Passive acoustic classification of underwater targets using unspervised representation learning schemes

Abstract

The detection and classification of underwater targets has gained significant research interest in the past few decades due to its strategic and commercial importance. As the human-assisted sonar classification turns out to be tedious and time-consuming, there is a need for automated intelligent classifiers that could analyze the received signals and identify the targets. The process of automated classification involves the extraction of characteristic features from the target signatures, followed by applying a classifier algorithm, mostly based on some form of pattern matching techniques. However, the intrinsic complexity associated with the passive sonar data due to the presence of numerous ambient noise sources and several channel-induced effects makes the target recognition task extremely challenging. newlineThis thesis attempts to address the challenges in underwater target classification using approaches primarily based on unsupervised representation learning, in order to identify the manifolds where the target classes are optimally separated. Several expeditions have been carried out in the Indian Ocean, especially along the shallow waters of the Arabian Sea, to acquire adequate data for the proposed work. Various distinctive hand-engineered features in conjunction with several classifier algorithms have been utilized to improve the recognition performance. Furthermore, to prevent overfitting in the absence of sufficient samples, appropriate regularization schemes are employed in the classifier design.A GPU-based high-performance computing cluster capable of doing sufficient parallel processing has been implemented to meet the computational requirements. A highly data-efficient classifier system based on unsupervised deep generative models has also been developed, yielding a classification accuracy much better than the classical signal processing pipeline. newline