Search

Aiming at solving the problems of low recognition accuracy and poor practicability caused by the small amount of data and insufficient training of the network weights when the deep neural network algorithm is directly used to classify the 4-class motor-imagery electroencephalogram signals(MI-EEG), We combine the one-versus-the-rest common spatial pattern (OVR-CSP) algorithm and a novel convolution neural networks (CNN) algorithm to extract features and classify 4-class MI-EEG signals. Firstly, the original EEG signal data is truncated and expanded by using a fixed-size overlapping window, the features of intercepted sub-signals are extracted by CSP algorithm and the obtained feature vectors are merged as the input sample matrix of CNN. Secondly, CNN algorithm performs second feature extraction and final classification on the input sample matrix. Finally, the validity of the proposed algorithm was verified by the datasets IIIa of the BCI2005 competition. The average recognition accuracy of the three subjects has reached 91.9%, this is an improvement over other algorithms in recent years.

A key issue in brain-computer interface systems (BCI) based on motor-imagery electroencephalogram signals (MI-EEG) is the classification accuracy of EEG signals. Although deep learning (DL) methods have achieved great success in many research fields, only a limited number of works investigate its potential in BCI application research. In order to optimize the classification performance of MI-EEG signals, we propose a deep learning end-to-end classification model which is combined with convolutional neural network (CNN) and stacked autoencoders (SAE). A new type of CNN is introduced into the model for learning generalized features from time and spatial domains and for dimension reduction. Finally, the features extracted in the CNN are classified by a deep network SAE. The effectiveness of the proposed approach has been evaluated by using datasets of BCI competition data III and BCI competition data Ⅳ. Our results show that DL should be considered as an alternative to other state of art approaches, if the amount of data is large enough.

In order to realize the fast and accurate automated detection and classification of EEG signals during the normal, inter-ictal and ictal periods of patients, we propose an automated classification method for feature extraction of epileptic EEG signals based on the sample entropy and fast-slow-wave energy ratio (FSR)of each frequency sub-band in this paper. EEG signals are decomposed into frequency sub-bands using wavelet packet decomposition (WPD)in this method. The SampEn and FSR of different sub-bands are calculated, which are used to form feature vectors and these vectors are used as inputs to three different classifiers: support vector machine (SVM), k-nearest neighbor (KNN), and probabilistic neural network (PNN), to evaluate four famous classification problems. Our results show that the SVM classifier using radial basis function (RBF)is able to distinguish the above four problems with high accuracy more than 98.67%.

As an important research direction in the field of artificial intelligence, emotion recognition has become a hot topic in current research. Because of the use of multi-channel EEG acquisition equipment nowadays, which brings many problems in practical use and post-calculation, we have also studied channel selection. In this paper, the DEAP database is used as EEG data. The multi-feature fusion in a time domain and the composite features based on wavelet feature and information entropy are used as EEG features for emotion recognition. The average recognition accuracy reached 72.03% and 71.7% respectively. We also use the ReliefF algorithm to select EEG channels. Under the premise of a slight loss of emotional recognition preparation rate, we selected the optimal combination of 6 and 13 channels, and the brain data was reduced from 32 channels to 13 channels. It lays a foundation for the development of portable, wearable devices.

Deep learning approaches have been used successfully in computer vision, natural language processing and speech processing. However, the number of studies that employ deep learning on brain-computer interface (BCI) based on electroencephalography (EEG) is very limited. In this paper, we present a deep learning approach for motor imagery (MI) EEG signal classification. We perform spatial projection using common spatial pattern (CSP) for the EEG signal and then temporal projection is applied to the spatially filtered signal. The signal is next fed to a single-layer neural network for classification. We apply backpropagation (BP) algorithm to fine-tune the parameters of the approach. The effectiveness of the proposed approach has been evaluated using datasets of BCI competition III and BCI competition IV.