On-line Calibration of Just in Time Learning and Gaussian Process Regression based Soft Sensor with Moving-Window Technology

Abstract

To handle time-varying, non-linear and multi-parameter characteristics of industrial processes, a new soft sensor modelling method by Gaussian process regression (GPR) with just in time learning (JITL) and moving window technology is proposed. Traditional soft sensors based on JITL only consider spatial characteristic of the query data point and select the best similar samples from a historical database for modelling, ignoring local temporal characteristics of industrial processes. That may result in some predictions relying too much on database. In the proposed soft sensor modelling method, firstly, JITL is used to build a GPR-based prediction model which gives output related to query data point. Then, a local temporal GPR-based model is built on the samples within the last given moving window. In the moving window, the prediction given by the JITL model is as the newest sample. Finally, the local GPR-based model is used to calculate output related to the query data point. This method takes into account not only spatial characteristic of a query data point but also local temporal characteristic of real-time process conditions. The proposed soft sensor is validated by an industrial Erythromycin fermentation process simulation. Results show that the proposed method has higher adaptability and predictive performance than traditional JITL based soft sensors.