Induction Motor Fault Diagnosis Based on Fuzzy Support Vector

Machine

Shuying Li

College of Electrical and Power Engineering Shanxi University, Taiyuan

Keywords: Induction motor, Fuzzy Support Vector Machine, Membership function, Wavelet analysis, Fault diagnosis.

Abstract: In order to solve the problem of correctly identifying fault classes in induction motor fault diagnosis and

improve the accuracy of the classification, a novel fault diagnosis method of the classification model based

on the Fuzzy Support Vector Machine (FSVM) is proposed in this paper. The fault pattern classifier was

trained, which the fuzzy membership of the feature vectors is computed by a controllable factors algorithm

membership function to overcome the sensitivity to noise and outliers. After the stator current was sampled,

the fault feature was extracted from the sampling data through wavelet analysis and used as the input of the

FSVM. A multi-class fault classifier was constructed to identify different faults, which was based on one to

one strategy and mixed matrix combination. Experiment results show that the Fuzzy Support Vector Machine

(FSVM) has good performance for classification over non-linear and high dimension and small sample sets.

This method improves the accuracy in rotor fault diagnosis.

1 INTRODUCTION

The induction motor has been widely used in many

industries, with its simple structure, low cost and

durability. However, due to heavy load or frequent

starting/braking, the connected region of rotor bars

and end rings is prone to breakage and other failures.

Therefore, early detection of the induction motor

rotor fault is significant value. Since the 1990s, many

scholars have applied artificial intelligence

techniques into induction motor fault diagnosis and

achieved remarkable results.

Papers (QIU Arui, 1999; Dong Jianyuan, 1998;

Filippetti F, 2000) show us better diagnostic results in

induction motor fault diagnosis technology, using

Radial basis function (RBF) neural network, BP

neural network and fuzzy neural network approach.

Essays (Cao Zhitong, 2004; Chen Liyuan, 2006; Fang

Ruiming, 2007) proposed motor fault diagnosis

technology based on the support vector machine

method, helping to deepen and develop research into

motor fault diagnosis technology so as to achieve a

better recognition effect.

Support vector machine (SVM) is a general-

purpose machine learning method based on statistical

learning theory, which effectively solved learning

problems, such as small samples, high dimensions

and nonlinearity (VAPNI V, 1995), overcoming the

difficulty of determining reasonable structure and the

presence of local optima in artificial neural network

learning. But SVM has some limitations, such as

sensitive reaction to the noise or outlier in training

samples, the accuracy of classification of samples not

completely belonging to one of the two categories

were low. The samples collected during fault

diagnosis often contain noise and outliers, these

samples including "abnormal" information are often

located near classification plane in the feature space,

thus affecting the diagnostic results (Computer

Engineering and Applications, 2009).

In this case, we adopted a fuzzy support vector

machine (FSVM) fault diagnosis method based on

subordinate degree function determining method of

controllable factor, different samples with different

weights punishment, determined the subordinate

degree of training sample, eliminate the effects of

noise and outliers in fault diagnosis. We input the

training set into FSVM classification method, trained

to get the fault diagnosis model, then input test set

into the fuzzy SVM model that sufficient trained to

achieve the identification of different fault types.

588

Li, S.

Induction Motor Fault Diagnosis Based on Fuzzy Support Vector Machine.

In 3rd International Conference on Electromechanical Control Technology and Transportation (ICECTT 2018), pages 588-592

ISBN: 978-989-758-312-4

2 FUZZY SUPPORT VECTOR

Taiwan scholars Chun and Shen proposed the concept

of fuzzy support vector machines, fuzzy logic was

introduced into the standard support vector machines,

assigned each sample a subordinate degree value,

adopted different penalties weighting factor for

different samples. In addition to the features and

generic identity of training samples, but also added

the fuzzy subordinate degree, showed the degree of

the sample belongs to a class, with a unique value.

Discriminant function corresponding to the

optimal classification plane is:

(1)

Where, is the kernel function.

Transfer the complex inner product on high

dimensional feature space, which was difficult to

achieve, into a simple function calculation easily to

implement on low-dimensional space.

3 FAULT DIAGNOSIS METHOD

BASED FSVM

3.1 The Basic Principle of Fault

Diagnosis

The basic principle of fault diagnosis and processes is

shown in Figure 1. After the signal processing of

collected information, computing the fuzzy

subordinate degree of training sample point, obtained

the trained fuzzy support vector machine network

model, then applicate FSVM fuzzy classification

rules in the fault diagnosis process.

