Mixture Model Based Soft Sensor Development And Applications In Fermentation Processes

Soft sensor technology is an effective method to realize online estimation of key process variables which are difficult to measure or immeasurable.Biological parameters are important variables in the process of industrial fermentation.It is difficult to accurately describe their process characteristics with a single model.Currently,mixture model based soft sensors have made a lot of achievements in the field of complex industry,which have higher prediction accuracy compared to single soft sensor models,but the problems of the mixture model structure and parameter optimization are still need to be solved.This paper focuses on the modeling methods of mixture model based soft sensors on the simulation data of a Penicillin fermentation process and industrial data of an Erythromycin fermentation process.The specific work in this paper is as follows:(1)Many single models based soft sensors(such as Partial least squares,Artificial neural network and Support vector machine)cannot guarantee prediction performance because of industrial process characteristics of non-linearity,coupling,dynamics and uncertainty.In this work,a novel mixture model based modeling method using Gaussian process regression(GPR)and principal component analysis was proposed to construct a soft sensor.In this method,principal component(PCs)extracted from original process data are used to build GPR based sub-models.Compared to single soft sensor models,not only the proposed model has better prediction accuracy,but also can show the confidence interval.(2)Traditional soft sensor models generally require process objects to satisfy the Gaussian distribution hypothesis and are limited to a single simple industrial process.However,a real industrial process is always with multiphase and multimode characteristic.So,Gaussian distribution assumption is not met,which may result in prediction error of traditional soft sensors.In this paper,a soft senor based on Gaussian mixture regression(GMR)model was proposed,which uses a finite number Gaussian components to describe the non-Gaussian distribution.Simulations show that GMR is superior to GPR model in prediction accuracy and prediction confidence interval.(3)As we all know,it is difficult to determine the number of mixture components and initialize model parameters of mixture models.In this paper,affinity propagation(AP)method and its typical variant,adaptive affinity propagation(adAP)method,were introduced to preprocess data for the GMR based soft sensor modeling,which have ability to automatically partition data space.In order to obtain the optimal components,the clustering results are taken as the initial value of the GMR model Simultaneously.Compared to GMR,simulations indicate that AP and adAP methods are both suitable to partition data for GMR models,and the latter performs better in the soft sensor modeling.