Process monitoring based on global-local multi-information integrated progressive graph convolutional network using causal inference and variable perturbation

Abstract

Background

Process monitoring in modern industrial processes is essential, however, few existing methods have been proposed to differentiate the scope of influence of these faults. Furthermore, incorrect fault traceability can be misleading to operators and negatively affect fault isolation.

Method

This paper proposes a process monitoring method named G-L MIIPGCN for unit-coupled industrial processes. First, the spatial topology graph of time-ordered correlation is constructed for local monitoring, and sequence-to-sequence latent variable forecasting is introduced to better capture the dynamic attributes. Second, the process monitoring indicators are constructed by fusing local information through the adaptive weighted summation mechanism (AWSM) and global feature selection (GFS). Then, the constrained path search algorithm (CPSA) is proposed to obtain fault propagation paths, and the path propagation selection indicator (PPSI) is introduced to obtain the dominant fault propagation path and an evaluation indicator is used to judge the trustworthiness of it.

Significant Findings

Our analysis indicates the inaccurate localisation of fault-generated effects significantly influences the monitoring performance. Experimental results demonstrate that G-L MIIPGCN exhibits excellent performance on the Tennessee Eastman dataset. This method effectively mitigates the problem caused by the coupled units and the smearing effect between variables, demonstrating its potential in process monitoring.