摘要
海洋水声信道具有时变、空变的特征,被动式声纳接收到的目标信号复杂多变,传统水下目标识别方法难以满足当前任务要求。提出基于注意力机制改进的多特征融合水下目标识别框架,在典型声学特征基础上,通过引入对比学习无监督特征加强目标数据的特征表达,使用多维自注意力机制,分别在特征维度和时间维度高效完成深层次目标弱信息特征抽取,显著提升识别效果。通过对真实录制的水声数据集进行实验对比,证明了所提出方法的有效性。
Since the underwater acoustic channel has the time-varying and space-varying characteristics, the target signal received by passive sonar is complex and changeable, so the traditional underwater target recognition method is difficult to meet the current task requirements. In this paper, an underwater target recognition framework is proposed based on an improved multi-feature fusion with the attention mechanism. Under the typical acoustic characteristics, the feature expression of target data is enhanced by introducing unsupervised features of contrastive learning, and the multidimensional attention mechanism is used to efficiently complete feature extraction from weak information of the deep-level targets in feature dimension and time dimension, respectively, which significantly improves the recognition effect. The effectiveness of the proposed method is proved by experimental comparison with real recorded underwater acoustic data sets.
作者
徐承
李勇
张梦
汪小斌
方磊
XU Cheng;LI Yong;ZHANG Meng;WANG Xiaobin;FANG Lei(Heifei iFlytek Digital Technology Co.,Ltd,Hefei 230088,China)
出处
《移动通信》
2022年第6期91-98,共8页
Mobile Communications
关键词
水下目标识别
无监督特征
特征融合
注意力机制
underwater target recognition
unsupervised features
feature fusion
attention mechanism
