상세정보

This dissertation presents the novel statistical signal processing algorithms for brain-computer interfaces (BCIs) which directly translate the intent reflected by brain signals into a control command without the use of muscles or peripheral nerves systems.

In the first part of this dissertation (chapter 2), we propose a novel feature extraction approach for motor imagery (MI) classification overcoming the weakness of conventional common spatial pattern (CSP) methods, especially for small sample settings. We consider local CSPs generated from individual channels and their neighbors (termed local regions) rather than a global CSP generated from all channels. The novelty is to select a few good local regions using interquartile range (IQR) or an ‘above the mean’ rule based on variance ratio dispersion score (VRDS) and inter-class feature distance (ICFD); instead of computationally expensive cross-validation method. Furthermore, we develop the frequency optimized version by extracting the VRDS and ICFD from local CSPs with filter-bank scheme. The proposed methods are tested on the three publicly available datasets and the performance is substantially improved in terms of classification accuracy.

The second part of this dissertation (chapter 3), we propose a novel MI classification algorithm using filter-bank CSP (FBCSP) features based on MI-relevant channel selection. Contrast to existing channel selection methods based on global CSP features, the proposed algorithm utilizes the Fisher ratio of time domain parameters (TDPs) and correlation coefficients: the channel with the highest Fisher ratio of TDPs, named principle channel, is selected and a supporting channel set for the principle channel that is consisted of highly correlated channels to the principle channel is generated. The proposed algorithm using the FBCSP features generated from the supporting channel set for the principle channel significantly improves the classification performance. 
 
The third part of this dissertation (chapter 4), we propose an optimal channel selection method to improve CSP related features for MI classification. In contrast to existing channel selection methods, in which channels significantly contributing to the classification in terms of the signal power are selected, distinctive channels in terms of correlation coefficient values are selected in the proposed method. The distinctiveness of a channel is quantified by the number of channels with which it yields large difference in correlation coefficient values for binary MI tasks, rather than by the largeness of the difference itself. For each distinctive channel, a group of channels is formed by gathering strongly correlated channels and the Fisher score is computed using the feature output, based on the FBCSP exclusively applied to the channel group. Finally, the channel group with the highest Fisher score is chosen as the selected channels. The proposed method selects the fewest channels on average and outperforms existing channel selection approaches. 

Finally (chapter 5), we propose a novel feature extraction method for EEG-based cognitive task classification based on the correlation coefficients of EEG channel pairs by introducing preprocessing of the EEG signals. The preprocessing attempts to optimally demix each pair of EEG channels using a two-dimensional rotation matrix in order to mitigate the inference between channel pairs and, consequently, to enhance the resulting correlation coefficient features for cognitive task classification. For the optimization, the following three criteria are proposed with an optimal rotation angle approximated for each criterion: i) maximum inter-class correlation coefficient distance (ICCD); ii) minimum within-class correlation coefficient distance (WCCD); and iii) maximum Fisher ratio (FR), which is the ratio of ICCD to WCCD. Performance evaluation based on the three publicly available cognitive task datasets shows that ICCD optimization with the ‘above the mean’ and 1.5 interquartile range (IQR) feature selection method yields the best classification performance in comparison with other existing cognitive task classification methods.

Abstract
Contents i
List of Figures iv
List of Tables vi
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Issues of classification for BCI systems . . . . . . . . . . . . . . . . 3
1.2 Dissertation Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Frequency-optimized local region common spatial pattern approach 7
2.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.1 Local region and local region CSP . . . . . . . . . . . . . . . . . . 9
2.1.2 Local region selection . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.3 Local region CSP features and classifier training . . . . . . . . . . 14
2.1.4 Frequency domain optimized local region CSP . . . . . . . . . . . 15
2.2 Data and experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.1 Data description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.2 Data processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Result and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.1 Performance comparison of LRCSP with CSPP . . . . . . . . . . . 22
2.3.2 Performance comparison of LRCSP with R-CSP . . . . . . . . . . 24
2.3.3 Performance comparison of LRFCSP with frequency-optimized CSP
based methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.4 Performance comparison of proposed methods with related methods
using cross-validation . . . . . . . . . . . . . . . . . . . . . . . . . 27
3 Time domain parameters and correlation coefficients based MI-related
channel selection 31
3.1 System model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.1 Principle channel selection . . . . . . . . . . . . . . . . . . . . . . . 33
3.1.2 Supporting channel set for the principle channel . . . . . . . . . . 33
3.1.3 FBCSP applied to the supporting channel set . . . . . . . . . . . . 34
3.2 Result and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.1 BCI competition dataset IVa . . . . . . . . . . . . . . . . . . . . . 36
3.2.2 BCI Competition IV dataset I . . . . . . . . . . . . . . . . . . . . 38
3.2.3 Experiment Results . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4 Optimal channel selection using correlation coefficient features for CSP 42
4.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1.1 System model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1.2 Distinctive channel selection based on the correlation coefficient . . 45
4.1.3 Supporting channel group of a distinctive channel . . . . . . . . . 46
4.1.4 Optimal channel set selection using Fisher score . . . . . . . . . . 46
4.2 Experimental study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2.1 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2.2 Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.3 Result and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3.1 Performance comparison for BCI competition III dataset IVa . . . 50
4.3.2 Performance comparison for BCI competition IV dataset I . . . . . 53
4.3.3 Complexity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5 A novel EEG correlation coefficient feature extraction approach based
on demixing EEG channel pairs 57
5.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.1.1 System model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.1.2 Cost functions of correlation coefficients . . . . . . . . . . . . . . . 59
5.1.3 Adaptive spatial filtering . . . . . . . . . . . . . . . . . . . . . . . 60
5.1.4 EEG channel pair optimization using rotation matrices . . . . . . 62
5.1.5 Feature selection rules . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.2 Data and experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.2.1 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.2.2 Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.3 Result and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.3.1 Performance comparison of proposed three optimization schemes . 73
5.3.2 Comparison of the performance of the proposed method with existing
connectivity based methods . . . . . . . . . . . . . . . . . . . . 76
5.4 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.4.1 ICCD maximization . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.4.2 WCCD minimization . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.4.3 FR maximization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6 Conclusion 84
Bibliography 86
Acknowledgement