Attention Mask-Guided Optimal Transport for Few-Shot Fault Diagnosis Under Strong Noise and Spatial Sensor Domain Shifts

Fault diagnosis across sensor locations becomes difficult when vibration measurements contain strong noise and only a few labeled target samples are available. This study develops an Attention Mask-guided Optimal Transport (AM-OT) framework for this setting. A wide-kernel multi-scale network with channel attention extracts vibration representations, while supervised noise-contrastive pretraining encourages class discrimination across noise perturbations. During target adaptation, a joint source–target attention mask reweights feature differences in the optimal transport cost, reducing the influence of low-weight channels on alignment. In experiments using measured CWRU bearing signals from three sensor locations with controlled Gaussian noise, the reported five-shot comparison achieves an average accuracy of 71.40% across six spatial transfer tasks using 20 labeled target windows, or 0.5% of the target domain. These results demonstrate the effectiveness of AM-OT for few-shot cross-sensor diagnosis on the noisy CWRU benchmark.