Embedding-Enhanced Graph Attention Networks for Imbalanced Industrial Fault Diagnosis

Intricate nonlinear interactions among measurements and imbalanced distribution of fault samples pose considerable challenges for accurate fault diagnosis in modern industrial processes. To address these challenges, this study proposes an embedding-enhanced graph attention network to learn efficient representations, which captures the delicate interactions among sensors with learnable node embeddings, facilitating accurate fault diagnosis. First, sensor interactions are transformed into graphs using K-nearest neighbors using the similarity of nodes, with node features and edge weights learned through embedding and attention mechanisms, respectively. Subsequently, a dynamic attention mechanism, incorporating both node embedding and measurements, is adopted to update sample representations for fault diagnosis. During training, the focal loss, instead of cross-entropy loss, is adopted to emphasize hard classes and address category imbalance. Finally, the proposed model is tested on two real-world industrial process datasets, and the results demonstrate the efficiency and superiority of the proposed model compared to existing state-of-the-art models.