A Physics-Inspired CNN-Transformer Synergistic Framework for Bearing Fault Diagnosis under Strong Noise Environments

Abstract Rolling bearings play a crucial role in the operation of rotating machinery within industrial systems. To ensure the stable operation of mechanical equipment, the accuracy of bearing fault diagnosis is particularly critical. In recent years, deep learning (DL) has been widely applied in research on fault diagnosis of rotating machinery. Although significant achievements have been made in this field, some challenges still persist. Traditional convolutional neural networks (CNNs), while adept at extracting local textures, lack a global perspective and struggle to fully characterize the overall health status of bearings. Moreover, under strong noise and coupled interference, their sensitivity to fault features can be limited, leading to the loss of key fault characteristic information. To address these issues, this paper proposes Efficient Multi-scale Convolutional-Transformer Network (EMCT-Net), a physics- inspired CNN-Transformer synergistic framework designed to decouple and extract transient impacts and global periodicities from bearing fault signals. First, the acquired raw one-dimensional vibration signals are converted into two-dimensional time-frequency images using continuous wavelet transform. Subsequently, to capture the local transient energy bursts characterizing early faults, a designed Multi-scale Cross-spatial learning Convolutional Attention (MCCA) fusion module is employed. The MCCA acts as an information bottleneck to restrict unconstrained noise. Finally, the sliding window mechanism of the Swin Transformer captures global information, compensating for the limitations of the convolutional module, and ultimately achieves fault diagnosis classification. EMCT-Net's superiority and robustness are validated on three bearing datasets from different perspectives. Comparative evaluations with various state-of-the-art methods demonstrate that the proposed approach exhibits superior robustness and accuracy, even under strong noise interference.