Fuzzy Inference and Temporal Logic Based Reliable Graph Neural Network for Industrial Fault Diagnosis

Fault diagnosis is crucial for intelligent control in complex industrial process. In data-driven fault diagnosis filed, graph neural networks (GNNs) can model the complex topological structure and signal temporal approaches can process the time-series data, thus, reasonably combining these two methods is a popular strategy and could achieve satisfactory performance. However, the black-box modeling leads to poor interpretability and reliability in industrial system applications. To overcome this shortcoming, this paper presents a fuzzy inference and temporal logic based reliable graph neural network for industrial fault diagnosis, in which fuzzy inference is extended to GNN, i.e., FIGNN, to improve the physical interpretability of information transmission and parameterized signal temporal logic (PSTL) executes the function of temporal feature extraction. A novel model that integrates the FIGNN and PSTL layers in an alternating way to mutually enhance their capabilities. Experimental results on the Tennessee Eastman process show the performance superiority of the proposed method in industrial fault diagnosis. Meanwhile, the entire diagnosis process can be visualized by a reliable interpreter, benefiting from fuzzy inference and temporal logic.