1. Introduction and Research Significance
With the intensification of global climate change, reducing greenhouse gas emissions has become a common consensus. The Paris Agreement sets ambitious targets to limit the global temperature rise, urging all nations to accelerate the transition toward a low-carbon economy. In this context, China has announced the “Dual Carbon” goals and has implemented comprehensive policies to promote energy conservation and emission reduction. The transport sector, as one of the major carbon emitters, has been undergoing a profound transformation. New energy vehicles, especially battery electric vehicles, have emerged as a critical pathway toward sustainable mobility. In 2024, the penetration rate of new energy vehicles exceeded 40%, and by 2025, the production and sales volume of new energy vehicles are expected to reach nearly 16.5 million units. Alongside this rapid growth, the safety and reliability of the EV battery pack have become paramount concerns for consumers, manufacturers, and regulatory bodies.
The EV battery pack is the core energy storage component of an electric vehicle. Its performance directly affects the vehicle’s driving range, power output, and service life. However, complex electrochemical systems, variable operating conditions, and potential manufacturing defects can lead to various faults, such as thermal runaway, internal short circuits, overcharge/overdischarge, and capacity fade. These faults are often hidden, progressive, and unpredictable. Without effective monitoring and early warning systems, they can escalate into catastrophic safety accidents, threatening lives and property while damaging public confidence in electrification. It has been reported that the fire incidence rate for new energy vehicles was 2–3 times that of conventional internal combustion engine vehicles during 2021–2023, with the majority of fires originating from the EV battery pack.
Battery fault diagnosis is a complex system-level problem involving root causes, external triggers, and management system failures. Internal material aging, manufacturing irregularities, mechanical abuse, electrical abuse, thermal abuse, and BMS (Battery Management System) malfunctions can all contribute to fault development. Therefore, developing accurate and proactive fault detection and diagnosis methods is of great significance. This paper focuses on data-driven approaches, leveraging large-scale real-world vehicle data to identify early signs of battery anomalies. The goal is to build an intelligent diagnostic framework that can precisely and early identify EV battery pack faults, thereby improving vehicle safety, reducing maintenance costs, and promoting the healthy development of the new energy vehicle industry.
2. Literature Review
Battery fault diagnosis methods can generally be categorized into three groups: model-based methods, signal-processing-based methods, and data-driven methods.
2.1 Model-Based Methods
Model-based methods construct mathematical models (e.g., equivalent circuit models, electrochemical models, thermal models) to simulate battery behavior. Faults are detected by comparing the model outputs with actual measurements. He et al. proposed a physics-layer model updated by a deep learning algorithm that advanced fault warning time by an average of 51 seconds. Zhang et al. developed an online least squares support vector machine tailored for battery fault diagnosis, effectively handling small and scattered fault data. Ma et al. introduced a multi-fault diagnosis strategy using parallel principal component analysis and kernel PCA to reconstruct fault waveforms. These methods are interpretable but heavily rely on accurate model parameters, which are difficult to maintain across the battery’s entire lifetime and varying operating conditions. Their generalization capability is often limited, and they struggle to capture early and gradual faults with high sensitivity.
2.2 Signal-Processing-Based Methods
Signal-processing-based methods analyze voltage, current, temperature, and other time series signals in the time, frequency, or time-frequency domains. Liao et al. proposed a method based on wavelet packet decomposition and singular value decomposition for fault diagnosis of lithium-ion batteries. Guan et al. developed a multi-fault diagnosis method using unsupervised domain adaptation and contrastive learning. Ren et al. used a non-redundant interleaved measurement circuit and improved fuzzy entropy to detect multi-faults in battery systems. Shang et al. applied modified sample entropy for real-time multi-fault diagnosis of early battery faults. Although these methods are computationally efficient and do not require complex models, they depend on manual feature extraction and expert knowledge, which limits their universality. Additionally, they often fail to capture the complex nonlinear interactions among multiple signals.
