Synergistic Similarity Graph Construction (SSGC) for Steel Plate Fault Diagnosis with Graph Attention Networks

Fault diagnosis in industrial production is vital as emerging technologies require innovative methods to identify subtle fault distinctions. Traditional machine learning approaches for steel plate fault classification inadequately exploit sample relationships, limiting accurate diagnosis. Thus, we proposed to use graph construction methods in conjunction with Graph Attention Networks (GAT) for steel plate fault classification. We introduced the following four techniques to generate adjacency matrices representing sample connections: k-nearest Neighbors (k-NN), threshold-based cosine similarity, threshold-based Mahalanobis distance, and Minimum Spanning Tree (MST), Additionally, we proposed a novel graph construction algorithm, Synergistic Similarity Graph Construction (SSGC), to fuse these adjacency matrices, leveraging the strengths of each technique. We evaluated the proposed techniques on the UCI Steel Plates Faults dataset, comparing them with traditional machine learning models. The results demonstrated that the combination of GAT and SSGC leads to superior performance over the best traditional machine learning models, improving the accuracy, precision, recall, and Macro-F1 scores by 4.8, 4.8, 4.6, and 4.7%, respectively. In conclusion, The proposed method is a novel approach for steel plate fault classification using GAT and effective graph construction techniques. The method expands Graph Neural Networks (GNNs) applicability to tabular datasets without explicit connections, enhancing fault classification performance in industrial production. This research showed how to use GNNs in fault diagnosis across diverse domains. Our code and data are available at https://github.com/AnguoCYF/SSGC