含噪盲源分离算法研究及其在水声信号中的应用

发布时间：2018-04-27 20:30

本文选题：含噪盲源分离 + 独立分量分析　；参考：《解放军信息工程大学》2014年博士论文

【摘要】：盲源分离是指在各个源信号均未知的情况下,根据某种条件和假设,从混合的观测信号中分离出这些源信号的方法。在过去的几十年中,盲源分离技术由于其潜在的应用价值得到了众多学者的关注,发展迅速。在实际应用中,无论是通信信号、语音信号、医学信号还是水声信号等等都不可避免地会被各种形式的噪声和干扰影响。大多数盲源分离算法在没有噪声时具有最优的性能,当观测信号受到噪声污染时其性能会急剧下降,严重时可能导致分离失败。由于无法获得更多关于源信号的先验信息,并且信道参数未知,使得含噪盲源分离问题较无噪时的盲源分离问题更加复杂。目前,对含噪盲源分离问题的研究成果相对较少。本文紧紧围绕含噪盲源分离问题进行深入研究,主要对基于偏差去除的含噪盲源分离算法和基于去噪的含噪盲源分离算法两部分内容进行了研究,取得了一定的成果,最后将含噪盲源分离算法应用于水声信号中,验证了含噪盲源分离算法对水声信号盲分离的有效性和可行性。本文的主要工作概括如下:1.研究了盲源分离算法中的源信号数目估计问题,针对已有的源数估计算法在信噪比较低时的性能较差的问题,提出了一种基于奇异值分解的转折点检测算法,实验结果表明该算法提升了低信噪比时的估计性能。对Fast ICA和RobustICA这两种算法的性能进行了详细的分析和比较,分别针对亚高斯源、超高斯源以及由亚高斯源和超高斯源组成的混合源在不同信噪比以及不同采样点数的情况下进行了仿真实验,结果表明总体而言Robust ICA算法的性能更加稳健。2.对基于偏差去除的含噪盲源分离算法进行了研究。针对盲源分离超定模型提出了基于特征值分解的准白化FastICA算法,该算法通过特征值分解获得噪声方差的估计,从而可以对含噪信号中的有用信号进行白化,去除噪声引入的偏差,同时降低信号空间的维数,将超定模型转化为正定模型。进一步针对每路含噪混合信号的信噪比不完全相同的情况,提出了一种基于迭代的二次白化FastICA算法,该算法能够准确估计出每路混合信号的噪声方差,从而更准确地去除噪声引入的偏差,使得对有用信号的白化更加有效。仿真实验表明了这两种算法解决含噪盲源分离问题的有效性。3.研究了去噪算法与盲源分离算法相结合的策略,目前主要有去噪预处理、去噪后处理以及去噪预处理与后处理的级联方式。本文对去噪预处理的算法进行了深入研究,并讨论了去噪预处理与后处理的级联方式下的含噪盲源分离问题的解决策略。由于已有文献中提出的串行级联方式和并行级联方式没有充分利用分离算法得到的分离矩阵以及分离出的估计信号,本文提出了一种改进的预处理与后处理的并行级联方式,该方式同时使用了并行级联方式中的两路输出信号,充分利用了能够获得的有用信息,在进行去噪后处理之前先提高了含噪信号的信噪比,从而减小了去噪后信号的失真程度。实验仿真表明该方法较已有的串行和并行两种级联方式的分离性能更优。4.对小波去噪算法作为去噪预处理算法进行了深入研究。针对较低信噪比条件下小波去噪算法性能不佳的问题提出了改进的基于平移不变量的小波去噪算法,该算法对小波去噪算法中的关键参数进行了优化,并提出了一种更加稳健的噪声方差估计算法——高频系数滑动窗口法,同时缩小了平移不变量的范围,在减少运算量的同时去噪效果几乎没有降低。将该去噪算法作为去噪预处理应用于含噪盲源分离问题中,仿真实验验证该算法的有效性。针对高斯色噪声的去噪问题,提出了一种小波去噪算法,该算法使用改进的分层gcv阈值估计算法,将其应用于含有高斯色噪声的盲源分离问题,仿真实验证明了该算法能够更加有效去除高斯色噪声,提升盲源分离算法的性能。5.对基于经验模态分解的去噪算法作为去噪预处理算法进行了研究。针对传统的经验模态分解去噪算法存在的去噪不彻底以及有用信号被当作噪声滤除的问题,提出了一种分段emd阈值去噪算法,该算法首先使用平均周期法将经验模态分解得到的若干个本征模态函数分成噪声主导部分和信号主导部分,对噪声主导部分使用已有文献中的阈值估计算法,对信号主导部分使用新的下降更快的阈值估计算法,然后使用改进的阈值收缩算法对每个本征模态函数进行阈值收缩处理,重构信号。该算法能够克服已有算法的缺点,具有更好的去噪性能。将该算法应用于含噪盲源分离中,能够显著提升盲源分离算法的性能。接下来研究了高斯色噪声条件下基于经验模态分解的去噪算法。研究发现,同高斯白噪声相比,高斯色噪声经过经验模态分解后第一个imf分量的幅值相对较小,其余幅值下降速度相对平缓,根据高斯色噪声的这些特性对已有的阈值算法中的参数进行调整,同时采用了分段阈值估计算法,将该去噪算法应用于含有高斯色噪声的盲源分离问题中,实验结果证明了该算法的有效性。6.研究了含噪盲源分离算法在水声信号处理中的应用。由于计算机仿真信号在科研中具有特殊的优势,本文首先对海洋环境噪声、水声测试信号、水声通信信号以及舰船辐射噪声进行模拟仿真,然后分别针对盲源分离的正定模型和超定模型进行了仿真实验。对正定模型而言,主要使用了针对高斯色噪声的基于分层gcv阈值估计的小波去噪算法和基于分段emd阈值去噪算法这两种算法作为去噪预处理,然后使用robustica算法进行分离,仿真实验验证了这两种算法能够有效去除水声信号中的海洋环境噪声,显著提升盲源分离算法的性能。同时还讨论了算法性能同采样率之间的关系,实验表明,由于提高采样率能够改变噪声在有用信号宽带内的分布从而有助于提升去噪算法的性能,进一步提升盲源分离算法的性能。针对超定模型,主要使用了本文提出的两种基于偏差去除的含噪盲源分离算法:基于特征值分解的准白化fastica算法和基于迭代的二次白化fastica算法。仿真结果表明,直接使用fastica算法,对于水声通信信号,snr32db时分离得到的信号均方误差降至10~(-2)数量级,使用本文提出的算法,SNR14dB时信号的均方误差就已降至10~(-2)数量级;对于舰船辐射噪声,SNR22dB时分离得到的信号均方误差降至10~(-2)数量级,使用本文提出的算法,SNR10dB时信号的均方误差就已降至10~(-2)数量级。因此在含有噪声条件下本文提出的算法能够获得令人满意的分离效果。同时验证了采样率的提升对于基于偏差去除的分离算法的性能几乎没有影响,这是因为该算法不受噪声分布状态的影响,因此该算法对于含噪条件下的水声信号盲分离十分有效。
[Abstract]:Blind source separation refers to the method of separating these source signals from mixed observation signals under certain conditions and assumptions. In the past few decades, blind source separation technology has been developed rapidly because of its potential application value. In practical applications, no matter communication is in communication. Signals, speech signals, medical signals or underwater acoustic signals are inevitably affected by various forms of noise and interference. Most blind source separation algorithms have the best performance without noise. When the observed signals are polluted by noise, their performance will drop sharply, and the separation failure may result. More about the prior information of the source signal, and the unknown channel parameters make the problem of noisy blind source separation more complex than the blind source separation problem when the noise is noiseless. At present, the research results on the problem of noisy blind source separation are relatively few. This paper focuses on the problem of noisy blind source separation, and mainly deals with the denoising based on the deviation removal. The blind source separation algorithm and the denoising blind source separation algorithm based on denoising two parts are studied, and some results are obtained. Finally, the denoising blind source separation algorithm is applied to the underwater acoustic signal, and the validity and feasibility of the blind source separation algorithm of the noisy blind source separation algorithm are verified. The main work of this paper is summarized as follows: 1. In order to solve the problem of the number of source signals