Mechanical Systems and Signal Processing 85 (2017) 278–295

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Frequency domain averaging based experimental evaluation of

gear fault without tachometer for fluctuating speed conditions
Vikas Sharma n, Anand Parey
Discipline of Mechanical Engineering, Indian Institute of Technology Indore, India

a r t i c l e in f o abstract

Article history: In the purview of fluctuating speeds, gear fault diagnosis is challenging due to dynamic
Received 23 April 2016 behavior of forces. Various industrial applications employing gearbox which operate un-
Received in revised form der fluctuating speed conditions. For diagnostics of a gearbox, various vibrations based
3 August 2016
signal processing techniques viz FFT, time synchronous averaging and time-frequency
Accepted 11 August 2016
based wavelet transform, etc. are majorly employed. Most of the time, theories about data
Available online 20 August 2016
or computational complexity limits the use of these methods. In order to perform fault
Keywords: diagnosis of a gearbox for fluctuating speeds, frequency domain averaging (FDA) of in-
Fault detection trinsic mode functions (IMFs) after their dynamic time warping (DTW) has been done in
Gear crack
this paper. This will not only attenuate the effect of fluctuating speeds but will also extract
Fluctuating speeds
the weak fault feature those masked in vibration signal. Experimentally signals were ac-
Dynamic time warping
Kurtosis quired from Drivetrain Diagnostic Simulator for different gear health conditions i.e.,
healthy pinion, pinion with tooth crack, chipped tooth and missing tooth and were ana-
lyzed for the different fluctuating profiles of speed. Kurtosis was calculated for warped
IMFs before DTW and after DTW of the acquired vibration signals. Later on, the application
of FDA highlights the fault frequencies present in the FFT of faulty gears. The result sug-
gests that proposed approach is more effective towards the fault diagnosing with fluc-
tuating speed.
& 2016 Elsevier Ltd. All rights reserved.

1. Introduction

Devastation of mass and material in industries due to post diagnostics techniques such as root cause failure analysis,
fracture analysis, etc. has lead towards the emergence of early diagnosis through vibration analysis. Thus, vibration analysis
has proved to be an efficient technique in the detection of faulty gears of various mechanical systems employed in different
industrial sectors [1–3]. But, in the real-time conditions, fluctuations in the speed at which the gearbox operates which
results in smeared frequency spectrum and as a consequence, it becomes difficult to locate the fault frequencies. Therefore,
such fluctuating condition has made the existing techniques such as, Fast Fourier Transform (FFT), Time Synchronous
Averaging (TSA), cepstrum, etc., ineffective [4,5]. Various experiments on spectral analysis [6–8] suggest to be incompetent
alone, as they produce erroneous results and are limited to fluctuating speed and load environment. All of the above
methods have made some achievements to health monitoring and fault diagnosis. However, each of these methods has their
own limitation of fluctuating speed environment.

n
Corresponding author.
E-mail addresses: phd1301103010@iiti.ac.in (V. Sharma), anandp@iiti.ac.in (A. Parey).

http://dx.doi.org/10.1016/j.ymssp.2016.08.015
0888-3270/& 2016 Elsevier Ltd. All rights reserved.
 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295 279

In real life, the wind velocity and the output load doesn’t remain constant, these parameters makes the wind turbine to
face fluctuations in speed. Under fluctuating speed conditions, it becomes difficult to diagnose the gear fault. Further, to
comment about the severity of fault and type of fault becomes challenging. Therefore, to diagnose faulty gear box under
fluctuating speed, an effective signal processing technique, should possess the following characteristics:

(a) Capability to process non-stationary, nonlinear signals: analysis based instinctively on data;
(b) Highly robust method;
(c) Should be fast and neither computationally complex nor expensive in order to perform on-line monitoring.

Method for fault diagnosis based on time-frequency resolution i.e. Empirical Mode Decomposition (EMD) which follows
Hilbert Transform (HT) is one of the efficient techniques. EMD with an aim of fault diagnosis is useful in the extraction of the
Intrinsic Mode Functions (IMFs), or mono-component functions that comprises the original signal. Various studies [9–11]
have highlighted that number of components (high-frequency component, low-frequency components, and noise) present
in a gear vibrational signal have a strong influence on the fault diagnosis capabilities. These multi-components mask the
fault features of the gear vibration signals. Liu et al. [12] used B-spline EMD followed by spectrum of HT to diagnose gearbox
faults and found that it can enhance the crack excited transients in a better way as compared to wavelet transform. Qin et al.
[13] did an investigation on iterated Hilbert Transform and enforced it to mechanical fault diagnosis containing multi-
component AM-FM signals. In 2009, Yang improved the conventional EMD method and proposed a new Hilbert spectrum
method [14].
The key component of EMD is IMF; the selection of the IMF is based on operation skills thus, there are frequent chances
of getting erroneous results are high. Furthermore, hypothetical data and computational complexity limits the use of these
methods. However, it has better representation in time-frequency domain and capability of separating the signal into
different components based on oscillations of signal [15]. Additionally, frequency domain averaging is found very effective
by decomposing and attenuating the noise components from the signal [16], and can prove to be benefiting as no key phasor
signal or reference signal is required. Thus the use of tachometer is non-essential. Therefore, utilizing IMF for FDA after DTW
can prove to be a novel approach towards, fault diagnosis under fluctuating speed conditions.
The scheme of this approach has been illustrated in Fig. 1. The organization of this paper is as follows: the main steps of
the EMD, DTW and FDA are discussed in Section 2 respectively. Section 3 illustrates a simulation study conducted to verify
the proposed approach of FDA by diagnosing the fault phenomenon generated in a gearbox by a localized gear defect under
fluctuating speed conditions. Subsequently, the performance evaluation of proposed method is also expressed. Section 4
illustrates the description of experimental test-setup. The experimental findings are presented and followed by the dis-
cussion in Section 5. Conclusions about the effectiveness and the diagnostics capability of the method are reported in
Section 6.

