Predicting UAV Class and Operating Mode using Machine Learning

Unmanned Aerial Vehicles (UAVs) present a growing threat in conflict zones worldwide. While various capabilities exist to neutralize UAVs, counter-UAV (C-UAV) operators face significant challenges in quickly and accurately characterizing aircraft beyond line-of-sight. This research aims to address these challenges by developing machine learning (ML) models trained on radar-derived air track data. Exploratory data analysis is applied to radar output to illuminate relationships between air track features and UAV class and operating mode. Key features include radar cross section (RCS), target length, altitude, speed, pitch, and heading. A recommended modeling pipeline is presented to demonstrate the flow of data from radar output to classification. By applying supervised ML time-series models, aircraft types and operating modes are classified on simulated data and validated on data from real aircraft. Using long-shortterm memory (LSTM) networks on simulated flight tracks, the model is able to achieve 98% accuracy when classifying UAV class and 97% accuracy when classifying operating mode. The model was validated on data from actual radar systems that predicted UAV class with 72% accuracy, however, limited data available for certain classes of UAVs hindered the model's performance. This study lays the groundwork for future validation of these algorithms using radar track data from real-world UAV flight paths. In application, these models may provide decision aids to help C-UAV operators neutralize threats with speed and confidence.