In my research, I focused on the safety, reliability, and predictive maintenance of the electric vehicle battery pack. As electric vehicles become increasingly common, the electric vehicle battery pack has become the central energy-storage unit that determines vehicle performance, driving range, and user safety. Because the electric vehicle battery pack operates under complex road, temperature, load, and charging conditions, voltage inconsistency, overcharge, overdischarge, internal short circuits, sensor faults, connection faults, and thermal runaway risks can appear gradually or suddenly. I therefore developed a data-driven framework that combines data mining, statistical diagnosis, machine learning, and deep learning to detect, locate, diagnose, and warn about faults in the electric vehicle battery pack.
My work began from a practical observation: modern electric vehicles are required to upload operational data to cloud monitoring platforms, and these platforms accumulate large volumes of real-world data from the electric vehicle battery pack. However, many platforms only perform simple alarm checks and cannot provide early warning. I aimed to transform these accumulated data into actionable safety knowledge. The main objectives of my study were to improve anomaly detection efficiency, accurately locate abnormal cells, classify fault probability, and predict future voltage behavior so that the electric vehicle battery pack can be protected before severe failure occurs.
Operating Principle and Fault Mechanisms of the Electric Vehicle Battery Pack
I first reviewed the electrochemical principle of lithium-ion batteries. During charging, lithium ions move from the positive electrode to the negative electrode, while electrons flow through the external circuit. During discharging, the process reverses. The main reactions can be summarized as follows:
$$LiCoO_2 \rightleftharpoons Li_{1-x}CoO_2 + xLi^+ + xe^-$$
$$C + xLi^+ + xe^- \rightleftharpoons Li_xC$$
$$LiCoO_2 + C \rightleftharpoons Li_{1-x}CoO_2 + Li_xC$$
For the electric vehicle battery pack, abnormal operation may accelerate side reactions. Overcharge and overdischarge can cause metal dissolution, structural damage, heat generation, and even thermal runaway. I classified faults into battery-body faults and other system faults. Battery-body faults include gradual faults and sudden faults. Other system faults include connection-component faults, sensor faults, and battery-management-system faults.
| Fault Category | Typical Examples | Main Consequence for the Electric Vehicle Battery Pack |
|---|---|---|
| Gradual battery-body fault | Capacity fade, resistance growth, cycle-life reduction | Reduced usable energy, increased inconsistency, accelerated aging |
| Sudden battery-body fault | Internal short circuit, thermal runaway, leakage | Rapid temperature rise, fire risk, severe safety hazard |
| Connection-component fault | Loose connector, oxidized contact, poor welding | Voltage distortion, local heating, abnormal resistance |
| Sensor fault | Temperature sensor error, voltage sensor drift, current sensor bias | Incorrect state estimation, false alarms, hidden risk |
| Battery-management-system fault | Software error, hardware failure, communication fault | Loss of protection, uncontrolled charging or discharging |
The triggering mechanisms of these faults are strongly coupled. Overcharge, overdischarge, high-rate charge or discharge, low temperature, high temperature, and inconsistency can all push the electric vehicle battery pack outside its safe operating window. For example, overcharge can lead to lithium plating and internal short circuits. Overdischarge can damage the positive electrode and increase resistance. High temperature accelerates chemical reactions and aging, while low temperature increases polarization and risks lithium plating. Inconsistency among cells causes some cells to be overstressed even when the pack-level voltage appears normal.
Data Source and Preprocessing for the Electric Vehicle Battery Pack
I used real-world operational data collected from a cloud monitoring platform for pure electric passenger vehicles. The data covered multiple vehicles over several years and included alarm data, battery monitoring data, vehicle location data, vehicle data, motor data, and other signals. The electric vehicle battery pack data included total voltage, current, state of charge, maximum temperature, minimum temperature, cell voltage, insulation resistance, and alarm levels.
