Measurement 125 (2018) 63–77

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Compound gear-bearing fault feature extraction using statistical features T

based on time-frequency method
⁎
Laxmikant S. Dhamandea, , Mangesh B. Chaudharib
a
Department of Mechanical Engineering, Sanjivani College of Engineering, Kopargaon, 423603 (SPPU, Pune), India
b
Department of Mechanical Engineering, Vishwakarma Institute of Technology, Bibwewadi, Pune 411037, India

A R T I C LE I N FO A B S T R A C T

Keywords: A small incipient fault in gear and bearing causes multiple faults in gear-bearing system leading to catastrophic
Compound gear-bearing fault failure. The purpose of present study is to explore more complex situation of compound gear-bearing fault. The
Feature extraction compound faults such as a fault in the inner and outer race of bearing along with two teeth of gear having corner
Multiple faults damage or three teeth of gear having corner damage, etc. are investigated using experimentation. To improve
Vibration analysis
the effectiveness of diagnosis, vibration measurement is done at different speed and load condition. This paper
Wavelet transform
proposes new compound fault features, extracted from continuous and discrete wavelet transform of vibration
signal. The methodology consist of proposing the features in time–frequency domain and comparison of its
diagnostic potential with respect to the features extracted from time and frequency domain for compound fault
identification using three different classifiers. The fault classification accuracy of these features is found to be
better than the conventional time and frequency domain parameters.

1. Introduction
Gears and bearings are the most important and critical machine
components in condition monitoring of rotating machineries like gear
boxes of all automobiles, machine tools, aircrafts, turbines, etc. The
failure of gears or bearings is one of the most frequent reasons for
machine breakdown. Number of mechanical system wear out failures
are due to bearings and shock loading and pitting failures of gears.
Therefore the detection of faults in gears and bearings is the important
task for maintenance engineer using condition based maintenance to
avoid catastrophic failure. The existing work in this area is con-
centrated on either gear or bearing fault diagnosis. It is observed that a
good amount of work has been carried out by number of researchers on
development of different techniques such as wavelet transform [1–4],
envelop analysis [5], empirical mode decomposition method [6,7] for
identification of single and multiple faults in bearings. Similarly the
techniques such as time-frequency distribution [8,9], wavelet transform
[10,11], statistical feature extraction [12–16], multi-scale morpholo-
gical filters [17] are satisfactorily applied for gear fault identification.
In addition to this multiple fault identification is carried out using
methods such as adaptive spectral kurtosis [18], spectrogram [19], and
time-frequency methods [20–25]. However, it is possible that the fault
may be present in both gear and bearing leading to compound fault in
transmission system [26–35] which needs effective diagnosis system.
The main intention of these techniques is to minimize the time and cost
of machine repairs, or its effective implementation. It is found that
these techniques are very lengthy and needs an expert in the field for
making decision. So it is needed to give a solution to this problem by
providing a methodology which is easy to understand and also possible
to apply effectively and automatically. In this work a methodology is
proposed for identification of compound fault using statistical features.
If there is a fault in bearing it will affect the gear vibration signature
and vice versa [27]. The analysis of individual bearing or gear fault by
the diagnosis system may generate erroneous results. In actual practice,
there may be faults in both gear and bearing, so it is required to extract
the statistical features from vibration signal which can detect the
multiple compound faults.
Various techniques have been studied by researchers for detection
of local and distributed faults in gear and bearing systems such as wear,
pitting of gear teeth, spalls and tooth fillet crack in gear, worn and
broken teeth, chipping, breakage of more than one tooth and complete
removal of one tooth, inner and outer race fault in bearing, etc. using
vibration signal. A more challenging task is to explore simultaneous
multiple faults in both gear and bearing. In machinery fault detection
basically two types of methodologies are used, visual inspection by
qualified personnel and feature extraction and classification by using
machine intelligence. Singh and Kumar [4] have shown how bearing
groove race fault width can be measured using wavelet decomposition

⁎
Corresponding author.
E-mail address: lsdhamande@rediffmail.com (L.S. Dhamande).

