Electric vehicles (EVs) have become a central pillar in the global strategy to reduce carbon emissions and dependence on fossil fuels. Among all EV components, the EV battery pack is the most critical and safety-sensitive subsystem. However, field operation data from large-scale EV monitoring platforms frequently reveal failures such as thermal runaway, short circuits, and voltage inconsistency. In my research, I focus on developing robust data-driven methods for fault diagnosis of the EV battery pack. The aim is to improve the safety and reliability of the EV battery pack by identifying abnormal cells, evaluating their severity, and diagnosing system-level faults using real vehicle data rather than laboratory simulations.
In this article, I present a comprehensive pipeline for EV battery pack fault diagnosis. First, I discuss the characteristics of common faults and the theoretical foundation of machine learning models. Then I describe data preprocessing procedures used to clean real-world records. Next, I explain how random forest importance is utilized for feature screening. After that, I detail a hybrid genetic-algorithm and Bayesian-optimization (GA-BO) scheme for tuning support vector machines and random forests. I then evaluate the optimized models on multiple fault datasets. Finally, I propose an improved Z-score methodology combined with a combined-weighting approach for abnormal single-cell detection and severity evaluation. Throughout the article, I highlight the central role of the EV battery pack in every analysis stage.
Fault Types and Fundamental Theories
Faults in the EV battery pack can be divided into several broad categories: state-of-charge (SOC) anomalies, cell performance inconsistency, energy-storage device failures, motor system faults, and electronic control system faults. Among these, cell inconsistency is particularly common, because manufacturing tolerances and unequal degradation cause capacity and resistance variations among cells. The inconsistency appears as static voltage offset, slowly evolving voltage drift, or sudden voltage jumps. These patterns can be detected using monitoring data captured by the on-board battery management system (BMS).
In my study, I distinguish the following fault types: cell consistency alarm, cell under-voltage alarm, SOC-low alarm, SOC-jump alarm, temperature-difference alarm, insulation alarm, brake-system alarm, drive-motor temperature alarm, drive-motor controller temperature alarm, and DC-DC converter status alarm. All these faults can degrade the performance of the EV battery pack or indicate an impending failure of the whole energy system.
For the construction of diagnosis models, I employ two classical machine learning classifiers: support vector machines (SVM) and random forests (RF). The SVM constructs an optimal separating hyperplane in a high-dimensional feature space by solving a constrained quadratic optimization problem. Given a training set of input-output pairs, the decision function can be written as
\[
f(x) = \operatorname{sgn}\left( \sum_{i,j=1}^{n} \alpha_i y_i K(x_i, x_j) + b \right)
\]
where \(\alpha_i\) are Lagrange multipliers, \(y_i\) are class labels, \(K(\cdot,\cdot)\) is a kernel function, and \(b\) is the bias. The SVM seeks to balance the margin width and the classification error through the penalty parameter \(C\).
Random forest is an ensemble method that trains a large number of decision trees on bootstrapped samples and selects random feature subsets for each split. The final prediction is obtained by majority voting (for classification) or averaging (for regression). In my work, RF serves two purposes: as a feature-importance evaluator and as a standalone fault classifier. The flexibility of RF is especially useful for high-dimensional datasets generated by EV battery pack monitoring systems.
Since raw voltage signals from the EV battery pack often contain noise caused by electromagnetic interference and sensor inaccuracies, I also use wavelet decomposition and reconstruction as a denoising pre-processing step. The wavelet transform separates a signal into approximation and detail coefficients at multiple scales. After thresholding selected detail coefficients, the inverse transform reconstructs a smoothed voltage curve that preserves essential fault trends while removing high-frequency noise.
Data Source and Data Preprocessing
My research uses data collected from an electric-vehicle monitoring platform operated by a commercial enterprise. The dataset covers the operation of several EVs throughout 2022. The recorded information follows the Chinese national standard for EV remote service data, with a sampling period of 10 seconds. The original data include static parameters, such as vehicle identification and alarm thresholds, and dynamic telemetry, such as vehicle speed, total voltage, total current, SOC, temperature values, motor status, and cell voltage lists.
