990 IEEE TRANSACTIONS ON TRANSPORTATION ELECTRIFICATION, VOL. 8, NO.

1, MARCH 2022

A Data-Driven Method for Battery Charging

Capacity Abnormality Diagnosis in
Electric Vehicle Applications
Zhenpo Wang , Member, IEEE, Chunbao Song , Lei Zhang , Member, IEEE, Yang Zhao ,
Peng Liu , and David G. Dorrell, Fellow, IEEE

Abstract— Enabling charging capacity abnormality diagnosis system complexity brings about high propensity of fault occur-
is essential for ensuring battery operation safety in electric rence that needs to be timely diagnosed for efficient safety
vehicle (EV) applications. In this article, a data-driven method management [7]–[9].
is proposed for battery charging capacity diagnosis based on
massive real-world EV operating data. Using the charging rate, Accurate metering of battery capacity bears significance for
temperature, state of charge, and accumulated driving mileage as ensuring system safety and is fundamental for other battery
the inputs, a tree-based prediction model is developed with a poly- control algorithms development [10]–[12]. However, this can
nomial feature combination used for model training. A statistics- be hardly done through direct measurement using commer-
based method is then used to diagnose battery charging capacity cially available onboard sensors. Instead, model-based and
abnormity by analyzing the error distribution of large sets of
data. The proposed tree-based prediction model is compared with data-driven methods have been intensively researched to tackle
other state-of-the-art methods and is shown to have the highest the problem. However, it is a nontrivial task as lithium-ion bat-
prediction accuracy. The holistic diagnosis scheme is verified teries undergo a sophisticated and varied degradation process
using unseen data. during actual EV operations. A discrete Arrhenius fading
Index Terms— Abnormity diagnosis, big data, charging capac- model and a filter-based data-driven model were respectively
ity, electric vehicles (EVs), machine learning. proposed for battery capacity estimation in [13] and [14]. As a
key indicator of EV charging process, the charging capacity is
competent to represent the state-of-health (SOH) of a battery
I. I NTRODUCTION
system [15], thus soliciting a rich body of literature exploring

A UTOMOTIVE electrification plays a vital role in trans-

forming the petroleum-powered fleet into a green ground
transport system [1]. Continuous technological advancements
battery SOH and remaining lifetime estimation based on
it [16]–[18]. Abnormal charging capacity may result in sudden
changes of state of charge (SOC), which, in turn, could give
and sustained governmental support have remarkably con- rise to severe safety issues. According to statistics, there
tributed to the rapid development and mass adoption of electric were a total of 124 EV fire accidents reported in 2020 in
vehicles (EVs) worldwide [2]. This growing success of EVs China, 23% of which were confirmed to originate from the
can be largely ascribed to the elevating performance and charging process [19]. As such, charging capacity diagnosis is
falling costs of lithium-ion batteries [3]. In EV applications, necessitated for ensuring the safe operation of EVs [20].
hundreds of battery cells are connected in parallel and/or Diagnostic approaches can be used to identify battery abnor-
series configurations in a battery system to meet the energy mity on the cell or pack level [21], [22]. Straightforward rule-
and power requirements [4]–[6]. However, the resultant high based diagnosis methods based on direct voltage or current
measurement cannot detect latent faults that could potentially
Manuscript received May 17, 2021; revised August 15, 2021; accepted
September 19, 2021. Date of publication October 4, 2021; date of current disrupt normal battery operations [23]. Various approaches
version March 22, 2022. This work was supported in part by the Ministry have been proposed in the literature, which can be roughly
of Science and Technology of the People’s Republic of China under Grant classified into three categories. They are knowledge-based,
2019YFE0107900 and in part by the National Natural Science Foundation of
China under Grant 51805029 and Grant U1764258. (Corresponding authors: model-based, and data-driven methods. For the knowledge-
Lei Zhang; Yang Zhao.) based methods, fault diagnosis rules are often designed using
Zhenpo Wang and Peng Liu are with the Collaborative Innovation Center prior expert knowledge and then applied to several battery
for Electric Vehicles in Beijing and the National Engineering Laboratory for
Electric Vehicles, Beijing Institute of Technology, Beijing 100081, China, and parameters that can be easily measured [24]. With this method,
also with the Beijing Institute of Technology Chongqing Innovation Center, the fault diagnosis for EVs of the same models can be realized
Chongqing 401120, China. using the same diagnosis rules [25]. For the model-based meth-
Chunbao Song, Lei Zhang, and Yang Zhao are with the Collaborative Inno-
vation Center for Electric Vehicles in Beijing and the National Engineering ods, physical models representing electrochemical or thermal
Laboratory for Electric Vehicles, Beijing Institute of Technology, Beijing dynamics are used for battery fault diagnosis [22]. For exam-
100081, China (e-mail: [email protected]; [email protected]). ple, in [26] and [27], multimodel adaptive voltage estimation
David G. Dorrell is with the School of Electrical and Information Engineer-
ing, University of Witwatersrand, Johannesburg 2050, South Africa. methods were put forward to detect overcharge or overdis-
Digital Object Identifier 10.1109/TTE.2021.3117841 charge faults of batteries. The implementation of both the
2332-7782 © 2021 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: National Cheng Kung Univ.. Downloaded on June 15,2024 at 01:00:21 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: DATA-DRIVEN METHOD FOR BATTERY CHARGING CAPACITY ABNORMALITY DIAGNOSIS 991

