Measurement 116 (2018) 402–411

Contents lists available at ScienceDirect

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

A novel fault diagnosis method for lithium-Ion battery packs of electric T

vehicles
⁎
Xiaoyu Li, Zhenpo Wang
National Engineering Laboratory for Electric Vehicles, School of Mechanical Engineering, Beijing Institute of Technology, Beijing 100081, China
Collaborative Innovation Center of Electric Vehicles in Beijing, Beijing Institute of Technology, Beijing 100081, China

A R T I C L E I N F O A B S T R A C T

Keywords: This paper focuses on fault detection based on interclass correlation coefficient (ICC) method for guaranteeing
Electric vehicles safe and reliable of electric vehicles (EVs). The proposed method calculates ICC values by capturing the off-trend
Fault diagnosis voltage drop and the voltages are extracted from Service and Management Center of electric vehicles. The ICC
Lithium-ion batteries value is employed to analyze battery fault by ICC principle. The ICC value not only has advanced fault resolution
Interclass correlation coefficient
by amplifying the voltage difference, but also can prolong the fault memory by setting moving windows.
Service and management center for electric
Moreover, a loop joints the first and last voltages is designed to locate faults in battery pack. In addition,
simulation and experiment are employed to validate and analyze the voltage faults. Based on the simulation
verification, the appropriate size of moving windows is set to ensuring sensitivity of fault detection method. The
experiment results indicate the method can appropriately detect fault signals for EVs.

1. Introduction [4,27,28]. The battery external/internal short circuit is more hazardous

Electric vehicles (EVs) have reigned supreme as the most popular
transportation applications for their capabilities to protect environment
including high performance, and non-pollution [1–3]. Battery pack is
the core component in EVs, which can not only provide driving force
but absorb braking energy. The battery pack typically consists of nu-
merous battery cells in various series-parallel patterns due to limita-
tions of voltage and capacity for each cell [4,5]. During the EV opera-
tion, it is vitally important to ensure the safety and reliability by
monitoring states of cells. Thus, a so-called battery management system
(BMS) has been employed to collect the identification data and to es-
timate states for cells [6–8]. In addition, some applications of the BMS
such as equalization and energy management are designed for the sake
of prolong the battery lifetime [9–11]. Generally, the identification data
refers to temperature, current and voltage, and the cell state includes
state of charge (SOC) [12,13], state of health (SOH) [14–16], state of
energy (SOE) [17] and remaining useful lifetime (RUL) [18,19]. How-
ever, the general BMS cannot detect the battery faults, which contribute
much to fire or explosion accidents of EVs [20–22]. Nowadays, the
electrical accidents can be classified into four types, i.e., over charge
[23], over-discharge [24], external short circuit and internal short
circuit [25,26]. The first two types mentioned above are voltage ab-
normity, which can usually be avoided by means of battery SOC esti-
mation rather than providing valuable information to customers
compared with over charge/discharge. The short circuit fault is more
likely to lead to high heat generation and induce thermal runaway
[25,29,30]. The thermal runaway is defined as a case that the solid
electrolyte interface (SEI) begins to decompose when the temperature
close to 90 °C. Then the negative material, positive material and elec-
trolyte start to interact [31–33]. Thus, it is necessary to detect the
battery fault during the EV operation process.
At present, based on features like temperature and voltage, many
researchers are dedicated to detecting battery faults and to predicting
battery system faults [30,32,34,35]. The battery pack is a largely sealed
space, in which the temperature of cells should be same or slightly
different. Generally, the temperature sensors are disturbed evenly in
pack to collect internal temperature and investigate temperature dis-
tribution of lithium-ion batteries. According to the collected tempera-
tures, two popular methods have been proposed to detect battery short
circuit faults: threshold and model methods. The temperature threshold
method will be classified into two parts. Firstly, the difference between
two temperature values should be calculated and then compared with
the threshold, when the difference value exceeds the preset value, the
fault-detection mechanism will give an alarm. Based on threshold
method, the temperature difference is usually set as less than 10 °C
[36–38]. The model method consists of two steps. The key step is to
establish temperature model for simulating the temperature distribu-
tion in pack. Afterwards, the normal operation temperature condition is

⁎
Corresponding author.
E-mail addresses: xiaoyu_li187@163.com (X. Li), wangzhenpo@bit.edu.cn (Z. Wang).

