This article is licensed under CC-BY-NC-ND 4.

http://pubs.acs.org/journal/acsodf Article

Anomaly Detection Method for Lithium-Ion Battery Cells Based on

Time Series Decomposition and Improved Manhattan Distance
Algorithm
Minghu Wu, Shufan Zhang, Fan Zhang,* Rui Sun, Jing Tang, and Sheng Hu

Cite This: ACS Omega 2024, 9, 2409−2421 Read Online

See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

ACCESS Metrics & More Article Recommendations

Downloaded via FUNDACAO UNIV FED DO ABC on May 27, 2025 at 18:22:55 (UTC).

ABSTRACT: Abnormalities in individual lithium-ion batteries can cause the entire

battery pack to fail, thereby the operation of electric vehicles is affected and safety
accidents even occur in severe cases. Therefore, timely and accurate detection of
abnormal monomers can prevent safety accidents and reduce property losses. In this
paper, a battery cell anomaly detection method is proposed based on time series
decomposition and an improved Manhattan distance algorithm for actual operating
data of electric vehicles. First, time series decomposition is performed on the voltage
data of all battery cells in the battery pack to obtain the voltage trend component of
each cell. Then, the improved Manhattan distance algorithm is utilized to calculate
and compare the Manhattan distance values between adjacent cell trend
components, to determine the abnormal cells inside the battery pack. Furthermore,
the Manhattan distance values at the same sampling moment are calculated within
the data sequence to detect the specific time when the abnormal cells malfunction.
The data analysis and experimental verification results based on actual vehicle operating conditions indicate that this method can
accurately identify an abnormal cell within the battery pack and diagnose the specific moment of abnormality in the battery cell at an
early stage of failure, with good robustness.

1. INTRODUCTION automated manufacturing processes can lead to changes in

Transportation has significantly expanded the scope of human battery performance,10−13 resulting in inconsistent cell
activities and facilitated exchanges among different countries parameters within the battery pack.14 These inconsistencies
can lead to uneven charging and discharging of individual cells,
and regions.1 New electric energy vehicles are playing an
as well as varying degrees of self-discharge in series−parallel
increasingly important role in decarbonization in the trans-
battery packs.15 As the charging and discharging cycle
portation industry. They constitute a promising solution to a
continues, these differences will progressively intensify,
set of global challenges such as climate change and air
resulting in accelerated degradation in capacity and
pollution.2 Developing new energy vehicles has been a global
power.16,17 Moreover, premature failure of battery monomers
consensus, and developing new energy vehicles with pure
can occur, ultimately causing the failure of the entire battery
electric characteristics has become China’s national strategy.3
pack.18 It may even cause serious thermal runaway accidents.19
Due to its advantages of high energy density, low self-discharge
Therefore, accurate and effective identification of abnormal
rate, high cycle life, and no memory effect,4−6 the lithium-ion
cells within the battery pack is an important guarantee for the
battery (LIB) has gradually replaced the nickel−cadmium
safe and reliable operation of electric vehicles.20
battery, nickel−metal hydride battery, and lead acid battery as
Currently, extensive research is being conducted in the field
a mainstream choice for an electric vehicle power battery.
of battery fault diagnosis.21 Three main categories of
However, safety accidents caused by battery system faults have commonly utilized battery fault diagnosis methods exist:
been occurring frequently in recent years. As a result, accurate knowledge-based methods, model-based methods, and data-
and timely fault diagnosis of battery systems has become
increasingly important to ensure the safety and reliability of
electric vehicle operation.7 Received: September 7, 2023
Because of the inability of individual batteries to meet the Revised: December 6, 2023
power requirements of electric vehicles, they are usually Accepted: December 11, 2023
connected in series and parallel to form battery packs to Published: December 29, 2023
provide power.8,9 However, impurities in the active materials of
battery monomers and the tolerance during manual and
© 2023 The Authors. Published by
American Chemical Society https://doi.org/10.1021/acsomega.3c06796
2409 ACS Omega 2024, 9, 2409−2421
ACS Omega http://pubs.acs.org/journal/acsodf Article

