Transference of Training for a DNN to Complete the Aerial Refueling Task

View Video Presentation: https://doi.org/10.2514/6.2023-0192.vid The United States Navy plans on increasing the number of uncrewed aircraft within the carrier airwing. Naval leadership desire that carrier based aircraft have the ability to aerial refuel to enable flexibility. Due to multiple issues with data links, there is not a way to guarantee a human will be in or on the loop to provide visual oversight when an uncrewed aircraft aerial refuels. For this reason, several research efforts are underway for certifying an artificial intelligence (AI) model to enable an uncrewed system to aerial refuel. In this work a deep neural network (DNN) based AI is trained to correctly recognize a refueling drogue and to provide 3D-spatial coordinates allowing to safely establish contact with the refueling coupler. In this work a state of the art DNN is trained to recognize the KC-130 wing store based drogue and coupler using flight test video data of a manned platform aerial refueling. The performance of the trained DNN is then evaluated on flight test video data of the same manned aircraft refueling on a F/A-18 center line air refueling store whose drogue is similar to the KC-130 drogue. When refueling behind a KC-130 wing station drogue, the background is generally filled with the sky allowing to clearly differentiate the object of interest. When refueling behind a F/A-18, the background behind the center line of the F/A-18 drogue will normally consist of the tanker aircraft partially occluding with the area behind the drogue. This paper documents the process by which a DNN was trained on KC-130 flight test data using data and how the trained DNN was able to accurate identify the drogue and coupler behind a F/A-18.