Future flight safety monitoring: comparison of different computational methods for predicting pilot performance under time series during descent by flight data and eye-tracking data

Introduction. Effective and real-time analysis of pilot performance is important for improving flight safety and enabling remote flight safety control. The use of flight data and pilot physiological data to analyse and predict pilot performance is an effective means of achieving this monitoring. Research question. This research aims to compare two forecasting methods (XGBoost and Transformer) in evaluating and predicting pilot performance using flight data and eye tracking data. Method. Twenty participants were invited to fly an approach using Instrument Landing System (ILS) guidance in the Future Systems Simulator (FSS) while wearing Pupil-Lab eye tracker. The deviation to the desired route, the pupil diameter and the gaze positions were selected for forecasting the flight performance indicator: the difference between the aircraft altitude and the reference altitude corresponding to the ideal 3-degree glide path. Utilize XGBoost and the Transformer forecasting technique to develop a forecasting model using the data from this research, and conduct a comparative analysis of the accuracy and convenience of both models. Results & Discussion. The result demonstrates that using XGBoost regression model had a higher prediction accuracy, (RMSEXGBoots = 42.29, RMSETransformer = 102.10) and its easier to achieve a high prediction accuracy than Transformer as Transformer forecasting method placed a high demand on debugging model and computing equipment. The deviation to desired route and the pupil diameter were more important in the XGBoost model. Conclusion. The use of machine learning and deep learning methods enables the monitoring and prediction of flight performance using flight data and pilot physiological data. The comparison of the two methods shows that it is not necessarily the newer and more complex technology that can build more accurate and faster prediction models, but building the right model based on the data is important for real-time flight data monitoring and prediction in the future.