Soft-sensor Modeling And Model Predictive Control For L-lysine Fermentation Based On Machine Learning

The L-lysine fermentation process is a complex non-linear,strongly coupled,and large timevarying system.To maximize the production efficiency and product quality of L-lysine,it is necessary to use advanced control technology to dynamically regulate and optimize the L-lysine fermentation process in real-time and to control it under the best process conditions.However,the complex characteristics of the L-lysine fermentation process cause difficulties in optimal operation and control due to the growth,reproduction and metabolic process of the strain.Some problems involved in the process are:(1)The internal mechanism of the L-lysine fermentation process is very complex and the kinetic model is highly non-linear,so it is difficult to establish a mathematical model that can accurately describe the kinetic characteristics of the strain growth and metabolic process;(2)The kinetic parameters of the fermentation process vary with the reaction time or batch,so the traditional kinetic-based control method can hardly meet the control performance requirements;(3)The key variables such as the concentration of cell,substrate and product,which reflect the quality of L-lysine fermentation process,are still lacking effective online measurement means,which has become the bottleneck of the optimal operation control.At present,these key variables are mainly measured by offline sampling and analysis,which has a large measurement delay and is difficult to implement real-time control,while the fermentation broth is easily contaminated by miscellaneous bacteria during the sampling process,which seriously affects the L-lysine production process.The optimized operation control of the L-lysine fermentation process is mainly through the detection of some environmental variables(such as dissolved oxygen concentration,p H,temperature,etc.)and manual operation by workers’ experience and senses,which has a low level of automation and backward technology and cannot meet the requirements of optimized operation control of Llysine fermentation process.Therefore,how to overcome the defects brought by the above problems has become one of the urgent problems to be solved in the L-lysine industry.In this work,with the support of National Natural Science Foundation of China(41376175),the Priority Academic Program Development of Jiangsu Higher Education Institutions(PAPD),Zhenjiang city key R \& D projects-social development projects(SH2020005),we focus on the soft measurement of key variables in L-lysine fermentation process,non-linear model predictive control method,etc.,to carry out research on the optimal operation control of Llysine fermentation process.The specific research of this thesis is as follows:The ICS-MLSSVM hybrid multi-output modeling method is proposed and applied to the soft measurement of key variables in L-lysine fermentation process to achieve online real-time detection of key biochemical variables(cell concentration,substrate concentration and product concentration)in L-lysine fermentation process.First,a multi-output least-squares support vector machine regressor(MLSSVM)model was constructed based on the multiple-input and multiple-output characteristics of the L-lysine fermentation process.Then,the important parameters($\gamma$,$\lambda$,$\sigma$)of the MLSSVM model are optimized using an Improved Cuckoo Search(ICS)optimization algorithm.Finally,the ICS-MLSSVM soft-sensor hybrid model was developed using the optimized model parameter values to measure the key biochemical variables of the L-lysine fermentation process online.The simulation results confirmed that the proposed regression model could accurately measure the key biochemical variables.In addition,the hybrid ICS-MLSSVM soft-sensor model outperformed the MLSSVM soft-sensor model based on standard CS(CS-MLSSVM),Particle Swarm Optimization algorithm(PSO-MLSSVM)and Genetic Algorithm(GA-MLSSVM)in terms of prediction accuracy and adaptability.A non-linear Model Predictive Control(NMPC)method is proposed for the problem of difficult real-time control of acceleration rate flow during L-lysine fermentation.A non-linear model predictive control model of L-lysine fermentation process based on machine learning algorithm(least squares support vector machine-LSSVM)is established and LSSVM is used as a process model to predict the product concentration in the non-linear model predictive control.The key parameters(penalty factor and kernel width)of the LSSVM predictive model are optimized using the Gray Wolf Optimization(GWO)algorithm to improve its prediction accuracy.GWO is also used to solve the non-convex optimization problem in nonlinear model predictive control(GWO-NMPC)for calculating optimal future inputs.Finally,the GWObased LSSVM prediction model(GWO-LSSVM)and non-linear model predictive control(GWO-NMPC)methods are compared with the PSO-based prediction model(PSO-LSSVM)and non-linear model predictive control(PSO-NMPC)methods in combination with the soft measurement model of key variables to verify their effectiveness.The comparison results show that the prediction accuracy,adaptability,real-time tracking capability,overall error and control accuracy of GWO-based predictive control are better compared to PSO-based predictive control.