JID: NEUCOM

ARTICLE IN PRESS [m5G;July 25, 2017;6:10]

Neurocomputing 0 0 0 (2017) 1–10

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

A neural network constructed by deep learning technique

and its application to intelligent fault diagnosis of machines
Feng Jia, Yaguo Lei∗, Liang Guo, Jing Lin, Saibo Xing
State Key Laboratory for Manufacturing Systems Engineering, Xi’an Jiaotong University, Xi’an 710049, China

a r t i c l e i n f o a b s t r a c t

Article history: In traditional intelligent fault diagnosis methods of machines, plenty of actual effort is taken for the man-
Received 28 May 2017 ual design of fault features, which makes these methods less automatic. Among deep learning techniques,
Revised 9 July 2017
autoencoders may be a potential tool for automatic feature extraction of mechanical signals. However,
Accepted 14 July 2017
traditional autoencoders have two following shortcomings. (1) They may learn similar features in me-
Available online xxx
chanical feature extraction. (2) The learned features have shift variant properties, which leads to the
Communicated by Hongli Dong misclassification of mechanical health conditions. To overcome the aforementioned shortcomings, a local
connection network (LCN) constructed by normalized sparse autoencoder (NSAE), namely NSAE-LCN, is
Keywords:
Normalized sparse autoencoder proposed for intelligent fault diagnosis. We construct LCN by input layer, local layer, feature layer and
Deep learning output layer. When raw vibration signals are fed to the input layer, LCN first uses NSAE to locally learn
Intelligent fault diagnosis various meaningful features from input signals in the local layer, then obtains shift-invariant features in
Local connection network the feature layer and finally recognizes mechanical health conditions in the output layer. Thus, NSAE-LCN
incorporates feature extraction and fault recognition into a general-purpose learning procedure. A gear-
box dataset and a bearing dataset are used to validate the performance of the proposed NSAE-LCN. The
results indicate that the learned features of NSAE are meaningful and dissimilar, and LCN helps to pro-
duce shift-invariant features and recognizes mechanical health conditions effectively. Through comparing
with commonly used diagnosis methods, the superiority of the proposed NSAE-LCN is verified.
© 2017 Published by Elsevier B.V.

1. Introduction [7]. Based on these steps, lots of effort has been taken on intelli-
gent fault diagnosis. Georgoulas et al. [8] designed the features of
With the development of industry, machines have been more motor faults based on time-frequency methods and employed Ma-
automatic and efficient, and their components are linked to each halanobis Distance classifier to recognize motor health conditions.
other inseparably [1]. Once a component has a fault, this fault Prieto et al. [9] proposed a method using statistical features and
would quickly produce chain reaction and lead to the damage of hierarchical networks to classify bearing health conditions. Amar
other components. Such unexpected faults would make machines et al. [10] proposed a feature enhancement procedure to obtain
break down, resulting in economic loss and even person safety features from vibration spectra and applied neural networks to di-
threat [2]. Therefore, the fault diagnosis of machines has received agnose the bearing faults. Wang [11] designed a feature extraction
lots of attention. algorithm that extracts redundant statistical features from differ-
Intelligent fault diagnosis is one of the powerful tools in the ent wavelet decomposition levels, and applied K-nearest neighbor
field of fault diagnosis [3]. Based on massive monitored signals algorithm to identify gear health conditions. Lei et al. [12] designed
of the machines, it is able to replace diagnosticians with artifi- two features for gearboxes specifically and used these features and
cial intelligent techniques like neural networks to rapidly process relevance vector machine to recognize the health conditions.
these signals and automatically recognize mechanical health con- Although the studies above achieved good results, they may
ditions [4–6]. Thus, intelligent fault diagnosis plays an irreplace- suffer the weakness as follows. In these methods, plenty of the
able role in modern industries especially when massive vibration actual effort is taken for the manual design of feature extrac-
signals are available. As we know, traditional intelligent fault diag- tion algorithms since traditional classifiers cannot extract the rep-
nosis has two main steps: feature extraction and fault recognition resentative features from raw signals [13]. Such feature design-
ing processes should make full use of human knowledge in sig-
nal processing and diagnostic expertise, which costs much human
∗
Corresponding author. labor and makes the methods less automatic. Among deep learning
E-mail address: yaguolei@mail.xjtu.edu.cn (Y. Lei).

http://dx.doi.org/10.1016/j.neucom.2017.07.032
0925-2312/© 2017 Published by Elsevier B.V.

Please cite this article as: F. Jia et al., A neural network constructed by deep learning technique and its application to intelligent fault
diagnosis of machines, Neurocomputing (2017), http://dx.doi.org/10.1016/j.neucom.2017.07.032
JID: NEUCOM
ARTICLE IN PRESS [m5G;July 25, 2017;6:10]

2 F. Jia et al. / Neurocomputing 000 (2017) 1–10

T1 also have shift variant properties, leading to the misclassification

Amplitude (g)
of mechanical fault samples.
(a) We propose a local connection network (LCN) constructed by
normalized sparse autoencoder (NSAE), namely NSAE-LCN, to over-
come the shortcomings of autoencoders. LCN is constructed by
T2 four layers, i.e., input layer, local layer, feature layer and output
Fault Impacts layer, where the local layer is trained by NSAE. So when raw vi-
Amplitude (g)

bration signals are fed to the input layer, LCN first uses NSAE in
the local layer to locally learn various meaningful features from
(b)
the vibration signals, then obtains shift-invariant features from the
learned features in the feature layer and finally recognizes me-
chanical health conditions in the output layer. The proposed NSAE-
Time (s) LCN is validated by a gearbox dataset and a bearing dataset re-
spectively, both involving different health conditions under various
Fig. 1. Two simulated samples for mechanical fault signals: (a) the first sample and operating conditions. And its superiority is verified by comparing
(b) the second sample.
with commonly used diagnosis methods.
The contributions of this paper can be summarized as follows.

