Machine Learning Driven Spectrum Sensing for CRWSN Using Logistic Regression, K-Nearest Neighbor and Few–Shot Learning Model

Abstract

Reliable spectrum sensing is a critical function in Cognitive Radio Wireless Sensor Networks (CRWSNs), where unlicensed Secondary Users (SUs) opportunistically access licensed bands without causing harmful interference to Primary Users (PUs). Classical sensing techniques, such as energy detection and cyclostationary analysis, suffer from poor robustness under low signal-to-noise ratio (SNR) and fading conditions and impose high computational overhead on resource-constrained sensor nodes. This paper presents a machine learning–driven spectrum sensing framework that comparatively analyzes Logistic Regression, k Nearest Neighbor, and a Few-Shot Learning (FSL)–based Prototypical Network. At the fusion center, sensing data are transformed into a ten-dimensional radio-frequency feature vector derived from multi-domain descriptors, including detected energy, estimated SNR, channel center frequency, RSSI variance, cyclostationary spectral correlation features, adjacent-channel power ratios, and average primary user presence probability. Logistic Regression and k-Nearest Neighbor serve as baseline models, highlighting the limitations of conventional supervised learning in non-linear and data-scarce environments. To address these limitations, the proposed FSL-based Prototypical Network learns a compact embedding space and class prototypes using episodic training, enabling robust classification with very few labeled samples. Simulations carried out over the 470–698 MHz UHF band under Rayleigh fading conditions demonstrate that the proposed FSL model significantly outperforms the baseline methods, achieving 94.0 % accuracy, a probability of detection of 93.8%, a probability of false alarm of 5.7 %, and an ROC–AUC of 98.0 %. These results indicate that the proposed approach is well suited for dynamic CRWSN deployments.