Sound Signal Classification in the New England Mud Patch

In ocean acoustics, finding acoustic signals within long recordings is usually time consuming. In order to optimize this process, this paper explores signal classification using two deep learning models. These models are designed to classify various sources from single-sensor, 60 second time-averaged spectral density levels. The training and testing datasets were taken from 32 channels (on two VLAs) during the Seabed Characterization Experiment 2017 in the New England Mud Patch. A balanced dataset consisted of randomly selected data samples for each of the three classes: ‘Tonals’, ‘Chirps’,and ‘Other’. A two-layer linear model and a four-layer one-dimensional convolutional neural network (CNN) were trained and then tested on data samples from different times. While the linear model achieved above 90% accuracy on the testing samples, the CNN had higher than 98% accuracy. This work shows the potential for deep machine learning algorithms to help identify underwater sound sources, when different signals are present in long audio files. The results of these tests imply that time averaging spectrograms have potential to improve signal classification.