(1) Based on sparse autoencoder, NSAE is proposed for automatic

feature extraction from the vibration signals of machines. Since
techniques, autoencoders may help fault diagnosis to handle the
an orthonormality constraint is used in NSAE, the weight ma-
weakness above since their basic motivation is to be fed with raw
trix trained by NSAE can be viewed as a set of basis showing
signals and accomplish the task of feature extraction automatically
time-frequency properties, which encourages the learned fea-
[14–16]. Currently, autoencoders have attracted attentions in the
tures of NSAE to not only have meaningful patterns but also be
field of fault diagnosis. Thirukovalluru et al. [17] employed denois-
dissimilar. Thus, NSAE performs well in mechanical feature ex-
ing autoencoder to extract high-level features from manual fea-
traction.
tures and two classifiers to recognize mechanical faults. Jia et al.
(2) We propose NSAE-LCN for intelligent fault diagnosis. It incorpo-
[18] used frequency spectra as the input of a deep network based
rates the processes of feature extraction and fault recognition
on autoencoders to recognize mechanical health conditions. Chen
into a general-purpose learning procedure. Therefore, the pro-
and Li [19] applied sparse autoencoder (SAE) to get representa-
posed method can be used to directly learn features from raw
tive features from statistical values of bearing signals and recog-
vibration signals and recognize the health conditions of ma-
nized the health conditions using deep belief network. Mao et al.
chines for various diagnosis tasks.
[20] proposed a fault diagnosis method using frequency spectra
and autoencoder extreme learning machines. It can be seen that The rest of this paper is organized as follows. In Section 2,
most of these studies, however, still used manual features as the sparse autoencoder is briefly described. Section 3 details the pro-
input of the neural networks, which may deviate from the basic posed NSAE-LCN. In Section 4, the diagnosis cases of a gearbox
motivation of the autoencoders. dataset and a bearing dataset are studied separately using NSAE-
The following two shortcomings of the autoencoders are the LCN. Finally, conclusions are drawn in Section 5.
main reasons why they are not easily used to learn features well
from raw vibration signals of machines. (1) The autoencoders can- 2. Sparse autoencoder
not be ensured to get various meaningful features from the vibra-
tion signals. The feature extraction process of an autoencoder can SAE is a widely used autoencoder that attempts to learn fea-
be regarded as the dot produce results between its weight matrix tures from raw data. It has symmetrical neural network with an
composed by a set of basis vectors and the vibration signals. So input layer, a hidden layer and an output layer [21]. The input layer
obtaining good features of raw data depends on the weight ma- and the hidden layer constitute the encoder of SAE, which trans-
trix of the autoencoder. In the well trained weight matrix, its basis forms the input data into features. And the hidden layer and the
vectors should not only have the own patterns like acting as Ga- output layer constitute the decoder of SAE, which reconstructs the
bor bases to produce the meaningful features, but also be differ- input data from the corresponding features.
ent from each other so as to produce various features. Traditional Given unlabeled data {xm }M m=1
where xm ∈ N × 1 , the encoder
constraints applied to autoencoders, such as sparse regularization uses a mapping function fSAE to calculate the feature hm from xm
and weight decay, could force the basis vectors of the weight ma- and hm has K dimensions.
trix to learn patterns but cannot force them to be different. So
autoencoders learn too many similar features and prevent their hm = fSAE (xm ) = σs (WSAE1 xm + b1 ) (1)
applications in intelligent fault diagnosis of machines. (2) The au- where σ s (z) is the sigmoid function, WSAE1 is the weight matrix of
toencoders cannot be directly used for feature learning when the the encoder and b1 is the bias vector. The decoder uses a mapping
data have shift variant properties. Unfortunately, the vibration sig- function gSAE for reconstruction
nals of a faulty machine always show such properties. In Fig. 1, we
use two samples that simulate the vibration signals of a mechan- xˆm = gSAE (hm ) = WSAE2 hm + b2 (2)
ical fault to illustrate the properties. It can be seen that when a where WSAE2 is the weight matrix of the decoder and b2 is the bias
fault occurs in the machines, the periodic fault impacts are excited vector. The reconstruction error can be calculated by
by the contact of the fault component and other components, and
1  
M
the contact period is T1 . Such impulse-like vibration behavior of
JSAE = xm − xˆm 2 . (3)
a vibration signal is an important characteristic for fault recogni- 2M
m=1
tion. When machines operate, their components contact with each
other in a time-varying way. So the fault impacts of the first sam- To get better features of the input data, sparse representation is
ple and the second sample would shift by T2 . Once we use the used in SAE. So Kullback–Leibler (KL) divergence function is ap-
autoencoders to extract features from these samples, the features plied to encourage the sparsity of the feature hm and learn the

Please cite this article as: F. Jia et al., A neural network constructed by deep learning technique and its application to intelligent fault
diagnosis of machines, Neurocomputing (2017), http://dx.doi.org/10.1016/j.neucom.2017.07.032
JID: NEUCOM
ARTICLE IN PRESS [m5G;July 25, 2017;6:10]

F. Jia et al. / Neurocomputing 000 (2017) 1–10 3

Fig. 2. The flow chart of the proposed NSAE-LCN.

sparse representation. The average activation of hm is with L1 norm to find the sparse solution of the autoencoder. The
two functions are both commonly used for sparse representation,
1 
M
pˆ k = hm (4) but KL divergence function has two parameters and L1 norm only
k
M needs one parameter. Finally, a soft orthonormality constraint is
m=1
added in the cost function of NSAE so as to force NSAE to learn
where k = 1, 2, . . . ,K.
dissimilar features from mechanical vibration signals.
The sparsity penalty term KL( p pˆ k ) based on the KL divergence
Given unlabeled data {xm }M m=1
, the encoder of NSAE uses a
function can be used to penalize pˆ k deviating from the sparsity
mapping function f to calculate the feature hm
NSAE
from xm .
parameter p.
 p 1− p NSAE = f (xm ) = σr (WNSAE1 xm )
hm (7)
KL( p pˆ k ) = p log + (1 − p ) (5)
pˆ k 1 − pˆ k where σ r is ReLU and it is described in Eq. (8).
SAE is finally trained by the optimization problem in Eq. (6). 
0, if z < 0
  σr ( z ) = (8)
K z, if z ≥ 0
min JSAE + β KL( p pˆ k ) (6)
W ,b
k=1 Then the decoder of NSAE reconstruct xˆm by a mapping func-
where β is the regular parameter controlling the weight between tion gNSAE .
the reconstruction error and the sparsity penalty term. xˆm = gNSAE (hm
NSAE ) = WNSAE2 hNSAE
m
(9)

3. The proposed method To learn various features from vibration signals, an orthonor-
mality constraint is added to the weight matrix of NSAE. The cost
This section details the proposed NSAE-LCN method for in- function of NSAE using a hard orthonormality constraint is shown
telligent fault diagnosis of machines, as shown in Fig. 2. In this in Eq. (10).
method, NSAE is proposed to overcome the shortcomings of au-
1   
M M
toencoders in learning various features, and LCN is developed to min xˆm − xm 2 + λ h m
NSAE 1
deal with the shift-variance classification problem in intelligent W 2M 2
(10)
m=1 m=1
fault diagnosis.
s.t. WNSAE1WNSAE2 = I
3.1. Normalized sparse autoencoder where λ is the regular parameter of NSAE.
Traditionally, it is difficult to solve this problem with the hard
Based on SAE, we propose NSAE for various meaningful features orthonormality constraint. To loose this hard constraint, we use
in intelligent diagnosis of machines. The differences between NSAE a soft constraint based on the tied-weight strategy. We replace
and SAE are listed as follows. Firstly, we use rectified linear units WNSAE1 and WNSAE2 with W and WT in Eq. (10), respectively, and
(ReLU) as the activation function in NSAE since ReLU allows for the optimization problem becomes minimizing the following cost
more efficient training for a network than sigmoid function and function
encourages sparse activation [22]. Secondly, bias is not used in
1   
M M
NSAE because the research results [23] show that bias is nonessen- JNSAE = W T σr (W xm ) − xm 2 + λ σr (W xm )1 (11)
tial in feature learning. Thirdly, KL divergence function is replaced 2M 2
m=1 m=1