driven methods. Knowledge-based fault diagnosis methods residual SOC of each cell in the battery pack was compared
primarily rely on subjective analysis techniques like manual with a preset threshold to determine and isolate the faulty
analysis, inference analysis, and logical judgment to qual- current or voltage sensor. Once an accurate mathematical
itatively diagnose faults. Such methods do not require models model is established, model-based methods can be significantly
and can be classified into three subcategories: expert system effective.34 However, these methods have high requirements
methods, 22 graph theory methods, 23 and fuzzy logic for model accuracy, hardware, and additional equipment.35 As
methods.24 Expert system based fault diagnosis methods utilize the complexity of the system increases, a significant amount of
the accumulated experience of experts to establish a knowledge time and effort is required to test and model different batteries
base and design computer programs that simulate the and faults.26,36 Meanwhile, due to the highly nonlinear nature
reasoning and decision-making process of human experts in of battery systems, most existing models are usually only
diagnosing faults.25 These methods can make use of the suitable for detecting specific types of faults and do not have
experts’ empirical knowledge to eliminate the need for universality,37 making it difficult to establish accurate models.
mathematical modeling the system, resulting in easily under- Additionally, to characterize the characteristics of battery packs
standable results. The fault tree method and the signed more accurately, some models have complex mathematical
directed graph (SDG) method are the main methods of fault forms, which increases the complexity of online anomaly
diagnosis based on graph theory. SDG is a widely adopted detection.
graphical model for describing the causal relationship of a Data-driven fault diagnosis methods, which mainly include
system, known for its features such as simple modeling, easy- machine learning,38,39 information fusion,40−43 and signal
to-understand results, and wide application.26 However, in processing,44,45 detect and separate faults by processing and
complex systems, the search process of such methods can analyzing the collected battery data. These methods do not
become extremely complicated, leading to a low correct require precise mathematical models and expert knowledge of
diagnosis rate. The fault tree based diagnosis method is a battery systems. Machine learning methods typically transform
result-to-cause analysis process. It starts from the fault state of problems into supervised learning models, which learn a
the system and conducts a step-by-step reasoning analysis to mapping relationship between constructed fault features and
determine the basic cause, degree of influence, and probability fault types. Nevertheless, the time-consuming process of
of occurrence of the fault.27 The fuzzy logic method is a acquiring appropriate fault labels and data renders the
mathematical method based on fuzzy set theory, which is used utilization of these methods challenging in practical applica-
to tackle fuzzy and uncertain problems. It quantifies the tions. Guo et al.46 proposed an unsupervised health scoring
authenticity of specific problems by introducing the concept of method for early diagnosis of battery faults. They exported new
fuzziness. However, these methods encounter challenges in features and four types of feature sets related to battery health
knowledge acquisition and rule creation. and fault status for each cell. Then a new strategy was
Model-based fault diagnosis methods involve establishing an proposed to transform typical classification problems into
accurate model that reflects the dynamic characteristics of the quantitative scoring problems through multiclustering, adopt-
battery. The output value of the model is compared with the ing three clustering algorithms based on different principles
actual measurement value to generate a residual signal. Then and randomly dividing features into feature subsets. Finally,
this residual signal is compared with a threshold to determine if early diagnosis of thermal runaway faults was achieved through
the battery has malfunctioned.28 Yang et al.29 utilized a the coupling of temperature information. With the use of the
fractional order model to describe the voltage characteristics of method of information fusion, the time and location of the
batteries. They compared the root-mean-square error and fault can be determined by analyzing the correlation
threshold between the estimated voltage from the model and coefficients between battery data or the abnormal degree of
the actual measured value. This method aimed to detect information entropy in battery operation data. However, these
external short circuit faults in the battery. Shimoga methods typically use data from the laboratory, and their
Muddappa30 proposed an observer based on an electro- results have not been verified with actual vehicle operation
chemical model and a real-time fuzzy logic algorithm and data. To achieve early fault localization, the signal processing
developed a battery internal fault diagnosis method using methods utilize various techniques such as wavelet decom-
residual relationships. Sidhu et al.31 employed the equivalent position and variational mode decomposition (VMD). These
circuit and impedance spectrum methods of lithium-ion techniques are employed to decompose the original signal into
batteries to construct multiple nonlinear characteristic fault different frequencies and extract fault features from each
models characterizing battery overcharging, discharging, and frequency. The extracted fault features are then combined with
other anomalies. They used the extended Kalman filter (EKF) anomaly recognition algorithms. Although such methods can
to estimate the terminal voltages of each model and generate accurately capture small signal fluctuations during early faults,
