水下圆柱壳冲击声的听觉辨识与特征提取

Transient noise radiated from underwater vechicles are diffcult to be degraded by current noise control techniques, which will become an important information source for next generation advanced passive sonar systems. Traditional time-frequency signal processing and auditory perception based features extracton approach cannot meet the requirement of practical acoustic target recognition because of both large number of features and lower anti-interference ability. Aiming at a cylindrical shell strcuture, which is considered as a tyical structure of underwater vechicles, this project proposes a noval sound source feature extraction and optimization approach based on sparse representation of impact sounds and auditory perception principle. Firstly impact sounds are sparsely coded, and constant acoustic cues are extracted to used as target features based on an information accuracy measure. In other hand, from the view of point of auditory identification, it is well known that experienced and excellent listeners have intrinsic advantages for extracting and intergrating acoustic cues in strong noise and complicated interference environment. Hence, we construct a decsion-making model from identification results of excellent listeners and take perception weights as a measure and criterion, which are used to select and optimize signal features, finally obtain sound source features related with sources’physical attributes and being robust to noise and interference. This project will result in a high performance system used for automatic recognition of underwater target in noisy and interferential environment.

水下航行器发出的瞬态声很难用现有声隐身技术加以降低，已成为下一代先进被动声呐系统重点捕获的信息源。传统的基于时频信号处理和听觉感知原理的稳态声特征提取方法存在特征数量大、抗噪声和抗干扰能力弱等缺点。本项目以圆柱壳作为水下航行器的典型结构，提出一种以冲击声稀疏表达和听觉辨识原理为基础的声源特征提取与优化方法，进一步提高水下目标自动识别率和抗干扰性能。首先，利用冲击声信号的稀疏性对其进行稀疏编码，然后以信息精确度为评价指标提取反映声源物理属性的恒定声线索作为目标特征。在听觉辨识中，有经验和优秀的人类听者具有从强烈噪声和复杂干扰背景中提取和整合声线索的先天优势，为此建立优秀听者分类决策模型，以感知权重为测度和评价指标，实现声源特征的选择与优化。本项目的完成，将获得与声源物理属性紧密相关、对背景噪声和干扰稳健的声源特征，为实现高性能的水下目标自动识别系统提供新思路。

本项目基于合成声模型和录制冲击声分别从时域、频域、时频域、稀疏分解以及听觉感知等方面分析并提取特征，最终实现了圆柱壳材料的自动识别。首先提出了基于FEA-MEM-BEM（有限元-模态展开-边界元）和DSS-MEM-DMATM（离散状态空间-模态展开-延时模态传递矩阵）的合成声方法，其中后者在保证实验精度的条件下具有更高的计算效率。然后在合成冲击声连续统的基础上分析了冲击声在时域、频域和时频域随材料参数变化的差异以及导纳特征随材料属性变化的规律，确定了冲击声中含有反映与声源物理属性相关的信息。最后分别基于时频特征、正交稀疏分解特征和听觉感知特征进行了材料识别实验。结果表明，三类特征在有阻尼圆柱壳的冲击声识别上都具有较高的识别率，三类特征进行组合后能够进一步提升圆柱壳材料的识别率。本项目的完成，获得了识别率较高的时频特征、抗噪性较好的系数分解特征以及物理意义明确的听觉感知特征，为含噪条件下的圆柱壳材料的高性能自动识别提供了新的思路。