Hybrid System of Convolutional Neural Networks and SVM for Multi-class MIEEG Classification

Deep learning methods are used in the field of brain-computer interface (BCI). We propose a model that combines convolution neural networks (CNNs) and support vector machines (SVMs) for multi-classification based on the use of motor imagery technique. The Stockwell transform is used to transform signals into feature maps covering different frequency ranges. Here, the feature map is input into CNN for conducting feature extraction, and the SVM is used as a classifier to classify the feature secondary training extracted by the CNN to form a complete MI processing framework. Batch normalization and drop layer are added to the CNN network to improve the over-fitting phenomenon that may occur. The sliding Gaussian time window enhances the data samples to improve the classification performance of the network. The superiority of the model is demonstrated through the results of the tests conducted on the BCI Competition IV dataset 2a. The average cross-validation accuracy rates of CNN and CNN-SVM are 82% and 84.1%, and the Kappa values are 0.76 and 0.79, respectively. The accuracy rate and Kappa value of CNN-SVM increased by 2.1% and 0.03, respectively. Experimental results show that the classification performance of CNN-SVM is higher than that of CNN and most existing classification algorithms.