knowledge- and model-based methods relies on deterministic Mathematical Big data platform Data visualization
fault types and known mechanisms. This may lead to limited models
feasibility for different EV models and application scenarios. Machine Data storage
learning and processing
To improve the adaptiveness to various scenarios and the Message ¾Storm
Statistical queuing ¾Redis
robustness to discrepant fault types, data-driven approaches tools system ¾HDFS
¾Hbase
based on real-world EV operating data are being intensively ¾Spark
Fault ¾MapReduce
studied as a promising method. For instance, Shang et al. [28] diagnosis
applied the sample entropy to examine the consistency and User-defined
models Vehicle data collection
fluctuation of cell voltages within a battery system for battery
Remote GB/T 32960
voltage fault diagnosis. In [29] and [30], the clustering analysis transmission Protocol
and outlier identification were combined for cell voltage abnor-
mality detection and fault prognosis. Li et al. [31] presented a Fig. 1. Schematic of the EV data processing framework.
method by combining the long short-term neural network and
the equivalent circuit model to improve voltage prediction and
risk assessment performance. However, the existing studies system, and other user-defined items. The static data include
on battery fault diagnosis mainly focus on battery voltage or the information of vehicle license plate, operating regions,
temperature, while the battery charging capacity as a direct vehicle model, and so on. The schematic of the EV data
health indicator has been rarely studied for battery fault processing framework is illustrated in Fig. 1. Vehicle data are
diagnosis. Besides, the presented models were only verified in collected from EVs equipped with telematics boxes and are
laboratory environments, yet their efficacy may be significantly transmitted to the big data platform in accordance with the
compromised in practice due to complex and ever-changing GB/T 32960 protocol. The vehicle data are then reprocessed
driving conditions. using specialized big data techniques. Mathematical models,
In this article, a data-driven diagnostic framework is pro- including the statistical and fault diagnosis models, are used
posed to determine the abnormal charging capacity in the for in-depth analysis, and the analytic and diagnostic results
form of probability based on massive EV operating data. can be displayed through data visualization tools.
A data segmenting approach is first used to extract the charging The data under different vehicle operation modes, including
segments from a colossal EV data reservoir. A machine driving, idling, and charging, need to be partitioned into
learning-based model for battery charging capacity prediction different segments. The data segmenting process and real-
is then developed. Four parameters, including battery SOC world battery capacity data are illustrated in Section IV.
change, ambient temperature, accumulative driving mileage, Some vehicle specifications are collected from the avail-
and charging rate, are considered as the inputs. By using the able EV manufacturer databases. The climate data used
established prediction model and the statistical distribution in this work are obtained from the Weather Underground
derived from large-scale EV operating data, a data-driven (https://www.wunderground.com) and the Reliable Prognosis
diagnostic approach is devised to determine the abnormal (https://rp5.ru).
charging capacity in the form of probability.
The remainder of this article is organized as follows. III. F RAMEWORK FOR EV BATTERY C HARGING
Section II gives a brief introduction about data collection. C APACITY FAULT D IAGNOSIS
Section III introduces the holistic diagnostic framework for The diagnostic framework used for detecting battery charg-
battery charging capacity diagnosis. Section IV presents the ing capacity abnormity is expounded in this section. When
data segmenting and preprocessing methods. Section V elab- EVs operate in normal states, their charging capacities main-
orates on the proposed fault diagnosis model. Section VI tain a certain statistical regulation. Under this premise, abnor-
offers the verifications results, followed by the key conclusions mal changes in battery charging capacities can be discerned by
summarized in Section VII. comparing actual ones with the statistical regulation. The basic
logic for diagnosing charging capacity faults is given as fol-
II. EV O PERATING DATA C OLLECTION lows. First, a prediction model is established to predict battery
The data used in this study are collected from the National charging capacity considering varied influencing factors. Then,
Monitoring and Management Center for New Energy Vehicles the statistical regulations of large-scale EV applications are
and the Open Lab of the National Big Data Alliance of calculated to determine the permissible interval of predicted
New Energy Vehicles (NDANEV). The center serves as the charging capacity. Finally, a fault detection mechanism is
national EV big data platform in China for ensuring vehicular developed to detect the charging capacity fault occurrence.
operation safety. It can collect real-time operating data of The data-driven diagnosis process is presented in Fig. 2.
EVs used for public service all over China. Using EV data The EV data obtained from the big data center are pre-
and cloud computing capability, vehicle operation manage- processed; the target features are selected, cleansed, and
ment, battery recycling, and financial subsidy verification are arithmetically processed to meet the requirements of data
being extensively investigated. The EV operating data include processing in the next step. The data segmenting methods
both dynamic and static data. The dynamic data primarily are used to partition the consecutive EV operating data into
consist of operating data, such as velocity and accumulative segments labeled with driving, charging, and idling states.
driving mileage of vehicle, voltage and current of battery The charging segments are extracted for charging capacity

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on June 15,2024 at 01:00:21 UTC from IEEE Xplore. Restrictions apply.
992 IEEE TRANSACTIONS ON TRANSPORTATION ELECTRIFICATION, VOL. 8, NO. 1, MARCH 2022

Fig. 2. Framework of the proposed data-driven battery charging capacity diagnosis method.

analysis. Note that some new features are generated, and the
charging data are cleansed again for obtaining better prediction
accuracy. The data segmenting approach will be described
in detail in Section IV. Multiple features are considered as
the inputs for the machine learning-based regression model
development. Polynomial features generation is used to better
account for the nonlinear and coupled effects of these features.
Charging power rate classification is also performed. The
detailed prediction process is illustrated in Section V. Finally,
the charging capacity fault diagnosis is executed based on the
statistical information derived from large-scale EV charging
data. The detailed process will be presented in Section VI.

