You are on page 1of 13

Sadhana Vol. 41, No. 2, February 2016, pp.

239251 c Indian Academy of Sciences




Detection of a fatigue crack in a rotor system using full-spectrum based


estimation
C SHRAVANKUMAR and RAJIV TIWARI

Department of Mechanical Engineering, Indian Institute of Technology Guwahati, Guwahati 781039, India
e-mail: rtiwari@iitg.ernet.in

MS received 1 December 2014; revised 7 May 2015; accepted 8 October 2015

Abstract. The force due to crack switching has multiple harmonic components of the spin speed. These com-
ponents excite the rotor both in the same and reverse directions of the rotor spin. A full-spectrum method using
complex Fast Fourier transform equations is developed to obtain force coefcients and displacement coefcients
(forward and reverse). These coefcients are then used as an input to developed identication algorithms. Fault
parameters identied are the additive stiffness due to crack, disc eccentricity due to unbalance, and system viscous
damping. An extended algorithm estimates the crack forces. The algorithms are numerically tested.

Keywords. Switching crack; model-based identication; full-spectrum; fault parameter estimation; switching
crack forces.

1. Introduction 2. Literature review

A rotor shaft system during its operation can be subjected In this section, a literature review on identication of cracks
to multiple faults; such as unbalance, crack, misalignment, in rotors, and application of full-spectrum signal-processing
looseness, rub and the bow. Fatigue cracks in rotor systems for fault diagnostics are presented.
are common but also likely lead to catastrophic faults. Acci-
dents due to cracks in turbine rotors have been reported Crack identication. Dimentberg [3] observed that the
since 1950s [1]. Hence, the online condition monitoring of 2 component in the frequency spectrum is a strong indica-
a rotor shaft for the crack propagation becomes a neces- tor of crack, which occurs because of the shaft crossing two
sity. Condition monitoring literatures available on cracked limiting stiffness twice during one revolution. The cracked
rotors could be broadly grouped into following methods: the shaft has a critical speed at half the natural frequency of the
model-based and signal-based methods, the modal testing, intact rotor. The crack identication from changes in eigen
and non-traditional methods Sabnavis et al [2]. parameters has been studied by Pandey et al [4]. Trend anal-
Two identication algorithms are developed in the present yses of 1, 2, and 3 response amplitudes are expected
work: one for the estimation of rotor crack parameter: i.e., to increase in direct proportion to the crack size, accord-
additive stiffness in the shaft due to crack, along with ing to Gasch [5]. Dharmaraju et al [6] and Sekhar [7] have
disc eccentricity due to unbalance and shaft viscous damp- studied model-based methods for identifying crack param-
ing. The other algorithm is for estimation of crack forces eters such as the stiffness, location, and depth are used
which contribute to its switching mechanism. Both the algo- based on an inverse problem approach. Darpe et al [8] and
rithms use force coefcients and displacement coefcients, Chasalevris & Papadopoulos [9] used coupling effects intro-
directly obtained from full-spectrum as their input. Using duced by the crack for identication. Signal-based methods
full-spectrum frequency responses for identication has for crack identication have also been used by Liu & Jiang
advantages such as reduced data handling, directly obtain- [10]. Al-Shudeifat & Butcher [11] modeled crack based
ing the forward and reverse displacement coefcients (which on the time varying area moment of inertia at crack sec-
tion. For each time step, new values of area moment of
are otherwise to be estimated), and exibility to consider
inertia are calculated, based on which the centroid location
the coefcients of higher harmonics of the crack force in
and stiffness matrix are updated. Shravankumar & Tiwari
identication algorithms.
[12] identied the rotor crack using model-based parame-
ter estimation method. Crack was identied by means of
the additive stiffness it introduces at the cracked section.

For correspondence The additive stiffness is negative or reduction in stiffness.

239
240 C Shravankumar and Rajiv Tiwari

The main limitation of the work was identication in time critical speeds. Approximate solutions by means of asymp-
domain, which requires handling of large amount of system totic method and amplitude variation curves of individual
data, and intermediate steps in the identication algorithm. vibration components were obtained by the complex-DFT
The present work overcomes these limitations of the authors method. They concluded it as the most suitable method for
[12] by the application of full-spectrum to model-based the analysis of non-stationary vibrations. Bachschmid et al
identication, which reduces data handling and in addition, [23] used the SDI method in order to analyze the ellipticity
the magnitude and phase information of various forward and of ltered orbit and also the amplitude and the inclination
reverse frequency components (harmonics) of crack force angle of the rotor orbit major axis in the study of a large
and the rotor response can be directly obtained. turbo-generator unit subjected to the rotor and stator rubs.
Shravankumar & Tiwari [24] have applied full-spectrum
method for model-based crack identication. This work only
Full-spectrum. The concept of full spectrum as forward identies crack as a reduction in exibility. The present work
and backward whirl frequencies has been in use for the
extends the crack identication to switching crack forces,
development of Campbell diagram. However, the exten-
estimates of which give the actual behavior of a fatigue crack
sion of Fast Fourier transform technique (FFT) to obtain
in a rotor system.
full spectrum is relatively new and has been developed
From the literature review, it is clearly observed that
since two decades. Now, the use of full-spectrum and
applications of full spectrum have been applied chiey for
related signal processing techniques in fault identication
signal-based fault identication methods. The full-spectrum
of rotor systems is elaborately reviewed. In order to over-
method could be applied for model-based identication
come defects in traditional methods of vibration signal
methods and it is the main motivation of the present work.
processing, Qu et al [13] and Jian [14] introduced the
Wigner distribution and Short Period Fourier Transform
(SPFT). These methods were a precursor to the develop-
3. System modeling of a cracked rotor with
ment of full spectrum. Qu et al [15] also introduced the
method of holospectrum. The full-spectrum was introduced
transverse crack
as plus and minus spectrum enhancement by Southwick
A Laval rotor with a centrally located rigid disc and a mass-
[16, 17] in the Transient Data Manager2 proprietary soft-
less elastic shaft is considered, which is supported on rigid
ware of Bently Nevada. Determining the forward and reverse
bearings (gure 1). Transverse translation displacements in
whirl direction of rotors, determining synchronous and non-
two orthogonal directions x and y have been considered.
synchronous vibration components, degree of orbit ellip-
Gyroscopic effect and shaft elastic coupling are ignored
ticity due to preload and other conditions were discussed
for simplicity. A switching crack function based on the
(1994). Full spectrum cascade plots of a full annular rub
hinge model (gure 2) is considered. The system equa-
with reverse precession were given for both the startup
tions of motion (EOMs) along with the time and frequency
and shutdown conditions. Full spectrum plots were also
domain solutions have been elaborately derived in earlier
illustrated for the preload and shaft crack conditions. Gold-
works [12, 24]. The nal equations are presented briey for
man & Muszynska [18] used full spectrum to display the
completeness.
phase correlation between the horizontal and vertical spec-
The system EOMs in complex form are obtained by com-
tral components of a rotor orbit. Full spectrum was used
bining the displacements along x and y directions, using a
to determine whether the rotor orbit is forward or back-
complex variable r = x + j y. Combining the EOMs in
ward in relation to the spin direction. Various symptoms of
complex form reduces the number of equations, and applica-
rotor faults that could be detected using full spectrum were
bility of response to full-spectrum which has a pre-requisite
highlighted.
of complex response with phase correlation of x and y dis-
Tuma & Bilos [19] identied the whirl frequency com-
placement data. For orthogonal directions, the relative phase
ponents and the uid induced instability components using
is 90 . The EOMs are obtained as
the full spectrum in their study of uid instability in rotor
systems with journal bearings. Lee & Han [20] integrated
1
the techniques of Wigner distribution and SPFT with full mr + cr + kr = wy s(t)k (1 + exp (j 2t))
spectrum and provided the directional Wigner Distribution 2
(dWD) in order to study non-stationary rotor vibration. They +me2 exp (j (t + )) . (1)
proposed the Shape and Directivity Index (SDI) to quantify
the shape and direction of instantaneous whirl orbits during Here, m is the disc mass, c is the viscous damping, k is
the rotor run-up. Patel & Darpe [21] used the full spectrum the intact shaft stiffness, r is the complex displacement
to investigate the directional nature of higher harmonics response, wy is the static deection of the shaft in y direc-
for detection of rub in cracked rotors. Ishida et al [22] tion, s (t) is the switching crack function, k is the additive
studied the non-stationary vibrations of a rotor shaft with stiffness due to crack, e is the disc eccentricity due to unbal-
nonlinear spring characteristics during acceleration through ance, is the rotor spin speed, and is the phase due to
Crack parameter identication using full-spectrum 241