https://doi.org/10.1016/j.measurement.2017.11.034
Received 24 July 2017; Received in revised form 18 October 2017; Accepted 13 November 2017
Available online 14 November 2017
0263-2241/ © 2017 Elsevier Ltd. All rights reserved.
X. Li, Z. Wang
compared with the simulated temperature and when the difference
value exceeds the preset value, the alert mechanism will be triggered
[39–41]. Obviously, these two methods based on the temperature fea-
ture show three defects for detection battery fault. First, it will take a
relatively longer time to show the differences between model and actual
temperature value, therefore, bringing certain latency. Specifically, it is
time-consuming process to obtain the corresponding output tempera-
ture value by simulation model for the current computation ability of
BMS. Second, since the temperature sensor is evenly distributed among
the lithium-ion batteries instead of equipping each cell with a sensor, it
cannot locate the specific faulted battery. Third, the methods are rather
vulnerable to signal disturbance during the operation. Typically, the
method applies the differences between the simulation result and actual
sampling data to detect faults, however, the actual sampling data easily
suffer from the environment interference and contribute to the differ-
ences over the preset threshold.
Voltage-based fault detection methods are often employed to diagnose
and locate faults by collecting certain voltage signals [35,42–44]. Many
researchers have proposed considerable approaches to detect battery short
circuit faults, which can be divided into two groups, i.e., the threshold-
based method and model-based method. When there exists an external
short circuit accident in the battery, the increasing resistance is consistent
with temperature rise and the terminal voltage goes up rapidly. The
threshold method captures and analyses the terminal voltages of batteries
whether or not go beyond the preset voltage values [45–48]. Although
some vehicle original equipment manufacturers (OEMs) and BMS manu-
facturers set different failure levels, the limitation of the method is that the
voltage of short circuit battery may not trigger any preset value. To ad-
dress this issue, an improved threshold method has been proposed
[34,43]. The method needs to calculate the voltage difference between
any two terminal voltages and choose the maximum difference value
within battery pack that is then compared with threshold [43,45]. For this
method, the most difficult place is to say an appropriate threshold. If the
threshold is too high, the maximum difference is hard to touch the
threshold and the fault may not be detected timely, however, if the
threshold is too small, the alarm mechanism can be triggered easily and
maybe contribute to the mistake fault information.
The model-based method, basically, needs extensive experiments to
be conducted, which typically contains indentation [49], nail penetra-
tion [50], fabrication with defect structures [51] and extreme high
temperatures [31] four parts. Based on the data of aforementioned
experiments, the models of short circuit was built. During the EV op-
eration process, the model output voltages are compared with the ac-
tual output voltages of batteries. If the absolute value of the difference
between actual output voltage and model output voltage goes beyond
the threshold, the alarm system will be triggered [52,53]. In order to
overcome defects above, numerous articles and publications have built
multi-model structure, which comprises battery parameter estimation,
temperature and voltage model, for the sake of improving the reliability
and stability for fault detection [54,55]. In this method, the result of
fault detection is affected by batteries inconsistences and has a high
computing cost. Thus, there also exist some barriers to diagnose and
detect battery fault for battery pack.
Nowadays, however, there exists few publication that focuses on
detecting and diagnosing battery fault for the EV applications. In order
to deal with high computing costs, the influence of unbalance among
batteries and improving safety and stability for battery system, this
paper employs an interclass correlation coefficient (ICC) method to
detect the battery system accident in battery packs. This method cap-
tures the terminal voltages and calculates the ICC values among cells.
The ICC values indicate how strong a cell is compared with other cells
in the same group. The ICC method, which eliminates the voltage
sensors noise and inconsistency in batteries, can amplify the fluctua-
tions in group, and the method can avoid false diagnosis with the
battery at different states. In addition, this work collects the real-time
voltage data from the service and management center for electric
403
Measurement 116 (2018) 402–411
vehicles system (SMCEVS) and the fault detection process on computer
center, which not only reduces the computing load for BMS but also
improves safety and reliability.
The reminder of the paper is organized as follows: Section 2 de-
scribes a general presentation of the big data platform and the main
functions of SMCEVS. Section 3 gives a specific fault detection method.
The proposed method is applied to simulate and analyze in Section 4,
followed by the voltage fault analysis based on SMCEVS in Section 5.
Finally, Section 6 presents a general conclusion of this work.
2. Big data platform
Currently, some accidents of EVs have been worldwide reported
such as fire incidents and explosion hazards, which indicates there are
many potential safe problems in EVs. The problems of safety will
weaken the confidence of users and retard the market-penetration and
large-scale expansion of EVs. Therefore, it is imperative to develop an
effective system for safe operation. The system, which is the so-called
big data platform, not only includes monitoring and analysis of the real-
time data such as temperature, current, voltage, state of charge (SOC)
and so on, but also diagnose and predict the operation conditions of EVs
for manufactures and consumers by employing big data techniques. The
detailed information of SMCEVS is based on big data platform shown in
Fig. 1. In general, the SMCEVS mainly consists of two parts: the real-
time monitoring data and estimation data. The real-time monitored
data reflects the external features of EV and the estimation data can
explore deep conditions of batteries and so on.
The big data platform provides all-weather service for various ve-
hicles, including passenger cars, commercial vehicles, logistics vehicles
and buses. The monitoring information about vehicles is further pro-
cessed using corresponding strategy or/and algorithm. In practical ap-
plication, the battery pack is comprised of various series-parallels
structures. The battery module is composed of many battery cells.
Generally, the BMS collects the terminal voltages from the first battery
module to the last and the detailed process is shown in Fig. 2.
3. Detection method description
3.1. Battery fault problem statement
Currently, the detection methods of battery faults are divided into
two aspects: the hardware detection and model detection. The hard-
ware detection for battery fault needs a number of electric elements,
which not only leads to the complexity of hardware system but in-
creases the system cost. Hence, it is an impractical method for detecting
battery faults. And the model detection usually establish a consistent
model by means of various algorithms. Then the actual output value is
compared with the model output value. It is the difficulty of this
method to construct an accurate model, however, most systems are
nonlinear in practical problems. To sum up, the aforementioned
methods focus on detection of special single battery fault or individual
problem. For the battery pack, it is worthwhile to note that the cells
forming battery modules in parallels can generate internal circuit
equilibration. The battery pack comprises many modules in series and
the modules share the same input/output current. In normal condition,
the battery modules should have the same voltage. The battery fault
detection method needs to diagnose and locate the battery faults in a
short time. Thus, a new fault detection method should be proposed to
deal with battery fault problems. Based on the detailed investigation
above, the battery pack construction system should be simple and the
ICC detection method is described in Fig. 3.
3.2. The ICC method description
The ICC conception was firstly proposed by Ronald Fisher [56].
Afterwards, Bartko used the ICC algorithm to measure and evaluate the
X. Li, Z. Wang Measurement 116 (2018) 402–411

Fig. 1. The main functions of SMCEVS.

