Self-adaptive graph convolution networks with application to industrial soft sensor modeling

In recent years, deep learning has been widely studied in soft sensor modeling. However, the prediction of the deep learning model is difficult to explain, and it is hard to incorporate prior experience into the model. These shortcomings of deep learning prevent its application in real industrial processes. In this article, we propose a self-adaptive graph convolution networks (SAGCN) for industrial soft sensor modeling. This model uses the graph convolution network to introduce prior knowledge and construct the displayed nonlinear relationship among variables. And the graph convolution network can aggregate information to extract features from data. Because it is difficult and highly subjective to rely on prior knowledge and mechanisms to obtain the graph structure, this article proposes a graph structure self-learning method to realize the joint learning of the nonlinear relationship among auxiliary variables and the regression relationship between auxiliary variables and quality variables. The proposed method is verified through the CO₂ absorption column process from a real ammonia synthesis process. Based on the results, SAGCN demonstrates high accuracy and a certain capacity to discover knowledge.