z
o
x

y
o x

t
x

y
y

(a) (b)

Figure 1. (a) A Laval rotor with a cracked shaft. (b) The unbalanced disc and cracked shaft in inertial and rotational coordinates.

intact shaft. The variation or reduction in intact shaft stiff-


1 ness is periodic and modeled using a switching function
s(t) [5]. Hence, this periodic variation of the shaft stiffness
0.8 multiplied by static deection of the shaft is considered on
the right side of the system EOMs as the crack force. The
0.6 present work aims to identify both these faults. The crack
force could be further generalized as
s(t)

0.4 +

wy k pi exp (j it). (3)
0.2 i=

Here, pi is the force coefcient of i th harmonic component


0 of crack force with direction same as the rotor spin. Like-
0 50 100 150 200 250 300 350 wise, pi is the force coefcient in the direction opposite
to the rotor spin. pi which along with wy k contribute
(deg)
the magnitude of the switching crack forces. pi can be
Figure 2. Variation of switching function s(t) with shaft rotation. obtained from the full-spectrum coefcients of the switching
crack function. These results are presented in a later section
on numerical illustrations. Now, Eq. (1) is a linear, second
unbalance. The rst term on the RHS of Eq. (1) was further order differential equation. Its closed-form solution can be
simplied to the following closed-form expression. obtained as the summation of the particular integrals due to
crack and unbalance forces. It is given as
1   
wx s(t)k22 1 + e2jt = wx k22 0.25 + 0.319ejt
2 +
 pi k wy
+0.106ejt + 0.25ej2t r(t) = exp (j it)
m (i)2 + c (ji) + k
+0.106ej3t 0.021ej3t i=
me2 exp (j )
0.021ej5t + 0.032ej5t + exp (j t) . (4)
j7t m2 + c (j) + k
+0.032e . (2)
In practice, the closed-form solutions of the rotor system
The rst and second terms on the right side of the EOMs may not be available. Also, in closed-form solutions there
(Eq. (1)) are the forces due to crack and unbalance faults. is limitation on the number of harmonic components based
For a fatigue crack, the crack faces are alternatively opening on the approximation of the Fourier series of the switching
and closing as the shaft spins. Therefore the shaft stiffness function. But, in case of time-integrated solution, the switch-
changes with time. The maximum stiffness is that of the ing function can be solved directly. Equation (1) is written
242 C Shravankumar and Rajiv Tiwari

in a state space form, which converts a second order differ- 4. Need for full-spectrum in crack identification
ential equation into two single order equations. The single problem
order equations can be solved using fourth order Runge
Kutta method of numerical integration. The equations in Conventional FFT gives magnitude spectrum as well as
state space form are written as phase spectrum. But they do not contain information about
the relative phase between two different vibration signals.
r = v Also, they have no information on the direction of frequency
+

m c 1  components (of crack force or response) with respect to
r = v v + wx k pi exp (j it)
k k k the shaft rotation. The rotor orbit is a plot of vibration sig-
i=
 nals along two transverse directions and contains multiple
+me2 exp (j (t + )) . (5) forward and reverse whirling components. Both magnitude
and phase information are together required to determine
the actual shape of the rotor orbit. A full-spectrum plot
Here, v is the velocity response. Both the closed-form and decomposes the rotor orbit into forward whirl (positive)
time-integrated responses have been used subsequently for and backward whirl (negative) frequency components. At
identication. The algorithms using closed-form solutions a glance, the full-spectrum plot shows whether the orbit
serve to bench mark the algorithm using time-integrated of a particular frequency component is forward or reverse
responses. These results are discussed in a later section. with respect to rotor spin direction [18]. This is because the
The numerical responses do not contain any noise. Hence, full spectrum uses the relative phase correlation of the two
in order to mimic a practical rotor response, random noise is vibration signals, which constitute the orbit.
generated and added to the numerically generated responses. In the present work, the crack force (gure 3) consists
Gaussian white noise is used for the simulation. It is dened of multiple harmonic components of spin speed. Some of
as a statistical noise that has its probability density function its harmonics excite the rotor in the same sense of rota-
equal to that of the normal distribution. It is a random signal tion of spin whereas the other harmonics excite the rotor
with a at power spectral density, uncorrelated and normally in the reverse sense of spin (refer gure 4). The harmon-
distributed with a mean zero and unit variance. The noising ics, which excite the rotor in the same direction of spin,
response signal could be obtained as are denoted positive and their magnitude is given by p+i

