Classification of flight types from GPS recordings

Zapisovalniki zapisov letov, ki beležijo pozicije GPS, so v letalstvu vedno pogostejši. Ustvarjene zapise iz teh naprav uporabniki navadno želijo naložiti na spletne platforme, s katerimi lahko podatke analizirajo ali pa jih delijo z drugimi uporabniki. Količina tovrstnih podatkov hitro narašča, zato se pojavlja potreba po pametnem razvrščanju naloženih podatkov, s katero se olajša delo ponudniku storitve ter pilotu zagotovi boljšo uporabniško izkušnjo. Cilj te diplomske naloge je razviti sistem za prepoznavanje vrste letalne naprave, v kateri je bil ustvarjen naložen zapis GPS. Na podmnožici celotne zbirke podatkov definiramo nabor atributov, ki jih uporabimo za grajenje in primerjanje modelov, zgrajenih z različnimi metodami strojnega učenja. Predstavimo pa tudi nekaj možnih načinov za dopolnjevanje učne množice. Rezultati testiranja so dobri: z najboljšimi modeli dosežemo točnost po meri F1 nad 0,97, kar jih naredi uporabne tudi v produkcijskem okolju, s povečanjem števila učnih podatkov lahko uspešnost še povečamo.Flight loggers that store GPS positions are becoming increasingly popular. Records created with such devices are usually uploaded to various web platforms that provide methods for further data exploration and integrate services for sharing the captured flights on the social media. The amount of uploaded files is continually increasing, thus creating the need for smarter classification of the data, which in turn creates a more optimised and user-friendly service. The goal of this thesis is to develop a system for the recognition of the aircraft type in which the flight was recorded. We define a set of attributes on a smaller subset of the entire flight database and use it for building and comparing the models created using different methods of machine learning. We also present a few methods that could further expand the original training set. Final results are mostly positive: most promising models achieve an F1 score of more than 0.97 which makes them suitable for the use in a production environment. Even better scores can be attained by increasing the number of learning samples