With EEG signals as the "gold standard" for fatigue detection, aiming at the lack of real-time detection and early warning of current fatigue driving detection based on EEG signals, we designed a simulated driving test to collect and record the "eSense" Attention and Meditation, blink frequency, and the power spectrum of the θ, α, and β waves in real time through the TGAM module and the Bluetooth module. The ratio of attention to meditation, the power spectral density ratio of(θ+α)/β, and blink frequency were used as the fatigue indexes to calculate and use the correlation coefficient of Attention and Meditation as classification features. The k-nearest neighbor algorithm(KNN) was used to classify three fatigue exponents with different fatigue degrees. The improved D-S evidence theory synthesis algorithm was used to integrate the accuracy of the three features into a comprehensive index m(θ) to judge fatigue. The results show that fatigue index aA/M, cPSD, and bBlink can reflect the change of driver's driving state. In the simulation experiment, subjects began to be fatigued after driving for about 55 min, and they were already in the fatigue driving state between 55 and 75 min. The fatigue index thresholds are: aA/M: 0.8-1; cPSD: 3.32-4.64; bBlink: 0.28-0.42. The accuracy of comprehensive fatigue index m(θ) is slightly higher than that of single fatigue index. This method provides an important scientific basis and technical support for real-time detection and prediction of fatigue driving in the future.