Translated Abstract
Brain and cognitive science has been one of the most popular research fields worldwide. In 2016, China has incorporated Brain Science and Brain-Like Intelligence Technology into the Major National Science and Technology Projects in the 13th Five-year Plan draft. As an important branch of brain and cognitive science, Brain Computer Interface (BCI) aims to implement Human-Machine interaction, which highly depends on the understanding of the brain activities. EEG is a non-invasive device for brain activity measurement. Since EEG is harmless to human body, of high time resolution and low cost, it has become the major data source in BCI field. The EEG recording of brain activity is non-stationary, nonlinear and of low signal-to-noise ratio, which brings great challenge to brain signal analysis. With the approaching of Big Data era, and especially the great success achieved by deep learning in pattern recognition field, we are inspired to explore the feasibility of deep learning in brain signal analysis. Deep learning algorithms could obtain excellent classification performance based on the discriminative features automatically extracted by deep network model.
Generally, deep network model could be constructed by stacking basic network unit and the learning problem of complex model could be solved by pre-training and error propagation via contrastive divergence. In this way, we can use multiple Restricted Boltzmann Machines to construct Deep Belief Network (DBN), and employ convolutional layer and pooling layer to build Convolutional Neural Network (CNN). Specifically, this thesis employed DBN and CNN for motor imagery classification and P300 signal recognition, and excellent classification performance has been obtained which is superior to that of the conventional methods. The major work of the thesis is described as follows.
1 Research on Application of Deep Belief Network in motor imagery classification.
Motor imagery is characterized by the phenomenon of event-related desynchronization (ERD) and event-related synchronization (ERS) which is discriminant at specific frequency band. Therefore, we choose to transform the brain signal to frequency domain after bandpass filtering. Frequency domain representations of EEG signals obtained via fast Fourier transform (FFT) and wavelet package decomposition (WPD) have been employed. After normalization of the FFT or WPD output, it is fed to the input of the DBN. Through adjusting the network structure and tuning the parameters, an average classification accuracy of 83% is obtained, which has improved the best competition result by 3%.
2 Research on Application of Convolutional Neural Network in P300 signal recognition.
The difference between P300 signal and non-P300 signal in time domain is relatively discriminant, so the time series is employed as the input to the CNN. After segmenting P300 signal sequence from the raw signal, filtering it with a bandpass filter and running a normalization processing, CNN Classification is performed. By introducing Caffe framework to brain signal analysis, this thesis has designed the CNN architecture and has obtained good performance on P300 signal recognition and motor imagery classification.
Corresponding authors email