The first challenge is the complexity of nested fields. For example, the cell voltage list for the EV battery pack cannot be used directly for cell-level analysis. I therefore split these lists so that each cell becomes an independent feature column. Table 1 illustrates the resulting structure for a typical EV battery pack with 96 cells.
| Cell 1 (V) | Cell 2 (V) | Cell 3 (V) | … | Cell 96 (V) |
|---|---|---|---|---|
| 3.756 | 3.775 | 3.770 | … | 3.773 |
| 3.765 | 3.771 | 3.769 | … | 3.770 |
After splitting, the data contain some fields with a high proportion of missing values. The missing fields include the number of drive motors, motor index, motor status, controller temperature, motor speed, motor torque, motor temperature, input voltage of the controller, and DC bus current. I handle these fields differently according to their physical meaning. The drive-motor number and index are constant, so I apply forward filling. Motor status, torque, input voltage, and DC bus current are strongly affected by driving conditions; I use the K-nearest-neighbor (KNN) algorithm to impute these missing entries. For controller and motor temperatures, which change slowly, I fill the gap with the average of the preceding and succeeding values. For motor speed and torque, missing values often correspond to the vehicle being off or charged, so it is reasonable to replace them with zero. After imputation, the integrity of the dataset is substantially improved.
Duplicate records can distort model training. They typically arise from transmission errors or repeated uploads. I delete duplicate rows while keeping the first occurrence for each unique timestamp. This step guarantees that each observation represents a single moment in time.
Anomalous values also appear because of communication errors. For example, vehicle state may be recorded as 255 although only 0,1,2 are valid, and the DC-DC state may be recorded as 254 instead of 1 or 2. I detect such anomalies using boxplot analysis. Since vehicle state and DC-DC status are quasi-static within short time intervals, I replace these abnormal values with the last valid non-missing value in chronological order. After these operations, the clean dataset provides a reliable foundation for the subsequent model training and validation.
Feature Selection for EV Battery Pack Fault Diagnosis
High-dimensional data may improve the expressiveness of a model but can also introduce redundant features that increase training time and the risk of overfitting. Therefore, before building the diagnostic model, I conduct feature screening using the random forest importance measure. The rationale is that random forest randomly selects different feature subsets for each decision tree, and the importance of a feature can be quantified by observing the increase in out-of-bag error when the feature values are randomly permuted.
The procedure is summarized as follows:
1. Import the cleaned fault dataset into the RF model.
2. Construct an RF classifier using the TreeBagger function, enabling out-of-bag predictions and variable importance estimation.
3. Compute the baseline out-of-bag error with oobError.
4. Evaluate feature importance via the difference in out-of-bag error before and after permuting each feature using OOBPermutedPredictorDeltaError.
5. Sort the importance scores and select the top-ranked features for each fault type.
For the cell-consistency alarm in the EV battery pack, I started with 24 candidate features. The random forest importance scores indicate that the cell voltage maximum/minimum, total voltage, motor controller input voltage, and SOC are the most influential features. When I trained an SVM using all 24 features, the classification accuracy was only 84.67%. After progressively removing low-importance features, the accuracy changed as displayed in Table 2.
| Number of features | SVM accuracy (%) |
|---|---|
| 24 | 84.67 |
| 15 | 87.17 |
| 9 | 89.20 |
| 5 | 86.35 |
The results show that using only the top 9 features yields the highest classification accuracy. Adding irrelevant features degrades performance, while too few features discard useful information. Thus, I repeat the same screening procedure for each of the ten fault types. Since several fault types have very few samples, I focus on the most representative faults and select feature subsets accordingly.
