Hybrid Spectral Graph Attention Learning for Intelligent Rolling-Bearing Fault Diagnosis Using Vibration Signals

Rolling-bearing defects alter contact dynamics and generate characteristic changes in vibration responses, making vibration analysis an effective tool for condition monitoring and fault diagnosis. However, conventional machine-learning and deep-learning approaches generally process vibration signals as regular vectors or sequences and therefore do not explicitly model the relationships among fault-sensitive spectral components. This study proposes a hybrid spectral graph-attention framework for intelligent rolling-bearing fault diagnosis. Vibration recordings from the Case Western Reserve University (CWRU) Bearing Data Center and the HUST bearing dataset were harmonized into four common classes: healthy, inner-race fault, outer-race fault, and rolling-element fault. HUST signals were resampled from 51.2 to 48 kHz, while all signals were segmented into 4096-sample windows with 50% overlap and standardized at the window level. Each window was converted into a 16-node spectral graph, in which nodes represented frequency bands and edges connected first- and second-order neighboring bands. Eleven time- and frequency-domain descriptors were assigned to each node. A GATv2 encoder was evaluated in an end-to-end configuration with a multilayer perceptron classifier and as a graph-feature extractor coupled with support vector machine, random forest, and logistic regression classifiers. A leakage-aware temporal split was applied to CWRU, whereas HUST was evaluated using independent recording-level splits. The GATv2–logistic regression model achieved the best performance on both datasets. On CWRU, it obtained 98.12% accuracy and 98.43% Macro-F1, compared with 95.31% accuracy and 96.10% Macro-F1 for end-to-end GATv2. On HUST, it achieved 92.97% accuracy and 92.92% Macro-F1, outperforming the end-to-end model, which obtained 91.87% accuracy and 91.83% Macro-F1. These results show that spectral graph attention effectively captures discriminative relationships among vibration-frequency regions and that coupling GATv2 embeddings with logistic regression provides a consistent framework for intelligent bearing condition monitoring across heterogeneous datasets.