With the rapid proliferation of electric vehicles worldwide, the safety of lithium-ion power batteries has emerged as a pivotal factor constraining the sustainable development of the EV industry. In complex real-world operating environments, EV battery packs are susceptible to various faults, ranging from gradual performance degradation to abrupt and catastrophic thermal runaway events. Thus, precise multi-fault diagnosis and timely risk warning are essential for ensuring safe and reliable operation. Leveraging large-scale real-world data from China’s National New Energy Vehicle Monitoring Platform, I conduct a comprehensive study in three interconnected stages: systematic data preprocessing, multi-fault diagnosis and severity estimation, and machine learning-based long-term risk warning for EV battery packs.
1. Introduction and Research Context
The global energy crisis and environmental pollution have accelerated the adoption of new energy vehicles. The international energy agency data and recent market statistics show that NEV sales have surged dramatically, with China’s NEV market penetration peaking over 50% in 2024. Lithium-ion batteries, prized for their high energy density, form the core power source of these EVs. However, the frequent occurrence of battery failures, particularly thermal runaway events, seriously threatens both property and human safety. My analysis of documented EV fire incidents indicates that a substantial portion — as high as 37% — occurs spontaneously during driving without prior collision, while charging-related incidents constitute about 23%. These statistics underscore the critical importance of developing robust fault diagnosis and early-warning systems tailored for EV battery packs to proactively monitor and mitigate potential hazards, especially those arising from internal battery degradation and spontaneous failures.
2. Battery Faults and Data Fundamentals
An in-depth understanding of fault mechanisms is foundational for developing effective diagnostic and predictive strategies for EV battery packs. The spectrum of possible failure modes is broad, and their impacts vary significantly, which dictates the approach for detection and mitigation.
2.1 Taxonomy of EV Battery Pack Faults
Faults within EV battery packs can be classified into several primary categories as follows: overcharge fault, overdischarge fault, cell inconsistency, internal short circuit (ISC), external short circuit (ESC), and external system-related faults.
Overcharge faults typically arise from voltage exceeding the specified upper limit or overly aggressive charging rates. This stressor leads to electrolyte temperature rise, active material dissolution on the cathode, and potential lithium dendrite formation, thereby accelerating capacity fade and posing a significant thermal hazard. The risk becomes prominent especially when charging beyond 120% SOC. Conversely, overdischarge faults occur when the voltage drops below safe thresholds, causing irreversible damage to active materials, increased internal resistance, and a thickening SEI film. This degradation pathway may ultimately induce internal short circuits. Many BMS designs incorporate SOC estimation and low-voltage protection to mitigate both these fault types.
Cell inconsistency arises from manufacturing variations and disparate operational environments. Inconsistent thermal conditions, varying discharge depths, and uneven aging rates across cells cause performance divergence in terms of voltage, internal resistance, and capacity. The voltage standard deviation across a series string is a valuable metric for quantifying this inconsistency.
External short circuits result from an unintended conductive path between the terminals. This condition triggers an abrupt and excessive current, leading to rapid heat generation and possible electrolyte decomposition. Monitoring voltage and current characteristics allows for the rapid identification of such faults. Internal short circuits, often resulting from manufacturing defects, mechanical abuse, or dendrite growth, are among the most hazardous failure modes. They cause excessive local heat generation and can quickly propagate to a full thermal runaway. External faults, including BMS hardware failures, sensor malfunctions, and contactor issues, can compromise the monitoring and protective functions, indirectly threatening battery safety.