Please cite this article as: F. Jia et al., A neural network constructed by deep learning technique and its application to intelligent fault
diagnosis of machines, Neurocomputing (2017), http://dx.doi.org/10.1016/j.neucom.2017.07.032
JID: NEUCOM
ARTICLE IN PRESS [m5G;July 25, 2017;6:10]

4 F. Jia et al. / Neurocomputing 000 (2017) 1–10

In Eq. (11), the nonlinear activation function ReLU forces the J = N/Nloc , ∃J ∈ N+ (16)
learned features to be non-negative, which means that the nega- j
tive values in Wxm would be zeros and some information of the where xm ∈ Nloc ×1
and j = 1, 2, . . . , J.
input data would be lost in the training process. However, the first Local layer: Learn the local features of the signals.
term in JNSAE forces NSAE to reconstruct the input data as well as In this layer, the local weight matrix Wloc ∈ Kloc ×Nloc of LCN is
j
possible. Thus, NSAE actually encourages the results of Wxm to be used to get the local features from xm . To train Wloc , we first se-
positive, and the first term could degenerate into W TW xm − xm . lect the segment xs ∈  Nloc ×1 from the training samples {xm }Mm=1
When the value of the first term is close to zero in the training randomly. These selected segments compose the local training set
process, the result of WT W becomes I. {xs }Ns=1
s
where Ns is the number of the segments. Then, NSAE is
Since L-BFGS is an adaptive optimization algorithm and free of employed to calculate Wloc . After the segment xs is preprocessed
learning rate selection, it is used to minimize the cost function in by whitening [25], the Wloc is used to transform xs from original
space into the feature fs , and Wloc T is used to reconstruct x as x
ˆs .
Eq. (11). We calculate the gradient of JNSAE with respect to W: s
Based on Eq. (11), this objective function is given as
∂ JNSAE  1   1
∇J = = W D + λ · sgn · σr (W x ) xT + σr (W x )DT Ns 
 2 
Ns 
Kloc
 
∂W M M min Wloc
T
σr (Wloc xs ) − xs  + λ σr (Wloc
k
xs ) (17)
Wloc 2 1
(12) s=1 s=1 k=1

where D = W T σr (W x ) − x, sgn is the result of the sign function of where Wloc k is the kth row vector of W . By minimizing the ob-
loc
σ r (Wx), σr is the derivative function of ReLU, and x is the matrix jective function in Eq. (17), Wloc can be obtained. LCN can get the
form of xm . local features of the signal xm using Wloc . Thus, the local feature
m, j j
Based on L-BFGS algorithm, the update process of W can be de- floc ∈ Kloc ×1 of the segment xm is calculated as
scribed as  
m, j
floc = σr Wloc · xmj . (18)
W (i+1) = W (i ) − η (i ) H (i ) ∇ J (i ) (13)
In order to express the process concisely, we rewrite it using
where i is the ith interval of the update process, H(i) is the inverse
the matrix form. The weight matrix W1 ∈ K × N of the local layer
of Hessian matrix, and η is the step size. The computation of H(i)
is composed by J local weight matrix Wloc .
is detailed in Ref. [24], and η can be automatically determined by ⎡ ⎤
Wloc 0 ··· 0
η (i) = arg min J (W (i) − ηH (i) ∇ J (i) ) (14) ⎢ 0 Wloc ··· 0 ⎥
η >0 W1 = ⎢ . .. .. ⎥ (19)
⎣ .. .
..
. .
⎦
After the training process of NSAE, its weight matrix is actu- 0 0 ··· Wloc
ally normalized by the soft orthonormality constraint. This is the
reason why we call such a method normalized sparse autoencoder. J = N/Nloc = K/Kloc (20)

3.2. Local connection network constructed by NSAE Feature layer: Get the shift-invariant features.
We use an average strategy to deal with the shift variant prop-
As shown in Fig. 1, although the first sample and the sec- erties of the sample xm and produce its shift-invariant features. In
m, j
ond sample share the shift variant properties, the local properties the local layer, we have obtained the local feature floc of the seg-
j
of the first sample and the second sample are similar. Thus, to ment xm of xm . So the shift-invariant feature fm of xm is obtained
m, j
deal with the shift-variance classification problem, we can first use by averaging J local feature floc . Thus, the feature fm of xm can be
NSAE to extract the local features from the fault signals, and then obtained by
averaged these local features into the features. In this way, the ob-
1  m, j
J
tained features can enhance local properties of the fault signals
fm = floc (21)
and suppress random noise carried by the signals and thus they J
j=1
can present the shift-invariant properties. Finally, the obtained fea-
tures are input into the softmax classifier for classification. We Output layer: Determine the labels of input signals.
model these processes as a form of neural network and call it LCN. Softmax layer is used as the output layer of LCN since it is
In this section, we detail the construction of LCN for intelligent easy to be implemented and computed fast. Suppose that we have
fault diagnosis using NSAE, namely NSAE-LCN. As shown in Fig. 2, the training set { fm }M m=1 with its label set {ym }m=1 where ym ∈
M

LCN has an architecture containing four layers: input layer, local {1, 2, . . . , R}. For each fm , the softmax layer attempts to estimate
layer, feature layer and output layer. Firstly, raw vibration signals the probability p(ym = r| fm ) for each label of r = 1, 2, . . . , R. Thus,
are fed into the input layer. Then, in the local layer, NSAE is used the softmax layer is trained by minimizing the objective function:
to extract local features from the vibration signals. Thirdly, the fea- ⎡ ⎤
ture layer is used to average the local features so as to learn the
1⎢  e( ) ⎥
M R T
W2r fm
shift-invariant features of the vibration signals. Finally, the shift-
J (W2 ) = − ⎢ 1{ym = r} log ⎥ (22)
invariant features are input into the output layer and the health M ⎣  (W l )T fm ⎦
R
m=1 r=1 e 2
conditions of the vibration signals are determined. The details of
l=1
these layers are described as follows.
Input layer: Use raw vibration signals as its inputs. where 1{ · } denotes the indicator function returning 1 if the con-
The raw vibration signals are collected from machines, and are dition is true, and 0 otherwise, W2 is the weight matrix of the
used to compose the training set {xm , ym }M m=1
. xm ∈ N × 1 is the softmax layer, and W2r is the row vectors of W2 .
mth sample having N data points, and ym is the mth label indi- In the previous training process of NSAE-LCN, the weight matri-
cating the health condition of xm . We alternately divide xm into J ces W1 and W2 are trained separately. After this training process,
segments and input the segments of xm into LCN. fine-tuning strategy can be chosen to simultaneously update the

weight matrices W1 and W2 of NSAE-LCN so as to improve the
xm = x1m , x2m , . . . , xJm (15) classification accuracies.