Fig. 1. Scheme of fault diagnosis method.

280 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295

2. Proposed method for gear fault diagnosis

2.1. Empirical mode decomposition

EMD is purely an adaptive decomposition. No assumption is required about signal periodicity, linearity or about sta-
tionary behavior. The principle behind EMD technique is that, a signal x(t ) is decomposed automatically into IMFs which are
a set of the band-limited functions. Only the main steps of the method are briefly presented here: for a complete summary
[15] can be referred. Each IMF should satisfy two conditions:

(1) In the complete signal, the number of extrema (extreme values) and zero-crossings must either equal or at most differ
by one.
(2) At any point, the mean value of the envelope defined by local maxima and the local minima is zero.

Therefore in accordance to the definition of EMD, a non-stationary signal x(t ) can be expressed as a linear sum of IMFs
and the residual components. Mathematically,
N
x( t ) = ∑ IMFn( t ) + rn( t )
n= 1 (1)
th
Where, IMFn(t ) symbolizes the n intrinsic mode function and rn(t ) the residual component. The EMD algorithm can be
further summarized as follows:

(1) Extraction of all extrema of x(t ).

(2) Interpolation between minima for obtaining two envelopes L min(t ) and L max(t ).
(3) Compute the mean: a(t ) = (L max(t )+L max(t )) /2.
(4) Extract the detail h( t ) = x( t ) − a(t ).
(5) Test if h(t ) is an IMF using below conditions:
a) If it yes, procedure from the step (1) about the residual signal f ( t ) = x( t ) − h(t ) is repeated.
b) If not, then replace x(t ) with h(t ) and repeat the procedure from step (1).

EMD decomposes a multicomponent signal into mono-component signals which are called as IMFs. The IMF1( t ) contains
highest frequency component and lower frequency components will appear in subsequent IMFs in such a way that the
lowest frequency component lies in IMFN( t ). These frequency components changes with the variation of signal x(t ). On the
other hand, residue rn( t ) shows the central tendency of the signal.

2.2. Dynamic time warping ( DTW)

DTW is an approach to achieve an optimal path among two time domain series. To find out the resemblance in between

A Awarped
B Bwarped
Amplitude
Amplitude

Sample Sample
Fig. 2. Alignment of two time series based on DTW: (a) two similar sequences, with different phase and (b) alignment by DTW [17].
 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295 281

n
Warp path

Time series B

j
1
1 i m

Time series A
Fig. 3. Two dimensional DTW matrix and optimal warping path by DTW [17].

both the time series, one of the time series need to be warped on other by stretching and shrinking [17]. This type of
shrinking and stretching could be non-uniform in nature. The concept of DTW could be understood by considering two time
series A and B. Let the time series A be of length m and time series B of length n, such that A ¼a1, a2, …, ai, …, am, and B ¼b1,
b2, …, bj, …, bn. With this a warp path W of length p is constructed as W¼ w1, w2, …, wk, …, wp. The kth element of the warp
path is given by wk ¼(i,j); and the indexes i and j are from time series A and B. Also a warp path must satisfy the following
three conditions:

1) The warp path must start at the beginning of both the time series i.e., w1 ¼(a1, b1).
2) The warp path must finish at the end of both the time series i.e., wp ¼ (am, bn).
3) In between the start and end of the warp path, every index in both of the time series must be used i.e., wk ¼(i, j) and
wk þ 1 ¼(i′, j′) ⇒ i′ ϵ (i, iþ1), j′ ϵ (j, jþ 1).

This way the best possible match in between two-time series A and B with least distance path can be established and is
given by means of recursive formula as mentioned below [18]:

DTWp( A, B) = p
w ( i, j ) (2)

Where, w( i, j ) is the cumulative distance expressed as:

p
w ( i, j ) = ai−bi {
+ min w ( i − 1, j − 1), w ( i − 1, j ), w ( i, j − 1) } (3)

Fig. 2 exemplifies the perception of DTW by considering time series A and B (one as test series and other as reference
series), in which initially time series A was shrunk and then stretched to fit on time series B. Fig. 3 shows the best possible
warp path on the basis of the alignment of two time-series signals (Fig. 2). DTW computes the distance of all points with
equal effect of each point irrespective of the phase.

2.3. Frequency domain averaging (FDA)

In some instances, the tachometer/key phasor signal is not easily accessible to trigger time synchronous averaging or
may be contaminated by noise. The key aspect of this approach is that it does not use any trigger signal even for performing
fault diagnosis at fluctuating grounds [19]. This method is used to solve such vital problems, therefore FDA is applied. Since
time and frequency domain both have linear properties and are related to Fourier transform, it implies that changes ob-
served in one domain will be reflected in other domain. Therefore, synchronous averaging in time domain must have
equivalent frequency domain function. Since FFT and iFFT are linear functions, then

F [ a] + ⎡⎣ b⎤⎦ = F ( [ a]) + F ⎡⎣ b⎤⎦

( ) ( ) (4)

F([a]) and F([b]) are the discrete time Fourier transforms of [a] and [b] time domain data blocks. It is clear that summing
in the time domain is equivalent to summing in the frequency domain. FDA was implemented in the following manner:

(1) Corrected/warped IMFs were collected.

