Domain adaptation graph convolution network for quality inferring of batch processes

Abstract

Developing a reliable quality inference model in batch processes remains a challenge. The performance of the model tends to decline due to batch-to-batch variations. Additionally, the spatial coupling relationships among multiple variables need to be mined to explain the model behavior. To this end, a domain adaptation graph convolution network is developed for quality prediction of batch processes. Under the constraint of the structural loss function, the model is firstly encoded to capture the topological correlations between variables. Moreover, the production time information is added into input variables to track the process temporal characteristics. Subsequently, the cross-batch characteristics are explored by the finetune technology to transfer the source knowledge and enlarge the prediction domain. Compared with several popular data-driven candidates, the proposed model shows its superiority in two batch processes. Remarkably, the variable relationships self-captured by the source model are visualized, which are consistent with the process prior knowledge to verify the model interpretability.