Applications of deep learning framework and localization to intelligent radio spectrum monitoring

Abstract

In this thesis, the author proposed an intelligent radio spectrum monitoring system with implemented deep learning framework. Deep learning is a powerful method to handle hard tasks automatically. The system uses the RTL-SDR USB dongle as the sensor to collect the spectrum data. This dongle is a low-cost device that can measure the signal from 500kHz to 1700MHz. The maximum bandwidth of measurement is 2MHz. The main functions of this system are to collect the spectrum data and then detect the representation signals and extract their characteristics, such as bandwidth, center frequency, modulation type, and capacity. The modulation classification task was done with the deep learning Long Short-Term Memory model. The model can achieve up to 92% accuracy in the validated dataset. Compared to the result in the paper Convolutional radio modulation recognition networks O’Shea, T. J., Corgan, J., & Clancy, T. C. (2016), the model shows higher accuracy (92% vs. 87.4%). In addition, the model can run with a wide range of Signal-to-Noise Ratio values, while the other research papers often analyzes with a specific value. Finally, an algorithm of localization was implemented in the system in order to find the position of the unwanted or illegal signals. The average error of the algorithm in this thesis is 450m.