You are on page 1of 11

Energy 214 (2021) 118866

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

An Intelligent Fault Diagnosis Method for Lithium Battery Systems


Based on Grid Search Support Vector Machine
Lei Yao a, b, Zhanpeng Fang a, b, *, Yanqiu Xiao a, b, Junjian Hou a, b, Zhijun Fu a, b
a
Henan Key Laboratory of Intelligent Manufacturing of Mechanical Equipment, Zhengzhou University of Light Industry, Zhengzhou, 450000, China
b
Henan Engineering Research Center of New Energy Vehicle Lightweight Design and Manufacturing, Zhengzhou University of Light Industry, Zhengzhou,
450000, China

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

Article history: For the safe operation of the electric vehicle, it is critical to quickly detect the safety state and accurately
Received 7 March 2020 identify the fault degree in battery packs. This article proposes a novel intelligent fault diagnosis method
Received in revised form for Lithium-ion batteries based on the support vector machine, which can identify the fault state and
2 September 2020
degree timely and efficiently. Due to the noise signal’s existence, firstly, the discrete cosine filtering
Accepted 16 September 2020
Available online 18 September 2020
method is adopted, and the truncated frequency is optimized based on the characteristic of white noise
to achieve reasonable denoising. Secondly, since the covariance matrix (CM) of filtered data is sensitive to
the current fluctuation, a modified covariance matrix (MCM) is proposed to reduce the influence of
Keywords:
Lithium-ion battery
current variation on the condition indicators. Thirdly, to ensure the accuracy and robustness of Support
Fault diagnosis Vector Machine (SVM), the grid search method is proposed to optimize the kernel function parameter
Grid search and penalty factor. Finally, the MCM and CM are respectively introduced into the model as the condition
Support vector machine indicators, and the results show that the former has high accuracy and timeliness. In summary, the
Modified covariance matrix proposed intelligent fault diagnosis method is feasible. It provides the theoretical basis for future fault
hierarchical management strategy of the battery system.
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction Ensuring the safe operation of Evs has become the core task for
the battery management system (BMS). The BMS can predict the
In the transportation field, electric vehicles (Evs) have been current working state of the battery by monitoring the voltage,
widely recognized as an imperative option for environmental is- current and temperature to maintain the greater security diagnosis
sues, which has attracted full attention [1e3]. The battery pack is accuracy [7,8]. Till now, many efforts have been devoted to devel-
the critical component in Evs that can provide high energy density oping various reliable BMS to improve the efficiency of battery
to meet the vehicle driving range requirements and offer high condition identification [9e11] and fault diagnosis [12e14]. How-
power density to fit the acceleration and hill-climbing scenarios. ever, most of the researches focus on the estimation of the state of
Due to the limitation of the single cell’s voltage and capacity, the charge (SOC) [9,10], state of health (SOH) [15,16], state of energy
battery pack is typically composed of many cells in various series- [17,18] and remaining useful life (RUL) [19] of the batteries. Re-
parallel modes. In recent years, a series of traffic accidents have searches on the reliability and application technology for battery
been reported many times in the complex operating environment pack are relatively scarce and not well-rounded.
of Evs, exposing lots of security risks of the battery pack, which In the practical application of power battery, only three pa-
should hinder the further promotion and application of Evs. So the rameters, voltage, current, and temperature, can be directly
safety of the battery pack has become a crucial issue that should be collected. However, the current data acquisition is difficult, espe-
urgently investigated [4e6]. cially the parallel branch’s value, which is affected by electromag-
netic interference, so many battery fault diagnosis and prognostics
approach based on the variation of voltage and temperature pa-
rameters have been proposed. Simultaneously, according to
* Corresponding author. Henan Key Laboratory of Intelligent Manufacturing of
Mechanical Equipment, Zhengzhou University of Light Industry, Zhengzhou, different data processing methods, research methods can be clas-
450000, China. sified into three categories: thermal-model-based, voltage-model-
E-mail address: 2015073@zzuli.edu.cn (Z. Fang).

https://doi.org/10.1016/j.energy.2020.118866
0360-5442/© 2020 Elsevier Ltd. All rights reserved.
L. Yao, Z. Fang, Y. Xiao et al. Energy 214 (2021) 118866

