A Novel Lightweight Rotating Mechanical Fault Diagnosis Framework With Adaptive Residual Enhancement and Multigroup Coordinate Attention

Fault diagnosis of rotating machinery is widely recognized as a challenging problem. Recent advances in combining convolutional neural networks (CNNs) and Transformers have expanded the capabilities of intelligent fault diagnosis. However, in real-world industrial settings, three critical challenges substantially affect fault diagnosis performance: resource-constrained deployment environments, variable operating conditions, and cross-domain adaptability requirements. These challenges frequently undermine the efficacy of existing algorithms, particularly in sustaining robust performance while meeting lightweight implementation requirements. To address these challenges, a novel lightweight fault diagnosis framework, termed adaptive residual enhancement-multigroup coordinate attention transformer (ARE-MGCAFormer), is introduced in this article. First, an ARE block is specifically designed to extract multilocal receptive field features from vibration signals. The integrated gating mechanism utilizes learnable Params to adaptively balance the contributions of deep separable convolutions and inverse residual blocks, dynamically adjusting to varying signal features and noise levels while maintaining a lightweight design. Second, a multigroup coordinate attention (MGCA) mechanism is incorporated to effectively extract critical detailed features across the entire signal range while reducing computational complexity by distributing attention across multiple feature groups. Experimental verification was conducted using the gearbox dataset from Xi’an Jiaotong University (XJTU) and the rolling bearing fault dataset from the University of Ottawa (OU). The results demonstrate that the proposed framework exhibits superior lightweight characteristics and robustness in fault diagnosis tasks compared to recent mainstream frameworks based on CNNs and Transformers.