Abstract—Aiming at the problems of low accuracy and slow diagnosis speed in the existing fault diagnosis model of
electrical machine bearing, this paper presents an electrical machine bearing fault diagnosis method based on Deep
Gaussian Process of particle swarm optimization(DGP). A total of 10 characteristics of 9 damage states and no fault
states of the bearing are determined, constructing a deep Gaussian process model for electrical machine bearing fault
diagnosis based on expectation propagation and Monte Carlo method, and use the particle swarm optimization
algorithm to perform parameter searching optimization for its induction point value. The experimental results show
that the fault recognition rate of DGP on the CWRU data set reaches 95%, significantly better than other deep learning
methods, integration methods and machine learning methods. DGP method can better diagnose electrical machine
bearing faults, provide technical support for the safe operation of the electrical machine which are important for real
industrial applications.
Keywords—Deep Gaussian Process; electrical machine fault diagnosis; particle swarm optimization.
Received: June 24, 2021. Revised: March 18, 2022. Accepted: April 21, 2022. Published: May 18, 2022.
1. Introduction
ith the deepening of machine learning research in the
field of artificial intelligence, machine learning
technology is increasingly used in the field of pattern
recognition[1]. Traditional recognition tasks mainly apply
machine learning models such as support vector machine
(SVM), neural network, and random forest [2]. In recent years,
deep learning has developed rapidly in academia and industry,
significantly improve the accuracy of recognition on many
traditional recognition tasks, demonstrates its superb ability to
handle complex recognition tasks, attracted a large number of
experts and scholars to conduct research on its theory and
application [3-5].
Electricity fault diagnosis technology can find faults in
electrical equipment at the early stage of fault diagnosis,
therefore, timely targeted maintenance can be carried out,
saving a lot of time and funds for repairing faults, while
avoiding production stalls, it also improves economic
efficiency. Also in the memory fault detection technology is also
necessary, Eitan Yaakobi [6] et al. proposed a structure of single
error correction WOM code with better WOM rate, this
structure can effectively update and store the data in the
memory. Liang Xi [7] et al. proposed a Multisource
Neighborhood Immune Detector Adaptive Model for Anomaly
Detection, solve various problems existing in the real-valued
shape-space under dynamic environment mentioned and
improve the overall detection performances, and got better
stability.
In today's production activities and daily life, the electrical
machine is the most important motive power and drive unit, and
it has been widely used in various fields of people's production
and life.
The fault detection of the electrical machine often needs to
detect the fault in a very short time, so as to carry out the
targeted maintenance in time, so it needs fast detection speed
and flexible detection method. The Gaussian Process has the
characteristics of low computational complexity and fast
convergence speed in a small sample space. The Gaussian
process is named after the German mathematician Carl
Friedrich Gauss to commemorate his proposal of the concept of
normal distribution, developed based on statistical learning
theory and Bayesian theory. In the following decades, rich
research results have been obtained. Ori Shental [8] et al.
proposed a Gaussian Belief Propagation Solver for Systems,
compared with the traditional method, this method has a faster
convergence speed. D Bickson [9] and others proposed a
Gaussian Belief Propagation Based Multiuser Detection
algorithm, compared with the previous formula, the new
Electrical Machine Bearing Fault Diagnosis Based on Deep
Gaussian Process Optimized by Particle Swarm
HAI GUO1,*, HAORAN TANG1, XIN LIU1, JINGYING ZHAO1, 2, LIKUN WANG3
1 College of Computer Science and Engineering, Dalian Minzu University,
18 LiaoheWest Road, Dalian Development Zone, Dalian, 116600, CHINA
2 Faculty of Electronic Information and Electrical Engineering, Dalian University of Technology,
Dalian, 116023, CHINA
3 College of Electronic and Electrical Engineering, Harbin University of Science and Technology,
Harbin 150080, CHINA
W
WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS
DOI: 10.37394/23201.2022.21.11
Hai Guo, Haoran Tang,
Xin Liu, Jingying Zhao, Likun Wang
E-ISSN: 2224-266X
100
Volume 21, 2022
algorithm reduces memory requirements, calculation steps and
the number of messages passed. The deep Gaussian process has
certain theoretical advantages and is suitable for the research of
electrical machine fault detection technology. The deep
Gaussian process model is a deep model that superimposes
multiple Gaussian processes, any number of Gaussian processes
can be superimposed. The Gaussian process controls the
mapping between layers and also has the advantages of the
Gaussian process. Zhao [10] et al. proposed Computer
Modeling of the Eddy Current Losses of Metal Fasteners in
Rotor Slots of a Large Nuclear Steam Turbine Generator Based
on Finite-Element Method and Deep Gaussian Process
Regression, the analysis results show that compared with the
independent finite-element analysis, this method reduces the
design cycle time and improves the design efficiency for a
large-capacity turbine generator. Guo [11] et al. proposed
Predicting Temperature of Permanent Magnet Synchronous
Motor(PMSM) Based on Deep Neural Network. This model
can effectively predict the temperature change of stator
winding, provide technical support to temperature early
warning systems and ensure safe operation of PMSMs. Wang
[12] et al. proposed Cuckoo Search Algorithm for
Multi-Objective Optimization of Transient Starting
Characteristics of a Self-Starting HVPMSM. Experiments show
that the optimization speed of this method is significantly faster
than other methods, and this method has a faster convergence
speed while ensuring accuracy.
