Just-in-time learning based soft sensor with variable selection and weighting optimized by evolutionary optimization for quality prediction of nonlinear processes

Abstract

Just-in-time (JIT) learning based soft sensors have been widely used for predicting product quality variables of nonlinear processes. They dynamically build online local models by selecting the samples most relevant to the query data from a historical database whenever an estimate is requested. However, building high-performance JIT soft sensors remains challenging due to difficulties defining similarity measures and the selection of input variables for facilitating efficient relevant sample selection and model building. In this study, we propose a novel soft sensing framework, referred to as JIT learning with variable selection and weighting (JIT-VSW). In this framework, a mixture weighted similarity (MWS) measure is defined by combining multiple weighted Euclidean distance (WED) based similarity measures. The MWS measure enables variable weighting embedded in WED measures to account for the relevance between input and output variables and facilitates the handling of highly complex process characteristics through the mixture-type similarity measure. Meanwhile, a wrapper optimization approach using evolutional algorithms is proposed for input variable selection. Further, the selection of input variables and the determination of MWS parameters, i.e., weights assigned to input variables and mixture coefficients of WED similarity measures, are formulated as a mixed integer optimization problem and solved simultaneously by using the mixed integer genetic algorithms (MIGA). The effectiveness and superiority of JIT-VSW are verified through three real-world applications.