IV. EV DATA S EGMENTING AND C HARGING

DATA P ROCESSING
Fig. 3. Flowchart of charging data segmenting.
The charging capacity data are obtained through data seg-
menting on the initial EV operating data. The charging data
processing module includes two parts: charging detection and
The data segmenting process is illustrated in Fig. 3. In the
charging statistics. The EV operating data contain three vehicle
original data matrix, the vector [ti , x 1 , . . . , x n ] includes one
mode states, including driving, charging, and idling. As the
timestamp ti and n features [x 1 , . . . , x n ]. Since the original
charging capacity is needed in this study, accurately extracting
data matrix is usually large, a fast-recording matrix is used
the charging sessions is necessitated in data processing. Con-
to improve computing efficiency, which mainly records the
ventional segmenting approaches utilize vehicle mode states
starting times (ti ), ending times (t j ), starting SOCs (x k ), and
to identify the starting and ending timestamps of charging
ending SOCs (x p ) during charging sessions. Then, a statistical
sessions for charging session extraction. However, this vehicle
matrix is established by combing the fast-recording matrix
mode state can be easily corrupted in real-world conditions so
and the statistical results. The data segmenting and calculation
that the starting and ending points of charging sessions can be
procedures are given as follows:
erroneously identified, thus failing the charging sessions seg-
1) The original data matrix is given by
menting. In this study, an intelligent data segmenting approach
is presented to improve the accuracy and robustness of data  T
Do = t T x1T x2T . . . xnT (1)
segmenting without solely relying on the vehicle mode states.
It first marks the starting and ending timestamps of charg- where Do is the data input matrix, t is the time series, and x
ing sessions using vehicle velocity and battery states. Once is the feature vector.
preliminarily detected as the starting and ending timestamps, 2) Analysis on EV States: When implementing mathematical
the adjacent timestamps are also examined to further verify calculations, the EV state should be determined in advance.
the data segmenting accuracy. Normally, the EV operation mode can be obtained directly

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on June 15,2024 at 01:00:21 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: DATA-DRIVEN METHOD FOR BATTERY CHARGING CAPACITY ABNORMALITY DIAGNOSIS 993

from EV operating data. However, the charging mode state

data can be compromised in real-world conditions. Thus,
multiple types of data are used to address this issue. If the
missing or fault ratio of the charging state data is less than
5%, the charging-state data are used to determine the vehicle
modes, and the procedure proceeds to Step (4). If the ratio
exceeds 5%, the procedure moves to Step (3) to determine the
vehicle operation modes in advance.
3) If the ratio of the missing charging state data exceeds
5%, the vehicle speed and the battery current are used for the
preliminary determination of the charging mode state. A new
array is used to record the charging mode state. The basic
logic for determining the charging mode state includes two
items. The first is that, when the EV is in the charging mode,
the current flows from the charger to the battery; the second
is that the EV should be in the idle state while in the charging
mode. By using these two constraints, the charging state can
be determined. In addition, energy recovery from generative
braking is considered. Since the brake maneuvers are mainly
executed when the vehicle is running and involve with a SOC
variation of 2% in most cases, the condition of a continuous Fig. 4. Real-world changes of (a) SOC and (b) charging capacity.
SOC increase of five frames is added in the determination
process.
4) Since the operating state is determined based on previous where Ui,t and Ii,t are the total voltage and current of the
steps, the charging sessions can be extracted. First, the first one battery system in the i th charging session and at the time t;
of three consecutive charging mode frames is determined to Ci and E i are the charging capacity (Ah) and energy (kWh)
be the starting frame of a charging session. When the state in the i th charging session; k is the half window required for
is no longer in the charging mode, or when the time interval the moving average algorithm; and t is the time interval of
between two frames exceeds 30 s, the record is marked as the data recording.
ending frame of this charging session. In between, the charging By adopting the above methods, massive charging session
session is extracted. Meanwhile, the starting time, ending time, fragments can be extracted from the EV operating data pool.
starting SOC, and ending SOC of the charging session are Fig. 4(a) displays the SOC variations during different charging
recorded in the fast-recording matrix D f as sessions. The irregular SOC changes suggest the randomness
 T in EV charging profiles during realistic operations. In addi-
D f = tsT teT xkT xTp (2) tion, the charging capacity of each charging segment can be
where ts and te represent the starting and ending times, obtained using an integral calculation of the charging current in
respectively; xk and x p are the starting and ending SOCs. Do the time domain. The charging capacity of charging sessions is
and D f can be used to generate a statistical data matrix with shown in Fig. 4(b). It shows that the charging capacity exhibits
more features through simple arithmetical calculations. This irregularity, and the SOC data are well recorded and included
gives in the battery charging capacity prediction and fault diagnosis.
  The real-world EV charging data processed here provide a
Dsta = tsT teT xkT xTp · · · (3) solid database for further mathematical manipulations.
where Dsta includes the starting time, ending time, start-
ing SOC, ending SOC, SOC variation, starting accumulative V. BATTERY C HARGING C APACITY P REDICTION M ODEL
driving mileage, and ending accumulative driving mileage.
In this section, the proposed battery charging capacity
A moving average algorithm is utilized to calculate the
prediction model is introduced. Various influencing factors
charging capacity and energy, which are given by
⎡ ⎤ and specifically designed regression models are incorporated
t+k
into the prediction model to improve prediction accuracy.
  ⎢ Ui, j ⎥
1⎢ ⎥ The considered variables and the prediction model structure
Ui,t ⎢ j =t−k ⎥
= ⎢ t+k ⎥ (4) are described in Section V-A. The modeling performance is
Ii,t 2k ⎢ ⎥
⎣ Ii, j ⎦ described and examined in Section V-B.
j =t−k
⎡ te ⎤
 A. Prediction Model and the Considered Variables
  ⎢ Ii,t t /3.6 × 10 3
⎥
Ci ⎢ ⎥ The selection and processing of the input features are
=⎢
⎢
t=ts
te
⎥
⎥ (5) important to improve charging capacity prediction accuracy.
Ei ⎣  6⎦
Ii,t Ui,t t /3.6 × 10 To properly select the input features, the factors that may have
t=ts an impact on the charging process are discussed. As described