(2) FFT was performed.
(3) The frequency domain data block was time translated so that the trigger frequency could be set at zero phase.
(4) The frequency domain data block was averaged into the result block.

The benefits of performing FDA are: (1) A trigger channel is not required so the cost of trigger acquisition probe is saved.
(2) Issues of diagnosis of intermediate/idle gears or shafts can still be performed.
282 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295

Amplitude
0

-5
0 1 2 3 4 5 6 7
Time (s)
5

Amplitude
0

-5
0 1 2 3 4 5 6 7
Time (s)
5

Amplitude
0

-5
0 1 2 3 4 5 6 7
Time (s)
5
Amplitude

0
-5
0 1 2 3 4 5 6 7
Time (s)

Fig. 4. The simulated signal: (a) the component x1(t ) , (b) the component x2(t ) , (c) the random noise part n(t ) , (d) the complete signal x(t ).

3. Simulation of a gear fault

A simulation study was conducted on a gear synthetic signal which consisted of both AM-FM signal [20–23] along with a
periodic double sided asymmetric transient to illustrate the effectiveness of fault diagnosing capability of the proposed
approach.

3.1. Simulation study

To consider the characteristics of the gearbox in operation and to constitute the vibration signal of gearboxes with a
localized fault in one of the gears, a synthetic signal x(t ) containing additive random noise noise(t ) can be written as:
x(t ) = x1(t ) + x2(t ) + noise(t ) (5)

M
x1(t ) = ∑ X m[1 + a m(t )] cos [2πmZfr t + Φm + bm(t )]
m=0 (6)

Where, x1(t ) is the AM-FM harmonic signal, fr is the gear rotating frequency and Z is the number of teeth on gear. Therefore,
gear mesh frequency (GMF) can be calculated by Zfr , m is the number associated with harmonic of GMF, Xm is the amplitude
of mth harmonic of GMF, Φm is the original phase and am(t ) and bm(t ) are the AM functions.
Ms
where, a m(t ) = ∑ A mn cos(2πfr t + αmn)
m=1 (7)
Ms
and bm(t ) = ∑ Bmn cos(2πfr t + βmn)
m=1 (8)
th th
Amn and Bmn are the amplitudes, αmn and βmn are the phases of n sidebands of the amplitude around m meshing
harmonic. Bonnardot et al. [4] suggested that fr can be substituted by fr (i ) under the condition of fluctuating speed,
therefore Eq. (6) can be rewritten as follows:
M
x1(t ) = ∑ X m[1 + a m(t )] cos [2πmZfr (i)t + Φm + bm(t )]
m=0 (9)
Amplitude [mm/s2]

2
1.75
1.5
1.25
1
0
100 200 300 400 500 600 700 800 900 1000
Frequency (Hz)

Fig. 5. FFT of signal at fluctuating speed conditions.

 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295 283

K = 3.182111 K = 3.398308
IMF1 [mm/s2] 4 3

IMF2 [mm/s2]
2 1.5
0 0
-2 -1.5
-4 -3
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time (samples) x 10
4 Time (samples) 4
x 10

K = 3.036681 K = 3.240067
2 1

IMF4 [mm/s2]
IMF3 [mm/s2]

1 0.5
0 0
-1 -0.5
-2 -1
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time (samples) 4 Time (samples) x 10
4
x 10
2 K = 2.248817
Residue [mm/s2]

1
0
-1
-2
-3
0 0.5 1 1.5 2 2.5 3
Time (samples) x 10
4

Fig. 6. Signal decomposed into IMFs before application of DTW.

Another component x2(t ) which represents the localized tooth fault phenomenon can be modeled as a double-side
asymmetric transient as follows:

x2(t ) = ∑ h(t − T0 − kT )
k (10)

Where, h(t ) is single transient and keeping time index T0 as 0.05 s and time period T as 0.1 s for time center (kT + T0) of kth
transient. To perform simulation for a single harmonic of a gear, the values of m ¼1, Ms ¼1 were considered. The values of T0
and T depend on the rotational speed of gears and are the part of transient impulse. For the present simulation case, this
transient impulse that will appear for less than a second, more precisely for 0.1 s as the gear was considered running at
10 Hz and one impulse will appear for one revolution. Therefore, the values of T0 and T were taken as 0.05 and 0.1 re-
spectively. Now, this h(t ) is expressed as follows:

⎧ −σ / 1 − σ l2 (2πf0 t )2
⎪e l cos(2πf0 t ), t ≤ 0
h(t ) = ⎨
⎪ e−σr / 1 − σr2 (2πf0 t )2
cos(2πf0 t ), t > 0
⎩ (11)

Where f0 is the fault frequency generated due to impulse of localized fault. For different meshing frequencies there will be
different fault frequencies. For simulation, value of f0 is assumed as 520 Hz with left and right damping ratio be 0.02 and
0.01 respectively. For simulation, the input frequency was taken as 10 Hz which fluctuated sinusoidally obeying an equation
fr = 10 + (sin(2π10t )); number of teeth on gear was taken as 32. The sampling frequency kept for the simulation was
4.096 kHz. The values of other variables for simulation purpose were assumed as Xm = 5, Amn = Bmn = 0.05, αmn = βmn = 0.
The signals developed from these equations have been displayed in different sections a, b, c of Fig. 4. In Fig. 4(a), a pure AM-
PM signal was observed under fluctuating speed conditions. Fig. 4(b) shows the time response impulses generated due to
localized fault. Fig. 4(c) represents the additive random noise of surrounding and Fig. 4(d) shows the complete signal x(t ) as
given in Eq. (5). While performing simulation, noise was kept high intentionally to mask the fault components presents in
the signals and to evaluate the proposed approach. Conventionally, some denoising algorithms would be required to at-
tenuate the noise content, but the proposed approach was tested without using any denoising algorithm (Fig. 5).
284 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295