| Data Group | Representative Signals | Relevance to the Electric Vehicle Battery Pack |
|---|---|---|
| Alarm data | Maximum alarm level, battery fault count, motor fault count | Indicates existing or imminent safety events |
| Battery monitoring data | Cell voltage, total voltage, current, SOC, temperature | Core inputs for anomaly detection and fault diagnosis |
| Vehicle location data | Longitude, latitude, positioning state | Supports operational context and route analysis |
| Vehicle data | Vehicle state, speed, mileage, insulation resistance | Reflects driver behavior and vehicle health |
| Motor data | Motor speed, torque, temperature | Provides load information for the electric vehicle battery pack |
Raw data from real vehicles inevitably contain duplicate records, missing values, outliers, and inconsistent timestamps. I therefore designed a preprocessing pipeline. First, I removed duplicate timestamps and retained the first valid record. Second, I checked continuity and deleted segments with missing or erroneous data lasting more than one minute. Third, I used cubic spline interpolation for short gaps because it preserves smoothness and avoids the oscillation problems of high-order polynomial interpolation.
For duplicate samples, the observed duplicate rates were small but not negligible. I summarized representative statistics as follows.
| Vehicle ID | All Samples | Duplicate Samples | Duplicate Rate |
|---|---|---|---|
| 1 | 487505 | 2207 | 0.452% |
| 2 | 548311 | 2706 | 0.494% |
| 3 | 361280 | 1214 | 0.336% |
| 4 | 706832 | 3406 | 0.482% |
| 5 | 202025 | 254 | 0.126% |
| 6 | 140840 | 209 | 0.148% |
For outlier detection, I selected different methods according to distribution. For battery parameters such as voltage, current, and SOC, the sample size is large, and the distributions are approximately normal. Therefore, I used the \(3\sigma\) rule. For non-normal variables such as speed, mileage, and fault counts, I used boxplot detection. The \(3\sigma\) rule is given by:
$$\mu = \frac{1}{n}\sum_{i=1}^{n}x_i$$
$$\sigma = \sqrt{\frac{1}{n}\sum_{i=1}^{n}(x_i-\mu)^2}$$
$$x_i \text{ is an outlier if } |x_i-\mu| > 3\sigma$$
The boxplot method uses quartiles and the interquartile range:
$$IQR = Q_3 – Q_1$$
$$Lower = Q_1 – 1.5IQR$$
$$Upper = Q_3 + 1.5IQR$$
After preprocessing, I obtained a cleaner data foundation for the electric vehicle battery pack. This was necessary because data mining and machine learning models are highly sensitive to noise and missing values.
Anomaly Detection and Location in the Electric Vehicle Battery Pack
To detect abnormal cells in the electric vehicle battery pack, I compared several dimensionality-reduction and clustering methods. Principal component analysis is linear and fast but may not preserve nonlinear local structure. Autoencoders can learn nonlinear compression but require more computation. The t-distributed stochastic neighbor embedding method, or t-SNE, is effective for visualizing high-dimensional data and preserving local similarities. I therefore used t-SNE to reduce the high-dimensional cell-voltage data into a two-dimensional space.
The t-SNE procedure I applied was as follows. I prepared the cell-voltage matrix, normalized the features, computed pairwise similarities, optimized the low-dimensional embedding, and visualized the result as a scatter plot. The visualization revealed main clusters, isolated clusters, and distant points. These patterns suggested potential anomalies in the electric vehicle battery pack.
After dimensionality reduction, I applied K-means clustering. K-means partitions data into \(K\) clusters by minimizing the within-cluster distance. I used the distance from each point to its cluster center as an anomaly indicator. The threshold was defined as:
$$T = \mu_k + k\sigma_k$$
where \(\mu_k\) is the mean distance to the cluster center, \(\sigma_k\) is the standard deviation, and \(k\) is a sensitivity coefficient. In my study, \(k=1.5\). Points beyond this threshold were marked as abnormal.
The K-means clustering result showed three main structures: a dense normal cluster, a small isolated abnormal cluster, and several abnormal points far from the normal cluster. These abnormal points came from a specific operating period of one electric vehicle battery pack. However, K-means clustering alone could not identify which cell was faulty. I therefore used a Z-score-based diagnostic coefficient.