I list selected feature combinations for the main fault types of the EV battery pack in Table 3. These features are not the same for all faults; they reflect the underlying physical relationships between the fault and the recorded telemetry.
| Fault type | Selected features |
|---|---|
| Cell consistency alarm | Cell voltage min/max, total voltage, controller input voltage, motor temperature, SOC, motor torque, DC bus current, insulation resistance |
| Cell under-voltage alarm | SOC, cell voltage min/max, accumulated mileage, motor temperature, longitude/latitude, temperature min/max, motor controller temperature, total voltage, total current, vehicle speed |
| SOC low alarm | Total voltage, total current, SOC, cell voltage min/max, temperature min/max, accumulated mileage |
| Drive motor temperature alarm | Motor temperature, controller temperature, total current, controller input voltage, SOC, motor speed, insulation resistance |
| Insulation alarm | Total voltage, longitude/latitude, cell voltage maximum, insulation resistance, motor temperature, total current, controller temperature, SOC |
| Temperature difference alarm | Controller input voltage, cell voltage maximum, total voltage, longitude/latitude, accumulated mileage, controller temperature, insulation resistance, motor temperature |
This feature-screening strategy improves the accuracy of the fault diagnosis model for the EV battery pack while reducing computational burden. In particular, it allows the classifier to focus on the most discriminative signals, such as voltage traces from individual cells, which directly reflect the internal state of the EV battery pack.
Hybrid GA-BO Optimization for RF and SVM Models
The performance of SVM and RF classifiers strongly depends on their hyperparameters. For SVM, I need to tune the kernel coefficient \(\gamma\) and the penalty parameter \(C\). For RF, I need to tune the number of trees \(n_\text{estimators}\) and the maximum depth \(d_\text{max}\). Grid search is a common choice, but it becomes computationally prohibitive as the parameter space grows. Bayesian optimization is efficient in sample number, but it may get trapped in local optima when the landscape is complex. Genetic algorithms have good global exploration ability, but they often need many generations to converge to high precision.
To combine the advantages of these approaches, I design a hybrid optimization algorithm named GA-BO. First, the genetic algorithm generates an initial population of hyperparameter candidates and evolves them through selection, crossover, mutation, and elitism for several generations. After this coarse global search, the best individuals from the GA population are used as the initial points for Bayesian optimization. Bayesian optimization constructs a Gaussian-process surrogate model of the objective function and employs an acquisition function to choose the next promising parameter combination. The new candidate is then fed back into the GA population to maintain diversity and accelerate convergence. This cycle is repeated until a stopping criterion is met, such as a maximum number of iterations or negligible improvement.
The flowchart of the GA-BO algorithm is as follows:
· Initialize the GA population randomly.
· Evaluate the fitness of each individual by cross-validation accuracy.
· Apply elitism that carries the best individual directly to the next generation.
· Perform selection, crossover, and mutation to create new candidate solutions.
· After a few GA generations, take the current best individual as the starting point for BO.
· Build a surrogate model of the cross-validation error and optimize the acquisition function to propose a new hyperparameter vector.
· Reintegrate the BO proposal into the GA population and continue the evolutionary search.
· Stop when the maximum number of generations is reached or the error does not improve significantly.
Let \( \mathcal{P} \) denote the hyperparameter space, and let \( \mathcal{L}(\theta) \) denote the cross-validation error for a hyperparameter vector \(\theta \in \mathcal{P}\). The hybrid optimization can be expressed as finding
\[
\theta^* = \arg\min_{\theta \in \mathcal{P}} \mathcal{L}(\theta)
\]
where the GA performs a broad exploration and the Bayesian subroutine refines the most promising regions of \(\mathcal{P}\).
For the SVM model, the kernel function used in my experiments is the radial basis function:
\[
K(x_i,x_j) = \exp\left(-\gamma \|x_i – x_j\|^2\right)
\]
The search range for \(\gamma\) is \([10^{-4}, 100]\), and the range for \(C\) is \([1, 100]\). Using the GA-BO algorithm on a representative cell-consistency dataset, the optimization converges to \(\gamma=100\) and \(C=9\). The resulting confusion matrix shows only a few misclassifications. For a brake-system alarm dataset, the optimized hyperparameters are \(\gamma=85\) and \(C=40\), yielding 99.25% classification accuracy.
For the random forest model, I tune \(n_\text{estimators}\) in the range \([1,100]\) and \(d_\text{max}\) in the range \([1,30]\). The GA-BO algorithm finds the optimal combination \((d_\text{max}=17, n_\text{estimators}=42)\) for the cell-consistency alarm, achieving 98.73% accuracy, and \((d_\text{max}=18, n_\text{estimators}=30)\) for the brake-system alarm, achieving 99.75% accuracy.
