Machine remaining useful life prediction method based on global-local attention compensation network

Accurate remaining useful life (RUL) prediction is essential for ensuring the safe operation of machinery. The extraction of high-level features that contain both global dependencies and local refinements can effectively improve the accuracy of RUL predictions. In order to extract high-level features, this paper proposes a global-local attention compensation network (GLACN) for RUL prediction. The proposed network integrates a global interaction-feature (GIF) mechanism, a long short-term memory network (LSTM), and a local attention enhanced residual compensation (LAERC) mechanism. Initially, the GIF mechanism is used to processed selected signals from multiple sensors to facilitate global information interaction and allocate channel attention weights. Subsequently, the LSTM is employed to extract global temporal features and establish long-term dependencies among them. Finally, the global temporal features extracted by LSTM are further refined by LAERC to mine local features. To address the potential weakening of long-term dependencies during feature refinement, the global temporal features from the last hidden layer of LSTM are utilized as compensation, concatenated with refined features to generate final features. The effectiveness of the designed model for RUL prediction is tested by two benchmark datasets. The results illustrate that the prediction performance of the GLACN outperforms some of some state-of-the-art (SOTA) methods.