Mechanical Systems

and
Signal Processing
Mechanical Systems and Signal Processing 21 (2007) 480501
Monitoring low cycle fatigue damage in turbine blade
using vibration characteristics
Sandeep Kumar, Niranjan Roy, Ranjan Ganguli

Department of Aerospace Engineering, Indian Institute of Science, Bangalore 560012, India

Received 19 July 2004; received in revised form 25 February 2005; accepted 28 February 2005
Available online 25 May 2005
Abstract
A nite element based approach is used to simulate the evolution of low cycle fatigue damage in a turbine
blade. The turbine blade is modelled as a rotating Timoshenko beam with taper and twist. A damage
growth model for low cycle fatigue damage developed using a continuum mechanics approach is integrated
with the nite element model. Numerical results are obtained to study the effect of damage growth on the
rotating frequencies. It is found that low cycle fatigue causes sufcient degradation in blade stiffness for
changes in rotating frequency to be used as an indicator to track damage growth. Continuum damage
mechanics models in conjunction with nite element analysis are used to develop thresholds for damage
indicators. By placing suitable threshold on the frequency change, it is possible to detect the onset of the
nal stage of damage in the structure before failure occurs.
r 2005 Elsevier Ltd. All rights reserved.
Keywords: Low cycle fatigue; Vibration; Damage detection; Condition monitoring; Turbine
1. Introduction
Turbine blades undergo cyclic loading causing structural deterioration which can lead to
failure. It is important to know how much damage has taken place at any particular time to
monitor the condition or health of the blade and to avoid any catastrophic failure of the blades.
ARTICLE IN PRESS
www.elsevier.com/locate/jnlabr/ymssp
0888-3270/$ - see front matter r 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ymssp.2005.02.011

Corresponding author. Fax: +91 080 334 5134.

E-mail address: [email protected] (R. Ganguli).
Several studies look at damage at a given time during the operational history of the structure. This
is typically called diagnostics and involves detection, location and isolation of damage from a set
of measured variables.
According to researchers [1], the detection function is most fundamental as it points out if the
damage is present or not. However, some level of damage due to microcracks and other defects is
always present in a structure. The important issue of indicating when to detect a damage depends
on how much life is left in the structure and has not received much attention in the literature. It is
not advantageous to detect small levels of damage in a structure. It would be useful if a damage
detection was triggered slightly before nal failure.
The subject of prognostics involves predicting the evolution of structural or vibration
characteristics of the system with time and is important for prediction of failure due to
operational deterioration. Some recent studies have considered dynamical systems approaches to
model damage growth based on differential equations [2] while others have used physics based
models [3].
The stiffness of the structure is gradually reduced with crack growth, and stiffness is related to
the vibrational characteristics of the structure. The decreased frequency shows that stiffness of the
structure is decreasing and thus serves as a damage indicator for monitoring crack growth in the
structure. Selected studies have looked at modelling turbine blades as rotating Timoshenko beams
with twist and taper [46]. Some studies have addressed damage in such beams using vibrational
characteristics [7,8]. However, these studies typically address damage at a given time point in the
operational history and do not look at the effect of damage growth on the vibrational
characteristics. In additions, turbine blades are designed to sustain considerable amount of
accumulated damage prior to failure. Therefore, it is desirable to indicate that a blade is
damaged at the point when its operational life is almost over.
To study the structural dynamic behaviour of the damaged beam, a damage model needs to be
integrated into the nite element analysis. A number of damage models are available in Ref. [9].
Some studies use fracture mechanics based approach to include a crack in the nite element
formulation [10]. The effect of damage growth can be modelled using crack growth models such as
the Paris law [11,12]. Another approach is to use a phenomenological model devised from
experimental measurements which relates the loss of stiffness at the continuum level of the
structure to the growth of damage. Typically, most steam turbine rotor fractures are caused by
low cycle fatigue (LCF) [13]. The high strains which cause LCF typically occurs during the cold
start and sliding parameter stop phases of steam turbine operation. Due to high strain rates
implicit in LCF, it is typically accompanied by plastic deformation. LCF is also the most
signicant life reducing failure mechanism for military aircraft engines due to the throttle changes
experienced during operation [14]. Therefore, the study of LCF effect on turbine blades is an
important problem.
Most work on LCF in based on linear damage accumulation theory [13]. However, this theory
results in crude damage accumulation calculations and the predicted results are quite different in
practice. In reality, LCF damage accumulation is a non-linear process. Continuum damage
mechanics (CDM) provides a way to describe the entire failure process from micro-rupture
initiation and visual crack formation to structure failure. A damage model based on CDM for
LCF damage analysis is chosen in this study. LCF occurs after relatively low number of cycles
(typically o10 000) of high stress amplitude. The CDM models are easier to include in nite
ARTICLE IN PRESS
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 481
element analysis and the nite element analysis and damage growth analysis can be effectively
decoupled. A recent overview of damage growth models is given in [9].
The CDM models use a damage variable D 1 E=E
0
which ranges from D 0 for the
undamaged material to D 1 for complete failure. The CDM models the evolution of D from 0
to 1. Since it is very difcult to measure D in an operational turbine blade, another damage
indicator which is dependent on D needs to be used. Frequency is one such indicator and several
methods have been developed in recent years for frequency measurement of turbine blades.
The simplest method of determining frequencies is by using blade mounted strain gages.
Typically, blade mounted gages can be used to detect the blade tip deection. Lawson and Ivey
[15] mention that two strain gages per blade can be used to detect the rst four modes of blade
vibration. They developed an electronic circuit to energise the blade mounted strain gages and
amplify the resulting signal. This system was developed for on rotor operation. More details are
available in [15].
Another promising approach for determining vibration characteristic of rotating engine blades
is the use of blade tip timing methods. These methods have evolved because of the disadvantages
of blade mounted strain gages which include a complex installation procedure, a limited number
of gages and the possibility of failure of the strain gauges due to high temperature experienced in
the turbines. Strain gauges can also require complex telemetry of slip ring systems and can
interface with the mechanical properties of the bladed assembly [16]. Tip timing methods typically
use optical probes mounted on the engine casing and can measure the motion of each blade. These
probes measure the time of arrival (TOA) of each blade relative to a once-per-revolution (OPR) or
multiple-per-revolution (MPR) probe mounted on the rotor shaft. The difference between the
TOA of a vibrating blade and the computed TOA of a non-vibrating blade provides the raw data
which is used to determine the instantaneous blade displacement. Optical systems can be used to
obtain vibration amplitude using curve tting and frequencies using Fourier analysis. According
to Carrington et al. [16], tip timing data acquisition in now its fourth generation and is sufciently
advanced. Very recent work has also looked at the potential of a dual use capacitance probe
sensor to measure tip timing and tip clearance as alternatives to optical devices [15]. It is therefore
possible to measure the frequencies of turbine blade quite accurately using the tip timing method.
The present study incorporates a CDM based damaged model for LCF damage into the nite
element analysis of a rotating turbine blade. Numerical results are obtained to track the changes
in frequencies with damage growth. A method of detecting the nal stage of damage based on
placing thresholds on the frequency change is proposed.
2. Formulation
2.1. Modelling of turbine blade
The turbine blade is modelled as a tapered, twisted and rotating Timoshenko beam. The
geometry of the beam is dened in Fig. 1(a), the degrees of freedom for the element in Fig. 1(b),
the cross-section in Fig. 1(c) and the nite element used is shown schematically in Fig. 1(d).
The total strain energy U of a beam of length l, due to bending and shear defor-
mation (in the plane and out of plane directions) including rotary inertia and rotation effects is
ARTICLE IN PRESS
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 482
given by [6]
U
_
l
0
EI
xx
2
q
2
w
b
qz
2
_ _
2
EI
xy
q
2
w
b
qz
2
q
2
v
b
qz
2

