech. Syst. Signal Process., vol.
70–71, pp. 201–208, 2016, doi: https://doi.org/10.1016/j.ymssp.2015.10.007.

[75] Y. Lei and M. J. Zuo, “Gear crack level identification based on weighted K nearest
neighbor classification algorithm,” Mech. Syst. Signal Process., vol. 23, no. 5, pp. 1535–
1547, 2009, doi: https://doi.org/10.1016/j.ymssp.2009.01.009.

[76] C. U. Grosse and M. Ohtsu, Acoustic emission testing. Springer Science \& Business
Media, 2008.

[77] C. K. Tan and D. Mba, “Identification of the acoustic emission source during a
comparative study on diagnosis of a spur gearbox,” Tribol. Int., vol. 38, no. 5, pp. 469–
480, May 2005, doi: 10.1016/J.TRIBOINT.2004.10.007.

[78] C. K. Tan, P. Irving, and D. Mba, “A comparative experimental study on the diagnostic

133
 and prognostic capabilities of acoustics emission, vibration and spectrometric oil analysis
for spur gears,” Mech. Syst. Signal Process., vol. 21, no. 1, pp. 208–233, Jan. 2007, doi:
10.1016/J.YMSSP.2005.09.015.

[79] P. Feng, P. Borghesani, W. A. Smith, and Z. Peng, “Model-based surface roughness

estimation using acoustic emission signals,” Tribol. Int., vol. 144, p. 106101, Apr. 2020,
doi: 10.1016/J.TRIBOINT.2019.106101.

[80] S. K. Roy, A. R. Mohanty, and C. S. Kumar, “Fault detection in a multistage gearbox by

time synchronous averaging of the instantaneous angular speed,” JVC/Journal Vib.
Control, vol. 22, no. 2, pp. 468–480, 2016, doi: 10.1177/1077546314533582.

[81] B. Li, X. Zhang, and J. Wu, “New procedure for gear fault detection and diagnosis using
instantaneous angular speed,” Mech. Syst. Signal Process., vol. 85, pp. 415–428, Feb. 2017,
doi: 10.1016/j.ymssp.2016.08.036.

[82] B. Li, X. Zhang, and T. Wu, “Measurement of Instantaneous Angular Displacement

Fluctuation and its applications on gearbox fault detection,” ISA Trans., 2018, doi:
10.1016/j.isatra.2018.01.034.

[83] J. D. Smith, Gear noise and vibration, 2nd Ed. New York: Marcel Dekker Inc., 2003.

[84] R. G. Munro, “A review of the theory and measurement of gear transmission error,” in 1st
International Conference on Gearbox Noise and Vibration, 1990, pp. 3–10.

[85] H. Endo and N. Sawalhi, “Gearbox simulation models with gear and bearing faults,” in
Mechanical Engineering, 2012.

[86] S. L. Harris, “Dynamic loads on the teeth of spur gears,” Proc. Inst. Mech. Eng., vol. 172,
no. 1, pp. 87–112, Jun. 1958, doi: 10.1243/PIME_PROC_1958_172_017_02.

[87] H. Walker, “Gear tooth deflection and profile modification (Part 1),” Eng., p. 410 and 435,
1938.

[88] H. Walker, “Gear tooth deflection and profile modification (Part 2),” Eng., pp. 102–104,
1940.

[89] R. W. Gregory, S. L. Harris, and R. G. Munro, “Dynamic behaviour of spur gears,” Proc.
Inst. Mech. Eng., vol. 178, no. 1, pp. 207–218, 1963, doi: 10.1177/002034836317800130.

[90] R. W. Gregory, S. L. Harris, and R. G. Munro, “A method of measuring transmission error

in spur gears of 1:1 ratio,” J. Sci. Instrum., vol. 40, no. 1, pp. 5–9, 1963, doi: 10.1088/0950-

134
 7671/40/1/303.

[91] R. G. Munro, “The DC component of gear transmission error,” Proceedings of the

International Power Transmission and Gearing Conference. pp. 467–470, 1989.

[92] C. Timms, “Recent developments in spur and helical gears,” Prod. Eng., vol. 39, no. 6, pp.
321–340, 1960, doi: 10.1049/tpe.1960.0037.

[93] H. Endo, R. B. Randall, and C. Gosselin, “Differential diagnosis of spall vs. cracks in the
gear tooth fillet region: experimental validation,” Mech. Syst. Signal Process., vol. 23, no.
3, pp. 636–651, 2009, doi: https://doi.org/10.1016/j.ymssp.2008.08.015.

[94] D. Peng, W. A. Smith, P. Borghesani, R. B. Randall, and Z. Peng, “Comprehensive planet

gear diagnostics: use of transmission error and mesh phasing to distinguish localised fault
types and identify faulty gears,” Mech. Syst. Signal Process., vol. 127, pp. 531–550, 2019,
doi: https://doi.org/10.1016/j.ymssp.2019.03.024.

[95] D. Peng, W. A. Smith, R. B. Randall, and Z. Peng, “Use of mesh phasing to locate faulty
planet gears,” Mech. Syst. Signal Process., vol. 116, pp. 12–24, 2019, doi:
https://doi.org/10.1016/j.ymssp.2018.06.035.

[96] O. Lundvall and A. Klarbring, “Prediction of transmission error in spur gears as a

consequence of wear,” Mech. Struct. Mach., vol. 29, no. 4, pp. 431–449, Nov. 2001, doi:
10.1081/SME-100107621.

[97] W. D. Mark, A. C. Isaacson, and M. E. Wagner, “Transmission-error frequency-domain-

behavior of failing gears,” Mech. Syst. Signal Process., vol. 115, pp. 102–119, 2019, doi:
https://doi.org/10.1016/j.ymssp.2018.05.036.

