A novel dynamic variational compensation network with tracking for quality prediction of multirate industrial processes

Abstract

Quality prediction holds significant importance in monitoring industrial processes, with soft sensors proving to be highly effective in this domain. However, industrial processes frequently exhibit multirate characteristics due to measurement and cost limitations. The characteristics lead to periodic missing and varying dynamics of variables at different sampling rates, further presenting substantial challenges to current soft sensor techniques. To tackle the obstacles, we propose a Multirate Dynamic Variational Compensation Network with Tracking (MR-TDVCN). Utilizing a generic preprocessor and dynamic variational inference, MR-TDVCN effectively captures and characterizes crucial and diverse temporal dynamics related to multiple sampling rates, enabling comprehensive dynamic modeling of inhomogeneous multirate data. Based on this, a feature prism dynamic compensation network is developed to process multirate sequences for local feature compensation and global temporal relationship correction hierarchically and progressively. This mitigates the information loss due to multirate sampling, providing richer and more holistic feature representations for quality prediction. Finally, a feature tracking strategy is customized for multirate processes to alleviate the label sparsity problem. MR-TDVCN demonstrates superior performance on the common debutanizer column dataset, outperforming existing models. It is further applied to the polyester esterification process dataset to address real-world multirate challenges.