residual signals. The probability of characteristic faults was the signal interference generated during the actual operation of
determined through the residual to achieve accurate judgment the battery system can affect the diagnostic results, leading to a
of characteristic faults. Dey et al.32 designed three sliding mode high false alarm rate.
observers using thermodynamic models and equivalent circuit Therefore, this paper proposes an anomaly detection
models to detect, isolate, and estimate voltage, current, and method based on time series decomposition and an improved
temperature sensor faults in lithium-ion batteries based on the Manhattan distance algorithm. First, the time series decom-
error of the equivalent output of the sliding mode observer. position algorithm is utilized to decompose the battery voltage
Xiong et al.33 proposed a model-based fault detection method data under actual operating conditions, extracting the
for current and voltage sensors. The state of charge (SOC) of decomposed voltage trend components. This algorithm can
the battery was estimated using a combination of least-squares avoid cells with inconsistent voltage while completing feature
recursion and unscented Kalman filtering, and the actual SOC extraction, thus eliminating their influence on subsequent
was calculated using the Coulomb counting method. Then, the anomaly detection. Then, the similarities between the trend
2410 https://doi.org/10.1021/acsomega.3c06796
ACS Omega 2024, 9, 2409−2421
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 1. Overall framework structure diagram of STL algorithm.

components of adjacent cells are compared to determine over time. STL is divided into two parts: inner loop and outer
whether a fault has occurred. This is based on the improved loop. The inner loop mainly completes trend fitting and
Manhattan distance algorithm. Compared to the original seasonal component calculation, while the outer loop is mainly
algorithm, the improved new algorithm amplifies the voltage used to adjust the robust weights in the inner loop, that is, to
variation difference between individual cells and can more remove outliers. The inner loop is mainly divided into the
accurately detect abnormal cells within the battery pack. following six steps:
Subsequently, the improved Manhattan distance algorithm is Step 1. Initialization. The inner loop starts with the trend
used to further determine the specific moment of abnormal cell component and first assigns an initial value. We set T(0) t = 0.
failure. Finally, the effectiveness of this method is verified using The inner loop starts after initialization is completed.
real vehicle data. Step 2. Removing Trend Component. Step 2 is subtracting
The remainder of this paper is organized as follows: section the trend component of the previous iteration from the time
2 provides a detailed introduction to the constructed Seasonal- series (total signal), namely Yt − T(k)
t .
Trend decomposition procedure based on Loess (STL) and Step 3. Cycle-Subseries Smoothing. The next step is
improved Manhattan distance algorithms. In section 3, the creating a subseries set of sample points (same position after
proposed method is experimentally validated using real vehicle cycle translation) located at identical positions within each
data and compared with other methods. Section 4 is the cycle. The number of such subseries is equal to Np (the
conclusion summary and outlook. number of samples in a cycle is Np). A locally weighted
scatterplot smoothing (Loess) process is performed on each
2. PRINCIPLE AND STRUCTURE FOR ALGORITHM subseries. One time point is extended before and after each,
2.1. Time Series Decomposition Algorithm: STL. A and then combined to obtain a time series C(k+1)t of length (N
time series is a collection of observed data arranged in + 2Np).
chronological order. By statistically analyzing the time series, Step 4. Low-Pass Filtering of Periodic Subseries. Non-
we extract historical data information, identify dynamic periodic signals are extracted from periodic subseries, and three
changes, and reveal future evolution trends in behavior.47 sliding averages are performed with lengths of Np, Np and 3 on
The voltage, current, temperature, and other data of batteries C(k+1)
t . Then a Loess process is performed once to obtain a
in electric vehicles are collected in chronological order, so time series L(k+1)
t of length N, removing periodic differences.
battery data belongs to a time series. The method of processing Step 5. Detrending of Smoothed Cycle Subseries. The
time series can also be employed for battery data analysis. periodic signals are purified in periodic subseries, which means
The STL algorithm, proposed by Cleveland et al.,48 is the trend removal of C(k+1)
t , namely
suitable for decomposing time series. The principle of the STL
algorithm is to smooth the time series through robust local St(k + 1) = Ct(k + 1) Lt(k + 1) (2)
weighted regression, decomposing it into the trend compo-
nent, seasonal component (also known as periodic component After purification, periodic signals are obtained and then
in some cases), and residual component. Its decomposition removed, namely Yt − S(k+1)t .
formula is Step 6. Trend Smoothing. Step 6 is extracting the trend
signals needed in the next iteration and performing the Loess
Yt = Tt + St + R t , t = 1, ..., N (1) process on the sequence after removing periodic signals to
where Yt is the observed value in moment t, Tt, St, and Rt are obtain T(k+1)
t . Next is determining if convergence is achieved. If
the trend component, seasonal component, and residual not, move on to the next iteration. If convergence occurs,
component in moment t, respectively.48 The overall framework output the final results for the trend component and seasonal
structure of the STL algorithm is shown in Figure 1. component of this iteration, namely
The STL algorithm can handle any seasonal time series, and
the decomposition scale of its seasonal components can change St = St(k + 1) (3)