Speed Battery Vehicle

Collect Monitoring Battery Safety
Motor
Control Mangement

Voltage
Mismatch Voltage
Over- Over- Operation service
Charge Discharge
Battery and management Current Cell Energy
Safety and center for electric Monitorin
Short Park vehicle
Circuit
Protection Design g
Temper
Heat Extreme
ature
Dissipation
SOC SOH
Estimation Estimation
Battery
SOE
Status SOP
Estimation Estimation

reliability [57]. It described the proportion of individual variability to where the IC(x1,x2) is the ICC value between x1 and x2 , which is an
the total variability that was less than 1. Generally, the value less than unitless value and ranges from 0 to 1. The denominator S 2 can be cal-
0.4 suggests the test data is incredible, while the value more than 0.75 culated as
illustrates that the data is basically credible [58,59]. Considering a data
n n
set consists of N paired data values (x i,1,x i,2) , for n = 1,2…N , the ICC 1 ⎡ ⎤
calculation formula can be described as,
S2 = ∑ (xi,1−x )2 +
2n ⎢ i = 1
∑ (x i,2−x )2⎥
⎣ i=1 ⎦ (2)

n
∑i = 1 (x i,1−x )(x i,2−x ) where the x is mean value of data x i,1 and x i,2 . The specific calculating
IC(x1,x2) = process is as follows,
nS 2 (1)

Fig. 2. The battery pack construction.

== == == ==

== == == ==

Battery cell
Battery module

== == == ==

== == == ==

404
X. Li, Z. Wang Measurement 116 (2018) 402–411

IC(V1,V2)

=
n
( 1 n
2 ∑i = 1 Vi,1− 2n ∑i = 1 (Vi,1 + Vi,2) )(V 1
i,2− 2n
n
∑i = 1 (Vi,1 + Vi,2) )
2 2
n
∑i = 1 ( 1
Vi,1− 2n
n
∑i = 1 (Vi,1 + Vi,2) ) +
n
∑i = 1 ( 1
Vi,2− 2n
n
∑i = 1 (Vi,1 + Vi,2) )
(7)

The ICC calculating process can be easily applied for online appli-
cation. Meanwhile, the data needs to accumulate during the initial
stage until the operation time i is more than the size of moving windows
(n > m) . Hence, the final recursive formulas after the transformation
can be expressed as

n
1
Ri,1 = Vi,1− 2m ∑ (Vi,1 + Vi,2)
i=n−m
n
1
Ri,2 = Vi,2− 2m ∑ (Vi,1 + Vi,2)
i=n−m (8)

Generally, the terminal voltages of two adjacent cells are the same
under the various driving cycles. The ICC value ranges from 0.75 to 1
for the normal operation condition. And in this range, the voltage
sampling error and noise factors can be filtered. When the ICC value is
Fig. 3. The ICC calculation for a particular vehicle. less than 0.75, there is likely to be potential fault in the cell. Hence, the
ICC value can reflect the battery fault when the EV is operating under
n abnormal condition. The specific schematic diagram is presented to
1
x =
2n
∑ (x i,1 + x i,2) explain the ICC method for battery pack fault detection as shown in
i=1 (3) Fig. 4. The overall process is divided into three parts including Ex-
Based on (2) and (3), the ICC calculation formulas is redefined as traction Data, Fault Detection and Fault Location. The three parts will
follows, be depth analysis in next sections.

IC(x1,x2)
4. Simulation analysis
=
n
( 1 n
2 ∑i = 1 xi,1− 2n ∑i = 1 (xi,1 + x i,2) )(x 1
i,2− 2n
n
∑i = 1 (xi,1 + x i,2) )
2 2
n
( 1 n
∑i = 1 xi,1− 2n ∑i = 1 (xi,1 + x i,2) ) n
( 1 n
+ ∑i = 1 xi,2− 2n ∑i = 1 (xi,1 + x i,2) ) 4.1. Normal condition
(4)
In order to verify the stability and feasibility, two cells connecting in
For the sake of iterative regression calculation, the numerator and series are chosen to acquire test data and the cell parameters are listed
denominator can be expressed by two symbols as in Table 1. Based on specific current changes in driving cycles, the cells
n charging/discharging experiment is carried out in Arbin tester, BT-2000
⎛ 1 ⎞
Ri,1 = ⎜xi,1− 2n ∑ (xi,1 + x i,2) ⎟ under ambient temperature. It is worth mentioning that the current
⎝ i=1 ⎠ load of driving cycles is extracted from SMCEVS, then the specific
⎛ 1
n
⎞ current load is transformed based on corresponding vehicle type (the
Ri,2 = ⎜xi,2− 2n ∑ (xi,1 + x i,2)⎟ number of series and parallel of battery system) for single cell. The
⎝ i=1 ⎠ (5) current load consists of acceleration, deceleration and idling three
Thus, Eq. (4) can be simplified as parts.
n
The corresponding two voltage response curves are shown in
2 ∑ Ri,1 Ri,2 Fig. 5(a). When the cells have strong consistency, differences between
i=1 two terminal voltages under the same output current are small. In depth
IC(X1,X2) = n n .
analysis, the small differences may result from measurement noise and/
∑ Ri2,1 + ∑ Ri2,1
(6) or small resistance diversity. During this condition, the availability of
i=1 i=1
proposed method is verified by calculating the ICC value and ICC value
For the online battery fault detection, the ICC value should be cal- is shown in Fig. 5(b). From the Fig. 5(b), the ICC value is more than
culated in a sequential manner. However, it is difficult to calculate all 0.75 in total operation time. This value demonstrates the batteries have
data during the sampling process. A forgetting mechanism is employed perfect consistency. Compared with the Fig. 5(a), although the voltage
to deal with the problem, which designs a moving window to update response curves have extreme fluctuations that contain high charging/
the history data. discharging rate, the ICC values are showing a tendency to stabilization
Generally speaking, the cells in battery packs are not in complete and close to 1. Additionally, during the idling process, the two voltage
uniformity because they contain many influencing factors such as curves have a gradually rising tendency. However, the ICC values are
production process, manufacturing techniques, utilization process and still remaining unchanged and only relate to the difference between two
environment. When short circuit fault occurs in the cell, the voltage voltages. The results indicate the ICC fault detection method has better
signs are able to change immediately. Based on the ICC method, N is set stability and strong robustness. It is also worth pointing out the ICC
as the total operation time in a driving cycle and m is the moving value remains almost constant at first thirty seconds because the size of
windows for simplifying calculation. The data set (x i,1,x i,2) expresses the moving window equals 30. During this process, the ICC value keeps
cell voltages of V1 and V2 in sampling time i . To sum up, the ICC method fixed value that is calculated at first step. This approach is designed to
is applied to battery fault detection and formula (4) can be redefined as form the initial group of the ICC method.
follows,

