Research On Epilepsy Signal Detection Method Based On Convolutional Neural Network

Epilepsy is a common intractable human brain disease,which has been widespread in human life.At present,the diagnosis of epilepsy by doctor is still based on their own rich clinical experience,using EEG to diagnose whether the patient is in epileptic discharge.This method of manual identification in epileptic lesions has the disadvantages of low efficiency and high rate of missed diagnosis.Convolutional neural network,as the main method of deep learning,has been successfully applied in the field of one-dimensional signal and two-dimensional image recognition.Therefore,in order to improve the efficiency and reduce the artificial missed diagnosis in epileptic lesions,convolution neural network is of great significance in the automatic recognition in epileptic EEG lesions.In this thesis,the detection of epileptic EEG signal is studied.One dimensional hole convolution is used to optimize the structure of deep dense convolution,and the automatic detection algorithm with deep dense convolution neural network is applied to detect and identify epileptic EEG signal for the first time.On the one hand,the acquisition and characteristics of epileptic EEG signals are analyzed,and the optimized Gan network is used to reconstruct and enhance the epileptic EEG data set.The deformable convolution structure was introduced to optimize the structure of the classic LeNet and AlexNet neural networks,the modeling and feature extraction were realized.On the other hand,in view of the development of convolution neural network,one-dimensional hole convolution was proposed to optimize,and deep dense convolution neural network was built based on dense connection and one-dimensional hole convolution to achieve the purpose of epileptic lesion recognition.This thesis describes the convolution neural network epileptic EEG feature recognition method from the following three parts.(1)Pretreatment and enhancement of epileptic EEG data set.Firstly,the physiological basis of EEG generation was introduced,and the current international standards for EEG acquisition were introduced.Then,the characteristics of EEG and epileptic signal and epileptic focus waveform were analyzed.Then,the structure and connotation algorithm of convolution neural network were deeply studied,and the convolution neural network model was explained in detail from the mathematical point of view.Secondly,the training process of convolutional neural network and the related problems that must be overcome in the training were studied.Then,through the acquisition of the EEG data set of the University of Bonn in Germany,the data set was analyzed,and the data set was labeled for preliminary processing and classification.Finally,the Gan network model was optimized to enhance the epileptic EEG data.(2)Traditional convolution neural network for epilepsy recognition.In order to clarify the feasibility and effectiveness of convolutional neural network in the field of epileptic EEG signal recognition,firstly,the data sets needed for deep learning network training are obtained,and the relevant data sets are preprocessed and reconstructed,so as to achieve the purpose of being directly used as network input.Then,the corresponding structure was built according to the classical neural network model,and the related recognition experiments were carried out to determine the feasibility and effectiveness of the convolution neural network.(3)Deep dense convolution neural network for epileptic EEG signal recognition.Firstly,the network structure method of introducing dense connection was studied.Then,the deformable convolution was analyzed from a one-dimensional point of view,and the one-dimensional hole convolution was introduced to optimize different DenseNet structures.The process of building deep dense convolution network model was described from a mathematical point of view.The advantages of this method in epileptic EEG detection were analyzed.Finally,the platform migration based on Android platform was designed.The experimental results show that this method is more accurate and feasible than the traditional convolutional neural network method.