天津科学技术出版社
考虑小样本不平衡的快速可靠充电桩故障预测
陈静绵
作者信息
  1.  中国人民大学 北京 100872

摘要

随着新能源电动汽车的广泛应用,充电桩故障诊断预测对充电桩的快速维修与安全运行有着重要的现实意义。充电桩故障预测中针对表格数据的传统机器学习和深度学习需要大量训练数据,且依赖复杂超参数调优和特征工程,难以适应小样本分类不平衡的设备早期故障预警场景和日新月异的充电桩技术带来的变化。本文提出一种相对可靠、高效、可快速适配不同场景的充电桩故障预测通用模型。通过在小样本分类不平衡的数据集上训练了TabPFN充电桩故障预测模型,其准确率、精确率、f1值和AUC值达到97.8%、99.5%、97.7%、99.6%。同时本研究还融合SHAP技术,对模型进行了全局与局部层面的解释性分析,有效克服了模型在实际应用中的黑箱问题,为充电桩故障诊断研究及运维决策提供新的技术路径和思路。

Fast and Reliable Charging Pile Fault Prediction Considering Small-Sample Imbalance
Chen Jing-mian
Author Information
  1.  Renmin University of China, Beijing 100872, China

Abstract

With the widespread adoption of new energy electric vehicles, fault diagnosis and prediction of charging piles are of great practical significance for rapid maintenance and safe operation. Most traditional machine learning and deep learning methods for tabular data in charging pile fault prediction require large amounts of training data and rely on complex hyperparameter tuning and feature engineering, making it difficult to adapt to early fault warning scenarios with small-sample class imbalance and to keep pace with the evolving charging pile technology. This paper proposes a relatively reliable, efcient, and rapidly adaptable general model for charging pile fault prediction. The TabPFN-based charging pile fault prediction model was trained on a small-sample imbalanced dataset, achieving accuracy, precision, F1-score, and AUC values of 97.8%, 99.5%, 97.7%, and 99.6%, respectively. Additionally, this study integrates SHAP technology to conduct interpretability analysis at both global and local levels, efectively overcoming the black-box problem of the model in practical applications, and providing new technical pathways and insights for charging pile fault diagnosis research and operational decision-making.
收稿日期2025-12-09
修订日期2025-12-19
接受日期2026-01-13