EEG Fatigue State Classification Method Based On Machine Learning

As one of the main causes of major traffic accidents,fatigue driving has caused great harm to society,families and individuals.Therefore,it is especially necessary to find an effective method or model to detect the fatigue state of the driver and avoid the serious harm caused by fatigue driving as much as possible.Among various fatigue driving detection methods,detection methods based on EEG signals are recognized as the most accurate and objective methods.In this paper,we collected EEG data through simulated driving experiments,and then classified the EEG data generated during the simulated driving process from the following three parts:(1)Aiming at the low signal-to-noise ratio(SNR)?instability of EEG signals?the problem of manual feature extraction leads to the loss of its important features and the low efficiency of Convolutional neural network training,this paper proposes a fatigue state classification method based on convolutional neural network and residual network.We use the idea of residual network to improve the convolutional neural network and get the model:EEG-Conv-R,and validated the model using the EEG signal data collected in the simulated driving experiment and came to two conclusions,one is the convolutional neural network method has better classification performance than Support Vector Machine(SVM)and Long Short-Term Memory(LSTM);another is EEG-Conv-R converges faster than the original model and has a higher classification accuracy.(2)Aiming at the problem of EEG-Conv-R has a high time complexity and relies on high-dimensional samples,this paper proposes a fatigue state classification method based on CSP+LightGBM.We designed a CSP-based feature extraction method and proposed a Lightweight classifier based on LightGBM:LightFD.Experimental comparisons with models such as support vector machine(SVM),convolutional neural network(CNN),Gated Recurrent Unit(GRU)and large margin nearest neighbor(LMNN)shows that LightFD has better classification performance,does not rely on high dimensional sample data and Decision efficiency,and can distinguish more types of fatigue states.(3)Aiming at the problem of large differences in EEG data between different subjects,this paper proposes a MEDA-based fatigue state transfer learning method.The ultimate goal of studying the fatigue state classification algorithm is to put it into practical application,and the premise of putting into practical application is the versatility of the model.The paper uses the improved CSP method as the feature extraction method.With the MEDA method,the distance between the features in the manifold space is reduced by adaptive distribution configuration,and the classification accuracy is much higher than the joint distribution adaptation(JDA)and metric transfer learning framework(MTLF).