Analysis Algorithm Optimization And Real-time Control System Establishment For Brain-computer Interface Based On Visual Evoked Potentials

In recent years, with rapid development of the neuroscience, electrophysiology, computer science, signal processing technology, biomedical engineering fields and requirement of the medical rehabilitation field, brain-computer interface (BCI) is becoming one of the hottest research topics of scientists around the world. BCI is a novel communication and control technique that allows direction connection between a human brain and a computer or other external device without depending on conventional information output paths (such as, peripheral nerve and muscle tissue) of the brain. With the BCI, people can express ideas and operate devices directly through thought activities in their brains without any speech and gesture. The BCI makes interaction communication between severely disabled subjects and external environment possible, and will improve effectively quality of their life. Also, the BCI provides an alternative means of communication to those traditional ones (such as, auditory, visual and tactile ways) for healthy subjects.BCI can be developed using various means of brain signals, such as, non-invasive Electroencephalogram (EEG), Magnetoencephalogram (MEG), Functional Magnetic Resonance Imaging (fMRI), Near-Infrared Spectroscopy (NIRS), and invasive Electrocorticogram (ECoG). Since the EEG-based BCI is more secure with high practicability, and requires relatively inexpensive equipments, it has been more widely studied and developed. This study adopts the non-invasive EEG technique and focuses on analysis algorithm optimization and real-time system establishment for the visual evoked potential-based BCI. The main results achieved by this study are summarized as follows:(1) Conception of multiway signal processing is introduced into canonical correlation analysis to solve the correlation reference signal optimization for SSVEP recognition. A multiway canonical correlation analysis (MCCA) is proposed. The MCCA algorithm implements correlation analysis between a three-order EEG tensor (channel X time X trial) and a two-way sine-cosine signal matrix, to find the optimal correlation reference signals. Optimized correlation reference signals contain important information of subject-specific variability and trial-to-trial consistency. Experimental results validate that such optimal correlation reference signals derived by the MCCA assist to improve the SSVEP recognition performance.(2) A sparse representation recognition model based on least absolute shrinkage and selection operator (LASSO) regression is proposed for SSVEP recognition. The LASSO recognition model implements regression between the EEG and composite reference signals to form a sparse representation. The solved sparse weight vector implies the dominant frequency components in EEG and hence the current target stimulus frequency. Analysis results indicate the proposed LASSO recognition model enhances robustness and decreases required time for the SSVEP recognition, and hence improves practicability of the SSVEP-based BCI.(3) Three spatial features optimization algorithms are proposed:P300fast extraction algorithm based on FastICA, which is to extract P300effectively using few trials EEG with muli-channel denoising; Channel selection algorithm based on discrete particle swarm optimization (DPSO) for P300classification, which is to search the optimal channel configuration yielding the best classification performance by swarm optimization; Fisher’s criterion regularizd common spatial pattern (FCCSP), which is to extract the most discriminantive event-related potential (ERP) features from single-trial by multi-channel spatial optimization, and hence to improve classification performance. The aforementioned three algorithms are validated based on the Hoffmann’s P300dataset and face perception ERP dataset, respectively.(4) Theoretically, more channels provide richer information of ERP features, and hence should yield better ERP classification performance. However, the feature vectors formed by concatenation of multi-point from muli-channel are typically high dimensional, and most probably result in the so-called curse-of-dimensionality that will depress seriously generalization capacity of the trained classifier and cause poor classification performance. To solve such problem, a sparsity regularization technique, which has been widely discussed in compressed sensing, is introduced into linear discriminant analysis to propose a l1-norm (sparsity) regularized linear discriminant analysis (SPLDA). Extensive experimental tests validate the SPLDA can classifify ERP effectively even with few high-dimensional features.(5) A spatial-temporal discriminant analysis (STDA) is proposed based on multiway extension of the traditional LDA to solve the curse-of-dimensionality caused by the vectorized high-dimensional features in the ERP classification. The STDA implements collaboratively optimization in the spatial and temporal dimensions of EEG to learn two projection matrices rendering the projected features have maximal discriminantive information between target and non-target classes. The two learned projection matrices are then used to transform the constructed spatial-temporal two-way samples to new one-way samples with much lower dimensionality, which improves significantly covariance matrices estimation in the subsequent discriminant analysis, and hence enhances generalization capacity of the learned classifier. The proposed STDA is validated with the BCI Competition-Ⅲ’s dataset and our own experimental dataset. Results show the STDA is effective to reduce system calibration time of the ERP-based BCI, which is considerably important to improve the practicability of BCI system.(6) Configural information processing of human faces, which has been widely researched in the field of cognitive neuroscience, is successfully introduced into the BCI design, such that a hybrid BCI system exploiting simutaneously multiple ERPs is developed. The proposed BCI system is based on an oddball paradigm using stimuli of facial images with loss of configural face imformation. The vivid facial images are ideally effective to resist fatigue and discomfort of subjects for long time use. Loss of configural information makes face perception more difficult and associated with higher-level cognitive functions, which encourages subjects to focus attension on the target more actively, and elicits significantly discriminative ERPs (VPP and N170). Extensive investigations on various types of stimuli validate effectiveness of the proposed hybrid BCI system exploiting simutaneously multiple ERPs (VPP, N170and P300).(7) With the Simulink/Matlab (Mathworks Inc., USA) plantform, three real-time BCI control systems are developed using the multiple ERPs-based hybrid BCI paradigm: Intelligent wheelchair control system; robot arm control system; humanoid robots interaction system. All the three systems require not any neuromuscular function but only attention function of brain to realize wheelchair navigation, robot arm operation (e.g., catch food and deliver goods), and humanoid robots control.