EI
yy
2
q
2
v
b
qz
2
_ _
2
_ _ _

mAG
2
qw
s
qz
_ _
2

qv
s
qz
_ _
2
_ __
dz
1
2
_
l
0
Pz
qw
b
qz

qw
b
qz
_ _
2
dz

1
2
_
l
0
Pz
qv
s
qz

qv
s
qz
_ _
2
dz
_
l
0
p
w
zw
b
w
s
dz

_
l
0
p
v
zv
b
v
s
dz, 1
ARTICLE IN PRESS
x
b
1
y
z
b
2
h
2
h
1
u
1
u
3
u
9
u
11
u
4
u
6
u
7
u
8
u
13 u
15
u
16
u
14
u
10 u
12
u
5
u
2
(a) (b)
x
y
x
y

x
x
y
y
x
L
e
x

Hub
Finite element
under consideration
z
e
o
z
x
l
(c) (d)
Fig. 1. (a) An element of tapered and twisted beam; (b) degrees of freedom of an element; (c) angle of twist y;
(d) rotation of tapered beam.
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 483
where
Pz
_
Le
ez
e
z
mO
2
x dx
rAO
2
2g
L e
2
e z
e
z
2

rAO
2
g
eL
1
2
L
2
ez
e

1
2
z
2
e
_ _
e z
e
z
1
2
z
2
_ _
, 2
p
w
z
rAO
2
g
w
b
w
s
, (3)
p
v
z
rAO
2
g
v
b
v
s
. (4)
The kinetic energy of the element T including the effect of shear deformation and rotary inertia is
given by
T
_
1
0
rA
2g
qw
b
qt