based, and data-driven-based methods. accuracy of model parameters, and the complexity of calculation
The thermal-model-based methods focus on detecting the cell increases.
surface temperature or predicting the internal value trend to Recently, the data-driven-based methods have been widely
determine the current working state. As the experimental process applied in the battery fault detection field due to the healthy
is relatively direct and smooth, it has been widely adopted. Vis- development momentum of big data processing technology. Zhao
wanathan et al. [20] measured the entropy changes with the et al. [33] proposed a machine-learning-based fault detection
various cathode and anode materials using an electrochemical method to detect abnormal changes in cells efficiently and accu-
thermodynamic measurement system and presented the reversible rately. Xia et al. [4] presented a fault detection method for short
heat generation rate was found to be a significant portion of the circuits based on the correlating efficiency of voltage curves, which
total heat generation rate. Fredrik et al. [21] reported that the cells’ has strong robustness and anti-jamming ability. Yao et al. [34]
heat release rate would increase with the rise of SOC, but the cor- compared three different entropy algorithms to identify the con-
relation between total heat release and SOC decreases gradually. necting faults, which achieved a relatively small amount of calcu-
Vantuch et al. [22] proposed that the changing trend of the battery lation. Wang et al. [35] proposed a fault detection method based on
surface temperature was measured by infrared thermal imager in a modified Shannon entropy difference of each cell voltage to
long term monitoring and fault detection, but the accuracy and determine the location and time of the fault. Based on the data-
robustness decreased. These methods are simple and easy to driven principle, Zhao et al. [36] and Yao et al. [37] proposed a
operate, which can be identifying the rent working state of the new fault diagnosis method using a recurrent neural network and
power battery to improve the fault identification rate. However, convolution neural network with the dynamic characteristics of
those mentioned above have strict requirements on environmental lithium-ion power battery. These methods do not need to build a
conditions and instruments, which cannot be applied to real ap- battery model and complex parameter identification, which has the
plications. To combine the internal reaction degree of the cell with advantages of high efficiency and reliability.
the external parameters’ change, Panchal S et al. [23] presented a This paper mainly studies the intelligent diagnosis of the
mathematical model to predict the transient temperature and connection fault of a series battery pack. The fault principle has
voltage distributions at different discharging rates. Huo et al. [24] been described in detail in Ref. [34]. Due to the complex non-linear
studied a 3D model of a lithium-ion battery’s thermal performance time-varying system with the absolute inconsistency of the batte-
to analyze the dynamic thermal behavior by utilizing various drive ries, it is difficult for the conventional methods to diagnose the cell
cycles. The electrochemical model of power battery can accurately parameters’ abnormal change and predict the time of faults
capture the intensity of the internal chemical reaction in different occurring in the initial stage. Meanwhile, lithium-ion batteries can
loading ratios, which is of great significance for fault diagnosis. be affected by road conditions, driving habits, environmental
Using the battery thermal model and the equivalent circuit model temperature, and other factors making it challenging to provide a
(ECM), Dey. S. et al. [25,26] proposed an identification scheme real-time, accurate, and timely diagnosis in real operated Evs.
based on the Luenberger observer to detect and isolate three kinds Therefore, to bridge these drawbacks, the condition indicators
of thermal faults: internal thermal resistance fault, convective based on the voltage data are mined and applied to support vector
cooling resistance fault, and thermal runaway. The cell surface machine (SVM) to diagnose the potential abnormality state and
temperature feature can represent the current internal reaction classify the severity of fault accurately. Specifically, four parts are
degree, but cell temperature can be affected by factors such as made in this article. Firstly, an optimal filtering method is adopted
location, internal resistance, cooling degree of the battery packs, using a discrete cosine filter method based on white noise char-
etc. Meanwhile, temperature change has specific hysteresis char- acteristics. Secondly, the covariance matrix (CM) of filtered data is
acteristics: the heat will pass through a period from the inside to analyzed, and a new condition indicator is determined. Thirdly, the
the surface, leading to the low timeliness of the thermal-model- grid search method is adopted to optimize the kernel function
based methods. parameter and penalty factor to ensure the model’s accuracy and
Compared with temperature signals, voltage signals are more robustness. Finally, through the training of SVM parameters by
comfortable to acquire and more time-efficient, which many re- condition indicators, the method can effectively realize the intel-
searchers favored. Therefore, more publications are based on the ligent fault diagnosis. Simultaneously, the MCM’s efficiency and
voltage signal to diagnose the faults. The voltage-model-based engineering practicability are verified from the identification rate
methods usually include four key steps. The first step is to estab- and diagnostic accuracy. The detailed test information of this paper
lish a model that can represent the dynamic characteristics of the is shown in Fig. 1.
cells. Secondly, the parameters of the model are identified by The rest of this paper is summarized as follows. The test work-
intelligent algorithms, such as the Kalman filtering, particle bench and cell are described in Section 2. The proposed fault
filtering methods, and so on [27e29]. Thirdly, the parameters of cell diagnosis method for a series battery pack is presented in Section 3.
fault characteristics are determined, and the corresponding The condition indicators and the optimal filtering method based on
threshold values are set [30]. Fourthly, the measured signal is white noise characteristics are both described in the following
compared with the threshold to determine whether the fault has section. The experimental results will be discussed in Section 5, and
occurred. Ouyang et al. [31] studied an internal short circuit finally, in Section 6, a summary of findings and the conclusion will
detection method using recursive least squares to estimate the be presented.
mean-difference model parameters, and the parameters of all cells
in the pack are compared with the set threshold to detect the fault 2. Experimental study
state. To avoid the occurrence of faults and accelerated cell aging,
Zheng et al. [32] investigated a differential voltage analysis method 2.1. Test platform
to identify the battery’s current working state accurately. Xiong
et al. [6] proposed a two-step equivalent circuit battery model that The test platform shown in Fig. 2 is to implement the battery
can provide a rapid and precise fault diagnosis result by separating pack fault diagnosis, which contains a battery test instrument
voltage variation caused by polarization and a significant voltage (Digtron Battery Test System: BTS-600), a vibrating test bench, an
drop caused by a large current. The accuracy of these diagnostic information collector, several voltage sensors and the data pro-
methods are relatively high but mainly depends on the prediction cessor. The battery tester, which integrates the power supply,
2
L. Yao, Z. Fang, Y. Xiao et al. Energy 214 (2021) 118866

