Professional Documents
Culture Documents
Adrian I. Cuc
Bachelor of Science
Politehnica University of Timisoara, 1996
____________________________________________________
2002
________________________________ ________________________________
Department of Mechanical Engineering Department of Mechanical Engineering
Director of Thesis Second Reader
_______________________
Dean of the Graduate School
ACKNOWLEDGMENTS
I would like to thank my academic advisor, Dr. Victor Giurgiutiu for his understanding,
I would also like to thank to all the professors who have shared with me their
knowledge throughout my education. Special thanks go to Dr. Leonid Gelman for his
In addition, I would like to sincerely thank my colleagues and friends Jing Jing
(Jack) Bao, Max Fitzgerald, Paulette Goodman, Florin Jichi, Greg Nall, Mattias Ohlsson,
Radu Pomirleanu, Andrew Rekers, and Andrei Zagrai for their support and for having
My thanks go also to the South Carolina Army National Guard, Mr. Lem Grant
and Tim Cook for knowledge they shared with me. The financial support of the South
Last but not least, I would like to express my deeply gratitude to my parents,
ii
ABSTRACT
The state of the art in vibration-based damage detection and health and usage
monitoring is reviewed. The damage detection concepts and the signal analysis
techniques are examined, explained, and compared. Latest advances in signal processing
methodologies that are of relevance to vibration based damage detection (e.g., Wavelet
processing methods play an important role in early identification of incipient damage that
can later develop in a potential threat to the system functionality. A brief mathematical
Health and Usage Monitoring Systems (HUMS) programs installed on the civil and
military helicopters are reviewed. The current procedures for Rotor Track and Balance
(RT&B) of the helicopters are explained and vibration data from the sensors mounted in
the cockpit showing the reduction in vertical vibrations are presented. HUMS
cracked specimen was investigated. The non-linear equations of motions were set up and
damage was simulated as a change in the stiffness of the specimen. The relative crack
size considered in this study was r = 0 (pristine case) r = 0.1 (small crack), r = 0.4
(medium crack), and r = 0.6 (large crack). Using MATLAB-SIMULINK software, the
iii
time-domain nonlinear vibration signal was generated. Two signal analysis methods were
used to process the time-domain signal: Short Time Fourier Transform (STFT) and
investigated, the Fisher criterion was calculated for pairs of classes representing the
frequency of the main harmonic for the pristine case and the frequency of the main
harmonic for various damage cases. The results showed that both WVD and STFT
methods predict the presence of the damage. The STFT method gives a better indication
of the higher harmonics of the main frequency due to the crack growth. On the other
hand, WVD method gives stronger sensitivity to the crack presence based on the changes
The presented work shows that structural damage can be detected from vibration
Monitoring.
iv
TABLE OF CONTENTS
ACKNOWLEDGMENTS ............................................................................................................II
ABSTRACT................................................................................................................................. III
1 INTRODUCTION................................................................................................................. 1
DETECTION ................................................................................................................................. 5
3.2.2.5 FM0....................................................................................................... 14
v
3.2.2.6 NA4....................................................................................................... 14
3.2.2.7 FM4....................................................................................................... 15
BALANCE (RT&B)..................................................................................................................... 41
vi
4.1.4 Track main rotor in flight ......................................................................... 47
6.4.1 Case 1: same aircraft, same state, and same sensor ................................ 84
6.4.2 Case 2: different aircraft, same state, and same sensor ........................... 85
vii
7 DAMAGE DETECTION OF A SIMULATED CRACK USING SHORT TIME
viii
LIST OF TABLES
Table 5.1 List of the fault types created in the test gearbox (Essaway et al., 1998)...... 64
Table 5.2 Neural Network Classification results using APSD features at 100% load
Table 5.3 Neural network classification results using wavelet features (Essaway et al.,
1998) .............................................................................................................. 65
Table 6.1 Shaft frequencies and mesh frequencies for the main components of the
Table 7.2 Fisher criterion calculated for STFT and WVD methods and the feature
comparison. Each of the the damaged cases were compared with the pristine
ix
LIST OF FIGURES
Figure 3.1 Processing flow chart for feature extraction methods for vibration analysis
Figure 3.3 STFT(t,w) as a subset of the entire two-dimensional functions (Qian and
Cheng, 1996).................................................................................................. 19
Figure 3.4 WVD representing the energy distribution of linear chirp with Gaussian
Figure 3.5 WVD cross-term at midway between the two auto-terms. It oscilates in both
Figure 3.6 Comparison of elementary functions for STFT (a) and WT (b). For STFT all
Figure 3.7 Comparisons of STFT and WT (a) For STFT time and frequency resolutions
are uniform for the entire time-frequency domain; (b) For WT there is good
time resolution for high frequencies, and good frequency resolution for low
Figure 3.8 Elementary illustration of Fisher Criterion showing how the projected
(b) a more favorable angle a = 35o yields better separation, but the variance
x
S1 and S2 are still large; (c) optimal projection angle a = 72o yields clearly
Figure 3.9 A schematic illustration of the two classes C1 and C2 their mean values and
the projection of the means onto the two axes (Bishop, 1995)...................... 32
Figure 3.12 Vibration analysis software for analysis, training and preventive
Figure 4.3 Main rotor pitch link adjustment (AVA Technical Manual) ......................... 46
Figure 4.4 Tab adjustment tool used to bed the tab of the helicopter blades for RT&B
Figure 4.5 Trim tab pockets for a helicopter blade (AVA Technical Manual)............... 48
Figure 4.7 VMU Data storage and directory configuration: on the VMEP server and on
Figure 4.10 Historical trends for vertical vibration for one aircraft and one particular test
xi
Figure 4.11 a) Polar plot for the vibration data between two RT&B flights b) Plot of the
amplitude values for the same flights for all test states. ................................ 54
Figure 4.12 Plot of vibration amplitudes showing an increased vibration level for the
Figure 4.13 RT&B flights showing how the vibration amplitudes reduces after making
recommended adjustments............................................................................. 56
Figure 5.1 A comparison of the DWT algorithm and the DWPT algorithm. Ai is
Figure 5.2 Mean square wavelet maps for day one and day nine. On day nine a
Figure 5.3 Average normalized power index (Samuel et al., 1998) ............................... 59
harmonic wavelet map: 10% crack with 5% noise (Kim and Ewins, 1999) . 60
Figure 5.5 Accelerometer data for an undamaged gear. (Polyshchuck et al., 2000) ...... 61
Figure 5.6 Accelerometer data for damaged gear (Polyshchuck et al., 2000) ................ 61
Figure 5.7 Hybrid system, fault detection performance. 60-point discrete Fourier
transform (DFT) was performed for fault detection, 128-point and 256-point
Figure 5.8 All analog system performance. For fault detection in the case of all analog
neural network. 60-point DFT was performed (Monsena et al., 1995) ......... 63
xii
Figure 6.2 Sensors location on Bell 412 Helicopter ....................................................... 71
Figure 6.3 IMD HUMS hardware configuration; (a) Main CH-53E IMD HUMS display;
(b) Location of the IMD HUMS on the Army UH-60A (Hess et al. 2001) .. 73
helicopter ....................................................................................................... 78
Figure 6.9 Vibration data flow diagram showing downloading of data from the VMU to
the ground station and further to the USC teradata computer ....................... 80
Figure 6.11 Frequency spectrum for the tail gearbox sensor for the same aircraft, same
Figure 6.12 Frequency spectrum for the intermediate gearbox sensor for the same aircraft,
Figure 6.13 Frequency spectrum for the #1 nose gearbox sensor for the same sensor, same
Figure 6.14 Frequency spectrum for the intermediate gearbox sensor for the same sensor,
Figure 7.1 Stretching and compression of the cracked object. In stretching crack is
xiii
Figure 7.3 Frequency change as the crack progresses .................................................... 91
Figure 7.7 Time domain data: (a) pristine case (r=0) and (b) damaged case (r=0.6)...... 96
Figure 7.8 Graphical User Interface (GUI) for the MATLAB program (Gelman et al.
2001) .............................................................................................................. 97
Figure 7.9 WVD and STFT comparison for the pristine case, r = 0 ............................... 98
Figure 7.10 WVD and STFT comparison for small damage case, r = 0.1........................ 99
Figure 7.11 WVD and STFT comparison, damaged case r = 0.4. .................................. 100
Figure 7.12 WVD and STFT comparison for the damaged case, r = 0.6........................ 101
Figure 7.13 Damage detection effectiveness of the two methods, STFT and WVD, based
xiv
1 INTRODUCTION
system that may affect its current or future performance. From this definition of damage
once can see that damage is not meaningful without a comparison between two different
states of the system, one of which is assumed to represent the initial (pristine) state, and
the other the damaged state. The definition of damage can also be limited to changes to
the material and/or geometric properties of the system, including changes to the boundary
conditions and system connectivity, which adversely affect the current or future
The interest in the ability to monitor a structure and detect damage at the earliest
possible stage is persistent throughout the civil, mechanical and aerospace engineering
magnetic fields methods, eddy-current methods, and thermal methods (Doherty, 1987).
All these experimental methods have the disadvantage that the location of the damage
must be known a priori, and the portion of the structure that is being inspected must be
accessible.
1
The basic premise in vibration-based damage detection is that damage will
which in turn, modifies the measured dynamic response of the system. One of the most
local phenomenon and may not significantly influence the lower-frequency response of
Chapter 2 presnts, the state of the art in machinery health monitoring and damage
detection. Studies conducted by various researchers to assess the damage and the health
of mechanical systems are discussed. Various signal processing techniques were used
such as short time Fourier transform (Kim and Ewins, 1999), wavelet transform
these techniques is given. Statistical analysis method to compare and classify the
presented.
Chapter 4 describes the helicopter vibration reduction through rotor track and
balance. The procedure for equipment installation, vibration data collection, and data
analysis is presented.
The vibration-based methods for damage detection and health monitoring for
2
Wigner-Ville distribution (WVD), neural network-based and neuro-fuzzy validation and
The meaning and description of HUMS systems is given, along with a review of the
current vibration HUMS systems installed on military and civilian helicopters. The main
part of the chapter emphasizes the Vibration Monitoring Enhancement Program (VMEP)
installed on the South Carolina Army National Guard helicopters. The main contractor of
is responsible for the cost benefit analysis of the VMEP program. The VMEP-HUMS
configuration, data collection, data analysis, and feature extraction are presented.
SIMULINK. The signal-analysis and damage detection capabilities of two methods, the
short time Fourier transform (STFT) and the Wigner-Ville distribution (WVD), were
investigated. The cracked component was approximated with a single degree of freedom
nonlinear equation. Solving the equations of motion in the time domain generated the
structural response. The simulation was carried out for four relative crack sizes: r = 0
(pristine case), r = 0.1 (small crack), r = 0.4 (medium crack), and r = 0.6 (large crack).
The simulated time-domain responses for the pristine and damage cases were
analyzed in the frequency domain using STFT and WVD methods. The features
investigated were the change in the amplitude of the main frequency of the component
and the apparition of additional effects, i.e. higher order harmonics of the main
3
frequency. The comparison of the outcome of the two methods is based on a statistical
Chapter 8 presents the conclusion of the study and the future work that is needs to
be done to fully understand the mechanism of damage detection and health monitoring of
mechanical systems, using vibration signal and joint time-frequency signal processing
methods.
4
2 STATE OF THE ART IN MACHINERY HEALTH MONITORING
activity for most engineering systems. Helicopters are continuously subjected to periodic
loads and vibrational environments that initiate and propagate fatigue damage in many
be monitored and replaced at fixed intervals. This constitutes an expensive procedure that
Health and Usage Monitoring Systems (HUMS) have been developed in recent
years to detect incipient damage in helicopter components, predict remaining life, and
create the premises for moving over from scheduled based maintenance to condition
evaluate the damage or undamaged state of a critical component using only the vibration
data signals recorded during flight and ground operation. With such capability, the need
for frequent disassembly and bench checking of certain critical components can be
reduced and ultimately eliminated, with important ancillary savings in the O&S costs.
However, to achieve such capability, advance vibration signal processing algorithms are
necessary that can distinguish the damage related features from the background and
5
kurtosis analysis, time-frequency domain analysis, Wavelet transform, neural network
systems have shown promising results (Randall, 1980; McFadden, 1985; Kemerait, 1987;
Monsena, 1994). However, the challenge remains to translate this knowledge into fault
prediction.