405
X. Li, Z. Wang Measurement 116 (2018) 402–411

Fig. 4. The schematic diagram of fault detection

Extraction Data Fault Detection Fault Location process.

Extraction Data Calculation ICC BMS

from SMCEVS value Fault Diagnosis

Capture Fault Confirmation

Data Cleaning Signal Vehicle Type

If No Salve system
Terminal voltage Fault location
signals ICC >0.75

Yes
Normal
Fault Alarm
condition

Table 1 whole process maintains 30 sampling points which always equals to the
Battery parameters. size of moving window. Afterwards, the ICC rapidly recovers to the
normal value. For the signals transmission between BMS and the big
Parameter type Cylindrical 26650
data platform, the fault retention time is much longer than signals
Rated voltage 3.3 V transmission and computation time. Thus, the battery fault detection
Rated capacity 5 Ah method can capture the fault signal during the actual operation process.
Charge cut-off voltage 3.65 V When short circuit accident occurs, the battery terminal voltage sud-
Discharge cut-off voltage 2.0 V
Max discharge rate 5C
denly drops in less than five seconds, however, the battery detection
method can extend the fault time and memorize the fault phenomenon.
During the fault detection process, the detection speed depends on the
4.2. Abnormal condition sampling speed to a certain extent. From another perspective, the result
of battery fault detection is also influenced by the size of moving
During the abnormal condition, the battery fault detection method windows. In this study, three different sizes of moving windows are
should appropriately and rapidly inform of the battery fault. Based on employed to verify the fault detection speed. Hence, three ICC values
the battery fault principle, the different voltage signals are built by are calculated based on different window sizes, as shown in Fig. 7. The
reducing one cell voltage in a short time. The fault signals between two battery fault memory time increases with increasing window sizes.
batteries are constructed by a sudden voltage drop at the one hundredth When the window size is set as 50, the fault memory time is extended to
sampling points and the max voltage drop is close to 100 mV. The 150 s. Although the fault memory time is extended to some extent, the
specific voltage faults signals are plotted in Fig. 6(a). The ICC value for sensitivity of fault detection decreases and the ICC value becomes
the fault voltage signal is calculated in Fig. 6(b). Obviously, the ICC smaller. When the moving window sizes are set as 30 or 40, the ICC
value suddenly drops below 0.75 at the 100th sampling point and the values are under 0.75 at 100th sampling point. However, the ICC value

1.005
3.36
V1
V2
3.34 1
ICC Coefficient

3.32
Voltage(V)

0.995

3.3
0.99
3.28
0.985
3.26

3.24 0.98
0 100 200 300 400 500 600 700 800 900 0 100 200 300 400 500 600 700 800 900

Time(s) Time(s)

(a) (b)
Fig. 5. Two cells voltage curves and corresponding ICC value.

406
X. Li, Z. Wang Measurement 116 (2018) 402–411

1
V1
3.32 FV2 0.95
3.3
0.9
3.28

ICC Coefficient
0.85
Voltage(V) 3.26
3.24 0.8

3.22 3.3 0.75

3.2 0.7
3.2
3.18
0.65
3.16 70 80 90 100 110 120

0 100 200 300 400 500 600 700 800 900 0 100 200 300 400 500 600 700 800 900
Time(s) Time(s)

(a) (b)
Fig. 6. The fault voltage signals and corresponding ICC value.

1 200
N=30

Current(A)
0.95 N=40 0
N=50
ICC Coefficient

0.9
-200
0.85
-400
0.8 0 100 200 300 400 500 600 700 800
0.9
0.75 0.8 205

Voltage(V)
0.7 0.7
0.65 200
100 150

0 100 200 300 400 500 600 700 800 900 195
Time(s) 0 100 200 300 400 500 600 700 800
Time(10s)
Fig. 7. The results based on different window sizes.
Fig. 10. The current and voltage of EV practical operation.

3.4
V1
FV2
moving window is set as 30, which has obvious advantages including
3.35 V3 not only a sensitive ICC value but also a faster fault detection speed.
Voltage(V)

3.3 4.3. Fault location

3.25 The fault location is another important part for battery fault diag-
nosis. When the battery fault happens, the ICC value suddenly drop and
3.2
the location of battery fault should be identified immediately. Three
terminal voltages data are captured from series cells under the same
charge/discharge current. A voltage drop around 100 mV is set as 100th
0 100 200 300 400 500 600 700 800 900 second of the second cell. The specific voltage profiles are plotted in
Time(s)
Fig. 8. The first and third cells are normal with good consistency, while
Fig. 8. The battery voltage signals. the second cell has a maximum difference with those two cells at the
100th second. Based on the ICC fault diagnosis method, the ICC values
are calculated depending on number of voltages or cells. The number of
ICC(1,2)

0.9 ICC values can be summarized as,

0.8
0.7
0 100 200 300 400 500 600 700 800 900
{ n, n > 2
1, n = 2 (9)
ICC(2,3)

0.9
0.8
0.7 where n is the number of voltages. Therefore, three ICC values are
0 100 200 300 400 500 600 700 800 900 calculated for location the battery fault and the values are described in
1
ICC(1,3)

Fig. 9. The ICC values of ICC(1,2) and ICC(2,3) are less than 0.75 after the
100th second. Those two values indicate that at least one cell happens
0.95 fault. In addition, the ICC value of ICC(1,3) is more than 0.75 during
0 100 200 300 400 500 600 700 800 900
Time(s)
overall testing procedure. The value manifests the first and third cells
are not occurring short circuit fault. In conclusion, the fault cell can be
Fig. 9. The ICC values between two battery voltages. located in second cell at 100th second.

