DEViT: Deformable Convolution-Based Vision Transformer for Bearing Fault Diagnosis

Motor bearing faults can lead to significant industrial losses and accidents, making their accurate diagnosis essential for safe motor operation. In recent years, transformer-based models have achieved good results for bearing fault diagnosis, which, however, suffer from several limitations: 1) they equally extract patch features during the embedding stage, neglecting to prioritize more important features, resulting in the extraction of less useful or redundant local features and degrades the model performance and 2) they have quadratic computation complexity in the traditional self-attention mechanism, leading to high computational cost. To address these issues above, this study introduces the patch feature extraction (PFE) module to extract local crucial and global features to improve feature extraction and representation in the early stage, instead of utilizing multilayer networks as traditional transformer-based models. Additionally, the prob-attention module is incorporated to further improve computational efficiency and diagnostic accuracy. Then, a novel transformer-based model named DEViT is proposed for bearing fault diagnosis, which inherits the strengths of deformable convolution (DCN) and vision transformer (ViT). DEViT model utilizes PFE patch embedding to accomplish segmentation and underlying feature extraction of the 1-D signal, and then, the embedded sub-signal is deeply characterized by a multihead attention mechanism. Experimental results on two benchmark datasets, Case Western Reserve University (CWRU) and Southeast University (SEU), demonstrate DEViT’s superior performance with top-1 accuracies of 99.81% and 97.32%, respectively. Furthermore, ablation experiments are conducted to confirm DEViT’s advantages in recognition accuracy, stability, and efficiency.