A swin transformer-based fault diagnosis method with dynamic feature routing and spectrum awareness

In industrial practical applications, the vibration signals of rolling bearings are often overwhelmed by strong background noise, resulting in a significant decline in the performance of traditional deep learning models. Although the Transformer is good at capturing global dependencies, its standard self-attention mechanism is sensitive to noise and has a fixed calculation path, making it difficult to adaptively handle variable features. To address these challenges, this paper proposes a novel fault diagnosis framework called Dynamic Routing Spectrally-Aware Network (DRSAN). This framework integrates two key innovations. First, the Spectrum Enhancement Attention (SEA) mechanism is designed, which incorporates frequency domain weights to focus on key fault frequencies and suppress noise. Second, the Dynamic Feature Routing (AMoE) mechanism routes features adaptively using a Mixture of Experts (MoE) and discards the rigid computational path. Finally, the features extracted by the above modules are fused with fine local features to complete the final fault classification. Validation on an open bearing dataset is given to illustrate the effectiveness of presented framework.