qw
s
qt
_ _
2

rA
2g
qv
b
qt

qv
s
qt
_ _
2

rI
yy
2g
q
2
v
b
qzqt
_ _
2
_

rI
xy
g
q
2
w
b
qzqt
_ _
q
2
v
b
qzqt
_ _

rI
xx
2g
q
2
w
b
qzqt
_ _
2
_
dz. 5
The deformations can be discretised in terms of shape functions. Cubic polynomials are used for
the out-of-plane bending w
b
and in plane bending v
b
. Cubic polynomials are also used as shape
functions for the shear deformation w
s
and v
s
. Using the energy expressions and the nite element
discretisation, the element level mass and stiffness matrices are calculated. After assembling the
matrices and applying the cantilever boundary conditions we obtain
K o
2
MF 0. (6)
The stiffness and mass matrices are obtained by using expressions for the strain energy and for
kinetic energy. The details of the formulation can be found in [6] and the nite element model is
validated by comparing with the results in [6].
2.2. Fatigue damage model
There are many models for predicting damage growth due to fatigue. Some LCF models
proposed in the past have limited validity to particular cases [9,10]. The physical meaning of the
model parameters are also not clear. For engineering applications, models based on continuum
damage mechanics appear useful. They do not require detailed models of crack growth using
fracture mechanics but capture the non-linear nature of damage growth. They are based on the
ARTICLE IN PRESS
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 484
continuum damage variable D which can be dened [17] as D 1 E=E
0
. For undamaged
material, E E
0
and D 0. For complete damage, E 0 and D 1. For prognostics, we are
interested in the path of damage growth as it evolves from D 0 to 1 and the effect of such
damage growth or damage indicators such as frequencies and mode shapes. CDM models are
phenomenological and are typically obtained from experimental data and provide functional
relationship of Dt. For fatigue problems, t N where N is the number of cycles. After a search
of the literature, a useful damage model was obtained from [18]. The model addresses LCF and is
developed using a damage potential function. The condensed derivation of the model from [18] is
given in Appendix A. The damage model is
D 1 1 D
0
1
N
N
f
_ _
N
f
a
(7)
There are three parameters in the above model: D
0
, N
f
and N
f
a. These parameters can be
determined from experiments. The damage model for steam turbine blade material 2Cr13
martensitic stainless steel are derived using experiments. In [18], a material testing machine was
used for strain controlled fatigue tests. There are several methods for measuring the damage
variable D, as mentioned by Lemaitre [19]. For LCF, the best mechanisms include elasticity
modulus followed by the cycle stress amplitude method. For measurements based on elastic
stiffness, a specimen of the material needs to be machined to run mechanical tests. Here D
1 E=E
0
is used. This method needs accurate strain measurements. Typically, strain gages are
used and E is most accurately measured during unloading. Another approach which is used by
[18] is called the cycle stress amplitude method. The one-dimensional law of cyclic plasticity at
stabilisation may be written as a power relationship between the amplitude of stress range Ds and
the amplitude of strain range D at cycles:
D
Ds
K
c
_ _
M
. (8)
The above relationship is for an undamaged material. K
c
and M are material parameters. For
damaged material, the relationship becomes,
D
Ds
K
c
1 D
_ _
M
. (9)
For a cyclic test at constant amplitude of strain D, initial stress being Ds
0
and damage D may
be assumed to be zero, hence
Ds
0
K
c
D
1=M
(10)
and also from Eq. (9),
Ds 1 DK
c
D
1=M
. (11)
Combining Eqs. (10) and (11) we obtain:
D 1
Ds
Ds
0
. (12)
ARTICLE IN PRESS
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 485
The models developed for damage growth are as follows:
1. For low strain of 0:35 per cycle the number of cycles (N
f
) to produce fatigue crack resulting
in failure was 6230 and the model was
DN 1 0:906 1
N
6230
_ _
0:058
, (13)
where, D is the damage level in the structure after N cycles. For N=N
f
0:9; D 0:21 with 623
cycles left to nal failure.
2. For moderate strain of 0:5 per cycle the number of cycles (N
f
) to produce fatigue crack
resulting in failure was 1950 and the model was
DN 1 0:903 1
N
1950
_ _
0:064
. (14)
For N=N
f
0:9; D 0:27 with 195 cycles left for failure.
3. For high strain of 0:70 per cycle the number of cycles (N
f
) to produce fatigue crack resulting
in failure was 844 and the model was
DN 1 0:923 1
N
844
_ _
0:113
. (15)
For N=N
f
0:9; D 0:23 with 84 cycles left for failure.
Note that the high strain conditions where LCF happens occur only occasionally in the life of
the actual machine. For example, the average design life of a steam turbine is about 30 years.
Assuming the frequency of both cold start and sliding parameter stop are three times per year, the
total number of cold starts and sliding parameter stop are 90 times in 30 years [13].
The damage curves are shown in Fig. 2. The plots show that D varies quickly at the last stages
of the whole cycling and slowly at the middle stage from 10% to 80% of the total cycles, which is a
characteristics of LCF damage. High strain leads to faster failure of the material. In general, the
LCF damage curve can be divided into three stages. Stage 1 occurs up to N=N
f
0:1 when
damage value increases due to changes in the dislocation substructures. In stage 2, there is a slow
increase in damage value upto N=N
f
0:8. In stage 3, the damage increases more quickly to 1
due to the beginning of damage localisation and formation of fatigue micro-cracks. From a
condition monitoring viewpoint, it is useful to know when the blade has reached stage 3. The
midpoint of stage 3 of N=N
f
0:9 is therefore selected in this study as the time where the life of
the blade is almost over and the blade needs to be replaced.
2.3. Beam with fatigue damage
We want to know if fatigue damage be detected in turbine blades before it becomes catastrophic
by monitoring natural rotating frequencies. For the numerical results, the damage model
discussed in Section 2.2 is included in the nite element model in Section 2.1. The stiffness (E) of
the beam and the fatigue damage growth with number of cycle is related with the expression [17]:
DN 1 E=E
0
. Thus, Youngs modulus for each case, i.e. for low, moderate and high strain
models can be written as follows:
ARTICLE IN PRESS
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 486
For low strain model:
EN E
0
0:906 1
N
6230
_ _
0:058
_ _
. (16)
For moderate strain model:
EN E
0
0:903 1
N
1950
_ _
0:064
_ _
. (17)
For high strain model:
EN E
0
0:923 1
N
844
_ _
0:113
_ _
. (18)
The expression of Youngs modulus for three cases (i.e. low, moderate and high strains) are used
in energy expressions to obtain stiffness and mass matrices for the damaged beam at any given
point in time N.
3. Numerical results
The baseline undamaged blade has a length of 0.254 m, depth at root of 0.000865 m, breath at
root of 0.0173 m, twist of 45
0
, r 7800 kg=m
3
and E 2:1 10
11
N=m
2
. The beam is divided
into 25 elements of equal length, resulting in elements of length equal to 4% of the beam length.
ARTICLE IN PRESS
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
D
a
m
a
g
e

(
D
)
Low strain damage, = 0.35
Moderate strain damage, = 0.5
High strain damage, = 0.7
Cycle ratio (N/N
f
)
Fig. 2. Damage variation with number of cycles for different strain load cases.
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 487
The root, inboard and outboard locations have ve elements each. For numerical results, 20%
length of the beam is damaged for each of three locations, i.e. considering stiffness reduction in 20%
length of the beam in each of root, inboard and outboard locations. The stiffness reduction with the
number of the cycles according to Eqs. (15) and (18) for high strain loading case is given in Table 1.
3.1. Finite element simulations
A nite element analysis of the beam considering reduction in stiffness at root, inboard and outboard
locations is carried out. It is found that for 25 elements the nite element solution converges even with
considerable reduction in stiffness at the selected root, inboard and outboard locations. Fig. 3 shows the
results of a convergence study for damage at root, inboard and outboard locations as the number of
nite elements increases to 25. In these cases the damage value of D 0:99 is considered at each level.
The frequencies we normalised with the values obtained with 25 nite elements. A computer program is
written for the calculation of the modal frequencies separately considering stiffness reduction at root,
inboard and outboard locations and for different strain conditions.
The Youngs modulus values in Table 1 are used in energy expressions to get the mass and
stiffness matrix. The eigenvalue problem is then solved to get the modal frequencies. The graphs in
ARTICLE IN PRESS
No. of elements

