A novel infrared thermography-based gearbox fault diagnosis under small samples and different working conditions via few-shot learning

Infrared thermography provides a non-contact solution for gearbox condition monitoring in complex industrial environments. However, reliable fault diagnosis from infrared images remains challenging under few-shot conditions. In this paper, a novel Multi-Scale SwinResNet-enhanced prototype network with CA-Mahalanobis distance metric (MSwinR-CMPNet) is proposed for infrared image-based gearbox fault diagnosis under small samples and multiple different working conditions. First, a Multi-Scale Squeeze and Excitation block (MSSE) is developed to realize multi-scale feature aggregation and channel-adaptive enhancement of shallow layers. Second, the residual convolution and Swin Transformer ensemble block (SwinResNet) is presented to extract local features and model global representations with long-range dependencies to improve the model's feature extraction ability. Third, the covariance-aware Mahalanobis (CA-Mahalanobis) distance metric is involved to replace the traditional Euclidean distance to enhance the class discrimination and confidence reliability of the prototype networks. Experimental results show that the proposed MSwinR-CMPNet model outperforms other benchmarks and achieves an infrared image-based gearbox fault diagnosis accuracy of up to 98.3%. The proposed MSwinR-CMPNet model can provide accurate, stable, and reliable fault identification for infrared-based gearbox monitoring in data-scarce and noise-corrupted industrial scenarios.