AI-Enabled Spectrum Sensing and Allocation in Cognitive Radio Networks

A Comprehensive Framework Integrating Deep Reinforcement Learning and Transformer-Based Sensing

Authors:

 ^1* Dr. Gyanesh Kumar Pathak, ^2* Ramanjeet, ^3*Himanshu Gupta

^1*Assistant Professor, MCTE, Mhow, Indian Army, MoD, Govt of India

Email id: gyanesh8kit@gmail.com

^2* Assistant Professor, MCTE, Mhow, Indian Army, MoD, Govt of India

Email id: rjeet8745@gmail.com

^3* Assistant Professor, MCTE, Mhow, Indian Army, MoD, Govt of India

Email id: himanshu3105gupta@gmail.com

Abstract: Cognitive Radio Networks (CRNs) represent a paradigm shift in wireless spectrum management, enabling unlicensed secondary users (SUs) to opportunistically access licensed spectrum bands without causing harmful interference to primary users (PUs). A fundamental challenge in realizing practical CRNs is the design of accurate, low-latency spectrum sensing mechanisms and efficient spectrum allocation policies that can adapt to dynamic radio environments. This paper proposes a unified AI-enabled framework that integrates a Transformer-based cooperative spectrum sensing module with a Soft Actor-Critic (SAC) deep reinforcement learning (DRL) engine for joint spectrum access decision-making. The Transformer encoder processes multi-band spectrum snapshots and computes attention-weighted sensing decisions across a cooperative SU cluster, achieving a probability of detection (Pd) of 0.98 at SNR = -15 dB with a false alarm rate of 0.02. The SAC allocation agent learns a maximum-entropy spectrum access policy that simultaneously maximizes SU throughput and enforces the PU interference temperature limit of -55 dBm. Extensive simulations using ITU-R M.1225 vehicular and Rayleigh fading channel models demonstrate that the proposed framework achieves 83.6% spectrum utilization, an aggregate throughput of 46.2 Mbps for 16 concurrent SUs, and a 33.1 dB reduction in PU interference compared to fixed spectrum allocation. These results represent improvements of 51.4%, 173%, and 21 dB respectively over the threshold energy detection baseline. The proposed system is benchmarked against seven state-of-the-art methods from SCI-indexed literature, consistently outperforming all comparators across detection, allocation, and energy efficiency metrics.

Keywords: Cognitive radio networks; Spectrum sensing; Dynamic spectrum access; Deep reinforcement learning; Transformer neural network; Cooperative sensing; Soft Actor-Critic; Primary user protection; 5G spectrum sharing; IEEE 802.22