The Gaussian probability density function for cell voltage was written as:
$$f(x)=\frac{1}{\sqrt{2\pi}\sigma}\exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right)$$
The probability density for cell \(i\) at time \(t\) was:
$$P_i(t)=\frac{1}{\sqrt{2\pi}\sigma}\exp\left(-\frac{(V_i(t)-\mu)^2}{2\sigma^2}\right)$$
The Z-score was:
$$Z_i(t)=\frac{V_i(t)-\mu}{\sigma}$$
To improve localization, I designed an anomaly diagnostic coefficient based on the deviation of each cell probability density from the average probability density:
$$Z_i(t)=\frac{P_i(t)-\bar{P}(t)}{\sigma_P(t)}$$
where \(\bar{P}(t)\) is the mean probability density among all cells, and \(\sigma_P(t)\) is the standard deviation of the probability densities. A larger absolute value indicates a more abnormal cell in the electric vehicle battery pack.
In one case, the seventh cell showed a clearly different voltage trajectory. Its voltage dropped rapidly, and the maximum voltage difference reached 236 mV. The minimum voltage reached 3067 mV. The Z-value of the seventh cell was much larger than those of the other cells. This confirmed that the proposed method could locate the abnormal cell in the electric vehicle battery pack.
Comprehensive Performance Evaluation of Cells in the Electric Vehicle Battery Pack
After locating abnormal cells, I evaluated their anomaly degree using an entropy weight and coefficient-of-variation method. This method combines objective weighting from entropy with variability weighting from the coefficient of variation. It provides a comprehensive score for each cell in the electric vehicle battery pack.
The voltage matrix was defined as:
$$U=\begin{bmatrix}
u_{11} & u_{12} & \cdots & u_{1t}\\
u_{21} & u_{22} & \cdots & u_{2t}\\
\vdots & \vdots & \ddots & \vdots\\
u_{n1} & u_{n2} & \cdots & u_{nt}
\end{bmatrix}$$
The proportion of cell \(i\) under time index \(j\) was:
$$p_{ij}=\frac{u_{ij}}{\sum_{i=1}^{n}u_{ij}}$$
The entropy of index \(j\) was:
$$e_j=-\frac{1}{\ln n}\sum_{i=1}^{n}p_{ij}\ln p_{ij}$$
The entropy weight was:
$$W_{1j}=\frac{1-e_j}{\sum_{j=1}^{m}(1-e_j)}$$
For the coefficient-of-variation method, I computed the mean, standard deviation, and coefficient of variation:
$$A_j=\frac{1}{n}\sum_{i=1}^{n}u_{ij}$$
$$S_j=\sqrt{\frac{1}{n}\sum_{i=1}^{n}(u_{ij}-A_j)^2}$$
$$V_j=\frac{S_j}{A_j}$$
The coefficient-of-variation weight was:
$$W_{2j}=\frac{V_j}{\sum_{j=1}^{m}V_j}$$
The combined weight was:
$$W_{cj}=\alpha W_{1j}+(1-\alpha)W_{2j}$$
In my study, I set \(\alpha=0.4\). The comprehensive score of cell \(i\) was:
$$Score_i=\sum_{j=1}^{m}W_{cj}u_{ij}$$
The upper and lower score limits were:
$$S_{max}=\sum_{j=1}^{m}W_{cj}u_{j}^{max}$$
$$S_{min}=\sum_{j=1}^{m}W_{cj}u_{j}^{min}$$
The average score and score difference were:
$$\bar{S}=\frac{1}{n}\sum_{i=1}^{n}Score_i$$
$$\Delta S_i=Score_i-\bar{S}$$
Representative results are shown below.
| Cell Number | Comprehensive Score | Upper Score Limit | Lower Score Limit |
|---|---|---|---|
| 15 | 0.9634 | 0.96334 | 0.96363 |
| 22 | 0.93894 | 0.93889 | 0.93927 |
| 23 | 0.92143 | 0.92139 | 0.92180 |
| 3 | 0.56955 | 0.56941 | 0.57078 |
| 29 | 0.50589 | 0.50574 | 0.50725 |
| 1 | 0.46549 | 0.46534 | 0.46688 |
The seventh cell had the lowest score and the largest score difference. The difference reached 0.727. This agreed with the Z-score diagnosis. I also compared the seventh cell across three dates. The score difference increased from 0.727 to 0.831 and then to 0.876. This indicated progressive performance degradation in the electric vehicle battery pack.
Fault Diagnosis Based on the \(3\sigma\)-MSS Strategy
To classify the abnormal behavior and estimate fault probability, I used a multi-stage screening strategy called \(3\sigma\)-MSS. Traditional \(3\sigma\) screening can be affected by outliers because the mean and standard deviation are sensitive to extreme values. The \(3\sigma\)-MSS method repeatedly removes outliers, recalculates the center, and obtains a more reliable center point. This improves the distinction between normal variation and true faults in the electric vehicle battery pack.
