Passive sonar automated target classifier for shallow waters using end-to-end learnable deep convolutional LSTMs

Abstract Automated target recognition systems are increasingly employed in sonar systems to reduce manning and associated challenges. Although passive acoustic target recognition is an exceptionally challenging endeavor especially in shallow water scenarios, it is being used by naval forces of the world by virtue of its inherent advantages compared to the alternatives. In order to address these challenges as well as to exploit the latent and subtle features in the signal stream from the hydrophones, an end-to-end differentiable architecture is proposed in this paper. Here the key strategy is to rely on the data, instead of relying on the prior knowledge about the data. The raw acoustic signals from the hydrophones are directly fed to a pre-initialized 1-dimensional convolutional layer followed by a cascade of 2-dimensional convolutional spectro-temporal feature learners. Various auditory scales are used for pre-initializing, so as to emphasize the frequencies of interest. In order to better capture the temporal relations, a Bidirectional-LSTM layer with a trainable attention module is employed. The best configuration of the proposed classifier system yields an accuracy of 95.2% on a large acoustic dataset, collected from the shallows of the Indian ocean.