2.3 Data-Driven Methods
Data-driven methods, particularly machine learning and deep learning techniques, extract patterns directly from historical data. Liu et al. developed a statistical analysis-based multi-fault diagnosis method for LFP battery packs. Guo et al. proposed a cloud-based data-driven hierarchical warning method. Kamboj et al. designed an intelligent fault detection system using K-means clustering and ensemble learning, achieving an accuracy of 0.94. Li et al. utilized local outlier factor and standard deviation for battery pack fault detection. Sun et al. introduced a sparse data observer for real-time multi-fault diagnosis. Deng et al. employed multiclass support vector machines (MC-SVM) to reduce dependence on sample size while improving diagnostic accuracy.
Despite the advances, several challenges remain. First, real-world battery fault data are scarce, expensive to obtain, and often severely class-imbalanced. For example, the ratio of fault samples to normal samples can be as high as 1:550. Second, most existing methods rely on a single predictive model, which may suffer from low accuracy and poor generalization. Third, the interpretability of deep learning models is often inadequate for safety-critical applications. Moreover, many studies validate their approaches using simulation data rather than actual field data, reducing their practical relevance.
3. Data Preprocessing and Fault Severity Reconstruction
The data used in this study are collected from a specific brand of pure electric passenger vehicles. A total of 10 vehicles were monitored for approximately six months, producing about 5.51 million data records. The sampling frequency is 0.1 Hz, and the recorded fields include timestamp, vehicle state, charging state, speed, operational mode, national standard SOC, temperature, cumulative mileage, and general alarm information. A sample of the raw data fields is presented in Table 1.
| Field | Description | Notes |
|---|---|---|
| vid | Vehicle ID | Unique identifier |
| yr_modahrmn | Timestamp | YYYY-MM-DD HH:MM:SS |
| vehicle_state | Vehicle state | Start, off, other, abnormal, invalid |
| charging_status | Charging state | Parking charging, driving charging, not charging, charging finished |
| speed | Vehicle speed | 0–220 km/h |
| mode | Operational mode | Electric, hybrid, fuel |
| standard_soc | SOC (national standard) | 0–100% |
| mileage | Cumulative mileage | 0–999999.9 km |
| isulate_r | Insulation resistance | 0–60000 kΩ |
| dcdc_stat | DC-DC converter status | Working, disconnected |
| max_alarm_lvl | Maximum alarm level | 0–3 |
| alarm_info | General alarm flags | 32-bit flag |
Data preprocessing begins with segmenting the continuous records into meaningful operating fragments. Based on vehicle state and charging state, we define three types of fragments: parking stationary, parking charging, and driving discharging. A segmentation point is set whenever the state remains unchanged for more than three minutes. This avoids including meaningless short transitions such as stops at traffic lights.
Several data quality issues are addressed. Duplicate records are removed. Missing values for key measurements such as maximum/minimum cell voltages and maximum temperature are handled according to the proportion of missing data: if a large portion is missing, the entire field is discarded; otherwise, the previous value is used to fill the gap. Infinity values are dropped. Finally, the records are reordered by timestamp to ensure temporal continuity.
The raw alarm data contain numerous irrelevant alarms, such as braking system faults, high-voltage interlock anomalies, and DC-DC converter faults, which are not directly related to the EV battery pack. Moreover, the fault codes are inconsistent, and some alarm bits are missing or delayed. To address this, we apply a fault tree analysis (FTA) methodology to systematically decompose the battery system faults. FTA starts from the top event (the overall system fault) and uses logical gates to connect intermediate and basic events. The reconstructed fault severity levels are defined in Table 2 according to the national standard.
| Level | Definition | Vehicle Impact |
|---|---|---|
| 0 | No fault | Normal operation |
| 1 | Minor fault | No effect on driving |
| 2 | Moderate fault | Limited performance; inspection/replacement needed |
| 3 | Serious fault | Vehicle stopped for repair |
We revised the alarm thresholds based on the national standard GB/T 38661-2020 and related literature. A new set of fault alarm flags is defined in Table 3.
