Abstract
In the field of radio communications, spectrum allocation is an important task. Automatic modulation classification (AMC) is a critical component in Cognitive Radio (CR) to detect the nearby emitters to prevent radio interference and improve spectrum efficiency. The aim of the AMC is to recognize the modulation types of received signals in CR without the a priori information of channel and signal. The AMC has a scope of performance improvement in classifying modulation used in radio communication by adding real-time synthetic data to the original dataset. The approach is to use the Conditional Generative Adversarial Network (CGAN) for data augmentation and observe the experimental result. Recent work on AMC using deep learning methods shows room for an improvement in classifying 11 different modulation types used in real-time radio communications using dataset RadioML2016.10A with a classification accuracy of 70%. Further work shows that data augmentation is a potential way to improve accuracy and achieve a 2.5% hike by adding rotation, flip, and Gaussian noise. GAN is an emerging technique for data augmentation and data editing on image datasets. Some researchers showed a crucial role of GAN in modulation recognition with an ideal dataset assuming a priori information such as channel and signals. Thus, it proves to be less accurate in a real-time environment using auxiliary classifier GAN (ACGAN). This project focuses on exploiting the ability of CGAN in this field to observe and improve the classification accuracy. The project observed undistinguishable generated synthetic data for all 20 different SNRs using CGAN. The CGAN takes additional information as input with the noise value helps recognize 11 different modulation received signals in radio communications. The accuracy of using data augmentation in AMC is significant. The approach of data augmentation using CGAN in AMC shows a classification performance improvement using the existing CNN model.