2411 https://doi.org/10.1021/acsomega.3c06796
ACS Omega 2024, 9, 2409−2421
ACS Omega http://pubs.acs.org/journal/acsodf Article

Tt = Tt(k + 1) (4) Compared with the original algorithm, the improved

Manhattan distance algorithm amplifies the difference in
R t = Yt Tt St (5) voltage changes between battery cells within the battery
pack, thereby more effectively detecting abnormal cells.
The flowchart of STL algorithm is shown in Figure 2. Detailed descriptions will be provided in the experimental
verification of section 3.

3. RESULTS AND DISCUSSION

3.1. Real-World Vehicle Data and Experimental
Process Description. The data used in this paper are derived
from the new energy vehicle big data management platform.
The platform currently serves over 200 000 vehicles and
collects more than 10 million pieces of data daily. In this
section, we choose three vehicles for experimental verification
of the algorithm: vehicle #C1, which has no faults during
driving, and two vehicles #C2 and #C3, which experience
internal short circuit faults. The vehicle descriptions are shown
in Table 1.

Table 1. Parameters on the Three Vehicles

no. vehicle type faulty cell no. no. of cells sample interval (s)
#C1 normal 96 10
#C2 potential fault #47 95 10
#C3 potential fault #31 96 10

The measurable parameters of new energy vehicle batteries

mainly include voltage, current, and temperature, which are
commonly used feature data in battery anomaly detection.
Many existing studies have shown that when there are various
abnormal faults in the battery, the voltage of the battery
exhibits more pronounced fluctuations compared to other data
during abnormal conditions. Therefore, voltage anomaly is an
extremely important fault indicator in battery anomaly
detection. In this section, we mainly use battery voltage data
from the data set to complete experimental verification.
Figure 3 provides a clear depiction of the proposed battery
Figure 2. STL algorithm flowchart. anomaly detection method.
3.2. Voltage Data Decomposition. As of the latest data
2.2. Improved Manhattan Distance Algorithm. The collection time, vehicle #C1 is a normal vehicle without any
Manhattan distance, coined by German mathematician faults, with experimental data selected as the final discharge
Hermann Minkowski in the 19th century, is a geometric process voltage data transmitted to the big data platform.
term employed in metric spaces. The Manhattan distance Vehicle #C1 consists of 96 battery cells connected in series, so
algorithm is a similarity measurement algorithm. The principle each battery has a different voltage value and the same current
of it is to determine the similarity of samples by calculating the value. Vehicle #C3 is the same as #C1, except that it is a faulty
“distance” between them. The smaller the Manhattan distance vehicle. Vehicle #C2 consists of 95 battery cells connected in
between two instance points, the higher their similarity is series and is also a faulty vehicle. Because there are no faulty
considered. In n-dimensional space, the Manhattan distance battery cells in #C1, the first cell in #C1 is chosen to display
formula between two points a(x1,x2,...,xn) and b(y1,y2,...,yn) is49 the results of a single cell for STL decomposition. Figure 4
n shows the decomposition results, which include four subgraphs
dmanh(a , b) = |yk xk| arranged in a top-to-bottom order: the original voltage data of
k=1 (6) the battery cell, the decomposed trend component, seasonal
component, and residual component.
Zhang et al. previously utilized the Manhattan distance The STL algorithm is utilized to decompose the original
algorithm for battery fault diagnosis in their studies.50 voltage data of all battery cells in #C1, #C2, and #C3. From
However, in order to enhance the detection performance of this decomposition, the trend component of each cell is
the similarity algorithm, this paper improves the Manhattan extracted. Parts (a) and (b) in Figure 5 show the unprocessed
distance algorithm. The Manhattan distance formula after discharge process voltage data and the trend component results
improvement between two points in the n-dimensional space is after composition of the battery cells in #C1, respectively.
n Similarly to Figure 5, Figures 6 and 7 display the original
dmanh(a , b) = (yk xk)2 voltage data and the decomposed trend components of #C2
k=1 (7) and #C3, respectively.
2412 https://doi.org/10.1021/acsomega.3c06796
ACS Omega 2024, 9, 2409−2421
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 3. Diagram of the anomaly detection process.

Figure 4. Decomposition results of the first cell in vehicle #C1.