https://doi.org/10.1016/j.measurement.2018.04.059
Received 24 January 2018; Received in revised form 30 March 2018; Accepted 16 April 2018
Available online 20 April 2018
0263-2241/ © 2018 Elsevier Ltd. All rights reserved.
L.S. Dhamande, M.B. Chaudhari
of vibration signal. Wang et al. [11] investigated the sensitivity and
robustness of techniques such as amplitude and phase demodulation,
beta kurtosis and wavelet transform for healthy gears, filed, chipped,
and cracked gears. It is found that the beta kurtosis is a very reliable
time-domain diagnostic technique, while phase modulation is sensitive
to imperfections in gears, and continuous wavelet transform provides a
good visual inspection. A time and frequency domain vibration signal is
generally used by the analyst for detecting faults in a system. Saravanan
and Ramachandran [15] have analyzed the gear faults such as face
wear and tooth breakage by extracting features using discrete wavelet
transform and artificial neural network is used as a classifier. Bordoloi
and Tiwari [20] have used features from frequency domain for single
gear fault identification using support vector machine classifier. Sa-
walhi and Randall [26] have prepared a simulation model to under-
stand the gear and bearing interactions. The comparison of inner and
outer race fault in bearing is presented by using simulated model and
experimental measurement by spectrum comparisons, Spectral Kurtosis
(SK) analysis and envelope analysis techniques. Dhamande and
Chaudhari [27] have used statistical features in time and frequency
domain for detection of combined gear-bearing fault using artificial
neural network. Any vibration signal collected contains noise, so for
analyzing the signal different de-noising or filtration methods are used.
Ibrahim et al. [28] have used the adaptive filtering technique with least
mean square algorithm for extraction of gear box fault features i.e. gear
mesh frequency and their sidebands. Gan et al. [30] have proposed a
multi-domain manifold method for extracting features from faulty gear
and bearing.
In the current work a novel thing is that, the more complex but real
situation is considered where the compound multiple faults in gear and
bearing i.e. multiple faults in multiple components are investigated,
which is an uncovered field of research. This is a very critical situation
which will cause the failure of the system quite early than the single or
multiple incipient faults in individual component. Here an effort is
made to extract useful features such as standard deviation, variance and
absolute maximum from wavelet coefficient plot of Continuous Wavelet
Transform (CWT) at scales corresponding to gear mesh frequency and
its harmonics and from discrete wavelet transform of the measured
vibration signal for compound fault identification.
2. Experimental data collection
2.1. Experimental set-up
The schematic of the experimental set up is as shown in Fig. 1(a). It
consists of a rotor bearing system driven by a 0.5 HP, 3000 rpm D C
motor. The rotor bearing system is a model of turbine. The power is
transferred to a single stage spur gear box through belt and pulley ar-
rangement. A gear box having 26 teeth on pinion and 46 teeth on gear
with a module of 2.11 mm and 20 mm face width is provided with a
dynamometer for applying load torque [27]. The bearing supporting
gear box driving shaft is 6004 type of single row deep groove ball
bearing. A single accelerometer having sensitivity 100 mV/g
(g = 9.81 m/s2) is placed on the gearbox casing in the radial (vertical)
direction over the faulty bearing. To get maximum diagnosis accuracy,
measurements are carried out at the location close to the faulty gear,
which is mounted on driving shaft and supported on faulty bearing.
Vibration amplitude in terms of velocity is measured with the help of a
four channel Fast Fourier Transform (FFT) analyser. The sampling fre-
quency used is 16,384 Hz. Total six compound faulty conditions of gear
and bearing are created. The fault on driver gear is created by removing
a small amount of material from the corner of tooth indicating a tri-
angular area of about 3 mm2 by filing as shown in Fig. 2(a). The two
faulty conditions of gears are corner of two teeth damaged and corner
of three teeth damaged at 1200 apart, while the third faulty condition
on gear is created by removing one tooth completely. A local circular
fault of 0.8 mm in size on the outer and inner race of the bearing is
64
Measurement 125 (2018) 63–77
created by Electro Discharge Machining (EDM) as indicated in Fig. 2(b).
The two faulty conditions on bearing are two holes on outer race and
one hole on each inner and outer race of bearing. As a representative of
these faults, two teeth of gear having corner fault and one fault (pit) on
each inner and outer race of bearing are as shown in Fig. 2(a) and (b).
To analyze the effect of compound fault on vibration signal, mea-
surements are done for seven different conditions of gearbox. The seven
different conditions considered are as indicated in Table 1. The effect of
atmospheric conditions such as temperature, moisture is accounted by
collecting vibration signals at regular interval of time. The experimental
setup is properly tightened, aligned and balanced to avoid the peaks at
shaft frequency (1X) and its harmonics, which may be introduced due
to presence of mechanical faults such as unbalance (at 1X), misalign-
ment (at 2X), and looseness (at 1X, 2X, 3X….…..etc.). The operating
speed of 1000 rpm with variation of speed between 970 and 1030 rpm
with a step size of 10 rpm is used. The loads are varied as 10 N, 20 N,
30 N, and 40 N respectively. The setup is run for some time and read-
ings are collected after attaining the stable condition. Initially the
readings are collected from healthy condition of gearbox. The pinion
shaft is removed and a faulty pinion with two teeth having corner fault
is mounted on it. The pinion shaft was initially supported on healthy
bearing, which is then replaced by a faulty bearing with two faults on
its outer race. The vibration signals are then measured from a first
compound fault condition where there are two faults on outer race of
bearing and two teeth of gear having corner damage. Similarly other
compound fault conditions are created in experimental setup and vi-
bration signals are measured from it. In the present study, results from a
test conducted on a healthy gear and bearing is presented and discussed
along with the faulty gear and bearing. A healthy gear vibration signal
contains several gear meshing harmonics with a few accompanying
sidebands but, of low amplitude. Here, the shaft speed is 1000 rpm i.e.
16.67 Hz, and gear mesh frequency is 433.42 Hz. The effect of each type
of fault on the vibration signal is different; hence it is possible to extract
the features from vibration signal which can detect the fault more
precisely.
2.2. Methodology and procedure
The methodology followed in this work (Fig. 3) consists of use of an
experimental setup where it is possible to introduce the multiple faults
in gear and bearing simultaneously. The different components used in
this setup such as bearing block, bearing, shaft, gear, input and output
pulleys are connected to each other, so the vibration due to gear in-
teraction is transferred from gears to shaft, shaft to bearings and from
bearings to bearing block or housing. This connectivity makes the vi-
bration signal measured by accelerometer at housing surface a complex
multi-component signal. This signal is generated by shaft rotation,
tooth meshing, bearing vibration, gearbox resonance vibration and
noise due to interaction of all the components. The actual measured
vibration signal is a mixed signal of gear and bearing, hence a novel
thing is, to detect more complex and real situation in transmission
machinery where multiple components have multiple faults in it. The
time and frequency domain signal measured using an accelerometer in
terms of velocity is used for analysis. For identifying the compound
fault in the gear box system, raw vibration signals are measured from a
single stage spur gear box in seven different conditions. In each faulty
condition minimum three and maximum five faults are present on gear
and bearing in combination.
The methodology followed in this work consists of following im-
portant steps:
1. It consists of preparing an experimental set-up where compound
multiple faults in gear and bearing can be introduced.
2. Measuring time and frequency response of vibration along with the
overall Root Mean Square (RMS) and range of amplitude using FFT
analyser. A single accelerometer is located at the top of gear box
L.S. Dhamande, M.B. Chaudhari
Fig. 1. Schematic of experimental setup.
casing above the faulty bearing for measurement.
3. Extracting thirteen different statistical features T1 to T13 from time
response and another fourteen features F1 to F14 from frequency
response for one revolution of gear from all the faulty conditions of
gear and bearing as given in Fig. 3 processing flow for feature ex-
traction.
4. In time-frequency domain, thirteen features (W1 to W13) are ex-
tracted from the continuous wavelet transform of the raw vibration
signal using db44 as a mother wavelet and thirteen features are
extracted from discrete wavelet transform (DWT) of the raw vibra-
tion signal.
5. Finally different features in each domain are compared for their
diagnostic potential in identifying the type of fault which will be
useful for classifying the fault more accurately using artificial neural
network.
3. Signal processing and extraction of features
The aim of this section is to extract a number of conventional as well
as advanced features in time, frequency and time- frequency domain.
The diagnostic potential of these features is checked for identification of
multiple faults in gear and bearing.
(a)
Fig. 2. Photographic view of seeded faults in gear and bearing, (a) two corner fault on gear and (b) fault on inner and outer race of bearing.
65
Measurement 125 (2018) 63–77
3.1. Conventional feature extraction
The RMS value of vibration in mm/s and amplitude range is noted
from experimental setup using accelerometer. The plot for RMS velocity
in mm/s versus speed in RPM and RMS velocity in mm/s versus radial
load in N for the faulty and healthy conditions is as shown in Fig. 4(a)
and (b). Fig. 4(a) shows that the value of RMS increases with increase in
speed and with an increase in the number of faults. It is observed that,
the range of variation of RMS value for every faulty condition is dif-
ferent. Fig. 4(b) indicates the increase in RMS value of vibration with
increase in load for faulty and healthy conditions. This may be useful
for predicting the presence of faults; however it is difficult to know
which type of fault leads to increase in RMS value. Distinguishing a
compound fault in gear and bearing is a challenging task using vibra-
tion signal. For healthy condition of gear box the amplitude of vibration
falls in the range of −5 to +5 mm/s whereas for the compound fault
type 3 (FT3) the range of amplitude varies from −11 to +11 mm/s for
a radial load of 40 N at 1000 rpm. Similarly, the increment in the vi-
bration amplitude range for compound fault type 6 (FT6) is −14 to
+14 mm/s and for FT5; it is −7 to +7 mm/s. It is observed that there
is increment in the value of range of amplitude for different faulty
conditions. The vibration signal in time domain for FT3 condition at
1000 rpm is as shown in Fig. 5 while Fig. 6 shows the frequency
(b)
L.S. Dhamande, M.B. Chaudhari Measurement 125 (2018) 63–77

Table 1
Compound fault conditions.
Condition Description

Healthy gear-bearing (H) Healthy condition of gear-bearing

Fault type 1 (FT1) Two pits on outer race of bearing and two teeth of gear having corner damage
Fault type 2 (FT2) Two pits on outer race of bearing and three teeth of gear having corner damage
Fault type 3 (FT3) Two pits on outer race of bearing and one tooth of gear missing
Fault type 4 (FT4) One pit on each inner and outer race and two teeth of gear having corner damage
Fault type 5 (FT5) One pit on each inner and outer race and three teeth of gear having corner damage
Fault type 6 (FT6) One pit on each inner and outer race and one tooth of gear missing

