12636 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 37, NO.

10, OCTOBER 2022

An Online Data-Driven Fault Diagnosis and Thermal

Runaway Early Warning for Electric Vehicle Batteries
Zhenyu Sun , Zhenpo Wang , Senior Member, IEEE, Peng Liu , Zian Qin , Senior Member, IEEE,
Yong Chen , Yang Han, Peng Wang, and Pavol Bauer , Senior Member, IEEE

Abstract—Battery fault diagnosis is crucial for stable, reliable, density, long cycle life compared with the counterparts, such as
and safe operation of electric vehicles, especially the thermal run- lead–acid batteries, nickel–cadmium batteries, nickel–hydrogen
away early warning. Developing methods for early failure detection batteries, etc. [2]. To meet the voltage, power, and energy re-
and reducing safety risks from failing high energy lithium-ion
batteries has become a major challenge for industry. In this article, quirements of the applications, a typical LIB system contains
a real-time early fault diagnosis scheme for lithium-ion batteries many battery cells connected in series and parallel [3], [4]. It
is proposed. By applying both the discrete Fréchet distance and is of high importance and also quite challenging to maintain
local outlier factor to the voltage and temperature data of the safety of battery packs under complex operating conditions [5].
battery cell/module that measured in real time, the battery cell Recently, the increasing number of reported EV fire accidents
that will have thermal runaway is detected before thermal runaway
happens. Compared with the widely used single parameter based has drawn much attention, most of which are due to battery
diagnosis approach, the proposed one considerably improve the system catching fire [6], [7]. As a result, the fault diagnosis and
reliability of the fault diagnosis and reduce the false diagnosis early detection of thermal runaway of EVs become essential.
rate. The effectiveness of the proposed method is validated with However, the failure mechanism of LIB system is extremely
the operational data from electric vehicles with/without thermal complicated because it is a nonlinear time-varying system with
runaway in daily use.
a mass of dynamic electrochemical and mechanical phenomena
Index Terms—Discrete Fréchet distance (DFD), fault diagnosis, [8]–[10]. Faults of the LIB system can be categorized into
lithium-ion battery (LIB), local outlier factor (LOF). internal and external ones [11]. The former include overcharge,
overdischarge, internal short circuit, external short circuit, and
I. INTRODUCTION
overheat. The latter include fault of sensors, connection, and
ITH the ambition to achieve carbon neutrality in 2050, cooling system. Both of the two types of faults can lead to
W vehicle electrification has been attracting a significant
amount of attention [1]. Lithium-ion batteries (LIBs) are widely
abnormal voltages, temperatures, and pressure in the battery
pack [12], [13]. Nonetheless, there is a higher possibility that
used for applications on electric vehicles (EVs) due to their rel- the internal faults cause a fire accident or explosion rather than
atively low self-discharge rates, high energy density, high power the external faults.
The fault diagnosis approaches of LIB systems can be classi-
Manuscript received September 10, 2021; revised December 28, 2021 and fied into three types: rule-based, model-based, and data-driven
March 3, 2022; accepted April 28, 2022. Date of publication May 6, 2022;
date of current version June 24, 2022. This work was supported in part by the methods. For rule-based methods, if the value of critical pa-
National Natural Science Foundation of China under Grant 52072040, in part rameters, such as voltage and temperature, exceed the spec-
by China Scholarship Council under Grant 202006030153, and in part by the ified threshold, an alarm will be tripped. For example, Pe-
Ministry of Science and Technology of the People’s Republic of China under
Grant 2019YFE0107900. Recommended for publication by Associate Editor T. saran et al. [14] suggested the maximum temperature difference
Mishima. (Corresponding authors: Zian Qin; Yong Chen.) between modules should be less than 5 °C. To diagnose external
Zhenyu Sun is with the National Engineering Laboratory for Electric Vehicles, short circuit fault, Xia et al. [15] proposed three diagnostic
Beijing Institute of Technology, Beijing 100811, China, and also with the
Department of Electrical Sustainable Energy, Delft University of Technology, criteria utilizing measurement parameters, such as the voltage,
2628 CD Delft, The Netherlands (e-mail: [email protected]). current, and temperature increase. However, the appropriate
Zhenpo Wang and Peng Liu are with the National Engineering Laboratory threshold is hard to define, due to the complex operating cir-
for Electric Vehicles, Beijing Institute of Technology, Beijing 100811, China
(e-mail: [email protected]; [email protected]). cumstances of the EV. If the threshold is too low, the alarm can
Zian Qin and Pavol Bauer are with the Department of Electrical Sustainable be too sensitive and easily faulty tripped; on the other hand, if
Energy, Delft University of Technology, 2628, CD Delft, The Netherlands (e- the threshold is too high, when the alarm is tripped, severe faults
mail: [email protected]; [email protected]).
Yong Chen is with the School of Electromechanical Engineering, Beijing may already happen and make the fault detection meaningless.
Information Science and Technology University, Beijing 100192, China (e-mail: For model-based methods, analytical models are used to
[email protected]). generate residuals, which can be monitored and analyzed to
Yang Han is with the Department of Mathematics, The University of Manch-
ester, M13 9PL Manchester, U.K. (e-mail: [email protected]). detect faults. With simulation on a three-cell battery string
Peng Wang is with the Zhejiang Geely Automobile Research Institute Com- model, Chen et al. [16] proposed a bank of reduced-order Lu-
pany, Ltd., Ningbo 315800, China (e-mail: [email protected]). enberger observers and the learning observers for fault isolation
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TPEL.2022.3173038. and estimation. Liu and He [17] developed adaptive extended
Digital Object Identifier 10.1109/TPEL.2022.3173038 Kalman filter to estimate output voltage of equivalent circuit

