【作者】 朱广平；

【导师】 孙辉；

【作者基本信息】 哈尔滨工程大学 ， 水声工程， 2009， 博士

【摘要】 主动声纳在探测沉底或掩埋目标时海底混响是主要干扰。本文围绕着混响背景中信号检测的主题,结合高阶统计分析、支持向量机和时频滤波的理论及技术,研究了非高斯混响干扰中的信号检测技术和基于时频滤波的宽带信号检测技术。在混响统计模型的基础上,研究了混响干扰的统计特性,推导了受非高斯分布混响干扰时目标回波信号的概率密度。通过分析混响和目标回波的高阶统计特性,研究了基于高阶统计分析的特征检测方法(HOSA-SVM)。该检测方法在高信混比下检测性能较好,然而,由于只是提取3、4高阶统计特性,并且信混比对特性差异的影响比较严重,在低信混比下HOSA-SVM检测性能不太理想。针对HOSA-SVM检测方法的不足,提出了具有样本选择的支持向量机(DE-SVM)检测方法。该方法直接利用原始数据构造检测器,采用单类支持向量机进行样本选择,在不降低检测性能的前提下,有效的减少了训练时间,解决了训练样本数量与检测性能之间的矛盾。分析了该检测器的性能,在非高斯混响中DE-SVM检测算法的检测性能优于匹配滤波检测器。然而DE-SVM中的核函数及其参数的选择对检测性能的影响较大。针对该问题,首先理论分析了核函数在特征空间中的作用及对检测性能的影响,提出了基于数据驱动的自适应核函数的设计原则。利用混响和目标回波的高阶统计特性上的差别,设计了基于数据高阶统计量的自适应特征核函数。数学证明了该核函数在满足Mercer定理的前提下,能有效扩大两类样本在特征空间中的欧氏距离。然后将自适应特征核支持向量机(AFK-SVM)应用于混响背景中的信号检测,结合实际应用给出了训练和检测算法,并进行了湖上实验研究。最后分析了其检测性能,当混响背景为非高斯分布时,其检测性能优于基于传统核函数的支持向量机以及匹配滤波检测器。为了提高LFM信号的Wigner-Ville Hough变换(WHT)检测方法在低信噪比下的检测性能,研究了两种时频滤波方法。第一种方法首先分析了时频域上噪声和混响干扰的统计特性,然后采用基于统计特性的二维均值滤波和Wiener滤波方法抑制干扰。当信噪(混)比较高时,对噪声和混响有一定的抑制作用；然而,在较低的信混比下,二维均值滤波和Wiener滤波效果均不理想。第二种方法根据LFM信号与噪声和混响干扰在时频域上能量聚集性的不同,提出了自适应轴向均值脊波变换(XWVD-M-FRIT)的时频滤波方法。该方法首先采用互Wigner-Ville变换(XWVD)代替Wigner-Ville变换(WVD),在时频域上提高了信噪比,且避免了信号为多分量时各分量间的交叉项干扰。由于信号和噪声的统计特性不同,考虑到滑动窗长对均值滤波的影响,设计了自适应轴向均值滤波器用于滤除噪声和混响。然后采用脊波变换滤波进一步滤除噪声。最后采用Hough变换检测信号。对该滤波及检测方法进行了实验室及海上实验研究。分析了该时频滤波方法的统计性能,在低信噪(混)比下,XWVD-M-FRIT滤波后能有效的抑制噪声或混响干扰,使得该方法与WHT检测方法相比能够更有效的检测到目标回波信号。更多还原

【Abstract】 The sea-floor reverberation is main disturbance when the active sonar detecting bottom or buried targets. In this paper, concerning about the subject of signal detection in reverberation, we studied technology of signal detection and wide-band signal filtering based on time-frequency analysis using theory of high-order statistic analysis, support vectors machine and time-frequency filtering.The probability density function of targets echo in non-Gaussian distribution reverberation was droved based on statistical model of reverberation. The high-order statistic of reverberation and target echo was analyzed. Referring to idea of pattern recognition and classification before detection, the method using high-order statistics and support vectors machine (HOSA-SVM) was studied for detecting targets echo in the reverberation. But in low signal to reverberation ratio, the performance is not good.For the deficiency of HOSA-SVM detection method, we studied on directly constructing detector from SVM using original data (DE-SVM). Using one-class SVM to choose effective data for reducing training time, the problem of needing magnitude of data to achieve good performance was solved. In non-Gaussian reverberation, the performance of DE-SVM is better than matched filter detector. But the kernel function and parameter of DE-SVM detector seriously effected on the detection performance. So effect of kernel in feature space was analyzed. And the principle of designing adaptive kernel function based on data driving was proposed. Using the difference of high-order statistics between targets and reverberation, the adaptive kernel function based on high-order statistics was designed. It is proved that it enlarges the distance of two kinds of samples using feature kernel, and the kernel also satisfies the Mercer theorem. The feature kernel support vector machine was applied for signal detection in Gaussian and non-Gaussian reverberation. The training and detecting algorithms in practice were given. The results of experiment and simulation show that when selecting statistics existing great difference as feature and the reverberation is non-Gaussian distribution, its performance is better than matched filter and support vector machines based on traditional kernel function.To solve the problem which the performance of detection was reduced in the low signal to noise ratio (SNR) using Wigner-Ville Hough transform (WHT), two methods were studied. In the first method, it is to restrain noise using two-dimension mean filter and Wiener filter. When the SNR is high, this method is effective, but in the low SNR, the performance is not good. In second method, the method of XWVD adaptive mean and ridgelet transform filtering (XWVD-M-FRIT) was proposed. In this method, firstly used XWVD instead of WVD for improving SNR and avoiding the cross-components when the signal are multi-components. Due to the power distribution of signal is different from noise or reverberation in time-frequency domain and considering effect of length of splitting window, so designed adaptive axial mean filter. Then it is to restrain noise or reverberation using ridgelet transform filtering. At last, it is to detect the signal using Hough transform. The results of real and simulation experiments show, compared with WHT, in the low SNR the new method is feasible to restrain noise or reverberation in time-frequency domain for improving the performance of signal detection.更多还原