Hindawi Publishing Corporation

Shock and Vibration

Volume 2015, Article ID 390134, 10 pages
http://dx.doi.org/10.1155/2015/390134

Research Article
Gearbox Fault Identification and Classification with
Convolutional Neural Networks

ZhiQiang Chen,1 Chuan Li,1,2 and René-Vinicio Sanchez3

1
Chongqing Engineering Laboratory for Detection, Control and Integrated System,
School of Computer Science and Information Engineering, Chongqing Technology and Business University, Chongqing 400067, China
2
Research Center of System Healthy Maintenance, Chongqing Technology and Business University, Chongqing 400067, China
3
Department of Mechanical Engineering, Universidad Politécnica Salesiana, Cuenca, Ecuador

Correspondence should be addressed to Chuan Li; chuanli@21cn.com

Received 11 March 2015; Revised 20 April 2015; Accepted 24 April 2015

Academic Editor: Dong Wang

Copyright © 2015 ZhiQiang Chen et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Vibration signals of gearbox are sensitive to the existence of the fault. Based on vibration signals, this paper presents an
implementation of deep learning algorithm convolutional neural network (CNN) used for fault identification and classification
in gearboxes. Different combinations of condition patterns based on some basic fault conditions are considered. 20 test cases with
different combinations of condition patterns are used, where each test case includes 12 combinations of different basic condition
patterns. Vibration signals are preprocessed using statistical measures from the time domain signal such as standard deviation,
skewness, and kurtosis. In the frequency domain, the spectrum obtained with FFT is divided into multiple bands, and the root
mean square (RMS) value is calculated for each one so the energy maintains its shape at the spectrum peaks. The achieved accuracy
indicates that the proposed approach is highly reliable and applicable in fault diagnosis of industrial reciprocating machinery.
Comparing with peer algorithms, the present method exhibits the best performance in the gearbox fault diagnosis.

1. Introduction have been proven effective to reflect the healthy condition

of the rotating machinery. In the vibration-based gearbox
Gearboxes play crucial roles in the mechanical transmission fault diagnostics, Wang et al. [15] proposed the application of
systems, are used to transmit power between shafts, and are local mean decomposition of the vibration signal to diagnose
expected to work 24 hours a day in the production system. a low-speed helical gearbox. Hong et al. [16] investigated
Any failures with the gearboxes may introduce unwanted the vibration measurements for the planetary gearbox fault
downtime, expensive repair, and even human casualties. detection. The vibration characteristics in both the time and
Therefore it is essential to detect and diagnose faults in the frequency domains were analyzed by Lei et al. [17] for the
the initial stage [1–4]. As an effective component for the diagnostics of the planetary gearboxes.
condition-based maintenance, the fault diagnosis has gained Various studies exist, of algorithms for detection and
much attention for the safe operations of the gearboxes [5, 6]. diagnostics of faults in gearboxes; among these are support
Machine fault identification can be done with different vector machines and artificial neural network. A support
methodologies such as vibration signature analysis, lubricant vector machines based envelope spectrum was proposed by
signature analysis, noise signature analysis, and temperature Guo et al. [18] to classify three health conditions of the
monitoring. The gearbox conditions can be reflected by such planetary gearboxes. An intelligent diagnosis model based on
measurements as vibratory, acoustic, thermal, electrical, and wavelet support vector machine (SVM) and immune genetic
oil-based signals [7–12]. Of the above the diagnostic by vibra- algorithm (IGA) was proposed for the gearbox fault diagnosis
tions is the most employed for the reason that every machine [19]. The IGA was developed to determine the optimal
is considered to have a normal spectrum until there is a fault, parameters for the wavelet SVM with the highest accu-
where the spectrum changes [13, 14]. The vibration signals racy and generalization ability. Tayarani-Bathaie et al. [20]
2 Shock and Vibration

