Implementation of Machine Learning for Classification of Flight Simulation Data

Abstract: The main objective of this thesis project is to explore the possibility of using machine learning as a way of categorizing flight simulation outcomes. If such a possibility exists, the process of analyzing whether a new configuration stays within given boundaries can be streamlined. At least provide a first indication of the behaviour of the aircraft purely based on previous verified knowledge. A large quantity of flight simulations are performed to create known data for training the machine learning models. Focusing on the boundaries for angle of attack (α) vs side slip angle (β), the machine learning models are trained on the parameters mach number, altitude, external load configuration and maneuver and classified based on the α and β output received from the simulation. A total of 6 machine learning algorithms have been implemented and compared to each other in terms of performance. The machine learning models are developed on an insensitive aircraftsimulation model and then tested on a simulation model based on a real aircraft. Different ways of improving the performance of the machine learning models are explored. The results indicates that machine learning have potential when it comes to analyzing flight simulation data. In terms of total accuracy and precision there are models which perform incredibly well on the vast majority of the data. However, it is often the outliers and rare cases, which are of interest during analysis of simulation data, where the approach taken in this project lacks in performance.