A self-attention integrated temporal convolutional network framework for spectrum sensing in cognitive radio

Optimal use of available frequency bands while minimizing cross interference among licensed (primary users) and unlicensed (secondary users) network participants. Traditional spectrum sensing techniques, though effective in controlled scenarios, tend to encounter serious challenges in dynamic wireless environments, particularly under conditions where the signal strength is weak relative to background noise. To overcome these constraints, this work suggests a Self-Attention Temporal Convolutional Network (SA-TCN) architecture that combines attention mechanisms with dilated causal convolutions to extract both local and distant dependencies in sequential signal data. The suggested architecture makes use of residual connections, spatial dropout and layer normalization to improve stability and generalization, while a hybrid Simulated Annealing Particle Swarm Optimization approach is utilized for hyperparameter tuning. Experimental tests on the RadioML2018.01 dataset show that SA-TCN performs better in detection accuracy, precision and F1-scores than advance Deep Learning (DL) models while maintaining stable performance even at −12 dB SNR. Statistical analysis also proves that it has a notable improvement compared with traditional classifiers. By combining efficiency with strong adaptability, the SA-TCN framework provides a scalable and reliable solution for next-generation CRNs. This work adds to the development of intelligent spectrum sensing methods compatible with evolving needs of advanced telecommunication technologies.