Sensor Fault Detection and Diagnosis Using Graph Convolutional Network Combining Process Knowledge and Process Data
Lei Guo, Hongbo Shi, Shuai Tan, Bing Song, Yang Tao
Abstract
The condition of sensors is critical to ensure the safe operation and product quality of industrial processes, but fault detection and diagnosis techniques for sensors have received little attention. To alleviate this problem, we introduce a novel deep-learning (DL) framework that combines process knowledge and graph convolutional networks (KDGCNs) for process sensor fault detection and diagnosis. We inject process knowledge into a data-based modeling approach through graph neural networks (GNNs) and use attention mechanisms to model the dependencies between sensors. We implement sensor fault detection using residuals and determine the location of the faulty sensor using a directed graph. Finally, we set up several sensor faults based on the Tennessee Eastman simulation, and the KDGCN shows satisfactory performance in both detection rate and diagnosis results, indicating that the injected knowledge and graph structure help to achieve accurate sensor fault detection and diagnosis.