spectrum plot of it. The conventional and most successful features in
time and frequency domain for faulty gear and bearing are extracted as
explained in the following sections.
3.1.1. Features extracted from time domain signal
The time domain signal is measured from experimental setup using
FFT analyzer. It is expressed as a time series x (n), with n = 1, 2, 3….
….N, where N is number of data points in a signal. The time domain
signal is segmented, to get signal for one revolution of pinion shaft i.e.
for 60 ms for all the conditions. The statistical parameters, which are
measure of central tendency such as mean (T1), standard deviation
(T2), variance (T3), root mean square (T4), and absolute maximum
(T5), are calculated from time domain signal using Eqs. (1)–(5) as in-
dicated in Table 2. These parameters are then combined to get the
features which are more significant than those parameters alone. The
derived features of time domain signal are crest factor (the ratio of
absolute maximum to RMS), margin factor (the ratio of absolute max-
imum to variance), A factor (the ratio of Absolute maximum to product
of standard deviation and variance), and B factor (the ratio of product
of kurtosis and crest factor to standard deviation). The additional
features extracted from time waveform are coefficient of skewness (T6),
kurtosis (T7), crest factor (T8), margin factor (T9), shape factor (T10),
impulse factor (T11), A factor (T12), and B factor (T13) using Eqs.
(6)–(13) respectively. All these statistical features are considered here
to know how vibration signature changes due to presence of faults and
to know their diagnostic potential. All these features are chosen, as they
have been proven useful for bearing and gear fault identification
[13,16].
3.1.2. Features extracted from frequency domain signal
The frequency domain signal for healthy and faulty conditions of
gear and bearing are obtained using FFT analyzer. The spectrums for
compound fault type 1 (FT1) and FT5 are as shown in Fig. 7. The
measured vibration signal contains periodic impulse components re-
lated to bearing, mesh components related to gear faults and all the
fault frequencies including shaft, rotor, etc. It is observed that the
amplitude of gear mesh frequency of compound fault type 5 (FT5) is
more than FT1.
The vibration signal collected in frequency domain in terms of
amplitude and frequency is expressed as y(m) for m = 1, 2, M, where M

Fig. 3. Processing flow for feature extraction.

66
L.S. Dhamande, M.B. Chaudhari Measurement 125 (2018) 63–77

RMS Vs Speed
6

5.5

5 H
RMS in mm/s
4.5 FT3

FT6
4
FT5
3.5

2.5

1.5
970 980 990 1000 1010 1020 1030
Speed in RPM

(a)

RMS Vs Radial Load

6

5.5 H
FT3
5
FT6
4.5 FT5
RMS in mm/s

3.5

2.5

1.5
10 15 20 25 30 35 40
Radial load in N
(b)
Fig. 4. Variation in RMS value of signal for different faulty conditions, (a) with speed and (b) with load.

Time response
15

10
Amplitude(mm/s)

-5

-10
0 10 20 30 40 50 60
Time (msec)

Fig. 5. A typical time response of 60 ms duration for fault type 3 (FT3).

67
L.S. Dhamande, M.B. Chaudhari Measurement 125 (2018) 63–77

Frequency Spectrum Plot

0.4

0.35

0.3
Velocity (mm/s)

0.25

0.2

0.15

0.1

0.05

0
0 200 400 600 800 1000 1200 1400 1600 1800 2000 2200
Frequency (Hz)

Fig. 6. A typical frequency spectrum plot for fault type 3 (FT3).

is the number of spectrum lines, fm is the frequency value of the mth
spectrum line. The numbers of spectrum lines are selected by con-
sidering first five gear mesh frequency harmonics and their sidebands
for feature extraction. The overall representation of spectrum obtained
for each type of faulty condition is different than the other. This is
utilized in diagnosis of faults by extracting features from frequency
domain for a bandwidth of 0–2200 Hz. The statistical parameters such
as mean (F1), variance (F2), third moment (F3), and forth moment (F4)
are calculated from frequency spectrum using Eqs. (14)–(17) as in-
dicated in Table 3. Additional features are extracted from this spectrum
by using combination of above standard parameters such as grand mean
(F5), standard deviation with respect to F5 (F6), C factor (F7), D factor
(F8), E factor (F9), G factor (F10), Third Moment with respect to F5
(F11), Forth Moment with respect to F5 (F12), H factor (F13) and J
factor (F14). The frequency domain statistical features F5 to F14 are
calculated by Eqs. (18)–(27).
3.2. Features extracted from continuous wavelet transform
The time domain signal is transformed into time-frequency domain
using wavelet transform. Fig. 8(a) and (b) shows the continuous wa-
velet transform of vibration signal collected at 1000 rpm of the input
shaft for healthy gear and bearing condition (H) and for compound fault
type 1 (FT1). It is a three dimensional plot in the time-frequency do-
main. The amplitude of vibration is perpendicular to the plane of the
paper; the black shade gives a positive value while the white shade
gives a negative value of amplitude. The majority of researcher’s have
used db2, db4, db5, db10 and db40 as a mother wavelet function and 3,
4, 8 and 10 level of decomposition of vibration signal for identification
of fault in gear and bearing. Rafiee et al. [13] in his research found
db44 as a suitable candidate having a similar shape for gear and bearing
signature. The continuous wavelet transform of the raw vibration signal
is calculated using db44 as a mother wavelet.
Most of the researchers have used wavelet transform for visually
identifying the faults in gear and bearing. In this work an attempt is
made to find some new features from continuous wavelet transform.
Generally fault in bearing can be identified by observing the increase in
amplitude of bearing fault frequencies. Similarly, the fault in gear can
be identified by observing the increase in amplitude of gear mesh fre-
quency and increase in number and amplitude of sidebands. If both gear
and bearing have fault present in it then there is effect of bearing faults
on gear mesh frequency amplitude and effect of gear faults on bearing
fault frequency amplitude [27]. In this work, it is decided to consider
the wavelet coefficient plot at gear mesh frequency and its first two
harmonics. This will consider the effect of faults in gear as well as in
bearing on vibration signal. It requires initially finding the value of
scale corresponding to gear mesh frequency on CWT plot. This is fol-
lowed by extracting a wavelet coefficient plot from CWT at the required
scale. Generally, for gears the energy is centered at gear mesh frequency
(GMF), hence the scale 25 corresponding to GMF is used for extracting
the wavelet coefficient plot. The wavelet coefficient plot for healthy
condition and compound fault (FT1) at gear mesh frequency is as shown
in Fig. 9(a) and (b). Wavelet coefficients plots at three different scales
namely 25, 12 and 9 are used for analysis, which corresponds to gear
mesh frequency and its first two harmonics [11]. It is observed that the
plot in Fig. 9(b) is quite differentiable with respect to the plot obtained
from the healthy gear condition as shown in Fig. 9(a). The statistical
features W1 to W13 are extracted from each scale of the wavelet
coefficient plot obtained from CWT along with energy of signal (ED)
using Eqs. (28)–(41) as indicated in Table 4. The time-frequency do-
main signal is expressed as a series z(n), with n = 1, 2, 3….….N, where
N is number of data points of signal.

Table 2
Features from the time domain signal.
Parameter Equation Parameter Equation

Mean ∑nN= 1 x (n) (1) Crest factor T8 =

T5 (8)
T1 = T4
N
Standard deviation ∑nN= 1 (x (n) − T1)2
(2) Margin factor T9 =
T5 (9)
T2 = T3
N−1
Variance N 2 (3) Shape factor T10 =
T4 (10)
T3 = ⎜ ⎛ ∑n = 1 |x (n)| ⎞ 1 N
⎟ ∑ | x (n)|
N N n=1
⎝ ⎠
RMS ∑nN= 1 (x (n))2
(4) Impulse factor T11 =
T5 (11)
T4 = 1 N
∑ | x (n)|
N N n=1
Absolute maximum T5 = max|x (n)| (5) A factor T12 =
T5 (12)
T2 . T3
Coefficient of skewness ∑nN= 1 (x (n) − T1)3 (6) B factor T13 =
T7 . T8 (13)
T6 = T2
(N − 1)T32
Kurtosis ∑nN= 1 (x (n) − T1)4 (7)
T7 =
(N − 1)T24

68
L.S. Dhamande, M.B. Chaudhari Measurement 125 (2018) 63–77

Frequency Spectrum Plot Frequency Spectrum Plot

0.35 0.7

0.3 0.6

0.25 0.5

0.2 0.4
Velocity

Velocity
0.15 0.3

0.1 0.2

0.05 0.1

0 0
0 200 400 600 800 1000 1200 1400 1600 1800 2000 2200 0 200 400 600 800 1000 1200 1400 1600 1800 2000 2200
Frequency Frequency

(a) (b)
Fig. 7. Frequency response of gearbox for analysis of compound faults, (a) frequency domain signal for FT1 and (b) frequency domain signal for FT5.