Figure 1 Fault diagnosis procedures based on FSVM

3.2 Determine the Subordinate Degree

Function

The basic idea of Fuzzy Support Vector Machine is

to endow subordinate degree belongs to a certain

class to different sample, subordinate degree reflects

the importance of sample to its training, select the

appropriate subordinate degree function in a given

problem directly affects the classification results.

Under normal circumstances, the principle of

determine the subordinate degree is the relative

importance of the sample in the class, according to the

actual problem, or need, select the appropriate fuzzy

subordinate degree function to calculate the

subordinate degree of each sample point.

Given sample set , assumed positive class

contains p samples , The average

for its category as the central positive class, negative

class contains q samples , The

average for its category as the central negative

class, .

Thus the center distance between two types,

Txx







, the distance of each positive sample

from the positive class center,

rxx





, t h e

  

sgn ,

xSV

xwKxxb



















1, 2,

ip 







1, 2,

jq 



















Induction Motor Fault Diagnosis Based on Fuzzy Support Vector Machine

589

distance of each negative sample from the negative

class center,

rxx





Assuming the sample point of the positive class

and the negative class can be included in a

hypersphere, respectively, the radius of hypersphere

where include the positive and negative class can be

calculated as follows:







max







，







max







(2)

Radius control factor



, w h e r e





, g i v e n a

small positive number



previously, as the noise and

isolated points of subordinate degree, and its function

can be obtained as follows:

if y





































































(3)

Where: is a small positive number, in order to

ensure .The subordinate degree values of

support vector improved in this way, meanwhile, also

reduces the value of subordinate degree values of

noise points.

According to experience of numerical experiment.



0.2 0.4







 

4 INDUCTION MOTOR FAULT

DIAGNOSIS SYSTEM BASED

ON FSVM

4.1 System Configuration

In this paper, several typical rotor failure experiments

are simulated using a rotor test rig. Stator current data

was collected for each failure mode is used as fault

samples at the rated speed. The hardware of the

diagnostic system structure is shown in Figure 2,

including current signal detection, signal processing

and FSVM diagnostic.

Figure 2 Configuration of fault diagnosis

1) Fault signal detection:

After stator current is collected by the Hall current

sensors, it is Converted into an AC voltage signal

between -3 ~ + 3V through Signal Conditioners.

Subsequently, it is converted to a digital signal by the

A / D conversion. The data is transferred to the

computer through the parallel port communications.

2) Feature extraction (signal processing):

To analysis stator current spectrum using wavelet

transform, and then to extract the fault feature, which

is used as the input of the FSVM system.

3) Fault identification (learning phase):

To study the sample, and achieving the fault

category learning, recognition.

4) Fault diagnosis (testing phase):

To complete the motor fault determination and

classification using FSVM system, diagnostic results

are displayed through man-machine interface.

4.2 Experiment and Analysis of the

System

4.2.1 Acquisition of the Data Sample

Two poles motor is used as a test prototype, 1. 5 kW,

50 Hz, 220 V. Eddy current dynamometer is used as

the motor load. Under the load stability conditions (s

= 0.06), the stator current signal are taken

respectively measured when the following states:

①

no fault

②

squirrel-cage rotor continuous two

broken bars

③

a fracture of the rotor end ring

④

rotor

static eccentricity

⑤

rotor dynamic eccentricity

⑥

broken rotor bar and eccentricity faults occur

simultaneously.

The sampling frequency is 4096Hz and each data

length of the sample is 512 points. Each failure mode

selected 40 sets of data samples, a total of 240 sets of

data, consisting of six training sample set.

4.2.2 Experiment and Analysis

(1) Feature extraction

After getting the training data, analysis the current

signal by wavelet transform. Wavelet energy

eigenvalues are extracted and the corresponding

samples was constructed.

(2) Select the penalty factor C and kernel function

Penalty factor C is used to describe the balance

between the largest classification border and the

classification error during the process of training.

When the greater the C is, the correct the

classification results of training samples is. But at the

same time the generalization ability decreased. In the

fuzzy support vector machine, the penalty factor is

fuzzy. The different penalty factor is chose according



s 

ICECTT 2018 - 3rd International Conference on Electromechanical Control Technology and Transportation

590

to the different samples, indicating the importance of

the samples during training SVM. The grid search and

the cross-validation method was used to get the

penalty factor C = 78.

