Deep Learning Models for Passive Sonar Signal Classification of Military Data

The noise radiated from ships can be used for their identification and classification using passive sonar systems. Several techniques have been proposed for military ship classification based on acoustic signatures, which can be acquired through controlled experiments performed in an acoustic lane. The cost for such data acquisition is a significant issue since the ship and crew have to be dislocated from the fleet. In addition, the experiments have to be repeated for different operational conditions, taking a considerable amount of time. Even with this massive effort, the scarce amount of data produced by these controlled experiments may limit further detailed analyses. In this paper, deep learning models are used for full exploitation of such acquired data, envisaging passive sonar signal classification. A drawback of such models is the large number of parameters, which requires extensive data volumes for parameter tuning along the training phase. Thus, generative adversarial networks (GANs) are used to synthesize data so that a larger data volume can be produced for training convolutional neural networks (CNNs), which are used for the classification task. Different GAN design approaches were evaluated and both maximum probability and class-expert strategies were exploited for signal classification. Special attention was paid to how the expert knowledge might give a handle on analyzing the performance of the various deep learning models through tests that mirrored actual deployment. An accuracy as high as 99.0 +/- 0.4% was achieved using experimental data, which improves upon previous machine learning designs in the field.