Please cite this article as: F. Jia et al., A neural network constructed by deep learning technique and its application to intelligent fault
diagnosis of machines, Neurocomputing (2017), http://dx.doi.org/10.1016/j.neucom.2017.07.032
JID: NEUCOM
ARTICLE IN PRESS [m5G;July 25, 2017;6:10]

F. Jia et al. / Neurocomputing 000 (2017) 1–10 5

Table 1 Table 2
The classification accuracies of NSAE-LCN using various The cost time of NSAE-LCN using various param-
parameters in the gearbox diagnosis case. eters in the gearbox diagnosis case.

Nloc Kloc (%) Nloc Kloc (min)

50 75 100 150 200 50 75 100 150 200

50 98.48 99.05 98.99 99.22 99.35 50 1.31 1.80 2.34 3.37 4.27
75 94.08 98.86 99.24 99.20 99.31 75 1.53 1.94 2.53 3.63 4.98
100 87.56 97.34 99.20 99.25 99.29 100 1.87 2.40 2.79 4.01 5.04
150 55.32 84.58 94.86 99.16 99.26 150 2.40 3.07 3.52 4.64 5.59
200 36.42 64.36 79.85 95.69 98.71 200 2.64 3.26 3.71 5.02 6.13

4. Fault diagnosis using NSAE-LCN

4.1. Case study 1: fault diagnosis of planetary gearboxes

4.1.1. Data collection

The data were acquired from a two-stage planetary gearbox.
This planetary gearbox operated under 7 health conditions, involv-
ing a flaked needle on the planet bearing of the first stage, a pit-
ted tooth on the sun gear of the first stage, a cracked tooth on
the planetary gear of the first stage, a bearing inner race fault of
a planetary gear of the first stage, a missing tooth on the sun gear
of the second stage, a chipped tooth on the sun gear of the second
stage, and the normal condition. The gearbox under the first health
condition operated under 6 operating conditions and the gearbox
in other health conditions dividedly operated under 8 operating
conditions: 4 different drive motor speeds (210 0 rpm, 240 0 rpm,
2700 rpm and 3000 rpm) combined with 2 loading conditions (no Fig. 3. Classification accuracies of NSAE-LCN for the gearbox dataset using different
load and the maximum load). The accelerometer mounted on the regular parameter.

gearbox was employed to collect data with a sampling frequency

Table 3
of 5120 Hz. There are 335 samples collected for each health condi- The comparison of classification accuracies (mean accuracy ± standard deviation)
tion under an identical operating condition, and each sample is a of the gearbox dataset.
signal containing 1800 data points. So the gearbox dataset contains
Methods Description Training accuracy Testing accuracy
18,090 samples and is a seven-class classification problem.
1 Features + Softmax 86.20% ± 0.609% 85.87% ± 0.404%
2 Features + SVM 97.14% ± 0.451% 95.54% ± 0.216%
4.1.2. Parameter study of the proposed NSAE-LCN 3 PCA + SVM 100% 41.04% ± 1.165%
There are totally three parameters in the proposed LCN, i.e., 4 Stacked SAE + Softmax 48.63% ± 2.086% 34.75% ± 0.436%
Nloc , Kloc , and λ. So we investigate the classification results of 5 SAE-LCN 98.89% ± 0.471% 94.41% ± 0.537%
NSAE-LCN using various values of these parameters. To classify the 6 NSAE-LCN 100% 99.43% ± 0.105%

health conditions of the planetary gearbox using NSAE-LCN, 25% of

samples are randomly selected for training, 40,0 0 0 segments are
selected from these samples to train NSAE and other samples are displayed. It can be seen that when λ equals 0.1–0.4, the mean
used for testing. It is noted that each of the following experiments accuracy of the proposed NSAE-LCN is over 98% and varies lit-
is conducted in 15 trials so as to reduce the effect of randomness, tle. Thus, λ can be selected loosely and is advised to be from 0.1
and the displayed results are averaged by the results of these 15 to 0.4.
trials.
Nloc and Kloc determine the architecture of NSAE-LCN. Nloc de- 4.1.3. Diagnosis results
notes the dimension of a signal segment and measures how many Based on the analysis in the previous section, Nloc and Kloc are
segments a signal would be divided into. Kloc determines the di- both set as 100 and λ is selected as 0.25. As shown in Table 3,
mension of the local feature, and it denotes how many bases Wloc the training accuracy of the proposed NSAE-LCN is 100% and the
contains. Fine-tuning is not used in this section since we pay at- testing accuracy is 99.43%, and the corresponding standard devi-
tention to self-learning ability of NSAE under different parameters. ation values are 0% and 0.105%, respectively. The results indicate
Table 1 shows the classification accuracies of the NSAE-LCN based that NSAE-LCN is able to directly classify the health conditions of
on various Nloc and Kloc when λ is fixed as 0.25. It can be seen the gearbox from the raw vibration signals and obtains high and
that the diagnosis accuracies are relatively high when Nloc ≤ Kloc . stable classification accuracies.
So NSAE-LCN is able to robustly obtain the good results if we set To show the effectiveness of NSAE-LCN, we compare it with
Kloc not less than Nloc . However, as shown in Table 2, the larger several diagnosis methods. In Method 1, five features specifically
Kloc is, the larger the architecture of NSAE-LCN is and the more designed for gearboxes [26], i.e., FM0, FM4, energy ratio (ER), side-
the time is spent. Thus, it is actually a tradeoff between higher band index (SI) and sideband level factor (SLF), five features com-
accuracies and less time when we choose the parameter Nloc monly used in gearbox diagnosis, i.e., root mean square, energy
and Kloc . operator, skewness, kurtosis and crest, and four energy ratios of
Another parameter, λ, is the regular parameter used in the wavelet coefficients are combined to characterize the health condi-
NSAE-LCN. The diagnosis accuracies using different values of λ are tions of the planetary gearbox. Based on these manual features, the
shown in Fig. 3. In this figure, Nloc and Kloc equal 100 and the ac- health conditions are classified by the softmax layer using these
curacy distributions of the 15 trials using different λ are clearly fourteen normalized features. In Method 2, the health conditions

Please cite this article as: F. Jia et al., A neural network constructed by deep learning technique and its application to intelligent fault
diagnosis of machines, Neurocomputing (2017), http://dx.doi.org/10.1016/j.neucom.2017.07.032
JID: NEUCOM
ARTICLE IN PRESS [m5G;July 25, 2017;6:10]

