Measurement: Xiaoxun Zhu, Dongnan Hou, Pei Zhou, Zhonghe Han, Yiming Yuan, Weiwei Zhou, Qianqian Yin

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Measurement 138 (2019) 526–535

Contents lists available at ScienceDirect

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

Rotor fault diagnosis using a convolutional neural network with


symmetrized dot pattern images
Xiaoxun Zhu, Dongnan Hou, Pei Zhou, Zhonghe Han ⇑, Yiming Yuan, Weiwei Zhou, Qianqian Yin
School of Energy Power and Mechanical Engineering, North China Electric Power University, Baoding 071003, Hebei, People’s Republic of China

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

Article history: Vibration failure is a common problem in most rotating machinery, and vibration fault diagnosis is an
Received 6 January 2018 important means of ensuring stable equipment operation. The present work proposes a rotor vibration
Received in revised form 26 January 2019 fault diagnosis approach that transforms multiple vibration signals into symmetrized dot pattern (SDP)
Accepted 10 February 2019
images, and then identifies the SDP graphical feature characteristic of different vibration states using a
Available online 13 February 2019
convolutional neural network (CNN). SDP images reveal different vibration states in a simple and intu-
itive manner. In addition, a CNN can reliably and accurately identify vibration faults by extracting the fea-
Keywords:
ture information of SDP images adaptively through deep learning. The proposed approach is tested
Deep learning
Convolutional neural networks
experimentally using a rotor vibration test bed, and the results obtained are compared to those obtained
Symmetrized dot pattern with an equivalent CNN-based image recognition approach using orbit plot images. The rotor fault diag-
Fault diagnosis nosis precision is improved from 92% to 96.5%.
Rotor Ó 2019 Elsevier Ltd. All rights reserved.

1. Introduction machine learning approaches, such as neural network and support


vector machine (SVM) algorithms [8,9]. However, the accuracy of
Vibration failure is a common problem for all heavy machinery, conventional machine learning approaches depends heavily on
particularly for rotating machinery. Vibration faults are directly the specific characteristics of the extracted signal features and
related to increased operation and maintenance costs and can even lacks the process of adaptive feature learning. Moreover, the accu-
affect the level of safety and result in system-wide failures, which racy of the diagnosis model is limited due to insufficient learning
can have substantial repercussions, particularly for critical infras- depth.
tructure like electric power plants employing heavy rotating In recent years, the emergence of deep learning has made sig-
machinery. The avoidance of system failures requires the assess- nificant contributions toward solving the problem of feature adap-
ment of equipment condition via vibration fault diagnosis through tive learning [10]. Among the available deep learning approaches,
monitoring vibration signals collected by sensors located at various the application of convolutional neural networks (CNNs) to vibra-
locations [1]. tion fault diagnosis has recently become a focus of research. For
Vibration fault diagnosis is a process of fault feature extraction example, a vibration fault identification model based on one-
and pattern recognition [2,3] essentially. First, vibration signals dimensional (1D) CNN has been proposed [11,12]. Here, the vibra-
obtained from, e.g., rotating machinery, must be transformed into tion signal time series was directly used as the model input, and
useful information via signal processing. Time-domain analysis the vibration state was obtained by the direct learning of the signal
and frequency-domain analysis [4] have been widely employed characteristics representative of vibration fault states. Due to the
in vibration fault diagnosis owing to the simplicity and intuitive- complexity and non-linearity of the vibration signal, it is difficult
ness of these approaches. However, the analysis of only time and to directly input the CNN network to extract the effective features,
frequency domain information is not sufficient for nonlinear and which affects the diagnostic accuracy. In addition, vibration fault
non-stationary vibration signals, which require a combined time- diagnosis techniques based on CNN and image recognition have
frequency analysis approach [5–7]. Primary signal features can also been investigated. Zeng and Li [13] proposed a time-
then be extracted on the basis of signal analyses, and those feature frequency spectrogram image recognition method based on a
characteristic of specific vibration states can be identified by CNN, and accordingly established a vibration fault diagnosis model
of an automobile gearbox. However, the time-frequency image is
overly dependent on the time-frequency analysis method. The
⇑ Corresponding author. time resolution and frequency resolution of time-frequency images
E-mail address: hanzhonghe@ncepu.edu.cn (Z. Han).

