电工技术学报2026，Vol.41Issue(2)：512-526,15.DOI:10.19595/j.cnki.1000-6753.tces.250208
基于多尺度特征融合卷积神经网络的牵引电机转子断条故障诊断方法
The Fault Diagnosis Method of Traction Motor Broken Rotor Bar Based on Multi-Scale Feature Fusion Convolutional Neural Networks
摘要
Abstract
A traction motor is a key component of the traction transmission system in high-speed trains,which converts electrical energy into mechanical energy and provides power for the train.Accurate diagnosis of a broken rotor bar fault in a traction motor is crucial for the safe operation of high-speed trains,and it is also a key aspect of fault prognostics and health management(PHM).When a broken rotor bar fault occurs in the traction motor,the fault frequency is close to the power supply frequency,and the amplitude is small,making it easy to be masked.The fault frequency varies widely,resulting in significant changes in feature scale.Extracting practical features and obtaining accurate diagnosis results are challenging.This paper proposes a fault diagnosis method based on multi-scale feature fusion convolutional neural networks(MSFFCNN). To eliminate the power frequency component in the current signal and enhance the fault features,a current preprocessing method based on Hilbert transform(HT)is proposed.Firstly,the envelope analysis is used to eliminate the power supply frequency and enhance the fault features.Then,the obtained envelope signal is converted into an image.A multi-scale feature extraction module with attention fusion is constructed.Multiple convolution kernels are used to extract features simultaneously.Efficient channel attention(ECA)is employed for the weighted fusion of multi-scale features to enhance relevant features and suppress irrelevant features.Then,the MSFFCNN is designed to identify broken rotor bar faults and their corresponding fault degrees.Experiments were conducted on two datasets of broken rotor bars,and ablation and comparison experiments were designed to verify the proposed method. The results of the ablation experiment show that,compared with the pretreatment without HT,the accuracy of the proposed method increases by 13.37％,1.40％,and 0.80％when the training ratio is 20％,40％,and 60％,respectively.Compared to the case without the ECA mechanism,the accuracy of the proposed method increases by 0.69％,0.62％,and 0.15％when the training ratio is 20％,40％,and 60％,respectively.The proposed method achieves a higher average diagnostic accuracy and F1 score on both data sets than the comparison method at all training set ratios.When the proportion of the training set is 60％,the average diagnostic accuracy of the proposed method on the two data sets reaches 99.85％ and 99.82％.The visualization results show that the feature boundaries of different fault categories extracted by the proposed method are clear,which can effectively distinguish the broken rotor bar fault under various loads and power supply frequencies. The following conclusions can be drawn.(1)HT is used for current preprocessing to eliminate the influence of the power frequency component,and the broken rotor bar fault features are enhanced.The generated images contain more detailed information,making it easier to extract practical fault features and improve diagnostic accuracy.(2)ECA fuses multi-scale features to automatically realize effective feature extraction and avoid overfitting,enabling the model to adapt to different loads and power supply frequencies.As a result,the diagnostic accuracy and generalization performance of the model are improved.(3)Compared with the related methods,the proposed diagnostic method shows strong feature extraction ability,noise resistance,and generalization performance.It can identify broken rotor bar faults more accurately,providing a reference for the targeted setting of traction motor maintenance plans.关键词
希尔伯特变换/多尺度卷积/注意力机制/故障诊断/转子断条Key words
Hilbert transformation/multi-scale convolutional/attention mechanism/fault diagnosis/broken rotor bar分类
信息技术与安全科学引用本文复制引用
丁卓,张和生,汤昳琮,洪剑锋..基于多尺度特征融合卷积神经网络的牵引电机转子断条故障诊断方法[J].电工技术学报,2026,41(2):512-526,15.基金项目
北京市自然科学基金项目(L191008)、中国国家铁路集团有限公司系统性重大项目(P2018J001)和北京交通大学研究生专业核心课程建设项目(YJSSQ20230328)资助. （L191008）