0885-8993 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Universidade Tecnologica Federal do Parana. Downloaded on February 04,2025 at 12:38:11 UTC from IEEE Xplore. Restrictions apply.
 SUN et al.: ONLINE DATA-DRIVEN FAULT DIAGNOSIS AND THERMAL RUNAWAY EARLY WARNING FOR EV BATTERIES 12637

models (ECMs) and detect sensor fault by residual evaluation. have not happened yet, but some battery cells already show
Dey et al. presented a thermal model to diagnose thermal faults abnormities. If they are detected, damage of the battery will
of LIBs by Luenberger observer [18], nonlinear observer [19], be avoided. Moreover, it will also buy enough time to remind
and partial differential equation [20]. Feng et al. [21], [22] raised the driver of the EV to go for maintenance. In the end, the cost
an electrochemical–thermal coupled model to internal short of the fault will be very much reduced.
circuit. In general, the model-based methods are more accurate However, identifying the abnormal voltages or temperatures
and reliable than threshold-based methods. Nevertheless, the is complicated as they marginally deviate from the normal value,
computational cost is much larger, and some sophistic and highly thus false alarm often happens, especially only single parameter
coupled electrochemical models can hardly be implemented for (voltage or temperature) is used for diagnosis. To improve the
online implementation. reliability of the fault diagnosis, the correlation of the voltage
In recent years, data-driven fault diagnosis methods are draw- and temperature is considered in the second layer by combing the
ing more and more attention, including statistical methods [23], discrete Fréchet distance (DFD) with the LOF. More details are
signal-based [10], machine learning [24], etc. Xia et al. [25] elaborated in Section II. The content of this article is structured
applied the correlation coefficient between cell voltages to iden- as follows: Section Ⅱ explains the data source and studied battery
tify the short circuit fault. On top of this, by incorporating systems. In Section III, diagnostic strategies based on voltage
also the measured battery cell data from the laboratory setup, and temperature data are described. Results and discussions are
Kang et al. [26] implemented an improved correlation coefficient presented in Section IV. Finally, Section V concludes this article.
method to diagnose several types of faults, such as short circuits,
sensor faults, and connection faults. To detect connection failure, II. DATA SOURCE AND ONLINE DIAGNOSIS PROCESS
Yao built general regression neural networks and grid search
support vector machine [27], [28]. Yao et al. [29] proposed A. National Big Data Alliance of New Energy Vehicles
local outlier factor (LOF) and Grubbs outlier filter to identify the As depicted in Fig. 2, the NDANEV is built to provide data
battery with abnormal internal resistance. The aforementioned resources of BEV for researchers and pursue a safe operation
methods however have not been verified by operational data from of EVs. Using real-operating data of battery systems from
vehicles in daily use. The fault diagnosis approaches by using NDANEV can solve the problem Of insufficient BMS calcu-
cloud data platform, which is in real-time gathering the opera- lation capabilities. According to the protocol named technical
tional data from vehicles in daily use, are thus studied. Based specifications of remote service and management system for
on statistics, Zhao et al. [30] proposed a 3σ multilevel screening EVs, vehicle status data and power battery data are collected
strategy to find the faulty cell. Based on this, a combined method from real running vehicles. To be more specific, power battery
of k-mean clustering, z-score, and 3σ screening approach is used data include voltage and current of battery pack, cell voltage,
to locate abnormal cells by Xue et al. [31]. Li and Wang [32] and temperature. Vehicle status data consist of speed, mileage,
utilized the interclass correlation coefficient method to detect and positional information.
voltage faults. Hong et al. [33] and Wang et al. [34] established
a fault diagnosis mechanism for voltage and temperature faults
using the modified Shannon entropy approach, respectively. In B. Object of Study
order to overcome the difficulty of selecting the calculation win- Two accident EVs (vehicle 1# and vehicle 3#) and one normal
dow, Hong et al. [35] proposed the modified multiscale sample EV (vehicle 2#) with a data sampling frequency of 0.1 Hz are
entropy to predict the thermal runaway of LIB. Li et al. [36] considered in this study. Some more information of the EVs is
combine long short-term memory neural network and ECM to shown in Table I and II. Vehicle 1# and vehicle 2# are of the
diagnose overvoltage and undervoltage. These methods focus same model, whereas vehicle 3# is a different model. Vehicle
more on the diagnosis of single-parameter faults, such as voltage 1# is used as the training detection approach. Vehicles 2# and
or temperature, rather than multiple parameters coupled with 3# are utilized to verify it. The battery pack of vehicles 1#
each other in actual battery operation. Thus, they also overlook and 2# consists 17 modules (95 cells in total) connected in
the information regarding fault reflected by other parameters. In series, and has 95 voltage sensors and 34 temperature sensors,
this article, a multiparameter-based battery early fault diagnosis as indicated in Fig. 3 and Table I. The LIB system of the vehicle
approach is proposed. The operational data of EV batteries are 3# is a bit different, and it has 84 cells connected in series
obtained from the National Big Data Alliance of New Energy and 28 temperature sensors. In this article, the algorithm code
Vehicles (NDANEV) in Beijing, including the voltages and was compiled based on Python 3.7.5 and implemented on a PC
temperatures of the battery cells. As shown in Fig. 1, the fault (processor AMD Ryzen 7 4800H with Radeon Graphics CPU
diagnosis approach has two layers. The first layer is for severe @2.9 GHz).
fault detection, in which the battery cell voltage and tempera- Taking vehicle 1# as an example, the voltage and temperature
ture are compared with predefined upper and lower limits (for of the accident vehicle 1# near the thermal runaway are shown
temperature only upper limit). Once the voltage or temperature in Fig. 4. Note that, since the sampling rate is 0.1 Hz, there is
is beyond the limit, alarm is tripped. Since severe faults already one frame per 10 s. However, as a rule of the NDANEV, the data
happen in this scenario, the battery management system will stop sampling only happens when the vehicle is running. Thus, the
the operation of the EV ASAP. As the first layer is nothing new, voltage and temperature shown in Fig. 4 are the same as the
it will not be further elaborated. The focus of this article and voltage and temperature in time series but with the data during
the major challenge are the second layer, where severe faults parking (except charging) of the vehicle cut out. Nonetheless, it
Authorized licensed use limited to: Universidade Tecnologica Federal do Parana. Downloaded on February 04,2025 at 12:38:11 UTC from IEEE Xplore. Restrictions apply.
 12638 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 37, NO. 10, OCTOBER 2022

Fig. 1. Schematic of the proposed method.

