Research On Motor Intention Decoding For Small Sample EEG Data Augmentation

Brain Computer Interface(BCI),as a multidisciplinary technology,has been developing rapidly.The brain-computer interface technology uses an Electroencephalogram(EEG)device to collect the electrical signals of the wearer’s head and uses algorithms to interpret brain intentions.The computer interprets the brain’s intentions into corresponding instructions to control external devices to respond,thus realizing information exchange between the brain and the device.This technology has broad application prospects in neuromedicine,sports rehabilitation,games and entertainment,military and so on.Therefore,the research on BCI technology has important theoretical and practical significance.Machine learning often requires a large amount of data to parse brain intentions,but in some cases the small sample data obtained is not enough to support machine learning.Generative Adversarial Networks(GAN)can generate artificial data with similar features to the original EEG data to expand the sample size to obtain enough data for training.This article expounds the research work carried out from the following three aspects:(1)In order to obtain high quality EEG signals and improve the efficiency of braincomputer interface system,a bowl-ball coupling motion model for high-complexity boundary-avoidance tasks was established based on the equilibrium equation and EulerLagrange equation and presented on a computer screen in the form of virtual task scenarios.This experimental paradigm is a closed loop system with visual feedback,so it can alleviate the problem that subjects are prone to fatigue and unable to concentrate in the process of EEG collection to a certain extent,successfully awaken high-quality EEG signals,and enable the collected data to carry more characteristic information.Furthermore,based on the experimental paradigm model,a complete experimental scheme of EEG information collection was designed.During the experiment,different levels of experimental tasks were completed with the upgrading of the difficulty,which improved the interest of the experiment and the participation of the subjects.In this experimental scheme,EEG data generated by different fingers on the same hand of the subject are collected and analyzed,which can decode the motion intention of the fine parts of the human body.(2)We separated the EEG signal by maximum noise component analysis to remove par the pseudo-trace noise in the EEG data.An adaptive notch filter was designed to remove50 Hz power-frequency interference,eliminate artifacts,and extract 8-13 Hz rhythm signals.The Common Spatial Pattern(CSP)algorithm was used for feature extraction of EEG signals,and the characteristic signals with high discrimination were obtained.Extreme Learning Machine(ELM)classifier is constructed based on Single Hidden Layer Feedforward Neural Network(SLFN).The collected EEG data set and open competition EEG data set were used for intention recognition,and the recognition results were compared with those of various classifiers.The average accuracy of five-fold cross validation for the interpretation of different finger movements of the same hand of 7 subjects was 90.4%,the average value of AUC was 0.93,and the average value of MSE was 0.0955.The average classification accuracy of the 4 subjects in the BCI competition was 90.8%,the average AUC value was 0.95,and the average MSE value was 0.0925.(3)Generative Adversarial Networks(GAN)with five fully connected layers and three fully connected layer discriminators are constructed.Droupout has been added to the discriminator connection layers,and leaky Re LU is used as the activation function between the full connection layers.Based on Generative Adversarial Networks(GAN),data augmentation of different multiples was performed on the collected small sample EEG data set and the open competition EEG data set.A variety of classifiers were used to identify the intention and evaluate the results of mixed EEG data containing real EEG data and artificial data with different expansion amounts.The average accuracy of the five-fold crossvalidation of the single-fold and 1.5-fold enhanced data of 7 subjects was 94.0% and 94.5%,the average values of AUC were 0.97 and 0.96,and the average values of MSE were 0.0718 and 0.0698,respectively.The average classification accuracy was 88.5% and 90.1% in the five-fold cross-validation of single-fold and 1.5-fold enhanced data of 4 subjects in BCI competition,the average values of AUC was 0.92 and 0.93,and the average values of MSE was 0.1150 and 0.0985,respectively.To sum up,this paper mainly studied the establishment of bowl-ball coupled motion model under high-complexity boundary avoidance task,EEG data preprocessing and feature extraction,small sample EEG data enhancement based on generative adjudgment network and motion intention decoding based on small sample EEG data enhancement.Theoretical analysis and experimental results verify the practicability and feasibility of the proposed method.