A multi-task sparse feature learning method for underwater acoustic target recognition based on two uniform linear hydrophone arrays

With the rapid development of underwater acoustic research, it becomes common that two or more hydrophone arrays are deployed in underwater acoustic experiments. The prominent target related discriminative structures such as line spectra on Time-Frequency spectrum are essential for underwater target classification task. These structures are often sparsely distributed on Time-Frequency spectrum and have characteristics of temporally correlated block structured sparsity. With the help of multi-task sparse Bayesian learning framework, these discriminative structures can be recovered or strengthened to build more noise robust features. In this paper, considering the configuration of two uniform linear hydrophone arrays placed 50m apart in a reservoir, we utilized multi-task sparse Bayesian learning method with multiple observations on two arrays to learn noise robust sparse features which can recover and strengthen the prominent structures on spectrum shared by tasks on two arrays. On an actual measurement database, the classification performance of the proposed feature was evaluated. The result shows the proposed feature learned with two-task sparse Bayesian learning framework with multiple observation is more robust to noise and can achieve higher classification accuracy compared with sparse features learned from 1-task learning with less observations.