5241-the frame (2019-07-05 14:27:32) in Fig. 4(a). But the

maximum temperature of this LIB system was 50 °C (below
55 °C), as depicted in Fig. 4(b), which proves the ineffectiveness
of the fault detection approaches that rely on one parameter.
After the accident, the battery system structure has been severely
damaged, as seen in Fig. 4(c).

C. Online Diagnosis Process

The first layer algorithm is applied every time point. The
voltage and temperature of one battery primarily measured at
Fig. 2. Data collection schematic. one time point is compared with two thresholds, upper and lower.
If the voltage/temperature at that time point is beyond the upper
threshold, the battery is considered as overvoltage/overheat; if
would not affect the fault detection, since faults rarely happen the voltage is below the lower threshold, the battery is considered
during parking while not charging. As seen in the figure, the as undervoltage. It is usually used for detection of extreme
battery pack encountered a thermal runaway during charging overvoltage, undervoltage/overheat. If no alarm occurred during
process. The No.17 cell of module 4# is the source that triggered the first level detection process, the second layer is activated.
the thermal runaway. An overdischarge warning (under 2.8 V) First, abnormal behavior has not been detected by the first layer.
occurred in the No.17 cell whose voltage was 0.03 V at the When maximum voltage range or temperature range in charging/

Authorized licensed use limited to: Universidade Tecnologica Federal do Parana. Downloaded on February 04,2025 at 12:38:11 UTC from IEEE Xplore. Restrictions apply.
 SUN et al.: ONLINE DATA-DRIVEN FAULT DIAGNOSIS AND THERMAL RUNAWAY EARLY WARNING FOR EV BATTERIES 12639

TABLE I
DETAILED INFORMATION OF ALL VEHICLES

TABLE II handling, a minor fault of battery system could develop into ther-
BATTERY PACK PROPERTIES OF NO. 1 VEHICLE
mal runaway [37]. Voltage and temperature are the two factors
controlling the battery reactions. To prevent thermal runaway
and other severe damage of the battery pack, the abnormal cell or
module should be identified before the severe damage happens.
In order to achieve this, the LOF of the battery cell voltages can
be calculated at each time being. The cell that has the largest LOF
is then considered as the abnormal cell. However, this approach
does not consider the correlation of the voltages of one cell at
different time beings. In fact, in the early stage of abnormal
operation, the cell voltage/temperature will gradually deviate
from the healthy cells. This gradual change in the cell volt-
age/temperature can be very useful for detecting the abnormal
operation. Therefore, the DFD between the voltage/temperature
curve of a cell and the median voltage/temperature curve is cal-
culated first, where Savitzky–Golay filter is employed to remove
the noise from the voltage and temperature data beforehand.
Then LOF is calculated based on DFD and standard deviation
of each cell voltage and module temperature in time series. The
LOF value is then used to judge if the cell is in early stage of a
fault or abnormal operation. More details are as follows.

A. Discrete Fréchet Distance

Owing to the discrete acquisition data, the DFD with origin
Fig. 3. Structure of the battery pack. (a) Battery cell. (b) One battery module. from Fréchet distance [38] needed to be used [39]. The DFD was
(c) Connection structure of the modules. defined by Eiter and Mannila in 1994 and was further expanded
by Mosig and Clausen in 2005 [40], [41].
discharging cycles is more than 0.06 V or 3 °C at time te , As mentioned earlier, the cell voltage and module temperature
all voltage/temperature data from the start of cycle (charging/ are monitored and recorded with 0.1-Hz sampling frequency.
discharging) to time te will be used as input for the DFD and The time series voltage and temperature are chopped into pieces
LOF. If the LOF values of voltage/temperature are less than the for analysis. The starting and ending point of each piece is
given threshold, no alarm will be triggered. On the contrary, actually the starting of a charging event and the next charging
the faulty cell/module can be diagnosed and effective measures event, respectively. As such, an abnormal operation detection of
should be taken to prevent further damage from happening. the battery is executed after each charging/discharging cycle, and
no data point is missing in the analysis. Taking the piece between
t0 – te as an example to elaborate the detection approach, the cell
III. EARLY STAGE ABNORMAL OPERATION DETECTION
voltages are expressed as
OF BATTERY
⎡ ⎤
Thermal runaway of battery system is always triggered by U1t0 −te
various kinds of abuse, including mechanical abuse, electrical ⎢ .. ⎥
⎢ . ⎥
abuse, and thermal abuse. The mechanical abuse condition in- ⎢ t −t ⎥
U t0 −te =⎢
⎢ Ui
0 e ⎥
⎥ (1)
cludes crash, penetration, and bend. Overheating and fire expo- ⎢ . ⎥
sure belong to the thermal abuse. Overcharge, overdischarge, ⎣ .. ⎦
and short circuit are electrical abuse. In the case of electrical Unt0 −te
abuse or faults, multiple failure modes can be coupled to result in
complicated mechanism. Without suitable diagnostics and fault Uit0 −te = [uti0 , uti1 , · · · , utie ] (2)

Authorized licensed use limited to: Universidade Tecnologica Federal do Parana. Downloaded on February 04,2025 at 12:38:11 UTC from IEEE Xplore. Restrictions apply.
 12640 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 37, NO. 10, OCTOBER 2022

Fig. 4. Vehicle 1#: (a) voltages (cell 1 to 95); (b) temperature (sensor 1 to 34); (c) battery pack and module 4 after fire.

where n is the total number of cells, so 0 ≤ i ≤ n. The median

voltage vector between time t0 and time te is indicated as
t0 −te 
U = u t0 , u t1 , · · · , u te
n n n
1 1 1
= uti0 , uti1 , · · · , utie . (3)
n i=1
n i=1
n i=1