[98] R. B. Randall, D. Peng, and W. A. Smith, “Using measured transmission error for
diagnostics of gears,” SIRM, Copenhagen, 2019.

[99] R. Randall, Z. Y. Chin, W. Smith, and P. Borghesani, “Measurement and use of

transmission error for diagnostics of gears,” Survishno, Lyon, 2019.

[100] R. Lu, M. R. Shahriar, P. Borghesani, R. B. Randall, and Z. Peng, “Removal of transfer

function effects from transmission error measurements using cepstrum-based operational
modal analysis,” Mech. Syst. Signal Process., vol. 165, p. 108324, 2022, doi:
https://doi.org/10.1016/j.ymssp.2021.108324.

[101] C. J. Li and H. Lee, “Gear fatigue crack prognosis using embedded model, gear dynamic
model and fracture mechanics,” Mech. Syst. Signal Process., vol. 19, no. 4, pp. 836–846,

135
 Jul. 2005, doi: 10.1016/J.YMSSP.2004.06.007.

[102] A. Palermo, L. Britte, K. Janssens, D. Mundo, and W. Desmet, “The measurement of gear
transmission error as an NVH indicator: theoretical discussion and industrial application
via low-cost digital encoders to an all-electric vehicle gearbox,” Mech. Syst. Signal
Process., vol. 110, pp. 368–389, 2018, doi: https://doi.org/10.1016/j.ymssp.2018.03.005.

[103] J. Mączak, “Local meshing plane analysis as a source of information about the gear
quality,” Mech. Syst. Signal Process., vol. 38, no. 1, pp. 154–164, 2013, doi:
https://doi.org/10.1016/j.ymssp.2012.09.012.

[104] C. J. Li, H. Lee, and S. H. Choi, “Estimating size of gear tooth root crack using embedded
modelling,” Mech. Syst. Signal Process., vol. 16, no. 5, pp. 841–852, 2002, doi:
https://doi.org/10.1006/mssp.2001.1452.

[105] Y. Chen, S. Schmidt, P. S. Heyns, and M. J. Zuo, “A time series model-based method for
gear tooth crack detection and severity assessment under random speed variation,” Mech.
Syst. Signal Process., vol. 156, p. 107605, 2021, doi:
https://doi.org/10.1016/j.ymssp.2020.107605.

[106] H. Chang, P. Borghesani, W. A. Smith, and Z. Peng, “Application of surface replication

combined with image analysis to investigate wear evolution on gear teeth – A case study,”
Wear, vol. 430–431, pp. 355–368, 2019, doi: https://doi.org/10.1016/j.wear.2019.05.024.

[107] H. N. Özgüven and D. R. Houser, “Dynamic analysis of high speed gears by using loaded
static transmission error,” J. Sound Vib., vol. 125, no. 1, pp. 71–83, 1988, doi:
https://doi.org/10.1016/0022-460X(88)90416-6.

[108] R. G. Parker, S. M. Vijayakar, and T. Imajo, “NON-LINEAR DYNAMIC RESPONSE

OF A SPUR GEAR PAIR: MODELLING AND EXPERIMENTAL COMPARISONS,”
J. Sound Vib., vol. 237, no. 3, pp. 435–455, Oct. 2000, doi: 10.1006/JSVI.2000.3067.

[109] M. Benatar, M. Handschuh, A. Kahraman, and D. Talbot, “Static and Dynamic

Transmission Error Measurements of Helical Gear Pairs With Various Tooth
Modifications,” J. Mech. Des., vol. 141, no. 10, May 2019, doi: 10.1115/1.4043586.

[110] P. J. Sweeney and R. B. Randall, “Gear transmission error measurement using phase
demodulation,” Proc. Inst. Mech. Eng. Part C J. Mech. Eng. Sci., vol. 210, no. 3, pp. 201–
213, 1996, doi: 10.1243/PIME_PROC_1996_210_190_02.

[111] R. B. Randall, “A new method of modeling gear faults,” Trans. Asme J. Mech. Des., vol.

136
 104, no. 2, Apr. 1982, pp. 259–267, 1982, doi: 10.1115/1.3256334.

[112] I. O. MacConochie and A. Cameron, “The measurement of oil-film thickness in gear teeth,”
J. Basic Eng., vol. 82, no. 1, pp. 29–34, Mar. 1960, [Online]. Available:
http://dx.doi.org/10.1115/1.3662549.

[113] C. Xie, L. Hua, X. Han, J. Lan, X. Wan, and X. Xiong, “Analytical formulas for gear body-
induced tooth deflections of spur gears considering structure coupling effect,” Int. J. Mech.
Sci., vol. 148, pp. 174–190, 2018, doi: https://doi.org/10.1016/j.ijmecsci.2018.08.022.

[114] Z. Chen, J. Zhang, W. Zhai, Y. Wang, and J. Liu, “Improved analytical methods for
calculation of gear tooth fillet-foundation stiffness with tooth root crack,” Eng. Fail. Anal.,
vol. 82, pp. 72–81, 2017, doi: https://doi.org/10.1016/j.engfailanal.2017.08.028.

[115] R. W. Cornell, “Compliance and Stress Sensitivity of Spur Gear Teeth,” J. Mech. Des.,
vol. 103, no. 2, pp. 447–459, 1981, doi: 10.1115/1.3254939.

[116] K. Feng, W. A. Smith, P. Borghesani, R. B. Randall, and Z. Peng, “Use of cyclostationary

properties of vibration signals to identify gear wear mechanisms and track wear evolution,”
Mech. Syst. Signal Process., vol. 150, Mar. 2021, doi: 10.1016/j.ymssp.2020.107258.

137