(forward force coefcients). The harmonics in the reverse
(R 0.5) Np
rnoise (t) = r (t) + r (t) . (6) direction as spin are considered negative and their magni-
100 tude is given by pi (reverse force coefcients). Similarly,
Here, R is a random scalar value with mean 0 and standard the response coefcient magnitudes r+i and ri are for-
ward and reverse respectively. It is not possible to directly
deviation 1. Numerical responses are added with 1, 2 and 5
obtain pi s or ri s (reverse force coefcients and reverse
percentages (Np /100) of its values as the noise. In frequency
displacement coefcients, respectively) using conventional
response analyses, the system response is computed at dis-
FFT. Hence, there is a need for using full-spectrum signal
crete excitation frequencies. The forcing is thus a function
processing in a crack identication problem.
of frequency. It gives a measure of magnitude and phase
It can be seen from Eq. (4) and Eq. (8) that there is a
of the vibration response (say, displacement) as a function
need to know the reverse force coefcients and displacement
of frequency, in comparison to the force. For a particular
coefcients in crack problem, as they constitute the time
harmonic i of the forcing function, the assumed solution is
and frequency domain solutions, which will also be used
ri () exp (jit). Since, system EOMs are linear, using the
principle of superposition, the assumed solution for each
C r a c k f o rc e ( N )

1
harmonic i can be summed up as
0.5
n

r(t) = ri () exp (j it). (7) 0
i=n 0 0.5 1 1.5 2
Time (s)
Substituting the assumed solution into the EOMs (Eq. (1)),
C ra c k fo rc e (N )

0.5
response equations in frequency domain are obtained as
0
follows.
  0.5
(i)2 m + {(jic) + k0 } ri = k22 wx pi + funb . 0 0.5 1 1.5 2
Time (s)
(8)
Similar to pi explained above, the ri is the displacement Figure 3. Variation of the crack force with time in the vertical
coefcient of the i th harmonic of crack response, obtained (top) and horizontal (bottom) directions at measurement speed of
using full-spectrum. 50 rad/s.
Crack parameter identication using full-spectrum 243

1/t. The sampling rate also satises the NyquistShannon


0.3 sampling theorem (i.e., fs 2fmax ) where fmax is the max-
imum (bandwidth) frequency. In a similar manner, response
0.25
coefcients ri s can be obtained as
0.2
N1
1 
|pi|

0.15
r (d) = {r (lt)} exp (2jdl/N). (10)
N
0.1 l=0

0.05 A full-spectrum can be effectively plotted by placing zero


0 frequency component (corresponding to DC) at the centre;
positive frequencies to the right and negative frequencies to
0.05
1000 500 0 500 1000 the left. The frequency at N /2 corresponds to the Nyquist
Frequency (rad/s) frequency. The maximum frequency or the bandwidth fre-
quency of the full-spectrum depends on the sampling fre-
Figure 4. Full-spectrum of the crack force shows force coef-
quency or the sampling time. Based on this, the number of
cients (forward as well as reverse pi s) at measurement speed of 50
harmonics components in the spectrum can be chosen.
rad/s.

to develop the identication problem. Also, in practice, the


crack switching function would be unknown, that means a 5. Crack identification algorithm
closed-form expression for switching function or crack forc-
ing cannot be assumed. Hence, it is more reliable to use the In the authors previous work [12] two identication algo-
full-spectrum of crack force and responses, and obtain their rithms have been developed: one for identication of fault
respective coefcients. parameters: namely, additive stiffness due to crack, and disc
A full-spectrum requires a complex time domain response eccentricity due to unbalance, along with the system vis-
in its input. For this purpose, system of EOMs in x and cous damping parameter. An extended algorithm was also
y directions are combined in a complex form as men- developed to identify the switching crack forces, which
tioned in Eq. (1), and the full-spectrum of the force function can be used in real case for obtaining an idea of open-
1 ing/closing behavior of a fatigue crack. The inputs for
2 s (t) {1 + exp (j 2t)}, as well as the response r (t) are
obtained. Resultant spectrums will contain magnitude and these identication algorithms, i.e., force coefcients were
phase information of both the forward and reverse force and obtained from the closed-form expression of crack force
displacement coefcients. Obtaining the full-spectrum plot and displacement coefcients were only estimated using dis-
using complex-FFT is explained next. The full-spectrum placement responses in time domain. This has a limitation
force and displacement coefcients can be obtained using on two factors: one is on choosing the higher harmonic
standard complex-FFT equations [25]. Full-spectrum force components of crack force. The closed-form expression of
coefcients p (d) (which are complex Fourier coefcients) switching function provides the pi s limited to the rst
are obtained as few harmonic components as in Eq. (2). In this closed-
form expression, crack force is limited to the harmonics
N1
1  i = 1, 3, 5, +2, +7. On the other hand, full-spectrum
p(d) = {s(lt)(1+ exp(2jlt))}exp(2jkl/N).
N can provide pi values for any required number of harmonic
l=0
(9) components of switching function. This result is illustrated
in a subsequent section and will be particularly useful prac-
The full-spectrum converts a time domain complex signal tical estimation of crack forces. Also, instead of a switching
of length N (sample length) into a frequency domain com- function any other time-varying function can also be eas-
plex signal of same length N . Its coefcients are complex. ily implemented using full-spectrum. The other factor is
In Eq. (9), the frequency index d (length of the full spectrum the large amount of time domain data required to estimate
signal) varies from 0 to N 1. The index d from 0 to N /2 cor- the displacement coefcients ri . This data usage is greatly
responds to p(+d) and N /2 to N 1 corresponds to p(d); reduced while directly obtaining ri from full-spectrum.
i.e., p(N /2 +d) = p(d). Let pi correspond to p(d) where Now, the generalized form of the identication algo-
2 d/n = i, and i = 1, 2,..., l is the number of discrete rithms developed in the previous work are written, which
data in the time history of s(t)(1 +ej 2t ) with t as the are suitable for application with full-spectrum. Equation (8)
time interval (sampling time). Here, i corresponds to vari- which is the system EOM in frequency domain is writ-
ous positive and negative harmonic components of the crack ten in a linear regression form and further used in the
force. The acquisition time (T ) of the time domain signal is identication problem. The regression equation is entirely
N t. Samples are acquired at a sampling rate (fs ) which is based on frequency domain data (i.e., force coefcients and
244 C Shravankumar and Rajiv Tiwari

displacement coefcients). The generalized identication say 1 to q , the regression equation (Eq. (12)) is modied
algorithm is written as follows, for n harmonic components. as

A1 (1 ) b1 (1 )
A1 x1 = b1 , (11)

A1 (2 ) b1 (2 )

x1(91) = . (14)

.. ..
where .  . 



