초록

Transfer learning demonstrates strong capability in transferring knowledge for fault diagnosis across various working conditions or systems. However, in practical situations, multiple source domains of various working conditions or systems are available, and relying solely on a single source for transfer may yield sub-optimal outcomes. To solve the problem, multi-source domain adaptation based on dual-level weighting is proposed for fault diagnosis of rotating machinery. Dual-level weighting consists of sample-level and domain-level weights. First, sample-level weights aregenerated by assessing the envelope spectrum energy at the bearing characteristic frequency for each health state. Second, domain-level weights are calculated by the relative magnitude of the sample-level weight across each domain. A statistical distance, based on dual-level weights, is established to quantify the distance between each soruce and target domain. Minimizing this gap enables the transfer of valuable information from each source domain to the target, thereby enhancing the generalization performance of the target domain. Bearing datasets obtained from various systems are utilized to validate the efficacy of the proposed method. The results show that the proposed method outperforms the existing multi-source domain adaptation-based fault diagnosis methods.