From a temporal evolution perspective, faults in EV battery packs can be classified into abrupt faults and gradual faults. Abrupt faults, such as external short circuits, are triggered within milliseconds and are impossible to predict from early precursor data. Conversely, gradual faults, such as those stemming from micro-internal short circuits or accelerating cell inconsistency, provide a window of opportunity for early detection and intervention. The table below summarizes the fault severity grades and involved alarm types referenced in the study.
| Fault Level | Definition | Typical Faults |
|---|---|---|
| Level 1 | Faults that do not significantly affect vehicle safety but may involve functional issues. | Cell inconsistency, low SOC, SOC jump, large temperature difference |
| Level 2 | Faults that have an impact on vehicle operation but do not require immediate stop. Maintenance is needed soon. | Cell overvoltage, cell undervoltage, storage device undervoltage |
| Level 3 | Critical system faults that may prevent normal operation or pose safety hazards. Immediate action required. | High battery temperature, battery overvoltage |
2.2 Data Acquisition and Systematic Preprocessing
To ensure the authenticity and accuracy of this research, I utilize data from ten specific pure electric vehicles, collected via the national platform. The data was stored at 10-second intervals and encompasses 53 distinct alarm codes, which include single-level and multiple-concurrent fault alarms. The fault data distribution is highly imbalanced: single-fault alarms total 556,968 occurrences, double-fault alarms 46,431, triple-fault alarms 3,164, and quadruple-fault alarms a mere 56 instances. This phenomenon complicates the methodology for designing robust models.
The raw data suffers from various quality issues, including missing values, duplicates, and outliers. Given that defective data can drastically distort the model’s performance, a robust preprocessing regime was established.
Data Screening: For my analysis, it was essential to filter out alarm data unrelated to battery function. Alarms associated with braking system faults, DC-DC converter status faults, high-voltage interlock anomalies, or motor temperature problems were systematically excluded. This step, focusing the dataset on battery-related events, reduced interference and improved the signal quality of subsequent analytical steps.
Duplicate Removal: Transmissions and sensor-read issues lead to exact duplicate data rows. These were identified and removed, and the duplicates were treated as missing values for the subsequent filling steps.
Missing Value Imputation: Different methodologies were adopted. For parameters like which change slowly over time, such as vehicle mode, charging state, and the lists of cell voltages, I employed a forward-filling method. For continuously changing quantities like vehicle speed, max/min temperature, and insulation resistance, the average of values just before and after the missing entry was used for imputation. For data fields with a substantial percentage of missing values (e.g., fault code lists) have low analytical relevance, a deletion approach was adopted. Particularly for parameters with complex, non-linear dependencies, such as the total battery voltage and current, I employed a K-means clustering-based imputation. This method clusters complete data samples, assigns the incomplete ones to the relevant cluster, and fills the missing features using their respective cluster center values, preserving the multidimensional data structure.
Outlier Removal: A dual-pronged approach was used. First, a 3-sigma rule based on normal distribution was applied to detect and remove outlier points that result from sensor failure. Secondly, to capture values that marginally deviate from the norm, a box-plot analysis was conducted, with boundaries defined by Q1 − 1.5 × IQR and Q3 + 1.5 × IQR.
Noise Reduction: To remove high-frequency electromagnetic interference and mechanical noise, I applied a two-stage denoising framework combining trajectory piecewise polynomial regression and an Exponential Weighted Moving Average. This step significantly improved the signal-to-noise ratio. Finally, a Z-score normalization was applied to standardize all input features to ensure consistent performance across the differing scales of voltage, current, and temperature.
3. Multi-Fault Diagnosis and Severity Estimation
My proposed diagnostic strategy is a two-layer approach designed to first locate abnormal cells and subsequently quantify their fault severity. The overall workflow integrates a multi-time-scale feature extractor, an Advanced Local Outlier Factor (ALOF) algorithm, and a Fréchet Distance-based severity estimator for EV battery packs.
3.1 Multi-Dimensional Feature Extraction
Faults in EV battery packs often manifest as anomalies in cell voltages. I propose deep mining of three distinct voltage features. First, the kurtosis of the voltage signal within a long window captures the overall volatility of the voltage. It is a statistical measure reflecting the sharpness of the data distribution. The kurtosis calculation for the c-th cell, \(k_c\), is shown in the following formulas:
$$\mu_c = \frac{\sum_{i=1}^{n_l} v_i}{n_l}$$
$$\sigma_c = \sqrt{\frac{\sum_{i=1}^{n_l}(v_i – \bar{v})^2}{n_l}}$$
$$k_c = \frac{1}{n_l} \sum_{i=1}^{n_l} \left( \frac{v_i – \mu_c}{\sigma_c} \right)^4$$
Here, \(n_l\) is the long-window size, and \(v_i\) represents the voltage at sampling time \(i\). Second, the voltage change rate, \(r_c\), is used to capture short-term voltage transitions or anomalies:
$$r_c = \frac{v_s – v_t}{n_s}$$
where \(n_s\) is the short-window size, \(v_s\) is the voltage at the start of the window, and \(v_t\) is the voltage at the current moment. Third, the instantaneous voltage value at the diagnostic moment itself is used to flag inconsistencies within the pack.
