Research On Integrated Just-in-Time Learning Soft Sensing Modeling Methods

Soft sensor technology is an effective approach to realize online estimation of difficult-to-measure parameters,which plays an increasingly important role in the monitoring,controlling and optimization of production processes such as metallurgy,papermaking,pharmaceuticals,petroleum,and chemical industry.Since practical industrial processes often exhibit process characteristics such as nonlinearity,time-varying behavior,multi-phase/multi-mode features,it is difficult to obtain satisfactory results using traditional global modeling methods.Therefore,just-in-time learning,as a typical local learning modeling method,has gained popularity in the field of soft sensor modeling.This dissertation focuses on the just-in-time learning soft sensor modeling technology,aiming to develop high-performance ensemble just-in-time learning soft sensor models.The main research contents of the paper are summarized as follows:(1)The major characteristic of traditional just-in-time models is the single model configuration,which is difficult to adapt to all process states.To tackle this problem,an ensemble just-in-time learning soft sensor method based on diverse subspaces and similarities is proposed,which uses the LWPLS algorithm as the base learner.The proposed method stimulates the diversity of the just-in-time learning base models through a multi-modal perturbation mechanism that combines input feature perturbation and similarity perturbation and balance diversity and accuracy of LWPLS models based on the evolutionary multi-objective optimization method.The stacking ensemble learning strategy is adopted to achieve the fusion of the base models.The effectiveness of the proposed method has been demonstrated through two industrial processes.(2)Traditional JIT learning soft sensor methods mainly focus on using single similarity measure,which may lead to failure in dealing with complicated process characteristics.To tackle this problem,a selective ensemble just-in-time learning soft sensor method based on heterogeneous similarity is proposed.This method constructs a similarity function library by defining multiple similarity functions through using evolutionary multi-objective optimization to select similarity functions for establishing diverse and accurate just-in-time learning models.Finally,the stacking integration strategy is used to fuse base models and obtain the final forecast output.The effectiveness of the proposed method has been verified through an industrial process.(3)To solve the difficulty of online measurement of Mooney viscosity in industrial rubber mixing process,a selective ensemble just-in-time learning soft sensor method based on diverse subspaces and heterogeneous similarity measures is proposed.This method enhances the diversity of the JIT learning models through a multi-modal perturbation mechanism combining the similarity measure perturbation and input variable perturbation.The evolutionary multi-objective optimization method is used to obtain a set of just-in-time learning base models that meet both accuracy and diversity.Moreover,an adaptive hybrid mechanism integration strategy is used to achieve the fusion of base models.The effectiveness of the proposed method has been verified through experimental results.