AUTHOR=Wang Yan , Zhang Hao , Huang Wei , Zhou Manli , Gao Yong , An Yuan , Jiao Huifeng 

TITLE=DWSTr: a hybrid framework for ship-radiated noise recognition

JOURNAL=Frontiers in Marine Science

VOLUME=Volume 11 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/marine-science/articles/10.3389/fmars.2024.1334057

DOI=10.3389/fmars.2024.1334057

ISSN=2296-7745

ABSTRACT=Passive ship-radiated noise recognition is a highly significant research domain due to its profound military and economic implications. However, its progress is hindered by the considerable challenges encountered in real oceanic environments. The interference of natural sounds and the attenuation/distortion of ship-radiated noise signals make it difficult to extract relevant acoustic characteristics. Ship type differentiation is also complicated by noise signature variability and realtime analysis demands. To address these issues, a new method called DWSTr is proposed in this paper. It merges a depthwise separable convolutional neural network and a Transformer framework.The convolutional component specializes in extracting local acoustic features, enabling the model to distinguish relevant information from interference. With its attention mechanism, the Transformer framework captures global and long-range dependencies in the data, helping to mitigate the effects of interference and distortion, as well as the variability of noise signature. Moreover, the network's computational efficiency enables real-time analysis, making it suitable for immediate processing and response. Experimental results on the ShipsEar dataset demonstrate an impressive recognition accuracy of 96.5% for the proposed method.