The earliest work toward detecting defects in helicopter gearboxes used a high-
resonance technique and was a off-line monitoring tool focused on finding the exact
location of the defect. Based on this technique, the frequency spectrum of the vibration
signal is search to find a change in the spectral line at one particular frequency associated
with any particular faults. This method has proved good results for simple bearing
configurations but is not satisfactory for complex configuration or for a large damage
(McFadden, 1984)
involves computing the time average of the vibration signal. This method removes all
signals not synchronous with the gear shaft frequency and reveals the individual tooth
Fraser and King in 1990 using kurtosis observed that when a gearbox component
is severely damaged multiple impulses will appear in the frequency domain, resulting in a
cumulative response that tend to reduce the kurtosis value. Another approach is to apply
into an easier to interpret form, thus providing a practical system for routine prognostic
6
Forester in 1990 demonstrates that time-frequency techniques, such as Wigner-
Ville distribution can describe how the spectral content of the signal is changing with
time and provided a framework for developing robust detection and classification
7
3 MATHEMATICAL TECHNIQUES FOR VIBRATION-BASED
into the time-domain analysis, the frequency-domain analysis, and joint time-frequency
domain analysis. Some of the parameters (signal features) used in these methods are
Customarily, for non-stationary processes the spectral analysis methods used for
analyzing the frequency content of the signal are namely: the short time Fourier
sinusoidal functions.
+
X (w ) e
jtw
x(t ) = dw (3.1)
-
8
+
1
x(t ) e
- jw t
X (w ) = dt (3.2)
2p -
efficient algorithm for obtaining the Fourier transform of discretized time signals.
Various implementations of the FFT algorithm have become the workhorse of signal
analysis packages, while FFT chips are incorporated in portable signal analyzers.
The classical Fourier analysis applied to these signals do not provide an explicitly
dependency between the time change and the frequency content change.
difficult to establish a relation between x(t) and X(w) based upon the conventional Fourier
analysis.
Defects or damage that occurs in different parts of machinery will increase the
machinery vibration levels. These vibration levels are captured and converted into
electric signals for data measurements by accelerometers. The information concerning the
health of the monitored machine is contained in this vibration signature. Hence, the new
whether the component that is being monitored is behaving normally or signs of failure
In practice, although desired, the direct comparison of the signature is very difficult.
Instead, a technique that involves first the extraction of features from the vibration signal
and then comparison of the features could be used. The advantages in using features lies
9
in the fact that that ideally these features are more stable and well behaved then the raw
signature data and also provide a reduced data set for pattern recognition.
Before any feature can be calculated on the raw vibration data, the data must be
based on the data acquisition unit, mean value removal, time synchronous averaging, and
filtering.
Signal DC Offset
Raw signal
Conditioning Removal Rms
Kurtosis
Crest Factor
Time Synchronous Enveloping
Tachometer
Averaging
FM0
NA4 NB4
Remove 1st
order sidebands
FM4
M6A
M8A
Difference Band-pass Mesh
Signal Signal
Figure 3.1 Processing flow chart for feature extraction methods for vibration analysis (Lebold et al.
2000)
Lebold et al. (2000) grouped the processing features for gearbox diagnostics and
prognostics into five processing groups. The five processing groups are:
10
1. Raw signal (RAW)
The Raw Signal Preprocessing (RAW): denotes features that are calculated from the
raw or conditioned data. The signal conditioning is achieved by multiplying the raw data
Time Synchronous Average (TSA): signal processing technique used to extract the
repetitive signals from the additive noise. The raw data is divided into segments of equal
length related to the synchronous signal (tachometer) and then averaged together. In this
way after sufficient averages the random noise is canceled and only the desired signal is
left. The parameters involved with this technique that can affect the results are: the
interpolation factor, the number of revolutions concatenated, and the number of averages.
recommended to use one revolution for data alignment, which will provide the most
accurate alignment with the shaft rate. Increasing the number of averages will enhance
the synchronous component of the signal but it will be more computational intensive.
Residual Signal (RES): this preprocessing technique calculates the residual signal
consisting of the time synchronous averaged signal including the mesh and shaft
components along with their harmonics removed. In practice this can be achieved by
11
using a high pass filter with the cut-off frequency lie somewhere between DC and the
Difference Signal (DIF): the difference signal is calculated by removing the regular
meshing components from the time synchronous averaged signal. The regular meshing
components consist of the shaft frequency and its harmonics, the meshing frequency and
its harmonics. Because the difference signal processing section comes after the residual
signal section in which the mesh and shaft components along with their components have
been removed, only the sidebands of the meshing frequency and its harmonics need to be
removed.
Band-Pass Mesh Signal (BPM): in this section the time synchronous averaged signal
is band pass filtered around the primary gear mesh frequency. Hilbert transform is then
applied to obtain the real and imaginary part of the signal, which represents the band-pass
signal and the Hilbert transform of the signal respectively. The envelope of the signal will
represent the amplitude modulation present in the signal due to the sidebands.
3.2.2.1 RMS
The root mean square (RMS) value of a vibration signal is a measure of the power
content in the vibration signature. This feature gives good results in tracking the overall
noise level in the signal but will not give any information about the component that is
going to fail.
12
N
x 2
k
RMS = k =1
(3.3)
N
3.2.2.2 Kurtosis
Is also known as the fourth moment of a distribution and measures the relative
provides a measure of the size of the tail of a distribution and can be used as an indicator
[ x( n) - m ]
4
k= n =1
(3.4)
N (s 2 )2
where x(n) is the time series, m is the mean value of the data, s2 is the variance of the data
To observe the changes in the early stages of gear and bearing damage crest factor is
a useful feature. It is defined as the ratio of the peak level of the input signal to the RMS
value of the signal. If any peaks appears in the time domain signal will result in an
increase of the crest factor. Usual values of crest factor are between 2 and 6. Any values
above 6 are associated with possible machinery problems. The best results using this
feature are obtained in the case of impulsive vibration sources such as tooth breakage on
Peak Level
Crest Factor = (3.5)
RMS
13
where Peak Level is the peak level of the raw time signal, and RMS is the root mean
3.2.2.4 Enveloping
systems to periodic impacts such as gear or bearing faults. The envelope or high
frequency technique focuses on the structure resonance to determine the health of a gear
3.2.2.5 FM0
FM0 is defined as the peak-to-peak level of the time synchronous averaged signal
divided by the sum of the amplitude at the gear-mesh frequency and its corresponding
harmonics. It is a simple method used to detect major changes in the meshing pattern.
App
FM 0 = n
(3.6)
A( f )
i =1
i
where App is the peak-to-peak amplitude and A(fi) is the amplitude of the gear-mesh
3.2.2.6 NA4
NA4 is calculating by dividing the fourth statistical moment of the residual signal
by the time-averaged variance of the residual signal, raised to the second power. It is used
to detect the initiation of damage and to react to the growth of the damage. The equation
is
14
N
N ( ri - r )
4
NA4 = i =1
2
(3.7)
1 m N 2
( rij - rj )
m j =1 i =1
where r is the residual signal, r is the mean value of the residual signal, N is the total
3.2.2.7 FM4
FM4 was developed to detect changes in the vibration pattern resulting from
damage on a limited number of gear teeth (Zakrajsek et al. 1993). The equation to
calculate FM4 is
N
1
N
(d i - d )4
FM 4 = i =1
2
(3.8)
N
2
(di - d )
i =1
where d is the difference signal, d is the mean value of difference signal, and N is the
3.2.2.8 NB4
NB4 is similar to FM4 except that instead of using the residual signal, NB4 uses
calculated by dividing the fourth statistical moment of the envelope signal to the current
run time averaged variance of the envelope signal raised to the second power, as shown
in Equation (3.9)
15
N
N ( Ei - E )
4
NB 4 = i =1
2
(3.9)
1 m N 2
( Eij - E j )
m j =1 i =1
where E is the envelope of the band-passed signal, E is the mean value of the envelope
signal, N is the total number of data points, and m is the current time record.
The majority of signal in the real life are non-stationary. Machinery operating in
non-stationary mode generates a signature, which at each instant of time has a distinct
frequency (Shin et al. 1993). It is desired in such cases to use a joint time-frequency
analysis technique that will give an inside of how frequency changes with time.
The frequency contents of the majority of signals in our daily life change over
time. Some of these signals are speech signals, vibration signals, and biomedical signals.
Because the sinusoidal basis functions vary over the entire time domain and are not
concentrated in time, it is hard to indicate how a signals frequency contents changes with
time by using the conventional Fourier transform. It comes naturally to try to characterize
a signal simultaneously in the time and frequency domains by comparing them with
elementary functions that are concentrated in both time and frequency domains, such as
the frequency modulated Gaussian functions (Qian and Cheng, 1996). Thus, in order to
overcome the deficiency of the classical Fourier analysis, the simplest way would be to
compare the signal with elementary functions that are localized in both time domain and
16
3.4 SHORT TIME FOURIER TRANSFORM (STFT)
The short time Fourier transform is the most widely used method for studying
non-stationary signals. The basic idea of the short time Fourier transform is to break up
the initial signal into small time segments and apply the Fourier transform to each time
segment to ascertain the frequencies that existed in that segment. The totality of such
The limitation of the short time Fourier transform lies in the impossibility to
achieve finer and finer time localization by using a smaller window functions. As the
time becomes smaller, the information content of the resulting spectrum decreases. In the
end, by decreasing the time interval, the answer that we get for that particular spectrum
becomes meaningless. The representation of the Short Time Fourier Transform (STFT)
x(t )g
*
STFT (t , w ) = (t - t )e- jw t dt (3.10)
-
Generally the function g(t) has a short time duration thereby it is also called the
window function. Equation (3.10) is called short time Fourier transform or windowed
Fourier transform (Qian and Cheng, 1996). To understand the procedure for computing
Because the window function g(t) has short time duration, the Fourier transform
of the product x(t )g * (t - t ) will reflect the signals local frequency content. In this case
17
by moving the window function g(t) and repeating the steps mentioned above, it is
possible to obtain information about how the signals frequency content evolves over
x(t)
g(t)
time
FT
FT
FT
frequency
STFT
time
be applied with respect to Equation (3.10). This will yield (Qian and Cheng, 1996):
+
1 1
STFT (t ,w )e x(t )g (t - t )e = x(t )g (t - t )d ( m - t )dt =
j mw j ( m -t )w
dw =
2p -
2p (3.11)
= x( m )g ( m - t )
where
1 for m = t
d (m - t ) = (3.12)
0 otherwise
is the Dirac function
Let m = t, we have
+
1
x(t ) =
2pg (0) -
STFT (t , w )e jtw dw (3.13)
18
Equation (3.13) implies that if STFT (t,w) is given for all t and w, the signal x(t) can be
completely recover.
Another way to look at the short time Fourier transform (STFT) is to view STFT
Figure 3.3 STFT(t,w) as a subset of the entire two-dimensional functions (Qian and Cheng, 1996)
For any time domain function x(t) and window function g(t), such mapping
always exists. It is true in this direction but the inverse may not be true. Given a window
function g(t) and an arbitrary two-dimensional function B(t,w), there may be no signal
x(t) whose STFT is equal to B(t,w). In this case, B(t,w) is not a valid short time Fourier
For digital signal processing applications, the STFT is discrete in both time and
L -1
STFT [k , n] = x[i ]g [i - k ]WL- ni (3.14)
i =0
where
2p
j
STFT [k , n] STFT (t , w ) t = k Dt ,w = 2p n /( LDt ) , WL = e L
(3.15)
19
where Dt denotes the sampling interval in the time domain, and g[k] = g(kDt) is the L-
a) Total energy
The total energy of a signal is obtained by integrating over all time and frequency.
Considering an initial signal x(t) if the properties of the signal at time t are of interest then
the original signal is multiplied by a window function, g (t), centered at the time t. The
xt (t ) = x(t )g (t - t ) (3.16)
where t is the time of interest and t is the running time.
