Determining the channel characteristics of communication systems on the aircraft by machine learning methods

Aviation communication is a very important part of today's communication systems. Aviation communication is based on wireless communication. Wireless communication, especially in unmanned aerial vehicles (UAVs), will be an integral part of future wireless communication systems. Real-time data transfer is very critical in UAVs. Because UAVs need to be constantly controlled from ground systems and payload data must be transmitted to ground systems in real time. In order not to interrupt the communication of the UAV, there are some points that should be considered during the design of the communication system. The most important issue to be considered is airframe shadowing. UAV antennas are prone to airframe shadowing. Aircraft surfaces where antennas are placed block the main line of sight path. This causes interruptions in communication. Another important issue is that the losses used in link budget analysis vary in real time. Since the losses taken into consideration during the design phase differ in practical applications, interruptions in communication occur. For these reasons, the end-to-end channels observed in practical applications, including all negative effects, are quite different from theoretical calculations. In this study, an end-to-end channel characterization is made, including all negative effects. For the channel characterization model, machine learning methods are applied to the data measured during past flights. Ground antenna location, aircraft position, aircraft Euler angles and aircraft receive signal strength indicator data on past flights are used. All negative effects of the signal, from the output of the ground modem to the input of the UAV modem, corresponding to each position and Euler angle data are calculated. Missing data are found with machine learning methods. It has been observed that among the models found by machine learning methods, the best results are obtained with ensemble bagged trees and exponential Gaussian process regression methods. The test results of the inputs found by processing the data were worse than the test results of the raw data. However, the results obtained with both processed and raw data are better than the results of theoretical analysis.