MapcQtNet: A Novel Deeply Integrated Hybrid Soft Sensor Modeling Method for Multistage Quality Prediction in Heat Treatment Process

In multistage industrial processes, critical quality variables are often difficult to measure in real time. Thus, a common solution is to measure them indirectly in a quality prediction manner through soft sensor methods. Most existing networks for multistage quality prediction focus on the overall transfer of process information between stages but do not consider exactly how the quality information is transferred. In addition, how to introduce mechanisms in data-driven models to maximize the use of prior knowledge and improve the models’ interpretability is another problem that needs to be addressed. Considering the above two problems, a novel deeply integrated hybrid soft sensor modeling method is proposed for multistage quality prediction, known as the mechanism-aware progressive constraint based quality transfer network (MapcQtNet). MapcQtNet consists of two key blocks: quality transfer units (QTU) and mechanism-aware progressive constraint (MAPC). With the addition of two quality transfer gates, QTU can simulate the flow of quality information between stages in more detailed ways. In addition, MAPC innovatively integrates the prior mechanism formulas into the network in a constrained manner, which helps the network to be aware of the process mechanism. With it, MapcQtNet can not only enhance its interpretability but also gain the ability to achieve unlabeled predictions for the intermediate variable and uncertain mechanism parameter. A real industrial case of the heat treatment process verifies the validity of the proposed MapcQtNet as an advanced soft sensor modeling method for multistage quality prediction.