Towards trustworthy remaining useful life prediction through multi-source information fusion and a novel LSTM-DAU model

Remaining useful life (RUL) prediction is a key part of the prognostic and health management of machines, which can effectively prevent catastrophic faults and decrease expensive unplanned maintenance. A good health indicator (HI) can ensure the accuracy and reliability of RUL prediction. However, most of the existing HI construction methods use only a single signal and rely heavily on prior knowledge, making it difficult to capture critical information about mechanical degradation, which in turn affects the performance of RUL prediction. To solve the above problems, a novel adaptive multi-source fusion method based on genetic programming is proposed for building a HI that can effectively reflect the health state of machines. Subsequently, a new LSTM model with a dual-attention mechanism is developed, which differentially handles the network input information and the recurrent information to improve the prediction performance and reduce the time complexity at the same time. Experimental validation is carried out on the real PRONOSTIA bearing dataset. The comparative results validate that the constructed fusion HI has better comprehensive performance than other fusion HIs, and the proposed prediction method is competitive with the current state-of-the-art methods.