suggested a dynamic neural network to diagnose the gas Section 5 shows the obtained results and their evaluation.
turbine fault. The artificial neural network combining with Finally some conclusions are drawn.
empirical mode decomposition was applied for automatic
bearing fault diagnosis based on vibration signals [21]. 2. Methodologies
Among all the typical classifiers, the support vector classifi-
cation (SVC) family (i.e., the standard SVC and its variants) In this section, we first present the representations of the
attracted much attention due to their extraordinary classifi- convolutional neural network. And then the approach of
cation performance. According to the researches, the SVM extracting the sensitive features is introduced, where some
family received good results in comparison with the peer classical statistical parameters are calculated from the time
classifiers. and the frequency.
Recently, deep learning received great success in the
classification field. The deep learning gained better classi- 2.1. Deep Learning with Convolutional Neural Network. Con-
fication performance owing to its “deeper” representations volutional neural network was inspired by the visual system’s
for the faulty features. Up to now, different deep learning structure [35] and in particular by the models of it proposed
networks such as deep belief network [22], deep Boltzmann by [36]. The first computational models are based on local
machines (DBMs) [23], deep autoencoder [24], and con- connectivity between neurons and on hierarchically orga-
volutional neural networks [25] have been introduced, but nized transformations of the image in Fukushima’s neocogni-
few been used for the fault diagnosis cases. Tran et al. [26]
tron [37]. LeCun and collaborators, following up on this idea,
introduced the application of the deep belief networks to
designed and trained convolutional networks using the error
diagnose reciprocating compressor valves. Tamilselvan and
gradient, where state-of-the-art performance was obtained
Wang [27] employed the deep belief learning based health
[38, 39] on several pattern recognition tasks. Modern under-
state classification for iris dataset, wine dataset, Wisconsin
breast cancer diagnosis dataset, and Escherichia coli dataset. standing of the physiology of the visual system is consistent
The limited reports used the deep learning structure for the with the processing style found in convolutional networks in
fault diagnosis, with commonly one modality feature. the literature [40]. To this day, pattern recognition systems
This paper presents a study for the application of the based on convolutional neural networks are among the best
convolutional neural network in the identification and classi- performing systems [41]. This has been shown clearly for
fication of the gearboxes fault. Convolutional neural network handwritten character recognition [38], which has served as
(CNN) is a type of feed-forward artificial neural network. Its a machine learning benchmark for many years.
individual neurons are tiled in such a way that they respond A typical convolutional neural network [38] is organized
to overlapping regions in the visual field [28]. CNN and in layers of two types: convolutional layers and subsampling
its variations are widely used models for image and video layers. Each layer has a topographic structure.
recognition [29, 30]. In this work, it is used as a classifier for At each location of each layer, there are a number of
the gearbox faults diagnosis. different neurons. Each has its set of input weights that
The most successful methods of vibration-based fault is associated with neurons in a rectangular patch in the
diagnoses are composed of two main steps: extracting the previous layer. The same set of weights, but a different input
sensitive features and classifying the condition patterns. In rectangular patch, is associated with neurons at different
the vibration-based fault diagnosis, the most commonly used locations.
features have been generated from the temporal [31], spectral
Figure 1 presents the architecture of typical convolutional
[32], wavelet [33], and other representations of the signals.
neural networks, in which the early analysis consists of
Different representations can be regarded as different obser-
alternating convolution and subsampling operations, while
vations on the vibration signals [34]. In this work, statistical
the last stage of the architecture consists of a generic mul-
measurements such as standard deviation, skewness, and
kurtosis are computed from the acquired time domain data. tilayer network: the last few layers (closest to the outputs)
In the frequency domain, the spectrum obtained with a will be fully connected 1-dimensional layers. CNNs work
FFT is divided into multiple bands. The root mean square on the 2-dimensional data, so called maps, directly, unlike
value is calculated for each band so the energy maintains normal neural networks which would concatenate these
its shape at the spectrum peaks. Vectors of the features of into vectors. Typically convolutional layers are interspersed
the preprocessed signal are formed, which are used as input with subsampling layers to reduce computation time and to
parameters for the CNN. It is important to point out that the gradually build up further spatial and configural invariance.
testing is performed under five different rotation frequencies A small subsampling factor is desirable in order to maintain
and for each one four different load conditions are applied, specificity at the same time.
which simulates the most likely scenario within an industrial Convolutional layers move forward with deriving the
application. back propagation updates in a network, which compose
The rest of this paper is structured as follows. The feature maps by convolving kernels over feature maps in
CNN model and method of extracting statistical features layers below them. At a convolution layer, the previous layer’s
are introduced in Section 2; Section 3 explains the mechan- feature maps are convolved with learnable kernels and put
ical conditions for the experiment; Section 4 presents the through the activation function to form the output feature
implementation of classifier based on the CNN model; and map. Each output map may combine convolutions with
Shock and Vibration 3

F5

C1 S1 C2 S2

Feature map 1 Subfeature Feature Subfeature

map 1 map 2

2D feature Subfeature Feature Classifier

Feature map 1 Subfeature
map map 1 map 2 (SoftMax)

Spreading
Input Output
Subfeature Feature Subfeature
Feature map 1 map 2
map 1

Convolutions Subsampling Convolutions Subsampling

Full
1D feature map connection

Figure 1: Architecture of convolutional neural networks.

multiple input maps. In general, it is calculated as follows the backpropagation algorithm as shown in [41]. It should
[41]: be clear that because kernels are applied over entire input
maps, there are many more connections in the model than
x𝑗𝑙 = 𝑓 ( ∑ x𝑖𝑙−1 ∗ k𝑖𝑗𝑙 + 𝑏𝑗𝑙 ) , (1) weights; that is, the weights are shared. This makes learning
𝑖∈𝑀𝑗 deep models easier, as compared to normal feedforward-
backpropagation neural nets, as there are fewer parameters,
where 𝑀𝑗 represents a selection of input feature maps; 𝑙 is the and the error gradients goes to zero slower because each
𝑙th layer in a network, k is a matrix of 𝑆 × 𝑆; here, 𝑆 is the size weight has greater influence on the final output.
of convolutional kernels; 𝑓 is a nonlinearity active function,
typically hyperbolic tangent or sigmoid function. Each output 2.2. Statistical Features of the Gearbox Vibration Signals.
map is given an additive bias; for a particular output map, the The gearbox condition can be reflected by the information
input maps will be convolved with distinct kernels k. That is included in different features in frequency and time domain.
to say, if output map 𝑚 and map 𝑛 both sum over input map From the set of signals obtained from the measurements of
𝑖, then the kernels applied to map 𝑖 are different for output the vibrations at different speeds and loads, the features in
maps 𝑚 and 𝑛. frequency and time domain are obtained. From the group of
A subsampling layer produces downsampled versions of graphs the values that can be used as input parameters for the
the input maps. If there are 𝑁 input maps, then, there will CNN are selected. Sixty percent of the samples set are used
be exactly 𝑁 output maps, although the output maps will be for the training of the CNN, and forty percent are used for
smaller. More formally [41], testing.
x𝑗𝑙 = 𝑓 (𝛽𝑗𝑙 down (x𝑗𝑙−1 ) + 𝑏𝑗𝑙 ) , (2)
2.2.1. Time Data Statistical Features. Usually, statistical
where down(⋅) represents a subsampling function. Typically parameters are good indices for extracting the condition
this function will sum over each distinct 𝑛-by-𝑛 block in the information. In this research, statistical measurements such
input feature map so that the output feature map is 𝑛 times as standard deviation, skewness, and kurtosis for each node
smaller along both spatial dimensions. Each output map is are used. Standard deviation, skewness, and kurtosis are
given its own multiplicative bias 𝛽 and an additive bias 𝑏. computed from the acquired time domain data; the formulas
To discriminate between 𝑇 classes a fully connected used for this are shown in Table 1, where 𝐸(𝑥) is the expected
output layer with 𝑇 neurons is added. The output layer takes value of 𝑥. Correction bias is used for the evaluation of
as input the concatenated feature maps of the layer below it, skewness and kurtosis. The standard deviation, skewness, and
denoting the feature vector, 𝑓V , kurtosis evaluated on each of the vibration signals are used
𝑂 = 𝑓 (𝑏𝑜 + 𝑤𝑜 𝑓V ) , (3) for training and testing of the CNN. The evaluation of these
is done using standard MATLAB functions.
where 𝑏𝑜 is a bias vector and 𝑤𝑜 is a weight matrix.
k𝑖𝑗𝑙 , 𝑏𝑗𝑙 , 𝑏𝑜 , and 𝑤𝑜 of the model are learnable parameters. 2.2.2. Fast Fourier Transform Banded RMS Value. Figure 2
Learning is done using gradient descent which can be imple- shows the vibration signal spectrum obtained during the test
mented efficiently using a convolutional implementation of under the following condition patterns: gear 𝑍4 with face
4 Shock and Vibration