The Fourier transform of the signal will reflect the distribution of frequency around that
time
1 1
e e
- jwt - jwt
X t (w ) = xt (t )dt = x(t )g (t - t )dt (3.17)
2p 2p
The energy density spectrum at time t is
2
2 1
e
- jwt
PSP (t , w ) = X t (w ) = x(t )g (t - t )dt (3.18)
2p
The total energy in the signal becomes
2 2
E = PSP (t , w )dtdw = x(t ) dt g (t ) dt (3.19)
Equation (3.19) shows that if the energy of the window is one, then the energy of the
20
The time marginal condition is obtained by integrating over the frequency domain. After
2 2 2
X (w )
t dw = x(t ) g (t - t ) dt (3.20)
Similarly, the frequency marginal condition is obtained by integrating over the time
domain:
2 2 2
x (t )
t dw = X (w1 ) g (w1 - w ) dw1 (3.21)
method. The WVD is one of the most general time-frequency analysis techniques, as it
provides excellent resolution for accurate examination in both time and frequencies
domains. Some of the problems encounters in using WVD are related to the aliasing
arising in the computation of WVD and to its nonlinear behavior. To avoid the aliasing
problem, the original real signal is transformed into a complex analytical signal.
The Wigner distribution was developed by Eugene Wigner in 1932 to study the
signal analysis by the French scientist, Ville 15 year later. It is commonly known in the
The behavior of the frequency of a signal can be of two types: linear representation such
as the Fourier transform (Short time Fourier transform) and quadratic representation such
as Wigner-Ville distribution.
21
Some success has been achieved by researchers who used the WVD of vibration
signals to detect gear-tooth faults. The joint time-frequency distribution of the faulty gear
shows the vibration energy dispersion at time when damaged gear teeth are in mesh. Such
energy change can also be detected by classical Fourier transform but only at an
effectiveness, excellent accuracy, and invariant to the window effects, the Wigner-Ville
distribution also has drawbacks such as the so called cross-term or interference terms.
The power spectrum of a signal, which is the square of the Fourier transform
characterizes the signals distribution in the frequency domain. According to the Wiener-
Khinchin theorem, the power spectrum can be considered as the Fourier transform of the
+
2
R(t )e
- jw t
PS (t , w ) = X (w ) = dt (3.22)
-
where:
+
x(t ) x (t - t )dt
*
R(t ) = (3.23)
-
The Fourier transform of the time-dependent auto-correlation function R(t,t) with respect
R(t ,t )e
- jwt
PS R (t , w ) = dt (3.24)
-
22
where the time-dependent auto-correlation function R(t,t) is given by:
+
1
A ( v, t ) A ( v, t ) e
jvt
R(t ,t ) = x g dv (3.25)
2p -
where Ax(v,t) represents the ambiguity function of signal x(t) and Ag(v,t) represents the
ambiguity function in the window function g(t) (Qian and Chen, 1996).
t t
R (t , t ) = x t + x* t - (3.26)
2 2
where x(t) is the analytic signal and x*(t) is the complex conjugate of the analytic signal
x(t).
signal x(t).
Based on the Equation (3.27) the WVD of a signal can be calculated at any time
point t simply by shifting the time signal origin to that point (Polyshchuk et al., 1999).
+
t t - jw t
x 2 x
*
WVDxx (0, w ) = - e dt (3.28)
- 2
1
-a 2
a 4 t + jbt2
x(t ) = e 4
(3.29)
p
1
- a t 2 + (w - 2 b t ) 2
a
WVDx (t , w ) = 2e
(3.30)
23
and the plot of the WVD is presented in Figure 3.4
Figure 3.4 WVD representing the energy distribution of linear chirp with Gaussian envelope (Qian
and Chen, 1996)
If the time is shifted by t0 then the WVD is shifted accordingly. For a signal x(t)
Lets consider the WVD of the signal x(t) as being WVDx(t,w). For the frequency-
conditions, that is
24
+
1 2
+
2
WVD (t ,w )dt =
-
x X (w ) (3.34)
As a consequence, using Parsevals relation, it can be shown that the energy contained in
the Wigner-Ville distribution, WVD(t,w), is equal to the energy possessed by the original
signal s(t).
Using Equation (3.27) to derive the WVD for the sum of the two signals yields
t t
WVD(t , w ) = x* t - x t - e - jtw dt =
2 2 (3.36)
= WVD1 (t , w ) + WVD2 (t , w ) + WVD12 (t , w ) + WVD21 (t , w )
where
t t
WVD12 (t , w ) = x1* t - x2 t + e - jtw dt (3.37)
2 2
is called the cross-term Wigner-Ville distribution. Even though the cross-term is
complex, because the WVD is always real, the sum WVD12(t, w) + WVD21(t, w) is real
(auto-term) but has the additional term 2Re{WVD12(t,w)}, term called the interference
term or the cross-term. Because the cross-term usually oscillates and its magnitude is
25
twice as large as that of the auto-terms, it often reduces the useful patterns of the time-
dependent spectrum.
Some time, the cross-terms are referred to as undesirable effects of the Wigner-
Ville distribution. The cross-term in fact reflects the correlation of the corresponding pair
of auto-terms. Its location and rate of oscillation are determined by time and frequency
centers of the auto-terms. If the position of the auto-terms is known, then the position of
Figure 3.5 WVD cross-term at midway between the two auto-terms. It oscilates in both time and
frequency directions (Qian and Cheng, 1996)
Wigner-Ville distribution. It is independent of time and thus can describe the local
behavior of the signal. Unlike the short time Fourier transform where complex sine and
cosine functions are used to map the time signal into the frequency domain, wavelet
26
transform uses a set of elementary functions that can be independently dilated and shifted
of the signal in the frequency domain is achieved. If the center frequency or the mean
frequency of the elementary function y(t) at time zero is w0, and if the time is dilated by
b and the frequency shifted by a, then the center of the dilated and shifted elementary
t -b
function y will be at time b and frequency w0/a. When the dilated and shifted
a
t -b
elementary function y is used, the resulting presentation is called the continuous-
a
+
1 t -b
x(t )y
*
W ( a, b) = dt , a0 (3.39)
a - a
where y(t) is also called the mother function. The parameter a represents the scale index
and parameters b represents the time shifting. The quantity W(a,b) reflects the signals
Figure 3.6 shows a comparison of the elementary functions for short time Fourier
transform and wavelet transform. Once an elementary function for the short time Fourier
transform is chosen, both the time and frequency resolution are fixed. For the wavelet
transform, the time resolution and the frequency resolution of the elementary function are
27
(a) (b)
Figure 3.6 Comparison of elementary functions for STFT (a) and WT (b). For STFT all elementary
functions have the same envelope, whereas the WT elementary functions have different
envelopes (Qian and Cheng, 1996).
In order to understand better the difference between the short time Fourier
transform and the wavelet transform, an example is considered. Suppose that we have a
signal that contains two pulses in the time domain, d(t-t1) and d(t-t2), and two pulses in
wavelet transform (WT) for the signal described by Equation (3.40). A major difference
between the two methods can be observed. Time and frequency resolutions are constant
in both time and frequency domains in the case of STFT, while they vary in the case of
WT.
28
w w
w2 w2
w1 w1
t1 t2 t t1 t2 t
(a) (b)
Figure 3.7 Comparisons of STFT and WT (a) For STFT time and frequency resolutions are uniform
for the entire time-frequency domain; (b) For WT there is good time resolution for high
frequencies, and good frequency resolution for low frequencies (Qiang and Cheng,
1996)
At high frequencies, the time resolution of the WT is better than the frequency resolution;
at low frequencies the frequency resolution is better than the time resolution. For STFT,
the time and frequency resolutions are the same for the entire time-frequency domain.
From the wavelet transform, the original signal can be recovered by using
1 1 t -b
x(t ) =
Cy a 2
W (a, b)y
a
(3.42)
2
1 Y (w )
Cy =
2p w
dw (3.43)
29
3.7 FISHER CRITERION
One widely used method for damage identification is the Fisher criterion. Fisher
criterion is a statistical analysis method used for discriminating between two statistical
used in analyzing the statistical distribution of the signal features and deciding if damage
is present or not. This is achieved by comparing the statistical distribution of the signal
features measured on the current machinery with the statistical distribution of the same
signal features measured on a as new (pristine) machinery. If the Fisher criterion has a
low value, the current machinery is as new, i.e., no damage is present. However, if the
Fisher criterion exceeds a certain limit, the machinery can be classified as damaged or
faulty.
multidimensional feature vector, one distribution corresponding to the pristine case, the
other to the damaged case (Figure 3.8). All points of the two distributions are projected
onto a line in the direction l, and this direction is varied, by modifying the slope of the
line (angle a) until the two samples are maximally separated, i.e. the distance between
the means of the two classes is maximized, while the variance in each class is minimized.
Figure 3.10 presents the case in 2-D, i.e., when the feature space consists of just
two classes C1 and C2. The features are projected onto the line y and form two samples of
projected data with the mean m1 and m2. The two means are separated by the distance d.
Figure 3.10 shows three cases, in which the parameter a varies and the dependency
between the angle a, the distance d that separates the two means, and the variance in the
30
2
Class 1
8
Class 2 1.8 Class 1 Class 2
line class1
1.6
line class 2
center 1.4
6 mean class 1
mean class 2 1.2
1
y 0.8
4
0.6
0.4
0.2
y
2 0
-4 -2 0 2 4 6 8
o
0 a = 5 F = 1.46, m1 = 0.76 , m2 = 2.49
-2 0 2 4 6 8
S1 = 0.98, S2 = 1.03
-2
(a)
2
Class 1
1.8 Class 1 Class 2
8 Class 2
line class1
1.6
line class 2
y center 1.4
mean class 1
mean class 2 1.2
6
1
0.8
0.6
4
0.4
0.2
0
y
2
-4 -2 0 2 4 6 8
o
0
a = 35 F = 8.92, m1 = 0.04 , m2 = 3.29
-2 0 2 4 6 8
S1 = 0.80, S2 = 0.81
-2
(b)
8
y Class 1
2
Class 1 Class 2
Class 2 1.8
line class1 1.6
line class 2
center 1.4
6 mean class 1
1.2
mean class 2
1
0.8
4
0.6
0.4
0.2
2 y
0
-4 -2 0 2 4 6 8
o
0
a = 77 F = 18.10, m1 = 0.97 , m2 = 2.28
-2 0 2 4 6 8
S1 = 0.23, S2 = 0.23
(c)
-2
31
Figure 3.8 Elementary illustration of Fisher Criterion showing how the projected distributions vary
o
with the projection angle a: (a) unfavorable projection angle a = 5 yields overlapping of
o
Class 1 and Class 2 projected distribution; (b) a more favorable angle a = 35 yields
better separation, but the variance S1 and S2 are still large; (c) optimal projection angle a
o
= 72 yields clearly separated sharp distributions with small variances
We can see from the above figure that as the angle a changes (increases) the means
of the two classes separates, the distribution of the data in the two sets becomes more
sharply which means that the standard deviation of the two means and thus the variance
is smaller.