in the blind source separation algorithm, a turning point detection algorithm based on singular value decomposition is proposed to solve the poor performance of the existing source number estimation algorithm when the signal to noise ratio is low. The experimental results show that the algorithm improves the estimation performance of low signal to noise ratio. The two algorithms for Fast ICA and RobustICA are calculated. The performance of the method is analyzed and compared in detail. The simulation experiments are carried out for the subGauss source, the super Gauss source and the hybrid source composed of the sub Gauss source and the super Gauss source at different signal to noise ratio and the number of sampling points. The results show that the performance of the Robust ICA algorithm is more robust on the basis of the deviation removal based on the deviation. The blind source separation algorithm is studied. A quasi whitening FastICA algorithm based on eigenvalue decomposition is proposed for the blind source separation model. The algorithm obtains the estimation of the noise variance by eigenvalue decomposition, thus whitening the useful signal in the noisy signal, eliminating the error introduced by the noise, and reducing the signal space at the same time. According to the condition that the signal to noise ratio of the mixed signal with noise is not exactly the same, an iterative two whitening FastICA algorithm is proposed. The algorithm can accurately estimate the noise variance of each mixed signal and remove the error introduced by the noise more accurately. The whitening of useful signals is more effective. The simulation experiment shows the effectiveness of the two algorithms to solve the problem of noisy blind source separation..3. studies the strategy of combining the denoising algorithm with the blind source separation algorithm. At present, there are mainly denoising preprocessing, de-noising post processing, denoising preprocessing and post-processing. The algorithm has been studied deeply, and the solution strategy of blind source separation in the cascade mode of de-noising preprocessing and post-processing is discussed. Since the serial cascading and parallel cascading methods proposed in the literature have not fully utilized the separation matrix and separated estimation signals from the separation algorithm, this paper proposes a new method. An improved parallel cascade of preprocessing and post-processing, which uses two output signals in parallel concatenation, makes full use of the useful information that can be obtained, and improves the signal to noise ratio of the noise signal before de-noising processing, thus reducing the degree of distortion of the signal after de-noising. Compared with the existing serial and parallel two cascading methods, the separation performance is better than that of the two cascaded methods. The wavelet denoising algorithm is studied as a denoising preprocessing algorithm. An improved wavelet denoising algorithm based on translation invariant is proposed for the poor performance of the wavelet denoising algorithm under the condition of low signal to noise ratio. The key parameters in the algorithm are optimized, and a more robust noise variance estimation algorithm is proposed - the high frequency coefficient sliding window method, which reduces the range of the translation invariants at the same time. At the same time, the denoising effect is almost not reduced. The denoising algorithm is used as denoising preprocessing to denoising blind source separation. In the problem, the effectiveness of the algorithm is verified by simulation experiments. A wavelet denoising algorithm is proposed for the denoising problem of Gauss color noise. The algorithm uses an improved hierarchical GCV threshold estimation algorithm to apply it to the blind source separation problem with Gauss color noise. The simulation shows that the algorithm can effectively remove Gauss color noise. The performance of the improved blind source separation algorithm (.5.) studies the denoising algorithm based on the empirical mode decomposition (EO) decomposition (EO) as a denoising preprocessing algorithm. A segmented EMD threshold denoising algorithm is proposed for the problem that the traditional empirical mode decomposition denoising algorithm is not completely de-noised and the useful signal