HCA-VGANet: A hybrid condition-aware VMD-GAT for robust bearing fault diagnosis under varying operating conditions

Deep learning-based fault diagnosis models for rolling element bearings often exhibit poor generalization when faced with varying operating conditions. To address this critical challenge, this paper introduces a novel Hybrid Condition-Aware Variational Mode Decomposition (VMD)-GAT Network (HCA-VGANet). The core innovation of this framework is the conceptualization of the diagnostic task as a graph-based relational learning problem. We first employ VMD to disentangle the raw vibration signal into multiple intrinsic mode functions which, along with the raw signal, form the nodes of a graph. Crucially, the operating condition, autonomously identified from the signal’s envelope, is integrated as an explicit guiding node within this graph structure. This graph is then processed by a condition-guided relational aggregator, which utilizes a dynamic graph attention mechanism to leverage the condition node to adaptively modulate the information flow and learn condition-invariant relational features. Extensive experiments conducted on the HUST bearing dataset under a rigorous cross-speed scenario validate the superiority of HCA-VGANet. The proposed model significantly outperforms various baseline and advanced methods in diagnosing faults under operating conditions unseen during training. The results demonstrate that explicitly modeling the operating condition as an active guiding component in a graph-based framework is a highly effective strategy for overcoming the domain shift problem in intelligent fault diagnosis.