Physics-Informed Neural Networks for Sensing Radio Spectrum for NextGen Wireless Networks
Srinu Sesham1, Nalina Suresh2, Abisai Fillipus Mateus Shilomboleni3
1Srinu Sesham, Senior Lecturer, Department of Electrical and Computer Engineering, University of Namibia, Ongwediva, Namibia.
2Nalina Suresh, Senior Lecturer, Department of Computer and Mathematical Sciences, University of Namibia, Windhoek, Namibia.
3Abisai Fillipus Mateus Shilomboleni, Lecturer, Department of Electrical and Computer Engineering, University of Namibia, Ongwediva. Namibia.
Manuscript received on 04 July 2025 | First Revised Manuscript received on 24 July 2025 | Second Revised Manuscript received on 18 August 2025 | Manuscript Accepted on 15 September 2025 | Manuscript published on 30 September 2025 | PP: 8-13 | Volume-14 Issue-3, September 2025 | Retrieval Number: 100.1/ijrte.C828614030925 | DOI: 10.35940/ijrte.C8286.14030925
Open Access | Editorial and Publishing Policies | Cite | Zenodo | OJS | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Abstract: Sensing radio bands to improve the spectrum sharing capability for emerging wireless networks is crucial. In recent years, numerous data-driven models have been applied to detect radio bands. However, these approaches often suffer from poor generalization due to limited and noisy training data. To address this, domain-specific physical knowledge is incorporated into the neural network training through a physics loss term that regularizes feature representations towards an ideal feature vector extracted from reference (noiseless high-SNR) signal. The feature vector comprises higher-order moments, including energy metrics derived from the received signal samples. The proposed physicsinformed neural network (PINN) jointly minimises a standard binary cross-entropy loss and a physics-based squared Euclidean distance loss, balancing empirical risk with physical consistency via a tunable hyperparameter. Extensive simulations over a wide range of SNR values and multiple physical regularization strengths demonstrate that PINN significantly outperforms conventional energy and artificial neural networks-based sensing models. The proposed PINN model can sense signals down to -12 dB at Pd ≥ 90% with a lower dataset size compared to traditional data-driven models, achieving the same performance. The proposed work highlights the benefit of integrating physical priors into neural network models for spectrum sensing. It opens pathways for enhanced cognitive radio designs capable of reliable signal detection under practical channel impairments.
Keywords: Physics-Informed Neural Network, Spectrum Sensing, Signal-to-noise ratio, Detection accuracy, ROC curves.
Scope of the Article: Artificial Intelligence and Methods