Resumen
Fatigue affects operators? safe operation in a nuclear power plant?s (NPP) main control room (MCR). An accurate and rapid detection of operators? fatigue status is significant to safe operation. The purpose of the study is to explore a way to detect operator fatigue using trends in eyes? blink rate, number of frames closed in a specified time (PERCLOS) and mouse velocity changes of operators. In experimental tasks of simulating operations, the clustering method of Toeplitz Inverse Covariance-Based Clustering (TICC) is used for the relevant data captured by non-invasive techniques to determine fatigue levels. Based on the determined results, the data samples are given labeled fatigue levels. Then, the data of fatigue samples with different levels are identified using supervised learning techniques. Supervised learning is used to classify different fatigue levels of operators. According to the supervised learning algorithm in different time windows (20 s?60 s), different time steps (10 s?50 s) and different feature sets (eye, mouse, eye-plus-mouse) classification performance show that K-Nearest Neighbor (KNN) perform the best in the combination of the above multiple indexes. It has an accuracy rate of 91.83%. The proposed technique can detect operators? fatigue level in real time within 10 s.