Enhancing Industrial Soft Sensing via Optimized Message Passing in Spatial-Temporal Graph Neural Network

Soft sensor technology aims to address how to utilize simple and readily available process variables to predict key variables constrained by factors like technology, cost, and environment. Some graph neural network (GNN)-based soft sensor models constructed from static graphs yield poor performance as they overlook the dynamic coupling relationship among process variables. Moreover, during the message passing stage, the GNN incurs inevitable information loss due to the incomplete propagation process. To tackle these issues, we propose a soft sensor model based on optimized message passing in a spatio-temporal graph neural network (OMP-STGNN). The adaptive graph module is employed to capture the spatial correlations among process variables and dynamically adjust the edge weights between them, enabling the graph structure to adapt to data changes. By using a parallel computation linear gated recurrent unit (GRU) to capture the temporal correlations among process variables, it circumvents the issue of low parallel efficiency resulting from recursive processes. To prevent information loss due to the propagation of graph structure information to process variables, the message passing process is optimized, ensuring that the process variables after message passing possess more comprehensive information. Results from a real-world industrial process dataset demonstrate that OMP-STGNN displays higher prediction accuracy than the comparison model.