5. Voltage fault analysis based on SMCEVS

is over 0.75 at 100th sampling point when the moving window size
equals 50. After 10 s, the ICC value gradually falls to less than 0.75. The cell datasets are extracted from big data platform and the
From the overall operation stage, the last ICC value is more than those specific fault vehicle: No. GA0083X. The data period was two hours
of the first two types. To sum up, the size of moving window should be with a data acquisition frequency of 0.1 Hz. It is well known that an
considered to fault detection speed. Compared with the cases when the electric vehicle consists of hundreds of battery modules. Therefore, a
sizes of moving windows are set as 40 and 50, in this study, the size of large number of voltages are collected by salve systems. Then, the

407
X. Li, Z. Wang Measurement 116 (2018) 402–411

3360 1.05
ICC1
V1
3340 ICC2
V2
1 ICC3
V3
3320 ICC4
V4

Voltage(mV)
ICC5
V5

ICC Value
0.95
3300 V6
ICC6
ICC7
V7
3280 V8
ICC8
0.9
ICC9
V9
3260 V10
ICC10
0.85 ICC11
V11
3240 ICC12
V12

3220 0.8
0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800
Time(10s) Time(10s)

(a) (b)
Fig. 11. The voltages and ICC values of the first salve system.

3360 1.05
V1 ICC1
3340 V2 1 ICC2
V3 ICC3
0.95
3320 V4 ICC4
Voltage(mV)

V5 0.9 ICC5

ICC Value
3300 V6 ICC6
V7 0.85 ICC7
3280 V8 ICC8
V9 0.8
ICC9
3260 V10 ICC10
0.75
V11 ICC11
3240 V12 0.7 ICC12

3220 0.65
0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800
Time(10s) Time(10s)

(a) (b)
Fig. 12. The voltages and ICC values of the second salve system.

3360 1.05
V1 ICC1
3340 V2 ICC2
V3 1
ICC3
3320 V4 ICC4
Voltage(mV)

V5
ICC Value

0.95 ICC5
3300 V6 ICC6
V7 ICC7
3280
V8 0.9 ICC8
V9 ICC9
3260
V10 ICC10
V11 0.85
ICC11
3240
V12 ICC12

3220 0.8
0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800

Time(10s) Time(10s)

(a) (b)
Fig. 13. The voltages and ICC values of the third salve system.

voltage data of salve systems is uploaded to BMS. In this study, the salve
system can collect twelve voltages each time. Based on the big data
platform, sixty groups voltages are extracted from different five salve
systems. The total voltage of sixty groups voltages and charge/dis-
charge current of EV practical operation are described in Fig. 10. The
overall driving cycle comprises of acceleration, deceleration and idle
three parts through analysis the current changes. The maximum abso-
lute value of charge/discharge current is around 200 A. The total vol-
tage changes with the current. The voltage decreases (increases) under
discharge (charge). The data can reflect the daily operation condition.
The sampling voltages of five salve systems are verified based on the
proposed ICC method. Twelve voltages of each salve system are set as
an accuracy to three decimal places, for the sake of manifesting the
voltages changes clearly. The data of each salve system is flagged for
locating the voltages. And twelve ICC values can be calculated between
two voltages. The specific fault detection analysis method is described
in Figs. 11–15. Fig. 11(a) shows the voltage curves of the first salve
system, obviously, the voltage curves are inconsistent under practical
operation situation. It is evident that the seventh voltage in black varies
greatly when comparing with others. Based on the proposed fault de-
tection method, the ICC values can demonstrate this difference clearly.
In Fig. 11(b), the ICC values of ICC(6,7) and ICC(7,8) have less values
compared with other values. According to the fault detection method,
the seventh voltage of first salve system has fault potential.
Meanwhile, the voltages of the second salve system are extracted to
analyze, and the voltages and ICC values are plotted in Fig. 12. Ob-
viously, Fig. 12(a) shows the tenth voltage of the second salve system is
higher (lower) than other voltages under charge (discharge) condition.
Similarly, the ICC values of ICC(9,10) and ICC(10,11) are between 0.95 and
0.85 on the whole. In the initial stage, the ICC value of ICC(9,10) is less

408
X. Li, Z. Wang Measurement 116 (2018) 402–411

3500 1.4
V1 ICC1
3000 V2 1.2 ICC2
V3 ICC3
2500 V4 1 ICC4
Voltage(mV)
V5

ICC Value
ICC5
2000 V6 0.8
ICC6
V7 0.75 ICC7
1500 V8 0.6
0.5 ICC8
V9
1000 0.4 0.25
ICC9
V10
ICC10
V11 0
500 V12
0.2 430 440 450 460 470 480 ICC11
ICC12
0 0
0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800
Time(10s) Time(10s)

(a) (b)
Fig. 14. The voltages and ICC values of the fourth salve system.