Experimental Validation on Real Vehicle Data
To comprehensively validate the proposed models, I use multiple metrics: accuracy, precision, recall, and the F1 score. Accuracy measures the overall proportion of correct predictions. Precision indicates the fraction of positive predictions that are actually true. Recall represents the fraction of actual positive samples that are correctly detected. The F1 score is the harmonic mean of precision and recall. These metrics are calculated from the confusion matrix, which also permits direct visualization of true positives, false positives, true negatives, and false negatives.
I compare eight model variants on the EV battery pack datasets: the original SVM, RF, SVM fine-tuned with grid search (SVM-GS), RF with grid search (RF-GS), SVM with Bayesian optimization (SVM-BO), RF with Bayesian optimization (RF-BO), SVM with the proposed GA-BO optimizer (SVM-GA-BO), and RF with GA-BO optimizer (RF-GA-BO). The grid search generally requires an enormous number of evaluations. For the cell under-voltage alarm, grid search spends 1001.64 seconds, whereas GA-BO only takes 121.34 seconds while achieving higher accuracy. Bayesian optimization itself consumes 143.09 seconds, and its diagnostic precision is slightly lower than GA-BO. Thus, GA-BO provides a better trade-off between optimization quality and computation time.
Evaluation on the cell under-voltage alarm dataset
Table 4 summarizes the performance of the three SVM variants and the three RF variants on the cell under-voltage alarm. The RF-GA-BO model achieves perfect scores of 100% for all metrics on both the training set and the test set, demonstrating outstanding diagnostic capability for this EV battery pack fault.
| Model | Data | Accuracy (%) | Precision (%) | Recall (%) | F1 score |
|---|---|---|---|---|---|
| SVM | Train | 88.40 | 68.60 | 62.20 | 0.6546 |
| SVM | Test | 86.80 | 66.80 | 60.00 | 0.6036 |
| SVM-GS | Train | 99.76 | 98.70 | 100.00 | 0.9935 |
| SVM-GS | Test | 99.29 | 99.20 | 97.10 | 0.9798 |
| SVM-BO | Train | 99.72 | 98.50 | 100.00 | 0.9924 |
| SVM-BO | Test | 99.50 | 98.40 | 98.80 | 0.9860 |
| SVM-GA-BO | Train | 99.88 | 99.30 | 100.00 | 0.9965 |
| SVM-GA-BO | Test | 99.71 | 99.20 | 99.20 | 0.9920 |
| RF | Train | 93.00 | 83.00 | 78.10 | 0.8048 |
| RF | Test | 92.15 | 79.20 | 74.00 | 0.7651 |
| RF-GS | Train | 100.00 | 100.00 | 100.00 | 1.0000 |
| RF-GS | Test | 99.86 | 100.00 | 99.30 | 0.9965 |
| RF-BO | Train | 100.00 | 100.00 | 100.00 | 1.0000 |
| RF-BO | Test | 99.93 | 100.00 | 99.60 | 0.9980 |
| RF-GA-BO | Train | 100.00 | 100.00 | 100.00 | 1.0000 |
| RF-GA-BO | Test | 100.00 | 100.00 | 100.00 | 1.0000 |
For the cell-consistency alarm, the RF-GA-BO model again yields the best F1 score of 0.9945 on the test set. The SVM-GA-BO model obtains an F1 score of 0.9844. In addition, RF-GA-BO takes only 98.23 seconds for hyperparameter optimization, whereas RF-GS requires 1454.12 seconds. The RF-GA-BO model is therefore both accurate and computationally efficient.
