Assessing the Predictive Performance of Two DNN Models: A Comparative Analysis to Support Reusing Training Weights for Autonomous Aerial Refueling Missions

The United States Navy aims to enhance its fleet by expanding the deployment of unmanned aircraft in carrier air wings. However, certifying the autonomous refueling of unmanned aerial platforms currently lacks a publicly available method. Ongoing research at the United States Naval Academy focuses on investigating certification evidence that would enable a deep neural network (DNN) to facilitate autonomous aerial refueling (AAR). This study explores training a DNN to accurately detect the drogue and coupler deployed by a KC-130 tanker and a tanker-configured F/A-18 jet. Both tankers have a similar drogue refueling system but differ vastly in image background noise and contrast, posing a challenge for object detection. Using salient metrics, the performance of a DNN model trained separately on video footage of both tankers is tested to enable the AAR task. Our results indicate that a DNN trained on developmental flight test videos of aircraft refueling from a KC-130 tanker effectively completes the aerial refueling task on a F/A-18 tanker compared to another DNN trained on video footage of the same tanker. These findings might validate the idea that a DNN trained on a specific aircraft dataset with a similar probe and drogue refueling system satisfactorily performs the aerial refueling task on various tankers, eliminating the need for additional training data for each tanker individually.