Fig. 1. Framework of fault diagnosis algorithm.

electric load, signal acquisition, and transmission, is used for bat- upper computer. To reduce the influence of electromagnetic inter-
tery experiments to mainly simulate the charge and discharge ference on the data acquisition system when the vibrating test
mode in the electric vehicle operation. The vibrating test bench will bench is working, the information collector will be covered in an
simulate the battery pack’s bumpy state and provide a real vibra- iron-box during the whole test process.
tion road environment. The information collector can obtain real- The test is conducted using the lithium-ion cells, which have a
time voltage values by some sensors and transmit them to the nominal capacity of 50 Ah, produced by Tianjin Lishen battery Co,

3
L. Yao, Z. Fang, Y. Xiao et al. Energy 214 (2021) 118866

Fig. 2. Picture of the experimental system.

LTD, and LiMn2O4 will be used as the positive electrode and


graphite as the negative electrode. The general characteristics of
the battery are summarized in Table .1.

2.2. Test standard

The test standard mainly refers to China’s automobile standard:


“General Requirement of Traction Battery Enclosure for Electric
Vehicles,” which has explicit requirements for safety and reliability
of the power battery system. The measurement method for each
performance verification is established, including the vibration
frequency, acceleration value, and ambient temperature. The vi-
bration frequency is 10 Hze55 Hz, and the vibration acceleration is
10 g at 298 K.
The current value with the highest frequency is extracted from
the actual record data of the E150 vehicle operation in Beijing.
Through appropriate simplification, it is set as the value of this load.
The specific test parameters are shown in Fig. 3.

3. Diagnosis method
Fig. 3. Current loading curve.

3.1. Discrete cosine transform


extracted from the collected data.
With the development of technology, the Discrete Cosine
The most common DCT definition of a 1-D sequence of length N
Transform (DCT) has the property of the orthogonal transform,
is
which is widely used in signal and image processing [38]. This
method can quickly transform data from the time domain to the  
frequency domain, attempt to de-correlate the collection data, and X
N1
pð2x þ 1Þ
CðuÞ ¼ aðuÞ f ðxÞcos u (1)
enable each transform coefficient to encode independently after x¼0
2N
de-correlation. On the basis of preserving the authenticity of the
original image to the greatest extent, it can improve the compres- For. u ¼ 0; 1; 2; …; N  1where
sion rate and reduce the amount of data, although the data is lost to
8
some extent. Based on this theory, the noise characteristics can be >
>
>
< qffiffiffiffiffiffiffiffiqffiffiffiffiffiffiffiffi u¼0
aðuÞ ¼ 1
=N
2
=N (2)
>
> 1uN1
Table 1 >
:
Characteristics of the tested cells.

Properties value
Where f(x) is the original signal, x ¼ 0,1,2…N-1; C(u) is the u-th
Length of cell 170 mm cosine transform coefficient, u is the generalized frequency vari-
Width of cell 43 mm
able; N is the number of original signals; aðuÞ is a compensation
Height of cell 100 mm
Nominal capacity 50 Ah
coefficient, which can make the DCT transform matrix to be the
Nominal voltage 3.6 V orthogonal matrix.
qffiffiffiP
Voltage range 3V-4.2 V
It is clear from (1) that foru ¼ 0, Cðu ¼ 0Þ ¼ N1 N1
x¼0 f ðxÞ. Thus,

4
L. Yao, Z. Fang, Y. Xiao et al. Energy 214 (2021) 118866

the first transform coefficient is the average value of the sample  


sequence. kx  xi k2
Kðx; xi Þ ¼ exp  (7)
Inverse discrete cosine transform (IDCT) is the inverse trans- 2s2
formation of DCT, which can obtain the corresponding time-
domain data curve from the DCT coefficient curve. Where s is the parameter of the Gaussian radial basis function
In this paper, the voltage curve is composed of many different (RBF).
frequencies and amplitudes, and the noise frequency is commonly Penalty factor C and kernel function parameter s are two vari-
located in the high-frequency band. Therefore, combining the ables, which determine the classification accuracy. The optimiza-
calculation principles of DCT and IDCT, the noise characteristics in tion of SVM parameters is explained in part 5.
the collected data can be gradually separated by selecting data of
different frequency stages. The core part is the determination of the
4. Battery data analysis process
different cut-off frequencies (CFF), which is elaborated in the fourth
part of this paper.
4.1. Preliminary analysis

