Online Adaptive Semi-supervised Soft Sensor Modeling Method Combined With Active Learning Strategy

The stable and reliable control of product quality in modern industrial processes is extremely important.In view of the complex working conditions,limited detection technology and harsh environment in the actual industrial production process,considering that some key quality parameters often rely on manual assay,the perceptual data show sparse characteristic.Therefore,the Semi-supervised Learning(SSL)technology is needed to build the soft sensor model of product quality based on data-driven technology.However,SSL itself is easy to add redundant samples,and when the label data is scarce,which influence the performance of SSL greatly.Therefore,it is necessary to establish a semi-supervised soft sensor model combined with Active Learning(AL)strategy to guide manual assay for improving the quality of label samples and reducing the demand for label samples.In addition,the actual industrial process usually has the characteristic of slow time-varying,and the established soft sensor model should also have the characteristic of smooth self-updating.In order to solve the above problems,this paper chooses the Random Vector Functional Link Network(RVFLN)as the soft sensor model,which has rapid learning speed and better generalization performance,and carries out the research of online adaptive semi-supervised modeling method combined with the active learning strategy,trying to improve the quality of the soft sensor model in industrial process to the greatest extent with only few label samples.This major research work and achievements are as follows:(1)For the problems that the Semi-supervised RVFLN(SS-RVFLN)model is susceptible to redundant samples and lack of label data,this paper fully analyzes the impact on degree of sample information,sample representation and process nonlinearity on modeling,and introduces an active learning strategy based on the Comprehensive Evaluation Index(CEI).Furthermore,an Active Semi-supervised RVFLN(ASS-RVFLN)modeling method is proposed,and two different versions of offline learning paradigms for the model are given respectively.Finally,the simulation results show that the proposed method not only achieves high estimation accuracy,but also effectively reduces the acquisition burden of key quality parameters.(2)In view of the slow time-varying characteristic in industrial processes,this paper proposes an Online Adaptive Active Semi-supervised RVFLN(OAAS-RVFLN)modeling method with weight deviation constraint and Fuzzy Inference System(FIS)based on ASS-RVFLN,and two different versions of online learning paradigms for the model are presented.Finally,large numbers of simulation experiments prove that the proposed online modeling method is superior to the traditional online modeling method in estimation accuracy and learning speed.(3)The established OAAS-RVFLN model is applied to the Coal Dense Medium Separation(CDMS)process in coal industry for estimating the ash content online,which is an important operation index in production process.A sequence of simulation experiments prove that the proposed modeling method in this paper can effectively estimate the ash content of clean coal,and further verify its comprehensive performance and the potential of industrial application.The thesis inchlude 25 figures,15 tables and 89 references.