Redundancy Modification and Potential Feature Reactivation Network for Predicting the Remaining Useful Life of Machines

Remaining useful life (RUL) prediction is an essential part of prognostics and health management (PHM) technology, which ensures the safe and reliable operation of machines. Recently, deep learning technology has been extensively employed for machine RUL prediction. Nevertheless, there are problems of information leakage and reuse of redundant information when using present models for prediction tasks. To address these problems, a long short-term memory framework with redundancy modification and potential feature reactivation (RMPFR-LSTM) is proposed for multiple sensor data-driven RUL predictive networks. First, the feature squeeze excitation (FSE) mechanism is utilized to achieve adaptive weight assignment to multisensor data at the feature level. Second, the LSTM network is utilized to obtain degradation information from optimized feature data. Meanwhile, the softmax temporal permutation selecting (STPS) mechanism will capitalize on the mid-hidden states of the LSTM for the reactivation of potential degradation features. Finally, multiple fully connected networks (FCNs) are utilized to map features to RUL values. Experimental analyses were performed on two datasets, and the results demonstrate that the method is better than existing methods.