https://doi.org/10.1016/j.measurement.2019.02.022
0263-2241/Ó 2019 Elsevier Ltd. All rights reserved.
X. Zhu et al. / Measurement 138 (2019) 526–535 527

cannot be guaranteed, and the diagnostic accuracy is affected. Fig. 1 illustration of r(i) as the polar radius, and h(i) and u(i) as
From the perspective of pattern recognition algorithms, the axis the two angles of the conventional polar space by which any point i
orbit, which reveals characteristic information regarding vibration in a time-series signal X = {x1, x2, . . ., xi, . . ., xn} is mapped onto a
faults in the form of plot images, and it is an important method for point S[r(i), h(i), u(i)] in polar coordinates in conjunction with a
visualizing the vibration signals of rotating machinery, has been plotting gain n (n  h).
employed in conjunction with a deep learning algorithm to xi  xmin
develop a rotating machinery fault diagnosis algorithm [14]. How- rðiÞ ¼ ð1Þ
xmax  xmin
ever, the recognition rate of axis orbit plot images is substantially
reduced for low signal-to-noise ratio (SNR) signals. Moreover, the xiþq  xmin
center offset and rotation of the orbit plot images introduce diffi- hðiÞ ¼ hs þ n ð2Þ
xmax  xmin
culties for CNN identification. In an approach related to orbit plot
image recognition, the application of symmetrized dot pattern xiþq  xmin
uðiÞ ¼ hs  n ð3Þ
(SDP) images for the identification of vibration states [15] is a rel- xmax  xmin
atively new approach that combines signal processing and image
In which, xmax and xmin are the greatest and the smallest values
analysis. SDPs are particularly useful because they represent differ-
in X ,hs is the sth rotation angle of a specified mirror symmetry
ent vibration states in a simple and intuitive manner. However,
plane, which is given as hs = 360 S/N, where s = 1, 2, . . ., n and n is
existing studies have adopted the conventional fault matching
the number of mirror symmetry planes, q is the time lag coeffi-
method for fault identification, which provides insufficient learn-
cient, and n is the gain of the plotting (n  h). This signal represen-
ing depth with a correspondingly reduced diagnostic accuracy.
tation technique was first conceived for the visual characterization
The present work combines the advantages of SDP images and
of speech waveforms, and is particularly effective for low SNR
CNNs in image recognition and feature learning to develop a rotary
signals.
machine vibration fault diagnosis approach. Here, a vibration sig-
From the above analysis, we note that a SDP image derived from
nal is transformed into a SDP image, and the graphical features
a signal is n-mirrored in polar coordinates according to angle hs.
of SDP images characteristic of different vibration states are
This indicates that each mirrored image can represent the charac-
learned and identified through a CNN to realize vibration fault
teristics of the signal. In addition, a SDP diagram of N signals col-
diagnosis. The proposed approach is tested experimentally using
lected from different positions can also be displayed in the same
a rotor vibration test bed, and the results obtained are compared
polar coordinates to realize the fusion of information. Considering
to those obtained with an equivalent CNN-based image recognition
a periodic sinusoidal signal as an example ðyðtÞ ¼ sinð2  p  50ÞÞ,
approach using orbit plot images. The results demonstrate the
where the sampling frequency is 1 kHz, resulting in 1000 sampling
validity and significant contribution of the approach.
points, the SDP provides 6 mirrors in polar coordinates at 60° inter-
vals, as illustrated in Fig. 2.

