An Improved Method for Predicting the Remaining Useful Life Using a Spatial-temporal Feature Extraction Network with Attention Mechanism

• Published in: IEEE access Vol. 12; p. 1
• Main Authors: Yan, Xiaojia, Liang, Weige, Sun, Shiyan
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: attention mechanism; Data mining; Datasets; Deep learning; Degradation; Feature extraction; Life estimation; Long short term memory; Mathematical models; Monitoring; Performance prediction; Predictive models; Remaining useful life; spatial-temporal feature; Spatiotemporal phenomena; unidirectional residual convolutional network; Useful life
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3396156

The prediction of the remaining useful life (RUL) of mechanical equipment is of vital importance to its operation and maintenance. Deep learning methods can effectively extract degradation information closely related to equipment RUL from extensive monitoring data. However, when the data is nonlinear, multi-dimensional, long-term, and large-scale, the critical degradation information for RUL prediction may be obscured. Traditional deep learning networks do not perform well in predicting RUL. Therefore, a method for predicting the remaining useful life of mechanical equipment using a spatial-temporal feature extraction network is proposed. The main innovations can be classified into two aspects. Firstly, by training a unidirectional residual convolutional network (URCNN), the deep spatial features of the monitoring data are extracted. This network does not disrupt the temporal relevance of the monitoring data and can effectively avoid the phenomenon of vanishing gradient during the training process. Secondly, the weight parameters of the bidirectional long short-term memory network (BiLSTM) for extracting time-related features are optimized by introducing an attention mechanism. The attention mechanism can effectively enhance the expression of crucial degradation information for RUL prediction. Eventually, the benchmark dataset and the specialized transmission mechanism dataset validates the effectiveness and superiority of the proposed method. The analysis results indicate that for multi-dimensional monitoring data with complex operating conditions and variable fault modes, the proposed method can accurately locate degradation temporal points and effectively improve the RUL prediction accuracy of long-term operating equipment.