Deep neural network based automatic modulation classification

Abstract

In recent times, Deep Neural Network (DNN) has drawn a lot of interest due to its exceptional performance in identifying complex structured data. In this thesis, the automatic classification of modulation types of analog and digital modulated signals is examined using the DNN method. Due to its crucial function in dynamic spectrum access, which can support fifth generation (5G) wireless communications, automatic modulation classification (AMC) is an unavoidable component of different intelligent communication systems. The AMC has been investigated for more than 25 years, but it has proven challenging to create a classifier that is effective in a variety of multipath fading scenarios and other limitations. AMC systems have recently embraced DNN or Deep Learning (DL) based approaches, and significant advancements have been noted. This thesis suggests AMC approaches based on Convolutional Neural Network (CNN), Residual Neural Network (ResNet), and Convolutional Long Short-Term Deep Neural Network (CLDNN). The RadioML2016.10a dataset has been used in this investigation. It comprises of synthetic signals with 11 modulation types: AM-DSB, AM-SSB, WBFM, GFSK, CPFSK, PAM-4, BPSK, QPSK, 8-PSK, 16-QAM, and 64-QAM. Google Colaboratory has been used as the foundation for all the simulations. Batch Normalization layers have been added after each convolutional layer and first dense layer in the CNN based AMC model and significant accuracy improvement is observed. An accuracy of 68.54% is achieved at high SNR for the CNN based AMC model with Batch Normalization. We have tuned the hyperparameter, dropout rate to 0.3 in ResNet based AMC model and achieved an accuracy of 73.773% at 6 dB SNR. The hyperparameter, batch size has been tuned to 32 and an accuracy of 30.682% has been achieved at 14 dB SNR in the CLDNN based AMC model, which is 12.876% better compared to the CLDNN model with batch size 128.