3.2 Advanced Local Outlier Factor (ALOF) for Fault Diagnosis
The traditional Local Outlier Factor (LOF) algorithm is a widely used technique for anomaly detection. The core concept of LOF is to measure the local density deviation of a data point. For a data object \(x\), the k-distance \(d_k(x)\) is defined as the Euclidean distance to its k-th nearest neighbor. However, a critical limitation of the classic LOF is that when multiple abnormal cells are clustered together, their local densities become high, causing the algorithm to misinterpret them as a normal cluster and miss the fault. In my analysis of EV battery packs, a cluster of abnormal cells with similar faults could easily be overlooked.
To resolve this, I improved the LOF algorithm into the Advanced LOF (ALOF) by modifying the fixed parameter k to be dynamic. This innovation allows the neighborhood size to be computed dynamically from the local density of an object, making the algorithm more sensitive to anomaly clusters. The average local density is computed as:
$$\bar{\rho} = \frac{1}{|D|} \sum_{x \in D} \rho(x)$$
From this, the dynamic neighborhood value \(k'(x)\) is derived, where \(\alpha\) is an adjustment parameter controlling the rate of change. The symbol \(\lceil \cdot \rceil\) denotes the ceiling function to ensure an integer value.
$$k'(x) = \left\lceil \alpha \cdot \left( \frac{\bar{\rho}}{\rho(x)} \right) \right\rceil$$
The new neighborhood \(N_{k’}(x)\), augmented to include the point itself to enhance stability, is defined as \(N_{k’}(x) = \{ y \in D \mid d(y, x) \leq d_{k’}(x) \} \cup \{x\}\). Based on this, the local density \(\rho'(x)\) becomes:
$$\rho'(x) = \frac{1}{ \frac{1}{|N_{k’}(x)|} \sum_{y \in N_{k’}(x)} d_{k’}(y, x) }$$
Finally, the ALOF value is calculated as:
$$ALOF(x) = \frac{1}{|N_{k’}(x)|} \sum_{y \in N_{k’}(x)} \frac{\rho'(y)}{\rho'(x)}$$
By computing ALOF values for the feature vectors of kurtosis, voltage change rate, and instantaneous voltage, I obtained three distinct ALOF sequences, \(ALOF_K\), \(ALOF_R\), and \(ALOF_V\). To robustly classify cells, a dynamic threshold is employed that adapts to battery aging and various operating conditions. The threshold formula is given by:
$$T_p = \mu_p + 3\sigma_p$$
Here, \(p\) represents the corresponding feature, \(\mu_p\) is the mean of ALOF values across all cells, and \(\sigma_p\) is their standard deviation. A cell is diagnosed with a specific fault signature if its corresponding ALOF value exceeds the dynamic threshold. This dual mechanism of detection and localization is enabled by the fact that ALOF retains the battery ID ordering, directly flagging which cell is anomalous – a valuable feature for analysts.
3.3 Fréchet Distance for Fault Severity Estimation
After fault localization, the next step is to quantify the severity. I implemented the Fréchet distance, a metric that naturally captures the shape similarity between two curves. This provides a more comprehensive measure compared to point-wise distances, as it considers the order and location of points along the entire curve. Real battery voltage curves are often not a static signal but a time-series waveform that has continuous and dynamic changes. To compute the Fréchet distance, I use a discrete dynamic programming approach.
