In recent years, the rapid growth of electric vehicles (EVs) in China has highlighted the critical importance of battery safety and reliability. As a core component of EVs, the power battery system accounts for approximately 60% of vehicle failures, with issues such as thermal runaway posing significant risks. Traditional fault diagnosis methods often struggle to adapt to real-world driving conditions, leading to increased demand for data-driven approaches. This paper addresses these challenges by proposing a novel fault prediction algorithm based on Bayesian optimization of Long Short-Term Memory (LSTM) networks, specifically tailored for China EV battery systems. By leveraging real vehicle data, we aim to enhance the accuracy and robustness of fault prediction while minimizing false alarms.
The proliferation of lithium-ion batteries in China’s EV industry has necessitated advanced monitoring systems. However, existing methods, including knowledge-based and model-based approaches, face limitations in handling complex real-world scenarios. Data-driven techniques, particularly those utilizing deep learning, offer promising solutions by learning from vast datasets collected from national EV monitoring platforms. Our work focuses on optimizing LSTM networks using Bayesian methods to predict battery voltage anomalies, which are key indicators of potential faults. This approach not only improves prediction precision but also reduces computational costs by modeling a representative cell voltage instead of the entire battery pack.
In this study, we first discuss data preprocessing steps to handle the complexities of real vehicle data, including noise reduction and feature selection. We then detail the integration of Bayesian optimization with LSTM to fine-tune hyperparameters, ensuring optimal performance. Experimental results on multiple EV datasets demonstrate the superiority of our method in terms of reliability and robustness. The following sections provide a comprehensive overview of our methodology, experimental validation, and future directions for improving China EV battery safety.
Data Preprocessing for EV Power Battery Systems
Real vehicle data from China’s EV monitoring platforms often contain inconsistencies due to transmission errors, environmental factors, and varying driving behaviors. To ensure data quality, we implemented a multi-step preprocessing pipeline. Initially, raw data underwent cleaning to remove duplicates, missing values, and outliers using statistical methods like box plots. The data were then sorted chronologically to maintain temporal integrity for time-series analysis. Given the high dimensionality of the dataset, feature selection was crucial to reduce computational overhead and prevent overfitting.
We employed Pearson correlation coefficients to identify the most relevant features for predicting cell voltage. The correlation between variables was calculated using the formula:
$$ r_{x,y} = \frac{\sum_{i=1}^{n} (x_i – \bar{x})(y_i – \bar{y})}{\sqrt{\sum_{i=1}^{n} (x_i – \bar{x})^2 \sum_{i=1}^{n} (y_i – \bar{y})^2}} $$
where $r_{x,y}$ represents the correlation coefficient, $x_i$ and $y_i$ are data points, $\bar{x}$ and $\bar{y}$ are means, and $n$ is the sample size. Features with strong correlations (absolute value above 0.8) were retained. For instance, total voltage and state of charge (SOC) showed high correlation with cell voltage, as summarized in Table 1.
| Feature | Correlation with Cell Voltage | Significance Level |
|---|---|---|
| Total Voltage | 0.92 | Strong |
| SOC | 0.91 | Strong |
| Total Current | 0.45 | Weak |
| Insulation Resistance | 0.32 | Weak |
Based on this analysis, we selected total voltage, SOC, and cell voltage as input features for the LSTM model. To address the computational challenges of processing entire battery packs, we introduced a cell selection strategy. Two methods were tested: using the average voltage of all cells per time frame and using the median voltage. Experimental comparisons revealed that the median voltage provided more stable predictions, as it is less affected by outliers. Thus, we defined a representative cell voltage as the median of all cell voltages at each time step, which served as the basis for model training and prediction.
Methodology: Bayesian-Optimized LSTM for EV Power Battery Fault Prediction
Long Short-Term Memory (LSTM) networks are well-suited for time-series data due to their ability to capture long-term dependencies. However, LSTM performance heavily depends on hyperparameter settings, which are often tuned manually, leading to suboptimal results. Our approach integrates Bayesian optimization to automate this process, enhancing the model’s accuracy and efficiency. The LSTM architecture includes input, forget, and output gates that regulate information flow, as described by the following equations:
Forget gate: $f_t = \sigma(W_f \cdot [h_{t-1}, x_t] + b_f)$
Input gate: $i_t = \sigma(W_i \cdot [h_{t-1}, x_t] + b_i)$
Cell state update: $\tilde{C}_t = \tanh(W_C \cdot [h_{t-1}, x_t] + b_C)$
Output gate: $o_t = \sigma(W_o \cdot [h_{t-1}, x_t] + b_o)$
Hidden state: $h_t = o_t \cdot \tanh(C_t)$
where $\sigma$ is the sigmoid function, $W$ denotes weight matrices, $b$ represents bias terms, $h_t$ is the hidden state, and $C_t$ is the cell state at time $t$.
