Research On Related Algorithms Of Motor Imagery Based Brain Computer Interface

Exploring the mysteries of the brain is the most significant challenge of natural science in the21st century. The brain is the most complex part of human body and the highest level of the central nervous system. Brain signals generated by electrical activities can reflect different state information, so research on brain signals is an important component of brain science field. Brain Computer interface (BCI) system, which provides communication and controls channel between user and external device through the analysis and processing of the electrical activities of human brain, is a new human-computer interaction. BCI system involves computer communication and control, biomedical engineering and rehabilitation medicine and has become a hot topic in interdisciplinary research.The motor imagery based BCI system, which regards brain signals evoked by the changes of brain motor cortex rhythm via motor imagery as input, classifies different motor imagery by signal processing, and translates motor imagery into commands by computer, can realize communication and control between brain and external device. The main component of motor imagery based BCI system is the signal processing stage including preprocessing, channel selection, feature extraction and classification. The performance of motor imagery based BCI system depends on the localization of the changes in the sensorimotor rhythms accurately and the classification of motor imagery tasks correctly. However, development and application of motor imagery based BCI system face severe challenges because the electrical activities of brain are weak signals, which are vulnerable to interference, and their characteristics are low signal-to-noise ratio, dynamic, transient and nonstationary. How to extract features from the brain signals effectively and choose corresponding classifiers are the major research issues in the signal processing stage of motor imagery based BCI system.This thesis is based on the research of the algorithms used in preprocessing of brain signals, channel selection, feature extraction and classification stages. The brain signals can be analyzed and processed in time, frequency, and spatial domain and via nonlinear dynamics domain. Extracted feature can be matched with optimal classifier, in the meanwhile, designing channel selection algorithm aims to reduce computational complexity. Several algorithms are proposed to conduct feature extraction of brain signals and classification of motor imagery tasks. International standard BCI Competitions dataset is utilized to evaluate the efficiency of the developed algorithms. Main contributions of the thesis are as follows:1. The brain signals can be analyzed and processed in time-frequency domains, and a feature extraction and classification algorithm based on the modified S-transform is proposed to identify two-class motor imagery tasks. The modified S-transform is introduced to extract feature from electrocorticogram (ECoG), and an optimal frequency-dependent window is selected by optimizing two factors and adjusting the size of variable window. This window can localize time-frequency information of sensorimotor rhythm synchronized to motor activities accurately. The power spectral density can be obtained from ECoG with modified S-transform. Finally, we can get the local power representation of ECoG in time-frequency domain. Compared with S-transform, the modified S-transform can provide better energy concentration. Compared with other time-frequency methods, the modified S-transform can obtain spectral density function with better representation of time-frequency distribution and higher resolution; Compared with other classifiers, feature extracted from the modified S-transform with the gradient boosting classifier based on ordinary least squares can get the best classification performance; meanwhile, algorithm designed for channel selection can reduce the computational complexity and improve the system performance significantly. The experimental results validate that the proposed algorithm can achieve good classification performance.2. The brain signals are analyzed in time-spatial domain, so a feature extraction and classification algorithm combining local binary patterns with autoregressive model is presented. Local binary patterns used for texture analysis of computer image and autoregressive model are employed to the analysis of one-dimensional brain signals. The combinational feature is fed into the gradient boosting classifier to classify motor imagery tasks based on ECoG. The distribution of histogram of rotation invariant local binary pattern operators and the second-order autoregressive coefficient from Burg’s method constitute combinational features of ECoG brain signals. The brain signals can be analyzed for multi quantization angular space and multi resolution. The aim is to describe the changes of sensorimotor rhythms in time-spatial domain and reflect event-related desynchronization/synchronization phenomenons observed over sensorimotor cortex during motor imagery procedure. This algorithm is evaluated by BCI competitions dataset. The experimental results indicate that the combinational feature can describe the brain signals based motor imagery effectively and get higher classification accuracy. Compared with other classifiers, the gradient boosting classifier in conjunction with ordinary least squares can achieve the best classification performance. The combinational feature may result in increasing the dimension of feature vector, so channel selection algorithm is designed to narrow down the set of input features. It is validated that this algorithm can get a brilliant tradeoff between the classification accuracy and computational complexity.3. On the basis of nonlinear dynamics domain, the fractal geometry theory is employed to analyze and process brain signals, and a feature extraction and classification algorithm based on fractal features and local binary patterns is developed. The blanket covering technique, which is widely used for analysis of gray image, is introduced to compute blanket dimensions, fractal intercepts and lacunarities of different covering levels of brain signals. The fractal intercepts and lacunarities are combined with local binary pattern operators to characterize ECoG signals so that they can be analyzed in different resolutions and angles. The combinational feature can measure the complexity of brain signals, and reflect the speed of amplitude variation. Compared with other features, this combinational feature can get better classification performance, and it can perfectly and exactly identify the changes of sensorimotor rhythms in motor imagery based brain signals. Moreover, channel selection procedure can reduce the feature dimensions to get lower computational burden. The experimental results validate that this algorithm can obtain excellent performance and can get a better tradeoff between the classification accuracy and computational complexity.The research work in this thesis contributes to the development of the study of motor imagery based BCI system in the aspects of techniques theories, algorithms and practical application. This thesis also promotes research of brain signal in time, frequency, spatial domain and its nonlinear dynamics in the application of motor imagery based BCI system.