Table 3
Features from frequency domain signal.
Parameter Equation Parameter Equation

Mean M y (m)
∑m =1
(14) D factor ∑mM f 4 y (m) (21)
F1 = F8 = =1 m
M M f 2 y (m)
∑m =1 m
Variance M (y (m) − F 1)2
∑m =1
(15) E factor ∑mM f 2 y (m)
=1 m
(22)
F2 = F9 =
(M − 1) ∑mM y (m) ∑M f 4 y (m)
=1 m=1 m
Third moment M (y (m) − F 1)3
∑m (16) G factor F10 =
F6 (23)
F3 = =1
F5
M ( F 2 )3
Forth moment M
∑m = 1 (y (m) − F 1)4 (17) Third moment1 M (f − F )3y (m)
∑m =1 m 5 (24)
F4 = F11 =
M (F 2)2 MF63
Grand mean M f y (m)
∑m (18) Forth moment1 M (f − F 5)4y (m)
∑m (25)
F5 = =1 m F12 = −1 m
M y (m)
∑m MF64
=1
Standard deviation1 M (f − F 5)2y (m)
∑m
(19) H factor M (f − F 5)1/2y (m)
∑m (26)
=1 m F13 = −1 m
F6 = M F6
M
C factor M f 2 y (m)
∑m
(20) J factor F14 = (F7 + F8)/F1 (27)
F7 = =1 m
M y (m)
∑m =1

3.3. Features extracted from discrete wavelet transform (ED) are extracted from each detail obtained from DWT for every faulty
condition as given in Eqs. (28)–(41). Similarly the time-frequency do-
The time domain signal is obtained using sampling frequency of main signal processing for extraction of features is carried out for other
16,384 Hz and for a frequency range of 0–6400 Hz. The total numbers faulty conditions. The diagnostic potential of each feature is determined
of data points for 250 ms sample are 4096. The time for one revolution to know which features are better for classification of faults using
is 0.06 s at shaft speed of 1000 rpm. The time domain signal for one method explained in following section.
revolutions is utilized for analysis i.e. 60 ms. The time domain signal is
decomposed into eight level details D1 to D8 and approximation A1 to
4. Diagnostic results and discussions
A8 using discrete wavelet transform. The original signal is having
maximum frequency of 6400 Hz. This is decomposed into approxima-
Many tests are carried out on an experimental set up to create da-
tions and details. The first level approximation A1 contains frequency
tabase. The results in terms of various thirteen features in time domain,
range of 0–3200, D1 contains 3200–6400, A2 contains 0–1600, D2
fourteen features in frequency domain and forty-one useful features in
contains 1600–3200, A3 contains 0–800, D3 contains 800–1600, A4
the time-frequency domain are extracted. From the recorded vibration
contains 0–400, D4 contains 400–800, A5 contains 0–200, D5 contains
waveforms in time, frequency and time-frequency domain the whole set
200–400, A6 contains 0–100, D6 contains 100–200, A7 contains 0–50,
of features as described above are obtained in MATLAB utilizing Ad hoc
D7 contains 50–100, A8 contains 0–25 and D8 contains 25–50 Hz. The
algorithms. In this work a new and simplified feature selection method
details D1, D4, D5 and D6 are showing good diagnostic potential. The
is proposed to obtain effective features from each domain, from diag-
bearing and gear defect frequencies lies in decomposition level D4;
nostic point of view based on two approaches as follows. 1. Percentage
hence D4 level decomposition is used for analysis. The forth level ap-
increment or decrement in the value of feature with respect to the
proximation (A4), and the detail coefficients D1 to D8 for healthy
healthy condition for number of readings corresponding to different
condition of gear are as shown in Fig. 10(a) while for compound fault
loads and speeds. 2. The pattern of variation of a particular feature with
condition FT3 in Fig. 10(b). It is observed that the compound fault
respect to the type of faults. The feature selection method is as ex-
condition can be separated from healthy condition since, the magnitude
plained below with reference to time domain features.
of D2, D3 and D4 for compound fault are quite different than healthy
The time domain signal is expressed as a time series
condition. The features are extracted from high frequency components
x (n) = [x1 x2 x3 ……. x n] with n = 1, 2, 3 …N, where N is number of data
of signal, i.e. details and low frequency components of signal i.e. ap-
points of signal. Here (p-1) types of faulty conditions are required to be
proximations, but in all the multiple faults the low frequency signal are
classified, along with the healthy condition. The total number of con-
nearly same, hence could not give any significant change but appreci-
ditions to be classified is denoted as p. The number of raw features
able change is observed in high frequency components of signal i.e.
extracted from vibration signal is denoted as q and these raw features
Details. The total fourteen features W1 to W13 and energy of signal
compose a feature vector.

69
L.S. Dhamande, M.B. Chaudhari Measurement 125 (2018) 63–77

Absolute Values of Ca,b Coefficients for a = 1 2 3 4 5 ...

248
235
222
209
196
183
170
157
scales a

144
131
118
105
92
79
66
53
40
27
14
1
100 200 300 400 500 600 700 800 900
time (or space) b
(a)
Absolute Values of Ca,b Coefficients for a = 1 2 3 4 5 ...
248
235
222
209
196
183
170
157
scales a

144
131
118
105
92
79
66
53
40
27
14
1
100 200 300 400 500 600 700 800 900
time (or space) b
(b)
Fig. 8. CWT of (a) healthy gear and bearing (H) and (b) FT1: processed on raw signal at 1000 rpm.

Fi,j = [T1j T2j T3j ……… Tij]T
where Tij = f (x (n)) , i denote for the ith raw feature, i = 1 to q and jth
health condition, j = 1 to p. It is assumed that the length of Ti under jth
condition is L i.e. number of readings collected for each condition.
Ti = [Ti (1) Ti (2) Ti (3) . ………. Ti (L)].
The method utilized for feature selection involves following im-
portant steps:
(42) Step 1: The amplitude of feature Ti is plotted against number of read-
ings L for all the conditions. It is represented by nearly hor-
izontal lines where it is possible to see the variation of a par-
ticular raw feature with respect to number of readings. It can be
compared with all other conditions also.
Step 2: The feature lines having clear differentiation are said to be
having more diagnostic potential. In this way few features
having enough diagnostic potential are selected at first stage.

Wavelet coefficient Plot Wavelet coefficient Plot

2 15

10
1

5
Scale
Scale

0
0

-1
-5

-2 -10
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time Time

(a) (b)
Fig. 9. Wavelet coefficient plot corresponding to GMF of raw vibration signal, (a) healthy condition and (b) compound fault (FT1).

70
L.S. Dhamande, M.B. Chaudhari Measurement 125 (2018) 63–77

Table 4
Features from the time-frequency domain signal.
Parameter Equation Parameter Equation

Mean ∑nN= 1 z (n) (28) Crest factor W8 =

W5 (35)
W1 = W4
N
Standard deviation ∑nN= 1 (z (n) − W1)2
(29) Margin factor W9 =
W5 (36)
W2 = W3
N−1
Variance ∑nN= 1 |z (n)| ⎞
2 (30) Shape factor W10 =
W4 (37)
W3 ⎛
=⎜ 1 N
⎟ ∑ | z (n)|
N N n=1
⎝ ⎠
RMS ∑nN= 1 (z (n))2
(31) Impulse factor W11 =
W5 (38)
W4 = 1 N
∑ | z (n)|
N N n=1
Absolute maximum W5 = max|z (n)| (32) A factor W12 =
W5 (39)
W2 . W3
Coefficient of skewness ∑nN= 1 (z (n) − W1)3 (33) B factor W13 =
W7 . W8 (40)
W6 = W2
(N − 1)T32
Kurtosis ∑nN= 1 (z (n) − W1)4 (34) Energy of signal N
ED = ∑n = 1 (z (n))2 (41)
W7 =
(N − 1)W 24

