Automating Vessel Detection with Passive Sonar Signals and Convolutional Neural Networks
Abstract
In recent years, new acoustic stealth platforms, which have the potential to operate invisibly from human sonar operators, have emerged from near-peer competitor nations. In response to the challenges presented by acoustic detection and classification in adversarial marine environments, we proposed a novel application of convolutional neural networks for autonomous passive sonar analysis. Neural networks have made significant strides in multiple fields due to their powerful image recognition abilities. Using time and location information from Automatic Identification System (AIS) data provided by the U.S. Coast Guard, we labeled acoustic signal data recorded by an underwater hydrophone to nearby vessels. We then converted the labeled acoustic data into spectrogram images detailing the frequency, amplitude, and timesteps. With these spectrogram images, we attempted to train several convolutional neural networks to recognize images indicating the presence of maritime vessels. Our results exhibited severe overtraining and unreliable classification of the spectrogram images. We then explored the possibility of converting the spectrogram images to mean frequency vectors and applying other machine-learning algorithms to these vectors. These algorithms produced much more promising classification rates than those of the convolutional neural networks. We hope that our research may be further developed in the future for practical applications in autonomous acoustic classification.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 2020
- Accession Number
- AD1126480
Entities
People
- John W. Kim
Organizations
- Naval Postgraduate School