Table 1: Formula for the evaluation of statistical values.

×106

FFT amplitude (mm/s2 )

Feature Definition 15
1 𝑁
Mean 𝑥 = ∑ 𝑥 (𝑛) 10
𝑁 𝑛=1

1 𝑁 5
Standard deviation 𝜎=√ ∑ (𝑥 (𝑛) − 𝑥)2
𝑁 𝑛=1
0
𝐸 (𝑥 − 𝑥)3 5
Skewness 𝑆= 4 5000
𝜎3 Fau 3 4000
lt 3000
1 𝑁 (𝑥 (𝑛) − 𝑥)4 num 2
1000
2000 ( Hz)
Kurtosis 𝐾= ∑ b er 1 0 uenc
y
𝑁 𝑛=1 𝜎4 Fre q

Figure 3: Frequency spectrum under five combinations of different

6 condition patterns.
×10
15
FFT amplitude (mm/s2 )

Table 2: Test’s conditions.

10
Characteristic (𝐶1 ) Value
Sample frequency 44100 [Hz] (16 bits)
5
Sampled time 10 [s]
0 Power 1000 [W]
3500 Minimum speed 700 [RPM]
3000 5000
Spe 2500 4000 Maximum speed 1600 [RPM]
ed 3000
(RP 2000 2000
1000 Hz) Minimum load 250 [W]
M) 1500 0 ue nc y (
Freq Maximum load 750 [W]
Speeds 1760, 2120, 2480, 2840, 3200 [mm/s]
Figure 2: Frequency spectrum in function of the speed, under the
following condition patterns: 375 W load, 𝑍1 with face wear 0.4 Loads 375, 500, 625, 750 [W]
[mm], 𝑍2 with face wear 0.5 [mm], 𝐵2 with 2 pitting on inner ring, Number of loads per test 10
and 𝐵3 with 2 pitting on outer ring. Type of accelerometer Uniaxial
Trademark ACS
Model ACS 3411LN
wear 0.4 [mm], gear 𝑍3 with face wear 0.5 [mm], bear 𝐵3 with
2 pits on inner ring, and bear 𝐵2 with 2 pits on outer ring for Sensibility 330 [mV/g]
5 different rotation speeds, and load at 375 W. Figure 3 shows
frequency spectrum under five combinations of different
condition patterns. The time domain signal was multiplied RMS values, standard deviation, skewness, kurtosis, rotation
by a Hanning window to obtain the FFT spectrum, in which frequency, and applied load measurements. In this work, the
a shift in the frequency and an increment in the amplitude frequency range is 0 to 22050 Hz and the size of the data
in function of the speed increment are noticeable during the vector in the frequency is 18000 samples. The spectrum is
test. The different spectrum graphs showed that the amplitude divided into 𝑁RMS frequency bands, 𝑁RMS = 251.
of each component increases in a proportional manner to the
load variation. Also in the spectra some accentuations and 3. Experimental Setup
attenuations were observed on certain spectral component,
which suggests dependency of the fault features with respect To validate the effectiveness of the proposed method, we
to the amount of load applied. carried out the experiments on a gearbox fault experimental
With the objective of reducing the amount of input data platform. Figure 4 indicates the internal configuration of
to the CNN the spectrum was split in multiple bands, because the gearbox and positions for accelerometers. There are 3
with this number of bands the root mean square (RMS) values shafts and 4 gears composing a two-stage transmission of
keep track of the energy in the spectrum peaks [42], where the gearbox. An input gear (𝑍4 = 27, modulus = 2, and
the RMS value is evaluated with (4), and 𝑁 is the number of Φ of pressure = 20) was installed on the input shaft. Two
samples of each frequency band. Consider intermediate gears (𝑍2 = 𝑍3 = 53) were installed on an
intermediate shaft for the transmission between the input
𝑁
gear and the output gear (𝑍1 = 80, installed on the output
FFTrms = ∑ FFT (𝑛) . (4) shaft). The faulty components used in the experiments
𝑛=1
included gears 𝑍1 , 𝑍2 , 𝑍3 , and 𝑍4 and bearings 𝐵1 , 𝐵2 ,
Vectors of the features of the preprocessed signal are 𝐵3 , and 𝐵4 as labeled in Figure 4(a). Test’s conditions are
formed as input parameters for the CNN as follows: 𝑁RMS described in the Table 2. The vibration signal is obtained from
Shock and Vibration 5

