| As a bridge of human-computer communication,electroencephalography(EEG)-based brain-computer interface(BCI)aims to directly decode the neural activities into different control commands,which has been proven successful in many fields.In recent years,there have been frequent reports of death due to overwork,and the deaths caused by fatigue construction during working have increased year by year.Therefore,finding an efficient and convenient fatigue detection method has become the key direction of engineering construction safety research.At the same time,in the field of medical rehabilitation,EEG signals are also widely used.For example,in medical rehabilitation treatment of paralyzed patients and brain-controlled robots,motor imagery can be analyzed by decoding EEG signals.Although EEG signals have been widely used in these fields,there are still many challenges,such as the constraints of application scenarios and the accuracy of motor imagery classification.Based on this,the main work of this thesis is as follows:Firstly,to combat the problems of high power consumption,high price,complex deployment,and unsuitable for industrial scenarios of traditional EEG signal acquisition equipment,this thesis proposes a fatigue detection method which might be applied to helmet in industrial environments.The low-power,single-channel EEG signal acquisition module(TGAM)is used to collect EEG signals.Considering the requirements of industrial environment,the deployment of EEG electrodes is discussed in detail for smart helmets.By trade of the reliability of signal acquisition and the feasibility of deployment in practical application scenarios,a fatigue detection method is designed for helmets.Experiment results showed that the accuracy might reach 83.3%.In addition,considering the requirements of the application scenario,a three-axis acceleration sensor is used to monitor the head movement,which is utilized to study the influence of movement on the detection results.Herein,the threshold is set to improve the detection reliability.Experimental results showed that the detection accuracy might reach 65% under the scenario of non-intense activity.Secondly,to combat the problems of low accuracy and low learning efficiency during the motor imagery classification tasks,this thesis proposes a classification method of motor imagery based on spatiotemporal features.Currently,most studies focused on spatial analysis of EEG signals,and neglected the strong time series features of EEG signals.Considering the excellent performance of CNN in spatial analysis and the high accuracy of LSTM in time series classification,a CNN-LSTM joint model is designed to achieve motor imagery classification.This model firstly preprocesses the original EEG signals,including filtering,removing artifacts and continuous wavelet transform.Then it inputs into the CNN-LSTM neural network.Experimental results showed that it can achieve 96.26% classification accuracy.Finally,the number of network layers and parameters on the classification accuracy are discussed through experimental analysis,and the optimal model parameters are determined. |