Existing deep learning models have been able to diagnose
electrical machine faults well, but there are still many problems
such as insufficient accuracy and slow training speed. For
example, the semi-supervised training method of deep belief
network has the problem of slow training speed, the
autoencoder network has problems such as limited expression
features and difficulty in reconstruction, convolutional neural
network training requires a lot of data, and the effect is not ideal
when processing industrial signals, RNN has problems such as
gradient disappearance [13]. Therefore, this paper uses the
strong recovery ability of the deep Gaussian process to outliers
and the strong non-linear problem processing ability to
construct a deep Gaussian process classification model
optimized by particle swarm optimization, and apply it to the
fault diagnosis of electrical machine rolling bearings, and early
warning of motor bearing faults based on abnormal changes in
the signal before the fault occurs, so as to avoid electrical
machine damage and reduce losses caused by this model.
2. Deep Gaussian Process Classification
Model
2.1 Deep Gaussian process
For a given N observed values,
1,...,()
T
N
y y y
and
D-dimensional coordinates
1,...,()
T
N
X X X
, The output of
each hidden layer of a DGP model with L layers can be
expressed as
1
1
{}
L
ll
H
. The number of columns in Hi is the
number of nodes in layer L. It is also called the dimension of
the layer and can be written as Di.,which can be expressed by
equation (1).
1,1 1,
,1 ,
l
l
ll
D
l
ll
N N D
hh
H
hh
(1)
in:
,1
,()
l l i l
n i n
h f h
(2)
i
f
is given by Gaussian priors, usually for ease of
understanding, latent variable dimensions are ignored and
l
H
is written as
l
h
,and
,()
li
f
is written as
()
l
f
. First,
set a zero-mean Gaussian prior for
( | )
ll
pf
of each layer,
for layers with multiple nodes, the prior function is an
independent Gaussian function inside each layer. Assuming
that the noise of i.i.d can be parameterized in the output of
each layer, the prior and dependent variables of the deep
Gaussian process can be regarded as equation (3) and
equation (4):
( | ) ( | 0, ), 1, ,
l l l l
p f GP f K l L
(3)
1 2 1 2 0
1
( | , , ) ( | ( ), ),
N
l l l l l
l n l n l n n
n
p h f h hN f h h x
(4)
l
K
represents the kernel matrix between a given input and
layer L, where
11
( , )
l l l
K k h h
, For the layer with
1L
,
the input will no longer be a certain value, and the
corresponding output will not obey the normal distribution.
When
1L
, the model will become a shallow Gaussian
process model . Finally, the conditional probability of a
given target value in the output layer is shown in equation
(5).
1 2 1 2
1
( | , , ) ( | ( ), )
N
L L L
L n L n L
n
p y f f hNhy
(5)
Figure 1 is an example of a two-layer model, where a hidden
layer and an output layer are used for a two-dimensional
problem. The number of nodes
()
L
D
of the output layer will be
equal to the dimension of the observation value
n
y
of the
regression problem, or equal to the number of classes of the
classification problem. Like the shallow Gaussian process
model, adding appropriate sparse technology to the deep model
can effectively reduce the computational complexity of the deep
Gaussian process.
