基于采样机制优化的随机森林充电桩故障诊断方法研究

Research on Fault Diagnosis Method of DC Charging Piles Based on Optimized Random Forest by Sampling Mechanism

摘要:

针对直流充电桩多类别故障数据不均衡、运行数据噪声高以及响应慢的问题，采用一种引入元代价敏感学习机制的三重伯努利改进随机森林方法。该方法在特征子空间选择、特征内软采样及分裂准则切换三个阶段嵌入受控随机性，并在特征选择与节点分裂中通过权重矩阵引入类别代价信息，显著增强对不平衡小样本类别的识别能力，同时保持模型多样性与泛化性能。在包含19类故障、9项运行特征的真实充电桩数据集上进行试验测试，在测试集准确率（96.76%）和F1分数（0.948）方面均优于C4.5-RF（66.01%、0.605）和CART-RF（86.33%、0.834），训练时间仅54.93s，较C4.5-RF缩短约92.8%。经消融试验证实，三重伯努利机制与代价敏感信息的结合实现了充电桩故障诊断准确率的提高和计算成本的降低。

Abstract:

To address the challenges of imbalanced multi-class fault data, high noise levels, and slow diagnostic response in DC charging stations, this paper proposes an improved Random Forest method, termed Sandwich-RF, which integrates a Triple-Bernoulli mechanism with a meta cost-sensitive learning strategy. The proposed method embeds controlled randomness into three stages of decision tree construction: feature subspace selection, intra-feature soft sampling, and splitting criterion switching. Furthermore, category cost information is incorporated into both feature selection and node splitting through a dynamic weight matrix, significantly enhancing the recognition capability for minority and high-cost classes while maintaining model diversity and generalization performance. Experiments conducted on a real-world dataset containing 19 fault categories and 9 operational features show that Sandwich-RF achieves a test accuracy of 96.76% and an F1 score of 0.948, outperforming C4.5-RF (66.01%, 0.605) and CART-RF (86.33%, 0.834), with a training time of only 54.93s—approximately 92.8% faster than C4.5-RF. Ablation studies confirm that the combination of the Triple-Bernoulli mechanism and cost-sensitive learning effectively improves accuracy while reducing computational cost, demonstrating strong potential for engineering applications in online fault diagnosis of DC charging stations.