A new dual-channel transformer-based network for remaining useful life prediction

Abstract Remaining useful life (RUL) prediction is crucial for enhancing the reliability of complex systems, preventing unexpected failures, and reducing the maintenance costs of electromechanical systems. Data-driven methods, particularly deep learning, have been extensively applied in RUL prediction. However, recent methods have focused solely on long-term features, neglecting the significant role of short-term features in long sequence processes, which leads to an inability to achieve high-precision RUL prediction. In this article, a novel dual-channel network is proposed, which is based on the transformer encoder. Additionally, the approach proposes a robust dual-channel feature extraction layer that integrates a multi-scale convolution block and a gating convolution block. This integrated system is capable of capturing both long-term and short-term features from the raw data simultaneously. Compared to convolutional neural networks, recurrent neural networks and other transformer-based methods, the proposed method is more effective at extracting profound features and focusing on important information in the input. Utilizing the self-attention mechanism, the proposed model captures degradation features and achieves precise RUL prediction. Experiments conducted on the CMAPSS dataset demonstrate that the proposed method outperforms recent RUL prediction methods. The average root mean square error of the proposed method is reduced by 6.8%, and the average score is reduced by 8.5%, proving the effectiveness of the proposed approach.