A1 q b 1 q

(nq9) (nq1)
j (n) rn 0 wy pn
. . . Thus, the present section algorithm would give some con-

. . . dence in usage of the full-spectrum in model-based identi-


. . .
cation. However, it has the limitation that the crack force
j () r1
0 wy p1
coefcients pi are unknown in the real case. So, the further
0 0 wy p0 idea is to illustrate estimating these coefcients using the
A1((2n+1)3) =
j () r1
m2 ej wy p1 cracked rotor response. Once by estimating pi the periodic
j (2) r2
0 wy p2 crack forcing can also be known (Eq. (3)). The following

. . .
section alleviates this difculty by developing estimation

. . .
algorithm for the crack force coefcients as well with the
. . .
help of full-spectrum.
j (n) rn 0 wy pn
 2 2 

n m k0 rn


.


6. Identification of switching crack forces




.





.

The generalized crack force is expressed as wy 
2
 
m k r



0 1
+


k r
0 0 
k22 pi exp (jit) in Eq. (3). Here, s(t) or pi denes
b1((2n+1)1) =  2 i=

 2m k0 r1 the crack force completely and it is unknown. This implies
4 m k0 r2


that by estimating k22 pi , the prole of the periodic crack



.



force can be obtained at any dened running speed and time



.




interval. For this purpose, the identication algorithm, i.e.
.






n2 2 m k0 rn

Eq. (12) is modied into the following estimation equations
containing terms k22 pi in the unknown vector, as shown in
c
Eq. (17). Noting that the regression equation contains terms
x1(31) = e . (12)

k22
to be estimated for the crack force, i.e.pi and the additive
crack stiffness k22 ; both terms are treated as an unknown,
Bold capital letters are used for matrices and bold small let- and the equation is rearranged into a new regression equation
ters are used for vectors. Refer Nomenclature for details. where pi terms of A1 are included in the unknown vector
Unique solution of x1 can be obtained if the matrix A1 is x2 together with k22 . It is given in the following form,
square and satises the condition for matrix inversion. In
A2 x2 = b1 . (15)
the general case, A1 may be any rectangular matrix and
cannot be inverted with sufcient conditions. Hence, an  
Here, A2((2n+1)(2n+3)) = D1 D2 , where
approximate solution of x1 (or estimates) is found using the
least-squares solution, which is a common form of the linear
0 0 0 0 0 0 0
regression. .
. . . . . .

 1 . . . . . . .
x1 = AT1 A1 AT1 b1 .

(13) . . . . . . .
0 0 0 0 0 0 0


wy 0 0 0 0 0 0

D2 = ;

For well-conditioning of the matrices, the identication 0 wy 0 0 0 0 0
algorithm is modied and written for multiple measurement 0
0 wy 0 0 0 0

speed ranges [26]. When the number of regression equations .
. . . . . .

is increased the estimation is also better averaged and the . . . . . . .

effect of noise is reduced drastically. Thus, for accommo- . . . . . . .
dating estimation with a set of multiple measurement speeds 0 0 0 0 0 0 wy
Crack parameter identication using full-spectrum 245

c

For accommodating estimation with a set of multiple



e



measurement speeds, the regression equation (Eq. (16)) is

k pn





modied as

.






. A2 (1 ) b1 (1 )







. A2 (2 ) b1 (2 )




x2(31) =

.


k p1
.. ..
x2((2n+3)1) = ;
.  . 

k p0




A2 q b1 q

k p1 (2n+1)q3 (2n+1)q1





k p2





(17)

.



From the full-spectrum of crack force thus estimated, pi s



.



can be obtained and used further in Eq. (12) estimating



.



crack stiffness, as described in the previous section. Now,



k pn

through numerical simulations identication algorithms will
be elaborated in order to illustrate effectiveness.

n2 2 m k0 rn




.






. 7. Numerical Illustrations






. 






2 m k0 r1 In this section, illustrations of identication algorithms have





 k0 r0  been presented through numerical simulation of a chosen




b1((2n+1)1) = 2 m k0 r1 rotor model. A Laval rotor, consisting of a simply supported
elastic shaft with a disc at the mid-span, is considered for
 2 
4 m k0 r2




numerical illustration as shown in gure 1. The rotor model




.

data are summarized in table 1. A crack near disc causing




.
variation in stiffness due to switching function s(t) as shown




.

in gure 2 is considered.




 2 2

n m k0 rn It should be noted that s(t) can have any time variation

and the present identication algorithm is not restricted by
this variation. Responses are generated using closed-form
j (n) rn 0 wy 0 0 0 0 expression (Eq. (4)) as well as by time-integration (Eq. (5)).

. . . . . . . The orbit plots of x displacement responses versus y dis-

. . . . . . . placement responses are shown in gure 5. The proposed

. . . . . . . identication algorithms (Eq. (12) and Eq. (16)) are tested
j () r1
0 0 0 0 0 wy
with the help of numerical responses obtained. The esti-
0 0 0 0 0 0 0 mates of damping, eccentricity and crack force parameters
D1= 2 ej
. (16)
j () r1
m 0 0 0 0 0 are obtained and their accuracy is analyzed.
j (2) r2 0 0 0 0 0 0


. . . . . . . Obtaining force coefcients. Figure 3 and gure 4 show

. . . . . . . the variation of crack force and its corresponding full-
. . . . . . . spectrum plot containing pi s, respectively. Figure 4 shows
j (n) rn 0 0 0 0 0 0 the full-spectrum plot with forward and reverse exciting

Table 1. Rotor system data for the numerical simulation.