I.e. features having better (relative amplitude difference be-
tween the healthy and faulty condition), Δmax = Tp−Tp − 1.
Step 3: Then these selected features are further processed to find the
better feature. This is done by calculating average of Ti
∑ [Ti (1) Ti (2) Ti (3). .......Ti (L)]
Average of Ti = μi =
L (43)
Average of Ti is calculated for all the features q and for all the
health conditions p and its pattern is observed. A particular type
of pattern of variation of these parameters is observed for dif-
ferent types of faults.
Step 4: The features having same pattern are listed and the features
having different pattern of variation are selected as final fea-
tures to be used for fault diagnosis.
Step 5: The step 1 to step 4 is repeated for time domain signal, fre-
quency domain signal and also for time-frequency domain
signal to select the best features in each domain.
In the following sections the behavior of the features from a diag-
nostic point of view are presented.
times, FT5 – 3 times and FT6 – 6.7 times the healthy condition. Table 5
indicates the diagnostic potential of time domain parameters in per-
centage of healthy condition. This indicates that these features can
capture class information in varying degrees, thereby confirming the
effectiveness of the features in fault identification.
Fig. 12 shows the variation of frequency domain features variance
(F2), grand mean (F5), standard deviation1 (F6), and C factor (F7). It is
observed that for every condition features F1, F2, F5, F6, F7, F8, F9,
F10, F12 have some diagnostic potential. The increment in the value of
parameter F1 for compound fault type 1 (FT1) is 2.3 times, for FT2 is
3.8 times, for FT3 is 2.7 times, for FT4 is 2.1 times, for FT5 is 2.5 times,
and for FT6 is 5 times the healthy condition. The feature F5 shows 20%
increment for compound fault type 1 (FT1), 25% increment for the FT5
and 5% increment for the FT6.
Table 6 indicates the diagnostic potential of frequency domain
features in percentage of healthy condition. It is observed that only a
few features in time and frequency domain have some potential to
detect compound faults in gear and bearing.
4.2. Diagnostic potential of CWT and DWT features

4.1. Diagnostic potential of time and frequency domain features
Step 1 is followed for all the features T1 to T13 extracted from time
domain signal. The diagnostic potential to detect specific faulty con-
dition lies in specific feature. Fig. 11 indicates the plot obtained by step
1 for time domain features, standard deviation (T2), variance (T3),
absolute maximum (T5), and A factor (T12) for ten number of record-
ings with respect to healthy condition, where seven different lines in-
dicate each faulty condition. After following the step 3 and step 4 it is
observed that in time domain, the parameters-kurtosis (T7), crest factor
(T8), margin factor (T9) and impulse factor (T11) have same kind of
pattern of variation, while standard deviation (T2), variance (T3), and
absolute maximum (T5) have same pattern. The parameter A factor
(T12) and B factor (T13) have a different pattern than the other para-
meters. Considering this, the minimum number of time domain features
having excellent diagnostic potential is selected. In this feature selec-
tion method the features having better values than normal condition
are selected and the feature which shows a different pattern of variation
with respect to the type of fault is also considered.
It is observed that for every condition (H, FT1 to FT6), T2, T3, T5
and T12 features have shown some diagnostic potential since the lines
drawn for different conditions for these features are having some dif-
ferentiation from each other. These features produce noticeable limits
and display a unique pattern to separate one class from the other. The
increment in the value of parameters T2 and T3 for compound fault
type 1 (FT1) is 2 times, for FT2 – 2.2 times, FT3 – 4.5 times, FT4 – 2.9
Using feature selection procedure as explained in Section 4 the
variation of time-frequency domain features W2, W3, W5 and W7 is as
shown in Fig. 13. It is found that three features standard deviation
(W2), variance (W3) and absolute maximum (W5) have some diag-
nostic potential. The diagnostic potential of standard deviation (W2),
variance (W3) and absolute maximum (W5) features extracted from the
wavelet coefficient plot for different scales in percentage of healthy
condition is as reported in Table 7. It is observed that for scale 25 the
increment in the value of the W2 feature for compound fault type 1
(FT1) is 0.7 times, for FT2 is 3.4 times, for FT3 is 1.7 times, for FT4 is
6.3 times, for FT5 is 2 times, and for FT6 is 11 times the healthy gear
condition, hence it is found that the CWT coefficients have very good
diagnostic potential at scale 25 as well as at scale 12 and scale 9.
Similarly Table 8 indicates the percentage increase in the value of
standard deviation (W2), variance (W3) and absolute maximum (W5)
features extracted from detailed coefficients D1, D4, D5, and D6 ob-
tained using DWT for compound fault in gear and bearing. It is found
that there is enough potential in these features for compound gear and
bearing fault. Also the energy of signal is calculated for all the details
ED1 to ED8. It is found that only ED5 is showing the diagnostic po-
tential. Thus it is found that out of the fourteen features in frequency
domain, six features have some diagnostic potential. New features ex-
tracted from CWT and DWT for detecting compound gear-bearing fault
are showing good diagnostic potential. The following section shows
how these new features are better than the conventional features using
artificial neural network as a classifier.

71
L.S. Dhamande, M.B. Chaudhari Measurement 125 (2018) 63–77

Approximation A4 Detail D1 Detail D2

1 0.4 2

0.5 0.2 1
Amplitude

0 0 0

-0.5 -0.2 -1

-1 -0.4 -2
0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60

Detail D3 Detail D4 Detail D5

1 0.5 0.4

0.5 0.2
Amplitude

0 0 0

-0.5 -0.2

-1 -0.5 -0.4
0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60

Detail D6 Detail D7 Detail D8

0.4 0.5 0.6

0.2 0.4
Amplitude

0 0 0.2

-0.2 0

-0.4 -0.5 -0.2

0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (ms) Time (ms) Time (ms)
(a)
Approximation A4 Detail D1 Detail D2
2 1 10

1 0.5 5
Amplitude

0 0 0

-1 -0.5 -5

-2 -1 -10
0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60

Detail D3 Detail D4 Detail D5

10 4 2

5 2 1
Amplitude

0 0 0

-5 -2 -1

-10 -4 -2
0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60

Detail D6 Detail D7 Detail D8

1 2 0.6

0.5 1 0.4
Amplitude

0 0 0.2

-0.5 -1 0

-1 -2 -0.2
0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (ms) Time (ms) Time (ms)

(b)
Fig. 10. Discrete wavelet transform of raw vibration signal, (a) healthy Condition and (b) compound fault (FT3).

4.3. Discussion and comparison with some previous works
Ample amount of work has been done on detection of gear and
bearing faults using different techniques. In general the total work is
divided into two categories. It consist of either analyzing the vibration
signal using advanced signal processing techniques or using an expert
system for automatic classification of faulty conditions using features
extracted from vibration signal. The former needs an expert in the field
for decision making while the later does not need human intervention
for decision.
Ali et al. [6] found Root Mean Square (RMS) (T4), and absolute
maximum (T5), coefficient of skewness (T6), kurtosis (T7), crest factor

72
L.S. Dhamande, M.B. Chaudhari Measurement 125 (2018) 63–77

5 3
2.5
Feature value T2

Feature value T3
4
2
3
1.5
2
1
1 0.5

0 0
0 2 4 6 8 10 0 2 4 6 8 10
Number of samples Number of samples

25 20
FT1

Feature value T12

FT2
Feature value T5

20
15 FT3
15 FT4
10 FT5
10 FT6
H
5
5

0 0
0 2 4 6 8 10 0 2 4 6 8 10
Number of samples Number of samples
Fig. 11. Diagnostic potential of time domain parameters for different fault conditions at 1000 rpm.

