19 May 2026 in Physical Communication
2 October 2026 in SSRN Electronic Journal
As unmanned aerial vehicle (UAV) categories in low-altitude environments continue to evolve, conventional closed-set training requires complete retraining on both previously learned and newly introduced UAV data, resulting in substantial storage and computational overhead as well as long model-update and deployment cycles. To address this issue, this paper proposes DRF-CLD (Drone Radio-Frequency Class-Incremental Learning Detector), a class-incremental detection framework for UAV radio-frequency (RF) time-frequency spectrograms. Built on YOLOv11, DRF-CLD dynamically expands the classification output channels while inheriting parameters associated with previously learned classes, thereby enabling smooth expansion of the detectable class space without unnecessarily reinitializing existing class parameters. A representative-exemplar selection strategy based on feature-mean approximation preserves a limited amount of historical supervision under a fixed memory budget, avoiding storage of the complete historical dataset. In addition, the old-class responses of the previous-stage model are used to construct a cross-stage knowledge-distillation constraint that suppresses degradation of previously learned class discrimination during sequential updates and prevents learning new classes from severely impairing established detection capability. Using the public DroneRFa raw RF data, a dataset containing 15 detection classes is constructed and a five-stage class-incremental learning scenario is designed. DRF-CLD achieves an mAP@0.5:0.95 of 0.8602 at the final stage, with an old-class mAP@0.5:0.95 of 0.8766 and an average forgetting value of 0.0406. Comparative and ablation experiments demonstrate that DRF-CLD effectively balances acquisition of new classes with retention of historical knowledge under limited access to past data, while substantially alleviating catastrophic forgetting during continuous class expansion.
11 November 2026 in Physical Communication
8 May 2026 in IEEE Communications Letters
To address the application demands for full-azimuth two dimensional (2D) direction-of-arrival (DOA) estimation of signal over a wide frequency range, we propose a method named geometry frequency adaptive modulation and attention network (GFAM-Net), which is tailored for uniform circular arrays (UCA). GFAM-Net introduces ratio of array radius to wavelength γ as a geometric-frequency parameter and employs its normalized form γ to modulate the output features of a standard convolution via a multi-layer perceptron (MLP), enabling frequency-adaptive virtual baseline extension. Subsequently, a convolutional block attention module (CBAM) emphasizes informative signal components after the extension, and the attention-enhanced features, fused with high-dimensional γ embeddings, are jointly optimized for 2D DOA estimation. Compared to existing deep learning (DL) baselines, GFAM-Net achieves superior DOA estimation performance with only 0.67M parameters.
30 March 2026 in SSRN Electronic Journal
To address time-frequency aliasing in radio frequency (RF) signals from multiple UAVs and co-channel interference from IoT devices (e.g., WiFi, Bluetooth)—both of which hinder the detection of multiple UAV RF signal targets—this paper proposes a lightweight detection model named YOLOv11-WDMA. First, Wavelet Transform Convolution (WTConv) is introduced into the C3k2 module of the backbone network. By multi-level frequency decomposition, WTConv provides multi-scale receptive fields and enhances the extraction of high-frequency detail features of small targets such as Bluetooth and remote control signals. Second, a Dual-Path Modulated Feature Enhancement with Learnable Fusion (DPMA-LF) module is proposed and embedded in the detection head. This module applies cascaded spatial and channel modulation to two parallel feature streams, then adaptively fuses the modulated features with the original neck features via learnable weights, amplifying feature-space differences between aliased signals and background noise. Experiments on a simulated multi-UAV time-frequency aliasing dataset constructed from the public DroneRFa dataset show that, within the signal-to-noise ratio (SNR) range of [-10, 10] dB, YOLOv11-WDMA achieves 96% mAP@0.5 and 87.4% mAP@0.5:0.95, surpassing the YOLOv11n baseline by 13.2% and 17.4%, respectively. The model has only 2.97M parameters and 6.9 GFLOPs, while outperforming all existing nano-scale YOLO models and achieving detection accuracy comparable to RT-DETR and small-scale YOLO models, making it well suited for deployment on edge computing terminals.