Research on Fault Diagnosis Method of Lithium Battery
Based on Fuzzy Neural Network
Meng
Chen
1
, Zongyang Liu
1
, Jiang Wu
1
and Kun Ma
2
1
School of Electrical Engineering , Xi'an Jiaotong University, Xi'an, China
2
Lanzhou University of Finance and Economics Accounting School, Lanzhou, China
Keywords: Fuzzy neural network; Lithium battery; Fault diagnosis
Abstract: Aiming at the problem that the battery lithium battery has poor generalization performance for different
health conditions, a fault diagnosis method based on fuzzy neural network is proposed. The neural network
is used to diagnose the battery fault, and then the fuzzy system rules are used to output three fault states of
the lithium battery, namely Corresponding Capacity reduction, Increase of internal resistance, SOC
Reduction, and finally simulation analysis. The effectiveness of this method for fault diagnosis of lithium
battery systems is demonstrated.
1 INTRODUCTION
The domestic power battery technology is not
fully mature, and the battery failure is not easy to
detect at the initial stage. Therefore, it is of great
practical significance to carry out fault diagnosis
research on the battery system to ensure that the
battery is in normal operation.
The paper use fuzzy neural network to combine
fuzzy logic and neural network. The learning
mechanism of neural network is used to
automatically extract rules from input and output
data, and the fuzzy system is easy to express human
knowledge
[1-3]
. The characteristics can improve the
traditional fuzzy controller which must rely on
human thinking to adjust the membership function
repeatedly to reduce the error and improve the
performance. Simulation analysis shows that the
fuzzy neural network can effectively judge the fault
state of the lithium battery.
2 FUZZY NEURAL NETWORK
Design a five-layer feedforward neural network,
see Fig 1.The first layer is the input layer, and the
input value indicates that the first node of the input
layer corresponds to its first component. The input
and output of this layer are:
jj
xI =
1
1
njxIO
jjj
,,2,1,
11
===
2
1
x
j
x
n
y
1
y
Fig. 1. Fuzzy neural networks
The second layer is the membership function
layer, and the membership functions
belonging
to the fuzzy sets of the values of the respective
language variables are calculated. The membership
functions are:
( )
( )
( )
2
2
/exp
ijijj
i
j
EUx =
3
In the formula, the number of fuzzy partitions
m
= 3,
ij
U
indicating the mean of the membership
function,
ij
E
indicating the standard deviation of the
membership function,
ij
U
and
ij
E
are both
adjustable parameters. Then the input and output of
the second layer are:
jijjijj
xwOwI
121122
==
4
( )
2
2
122
2/exp
ijijjijj
EUxwO =
5
In the formula,
12
ij
w
indicates the connection
weight of the first node of the first layer and the
second node of the second layer.
The third layer is the rule front layer. The input
and output of the
jth
node are:
( )
= =
==
n
j
ij
n
j
ijijjijijjj
wEUxwwOI
1
23
1
2
2
122323
2/exp
6
+=
=
n
j
ijjj
wOO
1
2323
exp1/1
7
Where,
23
ij
w
represents the connection weight of
the second node of the second layer and the
jth
node of the third layer.
The fourth layer is the rule back layer, the input
and output are:
= = =
=
==
n
j
n
j
n
j
ijjijijjj
njwOwwOI
1 1 1
232343434
,,2,1,exp/
8
,/
3
1
334
=
=
n
j
jjj
OOO
9
In the formula,
34
ij
w
indicates the connection
weight of the
ith
node of the third layer and the
jth
node of the fourth layer; 3 indicates that the
fuzzification level is level 3.
The fifth layer is the output layer, which
represents the output variable. After deblurring, the
network output is obtained:
=
=
n
j
jijj
OwO
1
4455
10
Where,
45
ij
w
represents the connection weight of
the
ith
node of the fourth layer and the
jth
node of
the fifth layer.
3 LEARNING TRAINING
ALGORITHM
Use BP algorithm network to adjust weights
12
ij
w
23
ij
w
34
ij
w
45
ij
w
ij
U
ij
E
.
Define The network output error function
[4-5]
:
( )
njbaO
j
jjp
,,2,1,6.0
2
==
(11
is the expected output of the network,
j
b
is
the actual output of the network, and
n
is the
number of output categories.