6 F. Jia et al. / Neurocomputing 000 (2017) 1–10

are classified by support vector machine (SVM) using these manual

features and the two parameters of SVM are optimized by the grid
search method. As displayed in Table 3, the diagnosis accuracies
of Method 1 are 86.20% for training and 85.87% for testing, and
the standard deviation values are 0.609% for training and 0.404%
for testing. The diagnosis accuracies of Method 2 are 97.14% for
training and 95.54% for testing, and the standard deviation values
are 0.451% for training and 0.216% for testing. Comparing Meth-
ods 1 and 2 with Method 6, it is illustrated that the features auto-
matically extracted by NSAE can help NSAE-LCN obtain higher and
more stable diagnosis accuracies in classifying the gearbox dataset.
In addition, the two compared methods require the manual fea-
ture extraction using human knowledge on the vibration signals of
planetary gearboxes, whereas NSAE learns the features automat-
ically from the raw vibration signals using less prior knowledge.
Thus, NSAE releases us from the tough work of designing feature
extraction algorithms.
In Method 3, principal component analysis is used for learning
features and SVM is used for classification. As shown in Table 3,
the training accuracy of the compared method using Method 3 is
100%, and the testing accuracy is 41.04% with 1.165% standard de- Fig. 4. ROC curves and AUCs of various methods for the gearbox dataset.
viation. It is seen that a serious overfitting phenomenon occurs
in Method 3. The main reason is that the features extracted by
PCA are linear, so they have limited abilities in characterizing the
health conditions of machines. In Method 4, stacked SAE and soft-
max layer constitute a deep neural network, whose training accu-
racy is 48.63% with 2.086% standard deviation and the testing ac-
curacy is 34.75% with 0.436% standard deviation. It indicates that
autoencoders are unsuitable for learning features from the raw vi-
bration signals directly because of the shift variant properties of
the signals. So in Method 5, SAE combined with the architecture of
Fig. 5. The encoder of NSAE.
LCN, namely SAE-LCN, is used for intelligent diagnosis of the gear-
boxes. The accuracies of SAE-LCN are 98.89% and 94.41%, and the
standard deviation values are 0.471% and 0.537%. It indicates that
k is the kth row vector of W , and f m, j
the LCN helps to produce shift-invariant features of vibration sig- where k = 1, 2, . . . , Kloc , Wloc loc k
m, j
nals and improve the diagnosis accuracies. By comparing Methods is the kth element of floc .
Eq. (23) indicates that the value of the
5 with 6, it shows that a certain over-fitting phenomenon occurs kth output neuron is obtained by the inner product operation of
j k . And the features in
in the training processes of SAE-LCN and the testing accuracies of the signal segment xm and the row vector Wloc
NSAE-LCN are higher than SAE-LCN, which demonstrates that NSAE Eq. (18) are just the nonlinear transformation form of the features
is able to learn better features than SAE. in Eq. (23). In this view, Wloc can be regarded as a set of normal
We also compare the receiver operating characteristic (ROC) orthogonal bases, and thus the local feature learning in the local
curves of Methods 1–6. ROC curves [27] are used to visualize the layer of LCN can be interpreted as decomposing the signal segment
performance of the classifier since they measure the ability of a in the normal orthogonal system Wloc .
classifier to produce good relative sample scores that serve to dis- We randomly select 15 row vectors of Wloc trained by NSAE
criminate positive and negative samples. Fig. 4 shows the ROC and plot them in Fig. 6(a). The corresponding frequency spectra of
curves of the methods. The best classifier will yield the point in these row vectors are displayed in Fig. 6(b). It can be seen that the
the upper left corner of the ROC space, representing no false neg- row vectors show local striking in the time domain and they have
atives and no false positives. An untrained classifier will achieve narrow spectral bandwidth in the frequency domain. So these row
the dashed line in the ROC space, representing the random perfor- vectors have the time-frequency properties and show similarities
mance of the classifier. To quantitatively measure the performance, to one-dimensional Gabor functions, which serve as good band-
we calculate the area under the ROC curve (AUC). So the bigger pass bases for mechanical signals. For comparison, we randomly
AUC of a classifier is, the better it will perform. As we can see, the plot 15 row vectors of Wloc trained by SAE in Fig. 6(c) and (d).
AUCs of NSAE-LCN are larger than other methods. It shows the su- It can be seen that these vectors also have some time-frequency
periority of the proposed NSAE-LCN in classifying the health condi- properties but the properties are too similar.
tions of the planetary gearbox under various operation conditions. We use WlocWloc T to measure the similarity of the row vectors

of NSAE and SAE and display the results in Fig. 7. It shows that the
4.1.4. Properties of NSAE-LCN inner product between the same row vector of NSAE approaches
Here, we try to understand the feature learning process of to 1 and the inner product between the two different vectors is
NSAE-LCN. As we know, a neural network is always viewed as a close to 0, which means that NSAE is able to encourage the learned
black box. And in the field of fault diagnosis, few papers attempt weight matrix to have clear and different patterns. However, the
to discover the patterns in the weights of autoencoders. So under- inner product results show that the different vectors of the weight
standing the properties of these patterns is extremely essential. As matrix of SAE have unclear and similar patterns. Such properties
shown in Fig. 5, we rewrite the encoder part of NSAE in Eq. (18) as prevent SAE from learning the various meaningful features. So SAE
the linear transformation form performs worse than NSAE. Considering the accuracies of Meth-
ods 5 and 6 in Table 3, the results demonstrate that the clearer
fkm, j = Wloc
k
· xmj (23) and more different the time-frequency properties of the trained

Please cite this article as: F. Jia et al., A neural network constructed by deep learning technique and its application to intelligent fault
diagnosis of machines, Neurocomputing (2017), http://dx.doi.org/10.1016/j.neucom.2017.07.032
JID: NEUCOM
ARTICLE IN PRESS [m5G;July 25, 2017;6:10]

F. Jia et al. / Neurocomputing 000 (2017) 1–10 7

to maximize the information about the frequency composition.

Wavelet or other time-frequency schemes are between the two ex-
tremes. For wavelet bases, the temporal width becomes narrower
as frequency increases, giving the bases the ability to capture the
self-similar structure of signals [28]. To analyze the time-frequency
properties of the row vectors of the weight matrix of NSAE, we use
the method described in Ref. [29] to tile these row vectors in the
time-frequency space, as shown in Fig. 8(a). It can be seen that
the distribution of these vectors is similar to the wavelet scheme:
the temporal width decreases as the bandwidth increases. This per-
spective can give us more insight to understand NSAE-LCN in the
fault diagnosis of machines and may inspire the researches to em-
ploy an alternative perspective on other neural networks used in
the fault diagnosis. Besides, in the previous analysis, we conclude
that Kloc should be equal or larger than Nloc . In other words, com-
plete or overcomplete representation [30] is better to be used in
LCN. Here, we use the tiling of the time-frequency plane to ex-
plain the reasons. In Fig. 8(a), the learned row vectors cover all
the range of times and frequencies and fulfill the time-frequency
plane when Kloc equals Nloc , producing the complete representa-
tion of the input signals. When Kloc is smaller than Nloc , the corre-
sponding time-frequency plane is shown in Fig. 8(b). It can be seen
that although the learned row vectors are approximately orthogo-
nal, their distribution cannot cover the ranges of low frequencies
in the time-frequency plane, leading to the lower diagnosis accu-
racies. So the complete representation strategy is chosen in this
paper.