In order to eliminate the influence of cell voltage bias error

on the DFD, the voltages of each cell are deducted by a bias
voltage, so that the initial voltages of all the cells will equal to
the median value of the initial voltages of all cells. The bias
voltage deduction for the ith cell is defined as
ubud,i = uti0 − ut0 . (4)
The calibrated cell voltages then become
⎡ t0 −te ⎤ ⎡ t0 −te ⎤
Ucali,1 U1 − ubud,1
⎢ .. ⎥ ⎢ .. ⎥
⎢ . ⎥ ⎢ . ⎥
⎢ ⎥ ⎢ ⎥
t0 −te ⎢ t0 −te ⎥
Ucali = ⎢ Ucali,i ⎥ = ⎢ Ui⎢ t0 −te
− ubud,i ⎥ (5)
⎥
⎢ . ⎥ ⎢ .. ⎥
⎣ .. ⎦ ⎣ . ⎦
t0 −te t0 −te
Ucali,n U n − u bud,n

t0 −te
where Ucali,i = [utcali,i
0
, utcali,i
1
, · · · , utcali,i
e
].
The DFD between the calibrated voltages of the i- and the
t0 −te
median voltage vector U is then calculated and shown as
follows:
Fig. 5. Variation of values at different k. (a) LOF value of cell 17 and remaining
cells’ maximums. (b) Cost function. DF Di = inf max |uα
cali,i − u |.
β
(6)
α,β∈[t0 ,te ]

Authorized licensed use limited to: Universidade Tecnologica Federal do Parana. Downloaded on February 04,2025 at 12:38:11 UTC from IEEE Xplore. Restrictions apply.
 SUN et al.: ONLINE DATA-DRIVEN FAULT DIAGNOSIS AND THERMAL RUNAWAY EARLY WARNING FOR EV BATTERIES 12641

Fig. 6. LOF value for vehicle 1# at different cycle. (a) Charging voltage data. (b) Discharging voltage data. (c) Charging temperature data. (d) Discharging
temperature data.

B. Local Outlier Factor where δ u and δ T are the LOF threshold of voltage and temper-
In the end, each cell should be represented by two parameters ature, respectively. Apparently, these two thresholds affect the
as input of LOF calculation. Beside DFD, the standard deviation sensitivity of the detection. Moreover, the LOF value relates to
of each cell’s voltage between t0 – te is used as the other k. The selection of k, δ u , and δ T is then elaborated as follows.
parameter, and it is calculated as
C. Selection of k and Threshold δ
 te
1 The goal of setting k, δ u , and δ T is that abnormal cells and
σi = 
2
(uti − μ) (7) modules can be screened out as outliers and require a low
n t=t
0
calculation time. The evaluation function is expressed as
e
where μ = n1 tt=t uti . So, ith cell now is expressed as ΔLOF(k)
0 E(k) = (10)
(DFDi , σi ). LOF of ith cell’s voltage is depicted as LOFk (ui ). tcos t (k)
More details of the LOF calculation can be found in [27].
Similarly, LOF of ith module’s temperature can be calculated, where tcos t (k) is calculation time, and ΔLOF(k) is the maxi-
and it is illustrated as LOFk (Ti ). Note that since the temperature mum LOF variation at k. Thus, larger value of E(k) indicates that
sensor resolution is 1 °C, the LOF values of temperatures are the abnormal cells are more differentiated from the normal cells,
reset to 0 if the temperature variation is less than 3 °C. A cell or and thereby the detection becomes easier. The idea here is to use
module will be labeled as abnormal, if LOFk (ui ) or LOFk (Ti ) the data from an accident vehicle to extract a proper value of k,
is beyond the threshold. The flag Fi u or Fi T is then set to 1, as δ u , and δ T . Fig. 5(a) shows the voltage LOF value of No.17 cell
follows: and other cells at different k of the accident vehicle 1#, where the
time piece includes the time being of the fault shown in Fig. 4.

1, if LO Fk (ui ) ≥ δ u Fig. 5(b) shows tcos t,(k) of No.17 cell as a function of k. As
Fui = (8) seen, E(k) has the maximum value at k = 8. Therefore, k in this
0, if LO Fk (ui ) < δ u
 article is taken as 8.
1, if LO Fk (Ti ) ≥ δ T
FT =
i
(9) The charging data are relatively stable, whereas the data
0, if LO Fk (Ti ) < δ T during the discharging process are more complex. In this article,

Authorized licensed use limited to: Universidade Tecnologica Federal do Parana. Downloaded on February 04,2025 at 12:38:11 UTC from IEEE Xplore. Restrictions apply.
 12642 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 37, NO. 10, OCTOBER 2022

Fig. 7. Last charging process of vehicle 1#. (a) Voltage curve (cell 1 to 95). (b) LOF value of voltage. (c) Temperature curve (sensor 1 to 34). (d) LOF value of
temperature.