| Flag | Alarm | Threshold | Severity Level |
|---|---|---|---|
| 0 | Temperature difference | > 5 °C | 1 |
| 1 | Battery high temperature | > 53 °C | 3 |
| 2 | Storage device overvoltage | Above rated upper limit | 3 |
| 3 | SOC low | < 10% | 1 |
| 4 | Cell overvoltage | > 4500 mV | 3 |
| 5 | Cell undervoltage | < 3540 mV | 2 |
| 6 | SOC high | Above threshold | 2 |
| 7 | SOC jump | > 2%/min | 1 |
| 8 | Cell consistency | > 80 mV difference | 2 |
| 9 | Insulation resistance | < 100 Ω/V | 3 |
| 10 | DC-DC temperature | > 85 °C | 1 |
| 11 | DC-DC state | Disconnected while driving | 2 |
After applying the fault tree analysis and removing irrelevant alarms, the reconstructed dataset contains approximately 1,339,250 normal samples, 69,214 level-1 faults, 42,838 level-2 faults, and 44,599 level-3 faults. The overall fault proportion is about 10.47%, which forms a meaningful but still imbalanced dataset. To prepare the data for model training, we also define a fixed observation window of 25 minutes (150 data points) before each fault event, capturing the trend leading to the fault while avoiding excessive noise.
4. Feature Engineering and Selection
Feature engineering is crucial for improving the performance of machine learning models. We derive features from multiple dimensions, including battery consistency, driving behavior, and temporal context.
4.1 Consistency Features
The EV battery pack consists of many cells. Inconsistency among cells is a major cause of premature failure. We extract the range and standard deviation of voltages and temperatures. To reduce the influence of measurement scale, we calculate the coefficient of variation (\(C_v\)) and the maximum deviation rate (\(D_v\)) as follows:
$$ C_v = \frac{1}{n} \sum_{m=1}^{n} \frac{|K_{t}^{m} – \bar{K}_t|}{\bar{K}_t} \times 100\% $$
$$ D_v = \frac{K_{t}^{\max} – K_{t}^{\min}}{\bar{K}_t} \times 100\% $$
where \(K_{t}^{m}\) is the voltage or temperature of the \(m\)-th cell at time \(t\), and \(\bar{K}_t\) is the mean value. Higher \(C_v\) or \(D_v\) indicates worse consistency.
4.2 Driving Features
The operating voltage and current of the EV battery pack are strongly affected by the driving pattern. We derive features such as vehicle speed, acceleration and deceleration intensities, and their standard deviations to reflect the driver’s behavior. These features help characterize the dynamic load profile imposed on the battery.
4.3 Temporal Features
Battery behavior also depends on ambient conditions and calendar time. We extract the season, whether the day is a weekday or weekend, and the time-of-day interval (midnight 00:00–05:59, morning 06:00–11:59, afternoon 12:00–17:59, evening 18:00–23:59). These temporal contexts provide indirect information about environmental temperature and usage patterns.
After generating many candidate features, we perform feature selection using both correlation analysis and mutual information. The Pearson correlation coefficient \(r_{XY}\) is defined as:
$$ r_{XY} = \frac{\sum_{i=1}^{n}(X_i – \bar{X})(Y_i – \bar{Y})}{\sqrt{\sum_{i=1}^{n}(X_i – \bar{X})^2}\sqrt{\sum_{i=1}^{n}(Y_i – \bar{Y})^2}} $$
where \(X_i\) and \(Y_i\) are the \(i\)-th observations, \(\bar{X}\) and \(\bar{Y}\) are the means, and \(n\) is the number of samples. To capture non-linear dependencies, we also compute the mutual information:
$$ MI(X;Y) = \sum_{x \in X} \sum_{y \in Y} p(x,y) \log \frac{p(x,y)}{p(x)p(y)} $$
For continuous variables, we approximate the probability distributions using histograms and kernel density estimation. By combining Pearson correlation and mutual information scores, we initially select the top 20 features. Further analysis with a correlation heatmap leads to a final set of 13 key features, which are listed in Table 4.
