<p>Aiming at the problems of strong dependence on artificial features, unbalanced samples and insufficient generalization ability in traction motor fault recognition, a fault recognition model combining convolutional neural network and Bayesian optimization (BO)-XGBoost is proposed. The traditional method relies on manual extraction of features, which is inefficient and affected by operators. The existing deep learning methods are insufficient to deal with unbalanced samples. The CNN-BO-XGBoost model is constructed by extracting time–frequency domain and fusion features from dual-branch CNN and identifying faults. The dual-branch CNN is pre-trained as a feature extractor. Combined with XGBoost to deal with data imbalance, BO is used to optimize its hyperparameters. Combined with various technical advantages, the processing ability of unbalanced samples and the accuracy of fault identification are improved, the accuracy of failure mode identification is improved, the high reliability of traction motor is realized, and the cost is reduced. The experimental results show that the model has an accuracy rate of 98.5% on six types of fault data sets, which is 4.45% higher than that of a single CNN model,especially for minority-class faults (defined as those with total sample size ≤ 200, rotor bar, air-gap eccentricity, and end ring faults). The recognition rate on the independent test set for the air-gap eccentricity fault is increased by 7.5% (from 87.5% to 95.0%), which effectively solves the problem of data imbalance.</p>