the charging and discharging conditions are studied separately. is 8 frames earlier than thermal runaway (298th frame, 14:27:22,
Charging and discharging processes with less than 50 frames are 2019-07-05). While by using LOF voltage-based detection (sec-
excluded from the preprocessing. As shown in Fig. 4, thermal ond layer), the faulty cell No. 17 can be detected at frame 277,
runaway happens to the cell No.17. To prevent thermal runaway, which is 21 frames earlier than the first layer detection. It might
the fault needs to be detected in its early stage. The data of vehicle be noted that in Fig. 7(c), the LOF is null before frame 256. That
1# 2 min before the thermal runaway (14:27:22, 2019-07-05) is because, LOF calculation of voltage is only activated when
is then used to train the detection algorithm, where both the the voltage difference is larger than 0.06 V, which is only the
voltage and temperature are within the safe range, and the layer case science frame 256.
1 is not tripped yet. The LOF of the voltage and temperature Fig. 7(b) and (d) shows the comparison between the first and
are then calculated and they are shown in Fig. 6. The voltage second layer detection in terms of temperature. Specifically, the
LOF of No.17 cell is greater than 20 during the 148th charging first layer detection is not tripped because the temperature is
cycle. Before it, the voltage LOF during charging is not higher always below the threshold [typically 55 °C for EV batteries,
than 5. The voltage LOF does not exceed 15 during discharging red line in Fig 7(b)]. However, with the second layer detection
cycle. Fig. 6(c) shows that the temperature LOF of No.7 sensor relying on LOF of temperature, the faulty cell is detected at
(which is in the same module as the No. 17 cell) within module frame 241 (event  4 ), which is 66 frames (660 s) earlier than
4# is much larger than 30, whereas others are less than 14 thermal runaway. Note that, each temperature sensor is shared
during charging. Fig. 6(d) shows that the temperature LOF of all two or three cells, and the temperature sensor No. 7 is the one
sensors during discharging is less than 28. Therefore, the voltage most close to the cell No. 17.
diagnostic factor δ V is set to 15, and the temperature diagnostic
factor δ T is set to 30. IV. RESULT AND DISCUSSION
A. Validity Analysis
D. Verifying the Approach
Vehicle 2# (same model as vehicle 1#, but with no fault hap-
Taking the voltage and temperature data of 148th charging pened) is selected to verify if the proposed detection approach
process for vehicle 1# as an example, all data are shown in can have faulty trip. Fig. 8(a) and (b) showed the voltage and
Fig. 7 and the vehicle caught fire at the 307th frame [14:28:52, temperature of the cells during a whole charging and discharging
2019-07-05, event 2 in Fig. 7(a)]. As shown in Fig. 7(a), the cell cycle. In the two processes, the maximum voltage and minimum
No. 17 was detected undervoltage (first layer) at event 
1 , which voltage in the battery system are, respectively, 4.24 and 3.63 V,
Authorized licensed use limited to: Universidade Tecnologica Federal do Parana. Downloaded on February 04,2025 at 12:38:11 UTC from IEEE Xplore. Restrictions apply.
 SUN et al.: ONLINE DATA-DRIVEN FAULT DIAGNOSIS AND THERMAL RUNAWAY EARLY WARNING FOR EV BATTERIES 12643

Fig. 8. Parametric curves of vehicle 2#. (a) Voltages values (cell 1 to 95).
(b) Temperature values (sensor 1 to 34).
Fig. 9. LOF value of (a) voltage data and (b) temperature data of vehicle 2#
during charging and discharging processes.
and the maximum temperature is 38 °C. These values are in the
normal range, so no alarm can be tripped in the first layer.
The voltage and temperature LOF are shown in Fig. 9. As The voltage and temperature LOF values of vehicle 3# for
seen, the voltage LOF is always less than 3, no matter in charging fault detection were shown in Fig. 10(b) and (d). At the 829th
(C1) or discharging (D1), and it is far lower than the threshold frame, (08:25:08, March 14, 2019) in  3 of Fig. 10(b), the No. 47
15. The temperature LOF is always less than 1. Therefore, it can cell is detected by the second layer, which is 1000 s earlier than
be concluded that battery that in normal condition will not trip the over low voltage detection in the first layer. As shown in 5
the proposed detection approach in the second layer. of Fig. 10(d), the temperature of No.13 sensor can be detected
at 145th frame (04:57:45, March 14, 2019). This abnormal
B. Universality Analysis temperature alarm lasted for two frames. Then, at 867th frame
(08:25:46, March 14, 2019), abnormal temperature alarm is
Vehicle 3#, which is a different model from vehicle 1# but activated again in  6 of Fig. 10(d). Because No. 47 voltage
also encountered thermal runaway, is selected to verify if the sensor is in the same module 4# as No. 13 temperature sensor,
proposed detection approach is effective in a generic way. The it proved that the fault of module 4# can be accurately detected
battery system of vehicle 3# caught fire during discharging on by the proposed approach and it has considerable leading time
March 14, 2019. The ignition source was No. 47 cell of module than the thermal runaway.
4# according to the incident investigation. From 20:48:19, on
March 13, 2019, to 12:22:59, on March 14, 2019, voltage
data of the last discharging cycle of vehicle 3# were shown in C. Different Methods Comparison
Fig. 10(a). Before the 928th frame (09:59:03, March 14, 2019), In order to evaluate the performance of proposed method,
voltage values of all cells were within normal range. After the other three methods, including correlation coefficient [25], 3σ
928th frame, voltage data of all cells, including No. 47, were multilevel screening strategy [30], and Shannon entropy [34],
1.44 V, which can trigger an undervoltage alarm on the first were used to compare. The data of the first 930 frames of the
layer because the voltage is lower than the cutoff voltage in  1 last process of vehicle 3# were selected. The length of each
of Fig. 10(a). At the 1064th frame (12:22:59, March 14, 2019), window in [25] and [34] was 30 and 100 sample points.
the vehicle 3# has experienced thermal runaway and data cannot The correlation coefficient of the selected voltage data was
be obtained in  2 of Fig. 10(a). The original temperature curve shown in Fig. 11(a). The calculation time of each window of
during the 3rd discharging process was shown in Fig. 10(c). At this method is 0.03 s. However, false alarms are given frequently
the 855th frame (08:25:34, March 14, 2019), the temperature of when the correlation coefficient value was below 0.9. It was
No.13 and No.14 sensors from module 4# reached 60 °C and an noted that this method had the short calculation time and low
overheat alarm of the first layer was triggered in 
4 of Fig. 10(c). accuracy.

Authorized licensed use limited to: Universidade Tecnologica Federal do Parana. Downloaded on February 04,2025 at 12:38:11 UTC from IEEE Xplore. Restrictions apply.
 12644 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 37, NO. 10, OCTOBER 2022

Fig. 10. Last discharging process of vehicle 3#. (a) Voltage curve (cell 1 to 84). (b) LOF value of voltage. (c) Temperature curve (sensor 1 to 28). (d) LOF value
of temperature.

