Novel multimodal data fusion soft sensor modeling framework based on meta-learning networks for complex chemical process

Abstract

Data-driven soft sensor has been widely used in industrial processes. However, complex industrial processes all exhibit nonlinear and multimodal characteristics due to varying operating conditions. Multimodal data characteristics will cause the deterioration of soft sensor performance. Therefore, in this article, a modified Model-Agnostic Meta-Learning (MAML) based on K-Means (KM) is proposed. Firstly, the KM method is introduced to cluster the multimodal processed data, and then extract the clustered data to form multiple tasks. After that, MAML method is adopted to train a group of initialization parameters. The sum of each task's loss function is introduced to adjust the initial parameters by one or more steps of gradient. The proposed model is finally applied and verified in the Purified Terephthalic Acid (PTA) solvent system. Compared to some conventional methods, the prediction accuracy is improved by more than 70%. The result demonstrates the superiority in the proposed method.