Enhancing Spectrum Sensing for Cognitive Radio: Radio Signal Classification using Neural Networks

Subray, Siddhartha

Graduate Thesis Or Dissertation

Enhancing Spectrum Sensing for Cognitive Radio: Radio Signal Classification using Neural Networks Public Deposited

Analytics

Citations

Citeable URL: https://scholar.colorado.edu/concern/graduate_thesis_or_dissertations/8g84mn59b

Abstract

With the rapid development of numerous wireless network technologies and the growing number of wireless devices in use globally, sharing the radio frequency spectrum has become a challenge that must be addressed. In recent years, methods for detecting and classifying features in photos, audio, and other types of data have been developed using Deep Neural Networks (DNN). DNN classification algorithms have demonstrated the ability to analyze audio signals with a similar structure accurately for a variety of applications including music recognition, speaker identification, earthquake detection, and sound localization. Recently, DNNs have found applications in the wireless networks domain, and radio frequency (RF) signal identification and classification is one of ideal applications for this machine learning (ML) technology. Given that widely used wireless technologies such as Wi-Fi, LTE, and 5G-NR share modulation schemes, it is beneficial to discern the type of signal, rather than simply identifying the modulation scheme of a signal in order to improve spectrum sensing capabilities. In this dissertation, a novel input feature engineering approach for processing signal I/Q data is proposed and evaluated using different types of supervised neural network architectures, such as the Deep Feedforward Neural Network (DFNN), Deep Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) Neural Network, to detect and classify between 5G-NR, LTE, and Wi-Fi transmissions. The dissertation demonstrates that the proposed feature engineering approach significantly outperforms existing methodologies and that with the appropriate input features, simple neural network architectures can achieve high signal classification accuracy.

Creator