The initial voltage matrix was:
$$U^{(0)}=(U_1^{(0)},U_2^{(0)},\ldots,U_t^{(0)})$$
The deviation from the center was:
$$\Delta V_i^{(0)}=V_i^{(0)}-\mu_t^{(0)}$$
After filtering, a new matrix was obtained:
$$U^{(1)}=(U_1^{(1)},U_2^{(1)},\ldots,U_t^{(1)})$$
This process was repeated \(m\) times until convergence:
$$|\mu_t^{(m)}-\mu_t^{(m-1)}| \lt F$$
where \(F=0.001\) in my study. The fault matrix was then defined as:
$$R_{i,t}=\begin{cases}
1, & |V_i^{(m)}-\mu_t^{(m)}|>\beta\sigma_t^{(m)}\\
0, & \text{otherwise}
\end{cases}$$
The fault probability of cell \(i\) over the observation period was:
$$P_i=\frac{1}{T}\sum_{t=1}^{T}R_{i,t}$$
I defined two broad fault types. Systemic faults are stable, position-fixed, and usually caused by design or manufacturing inconsistency. Sudden faults show a sharp increase in fault probability at a specific time and may be caused by accidents or unexpected events. The electric vehicle battery pack fault probability was analyzed for multiple vehicles.
| Fault Type | Probability Pattern | Typical Location | Possible Cause |
|---|---|---|---|
| Systemic fault | Low and stable, often below 2% | Relatively fixed cell positions | Design weakness, material difference, manufacturing inconsistency |
| Sudden fault | High spike, sometimes above 15% | Not fixed | Accident, abuse, sudden internal damage, external impact |
I compared \(3\sigma\)-MSS with the connectivity-based outlier factor and the local outlier factor. The three methods produced similar overall patterns, but their probability scales differed. The local outlier factor was sensitive in low-probability cases but unstable in high-probability cases. The connectivity-based outlier factor underestimated very low fault probabilities. The \(3\sigma\)-MSS method was more stable in both low and high fault-probability regions. The comparison is summarized below.
| Algorithm | Maximum Fault Probability | Minimum Fault Probability | Observation |
|---|---|---|---|
| \(3\sigma\)-MSS | 0.0213 | 0.0153 | Stable in low and high probability regions |
| COF | 0.0191 | 0.0040 | Underestimates low-probability events |
| LOF | 0.0247 | 0.0143 | Better at low probability but unstable at high probability |
I also analyzed the electric vehicle battery pack fault probability over seasons. The highest fault probabilities were 1.99% in spring, 4.95% in summer, 3.67% in autumn, and 9.52% in winter. The average values were 1.54%, 4.31%, 3.07%, and 4.59%, respectively. Summer and winter showed higher risk. High summer temperature accelerates side reactions and aging. Low winter temperature increases polarization and may cause charging difficulty or lithium plating. These results provide maintenance guidance for the electric vehicle battery pack.
| Season | Maximum Fault Probability | Average Fault Probability | Risk Interpretation |
|---|---|---|---|
| Spring | 1.99% | 1.54% | Relatively low risk |
| Summer | 4.95% | 4.31% | High-temperature aging and thermal risk |
| Autumn | 3.67% | 3.07% | Moderate risk |
| Winter | 9.52% | 4.59% | Low-temperature polarization and charging risk |
Voltage Prediction and Fault Warning for the Electric Vehicle Battery Pack
Detection and diagnosis are important, but early warning is more valuable for the electric vehicle battery pack. I therefore developed a voltage prediction and fault warning method based on deep learning. The goal was to predict future cell voltage and compare it with warning thresholds before the fault fully develops.
