Deep learning based automatic modulation classification in the presence of carrier phase offset and carrier frequency offset

Abstract

Automatic Modulation Classification (AMC) has emerged after the efforts of making the modulation classification process autonomous. Since then, various meth ods, algorithms, and tools have been used in the AMC field, such as likelihood-based methods, the goodness of fit tests, feature-based methods, machine learning-based methods, and deep learning-based methods. With the help of these methods, the mod ulation classification operation can be performed automatically without any human input. In this thesis, we survey these methods in detail and propose our methods to contribute to the AMC field. First, we proposed a blind feature-based algorithm that uses K- nearest neighbor (KNN) to perform classification. When the number of sym bols in each signal decreases, the classification process may encounter an error floor. The main goal of the proposed feature-based algorithm is to combat this error floor. Then, we proposed a novel polar coordinate approach in deep learning to classify the signals that are affected by carrier phase offset (CPO). The polar coordinate approach converts the rotational effect of CPO into the translational effect, which makes the classification easier. Finally, we propose a 2-staged deep learning-based classification algorithm under the presence of carrier frequency offset (CFO). In the first stage, the algorithm estimates the CFO amount and in the second stage, it classifies the CFO affected signals. Finally, we conclude the thesis by discussing the future works and possible improvements.