Suppose
i
j
G
is the error back-propagation signal
the
jth
node of the
ith
layer of the network, then
the layers can be expressed as:
Output layer
jj
j
p
j
ba
O
O
G =
=
5
5
(12)
Fourth floor
2
444454
/
=
L
k
iik
L
k
iikij
L
k
iikiijjj
OEOEUOEOUGG
(13)
In the formula,
ij
U
represents the mean value of
the membership function,
ij
E
represents the standard
deviation of the membership function,
L
is the
number of output categories.
the third floor
2
444453
/
=
n
j
L
k
iik
L
k
iikij
L
k
iikiijjj
OEOEUOEOUGG
(14)
Second floor
=
2
j
G
( )( )
n
j
ijijjj
EUGG
2
33
1
(15)
The weight is calculated as:
12
ij
w
is 1
=
j nx
j
nx
jij
GG
m
w
2223
/
(16)
34
ij
w
=
( )
( )
( )( )
2
3
3
exp1
exp
j
j
jj
G
G
ba
+
(17)
( )
=
j
j
ij
ij
G
U
w
445
exp
1
(18)
ij
U
ij
E
are the parameters in the second
layer, find a step:
( ) ( ) ( )
( )
+=+=+
ka
E
kaakaka
ij
p
ijijijij
1
(19)
( ) ( ) ( )
( )
+=+=+
kb
E
kbbkbkb
ij
p
ijijijij
1
(20)
When
p
E
less than a given threshold
,
learning training stops. If the desired output has not
been reached, the weight parameter is adjusted
according to the formula until the predetermined
requirement is reached.
4 SIMULATION
In this paper, the CBP2450 battery pack in
reference[6] is used as the research object. The
charge and discharge experiments are designed by
using the ITECH series of DC electronic load and
DC voltage source. The specific battery parameters
are shown in Tab 1.
Tab. 1. The Battery Parameters
Fuzzy neural network has 3 nodes in the first
layer, and selects voltage, current and temperature as
network input; the second layer has 8 nodes; the
third and fourth layers are fuzzy rules and the post is
also 10 nodes; the fifth layer has 4 nodes. It
corresponds to four states of lithium battery,
including three major fault states and normal states,
such as Corresponding Capacity reduction, Increase
of internal resistance, and SOC Reduction.
Fig. 2 Composite prediction result
Select 180 data to be as training and testing
samples, and predict the actual curve and error curve,
the result is Fig.2.
Fig. 3. Fuzzy neural networks prediction
day
Actual
prediction
curve
Error
curve
Normal
Corresponding
Capacity reduction
SOC Reduction
Increase of
internal resistance
Fig. 3 is a fuzzy neural network prediction, the
function output corresponds to the discharge
capacity. In the sample size, the battery discharge
capacity can reach 100%. In order to facilitate the
observation error, the error curve is drawn by the
difference between the predicted output and the
expected output to predict the lithium battery fault
diagnosis. According to the prediction output , the
lithium battery fault could be diagnosed and the
result is credible.
After constructing the network, 150 training
samples are input for training. In order to avoid
over-learning, the training error precision is set, and
the learning process is stable and convergent. After
75 cycles, the allowable error range has been
reached. The BP network training error curve is
shown in Fig. 4.
16
32
48
64
80
3
10
4
10
2
10
1
10
0
10
1
10
Performance=0.004227
Goal=0.0015
Training Error
Iterations
Fig. 4 Training error curve.
5 CONCLUSION
Based on the complexity and uncertainty of
lithium battery faults in electric vehicles, this paper
proposes a lithium battery fault diagnosis method
based on fuzzy neural network. The method makes a
preliminary diagnosis of lithium battery fault
through neural network, and then uses the
combination rule to fuse different neural network
outputs, which can successfully diagnose the fault
state of the lithium battery, and the diagnostic
accuracy is higher than the single fault diagnosis
method, and the diagnosis is improved. Sexuality,
the result of the diagnosis is met, and the accurate
judgment of the fault state of the lithium battery of
the electric vehicle is obtained.
REFERENCES
1. WEI HeNICHOLAS williardCHEN Chaochaoet
a1 State of Charge Estimation for Lion Batteries
Using Neural Network Modeling and Unscented
Kalman Filter-based Error Cancellation[J]
International Journal of Electrical Power and Energy
Systems2014.
2. MATrHlEU D LIAW B Y Identify Capacity
Fading Mechanism in a Commercial LiFeP04
Cell[J]Journal of Power Source 2009194541-
549
3. XU JieState of Charge Estimation Online Based on
EKFAn Method for Lithiumion Power Battery[C]
2009 International Congress on Image and Signal
Processing(CISP 09)2009111-5
4. Steve R. Support vector machines for classification
and regression [D]. 2004,University of Southampton
5. Baker J.E.. Adaptive selection methods for genetic
algorithms[J], Proc. of the Intel Conf. on Genetic
Algorithms and Their Applications, 2008:101~111
6. ZHOU Meilan, ZHAO Qiang, ZHOU Yangqin. State
of Charge Estimation for LiFeP04 Cell Using
Improved PSO-BP Neural Network [J], Journal of
Harbin University of Science and Technology,
2015(4):88~92