基于集成学习与数据驱动的电动汽车动力电池多维度故障预警

Data-driven fault early warning of electric vehicle power battery based on ensemble learning from multiple dimensions

针对汽车动力电池风险预测中单一算法在非均衡数据集上性能欠佳,以及分析维度不足等问题,提出一种建立在多维特征之上的数据驱动多模型集成学习方法以实现电池故障预警。根据内外部因素从电池状态、行驶状况、历史信息和时间环境多方面提取特征,以还原真实应用场景;通过FilterWrapper法剔除冗余无效信息来进行特征筛选,以提高鲁棒性;在Bayes优化结合网格搜索法完成超参数调优的基础上,构建了以SVM、LightGBM和XGBoost为初级学习器、LR为次级学习器的异质集成Stacking模型。结果表明:考虑了外部特征后的融合模型在评估指标上综合表现最好,其中在中、高等级的故障预警上召回率分别达到了85.81%和88.92%,相较传统仅考虑电池内部特征的单一模型有着更好的预测精度和泛化能力。

Aiming at the problems such as poor performance of single algorithm on unbalanced data set and insufficient analysis dimension in electric vehicle power battery risk prediction,a data-driven multi-model ensemble learning method based on multi-dimensional characteristics was proposed to realize battery fault warning.According to internal and external factors,the features from battery information,driving condition,historical situation and time environment was extracted to simulate real application scenarios.To complete feature filtering,redundant information was removed by Filter-Wrapper method,which improved the robustness.The heterogeneous Stacking integrated model was constructed with SVM,LightGBM and XGBoost as primary learners and LR as secondary learners based on Bayesian optimization with grid search to optimize hyperparameters. The results show that the ensemble model with the addition of external features has the best comprehensive performance in the evaluation index, among which the recall rate in the medium and high level fault warning reaches 85.81% and 88.92%, respectively. Therefore, the method has better prediction accuracy and generalization ability than the single model that only considers the internal characteristics of the battery.