Deep Learning for Source Localization in a Laboratory Tank

Deep learning has been applied to underwater acoustics problems in many forms. One difficulty in applying deep learning techniques to ocean acoustics is the spatially and temporally varying environmental properties. Another challenge is the lack of labeled data for training large networks. The overall goal of this work is to develop deep learning approaches for source localization that can be adaptable to different conditions. In this work, a convolutional neural network was trained on acoustic data measured in a water tank, while the water was at room temperature, to predict source-receiver range. Tests were done to ascertain ideal quantities of training data for the task at different frequency bands. For all test cases, a dataset size of 1,806 data samples was found to be sufficiently large. In addition, comparisons between regression and classification models were made. It was found that regression models performed better than their classification counterparts when tested for generalizability. However, in all cases, the trained neural networks failed to generalize when the appropriate environmental variability was not included in the training data. To improve network performance, transfer learning can modify a pre-trained network to make predictions on data that was recorded under different conditions than the original dataset. Transfer learning was used to update the pre-trained model with a smaller set of data measured at different water temperatures. The resulting model better generalizes to measurements at different temperatures.