is treated as noise filtering. Using the average periodic method, some eigenmode functions obtained by the empirical mode decomposition are divided into the leading part of the noise and the leading part of the signal, and the threshold estimation algorithm in the existing literature is used for the leading part of the noise, and the new reduced and faster threshold estimation method is used for the leading part of the signal, and then the improved threshold shrinkage algorithm is used. Each eigenmode function performs threshold shrinkage processing and reconstructs the signal. This algorithm can overcome the shortcomings of the existing algorithms and have better denoising performance. The algorithm is applied to the noise - containing blind source separation and can significantly improve the performance of the blind source separation algorithm. Then, the denoising based on the empirical mode decomposition is studied under the Gauss color noise condition. It is found that compared with Gauss white noise, the amplitude of the first IMF component of Gauss color noise is relatively small after the empirical mode decomposition, and the rest of the amplitude is relatively slow. According to these characteristics of Gauss color noise, the parameters in the existing threshold algorithm are adjusted, and the segmentation threshold estimation algorithm is adopted. The noise algorithm is applied to the blind source separation problem with Gauss color noise. The experimental results prove the effectiveness of the algorithm.6.. The application of the noisy blind source separation algorithm in the underwater acoustic signal processing is studied. The noise of the ocean environment, the underwater acoustic signal and the underwater acoustic communication are first introduced in this paper. The signal and ship radiation noise are simulated, and then the simulation experiments are carried out for the positive and overdetermined models of blind source separation. For the positive definite model, two algorithms are mainly used for the wavelet denoising algorithm based on the hierarchical GCV threshold estimation and the segmentation based on the segmented EMD threshold denoising algorithm for Gauss color noise. For denoising preprocessing, the robustica algorithm is used to separate them. The simulation experiments show that these two algorithms can effectively remove the ocean noise in the underwater acoustic signal and improve the performance of the blind source separation algorithm. At the same time, the relationship between the algorithm performance and the sampling rate is also discussed. The experiment shows that the noise can be changed because the sampling rate can be improved. The distribution of sound in the wideband of useful signal helps to improve the performance of the denoising algorithm and further improve the performance of the blind source separation algorithm. For the overdetermined model, two kinds of noisy blind source separation algorithms based on the deviation removal are mainly used in this paper: the quasi whitening FastICA algorithm based on eigenvalue decomposition and the two iteration based whitening based on the eigenvalue decomposition method. The simulation results show that, using the FastICA algorithm directly, the mean square error of the signal separated by the underwater acoustic communication signal and the snr32db is reduced to the 10~ (-2) order of magnitude. Using the algorithm proposed in this paper, the mean square error of the signal is reduced to the 10~ (-2) order of magnitude when SNR14dB, and the signals obtained from the radiation noise of the ship and the signals separated at SNR22dB are all. The square error is reduced to the 10~ (-2) order of magnitude. Using the algorithm proposed in this paper, the mean square error of the signal has been reduced to the magnitude of 10~ (-2). Therefore, the algorithm proposed in this paper can obtain a satisfactory separation effect under the condition of noise. Meanwhile, it is proved that the lifting of the sampling rate is almost the performance of the separation algorithm based on the deviation removal. This algorithm is not affected by noise distribution, so the algorithm is very effective for blind separation of underwater acoustic signals under noisy conditions.

【学位授予单位】：解放军信息工程大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TN911.4

【参考文献】