OSAM-Fusion: A Dual-Branch Feature Fusion Network With Transformer for Open-Set Drone RF Signal Recognition

Abstract

With the widespread application of drones in both civilian and military fields, reliable identification of their radio frequency (RF) signals has become a critical challenge for ensuring low-altitude airspace security. To address the problem of detecting unknown drone threats in open environments, this paper proposes an end-to-end open-set recognition framework called OSAM-Fusion. First, the framework constructs a dual-branch deep feature extraction network: the spatial branch, based on EfficientNet-B0, captures local texture and structural information from the spectrogram, while the temporal branch captures the evolution patterns of the RF signal spectrum along the time dimension via a Transformer encoder. The features from both branches are then fused to obtain a comprehensive multi-scale representation. Second, an open-set aware metric loss function is designed, which actively constructs a discriminative feature space during training through dynamic prototype learning and boundary constraint mechanisms. Finally, a category-adaptive multi-stage decision method is proposed, combining Mahalanobis distance and statistical thresholds to achieve precise classification of known categories and effective rejection of unknown threats. Systematic experiments on a public dataset show that OSAM-Fusion demonstrates outstanding performance in both open-set accuracy and decision reliability across seven hierarchical open-set evaluation scenarios. This research provides a complete and engineering-feasible technical solution for drone intrusion detection in complex electromagnetic environments.