Abstract

Abstract
A Novel Hybrid Deep Learning Approach for RFI Detection
Amin Nazarian Saralang
Radio Frequency Interference (RFI) remains a persistent and significant challenge to the optimal performance of modern wireless communication systems. While fifth-generation (5G) technologies have expanded capabilities through the use of a broader spectral range, their operation in certain frequency bands, such as the 5 GHz range (5.1-5.9 GHz U-NII bands), faces unique interference issues. This spectrum segment is vital for many 5G services yet is heavily shared with other widely used technologies, such as Wi-Fi and Internet of Things (IoT) devices. Such coexistence increases the likelihood of complex and dynamic interference scenarios, often exceeding the capabilities of conventional RFI detection methods.

To address these challenges, this study proposes and evaluates a novel hybrid deep learning framework designed for robust, real-time RFI detection in demanding 5 GHz environments. The architecture combines the object detection strengths of YOLOv8 in leveraging Feature Pyramid Network (FPN) structures for multiscale feature extraction from spectrograms with the deep feature representation capabilities of a ResNet-50 backbone. These parallel feature extractors are integrated through an adaptive fusion mechanism, followed by refinement via channel and spatial attention modules. A Transformer module is then employed to capture long-range temporal and spectral dependencies, thereby enhancing classification accuracy. In addition, noise suppression techniques are incorporated to improve adaptability to the unique characteristics of this spectrum.

A comprehensive dataset strategy is developed to ensure rigorous training and validation of the proposed framework. This includes generating synthetic RFI signals that simulate a wide variety of interference patterns in the 5 GHz range through direct simulation and data augmentation, alongside integrating real-world RFI signal data. This dual dataset approach is intended to strengthen model robustness, scalability, and real-world applicability. Performance is systematically assessed using established metrics such as detection accuracy, processing latency, and computational efficiency. The experimental results show significant improvements of the proposed framework over traditional techniques, achieving high precision with manageable computational demands. Furthermore, the modular design offers scalability for adaptation to other frequency bands, making this work a forward-looking contribution to enhancing the reliability of current and emerging wireless communication systems.