By omitting the dimension in the symbol and adding a
Gaussian prior to the induction point of each layer, the final
Sparse Depth Gaussian Process Model can be written as
equation (6)-equation (8).
WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS
DOI: 10.37394/23201.2022.21.11
Hai Guo, Haoran Tang,
Xin Liu, Jingying Zhao, Likun Wang
E-ISSN: 2224-266X
101
Volume 21, 2022
1
u ,u
(u | ) N(u | 0, ),l 1,...,L
ll
Ll
pK
(6)
12
h ,h
1
(h | u , h , ) (h | A u , K )
ll
nn
N
l l l l l l l
l n n n
n
p N Q
(7)
12
h ,h
1
( | u , h , ) (y | A u , K )
LL
nn
N
L L L L l
L n n n
n
p y N Q
(8)
Fig. 1 An example of a DGP
h1
Y
U1
U2
XZ0
Z1
Fig. 2 Gauss process model with two-layer depth sparseness
The sub-index of the covariance matrix k is their
corresponding output, for example,
u ,u
ll
K
represents the
covariance matrix of the induced point uL. It takes position
11
,(z ,z )
ll
ll
uu
Kk
as a parameter, the covariance matrix
,h
ll
h
K
of nodes in a layer, and uses the output of the previous
layer
11
,h (h ,h )
li
ll
h
Kk
to construct. We also define
matrices A and Q as equation (9) and equation (10):
1
,u u ,u
l l l l
n
l
nh
A K K
(9)
''
12
,u u ,u u ,
l l l l
nn
l
nl
hh
Q K K K
(10)
Figure 2 shows a 2-layer deep sparse Gaussian process
model. The hidden layer depends not only on the output of the
layer but also on the induced point variables.
2.1 Reasoning technology of deep Gaussian process
In the deep Gaussian process model, the output of inference
induction points
1
{u }L
ll
and hidden layer
1
{h }l
lL
is performed
by marginalizing latent variables, which can predict the
posterior probability of the test set and calculate the slight
possibility of hyperparameter adjustment. However, both
variables are difficult to handle. When considering a deep
Gaussian process model with L=2 layers, the joint distribution
of the model is shown in equation (11).
2
1 2 2 2 2 1 2 1 1 2
1 1 2 1
1
p , h ,{u } |{ , } , | u , h , h | u , , | ...(11)
l l l l
l l l l
y X p y p X p u
To simplify the description, all model parameters are now
grouped into equation (12).
1 2 2
0 1 1
{z } ,{ , }
ll
ll
(12)
The marginal possibility is obtained by the marginalization
2
1
{u }
ll
and the hidden variable
l
h
in equation (11) to obtain
equation (13).
1 2 1 2 1
1
y | , ,h ,{u } | ,
ll
p X p y X du du dh
(13)
However, some of the integrals in equation (13) are difficult
to handle because they involve calculating the covariance
function with respect to random variables [14]. Integration can
be achieved by joint distribution in extension, as shown in
equation (12). Starting from equation (10), a corresponding
distribution of the output of layer
11
h | hl
p
is obtained,
need to calculate a density-dependent nonlinear kernel
function (Nonlinear kernel function of density)
1
hl
.
Another thing that needs to be predicted is the posterior
distribution on the induction point, which also requires the
calculation of the integral (13) in the model evidence as shown
in equation (14).
2 1 2 1
11
1
{u } | , , ,h ,{u } | , ...(14)
|,
ll
ll
p X y p y X dh
p y X
This result can be generalized to the case of
L2
. For the
sake of simplicity, the layer dependency will be removed from
the symbol, and
1
u {u }
lL
l
and
1
1
{h }
lL
l
h
will be
abbreviated, for any number of layers, the following
(generalized) induction point becomes
u | , ,p X y
. In order
to calculate the marginal likelihood and posterior, approximate
reasoning techniques are needed.
Some work in the literature related to the deep Gaussian
process suggests the use of a general sampling algorithm [15] to
evaluate the logarithmic probability, the main difference from
the method explained in this section is that the sampling
algorithm does not set any distribution on the output (assuming
the induced output is fixed), so they are included as model
parameters. In this way, some of the benefits of regularization
are lost. They also proposed to train the model by maximum a
posteriori estimation (MAP), but the author did not compare the
method with any other state-of-the-art technology, and no
improvement was observed when the number of layers was
increased.