K K6 K3
K5 K2
K
J K4 K1
I
J
Z1
I
B6 B2

B5 Z3
Z2 B3
B1 B4

Z4

(a) (b)

Figure 4: (a) The internal configuration of the gearbox; (b) positions for accelerometers.

the measurements of a vertically allocated accelerometer in Table 3: Nomenclature of gears fault.

the gearbox case. Tables 3 and 4 present the description of
each fault condition of each component of the gearbox used Designator Description
in the experiment. We call them basic condition pattern. In 1 Normal
our experiment, a test case includes several basic condition 2 Gear with face wear 0.4 [mm]
patterns, which is a combination of multiple component 3 Gear with face wear 0.5 [mm]
faults. For example, the test case A shown in Table 5 includes 4 Gear with chafing on tooth 50%
the following information of faults:
5 Gear with chafing on tooth 100%
Gear 𝑍4 : gear with pitting on teeth. Gear with pitting on tooth depth 0.05 [mm],
6
Gear 𝑍3 : gear with face wear 0.5 [mm]. width 0.5 [mm], and large 0.05 [mm]
7 Gear with pitting on teeth
Bear 𝐵2 : bearing with 4 pits on outer ring.
8 Gear with incipient fissure on 4 mm teeth to 25%
Bear 𝐵3 : bearing with 2 pits on outer ring. of profundity and angle of 45∘
Gears 𝑍1 and 𝑍2 and bears 𝐵1 and 𝐵4 : normal. 9 Gear teeth breakage 20%
To evaluate the performance of the proposed method for 10 Gear teeth breakage 50%
gearbox fault diagnosis, first, we constructed 12 condition 11 Gear teeth breakage 100%
patterns as listed in Table 5. Each pattern with 4 different load
conditions and 5 different input speeds was applied during Table 4: Nomenclature of bears fault.
the experiments. For each pattern, load and speed condition,
we repeated the tests for 5 times. In each time of the test, Designator Description
the vibratory signals were collected with 24 durations each 1 Normal
of which covered 0.4096 sec. 2 Bearing with 2 pits on outer ring
3 Bearing with 4 pits on outer ring
4. Implemented Classifier Based on CNN and 4 Bearing with 2 pits on inner ring
Statistical Features 5 Bearing with 4 pits on inner ring
6 Bearing with race on inner ring
In this section, the implementation of classifier based on 7 Bearing with 2 pits on ball
CNN will be introduced. Figure 5 shows the block diagram 8 Bearing with 2 pits on ball
of the process followed in the processing of the signal. The
CNN-based classifier includes parameters as follows:
(1) The size of input feature map, 𝑆in , depends on the (3) The number of output feature maps of convolution
feature representation of the preprocessed signal. layer, 𝐶𝑖 , expresses 𝑖th convolution layer. 𝑛𝐶𝑖 is as the
number of output feature maps of 𝐶𝑖 .
(2) The number of alternating convolution and subsam-
pling layers that decides the architecture of CNN is as (4) The scale of subsampling layer, 𝑆, which means the
follows. Two schemes are investigated: one is two con- size of output feature map of subsampling layer, is 1/𝑆2
volutional layers and two subsampling layers; another of that of the input feature map. 𝑆𝑖 expresses that of 𝑖th
is one convolutional layer and one subsampling layer. subsampling layer; and 𝑆𝑆𝑖 is as the scale of layer 𝑆𝑖 .
6 Shock and Vibration

Table 5: Condition patterns of the experiment. Pattern definition

Basic faults (i) Fault combination 1
Number of patterns (ii) Fault combination 2
Gear faults Bear faults
···
𝑍4 𝑍3 𝑍2 𝑍1 𝐵4 𝐵3 𝐵2 𝐵1
A 7 3 1 1 1 2 3 1
Data collection
B 7 3 6 8 1 1 1 1
C 5 5 1 1 6 7 2 1 ···
D 7 1 1 1 6 7 2 1
E 1 2 1 1 1 6 3 1
F 1 3 1 1 1 5 3 1 Model feature representations
G 2 9 1 1 6 7 3 1
H 5 5 1 1 6 3 2 4 Load and speed Time Frequency
measure modality modality
I 2 6 1 1 6 5 2 1
J 1 11 1 1 1 3 4 1 (i) Standard deviation
FFT and Hanning
K 1 1 1 1 1 6 3 1 (ii) Skewness
window
L 1 1 1 1 1 1 3 1 (iii) Kurtosis

NRMS RMS values of

(5) For each input map convolve with corresponding 88 Hz band width
kernel and add to output map; the convolutional
kernel is usually a matrix of 𝑘 × 𝑘, where 𝑘 is called
convolutional kernel size. Constructing CNN classifier

