Fault-Influenced Variable Selection and Distributional Offset Evaluation of CVDs Graph Convolutional Neural Network for Nonlinear KPI-Related Process Monitoring

The majority of fault diagnosis techniques for root cause identification often overlook a critical aspect of complex industrial operations, which is reflected in the neutralization and masking of certain fault-impacted variables. This phenomenon complicates the accurate pinpointing of the root cause variable using conventional approaches. Therefore, a novel fault-influenced variable selection and distributional offset evaluation of causal variation differences graph convolutional neural network (FIVS-DOEGCN) is proposed for industrial KPI-related fault diagnosis. First, to address the high intervariable correlation in operating units and to manage the aggregation of variable information across spatial and temporal dimensions, a subblock feature fusion multitimestep graph convolutional network (SBFF-MTGCN) is proposed for KPI-related process monitoring. Second, a fault-influenced variable selection method (FIVSM) is presented to identify candidate variables closely linked to the root cause. Moreover, an evaluation of the distributional offset of the causal variation differences (CVDs) is designed to get fine-grained characterization of sensor dependencies and obtain accurate root cause diagnosis results. Finally, the experiments on the Tennessee Eastman process and three-phase flow facility describe the superiority of the proposed method.