Residual Attention-Driven Dual-Domain Vision Transformer for Mechanical Fault Diagnosis

Traditional fault diagnosis methods, which rely on single-vibration signals, are insufficient for capturing the complexity of mechanical systems. As neural networks evolve, attention mechanisms often fail to preserve local features, which can reduce diagnostic accuracy. Additionally, transfer learning using single-domain metrics struggles under fluctuating conditions. To address these challenges, this paper proposes an innovative adversarial training approach based on the Time–Frequency Fused Vision Transformer Network (TFFViTN). This method processes signals in both the time and frequency domains and incorporates a robust attention mechanism, along with a novel metric that combines Wasserstein distance and maximum mean discrepancy (MMD) to precisely align feature distributions. Adversarial training further strengthens domain-invariant feature extraction. Experiments on bearing and gear datasets demonstrate that our model significantly improves diagnostic performance, stability, and generalization.