关键词: 滚动轴承, 故障诊断, 迁移学习, 判别性特征, 最大均值差异

Abstract:  A deep transfer learning method was proposed to address the challenge of inconsistent feature distributions and difficulties in removing noise components in vibration data collected under different operating conditions for rolling bearings. The method utilized a combination of discriminative feature extraction and dual-domain alignment. Firstly， the labeled vibration signals and unlabeled vibration signals were segmented into fixed-length data sets using a data segmentation method. To mitigate the interference of noise signals in practical operating conditions， a channel attention mechanism known as SENet was employed. Additionally， a discriminative loss term was incorporated to assist the feature extractor in extracting features that exhibit discriminative properties. To handle the issue of inconsistent data feature distributions， the MMD was utilized to align the global domain distributions between the source and target domains. Furthermore， conditional adversarial learning techniques were employed to align the sub-domain distributions， resulting in dual-domain alignment. Experimental validation was conducted on two publicly available rolling bearing fault datasets collected under different operating conditions. The results show that the proposed method achieves an average recognition accuracy of over 98%. Comparative analyses with different diagnostic methods further demonstrate the effectiveness and superiority of the proposed method. 

Key words: rolling bearing, fault diagnosis, transfer learning, discriminant feature, maximum mean discrepancy（MMD）