MIMGFN: a novel multi-information multi-level gated fusion network for remaining useful life prediction of rolling bearings

Abstract Because traditional bearing remaining useful life (RUL) prediction methods often employ a single time series modeling method, making it difficult to account for both local signal variations and global trends. Furthermore, most of bearing RUL prediction techniques only consider a single domain information as the input, ignoring the importance of multi-domain information fusion for bearing RUL prediction. To deal with these problems, a novel bearing RUL prediction method based on a MIMGFN model is proposed. Firstly, the AEnTCN is designed by improving the temporal convolutional network (TCN) framework and embedding the squeeze-and-excitation attention mechanism, thereby enhancing its capability for time-series modeling of vibration signals. In addition, a multi-level gated fusion module (MGFM) is proposed to adaptively control the fusion ratio of multi-source information. By introducing corresponding statistical feature projections into the output of each layer of the AEnTCN residual block and dynamically weighting them using a gating mechanism, the cross-level and multi-scale feature fusion is achieved, enhancing cross-modal information flow and inter-layer collaborative expression. Afterwards, the designed MIMGFN model has a dual-branch architecture consisting of a MGFM branch and a gated recurrent unit branch. By jointly processing vibration signals and multi-domain statistical features, this design greatly improves the model’s capacity to learn and depict intricate degradation behaviors. Finally, the IEEE PHM 2012 challenge dataset and the XJTU-SY test dataset are employed to confirm the efficacy of the suggested approach. The experimental results demonstrate that the proposed model achieves high prediction accuracy in bearing RUL prediction.