Gaussian Mixture Model-based Time-varying Soft Sensor Development

Process industries are seeking to ensure the safety of the process as well as the quality of the product. For both process monitoring and control, it is imperative to gather real-time information of the key process variables. However, there are some key process variables that are difficult to measure online due to the limitation of process and technology. Although the online analytical instrument may be installed to measure such variables, online measurement is not always able to meet the requirements of real-time control and optimization due to unacceptable expenses of analytical instruments or long measurement delay. To solve this problem, soft sensor is a key technology for estimating these important variables in the process. In last decades, data-based soft sensor modeling methods have already been widely used for the quick response, easy maintenance and low cost.Even though a good soft sensor is obtained, the prediction performance will deteriorate gradually due to the changes in process characteristics, such as catalyst deactivation, process drift and changes in the state of the chemical plant. To update models automatically when process characteristics change, different kinds of recursive modeling methods have been developed, which update models with new samples that reflect the process changes. In this paper, a novel relevant sample selection strategy based on Gaussian Mixture Model (GMM) is proposed for JITL soft sensor development to deal with the non-Gaussian as well as nonlinearity. The main contributions are described as follows:(1) In this paper, a new GMM-based similarity criterion has been proposed for relevant samples selection in JITL soft sensor modeling to deal with the non-Gaussian and time-varying process. By taking the non-Gaussianity of process data and the characteristics of the query sample into account, a more suitable similarity criterion is defined for sample selection of JITL soft sensor and better modeling performance can be achieved. Case studies on a numerical example as well as an industrial process are demonstrated to evaluate the feasibility and effectiveness of the proposed method.(2) The adaptive Gaussian Mixture Model (GMM)-based locally weighted partial least-squares regression (LWPLS) soft sensor modeling method has been promoted to deal with nonlinearity and time-varying. The improved similarity criterion is developed for relevant samples selection as well as sample weight, thus local weighted model can be built for online prediction. Case study on an industrial process demonstrated the feasibility and effectiveness of the proposed method while dealing with the process nonlinearity.Finally, this paper summarizes the research results, and expounds the difficulties and breakthroughs in the future research work.