System-on-Chip Implementation of Multiband Sensing Using Machine Learning Algorithms for Next-Generation Wireless Networks

Abstract Multiband spectrum sensing is an essential component for next generation (NextGen) wireless systems used for 5G NR and beyond, enabling dynamic and heterogeneous spectrum access. Although various machine/deep learning techniques have been explored owing to its robustness and threshold independence, most existing works are limited to simulations and focused mainly on narrow-band/single-band detection. Furthermore, hardware implementations for multi-band sensing on embedded platforms remain scarce. This paper presents a multi-band detector (\((2^m)\)-band detector, \((m = 0,1,2)\)) that utilizes a supervised machine learning (ML) model, namely logistic regression. The primary detection/input features for ML model are the maximum eigenvalues that are computed from captured sub-band data. A functional prototype supporting simultaneous detection across up to four sub-bands has been implemented on a PYNQ SoC platform, leverage its reconfigurable hardware and Python-based acceleration. The model is trained and validated using synthetic datasets emulating the 5G FR-1 frequency band. The proposed detector exhibits improved performance over widely applied energy and eigenvalue-based approaches by 5 dB and 2 dB, respectively at \((P_d = 0.9)\). Moreover, further improvements in detection accuracy in low SNR is observed with larger sample sizes. The design is evaluated in terms of classification accuracy, hardware resource utilization, and power consumption. Consistent performance across simulation and hardware confirms the model scalability and its potential deployment for future wireless systems.