Another method explained in [16] involves using a stochastic
characterization vector
to approximate the kernel
function
,xkx
of the Gaussian process. The core can be
approximated as an inner product as shown in equation (15).
WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS
DOI: 10.37394/23201.2022.21.11
Hai Guo, Haoran Tang,
Xin Liu, Jingying Zhao, Likun Wang
E-ISSN: 2224-266X
102
Volume 21, 2022
,x T
k x x x
(15)
The results show that the deep Gaussian process model can
be regarded as a Bayesian neural network, and the output of the
layer is given by
g wx b
,
g
is the activation function,
w
is the probability distribution
pw
, and
b
is Bayesian
noise.
2.3 Deep Gaussian process based on expectation
propagation and Monte Carlo method
The deep Gaussian process based on the expectation
propagation and Monte Carlo method follows the method
explained in [17], and uses the binding factor constrained
expectation propagation algorithm to approach the posterior
inducing point. As shown in Figure 3, the process of calculating
ln n
Z
at a single point in a 2-layer deep Gaussian process
model. In the first level,
hl
q
is given by a normal distribution
(blue in the figure above) sampled from it. In the second layer,
the true distribution of
ln n
Z
(blue) is no longer a Gaussian
distribution, but is given by equation (16). The proposed
method calculates and propagates the samples through the
network (green) to make the model more flexible and able to
approximate the heterogeneous distribution.
The final form of
n
Z
approximated by s samples is given by
the Gaussian mixture as shown in equation (16).
1
1
1ˆ
|
SL
ns
s
Z q y h
S
(16)
Where
1
ˆL
s
h
represents
s th
samples from the
corresponding distribution
12
ˆ ˆ
( | )
LL
ss
q h h
, which can be
calculated by the above sampling technique.
Fig. 3 The deep gauss process model is an example of
network propagation
Contrary to the method proposed in [14], this method can
capture the complex dependencies between DGP layers. In the
literature [18], this method is also suitable for stochastic
gradient descent training, such as small batch training. The final
form of marginal likelihood approximation is shown in equation
(17).
11
( ) [(1 ) ( ) ( ) ( )] ln ... 17
B
Ll l l
q prior b
ln
N
F N N Z
B
Where α includes the model to be adjusted and the AEP
parameters, |B| is the selected mini-batch size and
b
Z
can be
calculated for each mini-batch using equation (16).
3. Particle swarm optimization optimizes
the deep Gaussian process classification
electrical machine rolling bearing fault
diagnosis model
The deep Gaussian process electrical machine bearing fault
diagnosis classification model constructed in this section uses
expectation propagation and Monte Carlo methods to
approximate the Gaussian posterior, and the particle swarm
algorithm is used to search the number of induced points in the
deep Gaussian process model in the range. The construction
model adopts a 5-layer network structure, namely the input layer,
the 3-layer hidden layer, and the output layer. The hidden layers
all use the square exponential kernel as the kernel function of
the Gaussian map, as shown in equation (18).
2
2
( , ) 2
SE
d
K x x l
exp -
(18)
The overall model network structure is shown in Figure 4:
x1
x2
x3
x4
x5
h11
h12
h13
h14
h15
f11(x)
f12(x)
f13(x)
f14(x)
f15(x)
h21
h22
h23
h24
f21(x)
f22(x)
f23(x)
f24(x)
h31
h32
h33
f31(x)
f32(x)
f33(x)
h41
h42
f41(x)
f42(x)
Y1
Y10
...
Inputs First hidden
layer
Second hidden
layer
Third hidden
layer Outputs
Fig. 4 Deep gauss process model network architecture
The particle swarm optimization algorithm (PSO) is a kind of
swarm intelligence algorithm, and its design is based on the
simulation of bird predation behavior. Assuming that there is
only one food in the target area (that is, the optimal solution in
the optimization problem), the goal of the flock of birds is to
find this food source. Throughout the entire search process, the
birds communicate with each other to let other birds find their
position, and through collaboration, they can judge whether
they find the optimal solution, and at the same time, they can
also obtain the information of the optimal solution. Passed to the
entire flock of birds, and finally the entire flock of birds can
gather around the food source, that is, we have found the
optimal solution, that is, the problem converges [19].