K = 4.038380 K = 5.150077
IMF1 [mm/s ] 4 3

IMF2 [mm/s ]
2

2
2 1.5
0 0
-2 -1.5
-4 -3
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time (samples) 4
x 10 Time (samples) x 10
4

2 K = 4.761897 K = 5.023855
1.2

IMF4 [mm/s ]
IMF3 [mm/s ]

2
2

1 0.6
0 0
-1 -0.6
-2 -1.2
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time (samples) x 10
4 Time (samples) x 10
4

Fig. 7. IMFs after application of DTW.

Table 1
Kurtosis of IMFs before and after DTW.

Kurtosis Before DTW After DTW

IMF1 3.1821 4.0383

IMF2 3.3983 5.1500
IMF3 3.0366 4.7618
IMF4 3.2400 5.0238

2.0 2.0
IMF1 IMF2
1.5 Fault frequencies 1.5
Amplitude
Amplitude

Fault frequencies
1.0 1.0

0.5 0.5

0 0
0 500 1000 1500 2000 0 500 1000 1500 2000
Frequency (Hz) Frequency (Hz)

2.0 2.0
IMF3 IMF4
1.5 1.5
Amplitude

Amplitude

Fault frequencies
Fault frequencies
1.0 1.0

0.5 0.5

0 0
0 500 1000 1500 2000 0 500 1000 1500 2000
Frequency (Hz) Frequency (Hz)
Fig. 8. FFTs of warped IMFs.

3.2. Performance evaluation

To study the fluctuating profile of operating speed, a synthetic signal which imitates the behavior of gearbox vibration
was used for simulation purposes. Sequential use of EMD, DTW and FDA illustrates the concept and performance of pro-
posed approach. To further analyze the effectiveness of the proposed approach in terms of fault diagnosing, kurtosis value
was calculated for IMFs both before DTW and after DTW.
First step deals with the use of EMD and the extraction of IMFs. The synthetic vibration signal was decomposed into five
IMFs where fifth IMF was considered to be a residue. This decomposition was performed to get an idea about the different
 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295 285

1.5 X: 318 X: 521

Amplitude
Y: 1.176 Y: 1.137
Fault frequencies
1

0.5

0
0 500 1000 1500 2000
Frequency (Hz)
Fig. 9. Averaged FFT.

Fig. 10. Experimental test rig.

oscillating components present in the signal. The decomposition technique of EMD is based on and deprived from signal,
predefined basis like basis of wavelet transform is not a requisite for the signal. High frequency phenomenon was present in
first IMF while the last one shows low frequency components of the signal due to fault. Above all, the IMFs derived from the
EMD usually have physical interpretation. So, the IMFs should highlight impulses or modulation because of fault in the
signal or for gearbox operating under fluctuating conditions. But in Fig. 6, the IMFs depicted the decomposition of signal

Zoomed view

Arrow showing the

Bearing direction of vibrations to
Accelerometer Wax
housing be measured
cable Accelerometer
Fig. 11. Accelerometer position.
286 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295

Fig. 12. (a) missing tooth gear in mesh with output gear; (b) healthy gear; (c) chipped tooth; (d) cracked tooth gear.

with fault however; the impulses were not clearly visible. This makes the IMFs to be tested once again statistically.
Therefore, a statistical condition indicator kurtosis (fourth order about the mean) was used on these IMFs to know more
about the transients or peaks appearing due to fault. The evaluation of kurtosis of IMFs marks the inefficient decomposition
by EMD of the faulty signal under speed fluctuating condition. Now, as mentioned in Section 2.2, the IMFs in Fig. 6 were
warped using the concept of DTW. For warping purposes, a healthy gearbox vibration signal under the same operating
conditions was considered as a reference signal.
The warped IMFs are shown in Fig. 7. To say that the warped IMFs are retaining the original information and there is no
loss of significant data regarding fault phenomenon, they must be reviewed statistically by some condition indicator. When
these warped IMFs are tested statistically using kurtosis, they were found to be improved. Also the peaks were visible more
clearly. This shows that process of DTW has the capacity to attenuate the non-synchronous components or the irrelevant
components recorded during the signal acquisition. Table 1 shows the values of kurtosis calculated for IMFs for both before
and after the application of DTW.
To highlight the fault frequencies, the second step plays important role. To show the fault frequencies, these warped IMFs
were transformed into the frequency domain. Fig. 8, shows the FFTs of the warped IMFs and it is clearly visible that the fault
frequency more or less appears at same location in FFT i.e. close to 520 Hz, whereas the highest amplitude peak varies close
to gear mesh frequency, i.e. 320 Hz due to speed fluctuation. The reason behind this phenomenon is the fluctuation in speed,
which makes the gearmesh frequency to vary.
In the last stage of analysis, these FFTs were averaged, to get a concrete idea about the fault appearing phenomenon and
the fault frequencies. The principle of frequency domain averaging has explained in Section 2.3. This concludes the simu-
lation study which exhibits the capability of the gear fault diagnosing approach. The result of the simulation study strongly
supports the fault appearing in the gearbox. Figs. 7 and 8 clearly shows the important peaks of both fault frequency and
gearmesh frequency. The frequency domain averaging again clears the remaining noise present in the signal, even if some
components remain un-attenuated by DTW. On comparing the kurtosis value and initial and final FFTs (Fig. 5 and Fig. 9), it is
worth noting that the proposed approached of associating DTW and FDA to IMFs has shown superior fault diagnosing
capability than just decomposing signal into IMFs, under fluctuating speed.

