Anomaly detection in unmanned aerial vehicle systems flight data with temporal and spatial graph networks

This The aim of this research is to develop a method to detect anomalies by analysing unmanned aerial vehicles (UAV) flight data temporally and spatially and to improve the safety and reliability of UAVs. In the research, firstly, a UAV platform was created. In this platform, PIXHAWK3, an open source and programmable autopilot hardware, was used. Functions such as flight parameters, sensor data, navigation information, mission planning and emergency management of the UAV were set through this hardware. Afterwards, the UAV was flown in 10 different scenarios and data on roll, pitch and yaw parameters were recorded during each flight. These data were transferred to the computer for analysis after the flight. Python programming language was used for the analysis of the flight data. Pandas library was used for data manipulation and analysis, Scikit-learn library for data normalisation and Isolation Forest algorithm for anomaly detection. In addition, a method for geographical analysis of flight data was developed. In this method, a weighted graph was created from the flight data using the 'networkx' library and this graph was visualised as a 2D graph. As a result of the research, anomalies were observed at a total of 105 points at different times and at different altitude ranges. It was observed that the flight scenario with the highest number of anomalies belonged to the first flight scenario and anomalies were detected at 23 points during this flight. The flight scenarios with the least anomalies belonged to the 2nd, 4th and 5th scenarios and anomalies were detected at 7 points in each of these flights. It was also concluded that anomalies were observed more at altitudes between 1130-1160 m. The results show that the temporal and spatial graph neural network method can successfully detect anomalies in UAV flights. This method can be used as an important tool for flight safety and performance.