2. SDP images of time-series signals 2.2. SDP images of simulated signals

2.1. Overview of SDP The distinct advantages of SDP signal representations are illus-
trated here using four simulated signals representative of different
Changes in the amplitude and frequency of time-series signals vibration states, including a simple periodic signal (Signal A),
can be depicted by means of an easy-to-understand visual repre- superimposed periodic signals (Signals B and C), and a complex
sentation by mapping a normalized time waveform into SDPs on superimposed periodic signal including a periodic impulse compo-
a polar graph to produce SDP images. For a time series signal X = nent (Signal D). The simulation signals are given as
{x1, x2, . . ., xi, . . ., xn}, any point i in X can be mapped onto a point
S[r(i), h(i), u(i)] in polar coordinates, where r(i) is the polar radius,
and h(i) and u(i) are the two angles of the conventional polar space,
as shown in Fig. 1. This mapping is conducted according to Eqs.
(1)–(3).

Fig. 1. Principle of SDP method. Fig. 2. SDP image.


528 X. Zhu et al. / Measurement 138 (2019) 526–535

xðtÞ ¼ hðt Þ þ sðt Þ þ nðt Þ: ð4Þ ences in size or shape, although human experts can easily
discriminate between the shapes presented by SDP patterns
In which, h(t) represents aperiodic signal consisting of K = 4
robustly. As a result, SPD images remain manually monitored by
possible periodic components with amplitudes of Ai and frequen-
human operators in many industrial settings. Therefore, a machine
cies of fi, and is given as
learning method to autonomously identify different SDP patterns
X
K generated by rotor to establish a more robust and automatic vibra-
hðtÞ ¼ Ai sinð2  p  f i  tÞ; ð5Þ tion fault monitoring system are proposed.
i¼1

