Fault diagnosis of rolling bearings via multi-scale symmetric graph convolutional auto encoder with self-supervised and meta-learning

In small sample learning scenarios, traditional graph embedding methods suffer from limited generalisation capability, single-node representation, and a lack of attention to node attribute extraction. To address these issues, we propose a Multi-Scale Symmetric Graph Convolutional Auto encoder (MS-SGCAE-SSML) method that integrates self-supervised learning and meta-learning. This method enhances feature representation capabilities through self-supervised learning, effectively utilising unlabelled data for feature learning, while meta-learning improves the model’s adaptability and generalisation ability. Based on this, we construct a relational structure graph of the original features, employing a symmetric auto encoder framework for graph embedding and feature dimensionality reduction. The method simultaneously leverages the graph’s topological structure and node attributes to generate node representations and utilises the auto encoder framework to extract deep latent representations of the graph’s geometric structure. This approach enhances the model’s representation ability on attribute graphs. Experimental results demonstrate that the MS-SGCAE-SSML significantly improves classification accuracy, stability, and robustness in rolling bearing fault diagnosis under small sample conditions, achieving a diagnostic accuracy of 96.3%. This indicates that the proposed method can effectively enhance diagnostic performance in small sample learning and unlabelled data environments, thereby demonstrating strong potential for practical applications.