An unsupervised subdomain adaptation of cross-domain remaining useful life prediction for sensor-equipped equipments

Abstract

Most existing remaining useful life (RUL) prediction models assume a similarity between the data distributions in the source and target domains, but this assumption is often challenging in practical applications. Although traditional transfer learning methods could relax this assumption by minimizing the distributional differences between domains, they often ignore the fine-grained features exhibited by equipment at different health state and the inherent properties of the domains. Additionally, most models ignore the importance of the location information of the sensors equipped on the equipment, and the real-world problem of the high cost of fully labeling the data. Addressing these challenges, we introduce a novel model of unsupervised subdomain adaptation and feature disentanglement for equipment RUL prediction. The model contains four components: a feature extractor, a feature similarity component, a domain adaptation component, and a multi-task module. Specifically, the feature extractor designs a dynamic fusion network to divide the domain’s private and shared features. The feature similarity component employs a cross-attention mechanism to constrain the extracted features. The domain adaptation enables the proposed model to extract deeper subdomain-shared features. The multi-task module simultaneously divides subdomains and performs RUL prediction. The proposed model is validated using multiple tasks and compared with existing works.