Research And Application Of Soft-sensor Modeling Technology Based On Data-driven

Data-driven based soft sensing technique encounters the following problems: large number of highly correlated variables, limited data samples, data subject to noise pollution, highly nonlinear and time-varying process and so on. This paper studies the soft-sensor modeling method based on Support Vector Machine (SVM) which is a kind of novel machine learning methods, theoretically based on statistic learning theory, and solves the above problems successfully. Main results and contributions of this paper are as follows:1. Because the performance of SVM has a close relationship with the kernel function type, kernel function parameters, as well as punishment coefficient C, and their different selection is directly related to generalization capacity of the SVM model. Since there are few analytical methods to choose the SVM parameters, an automatic parameters selection strategy based on Particle Swarm Optimization (PSO) algorithm is proposed. In this new method, each particle indicates a group of SVM parameters and k-fold cross-validation error is used as the fitness function of PSO. Simulations of real data show that the excellently global searching ability of PSO contributes the task of parameters selection greatly.2. In order to deal with the high nonlinearity of the industrial process, a hybrid Partial Least Squares and Support Vector Machines (PLS-SVM) method is proposed. This method has merits of both PLS and SVM, and thus enhances the ability to handle non-liner of model. Based on the research of the Bisphenol-A crystal tower mechanism, a soft-sensor model of the production process for Bisphenol-A (BPA) is created by this method. The simulation results show that the algorithm is feasible and effective.3. Because of limitations of fitting data capacity of SVM, single model can not extract the information from data well. To solve this problem, a combination of SVM modeling method based on fuzzy C-means (FCM) is proposed. This FCM-SVM method introduces a fuzzy C-means clustering method, then through grouping of data training, and sets up multiple sub-models, so as to saves a lot of model training time. Finally the degree of membership is used for combining several sub-models to obtain the finial result. The simulation results prove that the method has good practical value, multi-model also improve the utilization of the network robustness.