Soft sensor development using non-Gaussian Just-In-Time modeling

Abstract

This paper introduces a novel Just-In-Time (JIT) learning based soft sensor for modeling of non-Gaussian process. Most of JIT modeling uses distance based similarity measure for local modeling, which may be inappropriate for many industrial processes exhibiting non-Gaussian behaviors. Since most of industrial processes are non-Gaussian, a non-Gaussian regression (NGR) technique is used to extract non-Gaussian independent components that are correlated to response variable in the sense of mutual information. Support vector data description (SVDD) is then performed on the extracted independent components to construct a new similarity measure. Based on the similarity measure, a novel JIT modeling procedure is proposed. Application studies on a numerical example as well as an industrial process confirm that the proposed JIT model can achieve good predictive accuracy.