3360 1.05
V1 ICC1
3340 V2 ICC2
1
V3 ICC3
3320 V4 ICC4
Voltage(mV)

V5

ICC Value
0.95 ICC5
3300 V6 ICC6
V7 ICC7
3280 V8 0.9 ICC8
V9 ICC9
3260 V10 ICC10
V11 0.85
ICC11
3240 V12 ICC12

3220 0.8
0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800

Time(10s) Time(10s)

(a) (b)
Fig. 15. The voltages and ICC values of the fifth salve system.

than 0.75, which is the preset threshold of normal condition. Therefore,
the two-level alarm mechanism can be triggered and the process maybe
keep 300 s. Because the fault occurs at the initial stage and ICC value is
updated after the moving windows work. Compared with other ICC
values at the initial stage, the values are less than the values of practical
operation. To sum up, the ICC value are below 0.75 at the initial stage
should be neglected and this phenomenon may be contributed by the
measure noises.
Based on the proposed fault detection method and Eq. (9), the ICC
value is calculated in a loop from the first to last voltages for sake of
detecting the fault of first and last battery modules. It is a quite smart
method for locating battery fault detection. Fig. 13(a) shows the last
voltage of the fourth salve system has great fluctuation. However, it is
difficult to determine the fault degree of the voltage and locate the fault
just through voltage fluctuation condition. The ICC values of ICC(11,12)
and ICC(12,1) can easily find the voltage location and the specific ICC
value can reflect fault degree on some extent. The ICC values of the
third salve system are showing in Fig. 13(b).
The short circuit of the battery is collected by the fourth salve
system and the fault voltage signal appears and recovers in few sam-
pling points. In addition, the fault voltage signals may be ignored due to
the data transfers frequency. Thus, it is important for the fault voltage
signals to be retention. The moving windows of this proposed fault
detection method have significant effects on extension voltage signals.
The fault voltage signals are shown in Fig. 14(a), the signals just keep
few sampling point, specifically, the fault signals only last 20 s. How-
ever, the ICC values of ICC(2,3) and ICC(3,4) can capture the fault signals
and prolong the signals, which are specifically described in Fig. 14(b).
The fault signals are retention 30 sampling points that value equals to
the window size and the fault signals are extended to 300 s. Thus, the
fault can be detected and captured timely.
In the fifth salve system, the voltage curves fluctuate in normal
scope besides the tenth voltage curve. The specific voltages change are
shown in Fig. 15(a). The ICC values can explain the voltages condition
of the fifth salve system. Fig. 15(b) shows all ICC values are more than
0.75 and the values range from 0.9 to 1 on the whole. By special con-
siderations of the ICC values, the tenth voltage maybe has a fault po-
tential due to the lesser ICC values between ICC(9,10) and ICC(10,11) va-
lues. In addition, the primary reasons are measure noise, production
technology and grouping technique for ICC values are less than 1.
However, those problems can be neglected during the voltage fault
detection process.
6. Conclusion
In this paper, a novel voltage fault detection method is proposed
based on the service and management center for electric vehicles
system of electric vehicles. The concept of the interclass correlation
coefficient is first introduced and then the interclass correlation coef-
ficient is applied to analyze battery short circuit fault by capturing the
off-trend voltage drop. The coefficient value not only has advanced
fault resolution by amplifying the voltage difference, but also can
prolong the fault memory by setting moving windows. In addition, the
proposed method has a feasible calculation processing, which does not
need to build model and redundant hardware design. And a smart fault
location method is completed by designing a loop joints the first and
last voltages.
Moreover, simulation and experiment are employed to validate and
analyze the voltage faults. First, simulation is applied to verify the
method availability by collecting three group voltages from Service and
Management Center of electric vehicles. Meanwhile, the sensitivity of
fault signals is analyzed by setting three different sizes of moving

