Abstract: In view of the defects of inefficient classification and low recall rate caused by the class imbalance of fault data of electric vehicle, an improved Bagging ensemble fault diagnosis method based on LightGBM in the fault classification process for electric vehicle is proposed in this paper. In the Bagging integrated learning model, the training set sampled by Borderline_SMOTE to improve the data imbalance in the training subset and avoid missing information on the faulty class. By embedding weight coefficients and regularization terms into the loss function of LightGBM, the misclassification cost of faulty classes in model training is increased. Through comparative experiments, the results show that the proposed model can effectively improve the recall rate, macro average F₁ measure and AUC on fault diagnosis, in which AUC reaches 0.898 4, and the recall rate is 0.808 3. The fault classification performance of this model on the unbalanced data set of electric vehicles is significantly more effective in classifying faults of electric vehicle than the single model and other comparative algorithms.

Key words: fault diagnosis, LightGBM, Bagging ensemble learning, unbalanced data, Borderline_SMOTE