Using EEG Power Maps To Recognize Pilots’ Fatigue States

Pilots’ fatigue is an important issue in modern aviation operations.The data show that up to 11%-15% of aviation accidents are caused by pilots’ fatigue.Therefore,effective monitoring of pilots’ fatigue states is important to ensure aviation safety.So far,there is still no mature technology applied in aviation fatigue identification.And it is also one of the difficulties that must be overcome on the road of developing large aircraft industry in China.Many studies have shown that different rhythms of EEG will increase or decrease with the increase of humans’ fatigue.How to extract the effective features and develop the potential variables of these features is a key point in EEG researches.Traditional methods mainly focused on the analysis of EEG rhythms in time or frequency domain,and used shallow network to process features.Aiming at the difficulty of defining EEG fatigue states index,a brain fatigue evaluation index based on power spectrum area coupling brain rhythms was developed.Considering of the deficiency of shallow network,a fine feature learning and fatigue states recognition model based on deep contractive sparse auto-encoding network was established.Aiming at the defects of traditional brain activity mapping in fatigue evaluation,a new brain map evaluation method based on fatigue evaluation indexes was proposed,and a model of fatigue states recognition based on brain map sequences was established.Firstly,simulation flight experiment was designed and combined with fatigue scales to collect pilots’ EEG data under different fatigue states,and the four rhythms of δ,θ,α,and β were extracted.The power spectral density of each rhythm and its curve area was calculated.And three fatigue states evaluation indexes:( + )/,( + )/( + )and / were formed.In order to overcome the shortcomings of feature redundancy in deep auto-encoding network and offsets caused by input jitters,sparse and Jacobian constraints were added to the objective function to form a new type auto-encoding network.Sparse constraints made most of the hidden neurons in suppression states and enhance the ability of the network to find features.Jacobian constraints could reduce the offsets of hidden units caused by input jitters and strengthen the representation of features,thus improving the stability of the model and the effectiveness of feature extraction.Traditional brain workload detection was to analyze single channel or several channels’ EEG characteristics.It was not enough to utilize the whole information.Although brain topographic maps could reflect the global characteristics,they could not be directly used in fatigue states recognition.Therefore,a new algorithm for generating brain power map sequences based on fatigue indexes was proposed.The sequences of brain power maps were obtained by combining the areas of rhythms’ power spectral density curves with the projection and interpolation calculation of fatigue indexes.And the problems of unified modeling,optimization and learning of brain fatigue mapping sequences could be solved by combining convolutional neural networks with long short-term memory network.The ten-fold cross-validation method was used to train the model and recognize the fatigue states.The proposed deep contractive sparse autoencoding network could significantly improve the stability and recognition accuracy rates.The model based on brain power maps could extract more effective fatigue features and obtain better fatigue states recognition results.This topic could provide a certain technical support for the trial certification through the study of fatigue states recognition with EEG signals.