Table 2
Gear specification.

Parameter Input gear Output gear

Number of teeth 32 80
Pressure Angle (deg) 20 20
Dimeteral pitch (per mm) 0.6299 0.6299
Module (mm) 1.5879 1.5879
Face width 12.5 12.5
 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295 287

Fig. 13. Profiles of fluctuating speed.

4. Experimental setup

4.1. Test rig

To examine the performance of the proposed approach in real time speed fluctuating conditions, experiments were
carried out on a simple motor-drive-brake type gear test rig (Drivetrain Diagnostic Simulator) as shown in Fig. 10. De-
tachable gears in the gearbox were used for such measurements. The rig consists of two deep groove ball bearings mounted
in casing which helped gear shaft to rotate inside the gearbox.
An AC motor (3 phase, 2.237 kW) was used to drive the test rig through a flexible coupling. The programmable control
panel of the AC motor can provide a continuously variable speed from 0 to 3000 rev/min, using Motor Control module of NV
Gate. To isolate the test gear from surroundings and to reduce vibrational transmission, vibration rubber pads were placed
beneath the motor and gearbox casing. The center distance between gearbox shafts is 89 ± 1mm, which can be used for
other parallel transmissions as well. A magnetic brake type loading arrangement was used to apply a maximum radial load
of 88 Nm (at the test gear) through Brake Control module of NV Gate. A uniaxial accelerometer was mounted on the outer
end of bearing case on the test gear shaft. Only one accelerometer was used and the vibrations were measured in the radial
direction. Fig. 11 displayed below, shows the direction of accelerometer to measure the vibrations.

4.2. Test gears

Spur gears were used as research object in this study with involute profile of medium carbon steel (AISI1045). Different
gear tooth faults i.e., gears with initial crack, advanced crack, missing and chipped tooth as shown in Fig. 12 were used to
examine the proposed approach under fluctuating speed conditions. The details of gears are listed in Table 2. The test gears
were mounted on the main shaft. Faults like crack in the tooth root of gears were introduced by wire electron discharge
machining as suggested by the authors in ref [24,25].

4.3. Measurement conditions

As the prime object of the proposed approach deals with fault diagnosis under fluctuating speed conditions. So, signal
acquisition for all the test gears were performed only at two loading conditions, i.e. no load and 40% load of maximum load
(i.e. 35.2 Nm) were considered. Along with the loading conditions, limited speed fluctuations (20% of mean speed) were
considered to evaluate the proposed approach. The following fluctuating profiles of speed were used.

(a) Sinusoidally fluctuating profile of speed

(b) Quadratically fluctuating profile of speed

The profiles of fluctuating speed are shown in Fig. 13, ref. [26–28] enlightens the different fluctuating profiles of speed
that has been observed in wind turbines. The speed was fluctuated around 10 Hz, i.e. 600 rpm and the vibration signals were
acquired with a sampling rate of 12.5 kS/s.
288 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295

Healthy gear Cracked toooth

1 1

0.5 0.5
Amplitude

Amplitude
0 0
-0.5 -0.5
-1 -1
0 0.55 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time (sampless) 4
x 10 Time (sam
mples) x 10
4

Chipped tooth 1 Missing toooth

1

0.5 0.5

Amplitude
Amplitude

0 0

-0.5 -0.5

-1 -1
0 0.55 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time (sampless) 4
x 10 Time (sam
mples) x 10
4

Signals under siinusoidal fluctuatin

ng speed conditionns
Healthy geear Cracked toooth
1 1

0.5 0.5
Am plitude

Am plitude
0 0

-0.5 -0.5

-1 -1
0 0.55 1 1.5 2 2.5 3 0 0
0.5 1 1.5 2 2.5 3
Time (samplees) x 100
4
Time (samples) x 10
1
4

Chipped tootth Missing toooth

1 1

0.5 0.5
Am plitude
Am plitude

0 0

-0.5 -0.5

-1 -1
0 0.5 1 1.5 2 2.5 3 0 0
0.5 1 1.5 2 2.5 3
Time (samplees) x 10
4 Time (sampples) x 10
4

Signals under quadratic fluctuatting speed conditioons

Fig. 14. Acquired vibration signals of different gears under different speed conditions.

Fig. 15. FFT of different gear vibration signal at constant speed of 10 Hz.
 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295 289

Fig. 16. FFTs of acquired vibration signal; (a) sinusoidal fluctuating profile of speed; (b) quadratic fluctuating profile of speed.

5. Results and discussion

For fault diagnosis of gearbox under fluctuating speed without tacho signal, the proposed scheme of FDA for FFTs of
corrected IMFs using DTW was applied on experimentally acquired vibration signals from a fixed axis gearbox test rig
shown in Fig. 10. Time domain vibration signals acquired at different fluctuating profiles of speed for a various gear health
are shown in Fig. 14. Quick glimpses of the effect of fluctuating speed are shown in Figs. 15 and 16. Gearmesh frequencies
and the fault appearing phenomenon can be easily noticed in FFT plot when speed remains constant as shown in Fig. 15,
whereas, the effect of fluctuating speed can be noticed by smeared FFTs in Fig. 16(a and b). The upcoming sections discusses
about the fault diagnosis mechanism for different fluctuating profiles of speed.