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on June 15,2024 at 01:00:21 UTC from IEEE Xplore. Restrictions apply.
994 IEEE TRANSACTIONS ON TRANSPORTATION ELECTRIFICATION, VOL. 8, NO. 1, MARCH 2022

Nonlinear and coupled feature contributions can be included

using the polynomial combinations of features to increase the
complexity of the model inputs [37]. The highest polynomial
combination order is set to 2. The generated set of features
are given as
 
1, x i , x ik x 2−k
j |i, j ∈ [1, 2, 3, 4]; i  = j ; k ∈ [0, 1] (6)

where i and j are the indexes of features. In this study, the total
number of features in the generated set is 16.
The model training and prediction processes are shown
in Fig. 6. Regarding the four considered features, the charging
rates are discrete, while the other three are continuous. A clas-
sification process is used to deal with the charging processes
with distinct charging rates. For the training process displayed
on the left-hand side of Fig. 6, a classification process ζ is
Fig. 5. Considered featured in the fault diagnosis model development.
(a) Mileage variation with the increasing segment number. (b) SOC change applied to the training data [Xt , Yt ] to obtain several subsets
with the increasing segment number. (c) SOC change with time. (d) Average [Xti , Yti ]. The charging modes are distinct (corresponding to
temperature variation with the increasing segment number. various charging rates) so that
{[Xt1 , Yt1 ], [Xt2 , Yt2 ], . . . , [Xtn , Ytn ]} ← ζ ([Xt , Yt ]). (7)
in [32], these primary factors mainly include battery SOC
change, ambient temperature, battery aging state, and charging Each training data subset is fed into a regression model (ϕ)
power rate. For the battery SOC change, there is a strong corre- for training. After that, a set M of regression models with the
lation between the battery SOC change and the charged capac- same number of charging mode categories is established. This
ity during a charging session, and an explicit SOC imbalance means that
among cells may considerably reduce the charged capacity so
M ← {ϕ1 (x; θ 1 ), ϕ2 (x; θ 2 ), . . . , ϕn (x; θ n )} (8)
that the battery SOC change is considered in this study [33].
The ambient temperature can also greatly affect the charged where x is the input data and θ k is the parameter vector of the
capacity, as the underlying electrochemical reactions inside kth model.
EV batteries are highly sensitive to extreme temperatures [34]. In the prediction process, a single EV charging process
The capacity degradation is a direct indicator for battery aging record undergoes the model selection procedure to choose
state, which renders it an indispensable variable for battery the regression model based on the corresponding charging
charging capacity analysis [35]. Note that the accumulated mode. The selected regression model implements to obtain
driving mileage of the same EV model is used to represent the predicted charging capacity by
the battery aging state in this study. For the charging power
rate, it was revealed that different charging power rates would ŷr = ϕk (Xr ; θ k ). (9)
impose an impact on the evolutions of the charging voltage and
The regression model utilized here is the XGBoost regres-
capacity [36], and the difference of distinct charging power
sion model, which is a tree-based boosting model capable of
rates can be embodied by the time used for an isometric
dealing with sparse data [38]. Since the XGBoost model is
SOC increase. Hence, four variables are used as the inputs of
a mature model and has been extensively utilized in multiple
the battery charging capacity prediction model, i.e., the SOC
research fields, only the loss function L is given by
change, ambient temperature, accumulative driving mileage, ⎧
and minutes per SOC increase. ⎪
⎨L = l( ŷi , yi ) + ( f k )
The curves of the four considered features are displayed i k (10)
in Fig. 5. Fig. 5(a) shows the accumulative driving mileage ⎪
⎩ ( f ) = γ T + 1 λω2
with the increasing number of charging cycles. It is worth 2
noting that the accumulative driving mileage directly reflects where l is a differentiable convex loss function, ŷi and yi
the aging levels of both the vehicle and the battery system. are the prediction and the target value of the i th leaf, and 
Fig. 5(b) shows the variations of battery SOC during different is a penalty item to decrease the model complexity. Detailed
charging sessions. Fig. 5(c) compares the time consumed in modeling procedures can be found in [38].
these charging activities. The charging speed can be approxi-
mately classified into three fixed-value categories. This char-
acteristic is further considered in constructing the prediction B. Prediction Results and Mutual Verification
model. Fig. 5(d) displays the average ambient temperatures of The performance of the proposed prediction model is exam-
various charging sessions. The average ambient temperature ined and the mutual verification results are discussed. Seven
varies greatly among adjacent charging sessions. This indicates existing and extensively used prediction models are utilized
the importance of considering the ambient temperature in to compare and validate the prediction accuracy of the pro-
charging capacity prediction. posed prediction model. To measure the prediction accuracy,

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on June 15,2024 at 01:00:21 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: DATA-DRIVEN METHOD FOR BATTERY CHARGING CAPACITY ABNORMALITY DIAGNOSIS 995

Fig. 6. Training and prediction processes for the prediction model.