I first performed correlation analysis. Pearson correlation measures linear dependence:
$$\rho=\frac{\sum_{t=1}^{n}(x_t-\bar{x})(y_t-\bar{y})}{\sqrt{\sum_{t=1}^{n}(x_t-\bar{x})^2}\sqrt{\sum_{t=1}^{n}(y_t-\bar{y})^2}}$$
Spearman correlation measures monotonic dependence using rank order:
$$R=1-\frac{6\sum_{t=1}^{n}d_t^2}{n(n^2-1)}$$
The analysis showed that total voltage and SOC had strong correlations with cell voltage. I therefore used SOC, total voltage, and historical cell voltage as key inputs. Temperature, current, mileage, and speed were also considered because they reflect operating conditions and driver behavior for the electric vehicle battery pack.
| Feature | Correlation with Cell Voltage | Role in Prediction |
|---|---|---|
| Total voltage | Very strong | Pack-level electrical state |
| SOC | Very strong | Available energy and operating point |
| Current | Moderate to strong | Load and charging/discharging stress |
| Temperature | Moderate | Reaction rate and safety condition |
| Mileage | Weak to moderate | Approximate health indicator |
| Speed | Weak to moderate | Driver behavior and load context |
I designed a CNN-LSTM hybrid neural network. The CNN extracts local and spatial features from the input sequence. The LSTM captures long-term temporal dependencies. The CNN structure included an input layer, convolution layer, batch normalization layer, ReLU activation layer, max pooling layer, flatten layer, LSTM layer, another ReLU layer, fully connected layer, and output layer.
The convolution operation can be expressed as:
$$y_j=\sum_{i=1}^{m}w_i x_{j+i-1}+b$$
The LSTM equations were:
$$f_t=\sigma(W_{fx}X_t+W_{fh}h_{t-1}+b_f)$$
$$i_t=\sigma(W_{ix}X_t+W_{ih}h_{t-1}+b_i)$$
$$\tilde{C}_t=\tanh(W_{cx}X_t+W_{ch}h_{t-1}+b_c)$$
$$C_t=f_t\odot C_{t-1}+i_t\odot \tilde{C}_t$$
$$o_t=\sigma(W_{ox}X_t+W_{oh}h_{t-1}+b_o)$$
$$h_t=o_t\odot\tanh(C_t)$$
To optimize hyperparameters, I used an improved grey wolf optimizer, or IGWO. The standard grey wolf optimizer models social hierarchy and hunting behavior. The distance to prey and position update are:
$$D=|C X_p(t)-X(t)|$$
$$X(t+1)=X_p(t)-A D$$
$$A=2a r_1-a$$
$$C=2r_2$$
$$a=2-\frac{2t}{MaxIter}$$
The three best wolves guide the pack:
$$D_\alpha=|C_1 X_\alpha-X|$$
$$D_\beta=|C_2 X_\beta-X|$$
$$D_\delta=|C_3 X_\delta-X|$$
$$X_1=X_\alpha-A_1D_\alpha$$
$$X_2=X_\beta-A_2D_\beta$$
$$X_3=X_\delta-A_3D_\delta$$
$$X(t+1)=\frac{X_1+X_2+X_3}{3}$$
I improved the grey wolf optimizer with Cauchy mutation, nonlinear parameter adjustment, and variable inertia weight. These changes increased diversity, reduced the risk of local optima, and improved convergence. The IGWO optimized the CNN-LSTM hyperparameters for the electric vehicle battery pack voltage prediction task.
I evaluated the model with RMSE, MAE, and \(R^2\):
$$RMSE=\sqrt{\frac{1}{n}\sum_{i=1}^{n}(u_i-\hat{u}_i)^2}$$
$$MAE=\frac{1}{n}\sum_{i=1}^{n}|u_i-\hat{u}_i|$$
$$R^2=1-\frac{\sum_{i=1}^{n}(\hat{u}_i-u_i)^2}{\sum_{i=1}^{n}(u_i-\bar{u})^2}$$
The comparison of different neural networks is shown below. The IGWO-CNN-LSTM model achieved the lowest error and the highest fit. The RMSE was only 4.5086 mV, and the MAE was 3.5254 mV.
| Model | MAE (mV) | MAPE | RMSE (mV) | \(R^2\) |
|---|---|---|---|---|
| BiLSTM | 7.3997 | 0.0018962 | 9.6738 | 0.99647 |
| LSTM | 9.1646 | 0.0023348 | 11.594 | 0.99493 |
| CNN-LSTM | 6.2004 | 0.0016033 | 8.4264 | 0.99732 |
| CNN-GRU | 4.4459 | 0.0011321 | 5.4142 | 0.99889 |
| IGWO-CNN-LSTM | 3.5254 | 0.00090089 | 4.5086 | 0.99923 |
Fault Diagnosis Strategy and Maintenance Recommendations
After obtaining the predicted voltage, I compared it with alarm thresholds. I classified voltage faults into overvoltage, undervoltage, rapid voltage rise, rapid voltage drop, and poor consistency. The consistency indicator was based on the standard deviation of cell voltages. The diagnostic strategy is summarized below.