In addition, s(t) represents the following periodic impact com-


3.1. Deep learning and CNNs
ponent [16], consisting of M single component periodic impact sig-
nals of equivalent amplitude B:
Deep learning neural networks employ a hierarchical model
X
M structure that is similar to that of the biological cognitive processes
sðt Þ ¼ Bexp½bðt  mT p Þ  cos½2pf r  ðt  mT p Þuðt  mT p Þ; of animals. The features of the input data are extracted step by step
m¼1
through a series of levels to form an abstract high-level feature
ð6Þ representation. A CNN is a truly multi-layer structured deep learn-
where b is the attenuation coefficient, Tp is the impact interval, fr is ing model architecture developed for two-dimensional image
the resonance frequency, and u() is a unit step function. Finally, n (matrix) analysis [17] that employs relative spatial relationships
(t) represents a noisy signal to simulate random interference and pro- among image pixels to reduce the required number of parameters.
vide an SNR of 30 for all signals. The sampling frequency is 1 kHz, The essence of a CNN is to autonomously learn several feature fil-
resulting in 1000 sampling points over a 1 s period. The variables of ters (or kernels) that can extract and process the features hidden in
s(t) are set using the parameters M = 14, B = 1.5, b = 420, fr = 420 Hz, the input image data layer by layer. The hierarchical structure of
and Tp = 1/30 s. The remaining parameters of the four simulation sig- the CNN employed in the present study is illustrated in Fig. 5,
nals are listed in Table 1. The four simulation signals were trans- and its functionality can be summarized as follows.
formed into SDP images using the parameters q = 10, n = 6, and
n = 30°, and the SDP images are shown in Fig. 3. We note from the fig- (1) Input layer. The image to be identified is input to the CNN,
ure that the SDP images clearly exhibit characteristics representative and then the features of the image are extracted and learned
of the different vibration states of the simulated signals. in a training process to determine the kernels to be subse-
We note from Fig. 3 that, for a mirror surface, the characteristics quently employed for feature identification.
of a single periodic signal SDP image is a crescent shape, the char- (2) Convolutional layer (C layer). The C layer employs a number
acteristics of superimposed periodic signals are a smaller crescent- of kernels that are individually convolved with the previous
shaped figure within a larger figure, and superimposed periodic layer to form a corresponding number of feature maps. Each
signals with a periodic impact component are characterized by convolutional kernel is applied to the previous layer accord-
an irregular pattern. ing to the local receptive fields of the individual neurons,
which all employ the same kernel, and a feature map is gen-
erated in conjunction with a nonlinear activation function
2.3. SDP images of low SNR signals
f(). Although the receptive fields of individual neurons are
limited, these overlap with the receptive fields of neighbor-
Compared with the orbit plot method, the SDP method is a par-
ing neurons, which lends a high degree of size and shift
ticularly effective technique for visualizing low SNR signals. To
invariance to the subsequent identification results. Specifi-
illustrate this advantage of SDP images, both SDP and orbit plot
cally, the input uli and the output xli for the jth neuron at layer
images for a single periodic signal (h(i) = 60°) with different SNR
l for each kernel of the C layer can be expressed as follows:
values are plotted in Fig. 4.
The results clearly indicate that the orbit plot image representa- !
tion expresses no identifying signal characteristics for an SNR of   X l l
xlj ¼ f uj ¼ f
l
xi  kij þ bj ;
l1
ð7Þ
30, whereas the corresponding SDP image expresses the essential
i2Ri
characteristics of the signal. Accordingly, the following conclusions
can be drawn. The proposed method is more suitable for low SNR where xl1 is the output of the ith neuron at layer l  1, klij is
i
signals than methods based on conventional orbit plot images. the kernel weight from the ith neuron at layer l  1 to the
jth neuron at layer l, and blj is the scalar bias of the jth neuron
3. Recognition of vibration states in SDP images based on a CNN at layer l. The output of the C layer can be considered as a sin-
gle layer with a depth dimension corresponding to the num-
Although SDPs contain significant information regarding the ber of feature maps.
condition of rotating machinery, these images are presently not
well utilized in typical plant applications because of their compli- (3) Subsampling layer (S layer). The S layer is employed to
cated shape patterns arising from various causes. Moreover, the reduce the dimension of a feature map representation, cre-
numerical definition of features represented by specific SDP pat- ate invariance to small shifts and distortions, and avoid
terns is not readily accomplished when considering subtle differ- over-fitting. The output can be given as

Table 1
Parameters of simulated time-series signals for illustrating SDP images.

A1 f1 (Hz) A2 f2 (Hz) A3 f3 (Hz) A4 f4 (Hz) s(t)


A – – 0.6 50 – – – – N
B 0.2 25 0.6 50 – – – – N
C – – 0.8 50 0.6 100 – – N
D – – 0.8 50 0.2 100 0.2 150 Y
X. Zhu et al. / Measurement 138 (2019) 526–535 529

Fig. 3. SDP images for different simulated signals based on Eqs. (4)–(6) and the parameters in Table 1.

l
xlj ¼ gðulj Þ ¼ gðblj  downðxl1
i Þ þ bj Þ; ð8Þ (5) Output layer. In classification tasks, the feature vector
extracted from the CNN is classified and identified, and a
value representative of an identification probability is out-
where g() is the activation function in the S layer, blj is the
put for each class. The Softmax classifier was implemented
weight coefficient for the jth neuron at layer l, and down()
in this layer [20].
is the subsampling (or pooling) operation. Accordingly, new
features are obtained by pooling processes using commonly
The CNN uses the back propagation (BP) algorithm to imple-
employed pooling methods such as mean-pooling, max-
ment the training of network parameters [21]. The main parame-
pooling, and stochastic-pooling [18,19].
ters for the optimization of the CNN include the convolutional
kernel weight parameters k of the C layer, the weight coefficient
(4) Fully connected layer (FC layer). All neurons in the FC layer l
b of the S layer, the weight coefficient w of the FC layer, and the
are fully connected to the neurons of both layers l  1 and l
bias values b corresponding to each layer. Optimization of the net-
+ 1. The output of this layer can be given as
work parameters of each layer is conducted according to an objec-
tive function denoted as the sensitivity parameter, which is given
   
xlj ¼ r ulj ¼ r wlj xl1
l
þ bj ; ð9Þ as
j

1 X 2
where r() and wlj are the activation function and weight EP ¼ ypj  opj ; ð10Þ
2 j
coefficient of the FC layer, respectively.
530 X. Zhu et al. / Measurement 138 (2019) 526–535

