Study On Epilepsy EEG Signal Detection Algorithm Based On Deep Learning

Epilepsy is a common brain disease,and its attack will have serious negative impacts on patients.Timely diagnosis of epileptic seizures is essential to improve patients’ health level and the quality of their lives.During epileptic seizures,electroencephalogram(EEG)will produce special waveform changes.Experienced neurologists can judge whether a patient is experiencing seizures according to the waveform of EEG.However,due to the complex shape of EEG waveform,limited energy for manual detection and long detection time,it is necessary to introduce automatic detection algorithm for epilepsy.Traditional automatic detection methods require artificial design of complex features.At present,the popular deep learning methods can automatically learn the effective features from the data,have stronger expression ability,and can effectively improve the accuracy of epilepsy detection.Therefore,this thesis constructs a convolutional neural network(CNN)for epilepsy detection,and optimizes the training process and model structure to further improve the performance of the model.The research of this thesis mainly includes the following aspects:1)In order to improve feature learning ability of the network and make the model convergence better,a layer-wise pre-training mechanism is proposed.Firstly,a three-layer convolutional neural network is constructed for epileptic EEG detection.Secondly,the network is trained by layer-wise pre-training mechanism.This mechanism can obtain a better initial point for the parameters by pre-training the sub-network before training the entire network,which makes the training convergence of the complete network better and effectively improves the performance of the model.2)Since the duration of the characteristic wave of the epilepsy signal is indefinite,this thesis uses a multi-scale feature extraction module based on dilated convolution to replace the ordinary convolution operation in order to extract the features on different time scales effectively.The dilated convolution can change the size of the receptive field by adjusting the dilated rate without adding additional parameters.Compared with directly using a large convolution kernel,it can effectively reduce the calculation overhead.This module usesmultiple convolution kernels with different dilated rates to extract data features at different scales.This thesis also proposes a feature fusion scheme to combine features of different scales.The improved model which is based on multi-scale feature extraction and feature fusion allows the network to learn more expressive data features under the premise of effectively controlling the number of model parameters.This thesis uses the public data set of epilepsy EEG to conduct experiments and test the effectiveness of this model.The experimental results show that the model with layer-wise pre-training mechanism can significantly improve the accuracy of epileptic detection and make the model convergence better.Visualization analysis shows that the pre-training mechanism allows the network to learn better separable features.After using multi-scale feature extraction and feature fusion schemes to optimize the network structure,the accuracy of epilepsy detection has been further improved.