Weighted Local Adversarial Domain Adaptation With Wide Kernels for Bearing Fault Diagnosis Under Multiple Conditions

Domain adaptation is the mainstream method for bearing fault diagnosis under multiple conditions. The feature alignment between the source domain and the target domain faces three challenges. First, due to the noises, features cannot be aligned at the global level; second, due to the fine-grained differences in subdomains, features cannot be aligned at the local level; third, due to the high diagnostic difficulty of individual categories, the features cannot be aligned at the individual level. Considering these issues, the weighted local adversarial domain adaptation with wide kernels (WWLADA) method is proposed in this article, which can integrate global, local, and individual feature alignment mechanism and establish a unified three-level feature transfer framework. The effectiveness of the proposed WWLADA method is assessed in fault diagnosis transfer tasks across two datasets, achieving peak accuracy rates of 100% and 98% in the most challenging transfer cases, demonstrating its excellent diagnostic performance and transferability in rolling bearing fault diagnosis under multiple conditions.