The mathematical expression for the Fisher criterion for two classes C1 with N1
points and C2 with N2 is derived. A linear projection of the data onto a one-dimensional
space is considered so that an input vector x is projected onto a scalar value y as given by
y = w Tx (3.44)
where w is the projection vector of unit length with adjustable weight parameters as
x2
C1
m1
C2
m2 y
w m2
m1
x1
Figure 3.9 A schematic illustration of the two classes C1 and C2 their mean values and the
projection of the means onto the two axes (Bishop, 1995)
The two means m1 and m2 can be calculated as follows
1 1
m1 =
N1 nC1
x( n ) , m 2 =
N2
x
nC2
( n)
(3.45)
32
where m1 is the mean vector for class C1 and m2 is the mean vector for class C2. The
Figure 3.8 shows that the two classes C1 and C2 are well separated in the original
two-dimensional space (x1, x2). The projection onto the x1-axis of the two classes gives a
much larger separation for the two projected means m1 and m2 than does projection onto
x2-axis. On the other hand the separation of data projected onto x2-axis is much better
than onto x1-axis due to the different within-class spread of the two classes along the two
axis. The solution proposed by Fisher is to maximize a function, which represents the
differences between the projected class means, normalized by a measure of the within-
(y ) (y )
2 2
s12 = ( n)
- m1 , s 22 = ( n)
- m2 (3.47)
nC1 nC2
The total within-class covariance for the whole data set is s12 + s22 . By definition the
( m - m1 )
2
F (w ) = 2 (3.48)
s12 + s22
Using Equations (3.45), (3.46), and (3.47) Fisher criterion can be rewritten in the form
w TSB w
F (w ) = (3.49)
w TS W w
33
This expression follows from Equation (3.48) since,
T
(m2 - m1 ) 2 = w T (m 2 - m1 ) w T (m 2 - m1 ) = w TS B w (3.50)
S B = ( m 2 - m1 )( m 2 - m1 )
T
(3.51)
The total within-class covariance is
(x )( ) + (x )( )
T T
(n)
SW = - m1 x(n) - m1 (n)
- m 2 x(n) - m 2 (3.53)
nC1 nC2
The maximum value for F(w) is calculated by differentiating Equation (3.49) with
(w S w)S
T
B W ( )
w = w TS W w S B w (3.54)
From Equation (3.54) the optimum vector w that will give the maximum value for
Maximizing this criterion will maximize the distance between the means of the two
A set of randomly uniform distributed data for two classes C1 and C2 was
generated. The number of points in each class is N = 200. The two sets of data are
34
Class 1
8 Class 2
mean class 1
mean class 2
0
-2 0 2 4 6 8
-2
Using Equation (3.45) the mean vectors for the two classes were calculated
0.984 2.269
m1 = , m2 = (3.56)
1.55 6.812
The within-covariance for the two classes were calculated using Equation (3.55)
35
Using Equations (3.55), (3.56), and (3.58) the optimum projection vector w is
found to be
0.003438
w opt = (3.59)
0.015
The direction of the vector wopt is given by the angle a, which can be calculated, and the
value is
maintenance cost, reduce unscheduled down time, reduce catastrophic failures, improve
recommendations, automation (to reduce human error), a refined reporting system and
ease of operation and use. Commercial vendors, realizing the importance of creating
sound maintenance programs, have introduced software programs and products to combat
the high cost and complexities involved with equipment maintenance. In recent years due
to the changes in technology, software now exists that is capable of exploiting vibration
analysis algorithms and data analysis functions, presenting data with a user-friendly
warning/alarm level functions are only some of the features offered in vibration analysis
software today. This section will comparatively examine three different packages that
36
3.8.1 MOBIUS ILEARN INTERACTIVE
In order to be more effective when monitoring a maintenance program, the end user
should understand the actual signals being collected to properly interpret the results. A
thorough understanding of the relationship between the machine and the characteristics
observed will enable the end user to make educated interpretations. Mobius iLearn
condition monitoring and vibration analysis, regardless of his or her experience in the
field. The CBT is self-paced, incorporating simulations, animations, and samples of real
data and diagnostic requirements creating a realistic, interactive, and valuable learning
tool. The complete curriculum is split into five, non-progressive, separate modules,
37
iLearn Hands-On is a set of vibration measurements taken from a fault test rig.
Over two hundred tests covering dozens of fault conditions are analyzed. Analysis
can be done on the screen or downloaded into the end users data collector.
iLearn Case Histories is a library of spectra and waveforms taken from real
can be done on the screen or downloaded into the end users data collector.
iLearn Signals is a virtual signal generator and spectrum analyzer. This software
program will generate simple signals to teach waveforms and spectra. Advanced
iLearn Machine Faults allows the end user to model a machine to understand
frequencies. The ease of drag and drop technology enable the end user to create a
SVD Inc. provides free online courses on their web site. The Canadian based
company manufactures vibration analysis software and devices. The courses range from
the end users is a web browser and Macromedias Shockwave plug-ins (available as a
free Internet download). The curriculum currently consists of twelve courses and four are
described below:
38
(a) (b)
Figure 3.12 Vibration analysis software for analysis, training and preventive maintenance: (a)
ExpertALERT from Predict-DLI; (b) SpectraScope CAF from Spectral Visualization and
Development, SVD Inc.
Introduction to Machinery Signals the end user is taught the basics of data
acquisition, such as when and how to take measurements, aliasing and the alias
foldover effect, and identifying time and stationary signals. Deterministic stationary
signals and the processes that generate them are covered as well.
Introduction to DSP (Data Signal Processing): Time and Frequency Domain are
two separate courses that are offered by SVD Inc. Data acquisition issues such as
single channel and multi-channel analysis are taught along with unit of measurement.
The concepts of mean, average and correlation, and how they relate to stationary
signals are presented. The basic types of spectral plots and spectral analysis and their
units of measurement are covered in frequency domain along with spectral estimators,
39
3.8.3 PREDICT-DLI
Predict DLI provides onsite training at their Cleveland or Seattle training center or
customized training classes for companies interested in setting up their own onsite
training classes. Predict DLI offers training in vibration analysis, lubricant analysis,
thermography with digital imagery and visual inspections with digital imagery. The
Vibration Analysis I and Machine Balancing is a beginners course for end users
who have little or no experience in vibration data analysis. The course emphasizes
Vibration Analysis II and Laser Alignment this course is a follow-up to the basic
vibration spectra. Problems found in gearboxes and belt driven machines are used for
examples. Alignment tools and techniques are covered to include laser pre-alignment
checks.
Expert ALERT for Voyager is designed for individuals who have purchased
vibration analysis equipment from Predict DLI. Software commands and functions
are discussed as well as analyzing, fine tuning data and manipulation of various
plotting and display functions. Emphasis is placed on setting up the database and
40
4 HELICOPTER VIBRATION REDUCTION THROUGH ROTOR
In the military and civil helicopter community, one of the standard procedures
before an aircraft can be airborne and perform its mission is Rotor Track and Balance
(RT&B). The main reason for doing RT&B is to smooth the helicopter hover and in flight
vibrations. A low level of vibration is desired mainly because of the following reasons:
Increase components life: less vibrations induced from the rotor through the main
frame to the main components of the helicopter, will reduce the chances of
incipient damage to occur and hence will extend the life of components.
Increase maneuverability: a low level of vibration due to a good RT&B will make
the helicopter more maneuverable and will give pilots quicker time to react to the
Increase comfort of the pilots: a less shaking environment for the pilots will
make them more accurate in making decision and will have a reduced impact on
their health.
All the above advantages of a smooth aircraft, will also translate in long-time cost
savings for helicopters operators, increase readiness of the helicopters, and decreased
41
4.1 RT&B PROCEDURE
In the military helicopter community, the procedures for RT&B are standardized.
In the Army the equipment used for generating the necessary adjustments is called
followed when doing RT&B. This sequence is presented in the Table 4.1. Each step in
GNDTRK is a ground plan designed to get the rotor in a flyable track prior to
hovering. It consists of one measurement, and it adjusts the pitch links only. It should be
the first test that is performed after any component change or major work on the rotor.
INITIAL
INITIAL is a flight plan performed after the GNDTRK, and will balance the rotor
42
FLIGHT
vibration using tabs only. It is to be used after the GNDTRK and INITIAL flight plans,
TAIL is a flight plan designed to balance the tail rotor. It uses the optical RPM
tail rotor. A RT&B scenario consists of the following operations. First the helicopter
performs on the ground with the blades having a flat pitch at 100 % nominal rotation
(NR) of the main rotor. When the rotor has reached 100 % NR, data is collected from the
two accelerometers placed in the cockpit for measuring the vertical and lateral vibration.
If the vibration level is within the limits, then the next task will be performed. If not, a set
of adjustments will be generated to reduce vibration at FPG 100. After making the
The next task is hover, when the helicopter takes off from the ground and hovers
above the ground. Data is collected and, if needed adjustments, are made and then the test
is repeated.
After hover, the next task is flight. During this task the aircraft is performing a
forward flight at different air speeds varying from 60 Kts (Knots) to 140 Kts. Data is
collected during the flight and, if the vibration levels are above the allowable values, the
AVA software is used to generate a set of adjustments that will bring the vibration level
A schematic of the RT&B test plan is presented in the following Figure 4.1:
43
Flight
Vibration Hover Vibration FPG 100 Vibration
FPG 100 amplitudes amplitudes
amplitudes Hover
60-140Kts
Adjustments Adjustments
Adjustments
Pitch link adjustments: brings the blade in track by moving the entire blade up or
down by a specified amount. Pitch link adjustments will affect all the test states.
Trim tab adjustments: consists of bending thin aluminum tabs (pockets) located
on the trailing edge of the blade. The tabs are adjusted up or down with a
specified angle. This will make the blade to go up or down at certain air speeds.
Hub weight adjustments: addresses changes in weights locate at the root of the
the blades. It is desired that the blades that are in pairs be balanced in the same
way.
44
4.1.1 TEST EQUIPMENT INSTALLATION/CHECKOUT
Test equipment installation/checkout refers to all the steps necessary for RT&B
prior to flight and data acquisition. Detailed installation procedure is given for those parts
following: Data Acquisition Unit (DAU); Universal Tracking Device; Magnetic rpm
45
4.1.2 FLAT TRACK MAIN ROTOR ON THE GROUND
Track and balance of the main rotor should be performed when any of the
following occurs: one or more blades have been changed; one or more tip caps have been
changed; one or more pitch control rods or rod end bearings have been changed; the main
rotor blades have been removed and the hub has been disassembled. The procedure
consists of:
via the Auxiliary Power Unit (APU), and operate the aircraft at 100% Nr, flat
pitch
corrections will be performed.To flat track the main rotor on the ground, the main
correction that is performed is the main rotor pitch link adjustment. The main rotor pitch
Figure 4.3 Main rotor pitch link adjustment (AVA Technical Manual)
46
The standard procedure for pitch link adjustment is as follows: first the barrel and
upper rod end needs to be marked with a single vertical line. Next, the upper and lower
jam is loosed. Turn the barrel the amount specified in the smoothing solution provided by
the RT&B algorithm. Last, tighten and secure the jam nuts.
Preparation: first, flat track the main rotor on the ground as described in the
previous section
Main rotor in-flight vertical tracking should be performed when any of the
following occurs: 1/REV in-flight vibration is unacceptable; changes have been made to
the track of one or more main rotor blades as a result of performing ground tracking
Equipment installation
(FPG100, Hover, 60 Kts, 80 Kts, 100 Kts, 120 Kts, or 140 Kts) track and
vibration data are acquired. If the measured values exceed specifications perform
47
that are mare are the trim tab adjustments. To perform these adjustments a special tool,
Figure 4.4 Tab adjustment tool used to bed the tab of the helicopter blades for RT&B (AVA
Technical Manual)
An example to illustrate the trim tab maneuver is presented next. Suppose that the
recommended adjustments are 1 degree up at pocket 4/10. The tab adjustment tool should
be positioned as indicated in Figure 4.5 along the tab of the blade. The bend should start
Figure 4.5 Trim tab pockets for a helicopter blade (AVA Technical Manual).
48
4.1.5 TAIL ROTOR BALANCE
Tail rotor balance should be performed after any rework/replacement of tail rotor
Equipment installation: install the vibration measuring equipment and the optical
rpm sensor.
Perform tail measurements: operate aircraft at 100% NR flat pitch; acquire tail
rotor vibration data and process the data. If the results are less than the values
The data acquired from the accelerometers mounted on the helicopter structure is
preprocessed in the Data Acquisition Unit (DAU). The resulting data is divided into two
domains: synchronous order domain data (SOD) and asynchronous frequency domain
data (AFD).
For RT&B, the data used is the synchronous order domain. Raw data is acquired
from two accelerometer mounted in the cockpit, one for vertical vibrations and one for
lateral vibrations. The raw data is first preprocessed in the DAU where it is divided into
segments of equal length related to the synchronous signal (tachometer) and averaged
together. The synchronous signal is then transferred from the DAU to the crew chiefs
laptop and then to the ground base station (GBS). On the GBS the synchronous signal is
processed using a neural network. The solution of the neural network is a set of
49
recommended adjustments that will best reduce the vibration level (if high), using a
Recommended
Adjustments
Raw data
Ground Station
SOD data
Crew chiefs
laptop
Data
Raw data desensitizing
SOD data
USC-VMEP
Crew chiefs server
laptop
Engineering. During the transfer process data is being desensitize by removing all the
sensitive information. From the VMEP server, each user can download the data of
particular interest, and perform specific analysis on local computers. In this way the
integrity of the data is kept intact and the possibility to alter the data is minimized. For
security reasons the users have been divided into categories and each one of the
categories have specific privileges on the data. The VMEP directory structure for the
VMU data, and the locally stored structure are presented in Figure 4.7.
