Community inspired edge specific message graph convolution network for predictive monitoring of large-scale polymerization processes

Abstract

The polyester fiber production process is a typical example of a large-scale system. The monitoring and prediction of key process variables play a crucial role in the stability of the polyester fiber production processes. Especially during the esterification stage, there are multiple types of strong nonlinear dependencies among many sensor variables, which has always been a challenge in non-diffusive system modeling and control. Therefore, an Edge-Specific Message Graph Convolution Network (ESMGCN) is proposed to achieve separate modeling of specific dependencies between each pair of sensors individually, and to describe more accurately non-diffusive polyester fiber production systems. In addition, to address the problem of different degrees of dependencies between different sensor community clusters in large-scale systems, a community graph structure generation method is proposed to construct the graph structure within and between sensor communities, respectively. Finally, a multivariate time series prediction model for the polyester fiber polymerization process is constructed. Experiments are conducted in a real large-scale industrial system, and the results verify the model’s effectiveness.