Study of Convolutional Neural Network for Recognition of Baseband Signals
Christopher Gravelle, Ruolin Zhou · 2019
We study the behavior of a functional, pre-defined deep learning convolutional neural network (CNN) used to classify baseband signal modulation types adopted from MathWorks. We examine the behavior of the CNN as the architecture is altered, decreasing and increasing the volume of the hidden layers. We test the viability of the pre-constructed CNN as a solution to real-world communications problems by testing the deep learning networks ability to recognize modulation schemes with a signal-to-noise ratio (SNR) significantly lesser than that used to train the network. We investigate the ability of the pre-constructed CNN to be trained with waveforms of a lower SNR in the interest of more accurate modulation recognition overall. We also demonstrate the efficacy of the CNN model using Universal Software Radio Peripheral (USRP) radios. A long-term goal is to develop the optimized CNN on an SoC-FPGA (System-on-Chip Field Programmable Gate Array) to enable autonomous self-learning as well as self-defense of an intelligent radio system.