Analysis Of Steady-state Visual Fatigue EEG Based On Restricted Boltzmann Machines

The fatigue state of human body under overload often affects its daily life,and even leads to accidents.Therefore,it is of great significance to study the fatigue mechanism and effectively monitor the fatigue.EEG signals are considered by many scholars as the "gold standard" for studying fatigue.Analyzing fatigue characteristics through EEG signals is an objective method for effectively judging fatigue.Due to the characteristics of high signal ratio,easy operation and no training required,the electroencephalogram signal based on steady-state visual evoked potential(SSVEP)has been widely used,but users will produce fatigue after long-term use,and there are few fatigue studies on SSVEP at present.Therefore,in this paper,based on the steady-state visually evoked EEG signal experiment,we use an improved deep mixed model to study fatigue EEG signals,and compare with the algorithms commonly used to extract EEG fatigue features and classification.The main research work of this article is as follows:1)In this paper,we designed a fatigue EEG signal experiment based on steady-state visual evoked potentials.In the experiment,a black-and-white flipped chessboard interface was designed,and a low frequency band 9-13 Hz was used as a stimulus source.18 subjects with normal vision and healthy body were tested in four stages to induce fatigue EEG signals.We first collected the original EEG signals through a bandpass filter of 0.5-35 Hz to remove the artifacts in other frequency bands,and then used independent component analysis to filter out the artifacts such as electrooculogram and myoelectricity contained in the same frequency band,and finally obtained clean EEG signals.2)We propose an unsupervised Multi-Channel Restricted Boltzmann Machine(MCRBMs)network model to reconstruct EEG signals,which can automatically extract the intrinsic features of fatigue EEG signals.In this paper,we improve the unsupervised restricted Boltzmann machine network(RBM)model.For training multi-channel EEG signals,we will get multiple RBM networks.In order to reduce the complexity of the network model,we propose " Weight Sharing" mechanism that not only reduces model parameters,but also reduces noise in EEG signals.The EEG signals reconstructed effectively by the MCRBMs model have significant differences in the amplitudes of sobriety and fatigue at the corresponding stimulus frequency response,that is,the hidden layer can effectively retain the main features of the original EEG signals,and at the same time it can extract the internal fatigue Feature,so the hidden layer can be directly connected to the input of the classifier as a feature layer,which overcomes the shortcomings of traditional methods that require manual extraction and selection of features.3)Multi-Channel Restricted Boltzmann Machines network is combined with supervised Convolutional Neural Network to form MCRBMs-CNN deep hybrid model.The hidden layer of the fatigue EEG signal extracted by the MCRBMs model is directly used as the input of the CNN,and it is fully connected to the classification output after passing through the spatial domain filter layer and the temporal domain filter layer.Based on the SSVEP-based fatigue electroencephalogram dataset,the average accuracy of the MCRBMs-CNN deep hybrid model is 88.63%,which is 10% higher than the traditional fatigue EEG algorithm model.The experimental results prove the MCRBMs-CNN model classification better.