UAV Tracking Using Channel-Anomaly-Based Deep Learning in ISAC Systems

Abstract

While integrated sensing and communications (ISAC) systems maximize efficiency through a unified waveform, active sensing requires sophisticated signal processing and transmission, resulting in increased power consumption and potential interference. To address such limitations, this article proposes a novel ISAC architecture that passively detects uncrewed aerial vehicles (UAVs) by exploiting channel anomalies occurring as they enter the first Fresnel zone, thereby eliminating the need for dedicated sensing transmission. The proposed scheme employs a two-stage deep learning (DL) framework, a lightweight real-time detector that identifies UAV presence, followed by a bidirectional long short-term memory (LSTM) network with dynamic weighted attention to fuse heterogeneous features, including power, phase, time–frequency, and spatial data under physical constraints for accurate trajectory prediction. To reduce computational overhead, the article introduces a spatially smoothed two-dimensional multiple signal classification algorithm for joint azimuth and elevation estimation, and it derives the closed-form Cramér–Rao lower bound. Additionally, a channel-power-variation method estimates the times of entry and exit into the Fresnel zone. Simulation results confirm that the proposed DL pipeline accurately detects UAV presence and predicts trajectories. The obtained results further demonstrate the high angular resolution of the proposed estimation method and show negligible timing error at high signal-to-noise ratios.