Prior Graphs Approach to GCN-Based Quality Variable Prediction for Industrial Processes

Graph convolutional network (GCN)-based predic-tive models have become increasingly popular for quality variable prediction in recent years. They effectively incorporate prior knowledge in the form of graphs, enhancing the generalizability and interpretability while maintaining the high accuracy of deep learning (DL). This study explores three prevalent prior graphs in GCN-based predictive models, including the mechanism graph, process topology graph, and K-nearest neighbor (KNN) graph. Performance evaluation is conducted on a benchmark debu-tanizer process. Two practical production scenarios are consid-ered: limited samples due to high labeling costs and incomplete variables from sensor damage or maintenance. Experimental results demonstrate performance improvement using the mech-anism graph and process topology graph. The strengths and limitations of different prior graphs are summarized based on simulation results. Future research directions regarding GCN-based quality variable prediction are provided at the end.