Novel Just-In-Time Learning-Based Soft Sensor Utilizing Non-Gaussian Information

Abstract

This brief develops a novel just-in-time (JIT) learning-based soft sensor for modeling of industrial processes. The recorded data is assumed to exhibit non-Gaussian signal components, which are extracted by a non-Gaussian regression (NGR) technique. Unlike previous work on JIT modeling which uses distance-based similarity measure for local modeling, this brief introduces a new similarity measure for the extracted non-Gaussian components using support vector data description. Based on the similarity measure, a JIT modeling procedure called NGR_JIT is proposed. Application studies on a numerical example as well as an industrial process demonstrate the proposed soft sensor can give better predictive accuracy when the predictor and response sets are non-Gaussian distributed.