Search

A brain-computer interface (BCI) is a direct communication system between a brain and an external device in which messages or commands sent by an individual do not pass through the brain’s normal output pathways but is detected through brain signals. Some severe motor impairments, such as Amyothrophic Lateral Sclerosis, head trauma, spinal injuries and other diseases may cause the patients to lose their muscle control and become unable to communicate with the outside environment. Currently no effective cure or treatment has yet been found for these diseases. Therefore using a BCI system to rebuild the communication pathway becomes a possible alternative solution. Among different types of BCIs, an electroencephalogram (EEG) based BCI is becoming a popular system due to EEG’s fine temporal resolution, ease of use, portability and low set-up cost. However EEG’s susceptibility to noise is a major issue to develop a robust BCI. Signal processing techniques such as coherent averaging, filtering, FFT and AR modelling, etc. are used to reduce the noise and extract components of interest. However these methods process the data on the observed mixture domain which mixes components of interest and noise. Such a limitation means that extracted EEG signals possibly still contain the noise residue or coarsely that the removed noise also contains part of EEG signals embedded. Independent Component Analysis (ICA), a Blind Source Separation (BSS) technique, is able to extract relevant information within noisy signals and separate the fundamental sources into the independent components (ICs). The most common assumption of ICA method is that the source signals are unknown and statistically independent. Through this assumption, ICA is able to recover the source signals. Since the ICA concepts appeared in the fields of neural networks and signal processing in the 1980s, many ICA applications in telecommunications, biomedical data analysis, feature extraction, speech separation, time-series analysis and data mining have been reported in the literature. In this thesis several ICA techniques are proposed to optimize two major issues for BCI applications: reducing the recording time needed in order to speed up the signal processing and reducing the number of recording channels whilst improving the final classification performance or at least with it remaining the same as the current performance. These will make BCI a more practical prospect for everyday use. This thesis first defines BCI and the diverse BCI models based on different control patterns. After the general idea of ICA is introduced along with some modifications to ICA, several new ICA approaches are proposed. The practical work in this thesis starts with the preliminary analyses on the Southampton BCI pilot datasets starting with basic and then advanced signal processing techniques. The proposed ICA techniques are then presented using a multi-channel event related potential (ERP) based BCI. Next, the ICA algorithm is applied to a multi-channel spontaneous activity based BCI. The final ICA approach aims to examine the possibility of using ICA based on just one or a few channel recordings on an ERP based BCI. The novel ICA approaches for BCI systems presented in this thesis show that ICA is able to accurately and repeatedly extract the relevant information buried within noisy signals and the signal quality is enhanced so that even a simple classifier can achieve good classification accuracy. In the ERP based BCI application, after multichannel ICA the data just applied to eight averages/epochs can achieve 83.9% classification accuracy whilst the data by coherent averaging can reach only 32.3% accuracy. In the spontaneous activity based BCI, the use of the multi-channel ICA algorithm can effectively extract discriminatory information from two types of singletrial EEG data. The classification accuracy is improved by about 25%, on average, compared to the performance on the unpreprocessed data. The single channel ICA technique on the ERP based BCI produces much better results than results using the lowpass filter. Whereas the appropriate number of averages improves the signal to noise rate of P300 activities which helps to achieve a better classification. These advantages will lead to a reliable and practical BCI for use outside of the clinical laboratory.