| Dimension | Feature | Description |
|---|---|---|
| Battery state | total_volt | Total voltage |
| Battery state | max_temp_diff | Temperature difference within the pack |
| Battery state | max_temp | Maximum temperature |
| Battery state | mileage | Cumulative mileage |
| Battery state | total_current | Total current |
| Battery state | isulate_r | Insulation resistance |
| Battery consistency | cell_temp_CV | Temperature coefficient of variation |
| Battery consistency | cell_volt_DV | Voltage maximum deviation rate |
| Battery consistency | cell_volt_CV | Voltage coefficient of variation |
| Battery consistency | cell_volt_CV_std | Std of voltage coefficient of variation |
| Driving behavior | speed | Vehicle speed |
| Temporal context | day_section_1 | Morning (06:00–11:59) |
| Temporal context | day_section_0 | Midnight (00:00–05:59) |
5. Data Augmentation Strategies
Class imbalance is a fundamental challenge in fault diagnosis. The severity of imbalance can be categorized into several levels as shown in Table 5. In our dataset, the original fault-to-normal ratio is approximately 1:550, which is extremely imbalanced. Even after reconstruction, the overall fault proportion is about 10.47%, still considered a moderate-to-severe imbalance. To mitigate this, we explore two data augmentation strategies: an improved Border-SMOTE method and a WGAN-GP-based generative approach.
| Imbalance Level | Ratio | Recommended Approach |
|---|---|---|
| Mild | 1:2 – 1:5 | Traditional methods |
| Moderate | 1:5 – 1:20 | Specialized sampling |
| Severe | 1:20 – 1:100 | Comprehensive strategies |
| Extreme | > 1:100 | Domain knowledge + custom methods |
5.1 Improved Border-SMOTE
The synthetic minority oversampling technique (SMOTE) generates new minority samples by interpolating between existing minority samples and their neighbors. Given a minority sample \(x_i\), a randomly selected neighbor \(x_{zi}\), and a random number \(\alpha \in [0,1]\), the new synthetic sample is:
$$ x_{new} = x_i + \alpha (x_{zi} – x_i) $$
However, standard SMOTE blindly interpolates all minority samples, which may introduce noisy samples and fail to focus on the classification boundary. Border-SMOTE improves this by only oversampling the “dangerous” minority samples that lie near the class boundary. For each minority sample \(x_i\), we find its \(k\) nearest neighbors in the full training set. If more than half of these neighbors belong to the majority class, \(x_i\) is considered a boundary sample and selected for oversampling. In Border-SMOTE2, we interpolate both with minority and majority neighbors, but constrain the interpolation parameter \(\lambda \in [0, 0.5]\) to keep the new samples close to the original boundary sample, strengthening the boundary region without expanding into the majority territory.
We applied Border-SMOTE2 to generate additional fault samples for the minority classes. The resulting augmented dataset is roughly balanced across the four classes (levels 0 to 3). Figure 1 illustrates the distribution of selected key features before and after augmentation. The probability density curves of the original and generated data overlap closely, and the cumulative distribution functions are nearly identical, confirming that the generated samples follow the underlying distribution of the real fault data.
5.2 WGAN-GP
Generative adversarial networks (GANs) have shown remarkable capability in generating high-fidelity data. However, vanilla GANs are notoriously difficult to train due to the use of Jensen-Shannon (JS) divergence, which can lead to vanishing gradients and mode collapse. The Wasserstein GAN (WGAN) replaces the JS divergence with the Earth-Mover (Wasserstein) distance, providing smoother gradients even when the distributions do not overlap. To enforce the Lipschitz constraint, WGAN uses weight clipping, which often causes gradient explosion or vanishing. WGAN-GP addresses this by adding a gradient penalty term to the critic loss.