For an abnormal cell voltage sequence \(U = [u_1, u_2, \ldots, u_n]\) and a normal reference voltage sequence \(V = [v_1, v_2, \ldots, v_n]\), the Euclidean distance between each pair of points builds the distance matrix \(D\), defined as \(D[p][q] = |v_p – u_q|\). The recurrence relation for the Fréchet distance \(F(p, q)\) is given by:
$$F(p, q) = \max \left( \min \left( F(p-1, q), F(p, q-1), F(p-1, q-1) \right), D[p][q] \right)$$
However, to maintain the integral order of both sequences and properly model the coupling effect of the curve shape, a more appropriate formulation is given by:
$$F(p, q) = \max \left( \min \left( F(p-1, q), F(p, q-1) \right), D[p][q] \right)$$
where \(F(0,0) = D[0][0]\). Using dynamic programming ensures that fluctuations in both frequency and amplitude are encapsulated in the final computed value. For the reference, rather than selecting a single normal cell, I calculated a representative normal cell voltage sequence, \(V_{rep}\), which is the mean voltage at each time step across all healthy cells identified by the ALOF step. This reference is more robust and representative of the pack’s state, which improves the estimate’s relevance.
3.4 Experimental Validation of Diagnosis
To validate the performance of the proposed ALOF and Fréchet distance strategy, I designed a series of simulated internal short circuit fault experiments. A string of twelve cells, each with a nominal voltage of 3.75V, were used as the basis. Known shunt resistors were placed across individual cells to simulate various patterns of internal short-circuit faults, creating distinct test cases.
| Case | R1 (Ω) | R2 (Ω) | R3 (Ω) | R4 (Ω) | R5 (Ω) | R6 (Ω) |
|---|---|---|---|---|---|---|
| 1 | 50 | 0 | 0 | 0 | 0 | 200 |
| 2 | 50 | 50 | 50 | 50 | 50 | 0 |
| 3 | 200 | 200 | 200 | 200 | 200 | 50 |
Case 1 – Single Point Fault: When Cell 12 was connected to a 200Ω resistor to model a mild short circuit (with Cell 4 having a 50Ω resistor), the ALOF analysis, as seen in the sequence, clearly flagged cell 12 as anomalous across all three features. The ALOF_R value was particularly high, indicating pronounced voltage change rates. The Fréchet distance for cell 12 was calculated at 0.98, which was significantly higher than the healthy cells, corroborating the voltage deviation.
Case 2 – Multiple Adjacent Faults: This case involved Cells 4 through 8, each shunted with a 50Ω resistor, creating a contiguous cluster of faulty cells. This is a direct test of the LOF limitation. My proposed ALOF algorithm successfully identified all five cells as anomalous. The table shows the ALOF values for each cell. The Fréchet distance for these cells was on the order of 1.4 to 1.5, effectively delineating them as a cluster of faulty cells. This confirmed that ALOF effectively overcomes the multi-anomaly clustering problem.
| Cell ID | ALOF_K | ALOF_R | ALOF_V | Fréchet Distance |
|---|---|---|---|---|
| 1 | 0.749 | 1.901 | 1.050 | 0.143 |
| 2 | 1.666 | 4.096 | 1.224 | 0.249 |
| 3 | 1.416 | 0.041 | 0.733 | 0.032 |
| 4 | 3.265 | 3.425 | 3.399 | 1.423 |
| 5 | 3.967 | 3.608 | 3.693 | 1.515 |
| 6 | 3.864 | 4.582 | 4.130 | 1.372 |
| 7 | 4.279 | 3.584 | 3.617 | 1.534 |
| 8 | 3.312 | 2.570 | 2.368 | 1.415 |
Case 3 – Combined Faults: When both a fault cluster (Cells 4-8 with 200Ω resistors) and a single isolated fault (Cell 12 with 50Ω) were present, the ALOF algorithm was able to simultaneously detect the anomalous cluster and the isolated point. This demonstrates that the algorithm is not biased or saturated by the presence of multiple faults and maintains high sensitivity across the entire pack.
The results from these controlled experiments confirmed the diagnostic capability and severity estimation accuracy of the algorithm. Upon applying this methodology to actual platform data from a faulty vehicle, the diagnostic algorithm successfully flagged Cell 12 for inconsistency and Cells 53, 54, and 55 for voltage fluctuation. The subsequent Fréchet distances calculated for these flagged cells were 1.71, 2.37, 2.11, and 2.25, respectively, providing a clear quantitative metric of their respective fault severities in a real-world operational scenario.
