Adversarial representation learning for generalizable remaining useful life prediction

Predicting the Remaining Useful Life (RUL) through domain generalization has recently been introduced to address the challenge of domain bias in unseen conditions. Current approaches primarily focus on extracting domain-invariant features from multiple source domains. However, obtaining valuable failure data from different machines or operational settings is difficult, as most available datasets are derived from a single condition. This paper presents a novel Adversarial Representation Learning (ARLearn) framework, which includes three modules designed for cross-domain RUL prediction under unseen conditions using only one source domain. The first module, responsible for generating diversity, augments data to mimic the unseen target domain. To further enhance the generated data’s utility, a diversity enhancement module is employed to refine its distributional properties. Additionally, a discrimination preservation module is incorporated to retain the semantic information of the generated data, thereby improving the model’s resilience to noise. Extensive experiments conducted on two public datasets confirm the proposed method’s effectiveness and superiority.