EEG Recognition Based On Convolutional Neural Network

As an important biological signal,EEG signals are widely used in various industries such as medical treatment,entertainment,mechanical assistance,and monitoring.The analysis and mining of EEG signal data can translate human EEG signals into external action instructions and realize the information interaction between the brain and the device.However,due to the particularity of the EEG signal itself and the characteristics of the EEG signal obtained in a complex environment,EEG signal recognition still faces many challenges.On the one hand,the EEG signal is different from ordinary data,and its own particularity brings about two problems.First of all,the EEG signal itself is extremely weak,Even in the laboratory environment,it is extremely susceptible to noise interference and has the characteristics of low signal-to-noise ratio.This low signal-to-noise ratio should be fully considered when identifying;In addition,the EEG signal is a biological signal,and the EEG signal data collected from different subjects often varies from person to person,and there are serious differences in data distribution.General EEG signal recognition methods need to have a certain ability to generalize across subjects,and how to overcome this difference in data distribution between individuals is brain telecommunications Number recognition unique technical difficulties.On the other hand,different from traditional image recognition and speech recognition,the conditions for obtaining EEG data are more demanding.Data is limited.How to achieve high-precision classification and recognition from limited data is another challenge faced by EEG signal recognition.In order to solve the problems of low signal-to-noise ratio,individual differences,and limited training data in the above-mentioned EEG signal recognition task,the braincomputer based on EEG recognition Interface system.Based on the convolutional neural network model,this article proposes two EEG signal recognition methods,and successfully applied them to the actual brain-computer interface In the system.The main contents are as follows:1.This paper proposes a convolutional neural network based on metric learning and autoencoder.This method uses a convolutional neural network as the base model.Convolutional neural networks have the ability to effectively learn temporal and spatial features from EEG signals,saving complicated preprocessing steps.Autoencoder can efficiently learn for data compression and classification tasks simultaneously and can learn the potential characterization of the data.In this way,the proposed method is able to learn latent representations that preserve discriminative information of the original EEG data by fusing deep metric learning(DML).Through experiments on the BCI IV 2a data set,it is proved that the model is superior to the traditional FBCSP algorithm and popular Network(EEGNet).Through feature visualization,the feasibility of deep metric learning based on autoencoder in the field of EEG recognition is verified.2.This paper proposes a Riemannian network structure,which opens up a new direction for deep nonlinear learning of symmetric positive definite matrices(SPD matrices).SPD Matrix is often used as a feature in the recognition of EEG signals,but in the past the use of SPD matrix is often simply a classification based on metrics,without in-depth exploration of the representation ability.First,imitating the convolutional layer in the convolutional network,a bilinear mapping layer is designed to convert the input SPD matrix into a more ideal SPD matrix.The eigenvalue correction layer is used to apply a nonlinear activation function to the new SPD matrix,and an eigenvalue logarithmic layer is designed to perform Riemannian analysis on the SPD matrix of the regular output layer.In order to train the proposed deep network,a new stochastic gradient descent backpropagation is used on the SPD manifold to update the structured connection weights and the SPD matrix data involved.The experimental results show that in three typical EEG data classification tasks,the proposed SPD matrix network is simple to train and the performance is better than the existing SPD matrix learning and the latest methods.3.The two EEG identification methods proposed in this paper have been successfully applied in actual production environments.In accordance with the objective requirements of the brain-computer interface system,this paper improves the intersubject generalization ability and the learning ability in the case of limited data while ensuring the accuracy of EEG recognition.Experiments show that the two methods proposed in this article are effective,and compared with the existing EEG recognition methods,their accuracy is higher and intersubjects generalization ability is better.In the relevant application practice,the recognition method in this paper also showed a high recognition accuracy,which fully proved its practical value.Finally,according to the work line of this article,we can continue to carry out related research work.