ResearchonJust-inTimeLearningSoftSensorModelingMethodofTime-varyingIndustrialProcess

In some complex industrial processes,as product quality requirements continue to increase,there is a need for rigorous monitoring and controling of key variables that directly determine product quality.However,due to the high price of some measuring instruments or technical constraints,these variables cannot be measured by online instruments.For these difficult-to-measure variables,estimations and predictions can be made by establishing a soft sensor model.After the offline soft sensor model put into operation,the actual industrial production process often has time-varying due to catalyst activity deactivation and equipment aging.The working point of the production process will change,and the previously established global model will be invalid and can not be applied to the current working point.The predicted results cannot meet the accuracy requirements.Therefore,the Just-in Time Learning(JITL)strategy was adopted to overcome the shortcomings of the global model by continuously establishing local models.This paper is based on the JITL strategy for soft sensor modeling research.The specific research contents and results are as follows:For some time-varying industrial processes with non-Gaussian properties,a JITL soft sensor modeling method was proposed based on Bayesian Gaussian Mixture Model(BGMM).First,for the given training sample set,the number of components of the Gaussian mixture model was optimized by Bayesian Information Criterion(BIC);Then,for the new test samples,the BGMM similarity criterion was used to find the most similar set of samples from the training samples to establish a Gaussian process regression(GPR)model;Finally,the model was used to predict the test samples.The effectiveness and accuracy of the proposed method were verified by the soft sensor modeling of the butane concentration of debutanizer tower bottom.Considering the multi-stage characteristics of industrial processes,based on the BGMM-JITL single model,a multi-model modeling method is proposed.The method uses the GMM algorithm to cluster the training set.For the incoming test samples,BGMM similarity criteria were used in each category to select similar samples to establish multiple local GPR sub-models,and the test samples were predicted to obtain predicted values of multiple sub-models.Finally,using Bayesian posterior probabilities merge the predicted values of multiple sub-models.The prediction experiment of product concentration in penicillin fermentation process shows that the proposed method has higher prediction accuracy and reliability.Aiming at the problem of deteriorating the regression performance of reducing the just-in-time learning model update frequency,a just-in-time learning modeling method based on second-order similarity was proposed.The method considers the whole distribution characteristics of the sample set.It constructs a second-order similarity criterion based on the first-order similarity criterion,and selects similar samples which have the most same neighborhoods with the corresponding testing sample to build up a local model.The cumulative similarity factor was used to determine the data length for the local model,and the similarity threshold was used to judge whether the model needs to be updated at present.The simulation experiment of penicillin fermentation process and numerical example prove that proposed method can improve the accuracy of the model while reducing the time consumption of the model.