5.1. Behavior of signal acquired under fluctuating profiles of speed

Fig. 14(a) shows the vibration signal acquired under the sinusoidally varying speed for all the aforementioned gears.
Subsequently, Fig. 16(a) shows the FFT of the signals for different gears and it can be noticed that only with increasing fault
size, i.e., from cracked tooth to missing tooth, there was an amplitude gain, but to locate the frequencies of fault features and
the gearmesh frequency and their harmonics is difficult. A similar phenomenon of increase in amplitude with the fault
severity was observed for quadratic fluctuation profile of speed too. Fig. 14(b) and Fig. 16(b) shows the vibration signals and
corresponding FFTs respectively.

5.2. Decomposition of signals acquired under fluctuating profiles of speed

The signals acquired from gearbox for different gears under both sinusoidally and quadratically fluctuating profiles of
speed were decomposed into five IMFs using EMD. This decomposition was performed to segment the signal on the basis of
their oscillating frequency. The fifth IMF was the residue of the signal. Fig. 17(a) and (b) shows the IMFs of the vibration
signals of different gears acquired under varying speed conditions. To detect the variation in the amplitude of the vibration
signal kurtosis value was calculated. An increase in kurtosis value indicates a difference between the measured signal and
the reference signal to show the severity of the fault [29–31]. Therefore, along with this, kurtosis was calculated and FFTs
290 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295

Fig. 17. IMFs of vibration signals of different gears under different fluctuating profile of speed before DTW at 40% load. (a) Sinusoidal fluctuating profile of
speed. (b) Quadratic Speed Fluctuation.
 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295 291

Fig. 17. (continued)

were plotted for all the IMFs. The kurtosis value doesn’t exhibit any indication towards fault; similarly FFTs of IMFs doesn’t
show any fault components. With this, it can be summarized that the EMD shows insignificant decomposition of the ac-
quired vibration signal under fluctuating speed conditions. To deal with this issue, DTW of the IMFs was performed, by using
292 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295

Fig. 18. IMFs of different gears under different fluctuating profile of speed after DTW at 40% load. (a) Under sinusoidal fluctuating profile of speed.
(b) Under quadratic fluctuating profile of speed.

the healthy gear signal as a reference. Warping of the IMFs was performed as mentioned in Section 2.2. To satisfy the
condition of the warping, i.e., the continuity and the length of the signal must be same to meet the endpoints; a resampling
algorithm is required to restore the length. Fig. 18(a) and (b), shows the warped IMFs processed after DTW of different gears
under different fluctuating profile of speed. The modulation observed in the IMFs before and after DTW (Figs. 17 and 18)
exhibits the effect of fluctuating speed. The kurtosis calculated for these warped IMFs indicates significant increase in their
value, which in turn suggests the severity of the fault in the vibration signals. Table 3 lists the value of kurtosis of IMFs
before and after the application of DTW. It has been observed that the kurtosis ranges from 3.05 to 3.92 before DTW,
whereas after DTW it ranges from 3.07 to 6.59; DTW makes proper fault correlation with time waveform of IMFs. This result
in hike of kurtosis value and the physical interpretation of this increase of kurtosis indicate the fault severity or crack
propagation [30]. Thus in our case it implies that, increase in the values of kurtosis of warped IMFs expresses fault severity. It
 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295 293

Table 3
Value of Kurtosis before and after DTW at 40% load.

Kurtosis Before DTW After DTW

Healthy gear Cracked tooth Chipped tooth Missing tooth Healthy gear Cracked tooth Chipped tooth Missing tooth

Sinusoidal fluctuating profile of speed

IMF1 3.05 3.22 3.59 3.88 3.07 4.89 5.25 6.15
IMF2 3.15 3.12 3.39 3.73 3.17 4.05 5.58 6.24
IMF3 3.22 3.29 3.74 3.45 3.20 3.79 4.85 5.98
IMF4 3.19 3.54 3.44 3.60 3.18 4.98 5.67 6.19
Quadratic fluctuating profile of speed
IMF1 3.25 3.72 3.19 4.02 3.23 4.93 5.74 6.59
IMF2 3.35 3.17 3.69 3.87 3.35 3.86 4.97 5.95
IMF3 3.41 3.92 3.47 3.54 3.40 4.97 5.80 6.05
IMF4 3.21 3.45 3.49 3.17 3.26 3.88 4.87 5.91

Healthy tooth Cracked tooth

3 3 Gear mesh frequency
Gear mesh frequency Harmonics Fault frequencies
Amplitude

Amplitude
2 2

1 1

0 0
0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600
Frequency (Hz) Frequency (Hz)
Chipped tooth Missing tooth
3 3 increased sidebands in
Gear mesh frequency
Fault frequencies shaft rotation frequency Fault frequencies
Amplitude

Amplitude

2 2

1 1

0 0
0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600
Frequency (Hz) Frequency (Hz)

Fig. 19. FDA of FFTs of IMF at sinusoidal fluctuating profile of speed showing all faults.

