Incremental Support Vector Mchine Regression Training Algorithm And Its Applications In Control

This thesis studies the incremental training algorithm of support vector machine (SVM) and its application in process control. SVM, which was put forward by Vapnik et. al, is a new and outstanding learning machine. It's the kernel content of statistical learning theory, and is a valid machine-learning tool in dealing with small samples. SVM overcomes some shortcomings of neural network, such as slow convergence, unstable solution, and bad generalization. So it has been widely applied to many areas, such as pattern recognition, signal processing, automation, communication, etc. SVM has become a hotspot of research in the area of machine learning.Now some famous SVM training algorithms have gotten extensive applications, such as chunking algorithm, decomposing algorithm, and SMO algorithm, etc. But these algorithms are all offline ones, which couldn't be used online. There are few studies on online SVM training algorithms. Based on previous research results, incremental SVM training algorithm is deeply studied in this thesis. An advantage of the incremental algorithm is that it can be used to train SVM model online, which largely extends the application area of SVM.The major contents of this thesis are as following:First, some improved methods are proposed to overcome the shortcomings of regular incremental SVM training algorithm (e.g, regular algorithm has a very slow training speed, and algorithm is invalid when margin support vector set is empty), which highly improves performance and practicability of SVM. Simulation results show the validity of the improved algorithm.Secondly, the improved incremental SVM training algorithm is applied to the area of process control. Some new and effective control algorithms are developed by combining the improved incremental algorithm with several classical control algorithms such as straight inverse control, internal model control and predictive control. Simulation results show the validity of these new control algorithms.Finally, the problems on regular SVM model are analyzed, and a weighted SVM model and its incremental training algorithm are proposed to get over these problems. A new method to remove redundant samples is also discussed. Simulation results show the validity of the weighted SVM model and its incremental training algorithm.