Research On Recognition Of Multi-level Fatigue Based On EEG

As an increasingly common phenomenon,fatigue has not only become an important issue affecting work and life,but also threatens the safety of people’s lives and properties.Especially in recent years,major safety accidents caused by fatigue have occurred frequently.Therefore,the detection and recognition of fatigue is of great significance to improve work efficiency,relieve physiological discomfort and protect people’s health and safety.Electroencephalography(EEG),as a physiological detection method,can directly reflect the fatigue status of the human body.However,there are still some problems to be solved in fatigue detection and recognition based on EEG signals.For example,most of the research on fatigue stays at the recognition of the two-level state of wakefulness and fatigue;the EEG signal acquisition process still inevitably doped with a large amount of noise;the traditional fatigue detection and recognition portrays the mental state through multidimensional features,and the irrelevant features will affect the classification accuracy.For the above problems,this thesis studies the multi-level fatigue state recognition technology based on EEG signals.The main work and contributions of this thesis are as follows:(1)To address the problem that most studies are limited to two-level state recognition of wakefulness and fatigue,this thesis uses the Stroop experiment to induce multiple levels of fatigue in subjects,and determines the four mental states of the subjects through their subjective fatigue scores,to ensure the usability of the collected data.Finally,a sufficient number of EEG signals from the subjects are selected for subsequent studies.(2)To address the problem of doping noise in the EEG data acquisition process,this thesis proposes an optimized VMD-NLM denoising method combining Variational Mode Decomposition(VMD)and Non-Local Mean(NLM).In the optimal VMD decomposition,the number of modes and penalty factors in the VMD are determined adaptively with the minimum value of K-L divergence as the cut-off condition.The theoretical analysis results on the EEG dataset demonstrate that the reconstructed EEG signal using VMD has a correlation coefficient of 0.9969,which is better than those of the comparison method.By calculating the sample entropy of each intrinsic modal function of VMD,the noise component is identified,and the noise component signal is denoised by NLM.The theoretical analysis results on the EEG dataset show that the signal-to-noise ratio of the denoised EEG signal obtained using the optimized VMD-NLM method is 25.419 d B,and the root mean square error is 0.1403,which is superior to the compared denoising algorithm.(3)To address the problems of redundancy in multi-dimensional features and incomplete information extraction of single-dimensional feature,this thesis proposes a Multiple Multi-scale Permutation Entropy(MMPE)feature extraction method based on parameter optimization,and conducts classification analysis based on the MMPE features.In the feature extraction stage,in order to reduce the fluctuation between scales,MMPE uses the moving average method in the coarse granulation process to reduce the problem of information loss,and uses the delayed shelving fetching method to increase the randomness of multiple sequences.Then,the genetic algorithm is used to optimize the three parameters of scale factor,embedding dimension and delay time in MMPE.In the classification stage,in order to verify the recognition effect of MMPE features,different classifiers and different EEG channels are combined and compared.A highest recognition accuracy of 96.6%can be achieved when using the MMPE features of O1 and O2 channels with Support Vector Machine classifier for four-category classification,which is 4.5% higher compared with that of Multi-scale Permutation Entropy features.