Abstract by David Van Komen
David Van Komen
Physics and Astronomy
Identifying underwater acoustic source range and seabed type using convolutional neural networks
Finding acoustic sources in the ocean with an unknown seabed is challenging. Machine learning models make predictions by identifying patterns from data. In particular, convolutional neural networks (CNN) can learn patterns directly from a sound source. We built a one-dimensional CNN to classify sound source ranges and ocean environment types from a received signal. The CNN was trained on signals generated in different environments (sandy, muddy, or mixed-layer sediments on the ocean floor) for several range classes. We found significant potential for localizing sources in varied environments using a neural network of this type. This type of network provides an alternative for learning from frequency domain spectrograms, removing the necessity for transformations which reduces computational requirements. Time-domain learning might also be beneficial for real-time applications like active sonar.