| Fault Type | Classification Rule | Alarm Level |
|---|---|---|
| Cell overvoltage | \(4.3V \le U \le 4.8V\) | 3 |
| Severe overvoltage | \(U > 4.8V\) | 2 |
| Cell undervoltage | \(U \le 3.4V\) | 1 |
| Rapid voltage rise | \(\Delta U \ge 0.4V/s\) | 1 |
| Rapid voltage drop | \(\Delta U \le -0.4V/s\) | 1 |
| Poor pack consistency | \(\partial \ge 0.5V\) | 1 |
In validation, I selected 318242 samples for training and used a voltage-anomaly segment for testing. The model predicted a cell voltage above 4.3 V before the actual alarm appeared. This showed that early warning for the electric vehicle battery pack is feasible. When the predicted voltage exceeded the threshold, the system identified the voltage anomaly type and issued an alarm at the cell level.
For maintenance, I recommend that a level-3 alarm should lead to immediate stop, evacuation, and emergency service because the electric vehicle battery pack may be in a dangerous state. A level-2 alarm should lead to stopping the vehicle and inspecting the battery system before further operation. A level-1 alarm should lead to reduced load, inspection, and scheduled maintenance. These actions can reduce risk and extend the service life of the electric vehicle battery pack.
Conclusions and Future Work
In my study, I developed an integrated framework for anomaly detection, fault location, fault diagnosis, and early warning for the electric vehicle battery pack. I used real-world cloud monitoring data, preprocessing, t-SNE visualization, K-means clustering, Z-score localization, entropy weight and coefficient-of-variation evaluation, \(3\sigma\)-MSS fault probability analysis, and IGWO-CNN-LSTM voltage prediction. The main findings are as follows.
First, data preprocessing is essential for the electric vehicle battery pack because real vehicle data contain duplicates, missing values, and outliers. Cubic spline interpolation and distribution-aware outlier detection improved data quality.
Second, t-SNE and K-means clustering can reveal abnormal structures in cell voltage data. The Z-score diagnostic coefficient can accurately locate abnormal cells. The entropy weight and coefficient-of-variation method can quantify the anomaly degree and track degradation over time.
Third, the \(3\sigma\)-MSS method provides stable fault probability statistics for the electric vehicle battery pack. It distinguishes systemic faults from sudden faults and performs better than COF and LOF in the tested real-world conditions. Seasonal analysis showed higher fault risk in summer and winter.
Fourth, the IGWO-CNN-LSTM model achieved high accuracy for multi-step voltage prediction. The RMSE was 4.5086 mV, and the MAE was 3.5254 mV. This model supports early warning and fault classification for the electric vehicle battery pack.
| Research Stage | Method | Output for the Electric Vehicle Battery Pack |
|---|---|---|
| Data preprocessing | Duplicate removal, cubic spline interpolation, \(3\sigma\), boxplot | Clean and reliable operational data |
| Anomaly visualization | t-SNE | Low-dimensional voltage structure |
| Anomaly detection | K-means clustering | Abnormal clusters and points |
| Fault location | Z-score diagnostic coefficient | Specific abnormal cell number |
| Performance evaluation | Entropy weight and coefficient of variation | Comprehensive anomaly score |
| Fault diagnosis | \(3\sigma\)-MSS | Fault probability and fault type |
| Voltage prediction | IGWO-CNN-LSTM | Future cell voltage and early warning |
Future work should focus on identifying the exact physical cause of each fault, coupling multiple fault modes, and improving neural network generalization. I will also explore additional neural architectures, activation functions, and optimization algorithms to further improve voltage prediction accuracy and warning reliability for the electric vehicle battery pack. With more detailed alarm labels, higher-resolution measurements, and longer operational histories, the proposed framework can be extended to broader electric vehicle battery pack safety management applications.