A self-supervised dual-channel self-attention acoustic encoder for underwater acoustic target recognition

• Main Authors: WANG, Xingmei, WU, Peiran, LI, Boquan, ZHAN, Ge, LIU, Jinghan, LIU, Zijian
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2024
• Subjects: Dual-channel mechanism; Local self-attention; Self-supervised learning; Underwater acoustic target recognition; Artificial Intelligence and Robotics; Oceanography and Atmospheric Sciences and Meteorology
• Online Access: https://ink.library.smu.edu.sg/sis_research/8738
• Institution: Singapore Management University

Due to the considerable noise and high annotation cost of passive sonar acquisition data challenges, underwater acoustic target recognition (UATR) tasks require novel solutions. In this paper, we propose a self-supervised dual-channel self-attention acoustic encoder (DSAE) for UATR tasks. First, to address the annotation challenge, we design a dual-channel self-attention acoustic encoder to unify features into self-supervised learning. Meanwhile, we propose a dynamic positive sample memory module (DMM) to enable the training samples comprehensive and balanced in self-supervised learning. Second, to address the noise challenge, we utilize a time–frequency mask to obtain space–time enhanced features. Based on our experimental results, DSAE improves the recognition accuracy and anti-noise robustness compared with other advanced acoustic learning methods. The results demonstrate that DSAE has the potential to offer significant value as a tool for UATR tasks.