In WGAN-GP, the objective is formulated as:
$$ \min_{G} \max_{D} \underbrace{\mathbb{E}_{x \sim P_r}[D(x)] – \mathbb{E}_{z \sim P_z}[D(G(z))]}_{Wasserstein\; distance} – \lambda \mathbb{E}_{\hat{x} \sim P_{\hat{x}}} \left[ \left( \| \nabla_{\hat{x}} D(\hat{x}) \|_2 – 1 \right)^2 \right] $$
where \(\hat{x}\) is a random interpolation point between a real sample \(x\) and a generated sample \(G(z)\):
$$ \hat{x} = \epsilon x + (1 – \epsilon) G(z), \quad \epsilon \sim U[0,1] $$
The gradient penalty term encourages the critic to have a gradient norm of 1 everywhere, which is a sufficient condition for 1-Lipschitz continuity. This leads to stable and reliable training. In our implementation, the generator is a fully connected neural network with multiple hidden layers, and the critic is similarly structured. We trained the WGAN-GP model on the minority fault samples (levels 1 to 3) separately to generate synthetic samples for each class. The training dynamics showed stable convergence: the critic’s output for real data remained positive while that for generated data stayed negative, and the gradient penalty consistently stayed around 1.002, confirming that the model learned a smooth and meaningful distance metric.
6. Stacking Ensemble Diagnostic Model
Single models often have limited accuracy and generalization capacity. To overcome this, we construct a Stacking ensemble model that integrates multiple heterogeneous base learners. Stacking is a multi-layer ensemble learning scheme. In the first layer, we train several base models using k-fold cross-validation and collect their predictions. These predictions serve as the input features for the second-layer meta-learner, which learns the optimal combination of the base models’ outputs.
6.1 Framework
We selected three complementary base learners for the first layer: LightGBM, XGBoost, and Random Forest. LightGBM is a gradient boosting framework that uses histogram-based algorithms and leaf-wise growth to efficiently handle large-scale high-dimensional data. XGBoost applies the gradient boosting method with regularization and second-order derivative optimization, providing excellent predictive accuracy and robustness. Random Forest builds multiple decision trees and aggregates their predictions to reduce variance and improve generalization.
The second layer uses Logistic Regression (LR) as the meta-learner. LR is a lightweight and interpretable model that learns to weight the base model outputs effectively. This combination leverages the strengths of different algorithms: gradient boosting methods capture complex nonlinear patterns, while Random Forest provides diversity, and LR prevents overfitting by smoothing the final decision boundary.
The 5-fold cross-validation procedure is illustrated in Figure 2. The training data is split into five folds. For each base model, we train on four folds and predict on the remaining fold, generating out-of-fold predictions. This process is repeated five times so that every sample gets a predicted value. The out-of-fold predictions form the new training set for the meta-learner. Simultaneously, the base models predict the test set during each fold, and their predictions are averaged to form the test features for the meta-learner.
6.2 Hyperparameter Optimization
To ensure fair comparison and optimal performance, we use Bayesian optimization combined with 5-fold cross-validation to tune the hyperparameters of each base model. Bayesian optimization constructs a surrogate model of the objective function and selects new hyperparameter configurations based on expected improvement. This approach is more efficient than grid search or random search, requiring fewer evaluations to find good parameters. All models are tuned individually before being used in the Stacking ensemble.
6.3 Evaluation Metrics
We evaluate the models using standard classification metrics derived from the confusion matrix. Let TP, FP, TN, FN denote the true positives, false positives, true negatives, and false negatives, respectively. The accuracy \(A\) is:
$$ A = \frac{TP + TN}{TP + TN + FP + FN} $$
The precision \(P\) measures the proportion of correctly predicted positive samples among all samples predicted as positive:
$$ P = \frac{TP}{TP + FP} $$
The recall \(R\) measures the proportion of actually positive samples that are correctly predicted:
$$ R = \frac{TP}{TP + FN} $$
The F1-score is the harmonic mean of precision and recall:
$$ F1 = \frac{2 \cdot P \cdot R}{P + R} $$
Because the churn in our problem is multiclass (levels 0, 1, 2, 3), we compute the weighted average of these metrics across classes. The macro-average is used for the final comparison.
7. Experiments and Results
7.1 Experimental Settings
After preprocessing and feature selection, the dataset contains about 1.5 million samples. For the experiments, we randomly split the data into training and test sets with an 80:20 ratio. The training set is used for data augmentation and model tuning, while the test set is reserved solely for final evaluation. All models are trained on the augmented training set and evaluated on the untouched test set.
