Underwater Acoustic Target Recognition Based on ReLU Gated Recurrent Unit

Abstract

In general, the traditional acoustic models'(e. g. Gaussian mixture model, GMM) for underwater acoustic target recognition (UATR) performance in sequential data has so far been disappointing. In contrast, Recurrent Neural Network (RNN) is a powerful tool for sequential data. This paper investigates the Gated Recurrent Unit (GRU), which is a variant of the RNN. We use the Rectified Linear Unit (ReLU) activation, and carry out experiments on the real acoustic dataset of ships. The recognition accuracy on the dataset with noise reach 86.1%, on dataset without noise reach 96.4%. Both performances are better than other baselines.