Table 5 Table 6
Diagnostic potential of features extracted from time domain in percentage of Diagnostic potential of features extracted from frequency domain in percentage
healthy condition. of healthy condition.
Feature FT1 FT2 FT3 FT4 FT5 FT6 Feature FT1 FT2 FT3 FT4 FT5 FT6

T2 199.8 203.2 458.5 284.7 302.5 649.3 F1 228.8 387.1 274.9 210.6 250.3 504.9
T3 224.1 198.2 482.3 297.5 292.4 712.5 F2 466.0 558.6 617.7 797.1 458.3 3048.8
F5 20.8 21.7 13.5 17.4 24.5 4.9
F6 54.2 84.2 56.5 70.8 67.7 121.7
F7 9.8 12.6 3.4 10.7 13.7 0.1
(T8), margin factor (T9), shape factor (T10), impulse factor (T11) and
F8 −4.3 0.8 −7.6 1.1 −0.4 −0.9
two additional parameters in time domain better for single bearing fault F9 14.8 11.7 11.9 9.5 14.2 1.0
identification using ANN. Bordoloi and Tiwari [23] have found var- F10 27.3 49.3 37.5 45.1 34.4 111.0
iance (F2), third moment (F3), and forth moment (F4) parameters from F12 20.8 21.7 13.5 17.4 24.5 4.9
frequency domain, better for single gear fault identification using
support vector machine classifier. Rafiee et al. [13] have found

-3
x 10
8 1300
Feature value F2

Feature value F5

6 1200

4 1100

2 1000

0 900
0 2 4 6 8 10 0 2 4 6 8 10
Number of samples Number of samples

200 1700
FT1
1600 FT2
Feature value F6

Feature value F7

FT3
150 1500
FT4
1400 FT5
FT6
100 1300 H
1200

50 1100
0 2 4 6 8 10 0 2 4 6 8 10
Number of samples Number of samples
Fig. 12. Diagnostic potential of frequency domain parameters for different fault conditions at 1000 rpm.

73
L.S. Dhamande, M.B. Chaudhari Measurement 125 (2018) 63–77

3 1.5
Feature value W2

Feature value W3
2 1

1 0.5

0 0
0 2 4 6 8 10 0 2 4 6 8 10
Number of samples Number of samples

8 8
FT1
Feature value W5

Feature value W7
6 FT2
6 FT3
4 FT4
4 FT5
2 FT6
H
0 2
0 2 4 6 8 10 0 2 4 6 8 10
Number of samples Number of samples
Fig. 13. Diagnostic potential of time-frequency domain parameters for different fault conditions at 1000 rpm.

Table 7 single gear and single bearing fault diagnosis from the available lit-
Diagnostic potential of features extracted from wavelet coefficients in percen- erature. Loutas et al. [16] have extracted statistical features from vi-
tage of healthy condition at 1000 rpm. bration signal collected from single fault in gear. They found RMS (T4),
Feature Scale FT1 FT2 FT3 FT4 FT6 coefficient of skewness (T6), and kurtosis (T7) parameters in time do-
main; mean (F1), variance (F2), standard deviation with respect to
W2 9 418.8 296.4 552.6 282.6 570.5 grand mean (F6), G factor (F10), and H factor (F13) parameters in
W3 9 481.2 332.9 564.1 320.7 638.4
frequency domain and energy of detail signal D1 (ED1), ED2, and ED3
W5 9 348.2 253.0 541.9 251.7 534.6
W2 12 472.1 333.9 721.8 281.9 1178.6 parameters from time-frequency domain, are better for gear fault
W3 12 523.3 369.3 723.6 317.8 1303.9 identification. The diagnostic potential of these parameters is assessed
W5 12 359.6 277.2 657.1 216.9 925.0 by observing its percentage increment with respect to the crack growth
W2 25 69.7 337.1 170.0 631.5 1122.1 in gear tooth.
W3 25 79.1 373.1 183.0 725.5 1190.0
W5 25 42.6 249.6 161.3 439.5 1021.3
Table 9 indicates the work done by different researchers on single
gear and bearing fault identification. They have used features in time
domain, frequency domain and time frequency domain for it. In this
Table 8 work compound/multi-component fault is investigated. The different
Diagnostic potential of features extracted from 4th level wavelet decomposition features which are found successful for gear and bearing in time, fre-
in percentage of healthy condition at 1000 rpm. quency and time- frequency domain used by other researchers are ex-
tracted from vibration signal. These features are evaluated and checked
Detail Feature FT1 FT2 FT3 FT4 FT5 FT6
for its diagnostic potential for compound fault. It is found that these
D1 W2 71.9 187.3 167.5 146.3 187.3 572.1 features are not having enough diagnostic potential; hence some new
D1 W3 116.1 250.5 222.3 167.9 250.5 732.5 features are extracted from time-frequency domain for compound fault.
D1 W5 36.0 134.4 130.9 127.7 134.4 404.3
The new features variance (W3) at scale 9, 12, and 25 from CWT,
D4 W2 365.0 287.8 626.1 203.2 287.8 1469.6
D4 W3 379.9 313.6 623.8 224.7 313.6 1644.9 standard deviation at detail D4 (W2D4), variance at detail D4 (W3D4)
D4 W5 345.8 276.9 723.5 197.1 276.9 1204.7 and absolute maximum at detail D4 (W5D4), found to have good diag-
D5 W2 54.8 296.3 133.1 551.7 296.3 956.0 nostic potential.
D5 W3 70.9 351.8 159.6 679.7 351.8 1064.2
D5 W5 31.2 201.0 94.6 367.2 201.0 780.6
D6 W2 70.6 83.8 103.0 63.0 83.8 233.6
4.4. Comparison of features using artificial neural network (ANN)
D6 W3 85.3 84.8 94.8 69.4 84.8 231.4
D6 W5 44. 71.9 111.6 44.2 71.9 234.2 To compare the performance of new features extracted from CWT
and DWT, six features having maximum diagnostic potential are se-
lected each from time, frequency and time-frequency domain. A multi-
standard deviation (T2), variance (T3), and kurtosis (T7) parameters component fault identification system using ANN is trained using these
from time-frequency i.e. from DWT of vibration signal, better for de- features, as ANN is a more accurate robust and less time consuming
tection of single tooth fault in gear system using artificial neural net- classifier. In this section three different systems/networks are compared
work (ANN). Jena et al. [19] have found standard deviation (T2), Root for their classification accuracy using confusion matrix. Each network is
Mean Square (T4), kurtosis (T7), and crest factor (T8) parameters in trained using same number of features and architecture for proper
time-frequency domain i.e. from CWT of acoustic signal, better for comparison and validation.
identification of multiple teeth faults in gear system. In this paper, we
have extracted 13 different features in time domain and 14 different 4.4.1. Network using features from time domain
parameters in frequency domain along with some new features for In this section, a feed-forward network trained with back propaga-
compound fault identification. These features are found successful for tion is utilized, which is a supervised network, and uses a set of input

