Multiclass CSP Based On Channel Extension And Its Application In BCI

Brain computer interface (BCI), also named brain machine interface (BMI), is asystem that enhances neuromuscular output or provides a neural interface tosubstitute for the loss of normal neuromuscular outputs by enabling individuals tointeract with their environment through brain signals rather than muscles. The signalsthat deliver user’s communication or control intention are translated into an output,for example, cursor movement. With the rapid development of BCI technology, BCIhas showed its application prospect in many fields, such as assistive, auxiliary control,entertainment, etc. It is the key factor determining the BCI performance to recogniseelectroencephalogram (EEG) signal correctly. So it is important to study therecognition algorithm based on EEG signal.The common spatial pattern (CSP) algorithm has been widely studied indecoding the spatial patterns of the corresponding neuronal activities from EEGsignal patterns in BCI, and has got well performance in discrimination of two-classmotor imagery (the accuracy can almost reach above95%). Extending CSP tomulticlass paradigms is either done by performing two-class CSP on differentcombinations of classes (e.g., by computing CSPs for all combinations of classes(termed OVO) or by computing CSP for one class versus all other classes (termedOVR), or by approximate joint diagonalization (AJD)). The classification results ofthese three methods have been compared in this dissertation, and are unsatisfactorywith their low classification accuracy. An improved method based on OVO-CSP isproposed by uniting k nearest neighbor (KNN), and gets good result on classificationaccuracy and stabilization compared with the three algorithms of before.The pretreatment process of EEG signal includes all kind of filters. A fewpersonalized band selection algorithms have been proposed, but most of thesealgorithms choose the frequency band by the classification results which are notpractical. Here power spectrum analysis in frequency domain is used for choosing thesub-bands. And a method of grouping feature vectors by CSP algorithm after multiplesub-bands filtering is proposed, and takes good result for single session dataset. CSP algorithm is a kind of spatial analysis algorithm. It aims to choose thechannels which can respect the feature signals of some sort of motor imagery andreduce the dimensions of source signals by these chosen channels. Normally, it ismore easier to find fittest channels as the spatial filters for each motor imagery withmore channels, but it implies more recording electrodes are needed which willincrease the complexity of the experiment and the participants’ tension; this willaffect the data acquisition. Some studies have shown that EEG signals recorded byabout20electrodes are enough for analysis the intention of most participants. In thisthesis, channel extension without increasing the number of electrodes, is proposed toincrease the number of channels. According the optimal number of spatial channels,the optimal channels are chosen after channel extension, and the result shows thegood performance for single session dataset.