Multi-source soft sensor model for industrial processes based on graph convolution and dual-constraint adversarial domain generalization

Abstract To address the issue of poor generalization performance of soft sensor models owing to multimode characteristics within process data and insufficient labeling, this paper proposes a Multi-Source Graph Convolution and Dual Constraint Adversarial Domain Generalization (MSGCN-DADG) soft sensor model. Multimode data constitute multiple source domains, and treat the process variable in the source domain as a node in the diagram structure. By updating the graph convolution network with self-loops and adjacency information of nodes, multi-source domain features are obtained. Domain adversarial networks are employed to align these features, generating domain-invariant features common to all source domains. The proposed method improves the domain adversarial loss function with a dual constraint regularization using Maximum Mean Discrepancy (MMD) to reduce the differences among source domain features and between the source domains and the prior distribution. A fully connected layer is used to establish a model between domain-invariant features and source domain labels, and this model is generalized to the target domain through domain generalization. The proposed model is validated on two real industrial processes, Multiphase Flow Process (MFP) and Thermal Power Plant (TPP) are compared with several domain adaptation and domain generalization methods. Experimental results verify that the proposed method exhibits better prediction accuracy and generalization ability.