With the rapid growth of the electric vehicle industry, charging infrastructure has become a critical component for sustainable transportation. In China, the expansion of the electric vehicle market, often referred to as China EV, has highlighted the importance of reliable charging piles. However, these charging piles are prone to frequent failures, which can lead to safety hazards and operational inefficiencies. Traditional fault prediction methods, such as manual inspections and basic analytical models, often fall short in accuracy and scalability, especially as the complexity of electric vehicle charging systems increases. To address this, I propose a novel fault prediction approach that combines a cooperative game strategy for feature weighting with an optimized deep learning model, specifically the Dung Beetle Optimizer-Bidirectional Long Short-Term Memory-Attention (DBO-BiLSTM-Attention) network. This method aims to enhance prediction accuracy by emphasizing key features and optimizing model parameters, ultimately supporting the reliability of electric vehicle infrastructure in China and beyond.
The increasing adoption of electric vehicles worldwide, particularly in regions like China EV markets, has driven demand for robust charging solutions. Electric vehicle charging piles are subject to various operational stresses, leading to faults that can disrupt services and pose risks. Existing fault prediction techniques, including wavelet analysis, principal component analysis, and standard machine learning models, often struggle with data quality issues and parameter sensitivity. For instance, methods like random forests or support vector machines may not fully capture temporal dependencies in sensor data, while models without feature amplification can overlook critical indicators. My approach integrates multiple perspectives—subjective, objective, and machine learning-based—to assign feature weights through a cooperative game strategy, followed by a DBO-optimized BiLSTM network with attention mechanisms to handle long-sequence data effectively. This comprehensive framework not only improves prediction performance but also addresses the unique challenges of electric vehicle charging pile monitoring.
In the context of electric vehicle infrastructure, data preprocessing is crucial for accurate fault prediction. The dataset used in this study includes parameters such as K1K2 drive signals, electronic lock drive signals, emergency stop signals, access control signals, total harmonic distortion of voltage (THDV-M), and total harmonic distortion of current (THDI-M), which are common in electric vehicle charging systems. To ensure data integrity, I first handle missing values using mean imputation, as described by the formula: $$\hat{X} = \frac{\sum_{i=1}^{n} \gamma_i X_i}{n},$$ where $\hat{X}$ is the imputed value, $X_i$ represents the i-th feature data, $n$ is the number of features, and $\gamma_i$ is an indicator (1 for missing, 0 otherwise). For outlier detection, I analyze the statistical distribution of each parameter; for example, THDV-M values beyond 4000 are identified as anomalies and removed. Subsequently, data normalization is applied to scale features between 0 and 1 using: $$X_1 = \frac{X – X_{\text{min}}}{X_{\text{max}} – X_{\text{min}}} \times (\text{max} – \text{min}) + \text{min},$$ where $\text{max} = 0.9$ and $\text{min} = 0.1$. This step ensures that all parameters, such as those from electric vehicle charging piles, are comparable and suitable for model training.
The cooperative game strategy is employed to determine optimal feature weights, which amplifies the influence of critical parameters in fault prediction for electric vehicle charging piles. I utilize three distinct weighting methods: the Analytic Hierarchy Process (AHP) for subjective evaluation, the CRITIC weight method for objective assessment, and Random Forest from machine learning for data-driven insights. Each method calculates weights for the six key features, and their results are combined using a cooperative game approach to minimize biases. The correlation coefficient $L(i)$ between weights from method $i$ and the combined weights from other methods is computed as: $$L(i) = \frac{\sum_{j=1}^{n} [W_j(i) – \bar{W}(i)] [W_j(m_i) – \bar{W}(m_i)]}{\sqrt{\sum_{j=1}^{n} [W_j(i) – \bar{W}(i)]^2 \sum_{j=1}^{n} [W_j(m_i) – \bar{W}(m_i)]^2}},$$ where $W_j(i)$ is the weight of the j-th feature from method $i$, $\bar{W}(i)$ is the average weight, and $W_j(m_i)$ represents the combined weights from other methods. The overall combined weight $W$ is then derived as: $$W = \begin{cases} \sum_{i=1}^{n} W(i) L(i) & \text{if } i \geq 3, \\ \frac{W(1) + W(2)}{2} & \text{if } i = 2, \end{cases}$$ and normalized to produce the final weights. This process ensures that features like THDI-M and THDV-M, which are vital in electric vehicle charging systems, receive appropriate emphasis.
| Feature | CRITIC (M1) | AHP (M2) | Random Forest (M3) | Cooperative Game (M4) | Full Cooperative Game (M5) |
|---|---|---|---|---|---|
| K1K2 Drive Signal | 0.015 | 2.620 | 5.700 | 1.317 | 2.528 |
| Electronic Lock Drive Signal | 0.014 | 7.655 | 5.900 | 3.834 | 4.671 |
| Emergency Stop Signal | 0.014 | 16.947 | 5.900 | 8.480 | 8.535 |
| Access Control Signal | 0.012 | 6.003 | 6.100 | 3.007 | 4.034 |
| THDI-M | 42.73 | 33.388 | 70.500 | 38.059 | 45.840 |
| THDV-M | 57.216 | 33.388 | 5.800 | 45.302 | 34.392 |
After obtaining the combined weights, the feature matrix is amplified to highlight important parameters. For instance, the original normalized data for K1K2 drive signal (0.686) is multiplied by its weight from M5 (2.528) to yield 1.733, as shown in the amplified feature matrix. This amplification enhances the model’s ability to detect faults in electric vehicle charging piles by focusing on dominant features like THDI-M, which is crucial in China EV applications where power quality issues are common.