The particle swarm optimization algorithm simulates the
birds in a flock of birds by designing a massless particle. The
particle has only two attributes: speed V and position X. Speed
represents the speed of movement, and position represents the
direction of movement. Each particle searches for the optimal
solution individually in the search space, which is recorded as
the current individual extreme Pbest, and the individual extreme
value is shared with other particles in the entire particle swarm,
find the optimal individual extreme value as the current global
optimal solution Gbest of the entire particle swarm. All particles
WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS
DOI: 10.37394/23201.2022.21.11
Hai Guo, Haoran Tang,
Xin Liu, Jingying Zhao, Likun Wang
E-ISSN: 2224-266X
103
Volume 21, 2022
in the particle swarm adjust their speed and position according
to the current individual extremum Pbest found by themselves
and the current global optimal solution Gbest shared by the entire
particle swarm. The idea of particle swarm optimization
algorithm is relatively simple, mainly divided into: 1. Initialize
the particle swarm; 2. Evaluate the particles, that is, calculate
the fitness value; 3. Find the individual extreme value; 4. Find
the global optimal solution; 5. Modify the speed and position of
the particles. The particle swarm optimization algorithm is used
to search the induction point value of the deep Gaussian process
model, and the optimal parameter setting of the model is
determined to improve the accuracy of model classification.
The specific process is shown in Figure 5.
Start
Initialize population size, position and
speed
Calculate the current individual extremum
of particles at all levels and find the current
global optimal solution of the whole
particle swarm
Initialization depth Gaussian process model
Updates the speed and position of
individual particles
Whether the termination
conditions are met
Output optimal solution
End
N
Y
Fig. 5 Particle swarm optimization
4. Results and discussion
4.1 Sample Database
The test data comes from Case Western Reserve University
(CWRU) Rolling Bearing Data Center. The CWRU data set is a
world-recognized standard data set for bearing fault diagnosis.
At present, the bearing fault diagnosis algorithm is updated
quickly. In order to evaluate the superiority of the algorithm
proposed in this chapter, all experiments use CWRU bearing
data.
The CWRU bearing center data acquisition system is shown
in Figure 6. The test object of this test is the drive end bearing in
the picture. The model of the bearing to be diagnosed is the deep
groove ball bearing SKF6205, which is manufactured by EDM
under the load of 0HP, 1HP, 2HP, and 3HP. The sampling
frequency of the system is 12kHz. There are a total of 3 types of
defects in the diagnosed bearing, which are rolling element
damage, outer ring damage and inner ring damage. The
diameter of the damage is 0.007inch, 0.014inch and 0.021inch,
and the specific information of 9 kinds of damage states is
shown in the table. 4.1. In the experiment, 2048 data points are
used for diagnosis each time. In order to facilitate the training of
the deep Gaussian network, each segment of the signal is
normalized, and the equation of the normalized processing is
shown in (19).
min
max min
xx
xxx
(19)
Fig.6 Data acquisition system of CWRU rolling bearing
Select 10000 pieces of data under 0HP load as shown in
Table 4.1. There are 1000 pieces of data for each type of fault,
including 900 training samples and 100 test samples. The
training sample adopts the data enhancement method as shown
in Figure 7. The length of the training sample collected each
time is 2048, the offset is 1, and there is no overlap between the
test samples.
Fig.7 Data enhancement
Table 4.1 Test data set description
*Damage diameter is 0.007 inch (inch), **Damage diameter is
0.014 inch,***Damage diameter is 0.021 inch
4.2 Experimental conditions and evaluation
indicators
The CPU used in the simulation experiment in this section is
Intel i7-7130U, the memory is 4GB RAM, the programming
language used is Python, and the framework used is Tensorflow,
Keras and sklearn. Using 10,000 signals under zero load in the
CWRU data set as samples, they are divided into 9 types of
faulty bearings and 1 type of non-faulty bearings, totaling 10
categories. Each type of fault contains 1000 pieces of data, 90%
WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS
DOI: 10.37394/23201.2022.21.11
Hai Guo, Haoran Tang,
Xin Liu, Jingying Zhao, Likun Wang
E-ISSN: 2224-266X
104
Volume 21, 2022
of which are taken as the training set, and the remaining 10% as
the test set. The accuracy, precision, recall and F1-Score under
macro-average are used as the evaluation indicators of the
model. The specific formula is as follows.