To confirm the optimal architecture of CNN-based clas- Training

sifier for gearbox fault diagnosis, some parameter tunings fine tuning
are performed. Table 6 presents 11 schemes with different
parameters of the CNN-based classifier. They are applied to a
test case with the 12 patterns indicated in Table 5, using data Gearbox fault
with 12000 sample signals, where sixty percent of the samples diagnosis result
set are used for the training of the CNN, and forty percent are
used for testing. The classification rate and computation time Figure 5: Training and testing process block diagram.
(Intel Core i7-4710MQ CPU @2.50 GHz 2.50 GHz, Memory
8.00 GB) of each epoch training are recorded in Table 6.
Table 6: Parameters tuning of the architecture of CNN.
From Table 6, we can assume that the cases of 16 × 16
input feature map are superior to those of 28 × 28. The Architecture of CNN Classification Time
cases with one convolutional layer and one subsampling layer Number
𝑆in 𝑛𝐶1 𝑆𝑆1 𝑛𝐶2 𝑆𝑆2 𝑘 rate (s/epoch)
are superior to those of two convolutional layers and two
#1 28 × 28 6 2 12 2 5 86.73% 11.6 s
subsampling layers. #7∼#11 cases have very good classification
accuracy. #9∼#11 cases have less computation times. So we #2 28 × 28 8 2 8 2 5 88.48% 12.8 s
select a configuration for the proposed CNN-based classifier #3 28 × 28 12 2 12 2 5 92.50% 21.7 s
as follows: one convolutional layer and one subsampling layer, #4 28 × 28 8 4 — — 5 86.71% 8.00 s
𝑆in = 16 × 16, 𝑛𝐶1 = 8, 𝑆𝑆1 = 2, and 𝑘 = 5. The suggested #5 16 × 16 6 2 12 2 5 90.23% 3.90 s
architecture of the CNN-based classifier for gearbox fault #6 16 × 16 8 2 8 2 5 89.50% 3.80 s
diagnosis is described in Figure 6.
#7 16 × 16 6 2 6 1 5 95.71% 2.40 s
#8 16 × 16 6 1 6 1 5 98.77% 4.50 s
5. Experiment Evaluations #9 16 × 16 6 2 — — 5 96.71% 1.04 s
The training is done in first instance with the 12 patterns #10 16 × 16 8 2 — — 5 98.35% 1.30 s
indicated in Table 5. The used data have 12000 sample signals, #11 16 × 16 12 2 — — 5 98.20% 2.02 s
where sixty percent of the samples set are used for the
training of the CNN, and forty percent are used for testing.
For further tuning parameters, we consider the #9∼#11 cases is 89.46% of the pair of 𝑛𝐶1 = 12 and epochs = 50; the best
in Table 6 with different number of iteration epochs: 50, one is 98.35% of the pair of 𝑛𝐶1 = 8 and epochs = 200. In
100, 150, 200, 250, and 300, respectively. Table 7 indicates the following experiment, 𝑛𝐶1 and epochs are set to 8, 200,
the classification rate for the first instance. As shown in respectively.
Table 7, each parameters pair has excellent performance for Confusion matrix is an effective tool and is a visualization
the gearbox faults classification. The least classification rate tool of the performance of a classification algorithm. Each
Shock and Vibration 7

1D feature Full
nC1 feature map nC1 subfeature vector
map connection
C1 S1 1
#1 12 × 12 #1 6 × 6 2
feature map 3
subfeature
···
36
#2 #2 6 × 6
12 × 12 37
feature map subfeature 38
···
72
16 × 16 #3 12 × 12 #3 6 × 6
256 statistic
feature map feature map subfeature SoftMax
feature
···
#4 #4 6 × 6
12 × 12 Output
Input feature map subfeature
Reshape
···
··· ···
Convolutions
Kernel 5 ∗ 5
#nC1 #nC1
12 × 12 6×6
feature map subfeature

Subsampling
scale SS1 = 2

Figure 6: Implementation of classifier based on CNN and statistical features.

1 401 0 0 0 3 0 0 0 0 0 0 0 99.3%
2 0 421 0 0 2 0 0 0 0 0 0 0 99.5%
3 0 0 380 11 1 0 1 1 0 0 0 0 96.5%
4 0 0 30 382 1 0 5 0 0 0 0 0 91.4%
5 0 0 0 2 432 0 0 0 0 0 0 0 99.5%
Output class

6 0 1 0 0 0 410 8 0 0 0 0 0 97.9%
7 0 0 0 0 0 0 416 0 0 0 0 0 100%
8 0 1 0 0 0 0 2 382 0 0 0 0 99.2%
9 0 0 0 0 0 0 0 0 411 1 0 0 99.8%
10 0 0 0 0 2 0 0 0 3 396 0 0 98.8%
11 0 0 0 0 1 0 0 0 0 0 406 0 99.8%
12 0 0 0 0 2 0 0 0 0 1 0 284 99.0%
100% 99.5% 92.7% 96.7% 97.3% 100% 96.3% 99.7% 99.3% 99.5% 100% 100% 98.4%
1 2 3 4 5 6 7 8 9 10 11 12
Target class

Figure 7: Confusion matrix using CNN.

