Performance of Regression Algorithms for Predictive Pilot Manual Control in Standard Rate Turn
Date of Award
12-12-2024
Degree Name
M.S. in Computer Science
Department
Department of Computer Science
Advisor/Chair
Timothy Reissman
Abstract
Predicting pilot performance is crucial for enhancing flight safety, efficiency and training effectiveness. This study explores the predictive capabilities of regression–based machine learning algorithms for pilot behavior during a standard rate turn maneuver. Temporal data was collected from 16 certified pilots using flight simulators under different virtual reality visual conditions. The aim was to accurately predict six target variables that represented both final standard rate turn, or task, performance and manual control inputs used by pilots to achieve such performance. These six target variables included: (1) final heading error; (2) final yaw rate; (3) final altitude error; (4) maximum positive aileron input; (5) maximum negative aileron input; (6) total aileron input. Six regression-based machine learning algorithms were employed based on top performers within the literature with respect to predicting piloting behavior and general human movement. These six algorithms included: (1) Random Forest; (2) Gaussian Process Regressor; (3) Gradient Boosting Regressor; (4) Linear Regressor; (5) Decision Tree; (6) k-Nearest Neighbor. Gaussian Process Regression performed the best followed closely by Random Forest, with single–output models performing equally well as multi-output models indicating weak correlations among the target variables. Four of the six target variables were found to be determined with high predictive accuracy temporally early on, while the final altitude error and the maximum positive aileron input required the longest temporal information to achieve the highest predictive accuracy. These findings support the usage of such algorithms in automated pilot training systems for providing early indicators in identifying pilots who may require additional guidance. Additionally, environmental factors such as visual richness, or visibility with respect to percentage of cloud cover, seem to affect pilot performance but do not significantly impact the performance of such algorithms. With that said, performance metrics, such as final heading error, were more accurately predicted in cloudier environments whereas pilot manual control metrics like aileron inputs were better predicted in no-cloud environments. The study also identified key features influencing many of the target variable predictions such as yaw rate and heading error, providing insights for the multi-faceted factors that should be considered in training programs and in the development of mathematical models to represent pilot-aircraft interactions. Lastly, these findings highlight the potential of machine learning to enhance pilot training by providing early and accurate performance predictions.
Keywords
Pilot, Manual Control, Regression, Machine Learning, Pilot Training, Virtual Reality, Temporal-Based
Rights Statement
Copyright © 2024, author.
Recommended Citation
Babu, Diya, "Performance of Regression Algorithms for Predictive Pilot Manual Control in Standard Rate Turn" (2024). Graduate Theses and Dissertations. 7463.
https://ecommons.udayton.edu/graduate_theses/7463