Fig. 7. Prediction and validation results. (a) Prediction performance of the proposed model. (b) Mean absolute errors for the comparison models.

the mean absolute error (MAE) is adopted; its formulation is the rectified linear unit function used as the activation function.
given by For the RF model, the number of trees in the forest is set as
300, and the quality of a split is measured by the mean squared
N
error. For the SVR model, the kernel-type radial basis function
MAE = | ŷi − yi |/N. (11)
is employed with the regularization parameter set as 100.
i=1
For the AdaBoost model, the maximum number of estimators
After applying the proposed model to real-world EV data, and the learning rate are set as 400 and 1, respectively. For
the EV charging capacity prediction results for 50 charging the proposed prediction model, the XGBoost model is used
sessions are depicted in Fig. 7(a). It can be observed that for regression, and the number of gradient boosting trees,
the difference between the predicted and the real capacity is the maximum tree depth, and the learning rate are set as 300,
quite small. Despite the existence of irregular SOC changes in 6, and 0.1, respectively. It is worth mentioning that the inputs
different charging sessions, the predicted charging capacities of these models are the same as that of the proposed model
can still closely follow the real capacities. for a fair comparison.
To validate the superiority of the proposed prediction To analyze the prediction performance in real-world
model, seven prediction models are also used for comparison. scenarios, charging capacity predictions are carried out in real-
They are the linear regression model (Linear), the proportion world EV applications. The performance of different predic-
model (PM-linear model with only one SOC change feature), tion models using large-scale EV data is delineated in Fig. 7(b)
the Lasso regression model (Lasso), the neural network model with the average MAEs used for quantitative assessment. The
(NN), the random forest regression model (RF), the support proposed model achieves the lowest average MAE of 1.23.
vector regressor model (SVR), and the AdaBoost model (Ada). The model with SOC change as the only input has the highest
The hyperparameters for the NN, RF, SVR, and Ada models MAE. This suggests that roughly estimating charging capacity
are obtained using a grid search algorithm, and the key using SOC values is insufficient in practice. Although both
parameters are given as follows. For the NN model, it has the proposed model and the RF model can achieve better pre-
two hidden layers, each of which consists of ten neurons with diction performance than the linear models, their high model

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on June 15,2024 at 01:00:21 UTC from IEEE Xplore. Restrictions apply.
996 IEEE TRANSACTIONS ON TRANSPORTATION ELECTRIFICATION, VOL. 8, NO. 1, MARCH 2022

Fig. 8. Proposed fault determination mechanism.

Fig. 9. Statistics of the relative errors. (a) Distribution. (b) CDF.

complexity leads to more intensive computation. Utilizing big determination and the determination criteria calculation.
data computing frameworks, such as Spark, the computing In terms of the determination criteria calculation, the charg-
efficiency can be improved [39]. ing capacity prediction model is used to process large-scale
EV data. The cumulative distribution function (CDF) of the
VI. FAULT D IAGNOSIS FOR EV BATTERY relative error is calculated. Different intervals for the relative
C HARGING C APACITY error CDF are then investigated to determine the criteria.
In this section, the fault determination and diagno- By using the constructed intervals, the permissible relative
sis approaches of battery charging capacity are addressed. error intervals can be obtained. Regarding the fault determi-
To determine the charging capacity fault or abnormality occur- nation process, the obtained permissible relative error interval
rence in a charging session, the permissible interval of the is utilized for determining whether a new calculated relative
charging capacity value should be determined, and the center error is abnormal. Those exceeding the maximum or minimum
and range of the interval need to be obtained. The high- bounds of the interval are flagged as abnormal charging
accuracy charging capacity prediction model developed in capacities.
Section V can provide the center of the interval. Thereby, once In Fig. 9(a), the distribution of the relative error for the
the range of the interval is determined, the permissible interval charging capacity prediction is shown. The CDF is given
of the charging capacity value can be acquired. As different in Fig. 9(b). The Kolmogorov–Smirnov test is used to confirm
batteries are used in EVs, a data-driven fault determination whether the distribution of relative errors conforms to the
approach is preferred. In this context, statistical information normal distribution. A null hypothesis is made so that the
is utilized to determine the abnormality of the new charging distribution to be tested is identical to the normal distribution.
record inputs. The EVs used in this study are of the same However, if the p-value of the test approaches 0, the null
model powered by a ternary lithium-ion battery with a nominal hypothesis would be invalid, and thus, the distribution of rela-
capacity of 110 Ah and a total energy of about 38 kWh. tive errors does not conform to the normal distribution. Hence,
Fig. 8 presents the charging capacity fault determina- the probability intervals of CDF are used for choosing the
tion mechanism. It consists of two parts, i.e., the fault permissible interval of the relative error. Fig. 9 shows the three

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on June 15,2024 at 01:00:21 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: DATA-DRIVEN METHOD FOR BATTERY CHARGING CAPACITY ABNORMALITY DIAGNOSIS 997

Fig. 10. Charging capacity fault determination results.

