| Human-computer interaction is a study that includes the technology of information exchange between people and systems (including computers and mechanical equipments etc.). Brain-computer interface (BCI) is a new type of human-computer interaction, and it does not depend on normal output pathways of peripheral nerves and muscle tissue, and it can set up a direct communication channel between the brain and external devices. Currently, the main signal of a BCI is electroencephalogram (EEG). The acquisition equipment of EEG is non-invasive, which has the advantages of low cost, convenient operation and no damage and so on. In this paper, the thorough research of BCIs is based on spontaneous EEG.After fully analyzing the offline experimental data, an online BCI system is constructed for training, which can effectively improve the recognition rate of EEG signals.After analyzing spontaneous EEG signals, reading Chinese characters in mind is designed as speech imagery for a BCI, and EEG signals of 14 subjects are collected by an offline experiment. Event-related spectral perturbation (ERSP) is used in the time-frequency analysis of signals and the appropriate filter range of each subject can be also confirmed by ERSP. In the light of the filter range, event-related (de)synchronization (ERD/ERS) of the signals is calculated for each channel. The experimental results show that reading Chinese characters in mind can induce the ERD/ERS phenomenon of EEG in specific cerebral cortex, which is similar with motor imagery. The spatial source distribution of EEG signals can be calculated by equivalent dipole analysis, and then the dipoles associated with reading Chinese characters in mind are decided by ERSP. Common spatial pattern (CSP) is selected as the feature extraction algorithm. The feature vectors from reading two Chinese characters in mind respectively compared with the idle period are classified by support vector machine, and the classification results are 81.3% and 81.4%, respectively. CSP is easily affected by noise, and the two classification results are improved to 83.5% and 83.1% by optimizing the filter range.In order to improve the stability of spontaneous EEG, mental tasks with speech imagery is proposed as a multimodal BCI.10 subjects participate in the offline experiment that includes two sessions:the first session is mental tasks with speech imagery; the second session is only mental tasks. The stability of EEG signals is analyzed from frequency, time and spatial in the two sessions, and the methods of analysis are autoregressive power spectrum density, Cronbach’s alpha and spatial filtering by CSP, respectively. The results show that signals from mental tasks with speech imagery are more steadinesses. After attaching speech imagery, the classification results from two imaginary periods respectively compared with the idle period are improved by 4.2% and 7.2%. The accuracy between two imaginary periods is also increased by 6.0%. In order to further improve the classification accuracy of EEG signals for different mental tasks, a selection model of time-frequency range based on mutual information is proposed. After optimizing the time-frequency range of signals, the average accuracy is improved by 2.8%.The operational dimensions of the BCI based on motor imagery are limited, and it needs long-term training to get a satisfactory result when the BCI has more dimensions. Imagining left hand and right hand movements are combined with speech imagery, and a multimodal BCI is designed with three-dimensional operation. EEG signals from the three kinds of imaginations are analyzed in the time and frequency characteristics by ERSP for each channel, and the energy fluctuation of signals from these imaginations is different in different channels. CSP algorithm is extended to a three-class paradigm according to the calculation model of one versus one. To improve the classification accuracy, the energy eigenvalues from CSP are combined with synchronization eigenvalues. Synchronization eigenvalues are respectively calculated by cross-correlation function and phase locking value. A method of selecting the channels is proposed to calculate synchronization, which is the synchronization with the greatest difference between two mental tasks. The experimental results of 10 subjects:classification accuracy of speech imagery is the highest (74.3%), followed by the left hand motor imagery (71.4%), and the last one is the right hand motor imagery (69.8%).Currently, feedback training of the BCI is mainly for the experimental paradigm of motor imagery. Based on analyzing the offline experimental data of speech imagery, an online training system of the BCI is devised for speech imagery. Hardware equipments of the training system include a Quik-Cap EEG cap, a Neuroscan SynAmps 2 system and a desktop computer. The software platform is combined with SCAN4.5, BCI2000 and Matlab. Two channels of the online training are selected by Fisher criterion function. Filter ranges of preprocess are from time-frequency analysis of the offline experimental data. Eigenvalues are extraction by variance and power spectrum density, respectively. The classifier is extreme learning machine, and it has faster computing speed. The feature vectors and classification models are updated with the training, and this ensures that the algorithms are adaptivity. The results of feedback are determined by two-step decision fusion. The method has stronger interference immunity and robustness. The results of 4 trainings from 6 subjects show that they can operate BCI systems better after some feedback trainings. Compared with the first training, the results of the second, third and fourth trainings are respectively increased by 7.1%,7.8% and 7.3%.In summary, speech imagery based on reading Chinese characters in mind is the origin of this paper, and speech imagery is extended to two types of multimodal BCIs; The generation mechanism and manifestation of three types of EEG signals are analyzed from three aspects of frequency, time and spatial; The feature extraction and classification algorithms of EEG signals are studied by offline data, which provide the support of algorithms to set up BCI systems; The effect of feedback trainings is analyzed for the BCI based on speech imagery, which lays the foundation for the practical application of BCI systems. |
PDF Full Text Request