n
first mode
second mode
third mode
1.025
1.02
1.015
1.01
1.005
1
0.995
5 10 15 20 25 30
first mode
second mode
third mode
1.025
1.015
1.005
1
0.995
1.01
No. of elements
1.02
1.03
5 10 15 20 25 30
1.005
1.01
1.015
1
0.995
No. of elements
first mode
second mode
third mode
5 10 15 20 25 30
Fig. 3. Convergence trend with number of elements for different locations of the blade.
Table 1
Reduction in stiffness with number of cycles for high strain condition
N=N
f
E=E
0
0 0.923
0.2 0.9
0.4 0.871
0.6 0.832
0.8 0.77
0.9 0.712
0.97 0.612
0.999 0.423
0.9999 0.326
N
f
844 is number of cycles resulting in failure.
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 488
Fig. 4 show the variation of frequencies with number of cycles for high strain loading conditions.
The data in the gures are normalised with respect to baseline undamaged blade. For the root
location in Fig. 4, the maximum change is in the fourth mode frequency. For the inboard location
in Fig. 4, the third mode changes most and for the outboard location in Fig. 4 the higher modes
(greater than fth) change most. Fig. 4 also shows results for the low strain case and it is evident
from the gure that damage at the root location causes maximum frequency change in the rst
mode, while damage at the inboard and outboard locations effect the third mode and sixth mode,
respectively. The moderate strain results are also shown in Fig. 4. Here the root and inboard
damage location effect the third mode and the outboard damage location effects the higher
modes. These graphs give the deterioration curves for frequency with respect to damage and show
that rotating frequencies can be used as a virtual indicator of the damage variable D.
Tables 24 shows the actual rotating frequencies in Hertz as the damage growth progresses.
These results are for the high strain case and are based on the underlying data in Fig. 4. These
results show that measurable changes in the frequencies occur due to LCF damage. Consider the
location N=N
f
0:9, which is slightly before complete failure of the material. Table 5 shows the
ARTICLE IN PRESS
Fig. 4. Frequency variation with number of cycles for stiffness reduction at different locations and strain cases of the
blade.
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 489
ARTICLE IN PRESS
Table 2
Modal frequencies (Hz) for reduction in stiffness at root location of the beam for high strain condition
N=N
f
0 0.2 0.4 0.6 0.8 0.9 0.97 0.999 0.9999
Mode 1 165.99 165.16 164.35 160.34 157.44 153.43 151.81 136.34 120.84
Mode 2 298.23 296.40 294.319 292.03 285.36 280.43 268.34 242.06 222.33
Mode 3 635.75 633.44 630.95 625.83 618.30 610.84 598.08 568.04 549.72
Mode 4 1061.00 1057.34 1052.40 1045.80 1034.23 1022.99 1002.40 957.46 929.48
Mode 5 1544.12 1540.18 1535.22 1527.69 1515.62 1503.67 1481.66 1428.76 1391.09
Mode 6 2403.38 2397.82 2390.46 2380.28 2362.85 2345.16 2310.72 2221.29 2151.31
Mode 7 2955.10 2948.42 2939.81 2927.39 2906.61 2885.31 2843.53 2733.22 2646.56
Mode 8 4146.55 4136.28 4122.74 4103.51 4070.10 4035.28 3965.20 3783.75 3654.61
Mode 9 4996.86 4985.54 4970.65 4949.48 4912.88 4874.68 4797.22 4586.11 4421.04
Mode 10 6207.42 6190.74 6169.01 6138.42 6086.50 6033.97 5933.46 5702.74 5556.83
Table 3
Modal frequencies (Hz) for reduction in stiffness at inboard location of the beam for high strain condition
N=N
f
0 0.2 0.4 0.6 0.8 0.9 0.97 0.999 0.9999
Mode 1 171.42 169.67 169.07 168.65 165.85 164.54 163.69 155.97 145.54
Mode 2 300.38 299.89 299.78 297.23 294.56 290.50 287.83 269.31 255.32
Mode 3 637.40 635.54 631.90 628.15 620.45 612.45 598.34 562.40 536.74
Mode 4 1063.11 1060.04 1055.82 1049.87 1039.34 1028.62 1008.21 955.79 919.25
Mode 5 1548.80 1546.16 1542.16 1536.96 1527.48 1517.86 1498.82 1448.92 1410.14
Mode 6 2405.46 2400.38 2393.76 2384.55 2368.69 2352.65 2321.48 2243.40 2186.17
Mode 7 2953.70 2946.47 2937.04 2923.93 2901.42 2878.73 2834.77 2726.88 2649.36
Mode 8 4146.70 4136.44 4123.05 4104.23 4072.09 4039.24 3974.47 3808.27 3684.87
Mode 9 4992.09 4979.55 4963.22 4940.33 4901.29 4861.51 4783.67 4586.87 4440.85
Mode 10 6212.24 6196.72 6176.45 6147.90 6099.08 6049.07 5950.95 5706.31 5534.49
Table 4
Modal frequencies (Hz) for reduction in stiffness at outboard location of the beam for high strain condition
N=N
f
0 0.2 0.4 0.6 0.8 0.9 0.97 0.999 0.9999
Mode 1 177.89 176.96 172.20 171.42 169.76 169.22 168.30 166.78 166.10
Mode 2 304.94 304.66 304.01 302.89 302.52 302.47 302.20 301.50 300.07
Mode 3 643.57 643.01 641.91 641.23 641.22 638.11 634.75 626.66 617.21
Mode 4 1070.88 1069.79 1068.76 1065.92 1062.68 1059.22 1050.99 1024.61 999.16
Mode 5 1549.26 1545.91 1542.33 1537.03 1527.05 1515.47 1492.51 1424.29 1367.43
Mode 6 2401.17 2394.42 2385.67 2373.04 2350.85 2327.21 2279.32 2151.52 2056.57
Mode 7 2950.25 2941.48 2930.12 2914.02 2885.85 2855.87 2796.34 2641.95 2532.72
Mode 8 4134.33 4120.42 4102.24 4076.78 4034.36 3991.72 3911.27 3728.72 3610.68
Mode 9 4978.44 4961.59 4939.70 4908.93 4856.66 4803.74 4702.62 4470.47 4323.60
Mode 10 6203.51 6185.76 6162.79 6130.82 6077.88 6024.68 5922.72 5663.77 5454.94
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 490
reduction in frequencies at these points in Hertz. The table shows that damage growth at the root
