关键词: 故障诊断, 迁移学习, 开集, 自适应训练

Abstract: To address performance degradation in cross-domain fault diagnosis caused by significant feature distribution shifts and class distribution discrepancies between source and target domains,a multi-source weighted transfer learning method was proposed.A fault type perception strategy based on classifiers was designed,which reduced feature distribution alignment between the target domain and the source domain by identifying unique classes in the target domain during domain adaptation transfer learning.The complementary classifiers generated from multiple source domains based on the similarity scores of target domain data in the source domain was weighted,combining these weights to fuse diagnostic decisions from multiple source domains for higher diagnostic accuracy.Two experimental cases were used to verify the feasibility and compare the proposed method.The results showed that the method achieved higher diagnostic accuracy and stronger generalization performance in cross-condition and cross-machine fault diagnosis scenarios for rolling bearing.

Key words: fault diagnosis, transfer learning, open-set, adversarial training