TABLE I
D IAGNOSED R EAL -W ORLD C HARGING R ECORDS T HAT A RE D ETERMINED AS A BNORMAL O NES

probability intervals of CDF, i.e., [5%, 95%], [10%, 90%], charging records of the abnormalities are given in Table I.
and [25%, 75%], which can provide different options for fault It shows the start time, end time, average ambient temperature,
determination. With a close observation on Fig. 9(b), these accumulative driving mileage, start SOC, end SOC, charging
three probability intervals correspond to the respective relative capacity, charging capacity over SOC change (Ccds ), and
error intervals of [−5.10%, 6.08%], [−3.60%, 4.15%], and (Pcdf ). Note that the mean and standard deviations of Ccds for
[−2.02%, 1.72%]. Different probability intervals can be used this EV model are 109.2 and 6.3, respectively. The probability
to determine the stringent level of fault determination. In this of CDF for these records is less than 0.04, and it means that
study, the probability interval of [5%, 95%] is used as an exam- quite a few records (<4%) underperform this. Therefore, all
ple to display the fault determination process; this probability the records (114.6, 121.4, 121.0, and 119.1) can be considered
interval can cover 90% of the charging capacity data records. as significantly deviating from the mean. These verify the
The corresponding interval [−5.10%, 6.08%] is utilized as the effectiveness of the proposed diagnosis method. Overall, this
permissible interval of the prediction relative error. provides a quantitative method for identifying the battery
To examine the performance of the proposed diagnosis charging capacity abnormality in the form of probability.
method in real-world EV operations, realistic charging records
of light-duty EVs are utilized as a case study. The diagnosis
VII. C ONCLUSION
results are delineated in Fig. 10. The relative prediction error
is shown in Fig. 10(a). Two red dashed lines denote the This study presents a data-driven fault diagnosis method
upper and lower bounds of the previously obtained permissible for EV battery charging capacity. The method is developed
interval. Fig. 10(b) illustrates the fault flags of different using a machine learning-based prediction framework and a
charging records. If the fault flag is 1, an abnormal charging data-driven fault determination mechanism. First, by com-
capacity is detected; if the fault flag is 0, no abnormality bining a tree-based regression model and the charging rate
is judged. Combining the results from Fig. 10(a) and (b), classification, the proposed prediction framework considers
four charging capacity abnormalities are found. The real-world four important parameters, including the SOC change, ambient

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on June 15,2024 at 01:00:21 UTC from IEEE Xplore. Restrictions apply.
998 IEEE TRANSACTIONS ON TRANSPORTATION ELECTRIFICATION, VOL. 8, NO. 1, MARCH 2022

