Encoder-Enhanced Graph Convolutional Network for Cross-Domain Mechanical Fault Diagnosis

Graph Convolutional Network (GCN) has been widely applied in mechanical fault diagnosis due to its ability to extract features from data in non-Euclidean spaces. However, the local aggregation mechanism in GCN has difficulty capturing the dynamic propagation relationships between fault signals. Additionally, the class imbalance in fault data significantly hampers model performance. To address these challenges, a new Encoder-Enhanced Graph Convolutional Network (EncGCN) is proposed, which integrates GCN with the attention mechanism to capture long-range dependencies, overcoming the local homogeneity assumption of conventional GCN. Furthermore, a Hierarchical Combination Optimization (HCO) strategy is designed, which dynamically adjusts the learning rate by combining the Adam and SGD optimizers, significantly enhancing the performance of EncGCN in small-sample cross-device transfer tasks. Experimental results demonstrate that EncGCN achieves higher classification accuracy on three public bearing datasets. It also exhibits superior generalization in cross-domain tasks, with the HCO strategy playing a critical role in the transfer process.