Radar Signal Detection in Interference Using Transfer Learning for Cognitive Electronic Warfare
IEEE Access, cilt.14, ss.131804-131816, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 14
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3727414
- Dergi Adı: IEEE Access
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.131804-131816
- Anahtar Kelimeler: Cognitive electronic warfare, electronic support systems, multi-label classification, positional encoding, radar interference detection, transfer learning, vision transformer
- Ankara Üniversitesi Adresli: Evet
Özet
In this paper, we propose a radar signal detection approach for cognitive electronic warfare applications where signal of interest (SOI), namely a radar signal, is disturbed with interference. Our approach utilizes time-frequency images (TFI) of the received spectrum within a multi-label classification framework. Pretrained vision transformer (ViT) models are employed in a transfer learning setup and fine-tuned explicitly for this task. We introduce an additional recurrent layer to more effectively encode the sequential information inherent in TFIs, while employing position embeddings in ViTs. A comprehensive dataset is generated containing diverse pulsed radar signals with various intra-pulse modulations, corrupted by a variety of interference types, namely 5G NR, LTE, WLAN, and continuous wave (CW) signals. Classification performance is evaluated using F1 score for a range of signal-to-interference ratio (SIR) levels in a congested signal environment. In simulations, a maximum F1 score of 0.9772 is achieved, while the minimum score is 0.9117, even at low SIR values, demonstrating the robustness of the proposed framework. Using the proposed recurrent positional encoding layer, a performance improvement of ~3.5% in the F1 score is obtained, highlighting the importance of incorporating sequential information. It is demonstrated that an additional recurrent positional encoding layer enhances the classification performance of lightweight models, achieving performance comparable to state-of-the-art ResNet methods while maintaining lower computational complexity.