An Optimized Faster Region-Based CNN for 1D Spectrum Sensing and Signal Identification in Cluttered RF Environments

In this paper, a faster region-based convolutional neural network (FRCNN) has been optimized to allow processing 1D signals and improve the inference time of the architecture for spectrum sensing and signal identification in cluttered RF environments. The number of wireless devices in operation continues to grow, contributing to an ever increasing spectrum congestion. These devices may be competing for scarce resources, or may use new capabilities to interfere in spectrum bands they do not belong. Spectrum sensing is a quintessential ability for cognitive radios, allowing them to detect the presence of transmitters. This may be used to optimize spectrum usage, increasing throughput and performance, or for security purposes, such as monitoring for abnormal activity at an airport. In each of these scenarios, the ability to accurately and adeptly sense the spectrum is required. Object detection has been shown to be excellent at locating signals in a congested environment. Our research optimizes FRCNN object detection for processing 1D signals, dramatically reducing computational complexity. This transformed model is capable of handling 1D FFTs directly, instead of image inputs. We show that our approach achieves higher performance than other state of the art models, through both synthesized and over-the-air tests. Additionally, we show the ability of the detector to allow processing and identifying of multiple signals, by applying automatic modulation classification to each detected signal.