Fig. 4. SDP (above) and orbit plot (below) images for a single periodic signal with different signal-to-noise ratio (SNR) values.

where opj is the actual output and ypj is the ideal output of the j-th The SDP images with an image matrix of 128 * 128 was composed
layer. Here, the objective is to obtain an actual CNN output value by N signal fusion. Randomly selected part of the SDP image data
that approaches the ideal output. set as the training set of CNN. CNN can find more complicated
structures from complicated image and extract high-dimensional
3.2. Vibration state recognition method and more abstract representations and further, help computer to
establish complicated concept. Thus, the trained CNN model can
The method illustrated in Fig. 6 is proposed for identifying perform the classification of SDP image. To improve the recogni-
vibration states inherent in SDP images through the extraction tion accuracy of CNN in SDP image recognition, the remaining
and learning of SDP image features using the CNN. In the proposed SDP image data set is taken as the test set which is classified by
recognition method, we fuse the vibration data collected by m sen- using CNN model. By comparing the classification results, the effi-
sors at different positions as SDPs. All n signals are arranged in the ciency of the CNN model in the features extraction of SDP images
same polar coordinates according to a designated rotation angle. can be checked.
X. Zhu et al. / Measurement 138 (2019) 526–535 531

Fig. 5. Hierarchical structure of the CNN employed in the present work.

Fig. 6. Proposed vibration state recognition method based on SDP images and the CNN.

4. Experimental simulated by adjusting the bearing oil pressure of the equipment.


In the experiment, the rotating speeds were about 1000 rpm,
4.1. Experimental setup 1500 rpm, 1800 rpm and 3000 rpm respectively. We obtained
400 samples (sample length is 1024) for each rotor vibration fault
To better demonstrate the improved accuracy offered by the state, where 300 were employed as training samples and 100 were
superior information fusion ability of SDP images, vibration signals employed as test samples. The CNN employed in this study (Fig. 5)
were obtained using 4 eddy current sensors installed at different consisted of two C layers (C1 and C3), two S layers (S2 and S4), and
positions of a Bently RK4 rotor vibration test table, as shown in a FC layer and output layer. C1 and C3 employed 6 and 12 kernels,
Fig. 7, to obtain vibration fault data for four rotor vibration fault respectively, each with a uniform kernel size of 5, and the subsam-
states (unbalanced, oil whirl, static and dynamic rubbing, and pling factor employed in the down() operations of S2 and S4 was
misalignment). In the experiments, sensor #1 was perpendicular set to 2. The rectified linear unit (ReLU) was selected as the activa-
to sensor #2 and sensor #3 was perpendicular to sensor #4 with tion function in fully connected layers.
respect to the longitudinal rotor axis, and the signal data was nor-
malized. The sampling frequency was 1 kHz. The mass imbalance
of the rotor is caused by the counterweight screws of different 4.2. Experimental results and performance evaluation
mass. The mass counterweight screws of 0.1 g, 0.2 g, 0.4 g, 0.8 g,
1.0 g, 1.2 g, 1.6 g, 2.0 g and their combinations are installed in the Representative waveforms obtained from the four sensors for
different phases of the rotary mass block to simulate the quality the four separate vibration fault states are shown in Fig. 8, with
unbalance faults; there is a friction bolt (copper and plastic) on the parameters time lag l = 10, angle h = 90°, angular gain n = 30°,
the friction device. Firstly, the distance between the bolt and the and the corresponding SDP images are shown in Fig. 9. As an image
rotor is fixed to make it close. Turn on the motor and then, the rub- recognition tool, CNN has been greatly developed in recent years
bing experiment under different conditions is simulated by differ- and many models are proposed based on it, such as lenet-5, Alex-
ent bolts; When the motor is turned on, the Oil whirl fault can be Net, VGG, ResNet, GoogLeNet, etc., which are used to deal with
more complicated problems and greatly improve the accuracy of
image recognition. In order to solve the problem of vibration fault
recognition, this paper firstly proposed the SDP diagram of multi-
sensor information fusion, which enhanced the difference between
different fault types, and then, based on this, the CNN structure
was designed. In the design process, this paper investigated several
CNN models and found that the more complex CNN model (such as
VGG) can absolutely realize the fault signal of SDP image recogni-
tion. However, the training time is longer with an increased diffi-
culty in model parameter adjustment. Therefore, in this paper,
the model was designed on the basis of lenet-5 to obtain the model
structure, and the structure of the CNN model is shown in Table 2.
Fig. 7. Bently RK4 rotor vibration test table employed in the experimental testing. Compared with lenet-5, it is simpler and it short one fully con-
532 X. Zhu et al. / Measurement 138 (2019) 526–535