From Figure 4, we can see that, compared to the original extraction of battery data but also avoids inconsistent
voltage data of the cell, the decomposed trend component monomers. This point has already been mentioned in the
filters out the interference of seasonal and residual Introduction. On the other hand, in Zhang et al.’s study on
components. By comparing (a) and (b) in Figures 5−7, it predicting the remaining useful life of a lithium-ion battery,
can be observed that the trend component only characterizes trend sequences of battery data were also used.51 However,
the essential change trend of voltage (voltage data character-
while the variational modal decomposition (VMD) method
istic) during the discharge process of the battery cell and does
not include any other components. Therefore, the inconsistent used in their study has its strengths, it is unable to effectively
component (which is included in the residual component) is deal with outliers in sample data. Conversely, the STL
also filtered out, avoiding false positives caused by the voltage decomposition algorithm can tackle this specific issue, making
inconsistency problem of the battery cells. Therefore, the STL it advantageous for performing battery cell anomaly detection.
decomposition algorithm not only completes the feature To the best of our knowledge, the STL algorithm is presented
2413 https://doi.org/10.1021/acsomega.3c06796
ACS Omega 2024, 9, 2409−2421
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 5. Discharge process raw data (a) and decomposed trend components (b) of battery cells in #C1.

Figure 6. Discharge process raw data (a) and decomposed trend components (b) of battery cells in #C2.

Figure 7. Discharge process raw data (a) and decomposed trend components (b) of battery cells in #C3.

2414 https://doi.org/10.1021/acsomega.3c06796
ACS Omega 2024, 9, 2409−2421
ACS Omega http://pubs.acs.org/journal/acsodf Article

for the first time in the field of fault detection of the lithium- the abbreviation for the Manhattan distance value). Finally, the
ion battery. Manhattan distance values are compared with the predefined
3.3. Manhattan Distance Calculation. In section 3.2, it threshold to identify faulty cells. When the Manhattan distance
is mentioned that, compared to the original voltage data of values between a cell and its two adjacent cells both exceed the
battery cells, the trend component decomposed by the STL threshold, we determine the cell as an abnormal battery cell.
algorithm can better reflect the trend and characteristics of We calculate the Manhattan distance value between the
battery cell voltage changes. Therefore, in this section, we trend components of two adjacent monomers at the same
choose to calculate the Manhattan distance of the trend sampling time and treat the sum of the Manhattan distance
component. Considering the robustness of the proposed values at all sampling times as the Manhattan distance between
detection method, we determine whether the target cell is an two adjacent battery cells. Figures 9, 10, and 11 present the
abnormal monomer by the Manhattan distance values between comparisons between the original Manhattan distance
the target cell and adjacent cells, thereby reducing the algorithm and the improved Manhattan distance algorithm
probability of false positives. Because the cells in the battery proposed in this paper for vehicles #C1, #C2, and #C3,
packs of vehicles #C1, #C2, and #C3 are connected in series as respectively, to demonstrate the advantages and effectiveness
a group, they should have similar voltages and voltage variation of the proposed method. In Figures 9−11, (a) represents the
trends unless faulty. We consider the first and last cells in a experimental results of the original algorithm while (b)
series-connected LIB pack as being adjacent to each other to represents the experimental results of the improved algorithm.
calculate the Manhattan distance. Therefore, each cell has two In Figures 9 and 11, the serial number i (i ranges from 1 to 95)
Manhattan distance values with respect to its two neighboring in the horizontal axes of the figures represents the Manhattan
cells. The diagram is shown in Figure 8 (“MD” in this figure is distance between the ith battery cell and the (i + 1)th battery
cell, while serial number 96 represents the Manhattan distance
between the 96th battery cell and the first battery cell. Figure
10 shares similarities with Figures 9 and 11, differing only in
the range of serial numbers, which spans from 1 to 95. The set
thresholds in Figures 10 and 11, represented by the red dotted
lines, are determined using the 3-sigma guidelines in statistics.
From Figures 9−11, it can be seen that, compared to the
original Manhattan distance algorithm, the improved Manhat-
tan distance algorithm has a more significant detection effect.
According to the previous analysis, because all lithium-ion
batteries in the same battery pack operate under the same
conditions, the state parameters of each battery, such as
voltage, should exhibit similar trends. In other words, the
Manhattan distance between the normal cells should be
exceedingly small. This can be seen from (a) and (b) in Figure
9, and (b) highlights the similarity between normal monomers
more effectively than (a). On the other hand, the voltage
change of a faulty battery cell can differ from that of a normal
Figure 8. Circuit diagram. cell. This difference will be relatively small and not obvious in
the early stages of a malfunction. From the comparison
between (a) and (b) in Figures 10 and 11, we can observe that
the improved Manhattan distance algorithm amplifies the

Figure 9. Experimental results of the original algorithm (a) and improved algorithm (b) in #C1.

2415 https://doi.org/10.1021/acsomega.3c06796
ACS Omega 2024, 9, 2409−2421
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 10. Experimental results of the original algorithm (a) and improved algorithm (b) in #C2.

Figure 11. Experimental results of the original algorithm (a) and improved algorithm (b) in #C3.

Figure 12. Manhattan distance between adjacent cells at the same moment in #C2: (a) original algorithm; (b) improved algorithm.

voltage variation difference between cells, so that even in the Figure 9 (b), the Manhattan distance values between all
early stages of a fault the difference between abnormal cell and battery cells and their adjacent cells are very small, which is
normal cell can be captured, and the abnormal cell can be consistent with the analysis of high similarity in voltage
more effectively detected without causing false alarms. In changes between normal cells mentioned earlier. In vehicle
2416 https://doi.org/10.1021/acsomega.3c06796
ACS Omega 2024, 9, 2409−2421
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 13. Manhattan distance between adjacent cells at the same moment in #C3: (a) original algorithm; (b) improved algorithm.

Figure 14. Experimental result of the improved algorithm from 0 to 800th sampling moment in #C2.