The collected data usually includes normal operation data


3.2. Support vector machine (SVM) signal, noise signal, and various types of fault information, which
has the characteristics of high time-varying, nonlinear, and
SVM is a supervised learning method suitable for a small sam- complexity. According to the principle of DCT and IDCT, this section
ple, nonlinear and high-dimensional data. It is a linear two classifier describes the noise filtering process of the collected data in detail.
defined to maximize the interval in the feature space. For the As shown in Fig. 4, in real operated Evs, due to the existence of
nonlinear case, SVM mainly uses a kernel function to map the noise, when the current is constant, many high-frequency tiny
original data to the high-dimensional Hibert space to achieve linear burrs appear on the voltage curve, which often covers the initial
separability, so as to solve the issue of linear inseparability in the fault signal, increasing the difficulty of timeliness and accuracy of
original space [39]. fault diagnosis. Based on the difference between the frequency of
For a given linear separable training data set ðxi ; yi ; i ¼ 1; /; nÞ, loading current and noise frequency, combined with the DCT
where x2Rn and y2Rn the separation hyperplane is obtained by principle, the collected data is transformed from the time domain
maximizing the interval or solving the corresponding convex to the frequency domain, as shown in Fig. 5.
quadratic programming problem. In Fig. 5, the DCT coefficient is set to abscissa, and the amplitude
is ordinate. The location of the DCT coefficient n represents that the
uT x þ b ¼ 0 (3) frequency is 2n, and the amplitude is the vibration amplitude of the
voltage at this point n.
Where u is a parameter vector, x and b are respectively sample data To capture the change characteristics of the high-frequency
and offset. signal of experimental data under CFFs, the CFF function f is set,
The corresponding classification decision function is: as shown in Fig. 5, where i is the cut-off value. The amplitude-
  frequency characteristic data of the high-frequency stage is trans-
f ðxÞ ¼ sgn uT x þ b (4) formed into time-domain data through IDCT. The mean value and
variance of the data are calculated respectively in Fig. 6.
For linear-non-separable data sets, each sample point in- Fig. 6(a) and (c) show the mean and variance of high-frequency
troduces a relaxation variable xi  0 to represent a non-negative signals in the time domain at different CFFs, respectively. The lower
measure of the misclassification error. Then the linear-non- the CFF is, the higher the mean value and variance value of the cut-
separable SVM can be represented by the following optimization off date in the time domain are. With the increase of the CFF, the
problems: two curves decrease dramatically and approach to 0. To display the
change trend of mean and variance curve under high CFF more
1 Xn
min kuk2 þ C xi
2 i¼1

s:t:yi ðuxi þ bÞ  1  xi

xi  0 i ¼ 1; 2; /; n (5)

Where C is the penalty factor, which is generally determined by the


actual problem.
Combined with kernel function and soft interval maximization
principle, the classification decision function of non-linear SVM can
be obtained by using the dual function and Lagrange optimization
algorithm, as follows:
!
X
n
f ðxÞ ¼ sgn ai yi Kðx; xi Þ þ b (6)
i¼1

Where ai  0 represents the Lagrange multiplier, and Kðx; xi Þ is the


kernel function, which satisfies Mercer theorem. In this paper, the
radial basis kernel function is chosen. Fig. 4. The original voltage data curve.

5
L. Yao, Z. Fang, Y. Xiao et al. Energy 214 (2021) 118866

basically parallel to the abscissa. However, the variance curve


changes smoothly with the increase of the CFF coefficient. After
n ¼ 400, the data changes are also approximately balanced with the
abscissa, which is consistent with the change of the mean curve. In
order to further verify the conclusion, the power spectrum curve of
the original function is calculated, as shown in Fig. 7.
Fig. 7 shows a certain power value in the low-frequency stage,
which decreases sharply with the increase of the CFF. When
n ¼ 400, the power value is approximately 0 in Fig. 7(b). It is
verified that when the frequency at n ¼ 400 is selected as the CFF,
the high-frequency data can be set as white noise. Combined with
the IDCT principle, the corresponding filtered voltage curve is ob-
tained, as shown in Fig. 8.
The setting of cut-off frequency is very crucial for subsequent
fault identification. If the CFF is set too high, the noise cannot be
removed to a large extent, increasing the difficulty of subsequent
fault identification; if it is set too low, it is easy to cause the voltage
data to not retain its actual characteristics. Based on the above
Fig. 5. DCT coefficient curve.
theoretical analysis, this paper sets the cut-off voltage frequency as
the frequency value when the DCT coefficient is 400.
accurately, the changing trend of the two curves at the inflection
point is amplified locally, as shown in Fig. 6(b) and (d) n ¼ 25 is the
4.2. Condition indicators
turning point for the change of the mean curve. Before that, the
data fluctuates obviously and declines sharply, and when the CFF
Generally, a series of statistical parameters from time-domain
coefficient is higher than 25, the mean data tends to be flat,
and frequency-domain signals are extracted with data mining

Fig. 6. Mean and variance curve. (a) global mean curve, (b) locally amplified mean curve, (c) global variance curve, (d) locally amplified variance curve.

