Soft Sensor Framework Based on Semisupervised Just-in-Time Relevance Vector Regression for Multiphase Batch Processes with Unlabeled Data

Abstract

Soft sensors using just-in-time learning (JITL) have attracted much attention in the application of online prediction in batch processes because of the ability to perform adaptive updating and dynamic modeling. However, developing effective JITL-based soft sensors of batch processes remains challenging due to the unlabeled data caused by the expensive online measuring instruments and long time-consuming offline analysis. Besides, the multiphase and nonlinear characteristics of batch processes make this challenge more complicated. To cope with this challenge, a novel soft sensor framework, termed semisupervised just-in-time relevance vector regression (SJRVR), is proposed. The SJRVR integrates JITL, adversarial autoencoder (AAE), and RVR into a unified framework to address soft sensor modeling for multiphase batch processes with unlabeled data. In this framework, an unlabeled data processing strategy based on process mechanism and AAE (MAAE) is presented to utilize the useful information on unlabeled data. Moreover, a local regression model is constructed using JITL with a designed modeling data selection strategy and RVR to address the soft sensor modeling for process data with multiphase and nonlinear characteristics. The effectiveness of the SJRVR method is illustrated with a penicillin fermentation process.