4. Long-term Risk Warning Model based on Machine Learning
Beyond efficient fault diagnosis, early warning of impending faults is crucial for proactive safety management in EV battery packs. I developed a novel risk-forecasting model integrating multi-scale feature extraction and attention mechanisms. The framework is designed to provide an accurate and practical tool for long-term risk prediction.
4.1 Construction of Fault Warning Model
The foundational architecture is a hybrid Deep learning network composed of a Depth-wise Separable Convolution (DSC), Gated Recurrent Unit (GRU) networks, and a SimAM attention module. An overview of the core machinery is as follows:
Multi-Scale Feature Extraction via DSC: Standard Convolution can be computationally expensive and may not optimally capture multi-timescale features. DSC reduces the computational load and decouples temporal and cross-channel correlations. The standard convolution for a multi-channel input is approximated by a depth-wise convolution (applying a single filter per input channel, \(y_c = R_c * x_c + b_c\)) followed by a point-wise convolution (\(1\times1\)), which combines the outputs across channels: \(z_n = \sum_{c=1}^{C} P_n * y_c + b_n\). To capture faults occurring on different timescales, I implemented two parallel DSC branches: one with 3×1 kernels for short-term voltage fluctuations, and another with 7×1 kernels for long-term degradation trends. This dual-branch structure provides a comprehensive feature basis for the model, which is essential for the varying temporal signatures of battery faults.
Sequence Learning via GRU: The DSC-extracted features are fed into parallel GRU layers. GRUs are a lightweight and effective architecture for learning temporal dependencies. The key formulas governing the GRU’s gating mechanism are as follows:
$$r_t = \sigma(W_r \cdot [h_{t-1}, x_t] + b_r)$$
$$z_t = \sigma(W_z \cdot [h_{t-1}, x_t] + b_z)$$
$$\tilde{h}_t = \tanh(W \cdot [r_t \odot h_{t-1}, x_t] + b_h)$$
$$h_t = (1 – z_t) \odot h_{t-1} + z_t \odot \tilde{h}_t$$
Here, \(r_t\) is the reset gate, \(z_t\) is the update gate, \(\sigma\) is the sigmoid function, and \(\odot\) denotes element-wise multiplication. The GRU effectively follows the evolution of the state of the EV battery pack, allowing for an understanding of the gradual progression of symptoms preceding a fault.
Feature Selection via SimAM: The high-dimensional features output from the GRU layers are processed by a SimAM (Simple Attention Module). This module calculates an energy score for each feature. The energy function is defined as:
$$E(X_{hwc}) = \frac{(X_{hwc} – \mu_c)^2}{4 \cdot (\sum_{h,w}(X_{hwc} – \mu_c)^2) + \lambda} + 0.5$$
where \(\mu_c\) is the mean across the spatial dimensions for channel \(c\), and \(\lambda\) is a small constant to prevent division by zero. The final attention weights are obtained through a sigmoid function \(A(X_{hwc}) = \sigma(E(X_{hwc}))\). This mechanism dynamically prioritizes the most salient features related to the forthcoming fault, improving the model’s predictive precision.
Imbalanced Learning with Dynamic Focal Loss: Since fault events are rare, the dataset is extremely imbalanced. To prevent the model from becoming biased toward normal samples, I introduced a multi-label dynamic Focal Loss. This loss function addresses class imbalance by dynamically adjusting parameters. For each class \(t\) in a batch, the dynamic alpha is computed as:
$$\alpha_t = \frac{1 – freq(t)}{\sum_{i=1}^{C} (1 – freq(i))}$$
and the class weight is:
$$w_t = \frac{\frac{1}{freq(t)}}{\sum_{i=1}^{C} \frac{1}{freq(i)}}$$
The resulting loss function is defined as \(FL(p_t) = -w_t \alpha_t (1-p_t)^\gamma \log(p_t)\). This ensures that the model pays more attention to the hard-to-classify and rare fault samples in each patch, which is a key to the algorithm’s overall performance in real-world settings.