7.2 Results with Border-SMOTE
First, we compare the performance of the individual base models and the Stacking ensemble on the dataset augmented by Border-SMOTE. The results are presented in Table 6.
| Model | A% | P-1% | P-2% | P-3% | P-ave% | F1-1% | F1-2% | F1-3% | F1-ave% | R-ave% | t/s |
|---|---|---|---|---|---|---|---|---|---|---|---|
| LightGBM | 94.41 | 96.61 | 91.79 | 93.15 | 94.09 | 93.82 | 93.00 | 94.42 | 94.20 | 94.38 | 14 |
| XGBoost | 94.50 | 95.79 | 95.58 | 97.13 | 95.06 | 95.11 | 94.53 | 94.54 | 94.55 | 94.10 | 11 |
| Random Forest | 91.94 | 91.30 | 96.13 | 97.18 | 93.21 | 92.89 | 94.23 | 90.15 | 92.05 | 91.22 | 48 |
| Stacking | 97.20 | 98.56 | 96.21 | 97.79 | 97.22 | 98.43 | 96.19 | 97.48 | 97.17 | 97.13 | 85 |
The Stacking ensemble achieves an accuracy of 97.20%, which is 2.7 percentage points higher than the best single model (XGBoost with 94.50%). The average precision and average F1-score are also significantly improved. In particular, the Stacking model obtains high precision and F1-score for all three fault levels, indicating that it can effectively distinguish between different severity levels of EV battery pack faults. The training time is higher because of the k-fold cross-validation in the ensemble, but this is acceptable for offline training and periodic updates.
We further compare the Stacking model under two data augmentation methods: standard SMOTE and the improved Border-SMOTE. The results are shown in Table 7.
| Method | Accuracy (%) | Precision (%) | Recall (%) | F1-score (%) | t/s |
|---|---|---|---|---|---|
| SMOTE | 96.06 | 96.20 | 95.90 | 96.04 | 91 |
| Border-SMOTE | 97.20 | 97.22 | 97.13 | 97.17 | 85 |
The improved Border-SMOTE brings an improvement of more than 1.0 percentage point across all metrics, confirming that focusing on boundary samples helps the model learn a more accurate decision boundary and reduces overfitting.
7.3 Results with WGAN-GP
Next, we evaluate the models using the dataset augmented by WGAN-GP. The performance of the single models and the Stacking ensemble is summarized in Table 8.
| Model | A% | P-1% | P-2% | P-3% | P-ave% | F1-1% | F1-2% | F1-3% | F1-ave% | R-ave% |
|---|---|---|---|---|---|---|---|---|---|---|
| LightGBM | 93.08 | 92.66 | 95.28 | 95.26 | 93.56 | 93.20 | 93.66 | 95.15 | 93.37 | 93.23 |
| XGBoost | 99.17 | 99.03 | 99.41 | 99.49 | 99.23 | 99.06 | 99.45 | 99.35 | 99.22 | 99.21 |
| Random Forest | 98.82 | 99.11 | 99.41 | 99.51 | 98.97 | 99.09 | 99.27 | 98.55 | 98.85 | 98.73 |
| Stacking | 99.81 | 99.90 | 99.70 | 99.65 | 99.79 | 99.90 | 99.80 | 99.63 | 99.80 | 99.81 |
The Stacking ensemble again achieves the best overall performance with an accuracy of 99.81%. The precision and F1-scores for each fault level are above 99%, demonstrating exceptionally reliable fault classification for the EV battery pack. Interestingly, the performance gap between the Stacking ensemble and the strongest single models (XGBoost and Random Forest) has narrowed to less than 1.0 percentage point. This indicates that the high-quality synthetic data generated by WGAN-GP reduces the inherent difficulty of the classification task, allowing even relatively simple models to attain near-optimal performance.