Healthy gear Increased sideband of Cracked gear

3 3 gear mesh frequency
Shaft frequency Fault frequencies
Second harmonics
2 2
Amplitude
Amplitude

Gear mesh frequency

1 1

0 0
0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600
Frequency (Hz) Frequency (Hz)
Chipped tooth Missing tooth
3 3
Seond harmonics
increased sidebands of Fault frequencies
shaft rotation frequencies Fault frequencies Gear mesh frequency
Amplitude
Amplitude

2 2

1 1

0 0
0 200 400 600 800 1000 1200 1400 1600 0 200 400 600 800 1000 1200 1400 1600
Frequency (Hz) Frequency (Hz)

Fig. 20. FDA of FFTs of IMF at quadratic fluctuating profile of speed showing all faults.

can also be noticed from Table 3, that the kurtosis of IMFs calculated for healthy gear condition doesn’t contain any sig-
nificant statistical differences before and after DTW. Ideally, the value of kurtosis is 3 for a pure Gaussian distribution, and on
the basis of that it can be said that the value of kurtosis increases with the increase in fault [29–32]. For the fluctuating
294 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295

speed conditions the threshold value of kurtosis is considered as 3.5, if it increases from 3.5, then it can be interpreted as the
presence of fault in the gear. Thus, an inference can be drawn that processing of DTW optimally matches the measured IMFs
of faulty gears to reference IMF of healthy gear, which results in feature extraction using the proposed approach more
accurately.
Figs. 19 and 20, show the FDA signals of the FFTs of the warped IMFs. FDA shows the averaging in frequency domain, this
FDA will reduce the noise and irrelevant components present in the signal, which were trapped during vibration signal
acquisition. This FDA based FFT will show the, shaft rotation frequency, gear mesh frequency and their harmonics along with
the fault frequencies present in the signal. Since the speed fluctuates thus, the gear mesh frequency will appear according to
the presence of high amplitude frequencies, but in the range of speed rotating frequency. From both Figs. 19 and 20, the gear
mesh frequency was observed at 366 Hz, the peak highlighting input shaft was found to be around 11.5 Hz and fault fre-
quencies occurred in the range of 800–1000 Hz. Fault frequencies will be different for different faults but to indicate specific
fault frequency w.r.t. fault is difficult. However, the increase in amplitude and density of sidebands indicates fault phe-
nomenon. Also, with the increase of fault severity, the amplitude of side bands and the fault peaks were found increasing.
Under real life conditions, the type of fault will not be known only its FFT will be visible at the output. Increase in amplitude
and sideband density visible in FFT along with the increasing value of condition indicator, indicates the fault phenomenon.
In terms of gear fault diagnosis under fluctuating profile of speed, the effectiveness and capability are noteworthy on
comparing Fig. 15 to Fig. 19 and Fig. 20, as the proposed approach almost minimizes the phenomenon of smeared FFT. Also,
the fault frequencies are made visible clearly.

6. Conclusion

Under the circumstances of real-time operating condition speed fluctuates and under such conditions gearboxes gen-
erate non-stationary gear vibration signals. To track the fault in the signal under fluctuating speed condition is a serious
challenge to conventional Fourier transforms based approaches, which exhibit limitations towards the fault. To address this
practical problem of speed fluctuations, FDA of warped IMFs based a new approach to detect the local gear faults from
acquired vibration signal is presented in this paper. FDA is an effective method to perform synchronous averaging; it shows
sensitivity towards fault, thereby minimizing the irrelevant components trapped during signal acquisition. All the gear
vibration signals were decomposed by EMD into IMFs. IMFs were warped by using DTW and as a resulting of warping gain
in the kurtosis has been noticed along with fault transients can been seen in the IMFS. Analysis of the warped IMFs
highlights fault severity in signals more clearly as compared to decomposed IMFs alone using EMD or experimentally
acquired vibration signals. FDA is more impressionable as it exhibits the gearmesh frequency, its harmonics, sidebands and
the fault frequencies thereby attenuating noise.
Experimental results clearly indicates that (a) the proposed approach is able to detect gear faults and have potential to
assess the fault diagnosis of gears under fluctuating speeds, (b) kurtosis corresponding to warped IMFs associated to the
fault vibration signals are found sensitive and exhibits a significant increase in its value to reflect the presence of gear tooth
fault and additionally (c) cost of trigger / tachometer/ key phasor signal probe is saved. Reference signal acquisition probe is
not required.

Acknowledgment

The authors would like to thank the editors and the reviewers for their valuable comments and suggestions to improve
the manuscript.

References

[1] C. Zhang, Y.S. Liang, Condition-based maintenance of rolling element bearings, Int J. COMADEM 7 (4) (2004) 3–12.
[2] A. Grall, C. Berenguer, L. Dieulle, A condition-based maintenance policy for stochastically deteriorating systems, Reliab. Eng. Syst. Saf. 76 (2) (2002)
167–180.
[3] J.S. Mitchell, History of condition monitoring and condition based maintenance, Sound Vib. 33 (11) (1999) 21–28.
[4] F. Bonnardot, M. El Badaoui, R.B. Randall, J. Danière, F. Guillet, Use of the acceleration signal of the gearbox in order to perform angular Resampling
(with limited speed fluctuation), Mech. Syst. Signal Process. 19 (2005) 766–785.
[5] M.A. Jafarizadeh, R. Hassannejad, M.M. Ettefagh, S. Chitsaz, Asynchronous input gear damage diagnosis using time averaging and wavelet filtering,
Mech. Syst. Signal Process. 22 (2008) 172–201.
[6] N. Ahamed, Y. Pandya, A. Parey, Spur gear tooth root crack detection using time synchronous averaging under fluctuating speed, Measurement 52
(2014) 1–11.
[7] Z. Feng, X. Chen, M. Liang, Iterative generalized synchro-squeezing transform for fault diagnosis of wind turbine planetary gearbox under non-
stationary conditions, Mech. Syst. Signal Process. 52–53 (2015) 360–375.
[8] N. Baydar, A. Ball, Detection of gear deterioration under varying load conditions using the instantaneous power spectrum, Mech. Syst. Signal Process.
14 (6) (2000) 907–921.
[9] M. Amarnath, I.R. Praveen Krishna, Local fault detection in helical gears via vibration and acoustic signals using EMD based statistical parameter
analysis, Measurement 58 (2014) 154–164.
[10] J.P. Wang, J. Cui, Vibration Signal Analysis and Fault Diagnosis of Gears Based on Hilbert Marginal Spectrum, Adv. Mater. Res. 909 (2014) 121–126.
 V. Sharma, A. Parey / Mechanical Systems and Signal Processing 85 (2017) 278–295 295

