| Summary: | 碩士 === 國立陽明大學 === 生物醫學工程學系 === 104 === These years, machine learning has been considered the main tool for classifying the Electroencephalography (EEG) signals. Although considerable effort has been devoted to the motor imagery classification, there are still two major problems to be solved. First, the motor imagery reactive frequency bands vary across subjects and even trials across the same subject. Second, motor imagery trials are often contaminated with artifacts and false imagination. In this thesis, we proposed a single trial motor imagery classification method called Parzen windows-based spatiospectral patterns with trial pruning (PWSPTP), that can find the optimal spatiospectral filter and prune the contaminated trials simultaneously, by using Parzen windows, analysis of variance, and particle filter based method for spectral band and contaminated trials evaluation. In our experiment on a public database, PWSPTP gives the most clustered spatial patterns and outperforms the competing methods in classification between left-hand and right-hand motor imagery, rejecting the null hypothesis beyond the 95 percent confidence level. Experiments for parameter selection for the proposed method and classification performances between other motor imagery classes are also presented in this thesis.
|
|---|