temperature, accumulative driving mileage, and the minutes [17] S. Yang, C. Zhang, J. Jiang, W. Zhang, Y. Gao, and L. Zhang, “A voltage
per SOC increase, as the inputs and achieves a low charging reconstruction model based on partial charging curve for state-of-
health estimation of lithium-ion batteries,” J. Energy Storage, vol. 35,
capacity prediction MAE of 1.23. The prediction model and Mar. 2021, Art. no. 102271.
the statistical regulation, as derived from large-scale EV data, [18] Z. Deng, X. Hu, X. Lin, L. Xu, Y. Che, and L. Hu, “General discharge
are then combined to synthesize a fault determination mech- voltage information enabled health evaluation for lithium-ion batter-
ies,” IEEE/ASME Trans. Mechatronics, vol. 26, no. 3, pp. 1295–1306,
anism. The performance of the proposed prediction model is Jun. 2021.
examined using unseen datasets; it achieves higher accuracy [19] Analysis of Electric Vehicle Fire Accidents in 2020. Accessed:
compared with seven commonly used prediction models. The May 1, 2021. [Online]. Available: https://www.d1ev.com/kol/138145
[20] Z. Liao, S. Zhang, K. Li, G. Zhang, and T. G. Habetler, “A survey of
fault diagnosis results are also validated using real-world EV methods for monitoring and detecting thermal runaway of lithium-ion
charging records. It has the potential of being used for different batteries,” J. Power Sources, vol. 436, Oct. 2019, Art. no. 226879.
types of EVs to quantitatively diagnose the charging capacity [21] L. Jiang, Z. Deng, X. Tang, L. Hu, X. Lin, and X. Hu, “Data-driven
fault diagnosis and thermal runaway warning for battery packs using
abnormality. real-world vehicle data,” Energy, vol. 234, Nov. 2021, Art. no. 121266.
[22] M. Schmid, E. Gebauer, C. Hanzl, and C. Endisch, “Active model-based
R EFERENCES fault diagnosis in reconfigurable battery systems,” IEEE Trans. Power
Electron., vol. 36, no. 3, pp. 2584–2597, Jul. 2020.
[1] L. Zhang, Y. Wang, and Z. Wang, “Robust lateral motion control for in- [23] T. Lin, Z. Chen, C. Zheng, D. Huang, and S. Zhou, “Fault diagnosis
wheel-motor-drive electric vehicles with network induced delays,” IEEE of lithium-ion battery pack based on hybrid system and dual extended
Trans. Veh. Technol., vol. 68, no. 11, pp. 10585–10593, Nov. 2019. Kalman filter algorithm,” IEEE Trans. Transport Electrific., vol. 7, no. 1,
[2] L. Zhang, Z. Wang, X. Ding, S. Li, and Z. Wang, “Fault-tolerant pp. 26–36, Mar. 2020.
control for intelligent electrified vehicles against front wheel steering [24] R. Zhao, J. Liu, and J. Gu, “Simulation and experimental study on
angle sensor faults during trajectory tracking,” IEEE Access, vol. 9, lithium ion battery short circuit,” Appl. Energy, vol. 173, pp. 29–39,
pp. 65174–65186, 2021. Jul. 2016.
[3] C. F. Zou, L. Zhang, X. Hu, Z. Wang, T. Wik, and M. Pecht, “A review [25] Q. Xue, G. Li, Y. Zhang, S. Shen, Z. Chen, and Y. Liu, “Fault diagnosis
of fractional-order techniques applied to lithium-ion batteries, lead-acid and abnormality detection of lithium-ion battery packs based on statisti-
batteries, and supercapacitors,” J. Power Sources, vol. 390, pp. 286–296, cal distribution,” J. Power Sources, vol. 482, Jan. 2021, Art. no. 228964.
Jun. 2018. [26] A. Sidhu, A. Izadian, and S. Anwar, “Adaptive nonlinear model-based
[4] G. Dong, F. Yang, K.-L. Tsui, and C. Zou, “Active balancing of fault diagnosis of Li-ion batteries,” IEEE Trans. Ind. Electron., vol. 62,
lithium-ion batteries using graph theory and A-star search algo- no. 2, pp. 1002–1011, Feb. 2014.
rithm,” IEEE Trans. Ind. Informat., vol. 17, no. 4, pp. 2587–2599, [27] Q. Yu, C. Wan, J. Li, R. Xiong, and Z. Chen, “A model-based sensor
Apr. 2021. fault diagnosis scheme for batteries in electric vehicles,” Energies,
[5] K. Liu, K. Li, Q. Peng, and C. Zhang, “A brief review on key vol. 14, no. 4, p. 829, Feb. 2021.
technologies in the battery management system of electric vehicles,” [28] Y. Shang, G. Lu, Y. Kang, Z. Zhou, B. Duan, and C. Zhang, “A multi-
Frontiers Mech. Eng., vol. 14, no. 1, pp. 47–64, 2019. fault diagnosis method based on modified sample entropy for lithium-ion
[6] W. Han, T. Wik, A. Kersten, G. Dong, and C. Zou, “Next-generation battery strings,” J. Power Sources, vol. 446, Jan. 2020, Art. no. 227275.
battery management systems: Dynamic reconfiguration,” IEEE Ind. [29] D. Li, Z. Zhang, P. Liu, and Z. Wang, “DBSCAN-based thermal runaway
Electron. Mag., vol. 14, no. 4, pp. 20–31, Dec. 2020. diagnosis of battery systems for electric vehicles,” Energies, vol. 12,
[7] J. Kim, A. Mallarapu, D. P. Finegan, and S. Santhanagopalan, “Modeling no. 15, p. 2977, Aug. 2019.
cell venting and gas-phase reactions in 18650 lithium ion batteries
[30] X. Hu, K. Zhang, K. Liu, X. Lin, S. Dey, and S. Onori, “Advanced fault
during thermal runaway,” J. Power Sources, vol. 489, Mar. 2021,
diagnosis for lithium-ion battery systems: A review of fault mechanisms,
Art. no. 229496.
fault features, and diagnosis procedures,” IEEE Ind. Electron. Mag.,
[8] X. Feng, M. Ouyang, X. Liu, L. Lu, Y. Xia, and X. He, “Thermal
vol. 14, no. 3, pp. 65–91, Sep. 2020.
runaway mechanism of lithium ion battery for electric vehicles: A
[31] D. Li, Z. Zhang, P. Liu, Z. Wang, and L. Zhang, “Battery fault diagnosis
review,” Energy Storage Mater., vol. 10, pp. 246–267, Jan. 2018.
for electric vehicles based on voltage abnormality by combining the long
[9] B. Mao et al., “Thermal runaway and fire behaviors of a 300 Ah lithium
short-term memory neural network and the equivalent circuit model,”
ion battery with LiFePO4 as cathode,” Renew. Sustain. Energy Rev.,
IEEE Trans. Power Electron., vol. 36, no. 2, pp. 1303–1315, Feb. 2021.
vol. 139, Apr. 2021, Art. no. 110717.
[10] A. F. C. Borray, A. Garcés, J. Merino, E. Torres, and J. Mazón, “New [32] Y. Zhao, Z. Wang, Z.-J.-M. Shen, and F. Sun, “Data-driven framework
for large-scale prediction of charging energy in electric vehicles,” Appl.
energy bound-based model for optimal charging of electric vehicles with
Energy, vol. 282, Jan. 2021, Art. no. 116175.
solar photovoltaic considering low-voltage network’s constraints,” Int.
J. Electr. Power Energy Syst., vol. 129, Jul. 2021, Art. no. 106862. [33] W. Han, C. Zou, C. Zhou, and L. Zhang, “Estimation of cell SOC
[11] Y. Zhang, Y. Zhang, Z. Ai, Y. L. Murphey, and J. Zhang, “Energy evolution and system performance in module-based battery charge equal-
optimal control of motor drive system for extending ranges of electric ization systems,” IEEE Trans. Smart Grid, vol. 10, no. 5, pp. 4717–4728,
vehicles,” IEEE Trans. Ind. Electron., vol. 68, no. 2, pp. 1728–1738, Sep. 2019.
Feb. 2021. [34] J. Lindgren and P. D. Lund, “Effect of extreme temperatures on battery
[12] T. Ouyang, P. Xu, J. Chen, J. Lu, and N. Chen, “An online prediction charging and performance of electric vehicles,” J. Power Sources,
of capacity and remaining useful life of lithium-ion batteries based on vol. 328, pp. 37–45, Oct. 2016.
simultaneous input and state estimation algorithm,” IEEE Trans. Power [35] M. Jafari, A. Gauchia, S. Zhao, K. Zhang, and L. Gauchia, “Electric
Electron., vol. 36, no. 7, pp. 8102–8113, Jul. 2021. vehicle battery cycle aging evaluation in real-world daily driving and
[13] T. Sun, B. Xu, Y. Cui, X. Feng, X. Han, and Y. Zheng, “A sequential vehicle-to-grid services,” IEEE J. Mag., vol. 4, no. 1, pp. 122–134,
capacity estimation for the lithium-ion batteries combining incremental Mar. 2017.
capacity curve and discrete Arrhenius fading model,” J. Power Sources, [36] Y. Gao, J. Jiang, C. Zhang, W. Zhang, Z. Ma, and Y. Jiang, “Lithium-
vol. 484, Feb. 2021, Art. no. 229248. ion battery aging mechanisms and life model under different charging
[14] Z. Lyu, R. Gao, and X. Li, “A partial charging curve-based data- stresses,” J. Power Sources, vol. 356, pp. 103–114, Jul. 2017.
fusion-model method for capacity estimation of li-ion battery,” J. Power [37] L. Huang, J. Jia, B. Yu, B.-G. Chun, P. Maniatis, and M. Naik, “Pre-
Sources, vol. 483, Jan. 2021, Art. no. 229131. dicting execution time of computer programs using sparse polynomial
[15] J. Li, M. Ye, K. Gao, X. Xu, M. Wei, and S. Jiao, “Joint estimation of regression,” in Proc. Adv. Neural Inf. Process. Syst., vol. 23, 2010,
state of charge and state of health for lithium-ion battery based on dual pp. 883–891.
adaptive extended Kalman filter,” Int. J. Energy Res., early access, doi: [38] T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,”
10.1002/er.6658. in Proc. 22nd ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining,
[16] C. She, Z. Wang, F. Sun, P. Liu, and L. Zhang, “Battery aging assessment Aug. 2016, pp. 785–794.
for real-world electric buses based on incremental capacity analysis and [39] Z. Han and Y. Zhang, “Spark: A big data processing platform based
radial basis function neural network,” IEEE Trans. Transport. Electrific., on memory computing,” in Proc. 7th Int. Symp. Parallel Architectures,
vol. 16, no. 5, pp. 3345–3354, May 2019. Algorithms Program. (PAAP), Dec. 2015, pp. 172–176.