Bayesian optimization employs a probabilistic model, typically Gaussian processes, to approximate the objective function and guide the search for optimal hyperparameters. The acquisition function, Expected Improvement (EI), balances exploration and exploitation by evaluating potential improvements over the current best solution. The EI function is defined as:
$$ \text{EI}(x) = \begin{cases}
(\mu(x) – f(x^*)) \Phi(z) + \sigma(x) \phi(z) & \text{if } \sigma(x) > 0 \\
0 & \text{if } \sigma(x) = 0
\end{cases} $$
where $z = \frac{\mu(x) – f(x^*)}{\sigma(x)}$, $\mu(x)$ is the mean, $\sigma(x)$ is the standard deviation, $f(x^*)$ is the best-observed value, and $\Phi$ and $\phi$ are the cumulative distribution and probability density functions of the standard normal distribution, respectively.
We applied Bayesian optimization to tune key LSTM hyperparameters, such as dropout rate and L2 regularization coefficient, within predefined ranges. This process minimized the root mean square error (RMSE) on validation data, ensuring robust performance. The optimized hyperparameters were then used to train the LSTM model for predicting the representative cell voltage, which generalized to the entire battery pack. This method significantly reduced training time and computational resources compared to modeling all cells individually.
Experimental Validation and Results
We evaluated our algorithm using real vehicle data from multiple EVs in China, including both ternary lithium and lithium iron phosphate batteries. The datasets covered diverse driving conditions and time periods, with training data from 2022 and testing data from 2023. Performance metrics included RMSE, mean absolute error (MAE), mean relative error (MRE), and mean absolute percentage error (MAPE), calculated as:
RMSE: $E_{\text{RMS}} = \sqrt{\frac{\sum_{i=1}^{n} (T_{\text{predict}} – T_{\text{real}})^2}{n}}$
MAE: $E_{\text{MA}} = \frac{1}{n} \sum_{i=1}^{n} |T_{\text{predict}} – T_{\text{real}}|$
MRE: $E_{\text{MR}} = \frac{1}{n} \sum_{i=1}^{n} \frac{|T_{\text{predict}} – T_{\text{real}}|}{T_{\text{real}}}$
MAPE: $E_{\text{MAP}} = \frac{1}{n} \sum_{i=1}^{n} \left| \frac{T_{\text{predict}} – T_{\text{real}}}{T_{\text{real}}} \right| \times 100\%$
Comparative analyses between the proposed Bayesian-optimized LSTM (BO-LSTM) model and standard LSTM models demonstrated significant improvements. For instance, the BO-LSTM model achieved a 61.59% reduction in RMSE, 61.31% in MAE, and 60.94% in MRE compared to the whole-battery voltage prediction model. These results underscore the efficacy of our approach in enhancing prediction accuracy for China EV battery systems.
| Model | RMSE | MAE | MRE | MAPE (%) |
|---|---|---|---|---|
| BO-LSTM (Cell-Based) | 0.00397 | 0.00233 | 0.00070 | 0.0698 |
| LSTM (Whole-Battery) | 0.01033 | 0.00602 | 0.00179 | 0.1787 |
To ensure practical applicability, we established fault alert thresholds based on prediction errors. Alerts were categorized into three levels: Level 1 for minor deviations (0.12 V to 0.24 V), Level 2 for moderate deviations (0.24 V to 0.36 V), and Level 3 for severe deviations (above 0.36 V), indicating potential thermal runaway. Testing on faulty vehicles confirmed that our model could detect anomalies up to 22 hours in advance, providing critical lead time for preventive measures. Moreover, the model maintained high reliability by minimizing false alarms in normal operating conditions, as validated with data from multiple EVs.
Conclusion and Future Directions
This study presents a data-driven fault prediction algorithm for China EV power batteries, combining Bayesian optimization with LSTM networks to achieve high accuracy and robustness. By focusing on a representative cell voltage, we streamlined the modeling process while maintaining predictive performance. Experimental results on real vehicle data confirm the algorithm’s superiority in reducing errors and enhancing fault detection capabilities. The integration of Bayesian optimization effectively addresses hyperparameter tuning challenges, making the approach suitable for large-scale EV applications.
Looking ahead, several areas warrant further investigation. First, as China’s EV fleet expands, optimizing the balance between data volume and model training efficiency will be crucial. Future work could explore incremental learning techniques to adapt models without full retraining. Second, developing universal models that generalize across similar EV types could reduce deployment costs and improve scalability. Finally, incorporating additional sensor data, such as temperature and current profiles, may further refine fault prediction accuracy. By advancing these aspects, we can contribute to safer and more reliable China EV battery systems, supporting the sustainable growth of electric mobility.
In summary, our research underscores the potential of data-driven methods in addressing the complex challenges of EV power battery management. The proposed algorithm not only improves fault prediction but also lays the groundwork for future innovations in battery safety and performance optimization.