Research On EEG Signal Recognition Based On Machine Learning

With the development of brain sciences,accurate EEG signal recognition can not only help doctors to diagnose diseases but also help patients to predict the onset of diseases.The classification and recognition of EEG signals still exist some problems,for example,the algorithm with high complexity is not suitable in wearable devices,classification accuracy is poor when training data set is too small,the data imbalance leads to unstable performance.How to solve these key problems in the recognition of EEG signals has become an important factor to furtherly promote development of brain science.This paper first analyzes the main existing problems of EEG signal recognition,and meanwhile,based on the domestic and international research in related fields,the three problems mentioned above are solved in turn by proposing corresponding algorithms.Firstly,this paper extracts a variety of different types of time-frequency features from a plurality of sub-bands obtained by wavelet decomposition and proposes an improved correlation-based feature selection algorithm,which is applied to feature sets to furtherly obtain the optimal feature set.Then,EEG signals recognition of the ictal state,inter-ictal state,and normal state were performed using a variety of classifier methods.Finally,the Bonn University epilepsy data set is used to verify the performance of the algorithm.Among the five different classifiers,Logical Model Tree achieves the best performance of 97.2%.The algorithm successfully arrives the purposes which reduces the complexity of the algorithm as well as improves the classification accuracy by reducing the number of redundant features.Secondly,the pre-ictal state and the inter-ictal state seizure prediction data is with an extremely unbalanced problem.Therefore,this paper designs a time-domain decomposition and reconstruction method to generate pre-ictal state data to reach data balance.In addition,since the amount of this data does not belong to small data sets,a combined time-spatial model algorithm is designed to extract its time-frequency characteristics.Meanwhile,convolutional neural networks are used to further optimize features while performing classification prediction.The algorithm was finally applied to the CHB-MIT dataset and achieved 92.2% sensitivity and 0.11/h false positive rate.Finally,because the EEG signal data set is often small due to the complicated of the acquisition process and difficult to recruit the subjects,common spatial pattern is improved by the following three methods to make the algorithm more suitable for the classification of a small sample:(1)Reducing the target user’s covariance matrix estimation variance by adding non-target users’ covariance matrix,thereby implementing a regularization covariance matrix(2)Adding a penalty term to the objective function to regularize the objective function,which increases the robustness of the entire algorithm.(3)Considering the strong individual differences of EEG signals,wavelet packet decomposition is used to obtain sub-bands with the same length.Then the first two-step improved method is used to extract features on all sub-bands and the maximum correlation minimum redundancy algorithm is applied to optimize feature set.Finally,the higher accuracy obtained by the linear discriminant analysis algorithm shows the excellent classification ability of the small sample data.In view of the main existing problems in EEG signal recognition,we propose the three different classification and recognition algorithms,which are applied to different data sets and can fully meet the needs in different tasks facing different problems,meanwhile,provide some effective ideas for other classification tasks.