Topology-Informed Graph Convolutional Network for Fault Diagnosis

The development of an accurate data-driven fault diagnosis model faces several obstacles. One significant challenge is how to combine the model with the process mechanism rationally and improve the model interpretability. In this work, a fault diagnosis method based on a topology-informed graph convolutional network (TIGCN) is proposed. This model first uses the empirical knowledge that analyzes the operating state of the process and combines the operating data to construct a signed directed graph. Then the operating data is divided into three modalities in the time dimension. The model utilizes a well-designed spatial-temporal convolutional layer and signed directed graph to construct different modalities separately to effectively detect fault types. The superior fault diagnosis performance of the proposed TIGCN model is demonstrated on the Tennessee-Eastman process as compared to traditional approaches.