[11] R. Shao, W. Hu, J. Li, Multi-fault feature extraction and diagnosis of gear transmission system using time-frequency analysis and wavelet threshold de-
noising based on EMD, Shock Vib. 20 (2013) 763–780.
[12] B. Liu, S. Riemenschneider, Y. Xu, Gearbox fault diagnosis using empirical mode decomposition and Hilbert spectrum, Mech. Syst. Signal Process. 20 (3)
(2006) 718–734.
[13] Y. Qin, S.R. Qin, Y.F. Mao, Research on iterated Hilbert transform and its application in mechanical fault diagnosis, Mech. Syst. Signal Process. 22 (8)
(2008) 1967–1980.
[14] Y. Yang, Y.G. He, J.S. Cheng, D.J. Yu, A gear fault diagnosis using Hilbert spectrum based on MODWPT and a comparison with EMD approach, Mea-
surement 42 (4) (2009) 542–551.
[15] N.E. Huang, Z. Shen, S.R. Long, M.C. Wu, H.H. Shih, Q. Zheng, N.-C. Yen, C.C. Tung, H.H. Liu, The empirical mode decomposition and the Hilbert
spectrum for nonlinear and non-stationary time series analysis, Proc. R. Soc. Lond. Ser. A 454 (1998) 903–995.
[16] W.D. Mark, H. Lee, R. Patrick, J.D. Coker, A simple frequency-domain algorithm for early detection of damaged gear teeth, Mech. Syst. Signal Process. 24
(8) (2010) 2807–2823.
[17] L. Hong, J.S. Dhupia, A time-domain fault detection method based on an electrical machine stator current measurement for planetary gear-sets, in
Proc. IEEE/ASME Int. Conf. Adv. Intell. Mechatronics, 2013, pp. 1631–163.
[18] C.A. Ratanamahatana, E. Keogh. Everything you know about dynamic time warping is wrong, In: Third Workshop on Mining Temporal and Sequential
Data, in conjunction with the 10th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD-2004), Seattle, WA (2004).
[19] L.J. Davis, Frequency domain synchronous averaging, 2000, IMAC XVIII.
[20] W.D. Mark, C.P. Reagor, Static-transmission-error vibratory-excitation contributions from plastically deformed gear teeth caused by tooth bend-
ingfatigue damage, Mech. Syst. Signal Process. 21 (2007) 885–905.
[21] M. Eltabach, T. Vervaeke, S. Sieg-Zieba, E. Padioleau, S. Berlingen, Features extraction using vibration signals for condition monitoring of lifting cranes,
CETIM.
[22] B. Tang, T. Song, F. Li, et al., Fault diagnosis for a wind turbine transmission system based on manifold learning and Shannon wavelet support vector
machine, Renew. Energy 62 (2014) 1–9.
[23] W.D. Mark, H. Lee, R. Patrick, et al., A simple frequency-domain algorithm for early detection of damaged gear teeth, Mech. Syst. Signal Process. 24
(2010) 2807–2823.
[24] Y. Pandya, A. Parey, Failure path based modified gear mesh stiffness for spur gear with tooth root crack, Eng. Fail. Anal. 27 (2013) 286–296.
[25] Y. Pandya, A. Parey, Simulation of crack propagation in spur gear tooth for different gear parameter and its influence on mesh stiffness, Eng. Fail. Anal.
30 (2013) 124–137.
[26] Z. Feng, X. Chen, M. Liang, Iterative generalized synchrosqueezing transform for fault diagnosis of wind turbine planetary gearbox under non-sta-
tionary conditions, Mech. Syst. Signal Process 52–53 (2015) 360–375.
[27] M. Zhao, J. Lin, X. Wang, Y. Lei, J. Cao, A tacho-less order tracking technique for large speed variations, Mech. Syst. Signal Process. 40 (2013) 76–90.
[28] C. Li, M. Liang, Time–frequency signal analysis for gear box fault diagnosis using a generalized synchrosqueezing transform, Mech. Syst. Signal Process.
26 (2012) 205–217.
[29] A. Parey, N. Tandon, Impact velocity modeling and signal processing of spur gear vibration for the estimation of defect size‖, Mech. Syst. Signal
Process. 21 (2007) 234–243.
[30] A. Parey, M.E. Badaoui, F. Guillet, N. Tandon, Dynamic modelling of spur gear pair and application of empirical mode decomposition-based statistical
analysis for early detection of localized tooth defect, J. Sound Vib. 294 (2006) 547–561.
[31] A. Parey, R.B. Pachori, Variable cosine windowing of intrinsic mode functions: application to gear fault diagnosis, Measurement 45 (2012) 415–426.
[32] J. Yin, W. Wang, Z. Man, S. Khoo, Statistical modeling of gear vibration signals and its application to detecting and diagnosing gear faults, Inf. Sci. 259
(2014) 295–303.