Deep multiscale feature fusion network with dual attention for rolling bearing remaining useful life prediction

Aiming at the existing life prediction methods for rolling bearing degradation information mining is not sufficient, the critical time step information degree is insufficient, resulting in the loss of key degradation information, model prediction accuracy and model generalization ability is insufficient, this paper proposes a novel deep multiscale feature fusion network with dual attention for rolling bearing remaining useful life (RUL) prediction. First, multi-domain feature sets of rolling bearing vibration signals are acquired. Subsequently, it is proposed to use Squeeze-and-Excitation (SE) attention mechanism to calibrate and weigh temporal significance of feature sequences, thereby capturing critical temporal information. Then, a multi-scale feature extraction and fusion module with deep network is constructed, consisting of multiple identical multi-scale residual pyramid layers connected in series to further delve into the state information of rolling bearings. Additionally, a relative position encoding method suitable for time series prediction is introduced within the multi-head attention mechanism, a network architecture based on a dual attention-enhanced Transformer encoding layer is established. This enhancement significantly improves model prediction accuracy, generalization capability, and sequence data comprehension ability. Finally, the high-level features output from the feed-forward layer are mapped through a regression layer to obtain the final prediction results. Experimental results demonstrate the superior performance of the proposed method in terms of both prediction accuracy and generalization capability for rolling bearing life prediction. A more robust and accurate framework for RUL prediction in rolling bearings is provided.