Remaining Useful Life Prediction Based on Multi-channel Attention Bidirectional Long Short-term Memory Network

Abstract

Recently, a large number of deep learning-based prognostics methods for machinery remaining useful life (RUL) have been proposed. And massive monitoring data is the basis of deep learning-based RUL prediction methods. However, most existing methods usually assume that the monitoring data acquired from different sensors contain similar degradation information, and they lack consideration on effectively identifying the multi-sensor degradation information, which affects the prediction performance of deep networks. To overcome the weakness, this paper proposes a multi-channel attention bidirectional long short-term memory network (MCA-BiLSTM) for RUL prediction of machinery. In MCA-BiLSTM, sensitive features are firstly extracted and selected from the monitoring data of each sensor individually. Then, multi-channel bidirectional long and short-term memory (LSTM) network is constructed with the sensitive features as the inputs. Simultaneously, time attention and channel attention are utilized to realize the adaptive information fusion. Finally, the fusion representations are input into fully connected layers to predict RUL. The effectiveness of MCA-BiLSTM is verified by multisensor monitoring data from life testing of milling cutters.