Research Of Brain Alertness Detection Method Based On Machine Learning

Alertness refers to a person’s ability of observers maintaining his focus of attention and remaining alert on performing an operational task for a prolonged period of time.In the daily driving process,the driver needs to maintain a high degree of alertness at all times.The driver will have the imbalance of physiological function and psychological function after driving for a long time.And this driving state is prone to cause serious traffic accidents,which is harmful to people’s life and property safety.Therefore,it is of great research significance to detect the state of driver alertness.In this paper,we mainly use part of the method of machine learning to detect the state of alertness of the subjects.In order to better simulate the driver’s driving process,we build a 3-D simulated driving platform using Unity3D,design the simulated driving environment,and complete the programming of the experimental task,the car driving and the communication with EEG signal acquisition device through the programming language of C#.This paper studies brain alertness from two aspects of regression analysis and classification of machine learning.Firstly,the regression model is built to estimate the state of alertness.We get moving-averaged power spectral with EEG signals by the sliding window technique,calculate the correlation between the main features with the principal component analysis and the behavior data,and find the highest correlation of channels.Finally,we predict the behavior data of the subject through the EEG signal power spectrum,and analyze the correlation between the predicted behavior data and the real behavior data to estimate the state of alertness.The purpose of this paper is to detect the state of brain alertness and to ensure that the detection method is more practical.Thus,this paper continues to study the two-classification and multi-classification of the alertness state.Firstly,the EEG signals are processed by discrete wavelet transform to divide the EEG signals into several sub-band signals.Standard variance,amplitude logarithm,quartile and coefficient of variation are selected as the characteristics of EEG signals for feature extraction.Finally,the original EEG signals and the sub-band signals are separately classified by the support vector machine and the extreme learning machine.And the best classification of combination of channels,sub-band signals,feature vector combinations and classifiers are selected to classify the alertness of EEG signals.The experimental results show that the single channel can achieve higher classification accuracy by using coefficient of variation or the feature vector containing coefficient of variation as features to extract.The different sub-band signals of subjects also have a certain influence on the classification,which is obtained when the alertness is classified in the d:（31.3～62.5Hz）,d₄（15.7～31.3Hz）and d₆（4.0～7.9Hz）bands.And the combination of the channels also can improve the classification accuracy to some extent.This paper preliminarily discusses the detection of brain alertness by some methods in machine learning,and verifies the feasibility and effectiveness of choosing the coefficient of variation as a feature.It laid the foundation of theory and experiment for the study of multi-classification of alertness by other machine learning methods.