74
L.S. Dhamande, M.B. Chaudhari
Table 9
Comparison of this work with some earlier research.
Literature Domain used Features proposed
[6] Time, T4, T5, T6, T7, T8, T9, T10, T11, Add1, Add2
Time + Frequency IMF1 to IMF8
[23] Frequency F2, F3, F4
[13] Time + Frequency W2, W3, Coefficient of skewness from DWT
[19] Time + Frequency W2, W3, Kurtosis, Crest factor from DWT
[16] Time, T4, T6, T7;
Frequency, F1, F2, F6, F10, F13;
Time + Frequency ED1, ED2, ED3
This work Time + Frequency W3 at scale 25, 12, 9 from CWT
W2D4, W3D4, W5D4 from DWT
vectors and a set of associated desired output vectors called target
vectors. The input vector consists of six time domain features having
enough diagnostic potential such as T2, T7, T8, T9, T12 and T13. The
output vector called the target matrix for different faulty conditions is
prepared using zeros and ones in MATLAB workspace. The next step is
to create a network and train it till it learns the relationship between the
given inputs and targets. The network used is a four-layer feed-forward
back propagation network. For all the conditions, a total of 210 (30 ∗ 7)
samples, thirty feature vectors for each condition consisting of six fea-
ture value sets, selected from above analysis, are used for training and
testing of network. The neural network is created in MATLAB using
ANN tool box with six input neurons, two hidden layers and seven
output neurons. The sigmoid transfer function is used in hidden and
output layer, the training function used is Levenberg Marquardt, and
the performance function used is mean squared error.
The performance plot gives the mean squared error of 0.06532 at
epoch 12. The 210 inputs and target vectors are randomly divided into
three sets. A 60% of the vectors are used to train the network, 20% of
the vectors are used to validate how well the network is generalized and
finally, the last 20% of the vectors are used for testing of network. 10-
fold cross validation process is employed to evaluate the performance.
Also to improve the accuracy of results the numbers of hidden neuron
are increased from 2 to 20. The better results are obtained for a network
created with 14 neurons in the hidden layer. The overall training,
testing and validation found to be the best fit with R = 0.7821 while
only 18 epochs are required for this. The trained network is used for
testing the new samples which the network has never seen, using si-
mulation of the network. The performance of ANN in the percentage
prediction of gearbox faults using above network is calculated using
confusion matrix. The accuracy (AC) of classification is calculated by a
proportion of the total number of correct predictions as given below.
TP + TN
Accuracy (AC ) =
TP + TN + FP + FN (44)
where TP is the true positive i.e. proportion of positive cases that were
correctly identified by the classifier, TN is the true negative i.e. pro-
portion of negative cases that were correctly classified, FP is the false
positive i.e. proportion of negative cases that were incorrectly classified
and FN is the false negative i.e. proportion of positive cases that were
incorrectly classified. The confusion matrix is prepared using a total of
35 new samples. It is observed that total eight incorrect or confusing
predictions are given by the above trained network. There is at least one
incorrect prediction in each faulty condition; hence the neural network
can classify the faulty conditions with an accuracy of 77.14%.
4.4.2. Network using features from frequency domain
In this network, the input vector consists of six frequency domain
features having enough diagnostic potential such as F1, F4, F5, F6, F7,
and F8. The neural network consists of six input neurons, two hidden
75
Measurement 125 (2018) 63–77
Fault considered Multi component fault considered
Single fault in bearing No
Chipped tooth, No
Missing tooth
Single fault in gear No
(Worn tooth)
Single fault in gear No
(Broken tooth)
Gear tooth crack No
Compound Fault in gear and bearing Yes
layers with nine neurons in hidden layer and seven output neurons. The
performance plot gives the mean squared error of 0.0027 at epoch 10.
The overall training, testing and validation found to be the best fit with
R = 0.9919 while the only 16 epochs are required for this. It is ob-
served that total six incorrect predictions are given by this trained
network out of 35 new samples, hence the neural network can classify
the faulty conditions with an accuracy of 82.85%.
4.4.3. Network using features from CWT and DWT
In this network, the input vector consists of six time-frequency do-
main features having enough diagnostic potential such as W325, W312,
W39, W2D4, W3D4, and W5D4. The neural network consists of six input
neurons, two hidden layers with sixteen neurons in hidden layer and
seven output neurons. The performance plot gives the mean squared
error of 0.0289 at epoch 31. The overall training, testing and validation
found to be the best fit with R = 0.9408 while 36 epochs are required
for this. It is observed that total four incorrect predictions are given by
this trained network out of 35 new samples, hence the neural network
can classify the faulty conditions with an accuracy of 88.57%. The
performance of different networks trained; in terms of classification
accuracy is compared as shown in Fig. 14. It is observed from above
analysis that the new features extracted from CWT and DWT are better
than the conventional features from time and frequency domain.
4.5. Comparison of features with other classifiers
4.5.1. Support vector machine (SVM)
The six time-frequency domain feature data with seven different
faulty conditions is used for classification using SVM. The support
vector classifier is trained using John Platt’s sequential algorithm. This
algorithm replaces all missing values and transforms nominal features
into binary ones. The feature values are normalized by it. For proper
probability estimates the calibration model is fitted to SVM’s output.
The 10-fold cross validation is used to generate training data for
90
Classification Accuracy
85
80
75
70
Time Freq Time+Freq
Network trained
Fig. 14. Comparison of trained neural networks in different domains.
L.S. Dhamande, M.B. Chaudhari Measurement 125 (2018) 63–77

Table 10 For validating the performance of new features a neural network

Comparison with other classifiers. classifier along with SVM and NB is used to check their diagnostic
Type of Number of Training Testing Time to Time to potential. The proposed features such as standard deviation, variance
classifier features accuracy in accuracy in build test and absolute maximum from the CWT and DWT i.e. W225, W325, W512,
% % model model W2D4, W3D4 and W5D4 are having excellent diagnostic potential to
(sec) (sec) classify the compound fault in gear and bearing.
Neural 4 85.71 82.85 0.66 0.56
Network 5 84.76 88.57 0.73 0.66 Acknowledgement
6 86.67 82.85 0.73 0.63

Support 4 90.00 97.14 0.23 0.09 This work is supported by SRES’s Sanjivani College of Engineering
Vector 5 89.04 94.28 0.25 0.20 management and Vishwakarma Institute of Technology, Savitribai
Machine 6 90.00 91.43 0.73 0.17 Phule Pune University, India.
Naïve Bayes 4 85.71 91.42 0.01 0.01
Classifier 5 87.14 91.42 0.02 0.01 References
6 86.67 88.57 0.01 0.01