Parameters Value
Disc mass, m 2 kg
Intact shaft stiffness, k0 3.2135 105 N/m
Additive (negative) crack stiffness, k 1.61104 N/m
Viscous damping in rotor system, c 16.03 N-s/m
Phase of unbalance, 10( /180) rad
Shaft deection due to disc weight, wy wy = mg/k0 = 6.105 105 m
Disc eccentricity, e 0.1106 m
246 C Shravankumar and Rajiv Tiwari

6 6 7
x 10 x 10 x 10
10
2 5 10
0
2 0 8
y(t) d isp lacemen t (m)

4 5
5 0 5 5 0 5 10
6
(a) 6 (b) 6

|ri|
x 10 x 10
5 5
x 10 x 10 4
5 1
2
0 0
0
5 1
5 0 5 1 0 1 1000 500 0 500 1000
(c) 5 (d) 5 Frequency (rad/s)
x 10 x 10
x(t) displacement (m) Figure 6. The full-spectrum of the cracked rotor response show-
ing displacement coefcients (forward and reverse ri s) at measure-
Figure 5. Orbit plots at (a) 180 rad/s, (b) 190 rad/s, (c) 200 rad/s, ment speed of 50 rad/s.
and (d) 210 rad/s during passage through critical speed due to twice
per revolution excitation component.
Inverse problem. In order to test the algorithms, some
of the parameters are treated as unknown (grey-box model
crack force coefcients. These coefcients pi are tabulated identication) and are then estimated from the identica-
for multiple harmonics of crack force in table 2. The pi s tion algorithm using full-spectrum coefcients. This is the
obtained using full-spectrum are the same as that obtained inverse problem approach used in the model-based identi-
from closed-form coefcients. These are further used as an cation. The estimates can nally be compared with assumed
input for the identication algorithm (Eq. (12)) to estimate values (in the direct problem) for testing the accuracy of
crack and other unknown parameters. the proposed algorithms. In this work both closed-form
and time-integrated responses are tested for parameter esti-
Obtaining displacement coefcients. The displacement mation. Numerical responses are also added with different
coefcients ri are obtained from full-spectrum coefcients levels of 1, 2 and 5 percentages of random noise.
of the displacement responses of cracked rotor (closed- Now different measurement spin-speed ranges are cho-
form or time integrated), using Eq. (10). The full-spectrum sen (for Eq. (14) and Eq. (16)) for which the displacement
plot of r (d) against frequency shows various displacement responses are considered for identication. Figure 7 and
coefcients as in gure 6. gure 8 show the resonance plots of various frequency com-
ponents of crack force. To avoid measurement error near
7.1 Simultaneous estimation of crack and rotor critical speeds often measurement outside the half-power
parameters using full-spectrum frequency band is advisable. The half-power points are two
frequencies on either side of the resonant amplitude. They
In this section numerical testing of the developed identi- are often referred as side bands and their corresponding
cation algorithm (Eq. (12)) is performed for various cases amplitudes will be X = 0.707Xres , where Xres is the
and the estimates obtained are compared with the assumed resonant amplitude and X is the amplitude at half-power
parameters. points. For example, it is noted from gure 7 that the

Table 2. Values of pi corresponding to full-spectrum coefcients of crack


force.

I pi i pi i pi i pi
1 0.3183 0 0.2500 15 0.0016 15 0.0012
2 0.2500 1 0.1061 17 0.0013 17 0.0009
3 0.1061 3 0.0212 19 0.0009 19 0.0008
5 0.0212 5 0.0091 21 0.0008 21 0.0007
7 0.0091 7 0.0051 23 0.0007 23 0.0006
9 0.0051 9 0.0032 25 0.0006 25 0.0005
11 0.0032 11 0.0022 27 0.0005 27 0.0004
13 0.0022 13 0.0016 29 0.0004 29 0.0003
Crack parameter identication using full-spectrum 247

4
x 10 half-power point frequencies corresponding to resonance of
4 3 frequency component are 132.3 and 133.6 rad/s. From
2 these considerations, the range of frequencies is selected.
The general identication algorithm (Eq. (12)) contains har-
0
0 5 100 200 300 400 500 600 monics varying from n to n. Now, a particular number
Disp lacemen t (m)

x 10 of harmonics can be chosen for each measurement range


4
by making use of the resonance plots. It is observed that
2
the magnitude of higher harmonics is very low of the
0
0 5 100 200 300 400 500 600 order of 108 m. Also, the sampling time determines the
x 10 maximum frequency and harmonic in a full-spectrum plot.
1 For example, a sampling time of 0.001 s gives a sam-
0.5 pling rate of 1000 Hz and a maximum full-spectrum fre-
0 quency of 500 Hz. Based on these, in a given speed range,
0 100 200 300 400 500 600 the harmonic components until the magnitude of 108 m
Frequency (rad/s) (refer gure 7 and gure 8) are chosen and lower mag-
nitudes are neglected. The different measurement cases,
Figure 7. Plot for resonance of (a) once per revolution, (b) twice
per revolution, and (c) thrice per revolution harmonics of crack their corresponding range of frequencies, and the number
force. of harmonics considered in each case are summarized as in
table 3.
6 The estimates obtained for different cases of measurement
x 10
4 speeds are now discussed. For case A, estimates are obtained
2 for a variety of noise levels of 1, 2, and 5 percentages and
are summarized in table 4. The estimates are observed to
0
0 6 100 200 300 400 500 600 conform well to the assumed parameters. The maximum per-
Disp lacemen t (m)

x 10 centage error in estimates of damping is 1.4% for 5% noise


2
level; for additive crack stiffness, it is 2.6% at 5% noise level
1 and there is no error in estimates of the disc eccentricity at
0 any noise level. For cases B, C, and D also the estimates are
0 6 100 200 300 400 500 600
x 10 obtained similarly. The estimates of 5% noise are summa-
1 rized in table 5. The maximum percentage error in damping
0.5 estimates is 0.9% for case B, 5.3% for case C, and 2.2%
0
for case D. For eccentricity it is around 10% for cases B, C,
0 100 200 300 400 500 600 and D for 5% noise level. The maximum percentage error in
Frequency (rad/s) additive crack stiffness is 3.2% for case B, 3.1% for case C,
and 4.3% for case D.
Figure 8. Plot shows resonance of (a) ve per revolution, (b)
seven per revolution, and (c) nine per revolution harmonics of crack Now, the subsequent section presents results for estima-
force. tion of crack forces due to switching.

