Dynamic data reconciliation for enhancing the performance of kernel learning soft sensor models considering measurement noise

Abstract

In modern industrial processes, data-driven soft sensor models avoiding the limitations of measurement techniques and expensive costs are developed for process monitoring and quality prediction. However, historical datasets usually contaminated by measurement noise reduce the reliability of model prediction. To alleviate the negative effects of measurement noise on process modeling, dynamic data reconciliation (DDR) is developed to enhance the prediction performance. First, based on the kernel learning framework, the modeling methods of kernel partial least squares (KPLS) and just-in-time kernel learning (JKL) are designed to construct soft sensor models. Then, DDR is combined with the KPLS-based and JKL-based soft sensor modeling to provide improved datasets. By combining information from model predictions and measurements, DDR obtains useful information through Bayesian estimation. Finally, a numerical example and a polymerization process show the negative effects of measurement noise on kernel learning soft sensors are weakened by DDR.