6
L. Yao, Z. Fang, Y. Xiao et al. Energy 214 (2021) 118866

Fig. 7. Power spectrum curve. (a) global data curve, (b) local amplified data curve.

technology. The parameters are called condition indicators, which


need to reflect the current system’s working state quickly and ! !
1 Xm 1 Xm 1 Xm
honestly. Therefore, the selection of condition indicators is essen- cðv1 ;v2 Þ ¼ v  v vi;2  v
tial for the intelligent identification of the system state. m  1 i¼1 i;1 m i m;1 m i m;2
In data mining technology, covariance is the primary condition (9)
indicator, which has been widely used. Covariance is mainly used to
Therefore, the whole covariance matrix can be expressed as:

2 3
6 c11;2 c11;3 / c11;k c12;3 c12;4 / c12;k / c1k1;k 7
6 7

6 c21;2 c21;3 / c21;k c22;3 c22;4 / c22;k / c2k1;k 7
C ¼ C1;2 C1;3 / C1;k C2;3 C2;4 / C2;k / Ck1;k ¼6
6
7
7
6 « « « « « « « « « « 7
4 5
cNmþ1
1;2 cNmþ1
1;3 / cNmþ1
1;3 cNmþ1
1;3 cNmþ1
1;3 / cNmþ1
1;3 / Nmþ1
c1;3

(10)

measure the overall error of the two variables. If the changing


tendencies of the two variables are the same, the covariance be- Where cði;jÞ is the covariance value between cell i and cell j, k sug-
tween the two variables is positive; if the changing trend is the gests the number of the cells. N-Mþ1 expresses the number of the
opposite, the covariance will be negative. When the trend is inde- dimension of the covariance matrix.
pendent of each other, then the covariance is 0. For the data set X: Based on the above algorithm, the actual voltage covariance
(x1,x2) composed of two vectors, its length is n, and the covariance curve is obtained, as shown in Fig. 9. To some extent, the voltage
formula can be expressed as: covariance matrix reflects the correlation between each cell voltage
data. The overall covariance curve is consistent with the voltage
fluctuation curve. However, the loading mode greatly influences
the changing trend of terminal voltage, resulting in the data fluc-
tuation of variance matrix under different current, as shown in
Fig. 9(a).
! ! The fluctuation value of covariance data mainly includes two
1 Xn 1Xn 1Xn
cðx1 ;x2 Þ ¼ xi;1  x xi;2  x (8) values, which can be expressed as follows:
n1 i¼1 n i n;1 n i n;2
DC ¼ DC1 þ DC2 (11)
Where the cðx1 ;x2 Þ is the covariance value between x1 and x2, which
is a unitless value. Where DC1 is the relative difference, which is a function of the
Due to a large amount of data in the process of recording mul- internal resistance and the link impedance of the cell, and it is the
tiple sets of battery data, in order to reduce the amount of calcu- main parameter for fault identification. DC2 is the data difference,
lation and ensure the implementation of the algorithm, it is which is greatly affected by the current fluctuation.
necessary to set appropriate windows length. Set N as the total Because of the existence ofDC2 , the change range of covariance
running time and m as the moving window to simplify the calcu- data in the constant current mode is not apparent, increasing the
lation. The covariance formula can be expressed as: difficulty of accurate fault diagnosis. In order to eliminate the
7
L. Yao, Z. Fang, Y. Xiao et al. Energy 214 (2021) 118866

Fig. 8. The filtered voltage data curve.

Fig. 10. SVM parameters optimization diagram.

1 Xk X k
ccaði;jÞ ¼ caði;jÞ  ca ; a ¼ 1; 2; /N  m þ 1
ð1 þ kÞk i¼1 j¼1 ði;jÞ
jsi

(12)

Where caði;jÞ is the a-th unmodified covariance value between cell i


and cell j, and ccaði;jÞ is the corresponding modified value. k suggests
the number of cells. a is the a-th second value of the covariance
matrix.
As shown in Fig. 9(b), the modified covariance data is reduced by
order of magnitude. In the constant current stage, the fluctuation
degree of the curve has been improved to a certain extent, and the
difference between the cell voltage data has been expanded, which
will be beneficial to the identification of the current working state
of the battery.

5. Verification

5.1. SVM parameter selection

Based on the above description of the SVM principle, the penalty


factor C and kernel function parameter s are two variables that
determine the accuracy of the fault classification algorithm. If the C
value is too large, the classifier is easy to produce over-fitting,
which affects the algorithm’s generalization ability; if the C value
is too small, the classifier will not care about the identification er-
ror, resulting in low classification accuracy. While the parameter s
is an essential parameter in the kernel function, which affects the
complexity of the SVM classification algorithm. To summarize, C
and s are the key parameters that affect the performance of the
SVM classifier. To improve the accuracy and applicability of the
SVM classifier, this section uses a cross-validation (CV) and grid
search (GS) method to optimize its parameters.

Fig. 9. The voltage covariance data curve (a) unmodified data curve, (b) modified data
5.1.1. Cross-validation
curve.
CV is a statistical method used to verify the classifier perfor-
mance. The basic idea is to group the original data, one part as the
current influence, the mean value of all cells covariance data is used training set, and the other part as the verification set. The training
as the basis, and each group of information is compared with the set is used to train the classifier, and then the verification set is used
mean value. The equation expression is as follows: to test the accuracy of the training model, which is used to evaluate
the performance of the classifier.
8
L. Yao, Z. Fang, Y. Xiao et al. Energy 214 (2021) 118866

In general, the original data is divided into N groups, each of very high while the value of the test set is deficient indicates that
which is verified once, and the rest of the N-1 groups are used as the classifier has been in the state of over-learning, and the
training sets. In this way, N models are obtained, the average value generalization ability is weakened. In order to avoid the above
of classification accuracy for the final verification set of N models is situation, the smaller penalty parameter C is a better choice when it
used as the classifier’s performance index. This method can effec- can achieve the highest accuracy of classification verification. So
tively avoid over-learning and under-learning states, and the re- this paper selects the best parameters according to the above
sults have individual practicability. principle and marks the point with the red asterisk in Fig. 10.

