Enhanced Atrous Convolution-Gated Recurrent Unit for Spectrum Sensing in Cognitive Radio Network

Abstract

Spectrum sensing is a technology that aims to address the issue of low utilization of spectrum resources. However, traditional approaches for spectrum sensing might struggle to accurately identify the absence of a primary user within the authorized spectrum under conditions such as low Signal-to-Noise Ratio (SNR), computational cost, and the presence of multipath fading and noise. This paper proposes the Enhanced Atrous Convolution-Gated Recurrent Unit (EAC-GRU) as a classification task to determine whether the spectrum is occupied or vacant. The Gated Recurrent Unit (GRU) component contributes to the model's ability to effectively discern the presence or absence of signals at specific locations by learning from temporal patterns, preventing overfitting, and ensuring efficient training. Enhanced Atrous Convolution (EAC) enables the network to capture a broader context within the signal data, improving the detection accuracy of primary user signals. This method leverages EAC for capturing spatial features and GRU for learning temporal dependencies. Additionally, the spectrum sensing accuracy is enhanced by transfer learning, which improves robustness and reduces training time. The experimental results showcase the superiority of the EAC-GRU model in spectrum sensing tasks, outperforming baseline methods and state-of-the-art approaches. Precise performance metrics, such as detection probability (${p}_{d}$) and false alarm probability (${p}_{f}$), demonstrate the model's effectiveness across various signal characteristics and sample lengths. The proposed EAC-GRU model achieves a $({p}_{f})$ of 0.14% and a ${(p}_{d})$ of 53.49% at − 20 dB SNR for signals with 64 and 128 sample lengths. These findings underscore the importance of leveraging advanced deep learning techniques to enhance spectrum utilization and minimize interference in cognitive radio networks.