Research On Dynamic Modeling Method Of Fermentation Process Based On MKSVM And Its Application

Microbial fermentation process is a kind of very complex biochemical reactionprocess, with time-varying, nonlinear, uncertain, multivariable couplingcharacteristics, humans yet not fully understand its mechanism. And biological stateparameters of fermentation process are obtained by off-line analysis, which often hasa big lag, and is unable to feedback control information timely. Besides, each batch offermentation process has the difference and uncertainty, currently the establishedmodels based on SVM are pre or off-line model of the fermentation process, and onceproduction conditions change, existing models may not be able to adapt to the newenvironment conditions inevitably.The development of the machine learning theory provides a new direction for themodeling of the complex industrial process such as fermentation process. Accordingto the analysis and research of the existing fermentation process modeling, a newdynamic modeling method was proposed based on Multiple Kernel Support VectorMachine(MKSVM). The major research findings and innovations are as follows:(1) The establishment of the similar sample set for fermentation process. First,based on the analysis of the affinity, the affinity was introduced to the weighedEuclidean distance; Second，the weighted Euclidean distance based on the affinity andDTW were combined to select similar samples，then the similar training set wasconstructed.(2) The comparison of the similar sample selection methods. The similar sampleselection method based on the weighted Euclidean distance and DTW was used toestablish fermentation process dynamic model based on SVM. Compared with theSVM dynamic models based on other similar sample selection methods, the modelhad a higher prediction accuracy and the prediction time had also been shortened.(3) Construct the multiple kernel function. By analyzing and verifying the natureof common kernel function, the local kernel function and global kernel function wascombined linearly to construct the multiple kernel function, and then it was used tosupport vector machine fermentation process modeling. Simulation experimentsverified the advantage of MKSVM in the fermentation process modeling. (4) Use Improved Adaptive Genetic Algorithm(IAGA) to select MKSVMparameters. Through the analysis of the existing AGA, IAGA was proposed bymodifying the crossover and mutation probability. Simulation results showed that thismethod can guarantee the accuracy of the model while maintaining the generalizationability of the model.