location is easier to detect since it effects the lower modes. Outboard damage effects the higher
modes to a greater extent.
3.2. Damage detection
The reduction in frequency Do can be considered to be a health residual which can be tracked
for condition monitoring of the turbine blade. Ideally, the residuals should only be effected by
faults. However, even for an undamaged system, the presence of disturbance noise, and modelling
errors cause residuals to become non-zero and interferes with the detection of faults.
The two main sources of errors in this problem are due to nite element modelling and the
presence of measurement noise. Modern nite element methods are quite accurate. For example,
Lawson and Ivey [15] compare measured and nite element simulated frequencies of compressor
rotor blades. For the rst three modes, the measured frequency was 244, 736, 1471 Hz and the
simulated frequency was 243, 740, 1486 Hz showing a discrepancy of 0.4%, 0.5%, 1.0%. Strain
gage signals were used for the measurements. So it is possible that some of the difference can be
attributed to measurement noise and the rest to modelling errors. Finite element models can be
further improved using model updating methods [20] to match with the experimental results and
the errors due to modelling can be minimised.
The second source of error is the presence of measurement noise. To some extent, measurement
noise can be reduced by signal processing methods [3,21]. Another way to make fault detection
more robust is to establish thresholds on the residuals instead of just checking for non-zero values.
A key problem in fault detection is the establishment of thresholds on the residuals which can be
used to signal when a damage has become sufciently large to be dangerous. For example,
according to Gertler [22], for a scaler residual rN, a threshold T can be established as
if rNoT then no fault,
XT then fault.
ARTICLE IN PRESS
Table 5
Reduction in frequencies (Hz) at N=N
f
0:9
Mode High strain Moderate strain Low strain
Root Inboard Outboard Root Inboard Outboard Root Inboard Outboard
1 12.56 6.88 8.67 0.85 0.15 0.75 8.06 4.26 0.24
2 17.80 9.88 2.47 8.00 6.95 1.26 13.4 12.5 2.28
3 24.89 24.95 5.46 13.65 13.25 3.04 19.8 17.58 5.74
4 38.01 34.69 11.66 19.34 19.97 6.56 20.71 15.15 16.49
5 40.45 31.13 33.79 21.04 17.20 18.30 29.86 27.10 36.92
6 58.22 52.81 73.96 29.80 30.18 41.23 35.56 38.37 47.23
7 69.79 74.99 94.38 35.60 42.56 52.78 56.09 54.87 73.76
8 111.27 107.46 142.61 55.96 60.89 81.52 61.62 66.77 89.46
9 122.18 130.59 174.7 61.69 74.05 99.43 88.95 83.08 92.40
10 173.45 162.67 178.89 88.86 92.33 102.22 106.18 103.61 124.6
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 491
A schematic representation of a residual generator is shown in Fig. 5. Such a fault residual can be
obtained from experiments, numerical simulations or in-service data. Typically, a threshold can be
developed from numerical simulation and then rened based on in-service performance.
Table 5 can be used to develop frequency thresholds on the health residuals Do o
u
o
d
.
Note that for an ideal undamaged blade Do 0. However, turbine blade can take considerable
damage before nal failure. Therefore, the Do variations in Table 5 can be used to select a
somewhat high threshold based on N
f
0:9 where a possible end of life of the blade can be
triggered. Considering the high strain case, one can select the minimum Do between the root,
inboard and outboard location as the threshold. This will lead to many false alarm but minimise
missed alarms. On the other hand, if the maximum Do is selected, it will minimise false alarms and
lead to more missed alarms. In general, threshold selection involves a trade-off between missed
alarms and false alarms [22]. Furthermore, the presence of noise in the data can be addressed by
slightly increasing the thresholds. Table 6 shows the thresholds based on the maximum criteria
after being increased by 5%, which is a conservative design and minimises false alarms. These
residuals can then be used to develop a fault detection system as shown in Fig. 6. Here each
residual is tested separately against an individual threshold. A simple rule is that if any frequency
threshold is exceeded, an alarm is indicated. Thus if any of the outputs from the test box in Fig. 6
is 1 an alarm is indicated by the diagnostic system.
ARTICLE IN PRESS
Table 6
Frequency thresholds for damage detection
Mode High strain Moderate strain Low strain
1 13.19 0.89 8.46
2 18.69 8.40 14.07
3 26.19 14.33 20.79
4 39.91 20.31 21.74
5 42.47 22.09 38.76
6 77.66 43.29 49.59
7 99.09 55.42 77.45
8 149.74 85.59 93.87
9 183.43 104.40 97.02
10 187.83 107.33 130.83
Residual
Generator
faults
disturbances
noise
residuals
modeling errors
Fig. 5. Schematic representation of a residual generator.
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 492
For ideal data, the maximum thresholds in Table 6 will give zero false alarms. The effect of
increasing noise on the success rate shown in Fig. 7. Here noisy data is generated by adding noise
to the ideal signals is Table 5 using Do
noisy
i
Do
i
1 a=100r, where a is a measure of noise
ARTICLE IN PRESS
TEST
r
1
r
2
r
N
Generator
Residual
Plant
TEST
TEST
0/1
0/1
0/1
Fig. 6. A schematic representation of a fault detection system.
1 2 3 4
50
60
70
80
90
100

