A sequence ensemble method based on Sparse Transformer with bidirectional gated recurrent convolution unit for remaining useful life prediction

The remaining useful life prediction (RUL) of rotating machinery is crucial for Prognostic Health Management (PHM) tasks. Relying on deep learning methods, the RUL prediction task of mechanical facility has made good progress. However, existing methods often do not consider mechanical degradation rules and cannot effectively capture short-term and long-term feature dependencies. To solve such problems and achieve a better RUL prediction effect, this paper designs a sequence ensemble model. The model synergistically integrates two key components: a bidirectional gated recurrent convolution unit (Bi-GRCU) and Sparse Transformer. This integration enables comprehensive dependency learning across different time scales. Specifically, the Bi-GRCU combines the strengths of recurrent and convolutional networks to effectively capture short-term dependencies. Meanwhile, the Sparse Transformer employs long-range locality sparse attention (LRLS-Attention), which can efficiently identify long-term dependencies by filtering redundant computations and reducing dot-product complexity. By sequentially integrating these two components, the proposed model achieves a balanced representation of both local and global temporal patterns, significantly improving prediction accuracy and robustness. The related experiments are conducted by applying the CMAPSS aeroengine dataset and bearing wear PHM2010 dataset, and the results demonstrate that the predictive ability of proposed model is superior to other existing methods.