Abstract
In the realm of Next Generation (NextG) wireless communication systems, spectrum sharing emerges as a solution to spectrum scarcity, catering to high data rates and improved quality of service. A critical aspect of enabling such spectrum sharing lies in sensing the electromagnetic spectrum (EMS) and characterizing surrounding wireless signals. Machine learning (ML) emerges as a key tool for this purpose. Leveraging established techniques like convolution neural networks (CNN), region-based convolution neural networks (R-CNN), fast region-based convolution neural networks (Fast R-CNN), and faster region-based convolution neural networks (FR-CNN), multiple signals can be identified and extracted simultaneously from a channel. This paper delves into optimizing the FR-CNN tailored for 1-dimensional (1D) signal processing for EMS sensing over a wideband, testing and evaluating the scalability of the model as the bandwidth varies. Models are meticulously developed and compiled to operate across various platforms including CPU and GPU. Additionally, universal software radio peripheral (USRP), GNURadio, and Radio-Frequency Network-on-Chip (RFNoC) are used for wideband receiver design to perform the over-the-air tests. Through rigorous evaluation in simulation and over-the-air, it becomes evident that the optimized 1D FR-CNN is highly scalable in terms of locating and characterizing the active signals within band of interest. Furthermore, while R-CNN demonstrates reduced classification time, a notable compromise is observed in classification accuracy. This optimization journey underscores the intricate balance between speed and precision, crucial for effective spectrum utilization and management in the dynamic landscape of wireless communication systems.