4.2 Model Training and Optimization
Training was performed using the AdamW optimizer, chosen for its robust handling of sparse gradients. I incorporated a cosine annealing learning rate scheduler, starting from an initial learning rate of 0.001, to enhance convergence. Hyperparameters, including the number of GRU layers, the number of hidden units, and the dropout rate, were optimized using the HyperBand algorithm. The table below summarizes the final model’s primary structural parameters.
| Layer | Parameter | Value |
|---|---|---|
| DSC_1 (Short) | Channel Conv Kernel | 3×1 |
| GRU_1 | Num Layers / Hidden Units | 5 / 104 |
| DSC_2 (Long) | Channel Conv Kernel | 7×1 |
| GRU_2 | Num Layers / Hidden Units | 6 / 152 |
| Attention | Feature Weighting | SimAM Enabled |
| Output | Units / Activation | 3 / Sigmoid |
| Training | Optimizer / LR / Decay | AdamW / 0.001 / 0.2 |
| Training | Loss / Epochs / Batch | Dyn. Focal Loss / 200 / 64 |
4.3 Model Performance Analysis
The model’s performance was thoroughly evaluated through metrics including accuracy, precision, recall, and F1-score. The table below presents a performance breakdown across the different stages of the model training and usage lifecycle.
| Stage | Average Precision (%) | Average Accuracy (%) | Average Recall (%) | Average F1-Score (%) |
|---|---|---|---|---|
| Pre-training | 81.34 | 81.92 | 82.13 | 81.77 |
| Fine-tuning | 88.29 | 87.97 | 89.37 | 87.50 |
| Validation | 87.15 | 86.76 | 86.95 | 87.57 |
| Testing | 86.06 | 85.59 | 86.18 | 85.36 |
After fine-tuning, the model achieved a test accuracy of 85.59%. The model’s effectiveness was confirmed in its ability to classify the severity levels (1, 2, and 3) of faults with high accuracy, exceeding 85% for levels 1 and 2, and remaining above 80% for level 3. The lower performance on level 3 faults can be attributed to their limited representation in the training dataset.
Ablation Study: To justify my model’s architectural choices, I conducted an ablation study. The results are presented in the table below, showing the contribution of each component to the overall performance.
| Model | Parameters | Avg. Accuracy (%) | Avg. Prediction Time (s) |
|---|---|---|---|
| CNN-GRU-SimAM | 171,366 | 76.47 | 0.37 |
| DSC-LSTM-SimAM | 338,907 | 82.25 | 0.71 |
| DSC-GRU (No Attention) | 254,891 | 82.72 | 0.53 |
| DSC-GRU-SimAM (Proposed) | 254,891 | 85.59 | 0.57 |
The ablation study validates my design choices. The full model achieves the highest accuracy (85.59%), demonstrating the positive contributions of each component. The replacement of standard CNN with DSC significantly improves performance (76.47% to 85.59% for the full model), while the use of GRU over LSTM provides a similar accuracy (82.25% for LSTM vs 85.59% for GRU) with a significantly lower parameter count and faster inference speed. The addition of SimAM enhances the accuracy from 82.72% to 85.59% without adding any parameters, proving its efficiency in feature selection within the EV battery pack context.
5. Integrated Risk Assessment and Warning Results
To translate the model probabilities into an intuitive score for end-users, I designed a comprehensive risk evaluation strategy. For a given time t, the risk score R(t) is calculated as follows:
$$R(t) = \beta_i \times (w_1 \times P_1(t) + w_2 \times P_2(t) + w_3 \times P_3(t))$$
where \(P_1(t)\), \(P_2(t)\), and \(P_3(t)\) are the model’s predicted probabilities for the three fault types. \(w_1\), \(w_2\), and \(w_3\) are the initial weights, and \(\beta_i\) is a correction coefficient for combined faults. For a single fault, the weights are set such that a 90% probability yields an ideal risk score. For combinations, the coefficient \(\beta\) adjusts for the compounded risk, and for recurring faults, dynamic weight adjustment based on cumulative count \(f_i\) is applied:
$$w_i(f_i) = w_i^{initial} \times (1 + \beta \times (1 – e^{-\gamma f_i}))$$
This ensures that recurrent faults escalate the risk score more realistically, reflecting the deteriorating health condition of the EV battery pack.
