Pilot Maneuvering Performance Analysis and Evaluation with Deep Learning

Abstract

Quick access recorder (QAR) data have been used to evaluate pilot performance for decades. However, traditional evaluation methods suffer from the inability to consider multiple parameters simultaneously, and most of them need to select features manually in advance. To study the relationship between QAR data and pilot performance, this paper puts forward one-dimensional convolutional neural networks (1-D CNN) which consider QAR metrics in an integrated manner. This paper obtained indicators describing the operational status of an aircraft first. Then, the correlation between indicators and pilot performance (skill levels) was studied. Inspired by the fact that CNN can extract both local and global features, this paper has developed an approach to achieve the state-of-the-art result in pilot performance evaluation, which was tested on operating data of Boeing 737. The results prove that other methods do not work well, while the 1-D CNN improves the prediction accuracy of 5 pilot skill levels. Besides, when it is used on a binary classification problem, the result improves to 78.18%. Finally, the indicators were grouped into 5 common factors by factor analysis and fed into 1-D CNN in different combinations. Each common factor plays a different role in pilot performance evaluation, which can provide advice for the future.