Fig. 11(b) showed the cell faulty frequency by 3σ multilevel

screening strategy. The calculation time of all data is 0.86 s.
Cell 73 was considered the most likely faulty cell in battery
pack because its frequency was 0.23. But the faulty source was
No. 47 cell, which indicated that this method cannot give an
accurate fault cell of vehicle 3#.
Fig. 11(c) showed the abnormal coefficient based on Shan-
non entropy. The calculation time of each window was 0.1 s.
The threshold of the abnormal coefficient was±4 in [34]. The
abnormal coefficient of all cells was [–44], so no abnormal cell
will be given. This indicated that this method did not accurately
identify abnormal cell.
The calculation time of the proposed method is 0.09 s, and
can meet the practical application. And faulty battery can be
diagnosed by the proposed approach

V. CONCLUSION
An online data-driven EV battery fault diagnosis scheme
using both the cell voltage and temperature is proposed. Both
DFD method and LOF are used to quantify the correlation
between the different battery cells’ voltage and temperature. In
this way, the faulty cell can be detected considerably earlier
than thermal runaway happens, which gives sufficient time to
start maintenance or protection scheme, to very much reduce
the cost of fault. The data from vehicles with thermal runaway
are used to train the detection approach. The trained detection
approach is then applied to vehicles with or without thermal
Fig. 11. (a) Correlation coefficient, (b) 3σ multilevel screening strategy, and
runaway to check its sensitivity and it works perfectly. All the
(c) abnormal coefficient of Shannon entropy of vehicle 3#. data are obtained from NDANEV, which is a national cloud

Authorized licensed use limited to: Universidade Tecnologica Federal do Parana. Downloaded on February 04,2025 at 12:38:11 UTC from IEEE Xplore. Restrictions apply.
 SUN et al.: ONLINE DATA-DRIVEN FAULT DIAGNOSIS AND THERMAL RUNAWAY EARLY WARNING FOR EV BATTERIES 12645

