Research And Application On Soft Sensor Modeling Method Of Difficult-to-measure Parameters In The Industrial Process

At present,many key variables of industrial process can only be measured through offline laboratory analysis with huge delay,which is detrimental to real-time monitoring,control and optimization.In recent years,with the wide application of distributed control system and information management system,large amount of process data can be preserved,which makes data-driven soft sensor technology to be a breakthrough for real-time monitoring of the difficult-to-measure variables.However,the traditional global modeling method can not meet the expectation due to the large lag,nonlinear,time-varying,multimode/multiphase characteristics of chemical(or biochemical)engineering process.Therefore,local learning strategy becomes an effective way to solve such problems,such as multi-model modeling,integrated leaning and just in time learning.The main work of this dissertation is to build soft sensor based on local learning strategy considering the process characteristics,then verify with the actual industrial process data.Summarized as follows:(1)To address the nonlinear,multimode and multiphase characteristics of industrial process,an improved ensemble soft sensor based on sample partition method is proposed.Firstly,the history samples are divided into multiple regions by improved Gaussian mixture model;then,local Gaussian process regression models are built based on each local region.During the online prediction,the posterior probabilistic of the test data belongs to every local region is estimated by using Bayesian inference strategy,and the final prediction result is obtained by every local model's output though the corresponding posterior probabilistic.Two ways to make the Gaussian mixture model more reasonable are proposed.First,set a wide range of initial values to avoid the problem that the traditional Gaussian mixture model is easy to fall into local extremum.Second,combine the similar Gaussian components to eliminate the redundancy and speed up the regression time.The effectiveness of the proposed approach is verified through the actual fed-batch chlortetracycline fermentation process data.(2)Traditionally,building local models of many available ensemble soft sensors are only based on sample partition,and the diversity of the process variables is ignored.To solve this problem,a new soft sensor modeling framework based on variable partition is proposed.First,multiple input variable sets are obtained through the Bootstrapping algorithm and partial mutual information(PMI)criterion.Second,a set of local models are built based on the historical samples corresponding each input variable set.During the online prediction,through the Bayesian inference strategy,partial good local models are chosen to ensemble.The proposed soft sensor is applied to the total sugar concentration prediction of chlortetracycline fermentation process,and the prediction results verify the effectiveness of the variable partition based ensemble frame.(3)Multi-model ensemble learning method has been widely used in the difficult-tomeasure variable prediction of industrial process.However,the traditional multi-model construction method is only based on manipulating samples or only based on manipulating variables,the diversities of samples and variables are not considered simultaneously.Therefore,a novel soft sensor based on the hierarchical ensemble of Gaussian process regression models is proposed.Firstly,multiple input variable sets are constructed by random resampling method and PMI criterion.Secondly,sample partition based ensemble Gaussian process regression models(SP-EGPR)are built based on the local historical samples corresponding to each input variable set.Then,the partial least square regression(PLSR)method is used to prune the redundant local SP-EGPR models.During the online prediction,through the Bayesian inference strategy and finite mixing mechanism to estimate test sample belongs to each local SP-EGPR,and choose partial good SP-EGPR to integrate.The proposed soft sensor is applied to Tennessee Eastman process and industrial rubber-mixing process,and the prediction result verified its effectiveness.(4)Since most industrial processes are time variant,the off-line modeling method will degrade once being applied to actual process.Therefore,a novel adaptive soft sensor based on just in time learning and ensemble learning is proposed.Firstly,multiple input variable sets are constructed by principle component analysis(PCA).The PMI of the input and output variable is added to the covariance function of the PCA,which can improve the accuracy of the variable partition result.When a test sample is coming,some similar historical samples are selected based on each input variable set;and then,local GPR models are built based on each set of historical samples;finally,the mixture prediction is obtained based on Bayesian inference strategy and finite mixing machine.The effectiveness of the proposed method is verified by applied to industrial rubber-mixing process.