Robust Spectrum Sensing under Realistic Radio Impairments: A Unified Evaluation of Classical and Learning-Based Methods

Abstract A fundamental challenge in next-generation wireless communication systems is the lack of robust spectrum sensing methods that can maintain reliable detection under practical channel impairments and non-ideal receiver conditions. This challenge is further aggravated by the performance degradation experienced by many spectrum sensing algorithms in the presence of impairments such as noise uncertainty (NU), Rayleigh fading, carrier frequency offset (CFO), and phase rotation. This paper presents a unified impairment-aware benchmarking framework for the consistent evaluation of classical, machine learning (ML), and deep learning (DL)-based spectrum sensing techniques under controlled non-ideal wireless conditions. The framework explicitly models receiver and channel impairments that are commonly encountered in practical systems but are often simplified or ignored in prior studies. Within this framework, energy detection (ED) and matched filter detection (MFD) are systematically compared with support vector machine (SVM), k-nearest neighbours (KNN), convolutional neural network (CNN), and convolutional neural network-long short-term memory (CNN-LSTM)-based sensing techniques. To ensure fair comparison, a constant false alarm rate (CFAR) thresholding strategy is applied uniformly across all sensing methods. Simulation results demonstrate that ED suffers from the well-known SNR wall phenomenon under NU, while MFD improves detection performance at the cost of increased computational complexity due to CFO compensation. ML-based techniques provide improved detection performance compared with classical baselines but remain sensitive to feature representation and dataset uncertainty. DL-based detectors show strong robustness under the considered impairments, while the CNN-LSTM baseline further illustrates the trade-off between temporal modelling capability and computational cost. The results provide practical design insights into the performance-complexity trade-offs of spectrum sensing algorithms for next-generation dynamic spectrum sharing systems.