A feature fusion model based on Convolutional Sparse Transformer with convolutional recurrent gated unit for remaining useful life prediction

The remaining useful life prediction (RUL) of mechanical device is pivotal for the Prognostic Health Management (PHM) tasks in the industrial sector. Although data-driven methods have advanced RUL prediction, many existing approaches fail to align with mechanical degradation rules and cannot effectively capture feature dependency relationships among monitoring data relevant to RUL prediction. To address these limitations, this paper proposes a novel feature fusion model integrating two core components: a Convolutional Recurrent Gated Unit (CRGU) and a Convolutional Sparse Transformer. The CRGU leverages the functional traits of convolutional and recurrent networks to effectively capture short-term local dependencies, enabling the model to concentrate on context-aware local features in time-series data. Moreover, the Convolutional Sparse Attention is proposed and incorporated into the Encoder structure to form a Convolutional Sparse Transformer. This design efficiently captures long-term feature relationships while avoiding the memory and computational bottlenecks caused by global computation in traditional self-attention mechanisms. By fusing features from CRGU and Convolutional Sparse Transformer, the model comprehensively leverages local and global degradation information. Experimental results on the CMAPSS and PHM2010 datasets validate that the proposed model outperforms state-of-the-art methods in prediction accuracy.