calibration model. The complexity parameter is set to 100. A poly
kernel function is used for evaluation. The accuracy of SVM classifier is
evaluated for a data with 6, 5, and 4 features. It is observed that the
training and testing accuracy increases or decreases with increase in
number of features used for classification. The testing accuracy of this
classifier is excellent and it is giving 97.14% accuracy even when only
four number of features are used as input as specified in Table 10. These
four features are having maximum diagnostic potential to detect com-
pound faults. These are selected using proposed feature selection
method.
4.5.2. Naïve Bayes classifier
The compound gear-bearing fault data is classified using Naïve
Bayes (NB) classifier. This classifier uses numeric estimator precision
values based on analysis of training data. The NB classifier’s testing
accuracy is determined using data with 6, 5, and 4 features. The
training and testing accuracy of this classifier also increases or de-
creases with increase in number of features.
Here the accuracy comparison of ANN, Naive Bayes classifier and
Support Vector Machine is done by keeping same number of samples for
training and testing. The classification accuracy comparison by above
three methods with effect of number of features on classification ac-
curacy is as given in Table 10. It can be noted that the training as well
as testing accuracy is better for the SVM compared to ANN and Naive
Bayes classifier. It is observed that the training accuracy of neural
network and Naïve Bayes classifier is same while it is better for support
vector machine. The time required to build training and testing model is
found to be minimum for Naïve Bayes classifier as compared to ANN
and SVM. The SVM classifier is giving better testing accuracy as com-
pared to the ANN and NB.
To know the effectiveness of the proposed features an unknown
compound fault is considered for detection. It is found that these fea-
tures have detected this unknown compound fault effectively.
5. Conclusions
The method for compound gear and bearing fault identification has
been developed using time-frequency method based on CWT and DWT
of vibration signal, since the present literature on compound fault is
scarce. In this paper the most successful features in time and frequency
domain for gear and bearing fault identification are extracted and
evaluated for compound fault, but very few features have shown di-
agnostic potential, hence new features are extracted from CWT and
DWT of vibration signal. The new features such as standard deviation,
variance and absolute maximum are extracted from wavelet coefficient
plot of CWT at scales corresponding to gear mesh frequency and its
harmonics. Also these new features are extracted from different de-
composition levels of DWT of vibration signal. It is found that these
features have excellent diagnostic potential.
[1] X. Lou, K.A. Loparo, Bearing fault diagnosis based on wavelet transform and fuzzy
inference, Mech. Syst. Signal Process. 18 (2004) 1077–1095.
[2] A. Djebala, N. Quelaa, N. Hamzaoui, Detection of rolling bearing defects using
discrete wavelet analysis, Meccanica 43 (2008) 339–348.
[3] D. Wang, Q. Miao, X. Fan, H. Huang, Rolling element bearing fault detection using
an improved combination of Hilbert and Wavelet transforms, J. Mech. Sci. Technol.
23 (2009) 3292–3301.
[4] M. Singh, R. Kumar, Thrust bearing groove race defect measurement by wavelet
decomposition of pre-processed vibration signal, Meas. J. Int. Meas. Confed. 46
(2013) 3508–3515.
[5] V.N. Patel, N. Tandon, R.K. Pandey, Defect detection in deep groove ball bearing in
presence of external vibration using envelope analysis and Duffing oscillator, Meas.
J. Int. Meas. Confed. 45 (2012) 960–970.
[6] J.B. Ali, N. Fnaiech, L. Saidi, B. Chebel-Morello, F. Fnaiech, Application of empirical
mode decomposition and artificial neural network for automatic bearing fault di-
agnosis based on vibration signals, Appl. Acoust. 89 (2015) 16–27.
[7] A. Djebala, M.K. Babouri, M.N. Ouelaa, Rolling bearing fault detection using a
hybrid method based on Empirical Mode Decomposition and optimized wavelet
multi-resolution analysis, Int. J. Adv. Manuf. Technol. 79 (2015) 2093–2105.
[8] W.J. Wang, P.D. McFadden, Early detection of gear failure by vibration analysis-I
calculation of time frequency distribution, Mech. Syst. Signal Process. 7 (3) (1993)
193–203.
[9] P.D. Mcfadden, Detection of gear faults by decomposition of matched differences of
vibration signals, Mech. Syst. Signal Process. 14 (5) (2000) 805–817.
[10] J. Lin, L. Qu, Feature extraction based on Morlet wavelet and its application for
mechanical fault diagnosis, J. Sound Vib. 234 (2000) 135–148.
[11] W.Q. Wang, F. Ismail, M. Farid Golnaraghi, Assessment of gear damage monitoring
techniques using vibration measurements, Mech. Syst. Signal Process. 15 (2001)
905–922.
[12] L.S. Dhamande, M.B. Chaudhari, Bearing fault diagnosis based on statistical feature
extraction in time and frequency domain and neural network, Int. J. Vehicle Struct.
Syst. 8 (4) (2016) 229–240.
[13] J. Rafiee, P.W. Tse, A. Harifi, M.H. Sadeghi, A novel technique for selecting mother
wavelet function using an intelligent fault diagnosis system, Expert Syst. Appl. 36
(2009) 4862–4875.
[14] T.H. Loutas, G. Sotiriades, I. Kalaitzoglou, V. Kostopoulos, Condition monitoring of
a single-stage gearbox with artificially induced gear cracks utilizing on-line vibra-
tion and acoustic emission measurements, Appl. Acoust. 70 (2009) 1148–1159.
[15] N. Saravanan, K.I. Ramachandran, Incipient gear box fault diagnosis using discrete
wavelet transform (DWT) for feature extraction and classification using artificial
neural network (ANN), Expert Syst. Appl. 37 (2010) 4168–4181.
[16] B. Li, P.L. Zhang, Z.J. Wang, S.S. Mi, Y.T. Zhang, Gear fault detection using multi-
scale morphological filters, Meas. J. Int. Meas. Confed. 44 (2011) 2078–2089.
[17] T.H. Loutas, D. Roulias, E. Pauly, V. Kostopoulos, The combined use of vibration,
acoustic emission and oil debris on-line monitoring towards a more effective con-
dition monitoring of rotating machinery, Mech. Syst. Signal Process. 25 (2011)
1339–1352.
[18] Y. Wang, M. Liang, Identification of multiple transient faults based on the adaptive
spectral kurtosis method, J. Sound Vib. 331 (2012) 470–486.
[19] D.P. Jena, S.N. Panigrahi, Rajesh Kumar, Multiple-teeth defect localization in
geared systems using filtered acoustic spectrogram, Appl. Acoust. 74 (2013)
823–833.
[20] D.J. Bordoloi, R. Tiwari, Support vector machine based optimization of multi-fault
classification of gears with evolutionary algorithms from time-frequency vibration
data, Meas. J. Int. Meas. Confed. 55 (2014) 1–14.
[21] R. Shao, W. Hu, Y. Wang, X. Qi, The fault feature extraction and classification of
gear using principal component analysis and kernel principal component analysis
based on the wavelet packet transform, Measurement 54 (2014) 118–132.
[22] B. Assaad, M. Eltabach, J. Antoni, Vibration based condition monitoring of a mul-
tistage epicyclic gearbox in lifting cranes, Mech. Syst. Signal Process. 42 (2014)
351–367.
[23] D.J. Bordoloi, R. Tiwari, Optimum multi-fault classification of gears with integra-
tion of evolutionary and SVM algorithms, MAMT 73 (2014) 49–60.
[24] Y. Li, K. Ding, G. He, H. Lin, Vibration mechanisms of spur gear pair in healthy and
fault states, Mech. Syst. Signal Process. 81 (2016) 183–201.
[25] G. Yong, H. Qinkai, C. Fulei, A vibration model for fault diagnosis of planetary

76
L.S. Dhamande, M.B. Chaudhari
gearboxes with localized planet bearing defects, J. Mech. Sci. Technol. 30 (9)
(2016) 4109–4119.
[26] N. Sawalhi, R.B. Randall, Simulating gear and bearing interactions in the presence
of faults. Part I. The combined gear bearing dynamic model and the simulation of
localised bearing faults, Mech. Syst. Signal Process. 22 (2008) 1924–1951.
[27] L.S. Dhamande, M.B. Chaudhari, Detection of combined gear-bearing fault in single
stage spur gear box using artificial neural network, Procedia Eng. 144 (2016)
759–766.
[28] G. Ibrahim, A. Albarbar, A. Abouhnik, R. Shnibha, Adaptive filtering based system
for extracting gearbox condition feature from the measured vibrations, Meas. J. Int.
Meas. Confed. 46 (2013) 2029–2034.
[29] R.B. Randall, Vibration-based diagnostics of gearboxes under variable speed and
load conditions, Meccanica 51 (2016) 3227–3239.
[30] M. Gan, C. Wang, C. Zhu, Multiple-domain manifold for feature extraction in ma-
chinery fault diagnosis, Meas. J. Int. Meas. Confed. 75 (2015) 76–91.
77
Measurement 125 (2018) 63–77
[31] D. Yu, M. Wang, X. Cheng, A method for the compound fault diagnosis of gearboxes
based on morphological component analysis, Meas. J. Int. Meas. Confed. 91 (2016)
519–531.
[32] S. He, J. Chen, Z. Zhou, Y. Zi, Y. Wang, X. Wang, Multi-fractal entropy based
adaptive multi-wavelet construction and its application for mechanical compound-
fault diagnosis, Mech. Syst. Signal Process. 76–77 (2016) 742–758.
[33] W. Teng, X. Ding, X. Zhang, Y. Liu, Z. Ma, Multi-fault detection and failure analysis
of wind turbine gearbox using complex wavelet transform, Renew. Energy 93
(2016) 591–598.
[34] C. Luo, C. Shen, W. Fan, G. Cai, W. Huang, Z. Zhu, Research on the sparse re-
presentation for gearbox compound fault features using wavelet bases, Shock Vib.
(2015) 1–11.
[35] D. He, X. Wang, S. Li, J. Lin, M. Zhao, Identification of multiple faults in rotating
machinery based on minimum entropy deconvolution combined with spectral
kurtosis, Mech. Syst. Signal Process. 81 (2016) 235–249.