Selection of preferred models for different fault types
Depending on the fault characteristics and data distribution, either SVM-GA-BO or RF-GA-BO may be more suitable. I compare the two optimized models across the remaining fault types in Tables 5–9. For the SOC-low alarm, SVM-GA-BO has slightly higher metrics than RF-GA-BO; therefore, SVM-GA-BO is preferred. For the SOC-jump alarm, SVM-GA-BO also shows a better F1 score. Conversely, the RF-GA-BO model outperforms SVM-GA-BO for the drive-motor temperature alarm, insulation alarm, drive-motor controller temperature alarm, DC-DC status alarm, and temperature-difference alarm.
| Model | Data | Accuracy (%) | Precision (%) | Recall (%) | F1 score |
|---|---|---|---|---|---|
| SVM-GA-BO | Train | 100.00 | 100.00 | 100.00 | 1.0000 |
| SVM-GA-BO | Test | 99.98 | 100.00 | 99.80 | 0.9980 |
| RF-GA-BO | Train | 100.00 | 100.00 | 100.00 | 1.0000 |
| RF-GA-BO | Test | 99.95 | 99.80 | 99.60 | 0.9970 |
| Model | Data | Accuracy (%) | Precision (%) | Recall (%) | F1 score |
|---|---|---|---|---|---|
| SVM-GA-BO | Train | 93.40 | 92.80 | 95.10 | 0.9394 |
| SVM-GA-BO | Test | 92.50 | 91.20 | 94.20 | 0.9268 |
| RF-GA-BO | Train | 99.82 | 99.80 | 99.80 | 0.9980 |
| RF-GA-BO | Test | 99.52 | 99.00 | 99.20 | 0.9910 |
| Model | Data | Accuracy (%) | Precision (%) | Recall (%) | F1 score |
|---|---|---|---|---|---|
| SVM-GA-BO | Train | 94.55 | 92.80 | 100.00 | 0.9627 |
| SVM-GA-BO | Test | 92.97 | 88.90 | 100.00 | 0.9412 |
| RF-GA-BO | Train | 96.50 | 93.40 | 100.00 | 0.9659 |
| RF-GA-BO | Test | 94.50 | 90.20 | 100.00 | 0.9485 |
| Model | Data | Accuracy (%) | Precision (%) | Recall (%) | F1 score |
|---|---|---|---|---|---|
| SVM-GA-BO | Train | 100.00 | 100.00 | 100.00 | 1.0000 |
| SVM-GA-BO | Test | 99.75 | 99.20 | 100.00 | 0.9960 |
| RF-GA-BO | Train | 99.64 | 98.70 | 99.60 | 0.9915 |
| RF-GA-BO | Test | 99.10 | 96.80 | 99.00 | 0.9789 |
| Model | Data | Accuracy (%) | Precision (%) | Recall (%) | F1 score |
|---|---|---|---|---|---|
| SVM-GA-BO | Train | 100.00 | 100.00 | 100.00 | 1.0000 |
| SVM-GA-BO | Test | 99.10 | 97.00 | 99.20 | 0.9809 |
| RF-GA-BO | Train | 99.64 | 98.70 | 99.60 | 0.9915 |
| RF-GA-BO | Test | 99.75 | 99.00 | 100.00 | 0.9950 |
The above results demonstrate that the proposed data-driven diagnostic framework can reliably detect different faults of the EV battery pack. The choice between SVM and RF depends on the nonlinearity and noise level of the data. In general, random forest performs better for cell-level voltage-related faults because it captures non-monotonic relationships, whereas SVM excels in datasets where the boundary is relatively simple and the sample size is limited.
Abnormal Cell Detection and Severity Evaluation
A critical task in EV battery pack health management is the identification of abnormal single cells from a large string of series-connected cells. Since every cell carries a portion of the total load, a single weak cell may cause the whole EV battery pack to fail. In this part of my study, I first denoise the voltage data using wavelet decomposition. I choose the Daubechies family as the wavelet basis and set the decomposition level to 3. A lower decomposition level fails to separate noise, while a higher level over-smooths the signal and removes meaningful fault features. In my experiments, level 3 gives the best trade-off between denoising performance and computational efficiency.
Comparing the original and wavelet-denoised voltage curves, the filtered curves are significantly smoother, and spike-like disturbances are suppressed. In normal operation, all cells in the EV battery pack show nearly identical voltage trajectories. In abnormal conditions, one or several cells deviate noticeably from the median voltage profile. The denoised signal preserves this deviation, which forms the basis of the subsequent anomaly detection.
