Abstract
Automatic modulation classification (AMC) is the process of being able to automatically identify and classify the modulation type of signals, which can be done through extracting features from the signals. Recently, machine learning has been widely used in AMC due to low latency and high accuracy it provides in comparison with other traditional methods of AMC. Specifically, convolutional neural networks (CNNs), a type of deep learning method, is capable of learning signal features through large amounts of data and can outperform support vector machines (SVMs) and other methods of AMC. In an underwater acoustic communication system, there are many characteristics of the channel that makes underwater communication challenging such as multipathing, strong attenuation and noise, limited bandwidth, low data rates, and propagation delay. These physical limitations of the underwater medium end up making it a challenge to determine the modulation type of transmitted signals. Therefore, CNNs will be leveraged and optimized to automatically classify the modulation types of underwater acoustic signals in this project. This report begins with an introduction to deep learning and AMC as well as the recent development of underwater acoustic communication systems and neural networks. Information on the underwater datasets and my contribution to this thesis then follows. Next, there is a review of CNNs and their layers, and then the literature review of the type of work that has already been done using CNNs and modulation classification as well as its alternatives. The report ends with the testing of the CNN and other models on MATLAB using the datasets containing underwater acoustic signals, the results obtained through the experiment, the discussion of certain optimization and data augmentation techniques for the neural network, and the conclusion and suggestions for future work.