SDR-Based Emulation of Machine Learning-Enabled Spectrum Sharing in 5G Private Networks

As 5G private networks become increasingly prevalent, efficient spectrum-sharing strategies for collocated deployments are critical. This study evaluates supervised, unsupervised and semi-supervised machine learning (ML) approaches for interference prediction and dynamic frequency/channel assignment in collocated 5G networks operating in the 3.5 GHz band. Utilizing GNU Radio OFDM transceivers and Software Defined Radio (SDR)-based emulation on the COSMOS testbed, we develop a spectrum management framework that leverages ML for real-time spectrum monitoring and intelligent resource allocation. The system, powered by an XGBoost model with 85% accuracy, accurately predicts interference and adjusts network frequencies to enhance performance. Semi-supervised models also perform well, achieving similar results with 60% unlabeled data. Key performance indicators such as Packet Error Rate (PER) and Signal-to-Interference-plus-Noise Ratio (SINR) show significant improvements following frequency reassignment, validating the effectiveness of the model-driven approach. Designed for scalability, this system is adaptable to more complex environments involving multiple networks and channels.