Research On High Accuracy And Low Complexity Epilepsy Detection Algorithm Based On EEG Signal

Epilepsy,an extremely common neurological disease,has caused disturbances to the normal lives of numerous patients.Timely and accurately epilepsy detection has become an influential measure for the prevention and treatment of epilepsy diseases.To this end,this thesis designs a set of EEG signal epilepsy detection algorithm to achieve automatic detection of epileptic seizures,and optimizes the precision and complexity of the algorithm.In this thesis,firstly,it studied the progress of epilepsy detection related work at home and abroad,and after comparing and analyzing the existing methods,it was determined that convolutional neural network was used as the core framework of the algorithm,and a one-dimensional multi-channel convolutional neural network structure was proposed as the basic model of the epilepsy detection algorithm.On top of this,to lower the complication of the algorithm,this thesis uses a depth-separable convolution operation to replace the regular convolutional layers in the model,which greatly reduces the model computation and the number of parameters at the cost of a small loss of information.Meanwhile,this thesis also proposes a redundant channel removal scheme with a channel screening algorithm to remove the channels with redundant information from the original multi-channel input data,further reducing the overall computational effort of the model.In order to improve the accuracy of the algorithm,this thesis focuses on the specificity problem in epilepsy detection by designing a patient-specific relearning mechanism,which automatically collects patient-specific samples during the work of the algorithm through an unsupervised label generation algorithm to train specificity models for different patients,effectively improving the classification accuracy of the model for different patients.In addition,this thesis also realizes a set of reconfigurable neural network gas pedal hardware for the proposed epilepsy detection algorithm,which can complete all the calculations required for EEG signal epilepsy detection through the cooperation of control module,storage module and computation module,and can realize the calculation of channel screening,label-free supervised generation and other innovative points of the algorithm.In this thesis,the proposed epilepsy detection algorithm and hardware are tested and analyzed in detail,using a mixture of all patient data to train a comprehensive model,and then testing each patient data,focusing on the generalization ability of the model.For the basic depth-separable convolutional neural network model,the average accuracy for each patient was 85.48% and the model computation is reduced by 75.31%.After patientspecific training,the average accuracy is up to 92.5%.With the introduction of the redundant channel removal scheme,the average accuracy reduction is about 1%,but the overall model computation can be further reduced by 10%-15%.For the hardware gas pedal,the average hardware accuracy is reduced by less than 0.7% with all test set samples,and the overall dynamic power consumption of the circuit is less than 1 m W.In summary,this thesis proposes an EEG signal epilepsy detection algorithm that can effectively accomplish seizure detection,and makes a hardware circuit implementation of the algorithm,which provides technical support for future deployment of epilepsy detection on portable health monitoring devices.