Feature Extraction And Classification For EEG Signals

Electroencephalogram(EEG)is a voltage fluctuation that is generated by the ion current of neurons in the brain.It reflects the activity patterns of the brain's bioelectricity and contains a large number of physiological and disease information.Due to the direct correlation between EEG and consciousness,emotion and behavior of humans,a large number of researches and applications based on EEG signals are implemented.This includes brain-computer interface,emotion recognition,or fatigue detection.However,the clinical application is the most important application field of brain signals.EEG is used to diagnose a variety of brain diseases,such as stroke and brain tumors,which can change the frequency and rhythm of EEG during episodes.EEG analysis has become a necessary means for diagnosis and treatment of these diseases,but there are still many difficulties in meeting clinical needs.Feature extraction and classification methods for EEG have become a common and important subject for the fields of computer science,neuroscience and medicine.With the increasing application of EEG,especially the wide application of wearable devices for acquisition,the data of EEG is more abundant and varied.The feature extraction and classification of EEG signals put forward higher requirements in terms of accuracy,universality,and real-time performance.This work studies feature extraction methods and classification methods of EEG signals,using EEG datasets for epilepsy diseases as test data.New methods are introduced to improve the accuracy,universality and real-time performance of existing methods.Based on the process of computer processing of EEG signals,this thesis carries out research on two key processes:feature extraction and classification.Three methods in terms of the feature extraction are proposed.In this research,a more general classification method is put forward.The main approach includes the following four aspects:Firstly,a feature extraction method based on multifractal detrended fluctuation analysis(FE-MDFA)for extracting EEG signals is implemented.Because of the randomness and non-stationary and nonlinear of EEG signals,the analytical method is used to analyze the nonlinear characteristics,and obtain the multifractal spectrum of EEG signals.The feature extraction method extracts features that can be explained easily and have clear physical significance from the multifractal spectrum of the signal samples.And the EEG signals are classified via the features.In the classification,the hyperparameters of the classification model are selected by searching algorithm,in addition that the insufficient samples training is studied.Secondly,a feature extraction method based on frequency-domain feature search(GAFDS)is suggested.There are a large number of signal analysis methods that can generate many features in reality.But these features are difficult to select and use.The method uses genetic algorithms to search for features in favor of the classification in the frequency-domain of EEG that facilitates classification.The features are compared with some commonly used nonlinear characteristics,and the scalability of method search is studied.In the classification,the study is to improve the classification accuracy via expanding the range of feature search.Thirdly,a feature extraction method based on the deep convolution network and Autoencoders(AE-CDNN)is presented.It is different from the above two methods to extract features based on the existing feature analysis methods.This method is used to reduce dimension for EEG signals,thereby extracting signal features.In the study,we compare the principal component analysis methods and sparse random projection.At the same time,the generality and effectiveness of the features obtained by this method are studied,that is whether the features can improve the classification accuracy of existing research in different data sets.Fourthly,a classification model named CNN-E,which can adapt to EEG of different sampling frequencies,is proposed.Due to the abundance and variety of EEG data,it is usually very difficult for a classification method to fit various EEG.For the problem of classification of EEG signals with different sampling frequencies and lengths,the limitations of existing methods are analyzed and a new effective method is studied.The above three methods for EEG feature extraction have their own characteristics and applicable scope,and they all improve the classification accuracy of the existing studies.In the study,a new classification problem is proposed and the corresponding method is present.The method is universal,and suitable for shorter data classification to improve the real-time index.This work furtherly promotes the research feature extraction and classification for EEG signals.Our study can provide an important basis for detection of EEG in clinical practice.