platform that gathering operational data from vehicles in daily [22] X. Feng, Y. Pan, X. He, L. Wang, and M. Ouyang, “Detecting the internal
use with 0.1-Hz sampling rate, rather than laboratory data that short circuit in large-format lithium-ion battery using model-based fault-
diagnosis algorithm,” J. Energy Storage, vol. 18, pp. 26–39, 2018.
obtained in an artificial scenario. Therefore, the proposed early [23] M. Cheliotis, I. Lazakis, and G. Theotokatos, “Machine learning and data-
fault detection approach is very practical and promising for real driven fault detection for ship systems operations,” Ocean Eng., vol. 216,
application. 2020, Art. no. 107968.
[24] R. Xiong, S. Ma, H. Li, F. Sun, and J. Li, “Toward a safer battery
management system: A critical review on diagnosis and prognosis of
REFERENCES battery short circuit,” iScience, vol. 23, no. 4, Apr. 2020, Art. no. 101010.
[25] Y. S. Xia, T. Nguyen, and C. Mi, “A correlation based fault detection
[1] D. Chen, J. Deng, B. Zhang, Z. Wang, and S. Wang, “A dual-transformer
method for short circuits in battery packs,” J. Power Sources, vol. 337,
based hybrid semidual active bridge converter for wide voltage range
pp. 1–10, 2017.
applications utilizing simple segmented control,” IEEE Trans. Power
[26] Y. Kang, B. Duan, Z. Zhou, Y. Shang, and C. Zhang, “Online multi-fault
Electron., vol. 37, no. 2, pp. 1435–1446, Feb. 2022.
detection and diagnosis for battery packs in electric vehicles,” Appl.
[2] Z. S. Gan, Z. Zhang, S. Xu, P. Liu, and Z. Qin, “Data-driven fault diagnosis
Energy, vol. 259, 2020, Art. no. 114170.
of lithium-ion battery overdischarge in electric vehicles,” IEEE Trans.
[27] K. X. Chen, J. Wei, and G. Dong, “Voltage fault detection for lithium-
Power Electron., vol. 37, no. 4, pp. 4575–4588, Apr. 2022.
ion battery pack using local outlier factor,” Measurement, vol. 146,
[3] Q. Wang, Z. Wang, L. Zhang, P. Liu, and Z. Zhang, “A novel consis-
pp. 544–556, 2019.
tency evaluation method for series-connected battery systems based on
[28] L. Yao, Z. Fang, Y. Xiao, J. Hou, and Z. Fu, “An intelligent fault diagnosis
real-world operation data,” IEEE Trans. Transp. Electrific., vol. 7, no. 2,
method for lithium battery systems based on grid search support vector
pp. 437–451, Jun. 2021.
machine,” Energy, vol. 214, 2021, Art. no. 118866.
[4] Z. Sun et al., “Detection of voltage fault in the battery system of elec-
[29] L. Yao, Y. Xiao, X. Gong, J. Hou, and X. Chen, “A novel intelligent method
tric vehicles using statistical analysis,” Appl. Energy, vol. 307, 2022,
for fault diagnosis of electric vehicle battery system based on wavelet
Art. no. 118172.
neural network,” J. Power Sources, vol. 453, 2020, Art. no. 227870.
[5] X. Hu, H. Jiang, F. Feng, and B. Liu, “An enhanced multi-state estimation
[30] Y. Zhao, P. Liu, Z. Wang, L. Zhang, and J. Hong, “Fault and defect diag-
hierarchy for advanced lithium-ion battery management,” Appl. Energy,
nosis of battery for electric vehicles based on big data analysis methods,”
vol. 257, 2020, Art. no. 114019.
Appl. Energy, vol. 207, pp. 354–362, 2017.
[6] R. Xiong, W. Sun, Q. Yu, and F. Sun, “Research progress, challenges and
[31] Q. Xue, G. Li, Y. Zhang, S. Shen, Z. Chen, and Y. Liu, “Fault diagnosis
prospects of fault diagnosis on battery system of electric vehicles,” Appl.
and abnormality detection of lithium-ion battery packs based on statistical
Energy, vol. 279, 2020, Art. no. 115855.
distribution,” J. Power Sources, vol. 482, 2021, Art. no. 228964.
[7] Z. Liao, S. Zhang, K. Li, G. Zhang, and T. G. Habetler, “A survey of
[32] X. Li and Z. Wang, “A novel fault diagnosis method for lithium-ion battery
methods for monitoring and detecting thermal runaway of lithium-ion
packs of electric vehicles,” Measurement, vol. 116, pp. 402–411, 2018.
batteries,” J. Power Sources, vol. 436, 2019, Art. no. 226879.
[33] J. Hong, Z. Wang, and P. Liu, “Big-data-based thermal runaway prognosis
[8] D. P. Finegan et al., “The application of data-driven methods and
of battery systems for electric vehicles,” Energies, vol. 10, no. 7, 2017,
physics-based learning for improving battery safety,” Joule, vol. 5, no. 2,
Art. no. 919.
pp. 316–329, 2021.
[34] Z. Wang, J. Hong, P. Liu, and L. Zhang, “Voltage fault diagnosis and
[9] Z. Chen, R. Xiong, J. Tian, X. Shang, and J. Lu, “Model-based fault
prognosis of battery systems based on entropy and Z-score for electric
diagnosis approach on external short circuit of lithium-ion battery used
vehicles,” Appl. Energy, vol. 196, pp. 289–302, 2017.
in electric vehicles,” Appl. Energy, vol. 184, pp. 365–374, 2016.
[35] J. Hong et al., “Thermal runaway prognosis of battery systems using the
[10] X. Hu, K. Zhang, K. Liu, X. Lin, S. Dey, and S. Onori, “Advanced fault
modified multiscale entropy in real-world electric vehicles,” IEEE Trans.
diagnosis for lithium-ion battery systems: A review of fault mechanisms,
Transp. Electrific., vol. 7, no. 4, pp. 2269–2278, Dec. 2021.
fault features, and diagnosis procedures,” IEEE Ind. Electron. Mag.,
[36] D. Li, Z. Zhang, P. Liu, Z. Wang, and L. Zhang, “Battery fault diagnosis
vol. 14, no. 3, pp. 65–91, Sep. 2020.
for electric vehicles based on voltage abnormality by combining the long
[11] M.-K. Tran and M. Fowler, “A review of lithium-ion battery fault diag-
short-term memory neural network and the equivalent circuit model,” IEEE
nostic algorithms: Current progress and future challenges,” Algorithms,
Trans. Power Electron., vol. 36, no. 2, pp. 1303–1315, Feb. 2021.
vol. 13, no. 3, 2020, Art. no. 62.
[37] L. Jiang et al., “Data-driven fault diagnosis and thermal runaway warning
[12] L. Liu et al., “Comparative study on substitute triggering approaches for
for battery packs using real-world vehicle data,” Energy, vol. 234, 2021,
internal short circuit in lithium-ion batteries,” Appl. Energy, vol. 259, 2019,
Art. no. 121266.
Art. no. 114143.
[38] M. Frechet, “Sur quelques points du calcul fonctionnel,” Rendiconti del
[13] B. Liu et al., “Safety issues caused by internal short circuits in lithium-ion
Circolo Matematico di Palermo, vol. 22, no. 1, pp. 1–72, 1906.
batteries,” J. Mater. Chem. A, vol. 6, no. 43, pp. 21475–21484, 2018.
[39] T. Eiter and H. Mannila, “Computing discrete Fréchet distance,” Christian
[14] A. Pesaran, S. Santhanagopalan, and G. H. Kim, “Addressing the impact of
Doppler Lab. Expert Syst., TU Vienna, Vienna, Austria, Tech. Rep. CD-TR
temperature extremes on large format Li-ion batteries for vehicle applica-
94/64, pp. 636–637, 1994.
tions (Presentation),” Office of Entific & Technical Information Technical
[40] A. Mosig and M. Clausen, “Approximately matching polygonal curves
Reports, Golden, CO (United States), No. NREL/PR-5400-58145, 2013.
with respect to the Fréchet distance,” Comput. Geometry, vol. 30, no. 2,
[15] B. Xia, Z. Chen, and M. Chris, “External short circuit fault diagnosis
pp. 113–127, 2005.
for lithium-ion batteries,” in Proc. IEEE Transp. Electrific. Conf. Expo,
[41] T. Wylie and B. Zhu, “Following a curve with the discrete Fréchet dis-
Dearborn, MI, USA, 2014, pp. 1–7.
tance,” Theor. Comput. Sci., vol. 556, pp. 34–44, 2014.
[16] W. Chen, W.-T. Chen, M. Saif, and M.-F. Li, “Simultaneous fault isolation
and estimation of lithium-ion batteries via synthesized design of luenberger
and learning observers,” IEEE Trans. Control Syst. Technol., vol. 22, no. 1,
pp. 290–298, Jan. 2014.
[17] Z. Liu and H. He, “Sensor fault detection and isolation for a lithium-ion
battery pack in electric vehicles using adaptive extended Kalman filter,”
Appl. Energy, vol. 185, pp. 2033–2044, 2017. Zhenyu Sun received the B.Eng. degree in vehicle en-
[18] H. E. P. Dey and S. J. Moura, “Model-based battery thermal fault diagnos- gineering from Beijing Forestry University, Beijing,
tics: Algorithms, analysis, and experiments,” IEEE Trans. Control Syst. China, in 2015, and the M.Eng. degree in mechanical
Technol., vol. 27, no. 2, pp. 576–587, Mar. 2019. engineering in 2018 from the Beijing Institute of
[19] S. Dey et al., “Model-based real-time thermal fault diagnosis of lithium-ion Technology, Beijing, China, where he is currently
batteries,” Control Eng. Pract., vol. 56, pp. 37–48, 2016. working toward the Ph.D. degree in mechanical en-
[20] S. Dey, H. E. Perez, and S. J. Moura, “Model-based battery thermal fault gineering.
diagnostics: Algorithms, analysis, and experiments,” IEEE Trans. Control He is currently a Visiting Student with the Delft
Syst. Technol., vol. 27, no. 2, pp. 576–587, Mar. 2019. University of Technology, Delft, The Netherlands,
[21] X. Feng, X. He, L. Lu, and M. Ouyang, “Analysis on the fault features for since 2021. His research interests include electric
internal short circuit detection using an electrochemical-thermal coupled vehicle big data analysis, fault diagnosis, and safety
model,” J. Electrochem. Soc., vol. 165, no. 2, pp. A155–A167, 2018. management of lithium-ion battery system.

