Study Of SVM-based Soft Sensor Modeling Technology In Fermentation Process

Microorganism fermentation engineering is the foundation of bioengineering and biotechnology. Biomass concentration is an important variable in the fermentation process, which reflects the state of the fermentation process and impacts the quality and yield of the output. In order to control the fermentation process effectively, online measurement of metabolic product concentration is indispensable.Constrained by the sensor technology in fermentation, metabolic product concentration can not be measured by regular sensors. Because of the high price and high maintenance costs, the application of dedicated sensors is limited. With the development of computer technology, soft sensor technology, estimating unmeasurable variables or difficult-to-measure variables with measurable ones online, is applied in microorganism fermentation process. Recently, many methods have been used to model fermentation soft sensors, such as mechanism modeling, Kalman filter, multiple regression, artificial neural network and support vector machines (SVMs). SVMs have drawn extensive attention with their high precision and strong generalization ability using small samples.Based on simple analysis of soft sensor modeling for erythromycin fermentation process using SVMs, this paper proposed a approach for model optimization of SVMs based on genetic simulation annealing algorithm (GSA) and Akaike Information Criterion (AIC), which could select important fermentation auxiliary variables and set SVMs parameters simultaneous. Training a support vector machine requires solving a linearly constrained quadratic programming problem (QP). This problem often involves a matrix with an extremely large number of entries, which make off-the-shelf optimization packages unsuitable. In order to make the complex algorithm easy to implement in industry, particle swarm optimization (PSO) with high convergence is proposed to solve the problem, which could determine the support vector weight and support vectors of support vector machines in a simple way. Simulations show that soft sensors modeling methods for fermentation process based on SVMs could be applied more conveniently and efficiently with the proposed optimization methods.