Edge-Compressed and Energy-Optimized Dual-Domain Graph Network With Variational Gating for Industrial Soft Sensing

Data-driven soft sensor methods have emerged as powerful tools for addressing the challenges of directly measuring and modeling key variables in industrial processes. These methods benefit greatly from the strong feature extraction and nonlinear representation capabilities of deep learning. Recently, an increasing number of studies have explored the integration of graph neural networks (GNNs) into soft sensing tasks to model the complex relationships among process variables. However, there are three problems with existing GNN-based soft sensor models: 1) there may be redundant information in the edge features of graph structures, which interferes with the effective modeling of variable dependencies by graph neural networks, 2) how to obtain a graph structure that as accurately as possible describes the real coupling relationship between variables from a data-driven perspective, and 3) in the face of graph structures expressing different physical meanings, how to reasonably fuse their information and effectively utilized. To address the above problems, an edge-compressed and energy-optimized dual-domain graph network with variational gating (D2GN-VG) soft sensor model is proposed in this paper. Specifically, for 1), the improved JPEG compression algorithm that is more suitable for graph structures will be used for edge feature compression. For 2), starting from the perspective of frequency-domain energy, the graph structure construction is transformed into an optimization problem to obtain a more accurate graph structure. For 3), a variational gated dual-domain fusion mechanism is proposed to effectively fuse information from different graph structures. Finally, the algorithm and workflow of D2GN-VG were summarized, and the effectiveness and superiority of the proposed soft sensor model were verified in industrial process datasets collected from the real world.