Fig. 6. Row vectors of Wloc for the gearbox dataset: (a) the vectors of NSAE in the 4.2. Case study 2: fault diagnosis of motor bearings
time domain, (b) the vectors of NSAE in the frequency domain, (c) the vectors of
SAE in the time domain, and (d) the vectors of SAE in the frequency domain.
4.2.1. Data description
Here, we apply the proposed NSAE-LCN to diagnose the mo-
tor bearing dataset provided by Case Western Reserve University
weight matrix are, the better the methods would perform in the [31]. In the bearing dataset, the vibration signals were collected
intelligent fault diagnosis of machines. Therefore, NSAE encourages from the drive end of a motor in the test rig under four differ-
learned features to be meaningful and dissimilar so as to improve ent conditions: normal condition, outer race fault, inner race fault,
the classification accuracies. and roller fault. For each fault case, three different severity levels
One way to understand the time-frequency properties of the (0.18, 0.36, and 0.53 mm) were introduced. The test rig was oper-
weight matrix of NSAE is to tile them in a time-frequency plane. ated under four load conditions (0, 1, 2 and 3 hp) and the sampling
As we know, in the time-domain, frequency discrimination is com- frequency was 12 kHz. Therefore, this bearing dataset totally con-
pletely ignored to get the maximum temporal discrimination. And tains ten bearing health conditions under the four loads. There are
in the frequency domain, temporal discrimination is sacrificed 100 samples for each health condition under one load and each

T
Fig. 7. The results of WlocWloc for the gearbox dataset: (a) NSAE and (b) SAE.

Please cite this article as: F. Jia et al., A neural network constructed by deep learning technique and its application to intelligent fault
diagnosis of machines, Neurocomputing (2017), http://dx.doi.org/10.1016/j.neucom.2017.07.032
JID: NEUCOM
ARTICLE IN PRESS [m5G;July 25, 2017;6:10]

8 F. Jia et al. / Neurocomputing 000 (2017) 1–10

Fig. 8. Time-frequency analysis of Wloc of NSAE for gearbox dataset: (a) Kloc equals 100 and (b) Kloc equals 50.

Table 4
Classification comparison of the bearing dataset.

Methods Description No. of classes Training samples Testing accuracy

[32] EEMD features + SVM 6 25% 97.04%

[33] EMD-WKLFDA + SVM 10 40% 98.80%
[34] Bi-spectrum features + SVM 4 50% 96.98%
[35] MFDFA features + MDC 12 N/A 96.67%
[36] SAX based features + NN 10 50% 99.43%
[37] Multiple domain features + ensemble NNs 7 50% 99.07%
[38] Spectra + Deep neural networks 4 N/A 94.73%
Proposed NSAE-LCN 10 25% 99.92%

sample is a vibration signal of bearings containing 1200 data

points. Therefore, the dataset totally contains 40 0 0 samples and is
a ten-class classification problem.

4.2.2. Diagnosis results and comparisons

We use 25% of samples of the bearing dataset to train the pro-
posed NSAE-LCN and use the rest to test. The parameters of NSAE-
LCN used for the bearing dataset are the same as those used for
the gearbox dataset. The testing accuracy of NSAE-LCN is shown
in Table 4. It can be seen that NSAE-LCN obtains 99.92% testing
accuracy in classifying the bearing dataset. Fig. 9 shows the inner
product results of the weight matrix, and it can be seen that the
majority of the row vectors of the weight matrix are approximately
orthogonal.
Since the bearing dataset is a benchmark in mechanical fault
diagnosis, we compare the proposed NSAE-LCN with the methods
using the same dataset in the published papers, where all those
T
methods are thoroughly developed. The comparisons are shown Fig. 9. The results of WlocWloc of NSAE for the bearing dataset.
in Table 4. In Ref. [32], ensemble empirical mode decomposition
(EEMD) was used to extract features and an optimized SVM was
applied to classify the six bearing faults and finally 97.04% accuracy Based on symbolic aggregate approximation (SAX) features and a
was achieved. Van and Kang [33] integrated empirical mode de- neural network (NN), Georgoulas et al. [36] got 99.43% accuracy in
composition (EMD), wavelet kernel local fisher discriminant anal- classifying ten bearing health conditions. Xu et al. [37] used multi-
ysis (WKLFDA) and SVM to distinguish ten health conditions and ple domain features to characterize the health conditions of bear-
got 98.80% testing accuracy. Saidi et al. [34] used bi-spectrum fea- ings and applied ensemble fuzzy ARTMAP neural networks to rec-
tures to get features and SVM to classify four bearing health condi- ognize the health conditions, and they obtained 99.07% accuracy.
tions, and a classification accuracy of 96.98% was obtained. Lin and Lu et al. [38] applied deep neural networks for feature learning
Chen [35] proposed a method using multifractal detrended fluctua- and fault classification from the spectra of vibration signals and
tion analysis (MFDFA) to extract features and Mahalanobis Distance 94.73% accuracy was obtained. In the most of the methods above,
classifier (MDC) to diagnose faults, and they got 96.67% accuracy. plenty of researchers took their effort on manual feature extraction

Please cite this article as: F. Jia et al., A neural network constructed by deep learning technique and its application to intelligent fault
diagnosis of machines, Neurocomputing (2017), http://dx.doi.org/10.1016/j.neucom.2017.07.032
JID: NEUCOM
ARTICLE IN PRESS [m5G;July 25, 2017;6:10]

