Supervised contrastive learning-based domain generalisation network for rotating machinery cross-domain fault diagnosis

In recent years, intelligent diagnosis models based on domain adaptation have been widely applied in the variable working condition scenarios of rotating machinery. However, the need to access the target domain data in advance limits the application of these models in practical application scenarios. To address this issue, a supervised contrastive learning-based domain generalisation network (SCLDGN) is proposed for cross-domain fault diagnosis of rotating machinery. A hybrid pooling spatial attention (HPSA) module is developed to refine the feature information extracted from the original vibration signal and match the mean and covariance of feature distributions across all source domains to achieve global alignment. Subsequently, class-discriminative features are learned by minimising the supervised contrastive loss, while intra-class compactness is enhanced via a class centre loss, thereby improving the reliability of predictions under unknown working conditions. Diagnostic cases from two experimental scenarios demonstrate the effectiveness and superiority of the proposed approach. The proposed method achieves outstanding classification accuracies of 85.87% and 87.81% on the PU bearing dataset and the SQI gearbox dataset, respectively.