【作者】 张之猛；

【导师】 刘伯胜；

【作者基本信息】 哈尔滨工程大学 ， 水声工程， 2009， 博士

【摘要】 多径效应是影响水下被动声纳系统检测性能的重要因素。针对该问题,论文对水声信号源分离中的盲解卷积技术进行了研究。结合已有的盲信号处理研究成果,以线性卷积混合模型为对象,分别针对时间域算法、变换域算法以及源信号数目未知条件下的盲解卷积算法等问题进行研究。提出了几类适用于水声工程领域的盲源分离方法,论文的计算机仿真及实测水声数据分析处理,取得了良好的结果,验证了本文方法的可行性和有效性。论文的主要研究内容包括：(1)基于舰船辐射噪声实测数据,分析了水声信号的非高斯分布特性,为盲信号处理技术用于水声信号源分离,提供了先验知识积累。针对舰船辐射噪声的非平稳特征,研究了循环频率估计和经验模态分解两种非平稳信号处理方法,用于提取螺旋桨噪声中的低频线谱,为后续章节中的实测数据盲源分离实验所获得的源估计信号作进一步分析,奠定了理论基础。(2)研究了两种基于二阶统计量的水声多途信道时域盲解卷积算法。针对盲源分离中存在的幅度模糊问题,提出了一种应用参考信号估计信道参数,利用自适应逆滤波分离源估计信号的盲源恢复算法。该方法可同时获得源信号的时序结构与幅度信息,这对声纳信号处理是很有帮助的；为改善长时延下时域算法计算量过大问题,提出了一种时域与频域相结合的盲解卷积算法,在频域中进行自适应滤波器权向量的更新,在时域中完成信号波形的分离。(3)分析和研究了变换域中的水声信号盲分离问题。针对实际中源信号间存在的相干分量会造成盲源分离性能下降的情况,提出了一种基于小波包变换子带分解的盲解卷积算法,通过最小化源信号间的互信息来降低源信号间的依赖性。算法可分别在时间域或小波域内进行未知源的分离。(4)分析和研究了源信号数目未知,且系统为欠定条件下的稀疏源盲分离算法的整体思路。基于短时傅里叶变换,提出了一种接收传感器数目和空间分布任意情况下的利用二分提取的欠定稀疏源盲分离算法。最后设计了实际中盲源分离的总体框架,并用于舰船辐射噪声实测数据的盲源分离实验中。更多还原

【Abstract】 Multipath effect impacts passive SONAR detection performance seriously. To solve this problem, this dissertation explored multi-channel blind deconvolution (MBD) techniques for underwater acoustic signals separation. Combined with existing research work, we studied three kinds of algorithm with the linear convolutive mixture model, which are concerned with theories in time-domain, frequency-domain and the case of source numbers unknown. Several algorithms, which are applicable to underwater acoustic engineering for multi-channel phenomenon, were proposed. Simulation and measured underwater acoustic data processing results show the good performance. The main contents of this paper are as follows:(1) Based on the real data of warship radiated noise, non-gaussianity of the underwater acoustic signal was analyzed. It provided prior knowledge for the underwater acoustic source signals separation using blind signal processing (BSP). Then, aimed at the non-stationarity of warship radiated noise, two kinds of nonstationary signal processing methods, i.e. cycle frequency evaluation and empirical modal decomposition, were studied to extract low frequency line spectrum of propeller noise. It established theoretical basis for the following work of thoroughly analyzing the evaluated source signals from the real measured data via blind source separation.(2) Based on two-order moment, two kinds of acoustic multi-channel blind deconvolution algorithms were proposed in time domain. Aimed at the obscure amplitude during blind source separation (BSS), a new blind source signal restoration algorithm was proposed, which evaluated channel parameters through reference signals, and separ..ted source signals through inverse adaptive filtering. It could obtain both source signals’time structure and amplitude information, which is helpful to SONAR signal processing. Then, to improve large computation of BSS with huge delay in time domain, a fast multi-channel blind deconvolution algorithm was proposed. It updated weight vectors in frequency domain and separated the source signals in temporal domain.(3) MBD algorithm was investigated in transform domain. Owing to the actual source signals’mutual dependency, the performance of BSS algorithms descended heavily. So, a new algorithm, which used wavelet packets transformation for sub-band decomposition, was proposed to process statistically dependent sources. By minimizing the source signals’mutual information, the dependency of the source signals was reduced. The algorithm can be carried out both in temporal domain and wavelet domain.(4) After investigating the whole tenor of underdetermined sparse source separation technique when the source numbers unknown, a novel blind source separation algorithm was proposed, which utilized binary mask (BM) method, and extended the unrestricted receiving sensors both in dimension and the condition of spatial distribution, based on short-time Fourier transformation. In the end, a general framework for BSS problem was set up and applied in actual underwater acoustic signal processing for warship radiated noise seperation.更多还原