The Local SVR And Its Applications In The Prediction Of Spatiotemporal Chaos Sequence

The EEG signals are very complicated signals and serve as windows for us to understandthe cerebral activities because these signals are the synthetical reflection of the electricityactivities of cerebral tissue and brain function status. In fact, EEG signal has the characteristicsof chaos i.e. EEG signal is a nonlinear spatiotemporal chaos sequence. The theory of analyzingchaos can be used in the analyzing EEG. It provides useful information for clinic applicationsuch as cerebral disease diagnosis and disease prediction.Resent years, with the development of chaos theory and application technology study,analysis and prediction of chaotic time series have become a hot point of chaotic signalprocessing research domain. The chaos theory can deal with a lot of nonlinear signal processquestions in engineering practices, which are difficult to be done by linear signal processingmethods. In 1995, the support vector machine (SVM) was proposed by Vapnik etc. according toStatistical Learning Theory (SLT), which is used widely in classification and regressionproblems. The SVM method is basic on the structure risk minimization instead of the empiricalrisk minimization. So it has better characteristic feature of generalization, global optimizationand sparse solution. Now the SVM has been widely attended by the researchers.In this paper, research work focus on the application of the local support vector regression(SVR) and distributed support vector regression in spatiotemporal chaos sequence, and also inthe real EEG signal. The local SVR was proposed and used in the EEG signal. A distributedSVR was proposed through improving K-means clustering algorithm for overcoming thedisadvantages of local SVR. The content of this paper are as following:(1) The characters of chaos and spatiotemporal chaos sequence were expatiated. Accordingto the chaos character of EEG, SVM method was introduced into the modeling of the EEG.(2) Statistical Learning Theory (SLT) and the SVM theory were expatiated. Then theresearch actuality of SVR and its training algorithms were concluded. For solving the difficult oftraining SVR in large samples, a local SVR was proposed, which has the advantages of smalltraining samples, simplicity and high precision(3) Basic on the local method idea, a distributed SVR was proposed. To do this, animproved Rival penalized competitive learning (RCPL) clustering algorithm was used in the distributed SVR.(4) The two proposed methods were used in the real EEG signals.Finally, the work of this paper was concluded. Simultaneity, some problems and studydirection ideas in this aspect were pointed out for further study.