50
VMEP Server
VMU Data Directory
Locally stored
VMU data
Figure 4.7 VMU Data storage and directory configuration: on the VMEP server and on the local
computer
The first task to perform when studying the RT&B vibration data is to identify
which flight is a RT&B flight. In the first phase, this task is done manually by looking at
the different types of files that are stored in a particular flight folder. For example if we
look at Figure 4.8b for the flight ID 1002035095 there are three files: FLIGHT,
FPG100_SOD, and FPG100_TRK. The three files contains the following information:
FLIGHT: general information about the flight that was performed: aircraft type,
the cockpit, nose, lateral and vertical. Amplitude and phase of vibration are
recorded from 1/rev up to 12/rev. The state of the test is flat pitch ground 100
51
(FPG100) as explained in Table 4.1. Notice that the state of the test is also the
FPG100_TRK: track data for the four blades. Also includes general identification
a)
b) c)
with FPG100 and continue up to 140 Kts air speed, as shown in Figure 4.8c. Based on
this assumption, the RT&B flights were identified and stored separately for further
analysis.
The historical trend for vibration over a period of time is expected to behave as
presented in Figure 4.9. For a smooth helicopter the vibration level is low. As the flight
hours for that particular helicopter increases, the vibration level gradually increases above
the desired level low-vibration. When the vibration level reaches the level of allowable
maximum vibration, it is time for a new RT&B. After the new RT&B is performed, the
vibration level dramatically decreases below the desired level. This process is shown in
Figure 4.9.
52
Allowable max vibration level
RT&B RT&B
looked at the vibration data acquired for one particular aircraft, and for one particular test
state (FPG100) over several months. The results presented in Figure 4.10 represent the
vibration amplitudes versus time. The time is represented by the date in (yy/mm/dd)
format.
0.45
0.4
0.35
0.3
Amplitude [ips]
0.25
0.2
0.15
0.1
0.05
0
010502 010502 010601 010816 010816 011009
Date
Figure 4.10 Historical trends for vertical vibration for one aircraft and one particular test state,
showing an increased vibration over a couple of months
53
Figure 4.10 shows that the experimental data collected over a period of time is in
concordance with the theoretical behavior of the vibration level between RT&B flights.
The vibration level is low in the first month, May 02, 2001 (015002) and gradually
increases as time increases eventually crosses the desired level for vibration, which is 0.3
ips (inch per second). We conclude that it might be it time for a new RT&B of the
helicopter to bring down the vibration level below the desired level.
Of interest in the analysis of the vibration data is the polar plot. A polar plot is a
angle. Knowing the exact position of the magnetic sensor mounted on the main rotor with
respect to the blades, the location and the blade where the high vibration occurred can be
determined.
0.6
0.2
120K 0.5
140K Hover
100K
Amplitude [ips]
0 0.4
-0.2
0.2
100K 140K
0.1
-0.4
120K
0
Hover FPG 100 Hover 60K 80K 100K 120K 140K
-0.6 Test state
(a) (b)
Figure 4.11 a) Polar plot for the vibration data between two RT&B flights b) Plot of the amplitude
values for the same flights for all test states.
When looking at a polar plot, there are two things of interest: how spread out the
points are, and how far away from the origin. If we look at the Figure 4.11a we can see
that for the first flight (triangles) the points are scattered and out of the circle that delimit
54
the 0.3 ips vertical vibration level. The VMEP neural network algorithm generates
adjustments that will try to bring the vibration amplitudes towards the origin. After the
recommended adjustments were made, in the next flight (diamonds) we can see how the
points had moved towards the center indicating a low vibration level. The same
information is given in Figure 4.11b where, for the first flight the vibration levels at
hover, 100 kts, 120 kts, and 140 kts are all above the desired level of 0.3 ips. After
making two adjustments, the vibration level decrease dramatically and the result is a
smooth helicopter.
In the analysis performed, we were also interested to see how the RT&B
adjustments will affect the vibration level at different test states. The goal of the RT&B
VMEP algorithm is to bring down the vibration level to a desired limit by using fewer
adjustments and thus less RT&B flights. A reduced number of RT&B flights will
translate into an increase mission capabilities of the helicopter and in cost savings.
1.6
1.4
1.2
1
Amplitude [ips]
0.8
0.6
0.4
0.2
0
FPG 100 Hover 60K 80K 100K 120K 140K
Test state
Figure 4.12 Plot of vibration amplitudes showing an increased vibration level for the three flights
55
In Figure 4.12 we can see that the two flights performed on October 2, 2001 and
on October 4, 2001 have a significant increase in the vibration level, especially at high
speeds (80 Kts to 140 Kts). These are strong indicators that a new rotor track and balance
is needed.
0.9
0.8
0.7
Amplitude [ips]
0.6
0.5
0.4
0.3
0.2
0.1
0
FPG 100 Hover 60K 80K 100K 120K 140K
Test state
Figure 4.13 RT&B flights showing how the vibration amplitudes reduces after making recommended
adjustments
After the first set of adjustments the vibration level for 60 Kts, and 80 Kts is below the
desired level for vertical vibration of 0.3 ips. However is still above the desired level for
hover, 100 Kts, and 140 Kts (triangles) as it can be seen in Figure 4.13. The next set of
adjustments brought most test states below the desired level, with the exception of (120
Kts and 140 Kts) (doted line). The final adjustments that were made brought the vibration
amplitudes for all test states to a very low level (blue circle line).
In conclusion, with the VMEP neural network algorithm, it took only three RT&B
flights to bring the vibration level down to a desired low level for vibration.
56
5 VIBRATION-BASED DAMAGE DETECTION/HEALTH
MONITORING
The wavelet transform is a signal-processing tool, which allows both the time
Performing a wavelet transform consist of convolving the signal with time shifted and
dilated. The result of wavelet transform will be a set of coefficients, which are function of
time and frequency, also called scale. These coefficients can be used to form a unique
basic wavelet function to form sets of filters, each one consisting of lowpass and highpass
filter. The signal is pass through the first set of filters and the result will be two signal
each with half of number of coefficients as the original signal. The signal formed using
the lowpass filter and thus containing the low frequency information is known an the
approximation, and the second signal formed using the high pass filter and thus
containing the high frequency information is known as the detail. For the second
recursion the approximation is passed through the next set of filters and so on until an
57
Newland (1993) presented the harmonic wavelet basic function, which is a
completely band limited in the frequency domain. A consequence of the above is that
DHWT need not be restricted to octave frequency bands. This form of DWT is known as
compute DHWPT is the Mallat recursive algorithm and a comparison between the two
Signal Signal
A1 D1 A1 D1
Figure 5.1 A comparison of the DWT algorithm and the DWPT algorithm. Ai is approximation at
level i (low frequency) and Di is detail at level i (high frequency). (Samuel et al., 1998)
Samuel et al. (1998) collected and analyzed data from an OH-58A main rotor
transmission. The test was run at 6060 rpm (100% of the maximum speed), which
resulted in a mesh frequency of 573 Hz, for nine days, eight hours per day at a 117%
design torque as part of an accelerated fatigue test. The results were represented in mean
58
Figure 5.2 Mean square wavelet maps for day one and day nine. On day nine a significant
increase in the amplitude of neighboring frequency bands is observed as the fault
developed (Samuel et al., 1998)
The mean square wavelet maps clearly shown the presence of the fault in day nine. Using
the normalized power computed for the mash frequency and its accompanying frequency
bands the evolution of the fault from day one to day nine is presented in Figure 5.3.
rotor system during the run-up period with a speed ramp rate of 100[rpm/sec] for the case
59
of a rotor with 10% crack (transverse) depth relative to its diameter. To validate the
advantages of DHWT, short-time Fourier transform (STFT) has been applied to the same
set of data (noise contaminated signal) and the results are presented as follows:
Figure 5.4 a) Directional spectrogram: 10% crack with 5 % noise; b) Directional harmonic wavelet
map: 10% crack with 5% noise (Kim and Ewins, 1999)
The results presented in Figure 5-4 reveal the advantage of DHWT over the STFT,
because the results from DHWT are insensitive to the random noise while the STFT
Polyshuchuk and Choy, (1998) used the WVD techniques, introducing a new
parameter NP4, to experimental data obtained from a helicopter tail gear transmission.
The damage introduced is a single-tooth damage in the tail gear. Figure 5.5 shows the
WVD and the instantaneous power plot for an undamaged gear. The two large
components are the first and second harmonics of the gear-mesh frequency. The WVD
and the instantaneous power plot for a damaged gear are shown in Figure 5.6
60
Figure 5.5 Accelerometer data for an undamaged gear. (Polyshchuck et al., 2000)
Figure 5.6 Accelerometer data for damaged gear (Polyshchuck et al., 2000)
Those results proved that the use of WVD could be a good tool in fault detection and
failure prevention.
61
5.3 NEURAL NETWORK-BASED AND NEURO-FUZZY METHODS FOR
as an accurate off-line monitoring tool used to reduce helicopter maintenance costs, and
an all analog neural network as a real time helicopter gearbox fault monitoring. The
intermediate gearbox known as the Hollins TH-1L 42-deg tail rotor. Vibration data was
recorded from two Endeveco 2220C accelerometers. A low pass filter with a cutoff
frequency of 10 kHz was applied when data was generated. The main objectives of the
hybrid and analog neural network are: fault detection, fault classification and fault
detect faults and the analog neural network is presented in Figure 5.7 and Figure 5.8.
Figure 5.7 Hybrid system, fault detection performance. 60-point discrete Fourier transform (DFT)
was performed for fault detection, 128-point and 256-point DFT for fault identification.
62
Figure 5.8 All analog system performance. For fault detection in the case of all analog neural
network. 60-point DFT was performed (Monsena et al., 1995)
The results indicate that a system employing 60-point DFT was capable of solving the
fault detection problem. For the fault classification and identification problem, a 256-
point DFT was required for perfect system performances. The performance results by
using the all analog neural network suggest that it is possible to achieve 100% fault
technique for monitoring the health of helicopter gearboxes. This technique is based on
neuro-fuzzy algorithms for pattern clustering, pattern classification, and sensor fusion
(Cho, 1997). The vibration data used was obtained from an aft main power transmission
of a CH-46E helicopter. Frequency domain and wavelet analysis techniques were used to
analyze the data and prepare them for the neural network inputs. To train the neural
was used. A feedforward backpropagation neural network was used to classify the
different faults. In the preprocess part of the vibration data, both auto power spectral
63
density (APSD) and wavelet coefficients were used [12]. A list of the fault types is
presented in Table 1.
