Research On Recognition Of Multi-level Attention Based On EEG

Attention is closely related to students’ learning effectiveness.Attention is a complex cognitive ability,and it is difficult to determine whether cognitive functions are functioning simply by external features such as demeanour and eyes.The EEG signal,as an electrical signal that responds to an individual’s state of mind,offers a new possibility for the identification of attention levels.However,there are still some problems with EEG-based attentional level recognition,such as the fact that most of the current research on attentional level is limited to two levels of attention;the traditional attentional level recognition only uses a single feature to portray the attentional level,and the classification rate is not satisfactory.To address the above problems,based on the previous research results,this paper mainly accomplished the following work.Firstly,experiments are designed for data collection.Based on the theory of the "inverted U model" of attention,two attentional tasks and two non-attentional tasks are designed to induce different levels of attention by adjusting the difficulty of the tasks and combining them with subjective scales to ensure the usability of the data.The final EEG signals from 14 subjects at four different attention levels are obtained for the follow-up study.Secondly,CEEMDAN-PE combines with wavelet threshold denoising is used to remove the noise contained in the signals.The IMF components at different frequencies are obtained using the CEEMDAN algorithm,and the components containing high frequency noise are identified by the permutation entropy of the IMF components.Unlike the traditional method in which the noisecontaining components are removed directly,this paper uses the wavelet threshold denoising method to process the noise-containing components and participate in the reconstruction together with the noise-free components.This method outperforms the traditional method in terms of R,RMSE and SNR.Finally,a total of 10 parameters of time domain,non-linearity and energy ratio are extracted as features for the four types of samples after denoising,and two complementary feature selection algorithms,filtered and encapsulated,are combined to propose a hybrid feature selection algorithm of Relief F-MI combined with RFECV,which can remove irrelevant features and redundant features.With this approach,the feature dimensionality is reduced,computational costs are saved,and the recognition rate is improved.The feature vector is eventually reduced from 100 to 36 dimensions,while the accuracy on the SVM classifier is improved from 85.2% to 91.7%.