Comparative Analysis of Artificial Neural Networks for Classification of Real and Generated Hydroacoustic Signals

In this article, we present the results of a study of the use of selected artificial neural network architectures for the classification of marine objects generating acoustic signals. The training data were acquired from various types of platform, including manned underwater vehicles, autonomous underwater vehicles, diver propulsion vehicles, surface vessels, and high-speed motorboats. A total of 14 models were trained and evaluated on two complementary datasets, consisting of a simulated submarine dataset and a real-world hydroacoustic dataset acquired with a DigitalHyd TP-1 hydrophone within the NARLUGA system. We considered the following neural network architectures: multi-layer perceptron, long short-term memory, gated recurrent unit, and convolutional neural network. This paper provides a detailed description of the model architectures, the training parameters, and the preprocessing steps used to adapt the data representation to each type of model. On the simulated dataset, feature-based models (based on gammatone cepstral coefficients) achieved a test accuracy of above 99%, whereas recurrent models trained directly on long raw sequences did not converge under the training settings used here. On the real-world dataset, the best feature-based model reached a weighted test accuracy of approximately 85%. The results confirm the validity of deep-learning-based algorithms for passive hydroacoustic classification, and highlight the importance of feature representation for robust performance and practical deployment.