A two-stage subdomain adaptation network for machinery remaining useful life prediction

Many domain adaptation (DA) methods have been explored to address distribution discrepancy and knowledge transfer between the source and target domains of machinery. However, in existing DA methods, the DA operation is performed directly on the sensor data throughout the degradation stage, and the forced alignment of the data distributions in different degradation stages is likely to lead to negative transfer. To address this problem, a cross-domain RUL prediction method, i.e., a two-stage subdomain adaptation network (TSSAN) is proposed. First, the degradation stages segmentation (DSS) mechanism is designed to divide machinery degradation process into two stages: pre- and post-degradation. The sensor data of different degradation stages are divided into different subdomains to better align the data distribution between both domains. Then, subdomain adaptation is employed on the source and target domain sample at various stages of degradation to align the marginal and conditional probability distributions. Finally, the dynamic weighting block is designed to dynamically weight each sample for better domain alignment. The effectiveness of the proposed methodology is validated through extensive cross-domain RUL prediction experiments, using two public datasets.