Improved Z-score based on Hausdorff distance
Conventional Z-score detects outliers by measuring the number of standard deviations from the mean. However, this method can be sensitive to the distribution shape and may lead to many false alarms when applied to dynamic voltage curves. To improve the robustness for EV battery pack monitoring, I integrate the Hausdorff distance into the Z-score calculation. The Hausdorff distance measures the maximum distance between two sets of points and hence is suitable for quantifying the dissimilarity between each cell’s voltage curve and the median voltage curve.
Let \(A\) be the denoised voltage sequence of a given cell and \(B\) be the median voltage sequence among all cells. The forward and backward directed Hausdorff distances are denoted \(h(A,B)\) and \(h(B,A)\). The Hausdorff distance between the cell and the median curve is defined as
\[
H_i = \max\{ h(A,B), h(B,A) \}.
\]
Then I compute the mean of the Hausdorff distances after removing the maximum contribution so that a potential outlier does not bias the baseline:
\[
\overline{H} = \frac{1}{N-1}\sum_{i=1}^{N} (H_i – \max_{1\le j\le N} H_j).
\]
The standard deviation of the Hausdorff distances is
\[
s_H = \sqrt{\frac{1}{N}\sum_{i=1}^{N} (H_i – \overline{H})^2}.
\]
The improved Z-score for the \(i\)-th cell is thus
\[
Z_i^H = \frac{H_i – \overline{H}}{s_H}.
\]
If \(|Z_i^H|\) remains below the chosen threshold, the cell is considered normal. If \(|Z_i^H|\) exceeds the threshold, the cell is flagged as abnormal. In my validation on the real vehicle data, the conventional Z-score identifies several cells as abnormal due to random disturbances. In contrast, the improved Z-score with Hausdorff distance correctly points to a single cell (Cell 92) in the EV battery pack. The \(Z_i^H\) value of Cell 92 drops sharply after 600 seconds, showing an increasing deviation from the median. The normal-case plot displays no threshold violation for any cell, confirming the reduced false alarm rate.
Combined assignment weighting for severity evaluation
After identifying the abnormal cell, it is necessary to evaluate how severe the anomaly is. For this purpose, I combine the variation coefficient method and the entropy weight method to compute a combined weight for each cell. Let \(V_{i,j}\) be the normalized value of the \(j\)-th indicator at time \(i\). For a dataset with \(m\) cells and \(n\) time samples, the variation coefficient \(V_j\) and entropy weight method values are calculated as follows.
Let \(\bar{V}_j\) be the mean of the \(j\)-th indicator and \(\sigma_j\) its standard deviation. The coefficient of variation is
\[
v_j = \frac{\sigma_j}{\bar{V}_j}.
\]
Then the variation-coefficient weight is obtained by normalization:
\[
W_j^1 = \frac{v_j}{\sum_{j=1}^{m} v_j}.
\]
For the entropy method, I compute the proportion of each cell in the \(j\)-th indicator, \(p_{i,j}\), and the Shannon entropy:
\[
e_j = -\frac{1}{\ln n}\sum_{i=1}^{n} p_{i,j} \ln p_{i,j}.
\]
The redundancy degree is \(d_j = 1-e_j\), and the entropy-based weight is
\[
W_j^2 = \frac{d_j}{\sum_{j=1}^{m} d_j}.
\]
The combined weight of the \(j\)-th indicator is then defined as
\[
W_j^c = \alpha W_j^1 + (1-\alpha) W_j^2,
\]
where \(\alpha\in(0,1)\) is a preference coefficient. In my experiments, I set \(\alpha=0.5\) to balance both weighting schemes.
Using the combined weights, I compute the composite score of each cell in the EV battery pack as
\[
S_i = \sum_{j=1}^{m} W_j^c V_{i,j}.
\]
Let \(\bar{S} = \frac{1}{m}\sum_{i=1}^{m} S_i\). The absolute score difference
\[
\Delta S_i = |S_i – \bar{S}|
\]
measures how far a particular cell deviates from the average behavior. A large \(\Delta S_i\) indicates a high degree of abnormality.