S
R
High strain root
1 2 3 4
80
85
90
95
100

S
R
High strain inboard
1 2 3 4
70
80
90
100

S
R
High strain outboard
1 2 3 4
50
60
70
80
90
100

S
R
Moderate strain root
1 2 3 4
60
70
80
90
100

S
R
Moderate strain inboard
0 1 2 3 4
70
80
90
100

S
R
Moderate strain outboard
1 2 3 4
60
70
80
90
100

S
R
Low strain root
5 6 7 8
95
96
97
98
99
100
S
R
Low strain inboard

1 2 3 4
60
70
80
90
100

S
R
Low strain outboard
Fig. 7. Effect of increasing percent noise in data on damage detection success rate.
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 493
and r is a random number between 1 and 1 from a normal distribution. Here a 1 corresponds
to 1% noise in the measurement. It is clear from Fig. 7 that the detection algorithm has a very
high accuracy with about 12% noise level in the data. According to Friswell [23] frequency can
be measured to an accuracy upto 0.1%, so it can be said that the damage detection scheme based
on maximum thresholds is robust.
A schematic representation of the damage detection system is shown in Fig. 8. Rotating
frequencies are measured for the turbine blade and compared with undamaged results to obtain a
frequency residual Do. The residual is then low-pass ltered to remove noise. Appropriate
thresholds are calculated for each frequency based on the strain level, CDM models and nite
element analysis. The residual is then threshold tested to determine if a damage has occurred.
4. Closing remarks
The current work has shown that frequencies can be used to detect damage in a turbine blade
just before it becomes catastrophic. However, it should be noted that for any identication
procedure, the error between the ideal and the estimation frequency also exists and is unavoidable.
For example, the maximal reduction in percentage of natural frequency at N=N
f
0:9 occurs for
mode 1 and is about 7.57%. Other frequency reductions are about 4%. However, for modal
parameter identication, an accuracy of about 7.57% is practically not easy to achieve using
either frequency domain or time domain methods. Nevertheless, the accuracy of frequency
ARTICLE IN PRESS
Data acquisition
hardware
Residual generator
Fault detection
Filtering
Threshold tests
FEA Model/Tests

>T
Steam turbine
Strains
Threshold selection
CDM Models
FEA Models
T
Fig. 8. A schematic representation of a fault detection system.
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 494
measurement continues to improve rapidly based on newer sensors and signal processing
methods. Using smart sensors leads to much less noise than strain gages, etc. and therefore the
identication of frequencies from the vibration data can be more accurate. Recent studies have
looked at on-line estimation of frequencies. Rew et al. [24] investigated various methods for the
real-time estimation of multi-modal frequencies and validated through numerical results using
experimental tests. Oberholster and Heyns [25] developed a methodology for the on-line condition
monitoring of axial-ow fan blades by using mode shapes and frequencies which were extracted
from on-line blade vibration and strain signals. Methods to improve the accuracy of modal
estimation have been proposed by several other researchers [2629]. Furthermore, researchers
[3,30,31] have suggested ltering methods based on median and wavelet approaches which can
make the process of frequency extraction from raw data more accurate. In addition, though
frequencies are not sensitive to small damage, this can be viewed as an advantage as most real
structures are designed to take considerable damage before failure [32]. The suggested method in
this paper therefore depends on the availability of accurate sensors and signal processing methods.
There are some other limitations of this work which need to be addressed in future. The study
has focussed on damage detection and there is also a need to focus on damage isolation. However,
it should be noted that the damage detection problem is important for the turbine blade
problem as the detection corresponds to 90% of the life consumption of the blade and could result
in catastrophic failure. An experimental analysis also needs to the conducted to validate the
numerical simulations. Though the current work has used frequencies as the damage indicators,
the use of CDM models with nite element simulations can also be used to track changes in blade
response, strains, acceleration and other measurable variables for the health monitoring
applications.
The work marks a considerable advance over most works on damage detection which do not
address the issue of damage growth and focus on the identication and detection of damage at
only one time point. Many such papers using frequencies and mode shapes are published in
journals showing the feasibility of such methods [3335]. In such papers, the modal data is
used for damages of much less magnitude than the 90% degradation which we have considered in
this study.
5. Conclusion
The effect of LCF damage on the rotating frequency of a turbine blade is studied. The turbine
blade is modelled using a rotating Timoshenko beam with taper and twist and the frequencies
obtained using nite element analysis. A damage model derived using continuum damage
mechanics and identied from experimental data is used to accurately capture the non-linear
nature of LCF damage. It is found that LCF causes sufcient material degradation resulting in
stiffness loss as the damage growth progresses. The change in rotating frequency can be used as an
indicator to track damage growth. Since turbine blades are capable of sustaining considerable
accumulated damage before nal failure, the simulated deterioration curves relating frequencies to
damage are used to determine thresholds on the frequencies at the point where 90% of blade life is
consumed. By placing suitable thresholds for the residual frequency, it is possible to detect the
onset of the nal stage of damage in the structure before nal failure.
ARTICLE IN PRESS
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 495
Appendix A
A.1. Derivation of damage model
In continuum damage mechanics, a damaged coupled potential is used as a starting point,
C C
e

e
ij
; T; D; p. (A.1)
For linear elasticity and isotropic damage, coupled damage constitutive equations are [36]
s
ij
r
qC
q
e
ij
(A.2)
or

e
ij

1 n
E1 D
s
ij

n
E1 D
s
ij
d
ij
. (A.3)
The damage strain energy release rate variable Y associated with D is dened by
Y r
qC
qD

s
2
eq
R
n
2E1 D
2
, (A.4)
where R
n
is expressed for fatigue load [36] as
R
n

2
3
1 n 31 2n
s
H
s
eq
_ _
2
,
where s
H
is the hydrostatic stress dened by s
H
s
kk
=3 and s
eq
is the Von Mises equivalent
stress dened by s
eq