5.1 Performance across Time Horizons
The predictive capability of the model was tested with different lead times, as summarized in the table below. The model demonstrates high accuracy for immediate forecasts and retains a strong predictive capability even for a 10-hour horizon, making it a valuable tool for proactive maintenance.
| Lead Time (min) | Recognition Rate (%) | Avg. Prediction Offset (min) | Fault Type Recognition (%) |
|---|---|---|---|
| Level 1 / Level 2 / Level 3 | |||
| 60 | 97.4 | 6.2 | 98.24 / 94.65 / 91.34 |
| 300 | 90.7 | 15.5 | 92.92 / 88.78 / 85.51 |
| 600 | 81.1 | 39.6 | 87.54 / 80.06 / 74.85 |
This table offers valuable insights into practical forecasting. For a lead time of 60 minutes, the model achieved an exceptional recognition rate of 97.4% and minimal timing offset. This is within the window of opportunity for preventing thermal runaway, giving the driver ample time to stop the vehicle and isolate the pack. Even at a 10-hour lead time, the model can identify high-risk threats with 81.1% accuracy, a level of performance that is very promising for fleet-level safety management.
5.2 Multi-Vehicle Generalization
To validate the model’s robustness and generalization capability, I applied fine-tuning and evaluation across data from five separate vehicles. The results indicated consistently high-performance across all vehicles. The average last-hour warning accuracy was approximately 90%, while the average 1-hour advance forecast accuracy remained around 80%. The average prediction time offset for the fine-tuned models was 14.66 minutes. This indicates that the base pre-trained model is robust and can be quickly adapted to new vehicles or new battery types with a relatively small amount of data. This performance is crucial for large-scale deployment across diverse fleets of EVs.
6. Conclusion and Future Outlook
In this thesis, I have addressed the critical challenge of enhancing safety for EV battery packs by developing a comprehensive data-driven framework for multi-fault diagnosis and risk early warning.
I first established a robust foundation by systematically preprocessing real-world operational data collected from the national monitoring platform. This process involved the careful screening of alarms, rigorous cleaning of missing data via multiple strategies (mean, forward-filling, clustering), and the removal of outliers, resulting in a high-quality dataset crucial for model fidelity.
For the core diagnostic task, I proposed a two-layer strategy. The first layer utilizes an Advanced Local Outlier Factor (ALOF) method, characterized by a dynamic neighborhood adjustment, to accurately localize faulty cells within EV battery packs, especially those in clustered anomalies. The second layer leverages the Fréchet distance to quantify the severity of the detected anomalies by analyzing the voltage curve divergence. The experimental validation, using simulated short-circuit faults, demonstrated high diagnostic accuracy and the unique ability to detect adjacent faulty cells, a scenario where conventional LOF fails.
For proactive risk management, I developed a DSC-GRU-SimAM model. This model effectively extracts multi-timescale features and learns the temporal dependencies of battery states. Combined with a dynamic Focal Loss, a comprehensive risk assessment strategy was constructed to address class imbalances and reflect the evolution of battery health. The results were compelling: the model achieved a 97.4% accuracy in predicting faults one hour in advance with a 6.2-minute offset, and can provide moderate risk warnings up to 10 hours in advance.
Looking forward, several promising research avenues remain. First, incorporating edge computing to enable real-time, onboard diagnostic and warning is a critical direction. The computational cost of the current model is acceptable for cloud platforms, but implementing it on a BMS requires further optimization. Second, integrating additional external data, such as driving behavior, ambient temperature, and road conditions, could create a more holistic model. Third, exploring more sophisticated domain adaptation or meta-learning techniques may help the model to adapt to the operational conditions of different battery chemistries or topologies with even less fine-tuning, further enhancing the safety and reliability of future EVs.