5.2. Determination of SVM sample set


5.1.2. Grid search
The basic principle of the GS method is to divide the grid in a With the fault event’s occurrence, the voltage fluctuation in-
particular range and traverse all points in the network with the creases sharply with the increase of the charging and discharging
values of parameters C ands. Combined with the CV method, the time. In this paper, according to the different degrees of voltage
accuracy of the training set verification classification under this curve fluctuation during the charging and discharging cycle, the
parameter value is obtained. Finally, the parameters C and s with cycle data when n ¼ 4, 8, 10, and 13 are selected respectively, as
the highest accuracy are determined as the best parameters. The shown in Fig. 11.
results of parameter optimization are shown in Fig. 10, where the As shown in Fig. 11, there is also a significant difference in the
ranges of C and s are set to [-10, 20] and [-5, 10], respectively, and fault degree of the system over time. In order to facilitate the
the steps are 0.1.
Fig. 10 shows that the classification accuracy of parameters C
and s in a particular range is very high. Still, if the parameters are Table 2
not selected properly, the accuracy will be in a deficient range. The meaning of fault classification.
When the parameter values are appropriately selected and com- Serial number Grade m Meaning
bined, some points that can make the diagnostic accuracy reach the
1 1 No-fault
highest value can be obtained. In this paper, the asterisk is used to 2 2 Slight failure
mark them out. 3 3 Medium failure
The situation in which the training set classification accuracy is 4 4 Serious failure

Fig. 11. The charging and discharging voltage curve (a) n ¼ 4, (b) n ¼ 8, (c) n ¼ 10, (d) n ¼ 13.

9
L. Yao, Z. Fang, Y. Xiao et al. Energy 214 (2021) 118866

Table 3
The diagnostic accuracy and time consumption of SVM.

Serial number Diagnostic accuracy % Parameter SVM identification time/ Running time/s
identification time/s s

M U M U M U M U

1 97.1631 96.4539 167 498 0.038 0.615 168.02 498.65


2 97.8723 96.5427 172 511 0.052 1.549 172.45 513.01
3 97.6352 94.3262 163 505 0.053 1.167 163.70 506.71

observer to understand the battery fault, we divide the fault into data-driven is proposed for the lithium-ion battery system. Accu-
four levels. The fault degree from 1 to 4 increases gradually. There is rate and reliable experimental voltage data is essential to deter-
no fault at the first level, and serious fault at the fourth level, as mine the current working state of battery packs. However, due to
shown in Table 2. the noise interference’s real-time existence, it is challenging to
identify the fault state accurately in real-time. Therefore, a novel
implementation of the DCT denoising technique is firstly presented
5.3. Result analysis
for the actual voltage value, which mainly transforms the time-
domain signal of the battery system in the non-vibration state
Based on the above theoretical analysis, this paper takes four
into the frequency-domain signal. And the optimized truncated
cells in series as an example to verify the effectiveness and sensi-
frequency is obtained based on the working characteristics of white
tivity of this method. 677 sets of data are selected in the paper, of
noise, which can reduce the noise disturbance to the maximum
which 536 sets are set as training sets, which are mainly used to
extent without distortion of voltage data. Since the covariance
determine the optimal parameter value of SVM, and the remaining
curve of the collected data after denoising is sensitive to the fluc-
data is set as test sets to verify the accuracy of the algorithm and
tuation of load current, a modified covariance matrix is proposed as
prevent over-fitting. After further iterative training, several
the condition indicator of the SVM algorithm. The relative variation
different structures appear as classifiers of fault types and degrees.
of the covariance curve can be obtained by taking the mean of
The accuracy and immediacy of fault state identification are two
covariance as the cardinal number in the appropriate length win-
essential factors to evaluate whether the method has engineering
dow. By the cross validation and grid search, the kernel function
application value. Therefore, while obtaining high accuracy of fault
and penalty factor of SVM are optimized. Under the same diag-
diagnosis, the time consumed by state identification is also a critical
nostic accuracy, the smaller penalty factor is selected to improve
characterization parameter to evaluate the algorithm.
the algorithm’s generalization ability. Moreover, the fault degree of
In the paper, MCM and CM are used repeatedly as input vectors
the battery system varies in different working periods. Therefore,
to calculate the diagnostic accuracy and time consumed, as shown
based on the severity of voltage curve fluctuation, the system state
in Table .3. (M: Modified; U: Unmodified).
is divided into four levels. At last, through the comparison of the
Table .3 shows that when MCM is used as the input of SVM, the
diagnostic results of SVM with MCM and SVM with CM, both the
diagnostic accuracy rates are 97.1%, 97.8%, and 97.6%, respectively,
diagnostic accuracy is higher than 95%, which meets the engi-
with an average of 97.5%. There is a certain difference in each
neering requirements. But the consumed time is 167.38 s and
diagnostic accuracy value, mainly because the test set data and
505.8 s, respectively, saving 67%, which shows that the SVM with
training set data are randomly allocated to improve the robustness
MCM has high timeliness and can quickly identify the current
of the algorithm. The data of each training is not the same, and
working state of the system. To sum up, the proposed method not
there is a certain error in the diagnostic accuracy. When CM is used
only detects the voltage fault in time but also reflects the severity of
as the input, the above phenomena also exist in the multiple
the fault. It also provides the purpose and basis for the study of
diagnostic values of SVM, with the accuracy of 96.4%, 96.5%, and
battery system fault hierarchical management strategy, which will
94.3%, respectively, and an average value of 95.7%. The results show
be the focus of future further work.
that the accuracy of the fault diagnosis method based on GS-SVM is
higher than 95%, whether CM or MCM is used as the input variable,
which can meet the requirement of engineering application. Credit author statement
However, when the MCM is the input parameter, the precision
value is slightly better. Lei Yao: CRediT roles: Methodology; Data Curation; Writing-
In terms of immediacy, Table .3 shows the time spent in each Original Draft; Formal analysis. Zhanpeng Fang: CRediT roles:
step in two input quantities. When MCM is the input of SVM, the Writing-Review & Editing; Funding acquisition; Methodology;
time of parameter optimization and state identification is 167 s, Project administration; Resources; Supervision. Yanqiu Xiao:
172 s 163 s, and 0.038 s, 0.052 s 0.053 s, respectively. The total CRediT roles: Writing-Review & Editing; Formal analysis; Investi-
average time is 167.38 s. However, when MC is the input, the time is gation; Methodology. Junjian Hou: CRediT roles: Investigation;
498 s, 511 s, 505 s, and 0.615 s, 1.549 s, 1.167 s, respectively, and the Methodology. Zhijun Fu: CRediT roles: Formal analysis.
total average time is 505.8 s. The results show that the time of fault
diagnosis based on SVM is 33% based on CM, and 67% of the Declaration of competing interest
working time is saved.
To summarize, MCM is more suitable as the condition indicator The authors declare that they have no known competing
of state identification, making SVM reflect the system’s working financial interests or personal relationships that could have
state more quickly and accurately. appeared to influence the work reported in this paper.

