A hybrid framework driven by the integration of degradation knowledge and data for remaining useful life prediction of rolling bearings based on Transformer

Abstract With the advancement of condition monitoring technology, deep learning has gained widespread application in prognostics and health management (PHM). However, data-driven predictive methods often heavily rely on raw data leading to compromised accuracy under variable operating conditions and suffer from poor model transparency. Moreover, data-driven approaches may lead to inferences that do not align with physical laws. Poor generalization due to neglect of inherent degradation physics. To address these issues, this paper proposes a physical knowledge integrated Transformer network for RUL prediction. The proposed method employs a channel attention mechanism based on advanced signal processing techniques to learn the degradation process of the target unit from the frequency domain perspective, with the reconstructed features containing rich physical information. In addition, the irreversibility of the degradation process and variations in the degradation rate are viewed as prior knowledge, and a novel loss function is introduced to align the model’s predictions with actual degradation patterns. Experiments on bearing degradation datasets demonstrate the effectiveness and superiority of the proposed method.