Domain transfer spatiotemporal graph network for predicting key indicators in multimode chemical processes

Abstract

In modern chemical industries, soft sensors related to key indicators are crucial in production processes, particularly for multimode chemical processes. However, challenges arise from scarce labeled data and constant process shifts, which complicate the development of these sensors. Additionally, complex process variable interactions hinder prediction interpretability. This work develops a domain transfer spatiotemporal graph network (DTSGN) regression model, tailored for cross-domain modeling in multimode chemical processes. The DTSGN model first employs graph convolution to capture spatial dependencies and gating mechanism to track temporal relationships. A Bernoulli-distributed perturbation graph enhances model robustness. By embedding constraints into conditional distribution discrepancy, the parameters for DTSGN are calibrated under the fine-tuning paradigm for new mode. Therefore, the model extends its applicability to address process shifts in multimode chemical processes. A graph explainer finally ensures interpretability by aligning learned variable graph with process knowledge, and two multimode cases evidence the proposed model’s efficacy and superiority.