(a) Unbalanced (b) Oil whirl

(c) Rubbing

Fig. 8. Representative waveforms obtained from the four sensors for the four different vibration states.

nected layer. In our experiments, considering the four kinds of fault adversely affected. The recognition results for the orbit plot images
classification with a less type classification than lenet5, The CNN are listed in Table 4. The experimental results for this method yield
network with one fully connected layer can achieve the same clas- an overall recognition accuracy of 93.8%, indicating that the pro-
sification accuracy, and also can effectively improve the computing posed recognition approach based on SDP images yields a higher
speed. On the other hand, in order to select the network, a compar- recognition accuracy than that based on orbit plot images.
ison between the lenet-5 and similar VGG network model are con- In order to compare the identification accuracy of CNN and
ducted and results show that, the accuracy rate of VGG model and methods in Ref. [15] in SDP image, the test set used in this paper
lenet5 model is basically in the same level (the accuracy of not was diagnosed using the failure matching based on the clustering
optimized VGG model is even lower than the classification results templates. First, all SDP images of the test set were converted into
of lenent5 model) and the optimal parameters were selected by image gray matrix, and 30 SDP images under each fault state were
trial and error method to identify SDP images. selected as the fault template. The characteristic value of the clus-
Some degree of noise of the signals derived from the experi- tering center is calculated by summing the image matrixes of all
mental station is observed. The recognition results for the pro- the templates of each running state of rotor and then calculate
posed method on the 400 test samples of all four vibration fault the average value. The threshold value was assigned as 0.75 with
states are listed in Table 3. The experimental results yield an over- is the same with that in Ref. [15]. The diagnosis of the remaining
all recognition accuracy of 96.5%, which represents high reliability. 280 SDP images was shown In Table 5 with an accuracy rate of
For comparison, orbit plot images were also constructed using 91.4%.
the data obtained from two of the perpendicularly placed sensors Keep 0.75 as constant, increasing the number of fault templates
(i.e., sensors #1 and #2) for the four vibration fault states, and as that in Ref. [15], taking 50 SDP images of each fault (50% of the
the results are shown in Fig. 10. From this comparison, it was test set) as the fault template for clustering analysis. The diagnosis
observed that each orbit plot image presents the vibration charac- of the remaining 200 SDP images was shown in the Table 6 with an
teristics obtained from only a single position. Meanwhile, a single accuracy rate of 92%.
SDP image can display the vibration characteristics of different SDP diagram shows some unique features, the method in Ref.
positions, which clearly demonstrates the advantages of SDP [15] can easily complete the fault recognition. However, if these
images in terms of information fusion. Then, the CNN was similarly fault templates do not appear, the method in Ref. [15] is difficult
applied for vibration fault recognition based on the obtained orbit to identify. Compared with method in Ref. [15], the features of
plot images. Orbit plot images require some preprocessing, such as weight sharing enable CNN to extract high-dimensional and more
centralization and reorientation, prior to conducting vibration fault abstract representation and further improve the recognition
state recognition [14]. Otherwise, the recognition accuracy will be accuracy.
X. Zhu et al. / Measurement 138 (2019) 526–535 533