F. Jia et al. / Neurocomputing 000 (2017) 1–10 9

algorithms using signal processing techniques, like Fourier trans- [8] G. Georgoulas, V. Climente, J.A. Antonino-Daviu, I. Tsoumas, C. Stylios,
form, EEMD, high order spectrum and wavelet analysis. Although A. Arkkio, G. Nikolakopoulos, The use of a multilabel classification framework
for the detection of broken bars and mixed eccentricity faults based on the
such processes make full use of human knowledge on fault diag- start-up transient, IEEE Trans. Ind. Inform. 13 (2016) 625–634.
nosis, too much human labor would be spent in analyzing mechan- [9] M.D. Prieto, G. Cirrincione, A.G. Espinosa, J.A. Ortega, H. Henao, Bearing fault
ical signals and grasping their properties. As the collected data be- detection by a novel condition-monitoring scheme based on statistical-time
features and neural networks, IEEE Trans. Ind. Electron. 60 (2013) 3398–3407.
come larger, it is increasingly difficult to manually design the fea- [10] M. Amar, I. Gondal, C. Wilson, Vibration spectrum imaging: a novel bearing
ture extraction algorithm. Compared with these methods, the pro- fault classification approach, IEEE Trans. Ind. Electron. 62 (2015) 494–502.
posed NSAE-LCN integrates the feature learning and fault diagno- [11] D. Wang, K-nearest neighbors based methods for identification of different
gear crack levels under different motor speeds and loads: revisited, Mech. Syst.
sis in one network so as to extract features and classify the faults
Signal Process. 70 (2016) 201–208.
automatically, which would make the proposed method less de- [12] Y. Lei, Z. Liu, X. Wu, N. Li, W. Chen, J. Lin, Health condition identification of
pendent on prior knowledge and human labor. So the proposed multi-stage planetary gearboxes using a mRVM-based method, Mech. Syst. Sig-
nal Process. 60 (2015) 289–300.
method performs more intelligently compared with these methods
[13] N. Zeng, Z. Wang, H. Zhang, W. Liu, F.E. Alsaadi, Deep belief networks for quan-
and obtains higher classification accuracies in the bearing diagno- titative analysis of gold immunochromatographic strip, Cogn. Comput. 8 (2016)
sis case. 684–692.
[14] Y. Bengio, A. Courville, P. Vincent, Representation learning: a review and new
perspectives, IEEE Trans. Pattern Anal. Mach. Intell. 35 (2013) 1798–1828.
5. Conclusions [15] N. Zeng, H. Zhang, Y. Li, J. Liang, A.M. Dobaie, Denoising and deblurring gold
immunochromatographic strip images via gradient projection algorithms, Neu-
rocomputing 247 (2017) 165–172.
To release us from the task of designing feature extraction algo- [16] W. Liu, Z. Wang, X. Liu, N. Zeng, Y. Liu, F.E. Alsaadi, A survey of deep
rithms, a method using NSAE-LCN is proposed for intelligent fault neural network architectures and their applications, Neurocomputing 234
diagnosis of machines. In this method, LCN first uses NSAE to lo- (2017) 11–26.
[17] R. Thirukovalluru, S. Dixit, R.K. Sevakula, N.K. Verma, A. Salour, Generating fea-
cally learn various meaningful features from vibration signals, and ture sets for fault diagnosis using denoising stacked auto-encoder, in: Proceed-
then LCN produces shift-invariant features based on the learned ings of IEEE Conference on Prognostics and Health Management, 2016, pp. 1–7.
features and classifies the health conditions of machines. Since [18] F. Jia, Y. Lei, J. Lin, X. Zhou, N. Lu, Deep neural networks: a promising tool for
fault characteristic mining and intelligent diagnosis of rotating machinery with
NSAE-LCN uses vibration signals as the input and classifies me- massive data, Mech. Syst. Signal Process. 72 (2016) 303–315.
chanical health conditions automatically, this method does not rely [19] Z. Chen, W. Li, Multisensor feature fusion for bearing fault diagnosis using
on much human labor and can be used for the different diagnosis sparse autoencoder and deep belief network, IEEE Trans. Instrum. Meas. 66
(2017) 1693–1702.
cases directly. A gearbox dataset and a bearing dataset are used to [20] W. Mao, J. He, Y. Li, Y. Yan, Bearing fault diagnosis with auto-encoder extreme
verify the proposed NSAE-LCN. The results show that the learned learning machine : a comparative study, Proc. Inst. Mech. Eng. Part C: J. Mech.
features of NSAE are meaningful and dissimilar, and LCN helps to Eng. Sci. 231 (2017) 1560–1578.
[21] E. Li, P. Du, A. Samat, Y. Meng, M. Che, Mid-level feature representation via
produce shift-invariant features and recognizes mechanical health
sparse autoencoder for remotely sensed scene classification, IEEE J. Sel. Top.
conditions effectively. Through comparing with commonly used Appl. Earth Obs. Remote Sens. 10 (2017) 1068–1081.
diagnosis methods, the superiority of the proposed NSAE-LCN is [22] V. Nair, G.E. Hinton, Rectified linear units improve restricted Boltzmann ma-
chines, in: Proceedings of International Conference on Machine Learning, 2010,
verified.
pp. 807–814.
In this paper, the parameters of NSAE-LCN are determined man- [23] R. Memisevic, D. Krueger, Zero-bias autoencoders and the benefits of co-adapt-
ually based on the study of parameter selection. The automatic ing features, in: Proceedings of International Conference on Learning Represen-
selection of these parameters may make NSAE-LCN easily to be tations, 2015, pp. 1–11.
[24] D.C. Liu, J. Nocedal, On the limited memory BFGS method for large scale opti-
utilized in intelligent fault diagnosis of machines and even raise mization, Math. Program. 45 (1989) 503–528.
its performance. Thus, the optimization techniques such as PSO [25] A. Hyvärinen, E. Oja, Independent component analysis: algorithms and appli-
[39,40] will be used to adaptively adjust the parameters of NSAE- cations, Neural Netw. 13 (20 0 0) 411–430.
[26] P.D. Samuel, D.J. Pines, A review of vibration-based techniques for helicopter
LCN in future work. transmission diagnostics, J. Sound Vib. 282 (2005) 475–508.
[27] T. Fawcett, An introduction to ROC analysis, Pattern Recogn. Lett. 27 (2006)
861–874.
Acknowledgments [28] B.A. Olshausen, K.N. O’Connor, A new window on sound, Nat. Neurosci. 5
(2002) 292–294.
This research was supported by National Natural Science Foun- [29] M.S. Lewicki, Efficient coding of natural sounds, Nat. Neurosci. 5 (2002)
356–363.
dation of China (51475355 and 61673311) and National Program for [30] B.A. Olshausen, D.J. Field, Sparse coding with an overcomplete basis set: a
Support of Top-notch Young Professionals. strategy employed by V1? Vis. Res. 37 (1997) 3311–3325.
[31] X. Lou, K.a Loparo, Bearing fault diagnosis based on wavelet transform and
fuzzy inference, Mech. Syst. Signal Process. 18 (2004) 1077–1095.
References [32] X. Zhang, J. Zhou, Multi-fault diagnosis for rolling element bearings based on
ensemble empirical mode decomposition and optimized support vector ma-
[1] P.K. Wong, J. Zhong, Z. Yang, C.M. Vong, Sparse Bayesian extreme learning com- chines, Mech. Syst. Signal Process. 41 (2013) 127–140.
mittee machine for engine simultaneous fault diagnosis, Neurocomputing 174 [33] M. Van, H-J. Kang, Bearing defect classification based on individual wavelet
(2016) 331–343. local fisher discriminant analysis with particle swarm optimization, IEEE Trans.
[2] L. Guo, N. Li, F. Jia, Y. Lei, J. Lin, A recurrent neural network based health in- Ind. Inform. 12 (2016) 124–135.
dicator for remaining useful life prediction of bearings, Neurocomputing 240 [34] L. Saidi, J.B. Ali, F. Fnaiech, Application of higher order spectral features and
(2017) 98–109. support vector machines for bearing faults classification, ISA Trans. 54 (2015)
[3] F. Pacheco, J.V. de Oliveira, R.-V. Sánchez, M. Cerrada, D. Cabrera, C. Li, G. Zu- 193–206.
rita, M. Artés, A statistical comparison of neuroclassifiers and feature selection [35] J. Lin, Q. Chen, Fault diagnosis of rolling bearings based on multifractal de-
methods for gearbox fault diagnosis under realistic conditions, Neurocomput- trended fluctuation analysis and Mahalanobis distance criterion, Mech. Syst.
ing 194 (2016) 192–206. Signal Process. 38 (2013) 515–533.
[4] J. Liu, W. Wang, F. Golnaraghi, An enhanced diagnostic scheme for bearing con- [36] G. Georgoulas, P. Karvelis, T. Loutas, C.D. Stylios, Rolling element bearings di-
dition monitoring, IEEE Trans. Instrum. Meas. 59 (2010) 309–321. agnostics using the symbolic aggregate approximation, Mech. Syst. Signal Pro-
[5] P. Henriquez, J.B. Alonso, M.A. Ferrer, C.M. Travieso, Review of automatic fault cess. 60 (2015) 229–242.
diagnosis systems using audio and vibration signals, IEEE Trans. Syst. Man Cy- [37] Z. Xu, Y. Li, Z. Wang, J. Xuan, A selective fuzzy ARTMAP ensemble and its appli-
bern. A: Syst. 44 (2014) 642–652. cation to the fault diagnosis of rolling element bearing, Neurocomputing 182
[6] Y. Lei, F. Jia, J. Lin, S. Xing, S.X. Ding, An intelligent fault diagnosis method (2016) 25–35.
using unsupervised feature learning towards mechanical big data, IEEE Trans. [38] W. Lu, B. Liang, Y. Cheng, D. Meng, J. Yang, T. Zhang, Deep model based domain
Ind. Electron. 63 (2016) 3137–3147. adaptation for fault diagnosis, IEEE Trans. Ind. Electron. 64 (2017) 2296–2305.
[7] C. Shen, D. Wang, F. Kong, P.W. Tse, Fault diagnosis of rotating machinery based [39] N. Zeng, H. Zhang, W. Liu, J. Liang, F.E. Alsaadi, A switching delayed PSO op-
on the statistical parameters of wavelet packet paving and a generic support timized extreme learning machine for short-term load forecasting, Neurocom-
vector regressive classifier, Measurement 46 (2013) 1551–1564. puting 240 (2017) 175–182.

