Soft Sensing And Feeding Predictive Control Of Marine Alkaline Protease Fermentation Process Based On MW-ELWPLS

Marine alkaline protease is a new type of protease developed based on natural enzyme-producing strains and modern biotechnology.Since the enzyme-producing strain grows in the hard marine environment with high salinity,low temperature and high pressure,the marine alkaline protease is superior to the traditional terrestrial alkaline protease in terms of oxidation,compatibility and catalytic properties.It has been widely used in the fields of detergent,medicine,food national defense,and many others.The actual fermentation process of marine alkaline proteases is a complex nonlinear,strongly coupled,and large time-varying system.In order to maximize the product yield and quality,the fermentation process needs to be dynamically regulated and optimized in real-time using advanced control technologies.However,there are three main problems in the fermentation process optimization control of marine alkaline proteases:(i)The internal mechanism of the fermentation process is very complex,and it is difficult to establish a kinetic mathematical model that can accurately describe the strain growth and substrate metabolism;(ii)The key parameters that can characterize the current fermentation state of marine alkaline proteases(such as bacterial concentration,residual sugar concentration,relative enzyme activity,etc.)lack effective online measurement means.Currently,they are mainly analyzed by manual offline sampling,but this method has a large measurement delay and is difficult to detect in real-time;(iii)Traditional fermentation optimization control usually relies on manual experience or simple PID control,but PID control often cannot adapt to the fermentation process with strong nonlinearity,which can lead to low yield and quality of the product.Data-driven model predictive control(MPC)is a better solution.To address the above problems,this paper focuses on the soft measurement of key parameters of the marine alkaline protease fermentation process and the method of model predictive control for feeding model to carry out research on the optimal control of the Marine alkaline protease fermentation process.The specific research contents are as follows.Firstly,the Partial Least Squares algorithm(PLS)and the real-time learning algorithm are introduced,and then the Local Weighted Partial Least Squares(LWPLS)algorithm is introduced.Considering that the LWPLS algorithm will computationally find the entire historical data set when selecting samples,which leads to excessive computation,a sliding time window(MW)is used to narrow down the selection of similar samples.To further improve the prediction accuracy of the algorithm,data with different query point locations under the same window length including query sample points are selected as sub-datasets,and the cumulative similarity of each sub-dataset is calculated,and then the sub-datasets with diversity are filtered by the set cumulative similarity threshold to establish the MW-LWPLS sub-model.Finally,the weighted average method in ensemble learning is used to fuse each MW-LWPLS sub-model.The feasibility of the proposed MW-ELWPLS algorithm is verified by numerical simulation.Next,a soft measurement model based on MW-ELWPLS was constructed and applied to the soft measurement of key parameters of marine alkaline proteases.By analyzing the fermentation process of marine protease and the influencing factors among the parameters,the soft measurement model of marine alkaline protease key parameters was established by combining the nearest neighbor mutual information method to select the auxiliary variables of the soft measurement model.The proposed soft-measurement model of MW-ELWPLS was experimentally validated and results show that the prediction accuracy is higher by about 70% compared with PLS models,and it also paves the way for the predictive control of the feeding.Finally,aiming at the problem that it is difficult to control the flow acceleration rate in the marine alkaline protease fed batch fermentation process,the proposed MW-ELWPLS algorithm is applied to the prediction model,and a nonlinear multi-step fed batch predictive control model based on marine alkaline protease fermentation process(MW-ELWPLS-MMPC)is established.Since the conventional online rolling optimization generally adopts the Steepest descent(SD)method,which leads to the local optimum phenomenon when solving the optimal control sequence in rolling optimization,the quantum particle swarm(QPSO)bionic optimization algorithm is used to improve the rolling optimization.The soft measurement model in Chapter 3 is also applied to the predictive control of the feeding model to construct a marine alkaline protease feeding control system.The feeding experimental simulation of marine alkaline protease shows that the method has a good tracking effect on the actual optimal curve.