409
X. Li, Z. Wang
windows. The experiment is proposed to verify the method for battery
fault of electric vehicles and the voltage data is extracted from service
and management center of electric vehicles. To sum up, the proposed
method not only detects and located voltage faults, but also can reflect
fault potential. In addition, the method can neglect the inevitable ef-
fects such as measure noise, production technology and grouping
technique.
Further work will focus on battery current and temperature ab-
normality and those phenomena are detected by this proposed method.
In addition, the method can be applied to big data platform for other
fault detection fields.
Acknowledgements
The authors would like to acknowledge the funding support from
the State Key Program of National Natural Science Foundation of China
under Grant No. U1564206.
Appendix A. Supplementary material
Supplementary data associated with this article can be found, in the
online version, at http://dx.doi.org/10.1016/j.measurement.2017.11.
034.
References
[1] M.A. Hannan, M.M. Hoque, A. Mohamed, A. Ayob, Review of energy storage sys-
tems for electric vehicle applications: issues and challenges, Renew. Sustain. Energy
Rev. 69 (2017) 771–789.
[2] J. Du, D. Ouyang, Progress of Chinese electric vehicles industrialization in 2015: a
review, Appl. Energy 188 (2017) 529–546.
[3] C. Zou, C. Manzie, D. Nešić, A framework for simplification of PDE-based lithium-
ion battery models, IEEE Trans. Control Syst. Technol. 24 (2016) 1594–1609.
[4] Z. Chen, X. Li, J. Shen, W. Yan, R. Xiao, A novel state of charge estimation algorithm
for lithium-ion battery packs of electric vehicles, Energies 9 (2016) 710.
[5] M. Dubarry, N. Vuillaume, B.Y. Liaw, From single cell model to battery pack si-
mulation for Li-ion batteries, J. Power Sources 186 (2009) 500–507.
[6] L. Lu, X. Han, J. Li, J. Hua, M. Ouyang, A review on the key issues for lithium-ion
battery management in electric vehicles, J. Power Sources 226 (2013) 272–288.
[7] H.J. Bergveld, W.S. Kruijt, P.H.L. Notten, Battery Management Systems, in: Battery
Management Systems: Design by Modelling, Springer Netherlands, Dordrecht,
2002, pp. 9–30.
[8] Z. Wei, C. Zou, F. Leng, B.H. Soong, K.J. Tseng, Online model identification and
state of charge estimate for lithium-ion battery with a recursive total least squares-
based observer, IEEE Trans. Industr. Electron. (2017).
[9] Y. Wang, C. Zhang, Z. Chen, J. Xie, X. Zhang, A novel active equalization method for
lithium-ion batteries in electric vehicles, Appl. Energy 145 (2015) 36–42.
[10] J. Wei, G. Dong, Z. Chen, Y. Kang, System state estimation and optimal energy
control framework for multicell lithium-ion battery system, Appl. Energy 187
(2017) 37–49.
[11] Y. Wang, Z. Chen, C. Zhang, On-line remaining energy prediction: a case study in
embedded battery management system, Appl. Energy 194 (2017) 688–695.
[12] K.W.E. Cheng, B.P. Divakar, H. Wu, K. Ding, H.F. Ho, Battery-management system
(BMS) and SOC development for electrical vehicles, IEEE Trans. Veh. Technol. 60
(2011) 76–88.
[13] Y. Wang, C. Zhang, Z. Chen, A method for joint estimation of state-of-charge and
available energy of LiFePO4 batteries, Appl. Energy 135 (2014) 81–87.
[14] A. Bartlett, J. Marcicki, K. Rhodes, G. Rizzoni, State of health estimation in com-
posite electrode lithium-ion cells, J. Power Sources 284 (2015) 642–649.
[15] D. Andre, C. Appel, T. Soczka-Guth, D.U. Sauer, Advanced mathematical methods of
SOC and SOH estimation for lithium-ion batteries, J. Power Sources 224 (2013)
20–27.
[16] C. Zou, C. Manzie, D. Nešić, A.G. Kallapur, Multi-time-scale observer design for
state-of-charge and state-of-health of a lithium-ion battery, J. Power Sources 335
(2016) 121–130.
[17] Y. Wang, C. Zhang, Z. Chen, An adaptive remaining energy prediction approach for
lithium-ion batteries in electric vehicles, J. Power Sources 305 (2016) 80–88.
[18] H. Dong, X. Jin, Y. Lou, C. Wang, Lithium-ion battery state of health monitoring and
remaining useful life prediction based on support vector regression-particle filter, J.
Power Sources 271 (2014) 114–123.
[19] J. Wu, C. Zhang, Z. Chen, An online method for lithium-ion battery remaining
useful life estimation using importance sampling and neural networks, Appl. Energy
173 (2016) 134–140.
[20] H. Wang, A. Kumar, S. Simunovic, S. Allu, S. Kalnaus, J.A. Turner, et al., Progressive
mechanical indentation of large-format Li-ion cells, J. Power Sources 341 (2017)
156–164.
[21] J. Sun, J. Li, T. Zhou, K. Yang, S. Wei, N. Tang, et al., Toxicity, a serious concern of
thermal runaway from commercial Li-ion battery, Nano Energy 27 (2016) 313–319.
410
Measurement 116 (2018) 402–411
[22] S. Abada, G. Marlair, A. Lecocq, M. Petit, V. Sauvant-Moynot, F. Huet, Safety fo-
cused modeling of lithium-ion batteries: a review, J. Power Sources 306 (2016)
178–192.
[23] N. Meethong, H.Y.S. Huang, S.A. Speakman, W.C. Carter, Y.M. Chiang, Strain ac-
commodation during phase transformations in olivine-based cathodes as a materials
selection criterion for high-power rechargeable batteries, Adv. Func. Mater. 17
(2007) 1115–1123.
[24] H. Maleki, J.N. Howard, Effects of overdischarge on performance and thermal
stability of a Li-ion cell, J. Power Sources 160 (2006) 1395–1402.
[25] M. Chen, F. Bai, W. Song, J. Lv, S. Lin, Z. Feng, et al., A multilayer electro-thermal
model of pouch battery during normal discharge and internal short circuit process,
Appl. Therm. Eng. 120 (2017) 506–516.
[26] B. Xia, Y. Shang, T. Nguyen, C. Mi, A correlation based fault detection method for
short circuits in battery packs, J. Power Sources 337 (2017) 1–10.
[27] Z. Li, J. Huang, B.Y. Liaw, J. Zhang, On state-of-charge determination for lithium-
