Detect and Avoid AI System Model Using a Deep Neural Network

Numerous Detect and Avoid (DAA) systems have been researched in the past to enable integrated operation of Remotely Piloted Aircraft System (RPAS) in the existing airspace with manned aircraft. This paper describes a process of constructing a DAA system using a Deep Neural Network (DNN). Training data are generated using the Detect and Avoid Alerting Logic for Unmanned Systems (DAIDALUS). DAIDALUS calculates the alert levels defined by the RTCA DO-365 MOPS document and outputs the conflict bands represented by ranges of heading, altitude, ground speed, and vertical speed that are predicted to cause well-clear violations with one or more aircraft. As the training data are required to cover a wide range of encounter geometries, two different data sets are combined. The first set of data are generated based on the historic operations that can reflect the characteristics of the airspace. Recorded trajectory data in a highly congested airspace, Incheon FIR of Republic of Korea, is used. Second set of data are generated based on the test vectors given by the MOPS that contains numerous combinations of encounter angles or speeds among many others. The DNN based DAA model is tested through DAA simulations. For this purpose, a previously developed pilot decision model is used along with a aircraft dynamics model. Flight Scenarios are created by modifying some of the test vectors or adding additional intruders. The DNN based model kept the aircraft free of loss of well-clear situation for all the test cases. Especially, it handled the multiple intruder situations that were not part of the training set. However, when compared with the same simulations directly using DAIDALUS, DNN based DAA model resulted in significantly increased fuel consumption, which suggest that the avoidance solutions were less efficient.