Accurate remaining useful life (RUL) prediction is essential for prognosis and health management. However, existing methods for RUL prediction generally suffer from the issues of ignoring variable operating modes, requiring numerous training parameters and producing significant prediction errors. To address these challenges, a novel lightweight hybrid network model, CS-TCN-SAPINN is proposed for accurate RUL prediction. Specifically, firstly, raw data are clustered and standardized (CS) based on operating modes. Secondly, for mapping high-dimensional hidden features to the low-dimensional space and capturing long-term dependencies, a temporal convolutional network (TCN) is utilized. Subsequently, a self-attention mechanism assisted physics-informed neural network (SAPINN) is employed for regularizing the prediction network and mapping features to the RUL. Finally, the C-MAPSS dataset is used for validation and results show that, compared with the existing state-of-the-art methods, the proposed approach achieves advanced performance with the least number of trainable parameters.

remaining useful life,prognosis and health management,cluster standardization,temporal convolutional network,physics-informed neural network,self-attention mechanism