ion batteries, J. Power Sources 348 (2017) 281–301.
[28] H. Dai, P. Guo, X. Wei, Z. Sun, J. Wang, ANFIS (adaptive neuro-fuzzy inference
system) based online SOC (State of Charge) correction considering cell divergence
for the EV (electric vehicle) traction batteries, Energy 80 (2015) 350–360.
[29] P. Peng, F. Jiang, Thermal safety of lithium-ion batteries with various cathode
materials: a numerical study, Int. J. Heat Mass Transf. 103 (2016) 1008–1016.
[30] R. Zhao, J. Liu, J. Gu, Simulation and experimental study on lithium ion battery
short circuit, Appl. Energy 173 (2016) 29–39.
[31] X. Feng, M. Fang, X. He, M. Ouyang, L. Lu, H. Wang, et al., Thermal runaway
features of large format prismatic lithium ion battery using extended volume ac-
celerating rate calorimetry, J. Power Sources 255 (2014) 294–301.
[32] S. Dey, Z.A. Biron, S. Tatipamula, N. Das, S. Mohon, B. Ayalew, et al., On-board
thermal fault diagnosis of lithium-ion batteries for hybrid electric vehicle applica-
tion, IFAC-PapersOnLine 48 (2015) 389–394.
[33] T. Inoue, K. Mukai, Roles of positive or negative electrodes in the thermal runaway
of lithium-ion batteries: accelerating rate calorimetry analyses with an all-inclusive
microcell, Electrochem. Commun. 77 (2017) 28–31.
[34] A. Sidhu, A. Izadian, S. Anwar, Adaptive nonlinear model-based fault diagnosis of
Li-ion batteries, IEEE Trans. Industr. Electron. 62 (2015) 1002–1011.
[35] S. Dey, S. Mohon, P. Pisu, B. Ayalew, Sensor fault detection, isolation, and esti-
mation in lithium-ion batteries, IEEE Trans. Control Syst. Technol. 24 (2016)
2141–2149.
[36] Z. Li, J. Zhang, B. Wu, J. Huang, Z. Nie, Y. Sun, et al., Examining temporal and
spatial variations of internal temperature in large-format laminated battery with
embedded thermocouples, J. Power Sources 241 (2013) 536–553.
[37] C. Forgez, D. Vinh Do, G. Friedrich, M. Morcrette, C. Delacourt, Thermal modeling
of a cylindrical LiFePO4/graphite lithium-ion battery, J. Power Sources 195 (2010)
2961–2968.
[38] X. Lin, H.E. Perez, J.B. Siegel, A.G. Stefanopoulou, Y. Li, R.D. Anderson, et al.,
Online parameterization of lumped thermal dynamics in cylindrical lithium ion
batteries for core temperature estimation and health monitoring, IEEE Trans.
Control Syst. Technol. 21 (2013) 1745–1755.
[39] G.J. Offer, V. Yufit, D.A. Howey, B. Wu, N.P. Brandon, Module design and fault
diagnosis in electric vehicle batteries, J. Power Sources 206 (2012) 383–392.
[40] K.-Z. Lee, B.J. Dougherty, M.S. Sandhu, M.A. Lane, P.J. Reier, D.D. Fuller, Phrenic
motoneuron discharge patterns following chronic cervical spinal cord injury, Exp.
Neurol. 249 (2013) 20–32.
[41] G.-H. Kim, A. Pesaran, R. Spotnitz, A three-dimensional thermal abuse model for
lithium-ion cells, J. Power Sources 170 (2007) 476–489.
[42] B. Xia, C. Mi, A fault-tolerant voltage measurement method for series connected
battery packs, J. Power Sources 308 (2016) 83–96.
[43] A. Sidhu, A. Izadian, S. Anwar,“Model-based adaptive fault diagnosis in lithium ion
batteries: a comparison of linear and nonlinear approaches, SAE Technical Paper
0148-7191, 2017.
[44] C. Wu, C. Zhu, Y. Ge, Y. Zhao, A review on fault mechanism and diagnosis approach
for Li-Ion batteries, J. Nanomaterials 2015 (2015) 8–18.
[45] L. Yao, Z. Wang, J. Ma, Fault detection of the connection of lithium-ion power
batteries based on entropy for electric vehicles, J. Power Sources 293 (2015)
548–561.
[46] Y. Zheng, X. Han, L. Lu, J. Li, M. Ouyang, Lithium ion battery pack power fade fault
identification based on Shannon entropy in electric vehicles, J. Power Sources 223
(2013) 136–146.
[47] Z. Wang, J. Hong, P. Liu, L. Zhang, Voltage fault diagnosis and prognosis of battery
systems based on entropy and Z-score for electric vehicles, Appl. Energy 196 (2017)
289–302.
[48] G. Zuchang, C.S. Chin, W. Wai Lok, J. Junbo, T. Wei Da, Genetic algorithm based
back-propagation neural network approach for fault diagnosis in lithium-ion bat-
tery system, in: 2015 6th International Conference on Power Electronics Systems
and Applications (PESA), 2015, pp. 1–6.
[49] L. Greve, C. Fehrenbach, Mechanical testing and macro-mechanical finite element
simulation of the deformation, fracture, and short circuit initiation of cylindrical
Lithium ion battery cells, J. Power Sources 214 (2012) 377–385.
[50] H. Maleki, J.N. Howard, Internal short circuit in Li-ion cells, J. Power Sources 191
(2009) 568–574.
[51] C.J. Orendorff, E.P. Roth, G. Nagasubramanian, Experimental triggers for internal
short circuits in lithium-ion cells, J. Power Sources 196 (2011) 6554–6558.
[52] A. Singh, A. Izadian, S. Anwar, Model based condition monitoring in lithium-ion
batteries, J. Power Sources 268 (2014) 459–468.
[53] Z. Liu, H. He, Model-based sensor fault diagnosis of a lithium-ion battery in electric
vehicles, Energies 8 (2015) 6509.
[54] Y. Wang, B. Jiang, N. Lu, J. Pan, Hybrid modeling based double-granularity fault
X. Li, Z. Wang Measurement 116 (2018) 402–411

detection and diagnosis for quadrotor helicopter, Nonlinear Anal.: Hybrid Syst. 21
(2016) 22–36.
[55] G.R. Marseglia, D.M. Raimondo, Active fault diagnosis: a multi-parametric ap-
proach, Automatica 79 (2017) 223–230.
[56] R.A. Fisher, Statistical Methods for Research Workers, Genesis Publishing Pvt Ltd,
1925.
[57] J.J. Bartko, The intraclass correlation coefficient as a measure of reliability,
Psychol. Rep. 19 (1966) 3–11.
[58] P.E. Shrout, J.L. Fleiss, Intraclass correlations: uses in assessing rater reliability,
Psychol. Bull. 86 (1979) 420–428.
[59] M.W. Post, What to do with “Moderate” reliability and validity coefficients? Arch.
Phys. Med. Rehabil. 97 (2016) 1051–1052.

411