6. Conclusions Acknowledgements

In this study, an intelligent fault diagnosis method based on The authors are grateful to the financial support of National
10
L. Yao, Z. Fang, Y. Xiao et al. Energy 214 (2021) 118866

Natural Science Foundation of China (No. 51805491), and the Key 2019;417:188e92.
[20] Viswanathan VV, Choi D, Wang D, et al. Effect of entropy change of lithium
Scientific Research Projects of Higher Education Institutions in
intercalation in cathodes and anodes on Li-ion battery thermal management.
Henan Province (No. 18A460035) and the Doctor Research Foun- J Power Sources 2010;195(11):3720e9.
dation of Zhengzhou University of Light Industry (No. 2016BSJJ014). [21] Fredrik L, Petra A, Bengt-Erik M. Lithium-ion battery aspects on fires in
electrified vehicles on the basis of experimental abuse tests. Batteries
2016;2(2):9.
References [22] Vantuch Tom as, Fulne
cek Jan, Holusa Michael, et al. An examination of
thermal features’ relevance in the task of battery-fault detection. Appl Sci
[1] Feng F, Hu X, Hu L, et al. Propagation mechanisms and diagnosis of parameter 2018;8(2).
inconsistency within Li-Ion battery packs. Renew Sustain Energy Rev [23] Panchal S, Mathew M, Fraser R, et al. Electrochemical thermal modeling and
2019;112:102e13. experimental measurements of 18650 cylindrical lithium-ion battery during
[2] Teixeira ACR, Ricardo Sodre  Jose
. Simulation of the impacts on carbon dioxide discharge cycle for an EV. Appl Therm Eng 2018;135:123e32.
emissions from replacement of a conventional Brazilian taxi fleet by electric [24] Huo W, He H, Sun F. Electrochemicalethermal modeling for a ternary lithium
vehicles. Energy 2016;115(3):1617e22. ion battery during discharging and driving cycle testing. RSC Adv 2015;5(71):
[3] Kang Y, Duan B, Zhou Z, et al. A multi-fault diagnostic method based on an 57599e607.
interleaved voltage measurement topology for series connected battery packs. [25] Dey S, Perez HE, Moura SJ. Model-based battery thermal fault diagnostics:
J Power Sources 2019;417:132e44. algorithms, analysis, and experiments. IEEE Trans Contr Syst Technol 2019;27:
[4] Feng X, Lu L, Ouyang M, et al. A 3D thermal runaway propagation model for a 576e87.
large format lithium ion battery module. Energy 2016;115:194e208. [26] Kong X, Zheng Y, Ouyang M, Lu L, Li J, Zhang Z. Fault diagnosis and quanti-
[5] Wang Y, Tian J, Chen Z, et al. Model based insulation fault diagnosis for tative analysis of micro-short circuits for lithium-ion batteries in battery
lithium-ion battery pack in electric vehicles. Measurement 2019;131:443e51. packs. J Power Sources 2018;395:358e68.
[6] Xiong R, Yang R, Chen Z, et al. Online fault diagnosis of external short circuit [27] Xiong R, He H, Sun F, et al. Model-based state of charge and peak power
for lithium-ion battery pack. IEEE Trans Ind Electron 2020;67:1081e91. capability joint estimation of lithium-ion battery in plug-in hybrid electric
[7] Hamut HS, Dincer I, Naterer GF. Exergy analysis of a TMS (thermal manage- vehicles. J Power Sources 2013;229(Complete):159e69.
ment system) for range-extended EVs (electric vehicles). Energy 2012;46(1): [28] Wei Z, Meng S, Tseng KJ, et al. An adaptive model for vanadium redox flow
117e25. battery and its application for online peak power estimation. J Power Sources
[8] Hu X, Li SE, Jia Z, et al. Enhanced sample entropy-based health management of 2017;344:195e207.
Li-ion battery for electrified vehicles. Energy 2014;64:953e60. [29] Wang Y, Zhang C, Chen Z. A method for state-of-charge estimation of LiFePO4
[9] Bo N, Binggang C, Bin W, et al. Adaptive sliding mode observers for lithium- batteries at dynamic currents and temperatures using particle filter. J Power