Table 3. Different speed range and harmonics considered for estimation.

Cases Range of frequencies (rad/s) Harmonics (i) considered for algorithm


Case A 1132.3 7, 5, 3, 1, 0,1, 2, 3, 5, 7
Case B 133.6198.4 5, 3, 1, 0, 1, 3, 5
Case C 202.4396.9 3, 1, 0, 1, 3
Case D 405600 3, 1, 0, 1, 3

Table 4. Estimates of parameters in the measurement speed range of 0130 rad/s.

Parameters c(N-s/m) e ( 106 m) k ( 104 N/m)


Assumed values 16.03 0.1 1.61
Noise level in % 0 1 2 5 0 1 2 5 0 1 2 5
Estimates closed-form 16.03 16.07 16.12 16.25 0.099 0.099 0.099 0.097 1.61 1.60 1.59 1.57
Estimates time-integrated 16.17 16.18 16.18 16.18 0.100 0.100 0.099 0.098 1.61 1.60 1.59 1.57
248 C Shravankumar and Rajiv Tiwari

Table 5. Estimates of parameters for measurement cases B, C, and D with 2% and 5% noise in response.

135190 rad/s at 210390 rad/s at 410600 rad/s at


Measurement speed range 1 rad/s interval 1 rad/s interval 1 rad/s interval
Parameters Assumed Noise level 2% 5% 2% 5% 2% 5%
value in %
c(N-s/m) 16.03 CF 16.00 15.99 15.70 15.18 15.95 15.78
TI 16.18 16.17 16.41 16.42 16.37 16.38
e ( 106 m) 0.10 CF 0.093 0.091 0.097 0.095 0.101 0.112
TI 0.099 0.097 0.099 0.099 0.099 0.097
k ( 104 N/m) 1.61 CF 1.59 1.56 1.59 1.56 1.58 1.54
TI 1.59 1.57 1.59 1.57 1.59 1.57
(CF Estimates from closed-form responses; TI Estimates from time-integrated responses).

7.2 Simultaneous estimation of crack forces and rotor 1.5


(a) (b)
parameters using full-spectrum 1 0.5

Crack ex citatio n fo rce (N)


For case A, estimates are obtained for different noise levels 0.5 0
of 1, 2 and 5 percentage. The estimates from time-integrated 0 0.5
responses are presented in table 6. In order to check the accu-
racy of the estimates the assumed and estimated switching 0.5
0 0.02 0.04 0 0.02 0.04
crack forces are plotted for a convenient spin speed and com-
pared as in gure 9. The comparison is shown for the 2% 1.5
(c) (d)
and 5% noise levels for responses in both the horizontal and 1 0.5
vertical directions. For cases B, C, and D, the estimates are 0.5 0
similarly obtained for the 1, 2 and 5% noise levels. The esti-
mates for 5% noise level in response are presented in table 7 0 0.5
and table 8. Figure 9 shows the comparison of assumed and 0.5
estimated crack excitation forces at 170 rad/s for 2% and 5% 0 0.02 0.04 0 0.02 0.04
Time (s)
noise levels. Similar comparisons are made for cases C and
D at spin speeds of 300 rad/s and 500 rad/s for 2% and 5% Figure 9. Comparison of the assumed and estimated crack force
noise levels as shown in gure 10 and gure 11. From these at 170 rad/s. (a) With 2% noise in x response; (b) with 2% noise in
gures, it is seen that the estimated crack forces conform y response; (c) with 5% noise in x response; (d) with 5% noise in
well to the assumed ones. y response. Assumed and Estimated.

Table 6. Simultaneous parameter estimation in the range 0130 rad/s with noise.

Noise level in %
0% 1% 2% 5%
Parameters Assumed values CF TI CF TI CF TI CF TI
c(N-s/m) 16.03 16.03 16.18 16.07 16.18 16.12 16.18 16.24 16.18
e(106 m) 0.100 0.100 0.099 0.100 0.099 0.099 0.099 0.097 0.097
k p1 (N/m) 5114.4 5114.6 5115.0 5083.8 5089.6 5052.9 5064.1 4960.3 4987.7
k p2 (N/m) 4017.0 4017.0 4017.2 3993.0 3997.2 3969.0 3977.2 3896.9 3917.3
k p3 (N/m) 1704.8 1704.9 1704.7 1695.4 1696.2 1685.7 1687.8 1656.5 1662.3
k p5 (N/m) 341.0 341.0 342.1 337.9 340.4 334.6 338.7 324.9 333.6
k p7 (N/m) 146.1 146.2 146.2 144.0 145.3 141.7 144.5 135.0 142.0
k p0 (N/m) 4017.0 4017.0 4016.9 3995.4 3996.9 3973.7 3976.9 3908.8 3917.0
k p7 (N/m) 81.1 81.2 81.4 80.5 81.09 81.7 80.8 95.00 80.0
k p5 (N/m) 146.1 146.2 146.7 138.8 145.9 139.1 145.2 104.5 142.9
k p3 (N/m) 341.0 341.0 340.6 337.93 338.9 334.8 337.2 325.5 332.2
k p1 (N/m) 1704.8 1704.9 1704.8 1696.5 1696.4 1688.1 1688.0 1662.9 1662.7
(CF Estimates from closed-form responses; TI Estimates from time-integrated responses).
Crack parameter identication using full-spectrum 249

Table 7. Simultaneous parameter estimation in the range 135190 rad/s with noise.