Please cite this article as: F. Jia et al., A neural network constructed by deep learning technique and its application to intelligent fault
diagnosis of machines, Neurocomputing (2017), http://dx.doi.org/10.1016/j.neucom.2017.07.032
JID: NEUCOM
ARTICLE IN PRESS [m5G;July 25, 2017;6:10]

10 F. Jia et al. / Neurocomputing 000 (2017) 1–10

[40] N. Zeng, Z. Wang, H. Zhang, F.E. Alsaadi, A novel switching delayed PSO algo- Liang Guo received the B.S. and Ph.D. degrees in me-
rithm for estimating unknown parameters of lateral flow immunoassay, Cogn. chanical engineering from Southwest Jiaotong University,
Comput. 8 (2016) 143–152. Chengdu, P.R. China, in 2011 and 2016, respectively. He
is currently working as a Postdoctoral researcher at the
State Key Laboratory for Manufacturing System Engineer-
Feng Jia is currently working toward the Ph.D. degree in ing, Xi’an Jiaotong University, Xi’an, P.R. China. His current
mechanical engineering at the State Key Laboratory for research interests include machinery condition monitor-
Manufacturing System Engineering, Xi’an Jiaotong Univer- ing, intelligent fault diagnostics and remaining useful life
sity, P.R. China. He received the B.S. and M.S. degree in prediction.
mechanical engineering from Taiyuan University of Tech-
nology, P.R. China, in 2011 and 2014, respectively. His re-
search interests include machinery condition monitoring
and fault diagnosis, intelligent fault diagnostics of rotat-
ing machinery. Jing Lin received his B.S., M.S. and Ph.D. degrees from
Xi’an Jiaotong University, P.R. China, in 1993, 1996 and
1999, respectively, all in mechanical engineering. He is
currently a Professor with the State Key Laboratory for
Manufacturing Systems Engineering, Xi’an Jiaotong Uni-
versity. From July 2001 to August 2003, he was a Post-
Yaguo Lei received the B.S. and Ph.D. degrees in me- doctoral Fellow with the University of Alberta, Edmonton,
chanical engineering from Xi’an Jiaotong University, Xi’an, AB, Canada, and a Research Associate with the Univer-
P.R. China, in 2002 and 2007, respectively. He is currently sity of Wisconsin–Milwaukee, Milwaukee, WI, USA. From
a Full Professor of mechanical engineering at Xi’an Jiao- September 2003 to December 2008, he was a Research
tong University. Prior to joining Xi’an Jiaotong University Scientist with the Institute of Acoustics, Chinese Academy
in 2010, he was a Postdoctoral Research Fellow with the of Sciences, Beijing, China, under the sponsorship of the
University of Alberta, Edmonton, AB, Canada. He was also Hundred Talents Program. His current research directions
an Alexander von Humboldt Fellow with the University of are in mechanical system reliability, fault diagnosis, and wavelet analysis. Dr. Lin
Duisburg-Essen, Duisburg, Germany. His research interests was a recipient of the National Science Fund for Distinguished Young Scholars in
focus on machinery condition monitoring and fault diag- 2011.
nosis, mechanical signal processing, intelligent fault diag-
nostics, and remaining useful life prediction. Dr. Lei is a Saibo Xing is currently working for the Ph.D. degree in
member of the editorial boards of more than ten journals, mechanical engineering from Xi’an Jiaotong University,
including Mechanical System and Signal Processing and Neural Computing & Appli- P.R. China and received the B.S. degree in material sci-
cations. He is also a member of ASME and a member of IEEE. He has pioneered ence and engineering from Xi’an Jiaotong University, in
many signal processing techniques, intelligent diagnosis methods, and remaining 2015. He graduated from Hsue-shen Tsien Experimental
useful life prediction models for machinery. Class majoring in material science and engineering, and
became the graduated student without examination ma-
joring in mechanical engineering. His research interests
focus on intelligent fault diagnostics and prognostics of
rotating machinery.

Please cite this article as: F. Jia et al., A neural network constructed by deep learning technique and its application to intelligent fault
diagnosis of machines, Neurocomputing (2017), http://dx.doi.org/10.1016/j.neucom.2017.07.032