Fig. 9. SDP images corresponding to the waveforms given in Fig. 6 for the different vibration states.

Table 2 5. Conclusion
The structure of the CNN in this paper.

Layer Convolution kernels (High  width)/stride Feature maps


This paper proposed an approach for autonomously detecting
the fault vibration states of rotor based on a CNN using SDP images
Input 128  128
generated from the time-series data obtained from multiple sen-
C1 5  5/1 124  124  6
S2 2  2/2 62  62  6 sors. The proposed approach was tested experimentally on a rotor
C3 5  5/1 58  58  16 vibration test bed, and the results obtained were compared to
S4 2  2/2 29  29  16 those obtained with an image recognition approach based on an
FC 84
equivalent CNN using orbit plot images. Results demonstrate that
Output 4
the developed CNN can extract image features automatically

Table 3
Confusion matrix obtained using the proposed method.

Classification result
Unbalanced Oil whirl Rubbing Misalignment
Ground truth Unbalanced 99 1 0 0
Oil whirl 1 98 1 0
Rubbing 1 0 94 5
Misalignment 0 1 4 95
534 X. Zhu et al. / Measurement 138 (2019) 526–535

Fig. 10. Orbit plot images corresponding to the waveforms of only sensors #1 and #2 given in Fig. 6 for the different vibration states.

Table 4
Confusion matrix obtained using orbit plot images for CNN based recognition.

Classification result
Unbalanced Oil whirl Rubbing Misalignment
Ground truth Unbalanced 96 2 1 1
Oil whirl 4 94 0 2
Rubbing 2 0 92 6
Misalignment 1 1 5 93

Table 5
Confusion matrix obtained using the Ref. [15] method (280 SDP images).

Classification result
Unbalanced Oil whirl Rubbing Misalignment
Ground truth Unbalanced 65 0 1 4
Oil whirl 1 62 6 1
Rubbing 1 1 66 2
Misalignment 3 2 2 63

Table 6
Confusion matrix obtained using the Ref. [15] method (200 SDP images).

Classification result
Unbalanced Oil whirl Rubbing Misalignment
Ground truth Unbalanced 46 1 0 3
Oil whirl 1 44 3 2
Rubbing 0 2 47 1
Misalignment 2 1 0 47
X. Zhu et al. / Measurement 138 (2019) 526–535 535