Table 7: Parameters tuning of CNN. the instances in an actual class (target class). Figure 7 presents
the confusion matrix using CNN model for 12 patterns indi-
Epochs
𝑛𝐶1 cated in Table 5. As shown in Figure 7, the global percentage
300 250 200 150 100 50 of true positive classification of the 12 condition patterns
12 97.98% 98.0% 98.02% 96.71% 95.35% 89.46% of faults is 98.4% and the total error is 1.6%. The smallest
8 97.92% 98.19% 98.35% 97.98% 96.31% 91.19% percentage of true positive classifications is obtained for type
6 97.98% 97.27% 96.71% 96.71% 96.25% 93.1% 3; this is because this kind of conditions patterns with 6 basic
faults. This is evident by observing the confusion matrix in
which 30 times of type 4 are classified as type 3, noticing that
column of the confusion matrix represents the instances in mostly there is confusion between type 4 and type 3, in which
a predicted class (output class), while each row represents they have 4 same basic faults. The percentages of true positive
8 Shock and Vibration

1 320 8 2 36 12 0 0 14 12 0 0 0 79.2%
2 0 329 9 36 49 0 0 0 0 0 0 0 77.8%
3 0 0 222 150 18 0 4 0 0 0 0 0 56.4%
4 0 0 110 300 8 0 0 0 0 0 0 0 71.8%
Output class 5 0 0 35 85 313 0 1 0 0 0 0 0 72.1%
6 0 0 139 42 63 129 46 0 0 0 0 0 30.8%
7 0 0 162 82 25 7 140 0 0 0 0 0 33.7%
8 0 0 53 69 9 0 0 254 0 0 0 0 66.0%
9 0 0 0 0 0 0 0 0 412 0 0 0 100%
10 0 0 0 0 2 0 0 0 48 254 97 0 63.3%
11 0 0 0 0 1 0 0 0 9 0 397 0 97.5%
12 19 0 8 7 10 0 0 0 0 0 0 243 84.7%
94.4% 97.6% 30% 37.2% 61.4% 94.9% 73.3% 94.8% 85.7% 100% 80.4% 100% 69.0%
1 2 3 4 5 6 7 8 9 10 11 12
Target class

Figure 8: Confusion matrix using SVM.

Table 8: Comparisons of classification rate with SVM using 20 test cases.

Number 1 2 3 4 5 6 7 8 9 10
CNN 91.4% 98.5% 97.1% 98.0% 98.4% 98.6% 97.2% 98.8% 97.4% 98.9%
SVM 73.8% 77.4% 65.9% 67.5% 65.5% 79.2% 69.1% 81.5% 66.8% 72.0%
Number 11 12 13 14 15 16 17 18 19 20
CNN 98.5% 98.7% 94.2% 97.7% 93.5% 98.7% 97.7% 94.7% 92.7% 97.4%
SVM 72.3% 66.4% 55.9% 64.7% 63.5% 67.0% 62.1% 61.3% 64.0% 60.9%
Mean Std. Least Most Median
CNN 96.8% 2.57% 91.4% 98.9% 97.7%
SVM 67.8% 6.49% 55.9% 81.5% 66.8%

classification of Type 1, Type 6, and Type 12 are all 100%. These parameters were found through a search, aiming at
Confusion matrix in Figure 7 shows that the presenting CNN the best model for the SVM. Figure 8 shows the confusion
model has very high percentage of true positive classification. matrix using SVM method for 12 patterns indicated in
To further validate the robustness of the present CNN Table 5. The experiment results of 20 test cases using the
model, a fault condition pattern library was constructed, SVM method are shown in Table 8. Its global percentage
which has 58 kinds of combinations based on the basis of true positive classification of the 12 condition patterns of
patterns described in Tables 3 and 4. 20 test cases are used faults is only 69.0% and the total error is 31.0%. The smallest
to test the robustness of the present CNN method; each test percentage of true positive classifications is 30% of type 3.
case has 12 kinds of combinations that are randomly selected The mean, standard deviation, and median of classification
from the pattern library. The experiment results of 20 test rate using SVM for the 20 test cases are 67.8%, 6.49%, and
cases using the CNN are shown in Table 8. With regard to the 66.8%, respectively. Comparing with the deep learning CNN
CNN method, its smallest classification rate is 91.4% of 1st test method, the SVM method exhibits inferior performance for
case, and the largest one reaches 98.9% of 10th test case. The the gearbox fault diagnosis.
mean, standard deviation, and median of classification rate
using CNN are 96.8%, 2.57%, and 97.7%, respectively. 6. Conclusions
In addition, the CNN method was compared with “shal-
low” learning algorithms SVM. As for the SVM, one of In this paper, a deep learning technique based CNN for
the most important representatives in the “shallow” learning the vibration measurements has been proposed to diag-
community, good classification results can be found for the nose the fault patterns of the gearbox. The present CNN
gearbox fault diagnosis, which is similar with some existing method identifies and classifies faults in gearbox by using
researches (e.g., [43]). The algorithm SVM is applied using the vibration signals measured with an accelerometer. Fea-
the LibSVM [44]. The parameters for SVM are chosen as ture representations are selected as the input parameters of
𝐶 = 137.187 and core (kernel) given by a radial basis the CNN with a vector formed by RMS values, standard
𝐾(𝑢, V) = exp(−𝛾|𝑢 − V|2 ) function where 𝛾 = 1910.852. deviation, skewness, kurtosis, rotation frequency, and applied
Shock and Vibration 9

