New CNN Model for Feature Extraction and Prediction of BCI Signal on EEG Technology

Abstract

EEG signals were used for direct communication between the human bodies and newlineworldwide in BCI technologies with essential prospects of use in cognitive science and medical newlinecare. BCI based on MI (Motor Imagery) has been extensively utilized in exoskeleton newlinerehabilitation. BCI is a direct communication channel among brain signals of subject and newlineexternal devices. In practical use, the poor signal-noise ratio of electroencephalograms (EEG) newlineleads to low accuracy of identification in BCI. Classification of the EEG signal is essential in newlinecreating a specific BCI system. Numerous researches have thus focused on improving the newlinefeature extraction and classification methods. Several DL (Deep Learning) and ML (Machine newlineLearning) methods were utilized to classify EEG signals. Several of the studies covered time newlinedomain and Frequency domains, but many studies used time and spatial domain features newlineconcurrently to classify multiclass EEG signals. newlineIn this work, we have presented a novel method for features extraction and analysis for newlinesingle-trial MI EEG data dependent on a deep convolution neural network (CNN). At first, newlineAlexNet CNN has examined to classify motor imagery signals. EEG signals were augmented newlineto timeline images of source skull mapped images combining time and spatial domain features newlinein one image to be analyzed simultaneously. DL technology has obtained outstanding results newlinein the BCI method over the past few years, in particular through the use of CNN frameworks newlinein motor imagery signals recognition and evaluation. In this, we have established MI EEG newlinesignal spatial frequency features. Augmented images enabled the AlexNet to extract features newlineof EEG signal activity in terms of time and location of brain activation at the same time. Next newlineof this work, we have proposed transfer learning along with the DenseNet-121 model. This newlinemay be utilized for spatial frequency feature learning and MI EEG classification. newlineBCI Competition III dataset IVa is utilized to show the reliability of our suggested newlinetechniques. Outcomes show that changing EEG classi