Time-series spatial predictions of UAS using hybrid machine learning

The proliferation of Unmanned Aerial Systems (UAS) demonstrated in recent conflict zones have illustrated the need for robust Counter-UAS (C-UAS) capabilities that can accurately detect, identify, and track these highly maneuverable systems to protect critical assets. This proposal exploits recent advances in Artificial Intelligence (AI) and Machine Learning (ML) to recognize patterns in UAS trajectories and apply AI predictions to aid engagement with C-UAS. By using an AI-based approach to rapidly learn contextual target maneuvers, the C-UAS system-of-systems can offload intensive processing and thereby enable lower-cost, resource-constrained, expendable systems to defend against identified targets. To overcome issues with small, unbalanced datasets, we combined the benefits of a Convolutional Neural Network (CNN) with a Long Short-Term Memory neural network to make time-series spatial predictions. Our current dataset of model UAS trajectories features highly randomized, but physically plausible maneuvers. The randomness of the maneuvers substantially confounds the AI approach to a surprising degree, but the AI/ML model learns some of the velocity and acceleration behavior of the modeled vehicle. We also compare our predictions to traditional state estimation methods to assess baseline performance.