load. For evaluating the proposed CNN method, the gearbox [9] T. Toutountzakis, C. K. Tan, and D. Mba, “Application of
fault diagnosis experiments were carried out using different acoustic emission to seeded gear fault detection,” NDT & E
techniques. The results show that the present method has International, vol. 38, no. 1, pp. 27–36, 2005.
the outstanding performance of the gearbox fault diagnosis, [10] J. R. Ottewill and M. Orkisz, “Condition monitoring of gear-
comparing with peer methods. This type of classifiers could boxes using synchronously averaged electric motor signals,”
make a contribution to maintenance routines for industrial Mechanical Systems and Signal Processing, vol. 38, no. 2, pp. 482–
systems, towards lowering costs and guarantying a continu- 498, 2013.
ous production system, and, with the appropriate equipment, [11] A. M. D. Younus and B.-S. Yang, “Intelligent fault diagnosis
online diagnostics could be performed. of rotating machinery using infrared thermal image,” Expert
Systems with Applications, vol. 39, no. 2, pp. 2082–2091, 2012.
[12] C. Li and M. Liang, “Extraction of oil debris signature using
Conflict of Interests integral enhanced empirical mode decomposition and corre-
The authors declare that there is no conflict of interests lated reconstruction,” Measurement Science and Technology, vol.
22, no. 8, Article ID 085701, 2011.
regarding the publication of this paper.
[13] P. Jayaswal, A. K. Wadhwani, and K. B. Mulchandani, “Machine
fault signature analysis,” International Journal of Rotating
Acknowledgments Machinery, vol. 2008, Article ID 583982, 10 pages, 2008.
[14] A. K. S. Jardine, D. Lin, and D. Banjevic, “A review on machinery
This work is supported by Natural Science Foundation
diagnostics and prognostics implementing condition-based
Project of CQ CSTC (nos. cstc2012jjA40041 and maintenance,” Mechanical Systems and Signal Processing, vol. 20,
cstc2012jjA40059), Science Research Fund of Chongqing no. 7, pp. 1483–1510, 2006.
Technology and Business University (no. 2011-56-05), the
[15] Y. Wang, Z. He, J. Xiang, and Y. Zi, “Application of local mean
National Natural Science Foundation of China (51375517),
decomposition to the surveillance and diagnostics of low-speed
and the Project of Chongqing Innovation Team in University helical gearbox,” Mechanism and Machine Theory, vol. 47, no. 1,
(KJTD201313). pp. 62–73, 2012.
[16] L. Hong, J. S. Dhupia, and S. Sheng, “An explanation of
References frequency features enabling detection of faults in equally-
spaced planetary gearbox,” Mechanism and Machine Theory, vol.
[1] Y. Lei, M. J. Zuo, Z. He, and Y. Zi, “A multidimensional hybrid 73, pp. 169–183, 2014.
intelligent method for gear fault diagnosis,” Expert Systems with [17] Y. Lei, D. Kong, J. Lin, and M. J. Zuo, “Fault detection of plane-
Applications, vol. 37, no. 2, pp. 1419–1430, 2010. tary gearboxes using new diagnostic parameters,” Measurement
[2] Z. Yang, W. I. Hoi, and J. Zhong, “Gearbox fault diagnosis Science and Technology, vol. 23, no. 5, Article ID 055605, 2012.
based on artificial neural network and genetic algorithms,” in
[18] L. Guo, J. Chen, and X. Li, “Rolling bearing fault classification
Proceedings of the International Conference on System Science
based on envelope spectrum and support vector machine,”
and Engineering (ICSSE ’11), pp. 37–42, IEEE, Macao, China,
Journal of Vibration and Control, vol. 15, no. 9, pp. 1349–1363,
June 2011.
2009.
[3] A. Aherwar and M. S. Khalid, “Vibration analysis techniques for
[19] F. Chen, B. Tang, and R. Chen, “A novel fault diagnosis model
gearbox diagnostic: a review,” International Journal of Advanced
for gearbox based on wavelet support vector machine with
Engineering Technology, vol. 3, no. 2, 2012.
immune genetic algorithm,” Measurement, vol. 46, no. 1, pp.
[4] D. Wang, P. W. Tse, and K. L. Tsui, “An enhanced Kur- 220–232, 2013.
togram method for fault diagnosis of rolling element bearings,”
[20] S. S. Tayarani-Bathaie, Z. N. Sadough Vanini, and K. Khorasani,
Mechanical Systems and Signal Processing, vol. 35, no. 1-2, pp.
“Dynamic neural network-based fault diagnosis of gas turbine
176–199, 2013.
engines,” Neurocomputing, vol. 125, pp. 153–165, 2014.
[5] D. Wang, P. W. Tse, W. Guo, and Q. Miao, “Support vector
[21] J. B. Ali, N. Fnaiech, L. Saidi, B. Chebel-Morello, and F. Fnaiech,
data description for fusion of multiple health indicators for
“Application of empirical mode decomposition and artificial
enhancing gearbox fault diagnosis and prognosis,” Measure-
neural network for automatic bearing fault diagnosis based on
ment Science and Technology, vol. 22, no. 2, Article ID 025102,
vibration signals,” Applied Acoustics, vol. 89, pp. 16–27, 2015.
2011.
[6] C. Li and M. Liang, “Separation of the vibration-induced signal [22] N.-N. Ji, J.-S. Zhang, and C.-X. Zhang, “A sparse-response
of oil debris for vibration monitoring,” Smart Materials & deep belief network based on rate distortion theory,” Pattern
Structures, vol. 20, no. 4, Article ID 045016, 2011. Recognition, vol. 47, no. 9, pp. 3179–3191, 2014.
[7] S. Hou, M. Liang, Y. Zhang, and C. Li, “Vibration signal [23] B. Leng, X. Zhang, M. Yao, and Z. Xiong, “A 3D model
demodulation and bearing fault detection: a clustering-based recognition mechanism based on deep Boltzmann machines,”
segmentation method,” IMeche Proceedings Part C: Journal of Neurocomputing, vol. 151, pp. 593–602, 2015.
Mechanical Engineering Science, vol. 228, no. 11, pp. 1888–1899, [24] K. Katayama, M. Ando, and T. Horiguchi, “Model of MT
2014. and MST areas using an autoencoder,” Physica A: Statistical
[8] D. Wang, Q. Miao, and R. Kang, “Robust health evaluation of Mechanics and Its Applications, vol. 322, pp. 531–545, 2003.
gearbox subject to tooth failure with wavelet decomposition,” [25] T. N. Sainath, B. Kingsbury, G. Saon et al., “Deep convolutional
Journal of Sound and Vibration, vol. 324, no. 3–5, pp. 1141–1157, neural networks for multi-modality isointense infant brain
2009. image segmentation,” NeuroImage, vol. 108, pp. 214–224, 2015.
10 Shock and Vibration

