Fault Diagnosis of Industrial Control System With Graph Attention Network on Multi-view Graph

In this paper, a fault diagnosis framework for the industrial control systems is proposed based on graph attention networks (GAT). The proposed method models the complex relationship between sensor signals and fuses multivariate sensor signals with relational information. First, to reveal the different relationships between sensor signals, a graph with multi-view is constructed considering the similarity and correlation properties. Second, for each view of the graph, GAT is used to extract the graph feature and adjust the structure information by introducing an attention weight to each edge. Finally, the multi-view graph features are concatenated to obtain the final fault feature of multivariate sensor signals. In addition, the focal loss is introduced to balance the contribution of easy classification samples to the loss function. The superiority of the proposed method is demonstrated by extensive experiments on the three-phase flow facility data set.