Multi-Scale Attention Convolution Subdomain Adaption Network for Cross-Domain Fault Diagnosis of Machine

Transfer learning method represented by domain adaptation has achieved great success in cross-domain fault diagnosis of machine. However, there are still some issues to be addressed to further improve diagnostic performance. First, most existing methods extract features from a single scale, which may lose useful multi-scale feature information from vibration signals. Second, most of the existing methods majorly align the source and target distributions from a global perspective to reduce distribution discrepancy across domains, which ignores the relationship of the corresponding subdomains from the same category in different domains and leads to unsatisfying transfer learning results. To solve these issues, a multi-scale attention convolution subdomain adaption network is proposed for mechanical fault diagnosis under cross-domain conditions. Firstly, a multi-scale attention convolution block is built to extract and adaptively fuse multi-scale fault features. Secondly, the local maximum mean discrepancy metric is introduced to align the subdomain distributions of the source and target domains. The proposed method is evaluated based on six different transfer diagnostic tasks under variable speeds in case study, and the experimental results verify its adaptability and advantage over other advanced methods.