Noise level in %
0% 1% 2% 5%
Parameters Assumed values CF TI CF TI CF TI CF TI
c(N-s/m) 16.03 16.03 16.18 16.02 16.40 16.00 16.18 15.94 16.17
e( 106 m) 0.100 0.100 0.100 0.097 0.100 0.093 0.099 0.091 0.097
k p1 (N/m) 5114.4 5114.6 5115.3 5094.1 5090.7 5073.6 5064.1 5012.1 4987.2
k p2 (N/m) 4017.0 4017.0 4017.9 3994.8 3999.8 3972.5 3977.9 3906.3 3917.8
k p3 (N/m) 1704.8 1704.9 1706.0 1696.4 1693.9 1687.8 1688.9 1661.9 1663.1
k p5 (N/m) 341.0 341.0 342.9 343.0 342.9 344.9 338.9 350.7 332.9
k p0 (N/m) 146.1 4017.0 4017.0 3993.7 3996.3 3970.4 3977.0 3900.5 3917.3
k p5 (N/m) 4017.0 146.2 146.6 124.4 146.4 110.0 146.2 117.4 145.5
k p3 (N/m) 81.1 341.0 342.3 344.5 340.4 348.0 338.5 358.5 332.9
k p1 (N/m) 146.1 1704.9 1704.6 1697.4 1696.0 1689.9 1687.4 1667.9 1661.6
(CF Estimates from closed-form responses; TI Estimates from time-integrated responses).

Table 8. Simultaneous estimation of damping, eccentricity, and crack force parameters for measurement cases C and D with 2% and 5%
noise levels in response.

210390 rad/s 410600 rad/s


Noise level in %
Measurement speed range 2% 5% 2% 5%
Parameters Assumed values CF TI CF TI CF TI CF TI
c(N-s/m) 16.03 15.70 16.41 15.18 16.42 15.95 16.37 15.78 16.38
e( 106 m) 0.100 0.097 0.099 0.095 0.099 0.103 0.099 0.112 0.097
k p1 (N/m) 5114.4 5059.3 5064.9 4977.5 4988.0 5080.7 5089.2 5081.7 5012.3
k p2 (N/m) 4017.0 3971.2 3979.6 3902.9 3918.9 3945.6 3978.9 3839.0 3918.9
k p3 (N/m) 1704.8 1699.7 1684.4 1691.7 1656.9 1664.7 1697.9 1604.3 1681.3
k p0 (N/m) 341.0 3966.7 3976.3 3895.4 3916.3 3976.8 3975.3 3916.6 3914.7
k p3 (N/m) 146.1 351.8 343.2 366.9 344.3 271.9 321.2 311.2 302.4
k p1 (N/m) 4017.0 1685.1 1685.4 1655.5 1660.0 1683.3 1694.2 1648.0 1668.6
(CF Estimates from closed-form responses; TI Estimates from time-integrated responses).

1.5 1.5
(a) (b) (a) (b)
1 0.5 1 0.5
Crack ex citatio n fo rce (N)
Crack ex citatio n fo rce (N)

0.5 0 0.5 0
0 0.5 0 0.5
0.5 0.5
0 0.01 0.02 0.03 0 0.01 0.02 0.03 0 0.005 0.01 0.015 0 0.005 0.01 0.015

1.5 1.5
(c) (d) (c) (d)
1 0.5 1 0.5
0.5 0 0.5 0
0 0.5 0 0.5
0.5 0.5
0 0.01 0.02 0.03 0 0.01 0.02 0.03 0 0.005 0.01 0.015 0 0.005 0.01 0.015
Time (s) Time (s)

Figure 10. Comparison of the assumed and estimated crack Figure 11. Comparison of the assumed and estimated crack
forces at 300 rad/s. (a) With 2% noise in x response; (b) with 2% forces at 500 rad/s. (a) With 2% noise in x response; (b) with 2%
noise in y response; (c) with 5% noise in x response; (d) with 5% noise in y response; (c) with 5% noise in x response; (d) with 5%
noise in y response. Assumed and Estimated. noise in y response. Assumed and Estimated.
250 C Shravankumar and Rajiv Tiwari

It is seen that the estimated crack forces conform very further to obtain force and displacement coefcients. This
well to the assumed forces. It is also observed that the forces information is used for crack identication as described fol-
in the direction of crack are unsymmetrical about the mean lows. An identication algorithm has been developed for the
position, while the forces in the orthogonal direction are simultaneous estimation of additive crack stiffness, viscous
symmetrical about their mean position (gures 911). This damping, and eccentricity. An extended algorithm has been
behavior is due to crack switching. developed to estimate switching crack forces, along with
damping, and eccentricity. The algorithms are completely
based on frequency domain (full-spectrum) inputs. The full
8. General guidelines for crack identification in spectrum has been employed to obtain force coefcients as
rotor systems well as displacement coefcients, which are then input to the
identication algorithm. The algorithms have been written
The general guidelines for model-based identication of in the form of linear regression equations. Unknown param-
crack and other unknown system parameters are presented eters have been estimated using Least Squares approach. As
in this section. The steps for identication are as follows: the crack forces have excitation harmonic components both
The cracked rotor responses, namely, displacement ver- in the forward and reverse directions of spin, full-spectrum
sus time responses, and full-spectrum frequency responses is the suitable technique for frequency domain identication.
are obtained. These responses can be simulated for a chosen The algorithm is tested for different measurement spin-
data of system and fault parameters. In such a case, it consti- speed ranges and for different noise levels. The comparison
tutes the direct problem. The responses can also be obtained of estimated and assumed parameters shows robustness of
directly by means of an experimental set up. algorithms. The estimates of the crack force give an idea of
The second step is the Full-Spectrum analysis. Here, the crack opening/closing function as well. The algorithms
displacement responses in two orthogonal directions are can accommodate any crack opening/closing function and
combined in a complex form and fed as input to the full- also can handle required number of coefcients of crack
spectrum. The displacement coefcients ri are obtained force and response harmonics. The use of full-spectrum
from the full-spectrum. data as input has the advantage that the algorithms can be
The third step is the Identication of switching crack used for rotor response with reverse whirl conditions. The
forces. The input to the identication algorithm consists switching crack model illustrated for the present algorithms
of displacement coefcients of the previous step discussed makes it applicable for small or moderate cracks, which also
above, and known system parameters, such as the mass, means that they can be effectively used for early crack detec-
intact shaft stiffness, and static deection of the rotor sys- tion. Also, as the identication is purely done in frequency
tem. The estimated parameters are the terms of switching domain it helps in data reductions and gives the exibility
force (k pi ), which constitute the crack force mag- to consider higher harmonics of crack force, which can rep-
nitude, together with unbalance eccentricity, and viscous resent the crack breathing function more closely. In future,
damping. The plot of the switching crack forces can give the the algorithms can be extended for a multiple crack prob-
details of opening/closing prole of the fatigue crack in the lem together with location and quantication of crack by
actual case. nite element modeling of the rotor system. The algorithms
The fourth step is Full-Spectrum analysis of crack forces. will be experimentally veried by testing for the developed
From the full-spectrum, the force coefcients pi are laboratory and practical rotors with transverse fatigue crack.
obtained.
The fth step is Identication of crack stiffness. Inputs to
the identication algorithms include the displacement and Nomenclature
force coefcients (from the step two and the step four),
along with known system parameters. The unknown fault A regression matrix
parameter, i.e. additive stiffness due to crack, is estimated b regressand or known vector
for the condition monitoring of crack. In addition, the disc c viscous damping
eccentricity and the viscous damping are also estimated. d real number
D1 , D2 sub-matrices
e unbalance eccentricity
9. Conclusions exp exponential of
fs sampling rate
This paper discusses the need and application of full- fmax bandwidth frequency
spectrum for a cracked rotor problem. Initially, a Jeffcott FFT fast Fourier Transform
rotor with fatigue crack is considered. Closed-form solu- i = n, .., 0, 1, .. + n integer showing number of the
tion and numerical responses are obtained for unbalance and harmonic component
crack excitations. Effect of measurement noise is considered j= 1 imaginary number
to simulate experimental data. Full-spectrum method is used k intact shaft stiffness
Crack parameter identication using full-spectrum 251