In the classification problem, to analyze the effect of data and
classifiers, evaluation indexes can be used for auxiliary analysis.
In this paper, the following evaluation indexes are used to
comprehensively analyze and discuss the experimental results.
Accuracy is the most primitive evaluation index in
classification problems. The definition of accuracy is the
percentage of correct results in the total sample. The equation is
shown in (20):
TP TN
Accuracy TP TN FP FN
(20)
in:
True Positive (TP): Positive samples predicted to be
positive by the model;
False Positive (FP): Negative samples predicted to be
positive by the model;
False Negative (FN): Positive samples predicted to be
negative by the model;
True Negative (TN): Negative samples predicted to be
negative by the model.
Precision rate is the probability of the actual positive samples
among all the predicted positive samples, which can be
expressed by equation (21).
TP
Precision TP FP
(21)
Recall rate is the probability of being predicted to be a
positive sample in a sample that is actually positive, which can
be expressed by equation (22).
FNTP
TP
call
Re
(22)
F1-score is a weighted average of the precision and recall of
the model. The closer F1-score is to 1, the better the empirical
effect is. The evaluation index f1-Score can be expressed by
equation (23).
12
Pr Re
2Pr Re
ecision call
Fecision call
(23)
The specific parameters of the deep Gaussian process
electrical machine rolling bearing fault diagnosis model are set
as follows: the maximum iteration times is 500, the minimum
batch_size is 100, the learning rate is 0.01, and each node
contains The number of samples is 15, and the noise level is set
to 1e-5. On this basis, the particle swarm optimization algorithm
is used to search for the number of induced points in the range of
[10,100]. The parameters of the particle swarm optimization
algorithm are set as follows: population size is set to 100, the
maximum number of iterations is 150, and the inertia factor is
set to 2, and the weight factor is set to 0.5.
4.3 Experimental results and discussion
Under 10,000 data sets of electrical machine rolling bearings
under 0HP, the deep Gaussian process model based on particle
swarm optimization algorithm is used to classify faults. When
the induction point is 50 and the number of iterations is 40
times, the fault diagnosis classification accuracy reaches the
highest 0.95. Under the same experimental conditions, the
results of comparison with deep learning models such as Deep
RNN LSTM DNN DGP
0.0
0.2
0.4
0.6
0.8
1.0
Recognition accuracy
Model
Accuracy
Precision
ReCall
F1-Score
Fig.8 The accuracy of deep gauss process classification
model compared with other deep learning models
SGD KNN LR DT SVC GaussianNB DGP
0.0
0.2
0.4
0.6
0.8
1.0
Recognition accuracy
Model
Accuracy
Precision
ReCall
F1-Score
Fig.9 Comparison of accuracy between deep gauss
process classification model and other machine learning
algorithms
RF AdaBoost Bagging ET GB DGP
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Recognition accuracy
Model
Accuracy
Precision
ReCall
F1-Score
Fig.10 The accuracy of the depth gauss process
classification model was compared with other ensemble
learning models
WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS
DOI: 10.37394/23201.2022.21.11
Hai Guo, Haoran Tang,
Xin Liu, Jingying Zhao, Likun Wang
E-ISSN: 2224-266X
105
Volume 21, 2022
Neural Network DNN, Recurrent Neural Network RNN, and
long short-term memory network (LSTM) are as follows:
Shown in Figure 8.
It can be seen from Figure 8 that the accuracy of the deep
Gaussian process for bearing fault diagnosis under the samples
used in this chapter is up to 0.95, while LSTM and RNN also
maintain high accuracy rates of 0.93 and 0.88 respectively, the
accuracy of deep neural network is 0.74, and the accuracy of
deep Gaussian process model is higher than that of the above
deep learning model.
In the same experimental environment, machine learning
algorithms such as stochastic gradient descent (SGD), k-nearest
neighbor (KNN), decision tree (DT), support vector machine
(SVC), Gaussian NB and logistic regression (LR) are
compared.
The experimental results are shown in Figure 9. The
classification accuracy of the deep Gaussian process fault
diagnosis model is much higher than other commonly used
machine learning algorithms.