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on June 15,2024 at 01:00:21 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: DATA-DRIVEN METHOD FOR BATTERY CHARGING CAPACITY ABNORMALITY DIAGNOSIS 999

Zhenpo Wang (Member, IEEE) received the Ph.D. Yang Zhao received the M.S. degree in mechanical
degree in automotive engineering from Beijing Insti- engineering from Beijing Institute of Technology,
tute of Technology, Beijing, China, in 2005. Beijing, China, in 2016, where he is currently pursu-
He is currently a Professor with Beijing Institute ing the Ph.D. degree in mechanical engineering with
of Technology and the Director of the National the National Engineering Laboratory for Electric
Engineering Laboratory for Electric Vehicles. He has Vehicles.
published four monographs and translated books as His research interests primarily involve machine
well as more than 80 technical papers. He also holds learning algorithms, large-scale EV charging pre-
more than 60 patents. His current research interests dictions, energy consumption analysis, battery fault
include pure electric vehicle integration, packaging diagnosis, and intelligent transportation.
and energy management of battery systems, and
charging station design.
Prof. Wang has been 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 the Beijing Peng Liu received the Ph.D. degree in automotive
Municipal, China. engineering from Beijing Institute of Technology,
Beijing, China, in 2011.
Chunbao Song received the B.S. degree in mechani- He is currently an Associate Professor with Beijing
cal engineering from Beijing Institute of Technology, Institute of Technology and the Director of the Col-
Beijing, China, in 2013, where he is currently pursu- laborative Innovation Center for Electric Vehicles,
ing the Ph.D. degree in mechanical engineering with Beijing. He has published two monographs and five
the National Engineering Laboratory for Electric technical papers retrieved by SCI. He also holds
Vehicles. 17 patents. His current research interests include
His research interests mainly include battery battery fault diagnosis, intelligent transportation, and
state of health and deep learning algorithms’ big data analysis.
development. Dr. Liu was a recipient of numerous awards including the third prize
of Beijing Science and Technology Award and the first prize of China
Automotive Industry Science and Technology Award.
Lei Zhang (Member, IEEE) received the Ph.D.
degree in mechanical engineering from Beijing Insti-
tute of Technology, Beijing, China, in 2016, and
the Ph.D. degree in electrical engineering from
the University of Technology, Sydney, Australia, David G. Dorrell (Fellow, IEEE) was born in St
in 2016. Helens, U.K. He received the B.Eng. degree (Hons.)
He is currently an Associate Professor with the from The University of Leeds, Leeds, U.K., in 1988,
School of Mechanical Engineering, Beijing Institute the M.Sc. degree from The University of Bradford,
of Technology. His research interests lie in the area Bradford, U.K., in 1989, and the Ph.D. degree from
of control theory and engineering applied to electri- The University of Cambridge, Cambridge, U.K.,
fied vehicles with emphases on battery management in 1993.
techniques, vehicle dynamics control, and autonomous driving technology. He was a Professor of Electrical Machines with
Dr. Zhang is a member of the Institute of Electrical and Electronics The University of KwaZulu-Natal (UKZN), Durban,
Engineers (IEEE) and China Society of Automotive Engineers (CSAE). South Africa (2015–2020), and the Director of the
He serves for the Technical Committee on Vehicle Control and Intelligence EPPEI Specialization Center in HVDC and FACTS
and the Technical Committee on Parallel Intelligence in Chinese Associa- at UKZN (2016–2020). He is currently a Distinguished Professor with The
tion of Automation (CAA). He has served as a Guest Editor for several University of the Witwatersrand, Johannesburg, South Africa. He has held
journals, including International Journal of Vehicle Design, Chinese Journal positions with Robert Gordon University, Aberdeen, U.K., the University of
of Mechanical Engineering, and China Journal of Highway and Transport. Reading, Reading, U.K., the University of Glasgow, Glasgow, U.K., and the
He serves as an Associate Editor for Proceedings of the Institution of University of Technology Sydney, Sydney, Australia. He has worked in the
Mechanical Engineers Part C—Journal of Mechanical Engineering Science, industry and carried out several industrial consultancies. He is a Chartered
SAE International Journal of Connected and Autonomous Vehicles, and SAE Engineer in the U.K. His research interests cover electrical machines, renew-
Journal of Eletrified Vehicles. He is also on the editorial boards of Electronics, able energy, and power systems.
Sustainability, and China Journal of Highway and Transport. Dr. Dorrell is a fellow of the IET.

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on June 15,2024 at 01:00:21 UTC from IEEE Xplore. Restrictions apply.