#C2, we process the trend components of battery voltage in maximums, one being 2.49574 and the other being 2.27149,
the experiment to detect abnormal monomers more accurately. both exceeding the threshold. Accordingly, we can determine
This is necessary because there is a certain voltage difference that #Cell 31 is an abnormal monomer in vehicle #C3. These
between one part of the battery cells and another part of the experimental results conform to the description in Table 1, so
battery cells from the beginning of sampling. This voltage the effectiveness of proposed method is proved.
difference is due to the capacity jump caused by some batteries 3.4. Abnormal Time Detection. After identifying the
aging. By removing the voltage difference caused by the abnormal cells in the battery packs of the faulty vehicles (#C2
capacity jump, we can achieve a more accurate detection of and #C3), we can further determine the specific time when the
abnormal monomers in vehicle #C2. In Figure 10 (b), the
cell undergoes an abnormality through the improved
Manhattan distance values between the 47th cell (hereinafter
Manhattan distance algorithm. The algorithms are utilized to
referred to as #Cell 47) and two neighboring cells reach two
maximums, one being 0.21493 and the other being 0.166485, calculate the Manhattan distances between the trend
both exceeding the threshold. The Manhattan distance values components of all adjacent cells in #C2 and #C3 at the
between all other cells and the adjacent cells in #C2 are less same sampling time, as shown in Figures 12 and 13. The
than the threshold, and the values are exceedingly small. On figures present two experimental result graphs: (a) for the
the basis of the above analysis, we can determine that #Cell 47 original Manhattan distance algorithm and (b) for the
is an abnormal monomer. Similarly, in Figure 11 (b), the improved Manhattan distance algorithm. The comparison
Manhattan distance values between the 31st cell (hereinafter between the two graphs highlights the advantages of the
referred to as #Cell 31) and two neighboring cells reach two improved algorithm once again.
2417 https://doi.org/10.1021/acsomega.3c06796
ACS Omega 2024, 9, 2409−2421
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 15. Experimental result of the improved algorithm from 0 to 1900th sampling moment in #C3.

From (b) in Figures 12 and 13, at all sampling times, the monomers from 0 to the 1900th sampling time in Figure 13
Manhattan distance values between normal monomers are (b), and the experimental result is shown in Figure 15. The set
exceedingly small and the variation is also very stable. From thresholds in Figures 14 and 15, represented by the red dotted
Figure 12 (b), it can be observed that, at the 425th sampling lines, are also determined using the 3-sigma guidelines in
moment, the Manhattan distance values between #Cell 47 and statistics. These results indicate that the method proposed in
two adjacent cells suddenly increase sharply, indicating the this paper can detect abnormal cells accurately, namely #Cell
onset of #Cell 47 malfunction. Subsequently, the Manhattan 47 in vehicle #C2 and #Cell 31 in vehicle #C3, before the
distance values remain stable without significant changes until battery failure deteriorates, thereby preventing the occurrence
the 870th sampling moment, at which point the Manhattan of safety accidents.
distance values between #Cell 47 and two neighboring All experimental results also match the cause of the
monomers sharply increase again, with a greater magnitude malfunction of the two faulty vehicles. For vehicle #C2, the
and slope. This indicates a worsening fault condition of #Cell #Cell 47 in the battery pack showed a sudden voltage drop at
47. Therefore, we determine the specific time when #Cell 47 the 425th sampling moment, which was confirmed to be
malfunctions and deteriorates through the improved Manhat- caused by a weak internal short circuit in the battery cell.
tan distance algorithm. Although the voltage drop did not have a significant effect on
Similarly, in Figure 13 (b), a small sharp increase in the the other normal cells due to the equalization mechanism of
Manhattan distance between #Cell 31 and two neighboring the battery pack and other factors, it followed a similar trend.
cells is observed at the 350th sampling moment initially, However, at the 870th sampling moment, the internal short
indicating that #Cell 31 begins to malfunction currently, but circuit in #Cell 47 intensified, resulting in another sudden
this malfunction is relatively weak in the early stages. voltage drop and a more noticeable voltage difference. As a
Subsequently, the Manhattan distance values progressively result, the battery management system (BMS) issued an alarm.
increase until a significant and sharp increase in both range and Similar to vehicle #C2, at the 350th sampling moment, the
slope is witnessed at the 1960th sampling moment, signifying a battery #Cell 31 in vehicle #C3 experienced a weak internal
worsening in the malfunction situation. From this, we also short circuit and a sudden voltage drop. Subsequently, at the
determine the specific time when #Cell 31 malfunctions and 1960th sampling moment, #Cell 31 had a worsening internal
worsens through the improved Manhattan distance algorithm. short circuit and the voltage dropped again. Eventually, the
During the actual operation of new energy electric vehicles, BMS alarm was raised.
the battery failure in early stages is not obvious and is difficult 3.5. Analysis of the Detection Results of Other
to detect. When the malfunction worsens, the degree of Method. In previous studies, the Pearson correlation
abnormality in the battery will rapidly evolve, ultimately coefficient method52 has been applied to fault diagnosis of
leading to safety accidents. Therefore, we need to detect power battery packs. Analyzing the detection results of the
abnormal cells within the battery pack before the battery fault above method based on actual operating data of power
deteriorates. In Figure 12 (b), due to the deterioration of the batteries and comparing them with the method proposed in
fault in #Cell 47 at the 870th sampling moment, we calculate this paper were done to explore the advantages and
the Manhattan distance values of trend components between disadvantages of different methods in multiple dimensions.
all adjacent battery cells from 0 to the 800th sampling moment. The Pearson correlation coefficient method calculates the
The experimental result is shown in Figure 14. On the basis of correlation coefficient between the voltage sequence of a
the same experimental approach, we calculate the Manhattan battery cell and adjacent battery voltage sequences, then
distance values of trend components between all neighboring determining whether a fault has occurred based on the
2418 https://doi.org/10.1021/acsomega.3c06796
ACS Omega 2024, 9, 2409−2421
ACS Omega http://pubs.acs.org/journal/acsodf Article

Figure 16. Calculated correlation coefficients for single cells and adjacent cells by original data in #C3: (a) window length = 15; (b) window length
= 30.

