Novel Multimode Process Soft Sensing Methods Based on the Dynamic Mixture Variational Autoencoder Regression Model

Abstract

Modern industrial processes with increasing complexity not only contain nonlinear and multi-mode characteristics, but also are commonly the dynamic processes, which brought challenging problems to soft sensor modeling. In order to solve these problems, a dynamic mixture variational autoencoder regression (DMVAER) model is proposed for the nonlinear multi-mode modeling, which is suitable for industrial process quality prediction with multiple complex process characteristics. Furthermore, in order to deal with the problem of semi-supervised data with a large number of unlabeled samples, a semi-supervised dynamic mixture variational autoencoder regression (ssDMVAER) model is proposed, and the corresponding semi-supervised data sequence division method is adopted to make full use of the information in both labeled data and unlabeled data. Finally, in order to verify the feasibility and effectiveness of the proposed methods, the two models are applied to an actual industrial process of methanation furnace. The results show that the proposed methods have superior soft sensing performance than existing methods.