Semantic-aware dynamic graph neural network for zero-shot rolling bearing fault diagnosis

Rolling bearings are key components of rotating machinery, while faults occurring at different locations produce distinct vibration signatures. However, existing vibration-based fault diagnosis methods struggle to distinguish these fault locations due to the similarity of vibration signals and the interference of background noise. Moreover, these methods typically require large amounts of well-labelled fault data for training, which is frequently unavailable in actual industrial scenarios. In this paper, a Semantic-aware Dynamic Graph Neural Network (SD-GNN) is proposed to address these challenges within a zero-shot learning framework. Instead of using binary attribute matrices, SD-GNN encodes fault attribute descriptions via a pre-trained model BERT, which injects linguistic priors into the attribute space. A three-layer heterogeneous graph is constructed, consisting of sample nodes, attribute nodes, and class nodes. The sample-attribute and attribute-class edges over nodes explicitly establish a semantic bridge linking the feature space with the class space. Meanwhile, a gate-controlled adaptive edge weight fusion mechanism is designed. It dynamically fuses the Gaussian kernel similarity in the feature space and the cosine similarity in the semantic space through a lightweight gating network, allowing each edge to independently learn its optimal fusion ratio. Furthermore, SD-GNN is trained with a joint loss consisting of attribute prediction and semantic alignment. The semantic alignment loss aligns sample features with BERT embeddings, which enhances the accuracy of the proposed method. A strict zero-shot training paradigm is adopted, where the node features of unseen classes are set to zero vectors, and all associated attribute-class edges are removed during training, to prevent information leakage. Experiments on the CWRU and WPTB bearing datasets demonstrate that SD-GNN has superior performance in zero-shot diagnosis, with average accuracies of 97.72% and 86.38%, respectively. Ablation studies demonstrate the impact of each core component on accuracy separately. Comparisons with other ZSL methods verify the effectiveness of SD-GNN.