Figure 17. Calculated correlation coefficients for single cells and adjacent cells by trend components in #C3: window length = 15; (b) window
length = 30.

magnitude of the correlation coefficient. Taking vehicle #C3 as method application. When a larger shift window is utilized, the
an example, to make the experimental results appear clear and delay in detecting faults will be greater, and it is difficult to
not particularly confusing, we select 0 to the 500th sampling accurately determine the specific moment when the faults
moment to calculate the correlation coefficients. As shown in occur. The method proposed in this paper, compared with the
Figure 16, the correlation coefficients between single cell correlation coefficient method, exhibits higher robustness in
voltages and neighboring cell voltages in #C2 are calculated the face of real vehicle data. From Figure 13 (b), we can
using moving windows with lengths of 15 and 30. Here, we observe that the specific time when the fault occurs is
first use the original voltage data for calculation. In Figure 17, accurately detected from 0 to the 500th sampling moment.
the correlation coefficients are calculated using the voltage This method can accurately determine the specific moment of
trend components decomposed by the STL algorithm. The fault occurrence and deterioration while detecting an abnormal
results of Figures 16 and 17 both indicate that, as the length of cell. Additionally, the proposed method is not prone to false
the moving window increases, the curves become relatively positives due to signal fluctuations. To summarize, the method
smooth. By comparing Figures 16 and 17, it can be found that proposed in this paper is more suitable for anomaly detection
the correlation coefficients calculated using the trend of lithium-ion battery monomers in electrical vehicles than the
components are smoother than the results calculated using Pearson correlation coefficient method.
the original voltage data. This is because other components are
filtered out, but the detection results are not obvious enough. 4. CONCLUSION AND OUTLOOK
Due to the long sampling interval (10 s) of the actual operating Aiming at the phenomenon of individual battery abnormalities
data of electric vehicle batteries, the data fluctuates greatly, during the actual operation of electric vehicles, this paper
resulting in low robustness of the correlation coefficient proposes a lithium-ion battery anomaly detection method
2419 https://doi.org/10.1021/acsomega.3c06796
ACS Omega 2024, 9, 2409−2421
ACS Omega http://pubs.acs.org/journal/acsodf Article

based on the STL and improved Manhattan distance https://pubs.acs.org/10.1021/acsomega.3c06796

algorithms. First, the original voltage data of battery cells is
decomposed using the STL algorithm, which allows the Notes
extraction of the trend components. Afterward, the improved The authors declare no competing financial interest.
Manhattan distance algorithm is applied to calculate the
Manhattan distance values between the trend components of
all adjacent battery cells; thereby the abnormal monomer is
determined in the battery pack. Finally, by calculating the
■ ACKNOWLEDGMENTS
This work is supported by the Science and Technology Project
Manhattan distance values between neighboring cells at all of Hubei Province (2022BEC017), the Natural Science
sampling times, the specific moments of the occurrence and Foundation of Hubei Province (2022CFA007), and the
degradation of an abnormal cell fault can be determined. By Science and Technology Project of Hubei Province
analyzing the data of three actual electric vehicles in operation, (2022BEB016).
it is shown that the method proposed in this paper can
effectively and accurately detect an abnormal battery cell in a
lithium-ion battery pack. Compared with other methods, the
■ REFERENCES
(1) Sun, F. Green Energy and Intelligent Transportation�
proposed method has more advantages, and the results show promoting green and intelligent mobility. Green Energy and Intelligent
that this method exhibits strong robustness to disturbance Transportation 2022, 1 (1), 100017.
signals in real data. In future work, we will enhance our (2) Xiong, R.; Kim, J.; Shen, W.; Lv, C.; Li, H.; Zhu, X.; Zhao, W.;
investigation of anomaly detection algorithms to detect Gao, B.; Guo, H.; Zhang, C.; Sun, F. Key technologies for electric
abnormal monomers in the early stages of weak battery vehicles. Green Energy and Intelligent Transportation 2022, 1 (2),
100041.
failures. This will aid in providing early warning for battery (3) He, H.; Sun, F.; Wang, Z.; Lin, C.; Zhang, C.; Xiong, R.; Deng,
fault diagnosis. Additionally, our work in the future will focus J.; Zhu, X.; Xie, P.; Zhang, S.; Wei, Z.; Cao, W.; Zhai, L. China’s
on reducing the computational complexity of algorithms, battery electric vehicles lead the world: achievements in technology
thereby alleviating the computational pressure of real-time system architecture and technological breakthroughs. Green Energy
online detection of battery faults. We also want to enhance and Intelligent Transportation 2022, 1 (1), 100020.
sensitivity to fault signals, accelerate fault identification, and (4) Wang, S.; Fernandez, C.; Yu, C.; Fan, Y.; Cao, W.; Stroe, D.-I. A
refine the approach using a larger data set. novel charged state prediction method of the lithium ion battery
packs based on the composite equivalent modeling and improved
■ AUTHOR INFORMATION
Corresponding Author
splice Kalman filtering algorithm. J. Power Sources 2020, 471, 228450.
(5) Wang, S.; Jin, S.; Bai, D.; Fan, Y.; Shi, H.; Fernandez, C. A
critical review of improved deep learning methods for the remaining
Fan Zhang − Hubei Key Laboratory for High-efficiency useful life prediction of lithium-ion batteries. Energy Reports 2021, 7,
Utilization of Solar Energy and Operation Control of Energy 5562−5574.
Storage System, Hubei University of Technology, Wuhan (6) Lai, X.; Chen, Q.; Tang, X.; Zhou, Y.; Gao, F.; Guo, Y.; Bhagat,
430068, China; Hubei Engineering Research Center for R.; Zheng, Y. Critical review of life cycle assessment of lithium-ion
Safety Monitoring of New Energy and Power Grid batteries for electric vehicles: A lifespan perspective. eTransportation
Equipment, Hubei University of Technology, Wuhan 430068, 2022, 12, 100169.
China; Email: [email protected] (7) El Mejdoubi, A.; Chaoui, H.; Gualous, H.; Van den Bossche, P.;
Omar, N.; Van Mierlo, J. Lithium-Ion Batteries Health Prognosis
Authors Considering Aging Conditions. IEEE Transactions on Power Electronics
Minghu Wu − Hubei Key Laboratory for High-efficiency 2019, 34 (7), 6834−6844.
Utilization of Solar Energy and Operation Control of Energy (8) Wang, J.; Lu, S.; Wang, Y.; Li, C.; Wang, K. Effect analysis on
thermal behavior enhancement of lithium-ion battery pack with
Storage System, Hubei University of Technology, Wuhan
different cooling structures. Journal of Energy Storage 2020, 32,
430068, China; Hubei Engineering Research Center for 101800.
Safety Monitoring of New Energy and Power Grid (9) Li, W.; Xiao, M.; Peng, X.; Garg, A.; Gao, L. A surrogate thermal
Equipment, Hubei University of Technology, Wuhan 430068, modeling and parametric optimization of battery pack with air cooling
China for EVs. Applied Thermal Engineering 2019, 147, 90−100.
Shufan Zhang − Hubei Key Laboratory for High-efficiency (10) Rumpf, K.; Naumann, M.; Jossen, A. Experimental investigation
Utilization of Solar Energy and Operation Control of Energy of parametric cell-to-cell variation and correlation based on 1100
Storage System, Hubei University of Technology, Wuhan commercial lithium-ion cells. Journal of Energy Storage 2017, 14, 224−
430068, China; orcid.org/0009-0008-6666-8050 243.
Rui Sun − Hubei Key Laboratory for High-efficiency (11) Brand, M. J.; Kolp, E. I.; Berg, P.; Bach, T.; Schmidt, P.; Jossen,
A. Electrical resistances of soldered battery cell connections. Journal of
Utilization of Solar Energy and Operation Control of Energy Energy Storage 2017, 12, 45−54.
Storage System, Hubei University of Technology, Wuhan (12) Offer, G. J.; Yufit, V.; Howey, D. A.; Wu, B.; Brandon, N. P.
430068, China Module design and fault diagnosis in electric vehicle batteries. J. Power
Jing Tang − Hubei Key Laboratory for High-efficiency Sources 2012, 206, 383−392.
Utilization of Solar Energy and Operation Control of Energy (13) Wu, B.; Yufit, V.; Marinescu, M.; Offer, G. J.; Martinez-Botas,
Storage System, Hubei University of Technology, Wuhan R. F.; Brandon, N. P. Coupled thermal-electrochemical modelling of
430068, China uneven heat generation in lithium-ion battery packs. J. Power Sources
Sheng Hu − Hubei Key Laboratory for High-efficiency 2013, 243, 544−554.
Utilization of Solar Energy and Operation Control of Energy (14) Feng, X.; Zhang, X.; Xiang, Y. An inconsistency assessment
Storage System, Hubei University of Technology, Wuhan method for backup battery packs based on time -series clustering.
Journal of Energy Storage 2020, 31, 101666.
430068, China (15) Wang, L.; Fang, Y.; Wang, L.; Yun, F.; Wang, J.; Lu, S.
Complete contact information is available at: Understanding Discharge Voltage Inconsistency in Lithium-Ion Cells

