Research On Soft Sensor Modelling Of The Glutamate Fermentation Process

Bioindustry,including fermentation food,fermentation chemicals,fermentation pharmaceuticals,fermentation energy et al,which is pillar industry of national economy.It is widely used in food,feed,medicine,chemical industry et al.Glutamate is the amino acid with the largest production in the world,it is produced chiefly by fermentation process.During fermentation process,the real-time access of some important biochemical parameters such as biomass,substrate and product concentration are of great significance for process optimization and control.However,the fermentation process depends on multiple environmental factors featured as intensity nonlinear,time-varying property and strong coupling,until now there are no practical on-line sensors that can be used directly to measure them in real time.In the industrial production,biochemical parameters are mostly measured through sampling analysis.Soft sensor is a technique that estimates the hard-to-measure variables using easily available process variables,which provides an effective way for solving the above problems.In the past decades,soft sensor technology has become a hot topic in process control fields.Also,a lot of successful applications have been reported in industrial applications.The dissertation was derived from the Natural Science Foundation of China(61273131)"Modeling and optimization of on-line support vector machine in bioreaction process".Taking glutamate fermentation process as research background,in combination with the mechanism of glutamate fermentation process,deep research and discussion on the soft sensor modeling methods and related problems are performed for the key biochemical parameters that are hard-to-measure on-line in glutamate fermentation process.Major results achieved in this work are highlighted below.(1)For the key variables in fed-batch glutamate fermentation can not be measured online,which makes it difficult to control and optimize,the non-structural dynamics models are built and an improved genetic algorithm is used to identify their unknown model parameters.First of all the non-structural dynamics models for fed-batch glutamate fermentation are built based on Logistic model,Luedeking-Piret equation et al and the nonlinear programming,genetic algorithm and an improved genetic algorithm are used to identify the unknown model parameters respectively.The non-structural dynamics models can predict the key parameters in typical biochemistry reaction like cell,substrate and product concentration.And the validity of the models are verified through glutamate fed-batch fermentation experiment.(2)Aiming at the problem of the dynamical models with batches,poor prediction accuracy and mechanism modeling difficulty,a novel multi-phase support vector regression(MPSVR)based soft sensor model is presented for online quality prediction of glutamate fermentation process on the basis of inherently multi-phase features.The glutamate fermentation process can be divided into a sequence of several phases through combining moving window with Pearson correlation coefficient,the partition results are consistent with off-line assay.For each estimation phase,support vector regression soft sensor models are constructed and their performances are evaluated against fermentation data.The efficiency ofthe proposed soft sensor model for online product quality prediction has been demonstrated to be superior compared to that of reported techniques.(3)Aiming at the problem of support vector regression model with long training time,in addition,coupling problem of influence factors in glutamate fermentation process,based on the analysis of least square support vector machine theory,a prediction model based on the partial least squares and least square support vector machine is established to predict glutamate concentration online.First,correlation coefficient matrix is applied to analyze correlation of the input variables,the results indicate that strong correlations are existed among the input variables.Then variance inflation factors are applied to diagnose the co-linearity of variables,the results indicate that medium co-linearity are existed among the input variables,so it is necessary to filtrate correlated variables.Based on this,partial least squares is used to extract the features of the input variables to reduce the number of the variable dimensions and eliminate the correlations,through these to reduce model complexity and improve model performance.Furthermore,coupled simulated annealing is applied to optimize parameters of least square support vector machine.The results show that the proposed model has good prediction accuracy,it can offer an effective guidance for control and optimization of glutamate fermentation process.(4)Aiming at the problems of the support vector regression and least square support vector machine models with over-fitting,parameters setting issues and uncertain problems of prediction results,a self-correlation decision Gaussian process based on feature relevance is presented to predict glutamate concentration online,which combine the input variables selection,hyper-parameter adaptive acquiring and outputs with probability characteristic.Firstly,applying Gaussian process to train and determining hyper-parameter of covariance function under the framework of Bayesian,subsequently,making use of it to predict glutamate concentration.Secondly,sensitivity analysis suggest that fermentation time,CO2 evolution rate CER and O2 uptake rate OUR are closely correlated with glutamate concentration.Thirdly,analyze change relations among the variances of predition values and input variables,when fermentation tank temperature,CER and OUR anomalous change,we find the variances of predicted values with significant changes,it can be as a pointer whether the sensors or fermentation process status are unnormal operation.Lastly,a soft sensor based on self-correlation decision Gaussian process is built to predict glutamate concentration online.Simulation results show that the proposed method can offer higher precision prediction and small confidence limit,it can meet the demand of real-time control for glutamate fermentation process.(5)It is a challenge task to achieve optimal control for fermentation production process without accurate model on the base of complex internal mechanism because glutamate fermentation is a complex biochemical process.By the mechanism analyzing of inputs and outputs for glutamate fermentation,the implementation scheme of biochemical parameter soft sensor system and optimal control project design for dissolved oxygen concentration are constructed.Furthermore,apply the soft sensor system to recognize abnormal batches.On the basis of RSLogix5000 of Rockwell software platform,the optimal control strategy system for glutamate fermentation is developed.The results of practical application indicate that thiscontrol system improves automatic level of glutamate fermentation production process,and reduces the labor intensity of workers and increases the economic efficiency.