Real-time Spectrum Analysis with Convolutional Neural Network on Embedded/Edge Hardware Platforms

This paper presents a comprehensive approach to combining well-known technologies, readily available hardware, and open-source software to investigate and define the vector of parameters for describing the radio-frequency spectrum, which should be considered when designing practical, real-time applications. Spectrum analysis is vital in various applications, not only in wireless communication but also in numerous other essential aspects of human life, including security and surveillance, autonomous vehicles, medicine, and more. Advances in software-defined radio (SDR) technology allow unprecedented control over the entire radio-frequency processing chain. Advanced embedded systems, built with sophisticated hardware and software, enable devices to perform complex tasks at the edge. Convolutional Neural Networks (CNNs) have revolutionized image recognition by allowing computers to understand and interpret visual data with high accuracy. Bringing all of that together, we have developed a working system and compiled the vector of parameters that impact the trade-offs between latency, frame rate (FPS), energy consumption, network training efforts, and hardware throughput constraints that developers face when tackling real-world applications, such as cognitive radio.