2420 https://doi.org/10.1021/acsomega.3c06796
ACS Omega 2024, 9, 2409−2421
ACS Omega http://pubs.acs.org/journal/acsodf Article

via Statistical Characteristics and Numerical Analysis. IEEE Access (35) Wu, M.; Du, W.; Zhang, F.; Zhao, N.; Wang, J.; Wang, L.;
2020, 8, 84821−84836. Huang, W. Fault Diagnosis Method for Lithium-Ion Battery Packs in
(16) Han, W.; Zou, C.; Zhou, C.; Zhang, L. Estimation of Cell SOC Real-World Electric Vehicles Based on K-Means and the Frechet
Evolution and System Performance in Module-Based Battery Charge Algorithm. ACS Omega 2022, 7 (44), 40145−40162.
Equalization Systems. IEEE Transactions on Smart Grid 2019, 10 (5), (36) Lu, L.; Han, X.; Li, J.; Hua, J.; Ouyang, M. A review on the key
4717−4728. issues for lithium-ion battery management in electric vehicles. J. Power
(17) Zhou, L.; Zheng, Y.; Ouyang, M.; Lu, L. A study on parameter Sources 2013, 226, 272−288.
variation effects on battery packs for electric vehicles. J. Power Sources (37) Xia, B.; Shang, Y.; Nguyen, T.; Mi, C. A correlation based fault
2017, 364, 242−252. detection method for short circuits in battery packs. J. Power Sources
(18) Kim, J.; Shin, J.; Chun, C.; Cho, B. H. Stable Configuration of a 2017, 337, 1−10.
Li-Ion Series Battery Pack Based on a Screening Process for Improved (38) Li, D.; Zhang, Z.; Liu, P.; Wang, Z.; Zhang, L. Battery Fault
Voltage/SOC Balancing. IEEE Transactions on Power Electronics 2012, Diagnosis for Electric Vehicles Based on Voltage Abnormality by
27 (1), 411−424. Combining the Long Short-Term Memory Neural Network and the
(19) Wang, Q.; Wang, Z.; Zhang, L.; Liu, P.; Zhang, Z. A Novel Equivalent Circuit Model. IEEE Transactions on Power Electronics
Consistency Evaluation Method for Series-Connected Battery 2021, 36 (2), 1303−1315.
Systems Based on Real-World Operation Data. IEEE Transactions (39) Zhao, Y.; Liu, P.; Wang, Z.; Zhang, L.; Hong, J. Fault and
on Transportation Electrification 2021, 7 (2), 437−451. defect diagnosis of battery for electric vehicles based on big data
(20) Li, X.; Dai, K.; Wang, Z.; Han, W. Lithium-ion batteries fault analysis methods. Applied Energy 2017, 207, 354−362.
diagnostic for electric vehicles using sample entropy analysis method. (40) Wang, Z.; Hong, J.; Liu, P.; Zhang, L. Voltage fault diagnosis
and prognosis of battery systems based on entropy and Z-score for
Journal of Energy Storage 2020, 27, 101121.
electric vehicles. Applied Energy 2017, 196, 289−302.
(21) Zhang, F.; Xing, Z.-x.; Wu, M.-h. Fault diagnosis method for
(41) Liu, P.; Sun, Z.; Wang, Z.; Zhang, J. Entropy-Based Voltage
lithium-ion batteries in electric vehicles using generalized dimension-
Fault Diagnosis of Battery Systems for Electric Vehicles. Energies
less indicator and local outlier factor. Journal of Energy Storage 2022,
2018, 11 (1), 136.
52, 104963. (42) Yao, L.; Wang, Z.; Ma, J. Fault detection of the connection of
(22) Filippetti, F.; Martelli, M.; Franceschini, G.; Tassoni, C. lithium-ion power batteries based on entropy for electric vehicles. J.
Development of expert system knowledge base to on-line diagnosis of Power Sources 2015, 293, 548−561.
rotor electrical faults of induction motors. In Conference Record of the (43) Shang, Y.; Lu, G.; Kang, Y.; Zhou, Z.; Duan, B.; Zhang, C. A
1992 IEEE Industry Applications Society Annual Meeting; IEEE: 1992; multi-fault diagnosis method based on modified Sample Entropy for
pp 92−99. lithium-ion battery strings. J. Power Sources 2020, 446, 227275.
(23) Liu, F. F.; Zuo, H. F.; Su, Y.; Sun, J. Z. Review of research (44) Jiang, J.; Cong, X.; Li, S.; Zhang, C.; Zhang, W.; Jiang, Y. A
progresses for graph-based models in fault diagnosis method. China Hybrid Signal-Based Fault Diagnosis Method for Lithium-Ion
Mech. Eng. 2013, 24 (5), 696. Batteries in Electric Vehicles. IEEE Access 2021, 9, 19175−19186.
(24) Wu, C.; Zhu, C.; Ge, Y. A New Fault Diagnosis and Prognosis (45) Jiang, J.; Li, T.; Chang, C.; Yang, C.; Liao, L. Fault diagnosis
Technology for High-Power Lithium-Ion Battery. IEEE Transactions method for lithium-ion batteries in electric vehicles based on isolated
on Plasma Science 2017, 45 (7), 1533−1538. forest algorithm. Journal of Energy Storage 2022, 50, 104177.
(25) Hu, X.; Zhang, K.; Liu, K.; Lin, X.; Dey, S.; Onori, S. Advanced (46) Guo, W.; Yang, L.; Deng, Z.; Xiao, B.; Bian, X. Early diagnosis
Fault Diagnosis for Lithium-Ion Battery Systems: A Review of Fault of battery faults through an unsupervised health scoring method for
Mechanisms, Fault Features, and Diagnosis Procedures. IEEE real-world applications. IEEE Transactions on Transportation Elec-
Industrial Electronics Magazine 2020, 14 (3), 65−91. trification 2023, 1−1.
(26) Yang, F.; Xiao, D. Model and fault inference with the (47) Wei, W. W. S. Time Series Analysis. In The Oxford Handbook of
framework of probabilistic SDG. In 2006 9th International Conference Quantitative Methods in Psychology, Vol. 2; Little, T. D., Ed.; Oxford
on Control, Automation, Robotics And Vision; IEEE: 2006; pp 1−6. University Press: 2014.
(27) Lu, L.; Han, X.; Li, J.; Hua, J.; Ouyang, M. A review on the key (48) Cleveland, R. B.; Cleveland, W. S.; Terpenning, I.; McRae, J. E.
issues for lithium-ion battery management in electric vehicles. J. Power STL: A Seasonal-Trend Decomposition Procedure Based on Loess.
Sources 2013, 226, 272−288. Journal of Official Statistics 1990, 6 (1), 3−33.
(28) Chen, Z.; Xu, K.; Wei, J.; Dong, G. Voltage fault detection for (49) Strauss, T.; von Maltitz, M. J. Generalising Ward’s Method for
lithium-ion battery pack using local outlier factor. Measurement 2019, Use with Manhattan Distances. PLoS One 2017, 12 (1), e0168288.
146, 544−556. (50) Zhang, C.; Zhao, S.; Yang, Z.; He, Y. A multi-fault diagnosis
(29) Yang, R.; Xiong, R.; He, H.; Chen, Z. A fractional-order model- method for lithium-ion battery pack using curvilinear Manhattan
based battery external short circuit fault diagnosis approach for all- distance evaluation and voltage difference analysis. Journal of Energy
climate electric vehicles application. Journal of Cleaner Production Storage 2023, 67, 107575.
2018, 187, 950−959. (51) Zhang, C. L.; Zhao, S. S.; He, Y. G. An Integrated Method of
(30) Shimoga Muddappa, V. K. Electrochemical model based condition the Future Capacity and RUL Prediction for Lithium-Ion Battery
monitoring of a Li-ion battery using fuzzy logic. M.S. Thesis, Purdue Pack. IEEE Transactions on Vehicular Technology 2022, 71 (3), 2601−
University, 2014. 2613.
(31) Sidhu, A.; Izadian, A.; Anwar, S. Adaptive Nonlinear Model- (52) Li, X.; Wang, Z. A novel fault diagnosis method for lithium-Ion
Based Fault Diagnosis of Li-Ion Batteries. IEEE Transactions on battery packs of electric vehicles. Measurement 2018, 116, 402−411.
Industrial Electronics 2015, 62 (2), 1002−1011.
(32) Dey, S.; Mohon, S.; Pisu, P.; Ayalew, B. Sensor Fault Detection,
Isolation, and Estimation in Lithium-Ion Batteries. IEEE Transactions
on Control Systems Technology 2016, 24 (6), 2141−2149.
(33) Xiong, R.; Yu, Q.; Shen, W.; Lin, C.; Sun, F. A Sensor Fault
Diagnosis Method for a Lithium-Ion Battery Pack in Electric Vehicles.
IEEE Transactions on Power Electronics 2019, 34 (10), 9709−9718.
(34) Wu, M.; Sun, M.; Zhang, F.; Wang, L.; Zhao, N.; Wang, J.;
Huang, W. A fault detection method of electric vehicle battery
through Hausdorff distance and modified Z-score for real-world data.
Journal of Energy Storage 2023, 60, 106561.

2421 https://doi.org/10.1021/acsomega.3c06796
ACS Omega 2024, 9, 2409−2421