3:S
ij
S
ij
=2
_
and S
ij
is the stress deviator dened by S
ij
s
ij
s
H
d
ij
.
Assuming plastic deformation and micro-plastic deformation to cause damage and internal
energy dissipation, the dissipate potential f is
f f
p
s; R; D f
D
Y; _ p; _ p; T;
e
; D f
p
, (A.5)
where
f
p

s
eq
R
1 D
s
Y
.
There is little information about microplastical dissipation potential f
p
, which is not considered here.
Then the coupled damage constitutive equations and the dynamic damage evolution law can be
derived from the plastic dissipated potential f
p
and the damage dissipated potential f
D
as follows:

ij

e
ij

p
ij
,
_
p
ij
l
qf
p
qs
ij

3
2
l
1 D
S
ij
s
eq
,
_ p l
qf
p
qR

l
1 D
2
3
_
p
ij
_
p
ij
_ _
1=2
,
_
D l
qf
D
qY
,
ARTICLE IN PRESS
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 496
where l is non-negative proportion factor, which can be obtained from consistency condition,
f
p
0.
Fatigue damage is mainly caused by accumulated plastic strain. According to CDM theory,
LCF damage evolution law can be described by a suitable dissipation potential. At the basis of the
damage potential function Eq. (A.6), which is sufcient to model all the main properties within
the hypothesis of isotropy damage [36].
f
Y
2
2S
0
_ p
1 D
a
0
(A.6)
the damage character of LCF is conceded and a dissipation potential f is chosen as follows:
f
Y
2
2S
0
D_ p
1 N=N
f

1N
f
a
(A.7)
The term 1 N=N
f
, other than 1 D in Ref. [36], reects the inuence of accumulated
plastic strain, a is a parameter which describes the extent of accumulated damage, here it is the
plastic strain increment per cycle, which can be determined from monotonic tensile and cyclic
tensile stressstrain curve. Above Eq. (A.7) can be written as
_
D
qf
qY

Y
S
0
_ _
D_ p
1 N=N
f

1N
f
a
. (A.8)
From Eqs. (A.4) and (A.8) one can get
_
D
Ds
2
eq
R
n
2ES
0
1 D
2
D_ p
1 N=N
f

1N
f
a
. (A.9)
According to Lemaitres hypothesis of strain equivalence, the cyclic stressstrain relationship
coupled with damage should be written as follows:
Ds
eq
1 D
KDp
M
. (A.10)
Eqs. (A.9) and (A.10) give the general constitutive equation for LCF damage
_
D
K
2
R
n
2ES
0
D_ p
1 N=N
f

1N
f
a
. (A.11)
In the case of the proportional loading per cycle, R
n
can be considered as constant with respect
to time, and the damage during one cycle may be obtained through the integration of Eq. (A.11)
dD
dN

K
2
R
n
2ES
0
Dp
2M1
2M 11 N=N
f

1N
f
a
. (A.12)
Integrating Eq. (A.12) with the initial conditions: Dj
NN
0
D
0
; Dj
NN
f
1, gives
1 D
0

K
2
R
n
2ES
0
Dp
2M1
2M 1
1
a
(A.13)
ARTICLE IN PRESS
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 497
and
D D
0

K
2
R
n
2ES
0
Dp
2M1
2M 1
1
a
1 1
N
N
f
_ _
N
f
a
_ _
. (A.14)
Comparison of Eq. (A.13) with Eq. (A.14) gives the general LCF damage accumulation law
D 1 1 D
0
1
N
N
f
_ _
N
f
a
. (A.15)
Appendix B. Nomenclature
A area of cross-section
D damage variable
D
0
initial damage
b breadth of beam
e offset
E Youngs modulus
E
0
Youngs modulus of undamaged material
g acceleration due to gravity
G shear modulus
h depth of beam
I
xx
; I
yy
; I
xy
moment of inertia of beam cross-section about xx, yy and xy axis
K material constant
K
c
material parameter
K stiffness matrix
l length of an element
L length of total beam
M material constant
M
c
material parameter
M mass matrix
N number of cycles
N
0
number of cycles to produce initial damage D
0
N
f
number of cycles to produce fatigue failure
p plastic strain
Dp range of accumulated plastic strain per cycle
_ p rate of accumulation of plastic strain
pz axial force acting at section z
R isotropic hardening scalar variable
R
n
stress triaxiality factor
rN residual frequency
ARTICLE IN PRESS
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 498
S
0
temperature dependent material constant
S
ij
stress deviator
T temperature, threshold residual frequency
t time parameter
u nodal degrees of freedom
U strain energy
v displacement in xz plane
w displacement in yz plane
x; y co-ordinate axes
Y damage strain energy release rate
z co-ordinate axis and length parameter
z
e
distance of the rst node of the element from the root of the beam
o natural frequency
o=o
n
normalised modal frequency
a depth taper ratio h
1
=h
2
b breadth taper ratio b
1
=b
2
y angle of twist
r weight density
m shear coefcient
n Poisons ratio
C potential function
C
e
elastic potential function

