@article{WANG2026, 
author = {Lili WANG and Shuofeng YIN and Yue PAN},
title = {Controller fatigue discrimination algorithm based on facial features},
year = {2026},
journal = {Journal of Beijing University of Aeronautics and Astronautics},
volume = {52},
number = {4},
pages = {986-994},
keywords = {controller, fatigue discrimination, face keypoint detection, land and air call load, electroencephalogram signals},
url = {https://www.sciopen.com/article/10.13700/j.bh.1001-5965.2024.0057},
doi = {10.13700/j.bh.1001-5965.2024.0057},
abstract = {A real-time fatigue discrimination algorithm that takes into account the work characteristics of controllers is proposed in order to address the shortcomings of the current fatigue detection through facial information of controllers, such as the algorithm’s low robustness and infrequent integration of the actual scene of control. Firstly, the Attention Mesh algorithm is used to obtain the 3D coordinate information of 468 points on the face, and the thresholds of eye and mouth aspect ratios are calibrated sample by sample using the feature matching method. Secondly, three indicators are introduced, namely, the controller’s on-duty time, the real-time land and air call load, and the number of fatigue events, and these three indicators are dynamically mapped to the fatigue detection window through the exponential decay function and the fatigue frequency ratio of blinks within the dynamic decay time window is calculated through the calculation of the fatigue frequency ratio of blinks within the dynamic decay time window. The fatigue trend indicator is derived by calculating the percentage of blinking frequency within the dynamic decay time window. Finally, the EEG and facial video data of the post-shift control test of 30 mature release order controllers in the control room of a control unit are processed, and the fatigue indicator Fδ, obtained from the facial data, is correlated with the EEG fatigue indicator Fε in the time dimension. The findings demonstrated that the validity and reliability of the suggested algorithms were confirmed, and the overall Pearson correlation coefficients in the bivariate cross-correlation analysis results of the 30 subject samples ranged from 0.462 to 0.785. The Sig. two-tailed significance tests were found at the 0.01 level, indicating a significant correlation.}
}