Table 5.1 List of the fault types created in the test gearbox (Essaway et al., 1998)
Fault # Fault type Fault # Fault type
Fault 2 Planetary bearing corrosion Fault 6 Helical idler gear crack
propagation
Fault 3 Input pinion bearing corrosion Fault 7 Collector gear crack
propagation
Fault 4 Spiral bevel input pinion Fault 8 Quill shaft crack propagation
spalling
Fault 5 Helical input pinion chipping Fault 9 No defect
The results obtained using the first feature extraction method (129 points APSD) and
second extraction method (wavelet transform) are promising and they are presented in
Table 5.2 Neural Network Classification results using APSD features at 100% load (Essaway et
al., 1998)
Fault Load Acc1 Acc2 Acc3 Acc4 Acc5 Acc6 Acc7 Acc8
Fault 2 100% Train 100% 100% 100% 100% 100% 100% 100% 100%
Test 100% 100% 100% 100% 100% 100% 100% 100%
Fault 3 100% Train 100% 100% 100% 100% 93.3% 90% 100% 100%
Test 100% 100% 100% 100% 90.9% 100% 100% 100%
Fault 4 100% Train 100% 100% 100% 100% 86.7% 100% 100% 100%
Test 100% 100% 100% 100% 81.8% 100% 100% 90.9%
Fault 5 100% Train 100% 100% 100% 100% 70% 93.3% 100% 96.7%
Test 100% 100% 100% 100% 72.7% 100% 100% 100%
Fault 6 100% Train 100% 100% 100% 100% 80% 76.6% 100% 100%
Test 100% 100% 100% 100% 27.3% 90.9% 100% 100%
Fault 7 100% Train 100% 100% 100% 100% 100% 93.3% 100% 100%
Test 100% 100% 100% 100% 100% 90.9% 100% 90.9%
Fault 8 100% Train 100% 100% 100% 100% 96.7% 83.3% 100% 100%
Test 100% 100% 100% 100% 81.8% 63.6% 100% 100%
Fault 9 100% Train 100% 100% 100% 100% 90% 93.3% 100% 100%
Test 100% 100% 100% 100% 63.6% 100% 100% 81.8%
64
Table 5.3 Neural network classification results using wavelet features (Essaway et al., 1998)
Accelerometer# Fault2 Fault3 Fault4 Fault5 Fault6 Fault7 Fault8 Fault9
Acc 1 100% 100% 66.67% 100% 100% 100% 100% 88.89%
Acc 7 (14x14 SOM) 100% 88.89% 88.89% 100% 100% 100% 100% 100%
Acc 7 (9x9 SOM) 100% 88.89% 44.44% 100% 100% 100% 100% 100%
From the above results it can be seen that the neuro-fuzzy technique using both APSD
and wavelet transform, even though classification results was not perfect for all sensors,
65
6 HELICOPTER HEALTH AND USAGE MONITORING
SYSTEMS (HUMS)
In the past 10 years, the movement for Health and Usage Monitoring Systems
(HUMS) has become a hot topic in the helicopter community. This movement started in
the United Kingdom with systems being fitted to helicopters operating in hostile
environment of the North Sea. The driving force for behind the introduction of such
systems has been the need for improved helicopter safety and reliability. The HUMS are
providing valuable information about the health of the helicopters engines, drive train
airworthiness related arisings were detected on the North Sea helicopters, where an
arising is defined as an event, which has led to significant maintenance actions. The
planning purposes
Usage: manage the usage of the aircraft and controls life-expired component
replacement
66
Monitoring: main focus is directed toward providing diagnostic and prognostic
Integrated: HUMS will be an integral part of the aircraft maintenance and logistic
system, and the HUMS outputs will be fully integrated into the aircraft
maintenance policy
Due to some time and resources related restriction, civil operators have not been able
to gather cost related data to prove the benefits of HUMS. However there is no doubt that
- Simplified rotor track and balance procedures and reduced RT&B flights;
- Reduced airframe vibration levels, resulting in fewer avionics faults and other
67
program is sponsored by Army/Navy, and the main contractor is Goodrich
monitoring and fault diagnostic tool used to find faults in the helicopter rotors,
and the total number of maintenance test flights required to complete maintenance
in flight data obtained on a routine basis from line operations. The main
System). This project is a Stage I prototype demonstration under the Office of the
Goodrich Aerospace. The goal of the IMD HUMS project is to integrate and test a
68
commercial/military dual use mechanical diagnostic system on the H-53 and H-
health and usage monitoring system using field proven, off-the-shelf, airborne and
by detecting and diagnosing potential and actual failures, monitoring usage, and
Bell, Teledyne Controls, and Stewart Hughes have developed the 412 Bell
HUMS. The 412 Bell provides the essential tools for reducing direct operating costs,
The key features of the 412 Bell HUMS are: flight data recorder crash survivable
dynamic components; vibration monitoring of the drive train and engines; fully
69
AT) rotor track and balance system, including rotor fault diagnostic; monitoring and
capturing of aircraft limit exceedances; electronic log book data, including power
The potential benefits of the 412 Bell Hums can be listed as follows: reduced non-
revenue maintenance flights; insurance savings; log book accuracy; reduced false
The configuration of the 412 Bell helicopter HUM system is presented in Figure
6.1. The system architecture can be divided into to parts: on-board equipment and
ground-based equipment.
Computer are stored in the HUMS Processor Unit. Pilots have the possibility at any time
70
to monitor the functionality of the helicopter using the HUMS Display Panel. Optional
the data can be stored on a solid state Cockpit Voice and Flight Data Recorder.
After landing the helicopter, data can be transferred from the HUMS Display
Panel through a PC card to the HUMS Ground Station, or from the Processor Unit to the
HUMS Ground Station through the HUMS Data Retrieval Unit. From here, the data is
The number of sensors and data sources on the Bell 412 helicopter HUMS are:
A typical installation diagram for Bell 412 HUMS showing the location of the
sensors, and the data collecting and display hardware is presented in Figure 6.2.
71
Other HUMS products developed by Teledyne Controls and Stewart Hughes
currently in service in North America, Europe, Australia and Southeast Asia, are: North
set of functions for rotorcraft health and diagnostics. These functions support RT&B
usage, exceedance detection and processing and engine power assurance checks (Hess et
al., 2001).
The typical IMD HUMS functions are: engine performance assessment (EPA);
rotor track and balance; absorber tuning using the vibration information collected from
the accelerometers mounted in the cabin; mechanical diagnostics of the vibration data
The hardware consist of the Data Transfer Unit and Cockpit Display Unit (DTU
& CDU), the Main Processor Unit (MPU), and the Remote Data Concentrator (RDC) as
72
(a) (b)
Figure 6.3 IMD HUMS hardware configuration; (a) Main CH-53E IMD HUMS display; (b) Location
of the IMD HUMS on the Army UH-60A (Hess et al. 2001)
The IMD HUMS is composed of two major elements: the on-board acquisition
storing data obtained from sensors located throughout the aircraft. The principal element
of the OBS is the Main Processing Unit (MPU), which analyzes the input data for
exceedances and events, calculates various flight regimes, performs various diagnostic
algorithms, and stores the data to an onboard data cartridge. The Remote Data Collector
(RDC) is another element of the OBS system. It is used for the aircraft that do not support
modern avionics bus communication to collect the data from the sensors.
The Ground Station (GS) is the primary user interface with the HUMS system. It
is responsible for logging and maintaining all flight and maintenance data, generating
maintenance-due lists based on flight data, performing aircraft configuration and pars
tracking, generating engineering and managements reports, and archiving data (Hess et
al., 2001).
73
The IMD HUMS program is currently in progress on the Navy CH-53E, and the
Navy SH-60B. The CH-53E IMD HUMS was transitioned from development test into
operational test phase. It completed a 97 hours development test program in 2000 with
IMD HUMS transition from the development test phase to the operational test phase was
Other IMD HUMS programs will target the Army UH-60A and HH-60L, the
The VMEP system was intended primarily as an aid to the Army aviation
maintainer. The goal is to reduce unscheduled maintenance and the total number of
maintenance test flights required to complete maintenance activities. The intent of the
system is to capture, during operational flights activities, the vibration environment of the
major components of the helicopter. The vibration data is analyzed by the embedded
system to produce a set of specific machinery health indicators. These indicators are
stored for post flight use by the aircraft maintainers. Maintenance action are correlated to
these health indicators and communicated so that actions can be scheduled (Grabill et al.,
2001)
Rotor smoothing also known as Rotor Track and Balance (RT&B), which, is a
function that minimizes the fundamental harmonic of main rotor and tail rotor
vibration. Adjustments for weight, trim tab and/or pitch link length are provided.
74
Drive Train Diagnostics accelerometers measurements for all drive train
gearbox, and tail rotor gearbox. Classical machinery monitoring algorithms are
track and balance maintenance test flights; mitigate Apache Readiness Improvement
Program (ARIP) problems; reduce aircraft operation costs; increase aircraft availability;
To achieve the above stated goals and objectives, the methods and approach used
in the VMEP program are: to measure and record in-flight vibrations, process vibration
75
data, apply signal analysis methods and identify hot-spots, fine-tune track and balance
to reduce vibrations, and identify incipient failures and anticipate maintenance actions
The design and impact of the VMEP system on the aircraft must be minimal. The
requirements are to provide a system with open architecture, using Commercial off-The-
Shelf (COTS) hardware components and COTS software components. The key features
vibrations
monitoring
Data can be transferred to WEB-GBS for long term trends and aircraft
comparison
The VMEP HUMS consist of two systems, the On-board System and the PC
sensors that are connected to the Vibration Management Unit (VMU). The VMU is fully
76
integrated and qualified with the Army airframe. The VMU is a stand-alone system that
does not depend on, or adversely interact with, any other system on the aircraft. On the
helicopter airframe there are 18 accelerometers mounted, three tachometers and a blade
tracker.
hardware and software, which provides a low cost, easily upgradeable system for on-
board diagnostic such as rotor track and balance, drive train vibration monitoring, turbine
engine diagnostics, and gear box monitoring. The total weight of the VMU is 4 lbs and 5
77
Figure 6.6 VMEP configuration for AH-64A
Figure 6.7 Pictures of the accelerometer mounted on different components of the helicopter
78
PC-Ground Based Station (PC-GBS) uses a Windows based software that
receives measured data from the on-board system, stores this data into a database and
performs analysis of the data producing easy to understand corrective actions for the
maintenance personnel. The PC-GBS is also designed to operate on any computer with a
Pentium 133 MHz or faster processor. The PC-GBS transfers the aircraft data from the
VMU for at-aircraft maintenance. The PC-GBS uses a neural network based rotor
smoothing algorithm to calculate the maintenance actions required to optimize the rotor
data to provide a comprehensive indication of the health of the aircraft machinery such as
engine, drive-shafts, gearboxes, bearings, pumps, and generators (Grabill et al., 2001).
requires the user a minimum training to operate the software and perform the basic
functions without any prior expertise in rotor track and balance and vibration analysis.
79
6.2.2 VMEP HUMS VIBRATION DATA COLLECTION
The data collection process starts in the cockpit when the pilot/copilot interacts
with the VMU by selecting a data collection MODE and aircraft STATE and
pressing the data collection DO button. An indication of the status of the data
After completion of the flight the vibration data stored in the VMU is transferred
via the crew chief laptop to the PC-Ground Based Station (PC-GBS) as shown in Figure
6.9.
Data Desensitizing
Ground station
Figure 6.9 Vibration data flow diagram showing downloading of data from the VMU to the ground
station and further to the USC teradata computer
From the PC-GBS the data will be transferred to USC for cost benefit analysis.
components. After the desensitizing process is complete, the vibration data is transferred
80
to the USC teradata computer. From the secured server the data can be utilized by various
end users to perform specific analysis. It is important to mention that the raw data from
the VMU stored on the USC teradata computer is protected against any kind of
accidentally pollution of the data, i.e. deleting data or adding data, by hierarchiezd access
The first step when dealing with the vibration analysis of complex machinery like
helicopters is to devide the whole assembly into components, and to identify the main
vibration features of each component. Using the mechanical drive schematic presented in
Figure 6.10, the major vibration sources can be identified: gear-clashing tones, bearing
81
Based on the information furnished by the mechanical drive schematic, the shaft
frequencies and the mesh frequencies of the components can be calculated. A list of the
components and their frequencies for the AH-64A mechanical drive system is presented
in Table 6.1.
Table 6.1 Shaft frequencies and mesh frequencies for the main components of the mechanical
drive shaft of AH-64A helicopter
Gearbox Component rpm/ Mesh Frequency Shaft Freq
nr. of teeth Hz Hz
Tail Gearbox shaft1 3636 1333.2 60.6
gear1 22
Intermediate Gearbox tail rotor drive shaft 4815.25 2969.4 80.3
gear 37
#1 & #2 Nose Gearbox shaft 20952 10825.2 349.2
gear 31
shaft 9841 10661.1 164.0
gear 65
#1 & #2 Engine shaft 20952 349.2
Hydraulic Pump hydr. pump shaft 12635 6738.7 210.6
gear 32
AC Generator AC gen shaft 12251 6738.1 204.2
gear 33
Aft & Fwd Hanger Bearing tail rotor drive shaft 4815.25 0.0 80.3
Having this information, the spectral analysis of the vibration data recorded by
particular sensors is possible. In the frequency spectra, each component can be easily
identified using its characteristic frequency. Changes in the frequency spectra at certain
Monitoring changes in the frequency spectra can give indication about the life and health
Diagnostics algorithms that are embedded into the VMEP software generate
82
resonance of the system. A condition indicator uses a set of measured data as input and
produces a single real number as output. This single number can be thresholded, trended,
or analyzed, to provide an indication a fault presence, the location and the type of the
fault.
The data downloaded from the Vibration Monitoring Unit (VMU), is of two
types: synchronous order domain data (SOD) and asynchronous frequency domain data
(AFD). In the SOD data file are stored twelve vibration amplitudes acquired from three
sensors, nose, lateral and vertical located in the cockpit of the helicopter. The 12
vibration amplitudes correspond to the rotor frequency, 1/rev up to 12/rev. The AFD data
file contains 400 data points for a frequency range from 0 to 1000 Hz. The data is
The feature investigated in the present work is the amplitude of the frequency of
particular components. As mentioned in section 6-3, the rotational speed for each rotating
component was identified and the corresponding frequencies were calculated. The mesh
By plotting the vibration information from each sensor in the frequency domain, a
frequency spectrum is generated where the components can be identified based on their
own frequency.