To map the score difference into a fault level, I introduce the abnormal coefficient \(AF\):
\[
AF = \frac{\Delta S – \Delta S_{\text{median}}}{\sigma_{\Delta S}},
\]
where \(\Delta S_{\text{median}}\) is the median of the score differences and \(\sigma_{\Delta S}\) is the standard deviation calculated after excluding abnormal cells. Based on the absolute value of \(AF\), I define three levels as shown in Table 10.
| |AF| range | Fault level | Description |
|---|---|---|
| 0 – 3.5 | Normal | No action required |
| 3.5 – 4.0 | Second-level | Intermittent and potentially ignorable; monitor |
| > 4.0 | First-level | Severe anomaly; immediate action recommended |
Validation of anomaly detection on real vehicle data
I apply the proposed method to a dataset from an EV battery pack containing 96 series-connected cells. The composite score \(S_i\) is the lowest for Cell 92, and its absolute score difference \(\Delta S_i\) is the largest among all cells, reaching a value of 0.544438 on February 19, 2022. The abnormal coefficient \(AF\) for Cell 92 is -5.21521, whose absolute value exceeds 4.0. According to Table 10, this corresponds to a first-level fault. The remaining cells have \(|AF|\) values below 3.5, indicating normal operation.
To verify the consistency of this result, I analyze data from two additional dates. On January 22, 2022, Cell 92 already has the largest \(\Delta S\) among all cells, with an \(|AF|\) value of 3.96389. This places Cell 92 at the boundary between normal and second-level fault. Without intervention, the anomaly becomes more severe, and by March 2, 2022, the \(|AF|\) value exceeds 4.0, while the score difference \(\Delta S\) increases to 0.643147. This multi-date examination confirms that the proposed method tracks the progressive degradation of a single cell in the EV battery pack.
In summary, the combination of the improved Z-score method and the combined-weighting approach provides a reliable tool for both fault localization and severity assessment in EV battery pack monitoring. The introduction of the abnormal coefficient greatly improves the interpretability of the result. For a two-level fault, a driver or fleet manager can pay special attention to the EV battery pack and plan a maintenance action before the fault escalates to a dangerous level.
Conclusion and Future Work
In this research, I propose a complete data-driven framework for EV battery pack fault diagnosis, abnormal cell identification, and severity evaluation. The main contributions are as follows.
First, I design an RF-based feature screening strategy that selects the most influential features for each type of EV battery pack fault. The experimental results show that models trained only on these selected features achieve higher classification accuracy than those using all original features.
Second, I develop a hybrid hyperparameter optimization algorithm that combines genetic algorithms with Bayesian optimization. The GA-BO method is applied to both support vector machines and random forests. Compared to grid search and pure Bayesian optimization, GA-BO delivers better or comparable classification performance with significantly less computation time. By comparing the optimized SVM-GA-BO and RF-GA-BO models on ten fault types, I select the more suitable model for each fault scenario. The confusion-matrix analysis indicates that the GA-BO-enhanced models are particularly effective in detecting anomalous states of the EV battery pack.
Third, I propose an improved Z-score based on Hausdorff distance to identify abnormal single cells in the EV battery pack. This method reduces false positives and clearly pinpoints the problematic cell. I further develop a severity assessment pipeline that combines the variation coefficient method, entropy weighting, and the abnormal coefficient. The proposed method successfully distinguishes between normal cells, second-level abnormal cells, and first-level severe abnormal cells in an EV battery pack. It also tracks the progression of a weak cell over several months.
There are some limitations in this work. First, due to data availability, I only diagnose ten fault types, and several electronic-control faults are not fully explored. Future work will include these fault types by collecting more diverse real-world datasets from monitoring platforms. Second, the current implementation is an offline diagnostic algorithm. For practical deployment, I plan to improve the real-time capability of the proposed framework so that early warnings can be generated online as soon as new telemetry data arrive from the EV battery pack.
In conclusion, the data-driven methods presented in this article provide an effective and economical solution for EV battery pack fault diagnostics. They can assist engineers in detecting failures earlier, reducing repair costs, and ultimately improving the safety of electric vehicles.