[26] V. T. Tran, F. Althobiani, and A. Ball, “An approach to fault [44] C.-C. Chang and C.-J. Lin, “LIBSVM: a Library for support
diagnosis of reciprocating compressor valves using Teager– vector machines,” ACM Transactions on Intelligent Systems and
Kaiser energy operator and deep belief networks,” Expert Technology, vol. 2, no. 3, article 27, 2011.
Systems with Applications, vol. 41, no. 9, pp. 4113–4122, 2014.
[27] P. Tamilselvan and P. Wang, “Failure diagnosis using deep belief
learning based health state classification,” Reliability Engineering
and System Safety, vol. 115, pp. 124–135, 2013.
[28] DeepLearning 0.1. LISA Lab, “Convolutional Neural Networks
(LeNet)—DeepLearning 0.1 documentation,” August 2013.
[29] M. Matsugu, K. Mori, Y. Mitari, and Y. Kaneda, “Subject inde-
pendent facial expression recognition with robust face detection
using a convolutional neural network,” Neural Networks, vol. 16,
no. 5-6, pp. 555–559, 2003.
[30] Y. LeCun, “LeNet-5, convolutional neural networks,” November
2013.
[31] M. Amarnath and I. R. P. Krishna, “Local fault detection in
helical gears via vibration and acoustic signals using EMD based
statistical parameter analysis,” Measurement, vol. 58, pp. 154–
164, 2014.
[32] K. K. McKee, G. L. Forbes, I. Mazhar, R. Entwistle, M.
Hodkiewicz, and I. Howard, “A vibration cavitation sensitivity
parameter based on spectral and statistical methods,” Expert
Systems with Applications, vol. 42, no. 1, pp. 67–78, 2015.
[33] R. Yan, R. X. Gao, and X. Chen, “Wavelets for fault diagnosis of
rotary machines: a review with applications,” Signal Processing,
vol. 96, pp. 1–15, 2014.
[34] C. Lim, M. Liang, and T. Wang, “Criterion fusion for spectral
segmentation and its application to optimal demodulation
of bearing vibration signals,” Mechanical Systems and Signal
Processing, 2015.
[35] Y. Bengio, “Learning deep architectures for AI,” Foundations
and Trends in Machine Learning, vol. 2, no. 1, pp. 1–27, 2009.
[36] D. H. Hubel and T. N. Wiesel, “Receptive fields, binocular
interaction, and functional architecture in the cat’s visual
cortex,” The Journal of Physiology, vol. 160, pp. 106–154, 1962.
[37] K. Fukushima, “Neocognitron: a self-organizing neural net-
work model for a mechanism of pattern recognition unaffected
by shift in position,” Biological Cybernetics, vol. 36, no. 4, pp.
193–202, 1980.
[38] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based
learning applied to document recognition,” Proceedings of the
IEEE, vol. 86, no. 11, pp. 2278–2323, 1998.
[39] Y. LeCun, B. Boser, J. S. Denker et al., “Backpropagation applied
to handwritten zip code recognition,” Neural Computation, vol.
1, no. 4, pp. 541–551, 1989.
[40] T. Serre, G. Kreiman, M. Kouh, C. Cadieu, U. Knoblich, and T.
Poggio, “A quantitative theory of immediate visual recognition,”
in Computational Neuroscience: Theoretical Insights into Brain
Function, vol. 165 of Progress in Brain Research, pp. 33–56,
Elsevier, 2007.
[41] J. Bouvrie, “Notes on convolutional neural networks,” Tech.
Rep., MIT, 2006.
[42] I. A. Abu-Mahfouz, “A comparative study of three artificial
neural networks for the detection and classification of gear
faults,” International Journal of General Systems, vol. 34, no. 3,
pp. 261–277, 2005.
[43] D. L. de Souza, M. H. Granzotto, G. M. de Almeida, and L. C.
Oliveira-Lopes, “Fault detection and diagnosis using support
vector machines—a SVC and SVR comparison,” Journal of
Safety Engineering, vol. 3, no. 1, pp. 18–29, 2014.
 International Journal of

Rotating
Machinery

International Journal of
The Scientific
Engineering Distributed
Journal of
Journal of

Hindawi Publishing Corporation

World Journal
Hindawi Publishing Corporation Hindawi Publishing Corporation
Sensors
Hindawi Publishing Corporation
Sensor Networks
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Journal of

Control Science
and Engineering

Advances in
Civil Engineering
Hindawi Publishing Corporation Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Submit your manuscripts at

http://www.hindawi.com

Journal of
Journal of Electrical and Computer
Robotics
Hindawi Publishing Corporation
Engineering
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

VLSI Design
Advances in
OptoElectronics
International Journal of

International Journal of
Modelling &
Simulation
Aerospace
Hindawi Publishing Corporation Volume 2014
Navigation and
Observation
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014
in Engineering
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014
Engineering
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014
Hindawi Publishing Corporation
http://www.hindawi.com
http://www.hindawi.com Volume 2014

International Journal of
International Journal of Antennas and Active and Passive Advances in
Chemical Engineering Propagation Electronic Components Shock and Vibration Acoustics and Vibration
Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014