Unsupervised physics-informed domain-adversarial graph convolutional network for bearing fault diagnosis under varying operating conditions

In real-world industrial settings, bearing operation is influenced by varying operating conditions, resulting in vibration responses that exhibit significant multi-scale feature coupling characteristics. Moreover, reliable data for characterizing these responses remain scarce. Traditional methods lack the ability to adaptively integrate multi-scale information during feature modeling and fail to effectively incorporate physical constraints related to failure mechanisms, resulting in unstable feature representations and diagnostic results that lack physical consistency. To address these challenges, an unsupervised physics-informed domain-adversarial graph convolutional network is proposed for bearing fault diagnosis under varying operating conditions. Within this framework, a Kernel Selective Fusion Attention mechanism is introduced to selectively fuse the extracted multi-scale convolutional responses, thereby enhancing the stability and discriminability of graph features. Furthermore, a physical prior distribution is constructed based on the envelope power spectrum and fault characteristic frequencies, and the prior knowledge of fault mechanisms is incorporated into the model optimization process through a sample-level gating strategy and a delayed-start linear ramp-up weighting. The experimental results demonstrate that the proposed method outperforms the comparison methods in terms of fault diagnosis accuracy in the target domain, cross-domain feature alignment capability, and compound-fault recognition performance, thereby validating its effectiveness in unsupervised domain adaptation for diagnosis under varying operating conditions.