through deep feature learning, and can recognize image features [5] L.I. Shun, J. Wen, A model-based fault detection and diagnostic methodology
based on PCA method and wavelet transform, Energy Build. 68 (2014) 63–71.
associated with a variety of fault vibration states accurately. How-
[6] Z.H. Han, X.X. Zhu, W.H. Li, A false component identification method of emd
ever, compared to the fault vibration state recognition accuracy of based on kullback-leibler divergence, Proc. CSEE 32 (11) (2012) 112–117.
the CNN using orbit plot images, the accuracy was significantly [7] C.Y. Yang, T.Y. Wu, Diagnostics of gear deterioration using EEMD approach and
improved by employing SDP images owing to the ability of this PCA process, Measurement 61 (2015) 75–87.
[8] S.J.S. Hakim, H.A. Razak, S.A. Ravanfar, Fault diagnosis on beam-like structures
representation to fuse information obtained from multiple mea- from modal parameters using artificial neural networks, Measurement 76
suring points, and the more obvious representation of the features (2015) 45–61.
of different vibration states provided by SDP images, even for low [9] W.D. Jiao, S.S. Lin, Overall-improved fault diagnosis approach based on support
vector machine, Chin. J. Sci. Instrum. 8 (2015) 1861–1870.
SNR signals. This is also proved by comparing the diagnostic accu- [10] J. Schmidhuber, Deep learning in neural networks: an overview, Neural
racy of CNN with the fault matching based on clustering template. Networks 61 (2015) 85–117.
The experimental results therefore verify that the fault vibration [11] T. Ince, S. Kiranyaz, L. Eren, et al., Real-time motor fault detection by 1-D
convolutional neural networks, IEEE Trans. Industrial Electron. 63 (11) (2016)
state recognition approach proposed in this paper is more effective 7067–7075.
and reliable than a similar approach employing orbit plot images [12] O. Abdeljaber, O. Avci, S. Kiranyaz, et al., Real-time vibration-based structural
and the fault matching based on clustering template. damage detection using one-dimensional convolutional neural networks, J.
Sound Vib. 388 (2017) 154–170.
[13] X.Q. Zeng, J. Li, Time-frequency image recognition based on convolutional
Acknowledgments neural network, Mach. Electron. 5 (2016) 006.
[14] H. Jeong, S. Park, S. Woo, et al., Rotating machinery diagnostics using deep
learning on orbit plot images, Procedia Manuf. 5 (2016) 1107–1118.
The Fundamental Research Funds for the Central Universities
[15] X. Xu, H. Liu, H. Zhu, et al., Fan fault diagnosis based on symmetrized dot
(2018MS111). pattern analysis and image matching, J. Sound Vib. 374 (2016) 297–311.
[16] X.M. Chen, D.J. Yu, X. Li, et al., Early rub-impact diagnosis of rotors using
References morphological component analysis, J. Vib. Eng. 27 (3) (2014) 466–472.
[17] C. Liang, X.M. Deng, M.Q. Zhou, et al., Convolutional neural networks in image
understanding, Acta Autom. Sin. 42 (9) (2016) 1300–1312.
[1] F. Jia, Y. Lei, J. Lin, et al., Deep neural networks: a promising tool for fault [18] M.D. Zeiler, R. Fergus. ‘‘Stochastic pooling for regularization of deep
characteristic mining and intelligent diagnosis of rotating machinery with convolutional neural networks.” arXiv preprint arXiv:1301.3557 (2013).
massive data, Mech. Syst. Sig. Process. 72 (2016) 303–315. [19] Y.L. Boureau, F. Bach, Y. Lecun, et al. ‘‘Learning mid-level features for
[2] Z. Yin, Q. Chen, Y. Zhang, Deep learning and its new progress in object and recognition.” 2010 IEEE Computer Society Conference on Computer Vision
behavior recognition, J. Image Graphics 19 (2) (2014) 175–184. and Pattern Recognition, CVPR 2010. 2010.
[3] G.Q. Zhao, Q.Q. Ge, X.Y. Liu, et al., Fault feature extraction and diagnosis [20] S. Mei, J. Ji, J. Hou, et al., Learning sensor-specific spatial-spectral features of
method based on deep belief network, Chin. J. Sci. Instrum. 39 (7) (2016) hyperspectral images via convolutional neural networks, IEEE Trans. Geosci.
1946–1953. Remote Sens. 55 (8) (2017) 4520–4533.
[4] L. Ou, D.J. Yu, C.T. Wang, Correlation spectrum of Laplacian eigenvectors and its [21] P. Dayan, L.F. Abbott, Theoretical Neuroscience, 806, MIT Press, Cambridge,
applications to rolling bearing fault diagnosis, J. Vibr. Eng. 27 (5) (2014) 748– MA, 2001.
754.

You might also like