Because there is no baseline to compare with the data collected from the
accelerometers, first we tried to look on how the frequencies of the components modify
in time and so to establish a baseline. To do so we identify the variables that are part of
83
the feature extraction as being: aircraft ID, state status, and the sensor. Base on these
As Figure 6.11 and Figure 6.12 show, the main frequencies of the components
and their harmonics can be clearly identified in the frequency spectrum. The
abbreviations used in these plots are: tail rotor drive shaft (TRDS), engine drive shaft
Tail Gearbox
0.6
80.3 Hz
TRDS
0.5
0.4
Amplitude [g's]
0.3
164 Hz
408 Hz
Input DS 591 Hz
2 nd overtone of
7th overtone of
AC Generator
0.2 TRDS
490 Hz
3rd overtone of
Input DS
0.1
0
0 100 200 300 400 500 600 700 800 900 1000
Frequency [Hz]
Figure 6.11 Frequency spectrum for the tail gearbox sensor for the same aircraft, same test state
(FLIGHT), and the same sensor.
Both, Figure 6.11 and Figure 6.12 shows that the amplitude of vibration for the
components is changing over time. After sufficient data has been collected a baseline for
the allowable amplitudes of the vibrations can be establish and each frequency can be
84
compared to the baseline. If damage will occur in any of the components being
monitored, the frequency spectrum will change and looking at the frequency at which the
change is present, the component can be identified and appropriate maintenance action
can be taken.
80.3 Hz
2.5 TRDS
2
Amplitude [g's]
1.5
1
164 Hz 325 Hz 975 Hz
Input DS EDS 3 rd overtone of
408 Hz EDS
0.5 2 nd overtone of AC
Generator
0
0 100 200 300 400 500 600 700 800 900 1000
Frequency [Hz]
Figure 6.12 Frequency spectrum for the intermediate gearbox sensor for the same aircraft, same
test state (FLIGHT), and the same sensor.
The second case considered is to analyze the vibration data for the same test state,
the same sensor but different aircraft. By doing so we are interested to see if the
frequency spectrum for each sensor is changing from one aircraft to another aircraft and
Figure 6.13 and Figure 6.14 show that the frequency spectrum for the two sensors
considered is not changing (shifting of main frequencies), except for changes in the
85
#1Nose Gearbox
3
700 Hz
2 nd overtone of
Engine Drive
Sh ft
2.5
2
Amplitude [g's]
656 Hz
4 th overtone of
1.5 Input Drive Shaft
350 Hz
EDS
1
534 Hz
164 Hz 3 rd overtone of
Input DS Fuel Boost Pump
0.5
0
0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000
Frequency [Hz]
Figure 6.13 Frequency spectrum for the #1 nose gearbox sensor for the same sensor, same test
state (FLIGHT) but different aircraft.
Intermediate Gearbox
4
80.3 Hz
TRDS
3.5
2.5
Amplitude [g's]
803 Hz
1.5 10 th overtone of
TR Driveshaft
1 241 Hz 883 Hz
3 rd overtone of 11 th overtone of
TR Driveshaft 405 Hz
23.5 Hz 5 th overtone of TR Driveshaft
0.5 TR 1/rev TR Driveshaft
0
0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000
Frequency [Hz]
Figure 6.14 Frequency spectrum for the intermediate gearbox sensor for the same sensor, same
test state (FLIGHT) but different aircraft
86
7 DAMAGE DETECTION OF A SIMULATED CRACK USING
DISTRIBUTION
applied to fault detection and machine-health diagnosis. The signal processing methods
using only Short Time Fourier Transform (STFT). The spectral analysis developed for
these arguments a new approach is considered for machine health diagnosis for non-
done in the past years by various authors using WVD mainly for gearboxes diagnostics
and prognostics, with encouraging results. Only a few work has been done using WVD
for crack detection, therefore not too many results are available.
87
The purpose of this work is to perform a comparison between diagnostic features
based on the new WVD approach and the traditional power spectral density (PSD) of the
The basis for using the forced oscillation method for the fatigue crack diagnostics
lies in the fact that in a cracked object the stiffness is nonlinear and the nonlinearity level
changes with the crack size (Gelman et al. 2000). Assuming that the object under analysis
can be approximated with a single degree of freedom nonlinear oscillator, the equations
x
where X = , x is displacement, m and c are the object mass and damping coefficient
m
kS k c c
wS = , wC = C , z S = , zC = . (7.2)
m m 2 kS m 2 kC m
At compression, the crack is closed and the material behaves like a continuum;
hence, the compression stiffness is the same as that of the material without crack,
kC = k
At stretching, the crack is opened and the material is discontinuous, therefore the stiffness
88
Dk = k - k S
as shown in Figure 7.1
Stretching
a h
Crack opening
Compression
Crack closing
Figure 7.1 Stretching and compression of the cracked object. In stretching crack is opening and in
compression crack is closing
89
For the axial loading case considered in the present work, the stiffness change due
to the crack is related to the width of the object (cross section area) hence the linear
approximation holds. For bending, the linear approximation is not true anymore, the
stiffness change being proportional to the width of the object raised to the power of three
(moment of inertia).
The variables w f and j are the constant excitation frequency and random initial
phase (Gelman et al. 2000). The random initial phase is uniformly distributed in the
interval [0; 2p ] . The frequency of the cracked object depends on the crack size a
w s2 a
2
= 1- = 1- r (7.3)
wc h
where r is the relative crack size,
a
r= (7.4)
h
The period of oscillation for the cracked object is given by:
2p T1 T2 1 2p 1 2p p p
T= = + = + = + (7.5)
w 2 2 2 ws 2 wc ws wc
Hence,
w S wC
w =2 , (7.6)
w S + wC
Using Equation (7.3) the frequency of the cracked object becomes:
2 1- r
w =w n (7.7)
1+ 1- r
k
where w n is the natural frequency of the un-cracked object, w n = , and r is the
m
90
The relation between the frequency and the relative crack size is shown in Figure
7.3. As the crack size increases the frequency of the structure decreases, as a consequence
25
20
Frequency [Hz]
15
10
0
0 0.2 0.4 0.6 0.8 1
Relative crack size [%]
91
7.2 DAMAGE DETECTION STRATEGY
The damage detection strategy used in the present work is based on the forced
oscillation response and spectral analysis features. Crack presence in the structure
modifies (decreases) the effective stiffness and hence the main frequency as shown in
Figure 7.4. In addition, the nonlinear character of the equations leads to parametric
resonances, which appear as additional features in the frequency spectrum. Higher order
Main frequency
1st and 2nd harmonic
of main frequency
investigation, we tracked the main frequency changes for increasing relative crack size (r
= 0, 0.1, 0.4, 0.6). For each relative crack size, the nonlinear time-response signal was
generated using numerical solution of Equation (7.1). The, the power spectral density
(PSD) of the signal was computed using two methods: (a) short time Fourier transform
92
After computing the power spectral density (PSD), the peaks in the frequency
investigated.
The entire numerical simulation was carried out using MATLAB SIMULINK
a flexible environment for technical computing. The MATLAB open architecture makes
it easy to use in conjunction with its companion products to explore data, generate
algorithms, and create custom tools. MATLAB includes tools for: data acquisition, data
analysis and exploration, visualization and image processing, algorithm prototyping and
The toolboxes used in the present work were: the signal processing toolbox, the
Before starting writing the program codes in MATLAB language and perform the
simulation in SIMULINK, an understanding of the general purpose and steps that must be
followed was needed. First, a flow chart of the road map for the entire simulation was
developed. Next a list of variables was generated. The SIMULINK simulation started by
generating the time domain data for the pristine and damaged structures, followed by the
93
frequency domain analysis and finally the application of statistical methods to interpret
Time Frequency
Domain Domain 2
Input Simulation Analysis Statistical (m,s )r Fisher
Variables Analysis Criterion
(m,s2)pristine
Damage Level
r = 0, ,1
For each crack size the simulation was run 1000 times, each time with a different
randomly selected initial phase, j. The phase selected in each of the iteration was
randomly selected in the interval [0:2p] by a random white nose generator. In this way a
The frequency domain analysis consisted of two methods: the first method was
the short time Fourier transform (STFT), the second method was the Wigner-Ville
distribution (WVD).
Statistical analysis was used to compare the results by obtained by using the two
methods, STFT and WVD. The statistical method used for classification was Fishers
94
7.3.1.2 MATLAB SIMULINK Simulated Time Domain Data
The input variables for the simulation are presented in Table 7.1.
The simulation program was based on Equation (7.1). A screen capture of the
95
Figure 7.6 shows that the simulation starts with the generation of time domain
signal. The signal generated is a constant amplitude sine wave. The frequency of the
signal changes as the crack progresses. The signal is then split in two; one branch goes to
the workspace and is saved as the time domain data for the undamaged case, and the
other branch enters the block diagram where the crack is present, hence the time domain
The natural frequency of the object without crack is fn = 20 Hz (wn=40p rad/s) and
the damping coefficients for stretching and compression are zsws =zc wc = 10 rad/s. The
relative crack size values investigated here are r = 0; 0.1; 0.4; 0.6
The time domain data for two cases, pristine (r = 0), and damaged (r = 0.6) is
(a) (b)
Figure 7.7 Time domain data: (a) pristine case (r=0) and (b) damaged case (r=0.6)
For the frequency domain analysis, the transitory part of the time domain signal
was not taken into consideration. After the transitory part, the data was stored, and further
96
7.3.1.3 MATLAB SIMULINK Graphical User Interface
A graphical user interface (GUI) was created to facilitate data entering and
interaction with the different MATLAB files created for time domain, frequency domain,
The GUI was designed as a window where the user must enter the information
need by the program to run. It has also press buttons that allow the user to select the type
the type of plots of interest. The graphical user interface is shown in Figure 7.8
Figure 7.8 Graphical User Interface (GUI) for the MATLAB program (Gelman et al. 2001)
97
7.4 RESULTS
In the present work the simulation was carried out for relative crack size, r, starting
from zero (pristine case) and up to 0.6. In the next sections the results will be presented
for four cases: pristine case (r = 0), small crack (r = 0.1), medium crack (r = 0.4), and
The main frequency of the undamaged object is the same as the natural frequency
of the pristine object, fpristine = 20 Hz. The experimental results, Figure 7.9, show that
both, Wigner-Ville distribution (WVD) and short time Fourier transform (STFT) predict
(a) (b)
Figure 7.9 WVD and STFT comparison for the pristine case, r = 0
98
7.4.2 CASE 2: SMALL CRACK (r = 0.1)
As the crack grows, the main frequency is changing. The calculated frequency for
the relative crack size of 0.1 is f0.1 = 19.5 Hz. The simulation results, Figure 7.10a and b,
show that in both spectrums (WVD and STFT) there is a small change in the peak of the
Additional features start to appear in the frequency spectrum. For the Wigner-
Ville distribution spectrum, the half-tone interference is present (f1/2 @ 10 Hz). For short
time Fourier transform, the higher harmonic effects start to be noticeable. The first
(a) (b)
Figure 7.10 WVD and STFT comparison for small damage case, r = 0.1
99
7.4.3 CASE 3: MEDIUM CRACK (r = 0.4)
At higher relative crack size, r = 0.4, the calculated main frequency is f0.4 = 17.5
Hz. The experimental results, Figure 7.11, show the change in the main frequency for
In the WVD spectrum, Figure 7.11a, the 1st and 2nd half-tone interferences are present at
f1/2 @ 9 Hz, f3/2 @ 26 Hz. The amplitude of the features is also more prominent. In the
STFT spectrum, Figure 7.11b, the 1st and 2nd higher harmonics of the main frequency are
The only difference between the two spectra is that, while in the STFT spectrum
the 2nd higher harmonic of the main frequency is well delimited, in the WVD spectrum
the 2nd higher harmonic of the main frequency, f2, does not appear. Instead we notice the
presence of the 2nd half-tone interference, f3. This issue will be addressed later in the
discussion section.