To further verify this observation, we included three additional base models—AdaBoost (ADA), Gradient Boosting Decision Trees (GBDT), and Multi-Layer Perceptron (MLP)—under the same WGAN-GP augmented dataset. The results are presented in Table 9.
| Model | A% | P-1% | P-2% | P-3% | P-ave% | F1-1% | F1-2% | F1-3% | F1-ave% | R-ave% |
|---|---|---|---|---|---|---|---|---|---|---|
| AdaBoost | 97.73 | 98.83 | 98.37 | 97.03 | 97.79 | 97.63 | 97.79 | 97.86 | 97.70 | 97.64 |
| GBDT | 99.71 | 99.79 | 99.68 | 99.54 | 99.69 | 99.79 | 99.71 | 99.58 | 99.70 | 99.71 |
| MLP | 99.32 | 99.57 | 98.67 | 99.03 | 99.22 | 99.63 | 98.91 | 99.14 | 99.27 | 99.31 |
| Stacking | 99.81 | 99.90 | 99.70 | 99.65 | 99.79 | 99.90 | 99.80 | 99.63 | 99.80 | 99.81 |
GBDT performs exceptionally well, achieving 99.71% accuracy and above 99.5% for all metrics, which is very close to the Stacking ensemble. This supports the notion that WGAN-GP-generated data effectively enhances the separability of fault classes, thereby improving the performance of even a single strong learner. The Stacking model still provides the highest accuracy and the most balanced performance across all fault levels, making it the most robust choice for safety-critical applications.
7.4 Discussion
The experimental results demonstrate that both data augmentation strategies improve fault diagnosis performance for EV battery packs. Border-SMOTE focuses on boundary samples and yields a significant improvement over standard SMOTE. WGAN-GP provides even greater gains by generating realistic and diverse synthetic samples, which helps the model learn the intricate feature manifold of fault conditions. The Stacking ensemble effectively combines the strengths of diverse algorithms and consistently outperforms each individual base model, especially when the data are augmented with Border-SMOTE. With WGAN-GP, the ensemble still provides a marginal but consistent advantage, making it a reliable and safe choice for real-world deployment.
One of the key findings is that high-quality data augmentation can substantially reduce the model design complexity. When the data distribution is well represented, even a relatively simple model such as GBDT can achieve near-99% accuracy. This suggests that future research should not only focus on developing more sophisticated models but also on enhancing the quality and completeness of the training data. Data augmentation techniques like WGAN-GP can be a powerful tool to achieve this for rare and critical failure modes.
8. Conclusion and Future Work
This paper presents a comprehensive data-driven fault detection and diagnosis framework for EV battery packs. The main contributions are threefold. First, we reconstructed the fault information from raw vehicle data using fault tree analysis, defining clear 0–3 severity levels based on national standards and domain knowledge. This systematic restructuring eliminated irrelevant alarms and resulted in a well-defined dataset with a fault proportion of 10.47%. Second, we introduced two data augmentation strategies, namely improved Border-SMOTE and WGAN-GP, to address the extreme class imbalance in the dataset. Third, we constructed a Stacking heterogeneous ensemble model with LightGBM, XGBoost, and Random Forest as base learners and Logistic Regression as the meta-learner. Experimental results on real-world vehicle data show that the proposed method achieves an accuracy of 97.20% with Border-SMOTE augmentation and reaches as high as 99.81% with WGAN-GP augmentation, significantly outperforming individual base models.
Future work will explore several directions to further enhance the robustness and practicality of the diagnostic system. First, we plan to incorporate different combinations of base models and meta-learners, including deep learning architectures, to determine whether further performance gains are possible. Second, we aim to investigate the coupling mechanisms among different battery faults, considering additional contextual information such as driving behavior, weather conditions, and road types. This could enable more precise fault attribution and early warning. Third, the current model is trained offline; developing online adaptive algorithms that continuously update the model with new operational data is essential for tracking battery aging and evolving fault patterns. Finally, we intend to collaborate with industry partners to collect more diverse fault data across different battery chemistries and pack designs, thereby improving the generalization capability of the proposed framework across the EV ecosystem.
In summary, the proposed “data optimization – data augmentation – ensemble learning” diagnostic framework provides a robust and scalable solution for EV battery pack fault detection and diagnosis. It not only enhances the safety and reliability of electric vehicles but also offers a valuable reference for data-driven fault diagnosis in other industrial domains.