Compared with ensemble learning algorithms such as
RandomForest (RF), AdaBoost, Bagging, ExtraTree (ET) and
GradientBoosting(GB) in the same experimental environment,
the experimental results are shown in Figure 10. The
classification accuracy of the deep Gaussian process model for
bearing faults is higher than that of the above ensemble learning
algorithm, which is more suitable for the fault diagnosis of
electrical machine bearings under large samples.
5. Conclusion
A fault diagnosis classification model of deep Gaussian
process electrical machine rolling bearing based on particle
swarm optimization is proposed. The basic components and
structural parameters of the deep Gaussian process model are
introduced. The parameter propagation formula based on
expected propagation and Monte Carlo method is derived. The
proposed model is trained and tested on the CWRU rolling
bearing data set. The fault recognition rate of the trained model
on the test set can reach 95 %, which is higher than that of other
machine learning, ensemble learning and deep learning
algorithms. It can better diagnose the electrical machine bearing
fault and provide technical support for the safe operation of the
motor.
Acknowledgment
This work was supported only by the Science Foundation of Ministry
of Education of China (No.18YJCZH040).
References
[1] Zhang X Y, Bengio Y, Liu C L, “Online and offline
handwritten Chinese character recognition: A
comprehensive study and new benchmark,” Pattern
Recognition, vol. 61, no. 61, pp. 348-360, 2017.
[2] Xie Z, Sun Z, Jin L, et al., “Learning Spatial-Semantic
Context with Fully Convolutional Recurrent Network for
Online Handwritten Chinese Text Recognition,”. IEEE
Transactions on Pattern Analysis and Machine Intelligence,
vol. 40, no.8, pp. 1903-1917, 2018.
[3] J. Thiagarajan V B, “Designing Accurate Emulators for
Scientific Processes using Calibration-Driven Deep
Models,” Nature Communications, vol. 11, no.1, pp.
5622–5632, 2020.
[4] Sagheer A K M, “Unsupervised Pre-training of a Deep
LSTM-based Stacked Autoencoder for Multivariate Time
Series Forecasting Problems,”. Sentific Reports, vol. 9, pp.
19038, 2019.
[5] Kim R G, Doppa J R, Pande P P, et al, “Machine Learning
and Manycore Systems Design: A Serendipitous
Symbiosis,” Computer, vol. 51, no. 7, pp. 66–77, 2018.
[6] Y Aa Kobi E, Siegel P H, Vardy A, et al., “Multiple
error-correcting WOM-codes,” IEEE International
Symposium on Information Theory. IEEE, vol. 58, no. 4,
pp. 2220-2230, 2012.
[7] O. Shental, P. H. Siegel, J. K. Wolf, D. Bickson and D.
Dolev, “Gaussian belief propagation solver for systems of
linear equations,” 2008 IEEE International Symposium on
Information Theory, pp. 1863-1867, 2008.
[8] D. Bickson, D. Dolev, O. Shental, P. H. Siegel and J. K.
Wolf, “Gaussian belief propagation based multiuser
detection,” 2008 IEEE International Symposium on
Information Theory, pp. 1878-1882, 2008.
[9] L XI*, Rui-Dong Wang, et al., “Multi-source neighborhood
immune detector adaptive model for anomaly detection,”
IEEE Transactions on Evolutionary Computation, vol. 25,
no. 3, pp. 582-594, 2021.
[10] Jingying Zhao, Hai GUO, Likun Wang, Min Han,
“Computer Modeling of the Eddy Current Losses of Metal
Fasteners in Rotor Slots of a Large Nuclear Steam Turbine
Generator Based on Finite Element method and Deep
Gaussian Process Regression,” IEEE Transactions on
Industrial Electronics, vol. 67, no. 7, pp. 5349-5359, 2020.
[11] L. Wang, H. Guo*, F. Marignetti, C. D. Shaver and N.
Bianchi, “Cuckoo Search Algorithm for Multi-Objective
Optimization of Transient Starting Characteristics of a
Self-Starting HVPMSM,” IEEE Transactions on Energy
Conversion, vol. 36, no. 3, pp. 1861-1872, 2021.
[12] Guo, Hai, Qun Ding, Yifan Song, Haoran Tang, Likun
Wang, and Jingying Zhao. 2020. “Predicting Temperature
of Permanent Magnet Synchronous Motor Based on Deep
Neural Network,” Energies, vol. 13, no. 18, pp. 4782,2020.
WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS
DOI: 10.37394/23201.2022.21.11
Hai Guo, Haoran Tang,
Xin Liu, Jingying Zhao, Likun Wang
E-ISSN: 2224-266X
106
Volume 21, 2022
[13] Jia F G L, Lei Y, “A neural network constructed by deep
learning technique and its application to intelligent fault
diagnosis of machines,” Neurocomputing, vol. 272, no. 10,
pp. 619–628, 2017.
[14] Damianou A C L N D, “Deep Gaussian Processes”,
Computer Science, pp. 207–215, 2012.
[15] Depeweg, S. et al., “Learning and Policy Search in
Stochastic Dynamical Systems with Bayesian Neural
Networks,” Machine Learning, 2016.
[16] Cutajar K M P, Bonilla E V, “Random Feature Expansions
for Deep Gaussian Processes,” Machine Learning, no. 70,
pp. 884–893, 2016.
[17] Bui T D, Hernandez-Lobato D, Li Y, et al., “Deep Gaussian
Processes for Regression using Approximate Expectation
Propagation,” 33rd International Conference on Machine
Learning, vol. 48, pp. 76-85,2016.
[18] Salimbeni H D M, “Doubly Stochastic Variational
Inference for Deep Gaussian Processes,” Neural
information processing systems, no. 30, pp. 4588–4599,
2017.
[19] J. Kennedy and R. Eberhart, “Particle swarm
optimization,” Proceedings of ICNN'95 - International
Conference on Neural Networks, vol.4, pp. 1942-1948,
1995.
Hai Guo received the B.S. in Electronic
Engineering from the Heilongjiang
University, in 2000, the M.S. in Pattern
Recognition and Intelligent Systems
from the Kunming University of Science
and Technology, in 2004, and the Ph.D.
degree in Material Science from the
Harbin University of Science and
Technology (HUST), Harbin, China.
Since 2010, he has been an Associate
professor with the College of Computer Science and
Engineering, Dalian Minzu University. He has authored over 30
articles in international journals and conference proceedings.
His current research interests include pattern recognition and
their applications.
Haoran Tang recrived the B.S. degrees
from College of Mathematical, Shanghai
Normal University, Shanghai, China, in
2018, the M.S. degree in Computer
Science and Engineering, Dalian Minzu
University in 2020.He current research
interests include artificial intelligence,
machine learning and electrical
engineering.
Xin Liu received the B.S. degrees from
College of Computer, Cangzhou Normal
University, Hebei, China, in 2020. He is
currently working towards the M.S. degree
at Computer Science and Engineering,
Dalian Minzu University.He current
research interests include deep learning
and artificial intelligence.
Jingying Zhao received the B.S. and M.S.
degrees from School of Computer Science
and Technology, Changchun University of
Science and Technology, Jilin, China, in
2000 and 2003. Since 2013, she has been
an Associate professor with the College of
Computer Science and Engineering, Dalian
Minzu University. She is currently working
towards the Ph.D. degree at Faculty of
Electronic Information and Electrical Engineering, Dalian
University of Technology, Liaoning, China. Her current
research interests include pattern recognition and machine
learning method and their applications.
Likun Wang(M’17) received the B.Sc.,
M.Sc., and Ph.D. degrees in electrical
machinery and appliance from the Harbin
University of Science and Technology
(HUST), Harbin, China, in 2010, 2013,
and 2015, respectively. Since 2017, he has
been working as a Postdoctoral Fellow
with the Institute of Electromagnetic and
Electronic Technology, Harbin Institute of
Technology, Harbin. Since 2018, he has
been an Associate Professor of Electrical Machinery and
Appliance with the College of Electrical and Electronic
Engineering, HUST. His research interests include synthesis
physical fields and dynamic operation mechanism of electrical
machines and its system. Dr. Wang was the recipient of the first
prize of science and technology progress of colleges and
universities of Heilongjiang province in 2019.
WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS
DOI: 10.37394/23201.2022.21.11
Hai Guo, Haoran Tang,
Xin Liu, Jingying Zhao, Likun Wang
E-ISSN: 2224-266X
107
Volume 21, 2022
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Conflict of Interest
The authors have no conflicts of interest to declare
that are relevant to the content of this article.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
This work was supported only by the Science Foundation of Ministry
of Education of China (No.18YJCZH040).