The selection of the kernel function has not yet

formed a unified, effective rules, the most commonly

used kernel functions include linear, polynomial,

Gaussian and sigmoid kernel function. In this paper,

the Gaussian radial basis function is used the kernel

function.

(, ) exp

Kxx















,Parameters



。

(3)The fusion strategy of the sub-classifier output

Induction motor rotor fault diagnosis is a multi-

class classification. After the various sub-classifier

training is completed, the proper integration of the

various sub-categories is needed in order to obtain the

result of classification.

The voting decisions, binary tree, neural network

method and mixed matrix method is commonly used

the Fusion algorithm. The different fusion strategy

has the greater impact on the classification results. In

this paper, mixed matrix method can achieve more

satisfactory accuracy, consuming far less than the

neural network method.

(4) Diagnostic Analysis

For each failure mode, 10 samples were taken as

the training samples in order to establish the diagnosis

FSVM model.

The training sample is input the model, and the

correct diagnosis ratio was 100%.The classification

can be completely correct. Secondly, the test samples

(30 samples for each fault) are inputted FSVM

network in order to training model, At last , the

mixing matrix method is used to judge the output of

the model, determining the sample belong to what

type.

Table 1: Fault diagnosis result

Failure mode Diagnostic results

SVM FSVM

Normal 28 28

Broken bars(or end ring

fracture)

55 58

Eccentric(static or dynamic) 50 54

Broken bars and Eccentric 24 26

Accuracy (%) 87.2 92.5

The results can be inferred from the calculation:

the correct diagnosis ratio of the standard SVM was

only 87.2%.This shows the effect of a simple

diagnostic based on the SVM method is not ideal. The

fault diagnosis accuracy rate (92.5%) based on the

FSVM significantly improved. This verifies the

effectiveness and feasibility of the fault diagnosis

method based on FSVM.

5 CONCLUSIONS

This paper presents a new method, which fuzzy

support vector machine applied to the induction

motor fault diagnosis. The fuzzy support vector

machine classifier as a failure mode, using class mean

distance to determine the fuzzy membership

functions, therefore, it can distinguish between

different fault samples, effectively eliminate the

effects of isolated points and wild ideas on the

diagnostic results. Under the small sample

circumstances, the different failure of classification

can be achieved.

Induction motor rotor fault, throughout the

wavelet transform each band energy of the frequency

component of the stator current is normalized, used as

the fault feature vectors, input the support vector

machines for training. This weakens the impact of

load changes and noise on diagnostic accuracy. Test

results show that: fault diagnosis model based on

fuzzy support vector function can correctly diagnosed

induction motor rotor fault, thanks to structural risk

minimization principle, taking into account the

training error and generalization ability, with the

ability of good classification.

REFERENCES

QIU Arui, SUN Jian. Pattern recognition and diagnosis of

faults in electrical machines [J]. Journal of Tsinghua

University (Science & Technology), 1999, 39 (3):72-74.

Dong Jianyuan Duan Zhisha.The diagnostic method of a

synchronous motor based on the BP neural network

[J].Journal of Xi’an University of

Architecture&Technology.1998,30(2):159-162.

Filippetti F, etc. Recent developments of induction motor

drives fault diagnosis using AI techniques [J].IEEE

Transactions on Industrial Electronics, 2000,

47(5):994-1004.

Cao Zhitong, Chen Hongping, He Guoguan. Support Vector

Machine for Fault Diagnosis of the Rotor Broken Bars

of IM [J].Chinese Journal of Scientific Instrument,

2004, 25(6):738-741.

Chen Liyuan, Huang Jin. Motor Broken Rotor Bar Fault

Diagnosis With Support Vector Machine

[J].Transactions of China Electrotechnical Society,

2006, 21 (8):48-52.

Fang Ruiming, Zheng Lixin, Ma Hongzhong. Fault

diagnosis for rotor of induction machine based on

MCSA and SVM [J]. Chinese Journal of Scientific

Instrument, 2007,28 (2):252-257.

Induction Motor Fault Diagnosis Based on Fuzzy Support Vector Machine

591

VAPNI V, N. Statistical learning theory [M].New York:

Springer, 1995.

HA Ming-hu, PENG Gui-bing

，

ZHAO Qiuhuan. New

fuzzy support vector machine [J].

Computer Engineering and Applications, 2009, 45

(25):151-153.

ICECTT 2018 - 3rd International Conference on Electromechanical Control Technology and Transportation

592