Multi-Head Attention Network with Adaptive Feature Selection for RUL Predictions of Gradually Degrading Equipment

Abstract

A multi-head-attention-network-based method is proposed for effective information extraction from multidimensional data to accurately predict the remaining useful life (RUL) of gradually degrading equipment. The multidimensional features of the desired equipment were evaluated using a comprehensive evaluation index, constructed of discrete coefficients, based on correlation, monotonicity, and robustness. For information extraction, the optimal feature subset, determined by the adaptive feature selection method, was input into the multi-head temporal convolution network–bidirectional long short-term memory (TCN-BILSTM) network. Each feature was individually mined to avoid the loss of information. The effectiveness of our proposed RUL prediction method was verified using the NASA IMS bearings dataset and C-MAPSS aeroengines dataset. The results indicate the superiority of our method for the RUL prediction of gradually degrading equipment compared to other mainstream machine learning methods.