e
elastic strain tensor
D strain range
p microplasticity accumulated strain
s
ij
stress level
Ds stress range
s
eq
Von Mises stress
s
H
hydrostatic stress
s
Y
initial yield stress
f dissipate potential
f
P
Von Mises plasticity function
f
D
damage dissipated potential
f
p
microplastical dissipation potential
l proportion factor
F eigenvector
Do residual frequency
O rotational speed of the beam (rad/s)
Subscripts
b bending
s shear
ARTICLE IN PRESS
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 499
References
[1] K. Worden, G. Manson, N.R.J. Fieller, Damage detection using outlier analysis, Journal of Sound and Vibration
229 (3) (2000) 647667.
[2] D.E. Adams, M. Nataraju, A nonlinear dynamics system for structural diagnosis and prognosis, International
Journal of Engineering Science 40 (2002) 19191941.
[3] N. Roy, R. Ganguli, Helicopter rotor blade frequency evolution with damage growth and signal processing,
Journal of Sound and Vibration 283 (3-5) (2005) 821851.
[4] R.M. Krupka, A.M. Baumanis, Bending-bending mode of rotating tapered twisted turbo machine blades including
rotary inertia and shear deection, Journal of Engineering for Industry American Society of Mechanical Engineers
91 (1965) 1017.
[5] J. Thomas, B.A.H. Abbas, Finite element model for dynamic analysis of Timoshenko beam, Journal of Sound and
Vibration 41 (1975) 291299.
[6] S.S. Rao, R.S. Gupta, Finite element vibration analysis of rotating Timoshenko beams, Journal of Sound and
Vibration 242 (1) (2001) 103124.
[7] I. Takahashi, Vibration and stability of non-uniform cracked Timoshenko beam subjected to follower force,
Computers and Structures 71 (1999) 585591.
[8] J. Hou, B.J. Wicks, R.A. Antoniou, An investigation of fatigue failures of turbine blades in a gas turbine engine by
mechanical analysis, Engineering Failure Analysis 9 (2000) 201211.
[9] A. Fatemi, L. Yangto, Cumulative fatigue damage and life prediction theories: a survey of the state of the art for
homogeneous materials, International Journal of Fatigue 20 (1) (1998) 934.
[10] D. Kujawski, F. Ellyin, A cumulative damage theory of fatigue crack initiation and propagation, International
Journal of Fatigue 6 (2) (1984) 8388.
[11] J.S. Rao, Turbine Blade Life Estimation, Narosa Publishing House, New Delhi, 2000.
[12] K.R.Y. Simha, Fracture Mechanics for Modern Engineering Design, Universities Press, Hyderabad, 2001.
[13] J.P. Jing, Y. Sun, S.B. Xia, G.T. Feng, A continuum damage mechanics model on low cycle fatigue life assessment
of steam turbine rotor, International Journal of Pressure Vessels and Piping 78 (2001) 5964.
[14] M. Naeem, R. Singh, D. Probert, Implication of engine deterioration for a high-pressure turbine-blades low-cycle
fatigue (lcf) life-consumption, International Journal of Fatigue 21 (1999) 831847.
[15] C.P. Lawson, P.C. Ivey, Turbomachinery blade vibration amplitude measurement through tip timing with
capacitance tip clearance probes, Sensors and Actuators A Physical 118 (1) (2005) 1424.
[16] I.B. Carrington, J.R. Wright, J.E. Cooper, G. Dimitriadis, A comparison of blade tip timing data analysis
methods, Institute of Mechanical Engineers Proceedings Journal, Part G 215 (2001) 301312.
[17] J. Lemaitre, J.L. Chaboche, Mechanics of Solid Materials, Cambridge University Press, London, 1990.
[18] Z.J. Xiaohua Yang, N. Lit, T. Wang, A continuous low cycle fatigue damage model and its application in
engineering materials, International Journal of Fatigue 19 (10) (1997) 687692.
[19] J. Lemaitre, A Course on Damage Mechanics, Springer, Berlin, 1992.
[20] V.R. Akula, R. Ganguli, Finite element model updating for helicopter rotor blade using genetic algorithm, AIAA
Journal 41 (3) (2003) 554556.
[21] W.J. Staszewski, Intelligent signal processing for damage detection in composite materials, Composites Science
and Technology 62 (2002) 941950.
[22] J.J. Gertler, Fault Detection and Diagnosis in Engineering Systems, Marcel Dekker, New York, 1998.
[23] M.I. Friswell, P. Jet, Is damage detection using vibration measurements practical?, Vol. in: EUROMECH 65
International Workshop: DAMAS 97, Structural Damage Assessment Using Advanced Signal Processing
Procedures, Shefeld, UK, 1997.
[24] K.H. Rew, S. Kim, I. Lee, Y. Park, Real time estimation of multi-modal frequencies for smart structures, Smart
Materials and Structures 11 (2002) 3647.
[25] A.J. Oberholster, P.S. Heyns, On-line fan blade damage detection using neural networks, Mechanical Systems and
Signal Processing, in press, doi:10.1016/j.ymssp.2004.09.007.
[26] N.G. Park, Y.S. Park, Damage detection using spatially incomplete frequency response functions, Mechanical
Systems and Signal Processing 17 (3) (2003) 519532.
ARTICLE IN PRESS
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 500
[27] N.M.M. Maia, J.M.M. Silva, E.A.M. Almas, R.P.C. Sampaio, Damage detection in structures: from mode shape
to frequency response function methods, Mechanical Systems and Signal Processing 17 (3) (2003) 489498.
[28] P. Verboven, E. Parloo, P. Guillaume, M.V. Overmeire, Autonomous structural health monitoring, Part i: modal
parameter estimation and tracking, Mechanical Systems and Signal Processing 16 (4) (2002) 637657.
[29] P. Verboven, E. Parloo, P. Guillaume, M.V. Overmeire, Autonomous structural health monitoring, Part ii:
Vibration-based in-operation damage assessment, Mechanical Systems and Signal Processing 16 (4) (2002)
659675.
[30] Y.Y. Kim, J.C. Hong, N.Y. Lee, Frequency response function estimation via a robust wavelet denoising method,
Journal of Sound and Vibration 244 (4) (2001) 635649.
[31] S. Kim, Y. Park, Active control of multi-tonal noise with reference generator based on on-line frequency
estimation, Journal of Sound and Vibration 227 (1999) 647666.
[32] G.C. Larsen, A.M. Hansen, O.J.D. Kristensen, Identication of damage to wind turbine blades by modal
parameter estimation, Vol. Report Riso-R-1334 (EN), Riso National Laboratory, Roskilde, Denmark, 2002.
[33] O.S. Salawu, Detection of structural damage through change in frequency, a review, Engineering Structure 17 (2)
(1997) 113121.
[34] S. Alampalli, Effects of testing, analysis, damage, and environment on modal parameters, Mechanical Systems and
Signal Processing 14 (1) (2000) 6374.
[35] E.P. Carden, P. Fanning, Vibration based condition monitoring: a review, Structural Health Monitoring 3 (4)
(2004) 355377.
[36] D. Krajeinovic, J. Lemaitre, in: Continuum Damage Mechanics Theory and Application, Springer, Berlin, 1987,
pp. 3789.
ARTICLE IN PRESS
S. Kumar et al. / Mechanical Systems and Signal Processing 21 (2007) 480501 501