Dynamic Graph Learning Soft Sensor in Process Industry

The development of an accurate soft sensor modeling method in the process industry remains a great challenge because the coupling relationship between variables is always intricate and difficult to model. In this work, a dynamic graph learning (DGL) soft sensor is proposed to alleviate this problem. The proposed model realizes the ability of the soft sensor to perceive the coupling relationship in real time by automatically learning the dynamic graph. Then, a causal convolutional mechanism and a multi-hop graph attention mechanism are used to systematically construct the dependencies of variables in the spatial-temporal dimension and model their variation patterns effectively. Finally, the proposed method is tested on the penicillin fermentation process and shown to be feasible and effective. The results showed that the change of the dynamic graph in the spatial-temporal dimension was in line with the process mechanism.