摘要
In the process industry, data-driven soft sensors enhance transparency by mapping the interdependencies among variables triggered by complex reactions. Yet, they frequently fail to account for the global dependencies that span across spatial and temporal dimensions. To tackle this gap, we propose a global dependency graph network (GDGN) soft sensor, informed by both prior knowledge and process data. Initially, it estimates the Bernoulli distribution to gauge the probability of dependency among variables, followed by a random sampling to form the dependency graph. Importantly, the integration of even partially available prior knowledge can bolster the physical accuracy of this graph. Subsequently, GDGN applies a self-attention mechanism (SAM) to craft the global dependency within set constraints, utilizing graph convolution for quality variable prediction. Ultimately, it elucidates the modeling and prediction process. The efficacy and physical coherence of the GDGN's prediction logic were validated through two distinct case studies.