| Feature | Original Data | Normalized Data | Amplified (M4) | Amplified (M5) |
|---|---|---|---|---|
| K1K2 Drive Signal | 11.848 | 0.686 | 0.903 | 1.733 |
| Electronic Lock Drive Signal | 12.124 | 0.515 | 1.975 | 2.407 |
| Emergency Stop Signal | -0.068 | 0.544 | 4.617 | 4.646 |
| Access Control Signal | 12.058 | 0.500 | 1.504 | 2.017 |
| THDI-M | 12.604 | 0.104 | 3.965 | 4.776 |
| THDV-M | 43.381 | 0.111 | 5.022 | 3.813 |
The DBO-BiLSTM-Attention model is designed to leverage the amplified features for fault prediction in electric vehicle charging piles. The BiLSTM network processes sequences in both forward and backward directions, capturing temporal dependencies that are essential for monitoring charging pile operations. The forward LSTM hidden state $A_i$ and backward LSTM hidden state $B_i$ are updated as: $$A_i = f_1(\omega_1 x_i + \omega_2 A_{i-1}),$$ $$B_i = f_2(\omega_3 x_i + \omega_5 B_{i+1}),$$ where $f_1$ and $f_2$ are activation functions, $\omega$ represents weights, and $x_i$ is the input at time step $i$. The final output $Y_i$ is computed by combining these states: $$Y_i = f_3(\omega_4 A_i + \omega_6 B_i).$$ To address the limitation of long sequences, an attention mechanism is incorporated, which assigns weights to hidden states based on their importance. The attention weight $a_{ti}$ for hidden state $h_i$ at time $t$ is calculated as: $$a_{ti} = \frac{\exp(S_{ti})}{\sum_{k=1}^{t} \exp(S_{tk})},$$ where $S_{ti} = V \tanh(W h_i + U h_{t-1} + b)$, and $V$, $W$, $U$, $b$ are learnable parameters. The context vector $F_t$ is then: $$F_t = \sum_{i=1}^{t} a_{ti} h_i,$$ and the final feature vector $h’_t$ is derived through a function $f$ incorporating $F_t$, previous hidden state $h_{t-1}$, and input $y_t$. This allows the model to focus on critical time steps, improving fault detection for electric vehicle charging systems.
Furthermore, the Dung Beetle Optimizer (DBO) is applied to fine-tune hyperparameters of the BiLSTM network, such as initial learning rate, number of hidden neurons, and L2 regularization coefficient. DBO mimics the behavior of dung beetles, including rolling, dancing, breeding, foraging, and stealing, to efficiently search the parameter space. The algorithm initializes a population, evaluates fitness, updates positions based on boundary constraints, and iterates until convergence. This optimization enhances the model’s performance by avoiding local minima and accelerating convergence, which is particularly beneficial for the complex data patterns in electric vehicle charging pile faults.
For experimental evaluation, I use a dataset from a public competition on electric vehicle charging pile fault classification, which includes samples from various charging piles in China EV networks. The data is split into training and testing sets, and performance is measured using accuracy and F1-score, defined as: $$\text{Accuracy} = \frac{TP + TN}{TP + FN + FP + TN},$$ $$\text{F1-score} = \frac{2 \times \text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}},$$ where precision $= \frac{TP}{TP + FP}$ and recall $= \frac{TP}{TP + FN}$. Several models are compared, including basic classifiers like Naive Bayes and Support Vector Machines, as well as variants of the proposed approach. The results demonstrate that the full model (Model 8) with cooperative game-based feature amplification and DBO-BiLSTM-Attention achieves the highest accuracy and F1-score on both training and testing sets, underscoring its effectiveness for electric vehicle charging pile fault prediction.
| Model | Training Accuracy | Training F1-score | Testing Accuracy | Testing F1-score |
|---|---|---|---|---|
| Model 1: Naive Bayes | 0.70 | 0.68 | 0.68 | 0.66 |
| Model 2: SVM | 0.66 | 0.63 | 0.67 | 0.63 |
| Model 3: DBO-BiLSTM-Attention (No Amplification) | 0.85 | 0.85 | 0.85 | 0.84 |
| Model 4: BiLSTM (With M5 Amplification) | 0.85 | 0.84 | 0.83 | 0.82 |
| Model 5: DBO-BiLSTM (With M5 Amplification) | 0.88 | 0.88 | 0.87 | 0.87 |
| Model 6: BiLSTM-Attention (With M5 Amplification) | 0.87 | 0.87 | 0.87 | 0.87 |
| Model 7: DBO-BiLSTM-Attention (With M4 Amplification) | 0.87 | 0.87 | 0.88 | 0.87 |
| Model 8: Proposed Full Model | 0.89 | 0.89 | 0.90 | 0.90 |
The analysis shows that feature amplification through the cooperative game strategy significantly improves model performance, with Model 8 achieving a 5% increase in testing accuracy and a 6% increase in F1-score compared to Model 3 without amplification. Additionally, incorporating Random Forest into the cooperative game (M5) further enhances results over M4, validating the importance of integrating machine learning insights. The attention mechanism and DBO optimization also contribute to better handling of sequential data and parameter tuning, respectively, making the model well-suited for the dynamic environments of electric vehicle charging piles. In China EV contexts, where charging infrastructure is rapidly expanding, this approach offers a scalable solution for proactive maintenance and fault prevention.
In conclusion, the integration of cooperative game strategy and DBO-BiLSTM-Attention provides a robust framework for fault prediction in electric vehicle charging piles. By combining multiple weighting methods and optimizing deep learning parameters, the model achieves high accuracy and reliability, addressing key challenges in the electric vehicle industry. Future work will focus on improving computational efficiency and real-time applicability, particularly for large-scale deployments in China EV networks. As the adoption of electric vehicles continues to grow, such advanced prediction methods will play a vital role in ensuring the safety and sustainability of charging infrastructure.