Establish A LS_SVM-based Dynamic Model Of The Fermentation Process And Its Parameters Optimization

The microbial fermentation is a typical batch-based technical-intensive industry,involving the pharmaceutical, food, fodder, chemical, environmental governanceand other industries, and it has a more and more important role in the economicdevelopment. With the rapidly development of the above industries as well as thecompetition demand for more varieties, small batch and high-quality, there is anurgent requirement for fermentation process control and its optimization.However, compared with the general physical and chemical processes, thefermentation process has very different kinetic characteristics, the kinetic model ishighly nonlinear and strongly time-varying, most biological variables are difficult tomeasure online. What is more, the measurement of biomechanical parametersusually has a big lag which can’t feedback control information in time. Therefore,the establishment of the high-precision prediction model that adapts to thecharacteristics of the fermentation process has crucial role of improving the overallperformance of the fermentation process, improving the product and the use ofmaterials.The development of the machine learning theory provides a new direction forthe modeling of the complex industrial process such as fermentation process.The fermentation process is strongly nonlinear, time-varying, high dimension,and its biomasses are difficult to measure on-line multivariable couplingcharacteristics, according to the analysis and research of existing fermentationprocess modeling, the idea of local learning is introduced into the least squaressupport vector machine and a new local modeling method is proposed based onSupervised Locality Preserving Projection (SLPP) and least squares support vectormachine (LS_SVM). The major research findings and innovation are as follows.1)The improvement of the LS_SVM that based on dynamic time warping(DTW).Change the length of the sliding window to be1to simplify the dynamic timewarping which will not only reduces searching time of the similar samples but alsoimproves the prediction accuracy.2)The selection of similar training sample set based on SLPP.We introduce a new similarity evaluation criterion that includes both the inputand output information to SLPP, then there comes a SLPP adapted to regression process. By constructing the projection matrix, the sample in the original space ismapped to a low-dimensional space, then use the Euclidean distance in theprojection space to search the samples with the highest similarity of the test sampleform the historical data to be the training sample set. The SLPP achieves theextraction of characteristic variables and improves the computing speed byeffectively reducing dimension.3)Selection method of model input variables and parameters sensitivityanalysis.By analyzing the mechanism and the simulation experiments, chooses the meansquare error (MSE) of the model as the evaluation index, then I analyzes the impactof different input variables and the kernel function on the modeling effect, andfinally select the RBF to be the kernel function. In the last by analyzing the impactof C and2on accuracy of the model, I determine the range of values of themodel parameters.4) Super parameters optimization method based on FLOO-CV.The cross-validation method is usually used to select the super parameters, byanalysis of cross-validation method, the Fast Leave One Out Cross-Validation（FLOO-CV）is proposed,which aims to minimize the LOO-CV error of the model.The computation amount of FLOO is only one-n of the normal LOO for there is onlyonce inverse operation (partial model number of training samples). The FLOO-CVmethod can guarantee the accuracy of the model while maintaining thegeneralization ability of the model.