Self-attention Mechanism Network Integrating Spatio-Temporal Feature Extraction for Remaining Useful Life Prediction

Abstract

Prognostics and health management technology for industrial equipment heavily relies on the accurate prediction of the remaining useful life (RUL). As commonly used RUL prediction approaches, the conventional convolutional neural network, and long-short term memory network are not only difficult to realize the extraction process of spatio-temporal features, but also cannot reflect the difference between the data at different moments in the RUL prediction results. Aimed to deal with these problems, a self-attention mechanism network integrating spatio-temporal feature extraction (SAMN-STFE) is proposed to predict RUL, which can deliver higher weight to the significant moments. In detail, feature selection and noise reduction are performed on the data picked up by the multiple sensors during the working process. The self-attention mechanism network assigns corresponding weights to different time points in the time window. Afterward, the spatial features are extracted by one-dimensional convolutional neural network. The temporal features are extracted by bidirectional long short-term memory networks. Ultimately, the trained SAMN-STFE can be utilized for online RUL prediction. To validate the proposed approach for predicting RUL, the dataset of aircraft turbofan engines, furnished by NASA Ames Prediction Center is employed. Experimental results represent that the proposed approach has excellent RUL prediction performance.