k additive stiffness parameter due to crack [9] Chasalevris A C and Papadopoulos C A 2013 Experimen-
along direction tal detection of an early developed crack in rotor-bearing
m disc mass system using an AMB. Third International Conference of
Engineering against Failure, June 2628, Kos, Greece
N sample length of time or frequency
[10] Liu C and Jiang D 2013 Experimental study on lateral and
response torsional vibration of cracked rotor with torsional excitation.
Np percentage of Noise ASME Turbo Expo 2013, June 37, 2013, Texas, USA
pi crack force coefcients from full-spectrum [11] Al-Shudeifat M A and Butcher E A 2011 New breathing
q number of measurement speeds functions for the transverse breathing crack of the cracked
r (t) = x + j y complex displacement rotor system: Approach for critical and subcritical harmonic
ri displacement coefcients from analysis. J. Sound Vibr. 330: 526544
full-spectrum [12] Shravankumar C and Tiwari R 2013a Identication of stiff-
rnoise time response with noise ness and periodic excitation forces of a transverse switching
s (t) periodic, time-varying crack switching crack in a Laval rotor. Fatigue Fracture Eng. Mater. Struct.
function 36(3): 254269
[13] Qu L, Xie A X and Li X 1993 Study and performance evalu-
t time
ation of some nonlinear diagnostic methods for large rotating
t sampling time machinery. Mech. Math. Theory 28(5): 699713
wx static deection [14] Jian M 1996 Some advanced techniques in fault feature
x translational displacement in x direction extraction for large rotating machinery. PhD dissertation.
y translational displacement in y direction Xian Jiao Tong University
rotor spin speed [15] Qu L, Liu X and Chen Y 1989 Discovering the holospec-
order of frequency trum. J. Noise Vibr. Worldwide 5862
phase angle between crack front direction [16] Southwick D 1993 Using full spectrum plots. Orbit 14(4): 12
and unbalance 16
. rst order time derivative [17] Southwick D 1994 Using full spectrum plots, Part 2. Orbit
.. second order time derivative 15(2): 1016
[18] Goldman P and Muszynska A 1999 Application of full spec-
state space co-ordinate for velocity
trum to rotating machinery diagnostics. Orbit 1721
[19] Tuma J and Bilos J 2007 Fluid intability of rotor systems with
journal bearings. Eng. Mech. 14: 6980
References [20] Lee C W and Han Y S 1998 Use of directional Wigner dis-
tributions for identication of the instantaneous whirling orbit
[1] Ishida Y 2008 Cracked Rotors: Industrial machine case in rotating machinery. 7 th ISROMAC. 2226
histories and nonlinear effects shown by simple Jeffcott [21] Patel T H and Darpe A K 2011 Application of Full spec-
rotor. Mech. Syst. Signal Process. 22: 805817 trum analysis for rotor fault diagnosis. IUTAM Symposium on
[2] Sabnavis G, Kirk R G, Kasarda M and Quinn D 2004 Cracked Emerging Trends in Rotor Dynamics 1011: 535545
shaft detection and diagnostics: A literature review. The Shock [22] Ishida Y, Yasuda K and Murakami S 1997 Nonstationary
and Vibration Digest 36(4): 287296 oscillations of a rotating shaft with nonlinear spring charac-
[3] Dimentberg F M 1961 Flexural vibrations of rotating shafts. teristics during acceleration through a major critical speed (A
London, England: Butterworths discussion by the asymptotic method and the complex-FFT
[4] Pandey A K, Biswas M and Samman M M 1991 Damage method). J. Vib. Acoust. 119(1): 3136
detection from changes in curvature mode shapes. J. Sound [23] Bachschmid N, Pennacchi P and Vania A 2004 Diagnostic sig-
Vibr. 145(2): 321332 nicance of orbit shape analysis and its application to improve
[5] Gasch R 1993 A survey of the dynamic behaviour of a simple machine fault detection. J. Brazilian Soc. Mech. Sci. Eng. 2:
rotating shaft with a transverse crack. J. Sound Vibr. 160: 313 200208
332 [24] Shravankumar C and Tiwari R 2013b Model-based crack
[6] Dharmaraju N, Tiwari R and Talukdar S 2004 Identication identication using full-spectrum. ASME 2013 Gas Turbine
of an open crack model in a beam based on force-response India Conference. December 56, 2013, Bangalore, Kar-
measurements. Comput. Struct. 82: 167179 nataka, India (GTINDIA 2013-3756)
[7] Sekhar A S 2004 Crack identication in a rotor system: a [25] Smith S W 1997 The scientist and engineers guide to digital
model-based approach. J. Sound Vibr. 270: 887902 signal processing. California Technical Publishing, San Diego
[8] Darpe A K, Gupta K and Chawla A 2004 Coupled bend- [26] Tiwari R 2005 Conditioning of regression matrices for simul-
ing, longitudinal and torsional vibrations of a cracked rotor. J. taneous estimation of the residual unbalance and bearing dy-
Sound Vibr. 269: 3360 namic parameters. Mech. Syst. Signal Process. 19: 10821095