A customized dual-transformer framework for remaining useful life prediction of mechanical systems with degraded state

Abstract

Remaining Useful life (RUL) prediction is important to ensure the stable operation of mechanical systems. Recently, deep learning (DL) has achieved success in RUL prediction tasks of mechanical systems. However, existing DL-based RUL prediction methods face two significant limitations: (1) they are difficult to perceive the change time from normal state to degradation state. (2) their prediction performance is limited since they fail to capture multi-term patterns in the degraded state. To address these problems, a customized dual-Transformer framework is proposed for RUL prediction of mechanical systems by considering the degraded state. First, a contrastive Transformer network is designed to learn representation discrepancy of operation state for determining uncertain change time. Moreover, a versatile Transformer network is developed to capture multi-term dependencies for RUL prediction beyond the state change point. Finally, a self-built IR experiment and IMS bearing datasets are implemented to validate the effectiveness and superiority of the proposed method. The experimental results demonstrate that our proposed method can effectively determine state change time in advance and achieve high-precision RUL prediction of mechanical systems.