Learning robust features from underwater ship-radiated noise with mutual information group sparse DBN

Abstract: Classifying Underwater Acoustic Target (UAT) based on ship noise datasets is a small-sample-size classification problem. Generalization ability of features depending on prior knowledge and traditional signal processing is seriously restricted for lack of labeled data. Deep Belief Networks (DBN) can make use of abundant unlabeled data and extract features spontaneously. But the resulting features exhibit statistical dependencies. In this paper, we propose a feature learning model called sparse group DBN for ship noise. Firstly, a standard Restricted Boltzmann Machine(RBM) is pre-trained in an unsupervised learning way, then group the hidden units adaptively according to mutual information and punish the activation degree within and between the groups. Secondly, stack the improved RBM by performing a greedy layer-wise training phase, followed by the discriminative fine-tuning to achieve the sparse representation of multilayer features. Recognition experiments on real-world ship noise datasets show that, compared to the traditional features, features learned by the proposed model can improve classification performance and generalization ability significantly.Keywords: underwater acoustics, ship noise, machine learning, deep belief networks, mutual information