Authorized licensed use limited to: Universidade Tecnologica Federal do Parana. Downloaded on February 04,2025 at 12:38:11 UTC from IEEE Xplore. Restrictions apply.
 12646 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 37, NO. 10, OCTOBER 2022

Zhenpo Wang (Senior Member, IEEE) received the Yang Han received the Ph.D. degree in statistics from
B.Eng. degree in automotive engineering from Tongji the University of Southampton, Southampton, U.K.,
University, Shanghai, China, in 2000, and the Ph.D. 2014.
degree in automotive engineering from the Beijing She is currently a Senior Lecturer (Associate Pro-
Institute of Technology, Beijing, China, in 2005. fessor) with the Department of Mathematics, The
He is currently a Professor with the Beijing Insti- University of Manchester, Manchester, U.K. Her
tute of Technology and the Director of the National main research interests include simultaneous infer-
Engineering Laboratory for Electric Vehicles. He has ence and multiple comparison procedures.
authored or coauthored four monographs and trans- Dr. Han was the recipient of a Distinguished
lated books and more than 80 technical papers. He Achievement Award for Teacher of the Year 2019 at
also holds more than 60 patents. His current research The University of Manchester.
interests include pure electric vehicle integration, packaging and energy man-
agement of battery systems, and charging station design.
Prof. Wang was the recipient of numerous awards, including the second
National Prize for Progress in Science and Technology and the first Prize for
Progress in Science and Technology from the Ministry of Education, China and
the second Prize for Progress in Science and Technology from Beijing Municipal,
China.
Peng Wang received the B.S. degree in measuring
and control technology and instrumentations from
Zhengzhou University, Zhengzhou, China, in 2009,
Peng Liu received the B.S. and M.S. degrees from
and the M.S. degree in engineering management from
Chang’an University, Xi’an, China, in 2005 and 2008,
the East China University of Science and Technology,
respectively, and the Ph.D. degree from the Beijing
Shanghai, China, in 2018.
Institute of Technology, Beijing, China, in 2011. At present, he is responsible for the development of
He is currently an Associate Professor with the
the battery, motor, and electric control system of elec-
School of Mechanical Engineering, Beijing Institute
tric vehicles in Zeekr Automobile (Ningbo Hangzhou
of Technology. His research interests include big data
Bay New Zone) Company, Ltd., and holds more than
analysis and safety management of electric vehicles. five patents.

Zian Qin (Senior Member, IEEE) received the B.Eng.

degree in automation from Beihang University, Bei-
jing, China, in 2009, the M.Eng. degree in control
science and engineering from the Beijing Institute of Pavol Bauer (Senior Member, IEEE) received the
Technology, Beijing, China, in 2012, and the Ph.D. master’s degree in electrical engineering from the
degree from Aalborg University, Aalborg, Denmark, Technical University of Kosice, Košice, Slovakia, in
in 2015. 1985, and the Ph.D. degree from the Delft University
He is currently an Assistant Professor with the of Technology, Delft, The Netherlands, in 1995.
Delft University of Technology, Delft, The Nether- He is currently a Full Professor with the Depart-
lands. In 2014, he was a Visiting Scientist with ment of Electrical Sustainable Energy, Delft Univer-
Aachen University, Aachen, Germany. From 2015 to sity of Technology and the Head of DC Systems, En-
2017, he was a Postdoctoral Research Fellow with Aalborg University. His ergy Conversion and Storage group. He has authored
research interests include wide bandgap devices, power electronics based grid, or coauthored more than 72 journal and almost 300
and Power2X. conference papers in many field (with H factor Google
Dr. Qin served as the Technical Program Chair of IEEE-ISIE-2020, the Tech- scholar 43, Web of science 20), he is an author or coauthor of eight books, holds
nical Program CO-CHAIR of IEEE-COMPEL 2020, and the industrial session four international patents, and organized several tutorials at the international
CO-CHAIR of ECCE-Asia 2020. conferences. He has worked on many projects for industry concerning wind
and wave energy, power electronic applications for power systems, such as
Smarttrafo; HVdc systems, projects for smart cities, such as PV charging of
electric vehicles, PV and storage integration, contactless charging; and he par-
ticipated in several Leonardo da Vinci and H2020 EU projects as Project Partner
Yong Chen received the Ph.D. degree in vehicle
(ELINA, INETELE, E-Pragmatic) and Coordinator (PEMCWebLab.com-Edipe,
engineering from the Beijing Institute of Technology,
Beijing, China, in 2002. SustEner, Eranet DCMICRO).
Prof. Bauer was the recipient of the title “Prof.” from the President of Czech
He is currently a Professor with the School of Me-
Republic at the Brno University of Technology, Brno, Czechia, in 2008 and Delft
chanical and Electrical Engineering, Beijing Infor-
University of Technology in 2016.
mation Science and Technology University, Beijing,
China. He is also a Researcher with the Collabora-
tive Innovation Center of Electric Vehicles, Beijing,
China. His research interests include modeling, simu-
lation and control of new energy source vehicles, and
vehicle system dynamics.

Authorized licensed use limited to: Universidade Tecnologica Federal do Parana. Downloaded on February 04,2025 at 12:38:11 UTC from IEEE Xplore. Restrictions apply.