(a) (b)
100
7.4.4 CASE 4: LARGE CRACK (r = 0.6)
For the last case, when the relative crack size reach 0.6, the calculated main
frequency is f0.6 = 15.5 Hz. The results show the decrease in the main frequency of the
Additional features are well represented in the frequency spectrum. Figure 7.12a
shows the additional features present in the WVD spectrum. The 1st and 2nd half-tone
interferences,(f1/2 @ 9 Hz, f3/2 @ 26 Hz) as well as the 1st higher harmonic of the main
frequency (f2 @ 35 Hz) appear in the power spectrum. Figure 7.12b shows additional
features present in the STFT spectrum. As anticipated, the 1st and 2nd higher harmonics of
the main frequency appear (f2 @ 35 Hz, f3 @ 53 Hz) in the power spectral density. It is also
noticeable that the amplitude of the harmonics as has grown considerable in comparison
with the previous cases that had a smaller relative crack size.
(a) (b)
Figure 7.12 WVD and STFT comparison for the damaged case, r = 0.6
101
7.5 DIAGNOSTIC EFFECTIVENESS
The Fisher criterion, described in section 3.7, was used to compare the
effectiveness of the two methods investigated in the present work. The Fisher criterion
was calculated for pairs of classes representing the frequency of the main harmonic for
the pristine case and the frequency of the main harmonic for various damage cases. The
change in the criterion as the damage progresses was examined. Table 7.2 presents the
values of the Fisher criterion for the two methods, short time Fourier transform (STFT)
Table 7.2 Fisher criterion calculated for STFT and WVD methods and the feature comparison.
Each of the the damaged cases were compared with the pristine case
Method Fisher criterion value
Pristine vs. Pristine vs. Pristine vs.
0.1 relative crack 0.4 relative crack 0.6 relative crack
STFT 0.001 1.476 4.119
WVD 0.666 3.297 43.367
The results of Table 7.2 are plotted in Figure 7.13. It can be noticed that the damage
detection capabilities of WVD algorithm are definitely better than those of STFT
102
STFT WVD
50
45
40
35
Fisher criterion 30
25
20
15
10
5
0
0 - 0.1 0 - 0.4 0 -0.6
Relative crack size
Figure 7.13 Damage detection effectiveness of the two methods, STFT and WVD, based on the
Fisher criterion
7.6 DISCUSSION
investigated using two signal-processing techniques: short time Fourier transform (STFT)
interest. Solving the equations of motion for a one degree of freedom oscillator generated
the time domain signal. The solution was obtained for four crack sizes: r = 0 (pristine), r
= 0.1 (small damage), r = 0.4 (medium damage), and r = 0.6 (large damage).
After the time domain signal was obtained, frequency domain analysis techniques
were used to investigate the change in frequency as a result of the crack growth. The
feature of interest was the amplitude of the main frequency as well as shifting of the other
103
The results of this investigation showed that, as the crack progresses, the
frequency spectra changes. According to the non-linear analysis, when damage is present,
the frequency spectra will change and higher harmonics and subharmonics of the main
frequency will appear. Our results clearly indicate the changes due to the presence of the
crack. The appearance of higher harmonics as the crack grows is well indicated by the
STFT. The 1st and 2nd harmonic of the main frequency (f2 @ 35 Hz, f3 @ 53 Hz) could be
easily identified on the power spectra for r = 0.4 and r = 0.6 (Figure 7.12b).
While the STFT method can detect well the higher harmonics of the main
frequency, the WVD method is less sensitive to this feature. In WVD power spectrum,
the 1st higher harmonic of the main frequency appears only for large crack size (relative
interference terms (cross-terms), are present in the frequency spectra of the WVD but do
not appear in the STFT. As the crack grows, the cross-terms are more visible and their
amplitudes are larger. This phenomenon led us to the idea of using the cross-terms as an
indicator of crack propagation. This idea is not investigate any further in the present
To classify the outcome of the two methods, a statistical tool was needed. The
Fisher linear discriminant method (Fisher criterion) was applied to classify the outcome
of the two damage detection techniques. The spectral features used in this classification
were chosen to be the change in amplitude of the main frequency for the pristine cases (r
= 0) and for the damaged case (r = 0.1), (r = 0.4), (r = 0.6). The main frequencies for the
damaged cases where compared one by one with the main frequency for the pristine case,
and the Fisher criterion was calculated for each case. This method was applied to the data
104
obtained using STFT and WVD. The results were plotted on the same graph and
Figure 7.13 shows that, for small cracks the detection of damage in the structure is
done poorly by both methods (STFT and WVD). As the crack size increases, the
detection of the damage becomes better, as indicated by the increase in the Fisher
criterion values. As expected, the best damage detection is achieved when the crack size
is large as shown for r = 0.6 in Figure 7.13. For small and medium crack sizes (r = 0.1
and r = 0.4) the two methods are comparable. For medium crack size (r = 0.4), the Fisher
criterion for STFT and WVD being FSTFT 0.4 = 1.5 and FWVD 0.4 = 3.3. For larger crack size
(r = 0.6), the WVD prediction is much better than the STFT, FSTFT 0.6 = 4.2 and FWVD 0.6 =
43.4. These results show that the overall the damage detection sensitivity using the WVD
In conclusion, both WVD and STFT methods predict the presence of the damage.
The STFT method gives a better indication of the higher harmonics of the main
frequency due to the crack growth On the other hand, WVD method gives stronger
sensitivity to the crack presence based on the changes in the main frequency. This latter
observation has also been substantiated by statistical analysis using Fisher criterion.
The method to be used for damage detection is entirely up to the person who is
performing the signal analysis. Its choice must be based on the application type, and on
the experience of the operator one method or the other can be used. The short time
Fourier transform method is a relative known method and is widely use in machinery
method and the full advantages it can offer are yet to be discovered. The present work is
105
only a small step taken on the long road, and further research needs to be conducted to
106
8 CONCLUSIONS
This thesis has presented vibration-based methods used for damage detection and
health monitoring of mechanical systems. The review of the state of the art revealed that,
in the past years, damage detection and health and usage monitoring have been proactive
topics in the researchers community. Extensive work has been recently done to study
linear and non-linear vibration phenomena, using various signal processing techniques.
Although the traditional methods, i.e. fast Fourier transform and short time Fourier
transform are still used, the focus of recent research is shifting towards more advanced
methods and signal processing techniques such as wavelet transform and Wigner-Ville
distribution.
vibration of helicopters was the focus of the present work. An understanding of the rotor
track and balance, the adjustments necessary to smooth the helicopter main and tail
rotors, and the procedures needed to be followed to achieve this goal have been
presented. Using VMEP neural network based algorithm for rotor smoothing, the
vibration levels of the helicopter are reduced by performing fewer adjustments and hence
less number of maintenance test flights. Maintaining a low vibration level increases the
operational time of helicopters, increases the life time of components, and reduces the
107
The health and usage monitoring aspect has also been presented in this thesis.
commercial helicopters with health and monitoring systems (HUMS) capable to provide
information about the health and the remaining useful life of major components. A brief
description of the programs that are currently in progress was given, with emphasis on
Army and South Carolina National Guard. The main contractor is Intelligent Automation
Corporation (IAC), and University of South Carolina is responsible with the cost benefit
analysis of the VMEP project. A description of the purpose of the project as well as the
software and hardware configuration was given. The vibration data collection, the
transfer of the data from the helicopter to the ground based station and further to the USC
A simulation for a cracked component was designed and carried out using
the structure for axial loading was used. The crack was simulated as a change in stiffness.
The signal processing techniques used for damage detection were short time Fourier
transform and Wigner-Ville distribution. The reason for using joint time-frequency
methods was the non-linearity of the structure response. The crack presence in the
structure modifies the effective stiffness and hence the main frequency. The relative
crack size considered in this study was r = 0 (pristine case), r = 0.1 (small crack), r = 0.4
(medium crack), and r = 0.6 (large crack). The natural frequency of the pristine object
was fn = 20 Hz. The results showed that both methods, short time Fourier transform and
Wigner-Ville distribution, predicted the presence and progress of damage. The feature
108
used to predict damage were, the change in the amplitude of the main frequency and
apparition of higher harmonics of the main frequency in the frequency spectrum. The
STFT method gives a better indication of the higher harmonics of the main frequency due
to the crack growth. On the other hand, WVD method gives stronger sensitivity to the
To quantify the effectiveness of the outcome of the two methods, the Fisher
criterion was used. The results showed that the Fisher criterion values for the Wigner-
Ville distribution method were slightly higher than those for the short time Fourier for
small and medium crack size (FSTFT 0.4 = 1.5 and FWVD 0.4 = 3.3), and there was a
significant difference between the values for large crack size (FSTFT 0.6 = 4.2 and FWVD 0.6
= 43.4). This is an indication that the Wigner-Ville distribution method predictions are
more accurate than the predictions of the short time Fourier method. The Fisher criterion
was applied to the vibration data representing the changes in the amplitude of the main
frequency.
the cross-terms apparition. For small crack size and medium crack size subharmonics
generated by the cross-terms are present in the frequency spectrum. Apparition of the
cross-terms in the frequency spectrum can be related to the apparition of damage in the
structure.
Further work needs to be done in order to better understand the outcome of the
Wigner-Ville distribution and to relate the apparition of cross-terms and the presence of
damage.
109
BIBLIOGRAPHY
Bouraou, N., Gelman, L. Theoretical bases of the free oscillation method for acoustical
University, 1997.
Cho, S.B. Neural network classifiers for recognizing totally unconstrained handwritten
numerals, IEEE Transactions on Neural Network, Vol. 8, No. 1, pp. 43-53, 1997
Essaway, M.A.; Diwakar, S.; Zein-Sabatto, S.; Garga, A.K., Fault Diagnosis of
Practices and Trends in Mechanical Failure Prevention, edited by H.C. Pussy and
Fraser, K.F., and King, C.N., Helicopter gearbox condition monitoring for the
Gelman, L., Gorpinich, S. Non-linear vibroacoustical free oscillation method for crack
detection and evaluation, Mechanical Systems and Signal Processing, 14(3) (2000)
343.
110
Gelman, L., Zinkovsky, Y., Petrunin, I. Spectral analysis of nonstationary narrowband
Giurgiutiu, V., Zagrai, A.N., Bao, J.J. Damage Identification in Aging Aircraft
Grabill, P., Berry, J., Grant, L., Porter, J. Automated helicopter vibration diagnostics for
the US Army and National Guard, Proceedings of the 57th Annual Forum of the
Haas, D.J., Baker, T., Spracklen, D., Schafer, C.G. Joint Advanced Health and Usage
Hess, R., Duke, A., Kogut, D. The IMD HUMS as a tool for rotorcraft health
iLearn Interactive (2000). Mobius iLearn Interactive, Practical and theoretical training
287-297.
111
Larder, B., Azzam, H., Trammel, C., Vossler, G. Smith Industries HUMS: Changing the
the 53rd Annual Forum of the American Helicopter Society, Virginia Beach, VA, Apr
55th Annual Forum of the American Helicopter Society, Montreal, Quebec, May 25-
May 27 1999.
McFadden, P.D., Smith, J.D., "A signal processing technique for detecting local defects
in gear from the signal average of the vibration", Proc. Institute of Mechanical
Monsena, P.T., Dzwonczyk, M., "Analog neural network-based helicopter gearbox health
Newland, D.E., Wavelet Analysis of Vibration, Part I: Theory, Journal of Vibration and
112
Newland, D.E., Wavelet Analysis of Vibration, Part II: Wavlet Maps, Journal of
Polyshchuck, V.V., Choy, F.K., Braun, M.J., New gear-fault-detection parameter by use
of joint time-frequency distribution Journal of Propulsion and Power, Vol. 16, No.
2, 2000.
Qian, S., Chen, D., (1996) Joint time-frequency analysis: methods and applications
Roe, J.D., Astridge, D.G. The health and usage monitoring of helicopters the next
Samuel, P.D., Pines, D.J., Lewicki, D.G. Fault Detection in the OH-58A Main
Transmission Using the Wavelet Transform, Proceedings of the 52nd meeting of the
Shin, Y.S. and Jeon, J.J., "Pseudo Wigner-Ville Time-Frequency Distribution and its
Vol. 1(Inauguration Volume), pp. 65-76, Nov. 1993, John Wiley & Sons, Inc.
Trammel, C., Vossler, G., Feldmann, M. UK Ministry of Defence generic health and
113