ion battery state estimation based on parameters identified online. Energy Sources 2015;279:306e11.
2018;153:732e42. [30] Xiong R, Yu Q, Shen W, et al. A sensor fault diagnosis method for a lithium-ion
[10] Li Y, Wang C, Gong J. A combination Kalman filter approach for State of Charge battery pack in electric vehicles. IEEE Trans Power Electron 2019:9709e18.
estimation of lithium-ion battery considering model uncertainty. Energy [31] Ouyang M, Zhang M, Feng X, Lu L, Li J, He X, Zheng Y. Internal short circuit
2016;109:933e46. detection for battery pack using equivalent parameter and consistency
[11] Ye, Min, Guo, Hui, Xiong, Rui,等. A double-scale and adaptive particle filter- method. J Power Sources 2015;294:272e83.
based online parameter and state of charge estimation method for lithium- [32] Zheng L, Zhu J, Wang G, et al. Differential voltage analysis based state of
ion batteries[J]. Energy, 144:789-799. charge estimation methods for lithium-ion batteries using extended Kalman
[12] Chen Z, Lin F, Wang C, et al. Active diagnosability of discrete event systems filter and particle filter. Energy 2018;158:1028e37.
and its application to battery fault diagnosis. IEEE Trans Contr Syst Technol [33] Zhao Y, Liu P, Wang Z, et al. Fault and defect diagnosis of battery for electric
2014;22(5):1892e8. vehicles based on big data analysis methods. Appl Energy 2017;207:354e62.
[13] Liu Z, Ahmed Q, Zhang J, et al. Structural analysis based sensors fault detection [34] Yao L, Wang Z, Ma J. Fault detection of the connection of lithium-ion power
and isolation of cylindrical lithium-ion batteries in automotive applications. batteries based on entropy for electric vehicles. J Power Sources 2015;293:
Contr Eng Pract 2016;52:46e58. 548e61.
[14] Dey S, Biron ZA, Tatipamula S, et al. Model-based real-time thermal fault [35] Wang Z, Hong J, Liu P, et al. Voltage fault diagnosis and prognosis of battery
diagnosis of Lithium-ion batteries. Contr Eng Pract 2016;56:37e48. systems based on entropy and Z-score for electric vehicles. Appl Energy
[15] Cui Y, Zuo P, Du C, et al. State of health diagnosis model for lithium ion bat- 2017;196:289e302.
teries based on real-time impedance and open circuit voltage parameters [36] Zhao R, Kollmeyer PJ, Lorenz RD, et al. A compact unified methodology via a
identification method. Energy 2017;144:647e56. recurrent neural network for accurate modeling of lithium-ion battery voltage
[16] Berecibar M, Garmendia M, Gandiaga I, et al. State of health estimation al- and state-of-charge[C] 2017 IEEE Energy Conversion Congress and Exposition
gorithm of LiFePO4 battery packs based on differential voltage curves for (ECCE). IEEE 2017:5234e41.
battery management system application. Energy 2016;103:784e96. [37] Yao L, Xiao Y, Gong X, et al. A novel intelligent method for fault diagnosis of
[17] Wang Y, Pan R, Liu C, et al. Power capability evaluation for lithium iron electric vehicle battery system based on wavelet neural network. J Power
phosphate batteries based on multi-parameter constraints estimation. J Power Sources 2020;453:227870.
Sources 2018;374:12e23. [38] Song BC, Jeong SC, Choi Y. Video super-resolution algorithm using Bi-
[18] Wang Y, Zhang C, Chen Z. An adaptive remaining energy prediction approach directional overlapped block motion compensation and on-the-fly dictio-
for lithium-ion batteries in electric vehicles. J Power Sources 2016;305:80e8. nary training. IEEE Trans Circ Syst Video Technol 2011;21(3):274e85.
[19] Zhou X, Pan Z, Han X, et al. An easy-to-implement multi-point impedance [39] Vapnik V. Statistical learning theory. New York: Wiley; 1998.
technique for monitoring aging of lithium ion batteries. J Power Sources

11

You might also like