Attention-augmented recalibrated and compensatory network for machine remaining useful life prediction

Deep learning methods play an increasingly important role in RUL prediction for machines due to their powerful nonlinear mapping capabilities. However, these methods often suffer from information leakage and correlation loss between features and data during the mapping process. A novel attention-augmented recalibrated and compensatory network (ATRCN) is proposed for RUL prediction, which contains a local interaction-feature (LIF) mechanism and a global compensation-information (GCI) mechanism. Firstly, the LIF mechanism strengthens the correlation between features and attention weights and recalibrate multidimensional feature. Then, the GCI mechanism is used to compensate for the information leakage of the long short-term memory (LSTM) network by adding the information of the intermediate hidden states to the last hidden state according to the attention compensation factor. The proposed method is verified by two benchmark datasets. Experimental results demonstrate that the prediction performance of the ATRCN is better than some existing approaches.