Multiscale dilated attention convolutional neural network for remaining useful life prediction of rolling bearings in noisy environments

Abstract Accurate prediction of the remaining useful life (RUL) of rolling bearings is crucial for the safe and efficient operation of mechanical equipment. However, most existing studies only predict the RUL in ideal or in low-noise environments, and fail to take into account the complex noise interference in real environments. Therefore, a novel multiscale dilated attention convolutional neural network (MSDA-CNN) is proposed to predict RUL of rolling bearings. In our approach, the multiscale dilated causal convolution module (MDCM) leverages dilated causal convolutions to capture degradation characteristics across multiple scales. Furthermore, we incorporate a dynamic attention module into the MDCM to refine feature selection and highlight key degradation patterns. To further enhance feature representation, a selective attention fusion module is designed to adaptively calibrate and fuse cross-scale features while suppressing redundant and noisy information. In addition, the dynamic feature fusion module integrates multi-path features through convex-weighted gating instead of high-dimensional concatenation, thereby improving efficiency. Extensive experiments on the PHM 2012 bearing dataset under varying noise scenarios demonstrate that the proposed MSDA-CNN substantially outperforms several advanced models in terms of prediction accuracy and noise robustness, thereby confirming its effectiveness and superiority for industrial applications.