LSSVM Modeling For Fermentation Process Based On Dividing Stages

The fermentation process is a highly nonlinear and complex dynamic process. As the rapid growth in fermentation industry, to build a accurate and rapid mathematical model for fermentation process is an important issue to study on microbial fermentation. It’s mainly because a good mathematics model lays a foundation for the subsequent soft measurement and optimization work, which are able to monitor and optimize fermentation process, and then promote the level of fermentation industry. Therefore, the research on modeling for fermentation process has actual meaning and application values.The internal mechanism of fermentation process is very complex and the fermentation kinetics mechanism of some fermentation products is not clear. The mechanism model is difficult to require. The black-box modeling methods are generally used to establish a data-driven empirical model for the fermentation process. Least Squares Support Vector Machine (LSSVM) modeling has high prediction accuracy, fast modeling speed and strong learning abilities. But when using LSSVM to establish fermentation process model, we found the following problems:(1) The fermentation process often experiences several stages. Global modeling of LSSVM treats entire fermentation process as a study object. It is difficult to guarantee the prediction accuracy.(2) Due to the values of variables are different and the correlation between variables changes with stage transition, one kernel may or may not built the optimal space mapping relation and space construction of all local models of stages when mapping learning data into higher feature dimension. The above problem further causes the prediction accuracy become lower due to learning algorithm model is limited by the specific hypothesis space. Therefore, this paper does some research on staged modeling and LSSVM with optimal mixtures of kernels.Firstly, we introduce LSSVM and mixtures of kernels. Secondly, we study on staged modeling method and Fuzzy C-means (FCM) clustering algorithm which’s commonly used to divide stages and give the basic steps of staged modeling. The staged modeling is validated based on Pensim which is a standard simulation platform of penicillin fermentation. Thirdly, LSSVM with mixtures of kernels is applied into the modeling for fermentation process for its good performances, so we propose the staged modeling of LSSVM with mixtures of kernels and based on that, staged modeling based LSSVM with optimal mixtures of kernels is proposed. First of all, sample data is classified using FCM clustering algorithm and fermentation process is divided into several stages. Then, local models of stages are built using LSSVM and HGA is employed to find optimal kernels from the given four types mixture of kernels and parameters of models. In the end, the integral process model is constituted with these local models. This method is preliminarily validated by being applied to model the penicillin fermentation process. In the end of this paper, we design the Protein Fermentation Development System including database and interface design and this system is implemented based on Visual Studio2008and SQL Server2005. Based on the above system, the LSSVM model for recombinant E.coli fermentation process in the interleukin-2production is developed to further improve the effectiveness of the proposed method. Compared with the global modeling, the staged modeling method has higher predicting accuracy and lower time complexity which proves that the proposed modeling method can provide a superior model for fermentation process. It has important reference value for development and production of other engineered bacteria.