(A) study of wearable devices’ efficiency in evaluating fitness-for-duty in nuclear power plants

Abstract

Among the events at domestic nuclear power plants, incidents caused by human error account for 75% of Level 2 incidents and 36% of Level 1 incidents in South Korea. To reduce incidents caused by human error, the U.S. NRC requires a Fitness-For-Duty (FFD) program that includes alcohol & drug tests, and fatigue management etc. In Korea, the inspection is conducted on a daily basis for workers, but there is a limit to the subjective judgment of the inspector. Therefore, in this study, we propose a machine learning classification model that can assist FFD through objective biosignals. An FFD classification model is proposed using wearable devices that can be used in actual nuclear power plants. Biosignals were collected from 12 domestic/foreign students at KAIST in drinking and non-drinking conditions. Focusing on the individual's different bioreactions during drinking, the experiment was conducted in a within-subject method in which one person was measured a total of 24 times for 5 minutes respectively. The experiment was conducted when the blood alcohol concentration was 0.03% (2 cans of beer) or higher. Meanwhile, three games were conducted to evaluate cognitive abilities related to NPP between biosignal measurements, and the scores of drinking and non-drinking states were compared. As a result of the cognitive game scores, there was no statistically significant difference in the two simple games (resilience and situational judgment), but there was a significant difference in the case of a rather complex game (planning ability). This suggests that drinking may increase the possibility of human error in abnormal situations requiring complicated accidents in NPPs. On the other hand, as a result of binary classification of representative machine learning models, 10 out of 12 people achieved 100% in terms of sensitivity, and 2 people achieved about 70% due to race and individual body characteristics.