【作者】 李洪升；

【导师】 赵俊渭；

【作者基本信息】 西北工业大学 ， 水声工程， 2004， 博士

【摘要】 波束形成技术在声呐、雷达、移动通信和电子对抗与侦察等领域中有着广泛的应用。常规的波束形成需要知道阵列流形等先验知识,而且在实际应用中,阵列模型不可避免地会存在误差,即使是很小的误差,也会引起系统性能的急剧下降。尤其在水声环境中,由于海水温度、深度等条件的变化,使得实际水声阵列的结构尺寸等参数发生改变。由于实际的阵列流形与假设的不尽一致,因而导致了系统波束形成性能变差。所以在实际应用中需要进行误差校正。由于阵列校正必须频繁多次进行,且每次都要存贮大量阵列流形信息,同时要求提供校正信号源,所以,给工程应用带来许多不便。盲波束形成方法,利用了信号自身的特性,通过算法以解决模型误差问题,有效地增强了系统的自适应性和系统的稳健性。但是,盲波束形成算法实现过程中,多数都用到了矩阵求逆与/或矩阵特征分解(或奇异值分解)等运算,随着阵元数的增大使算法运算量急剧上升,难以满足实时性的要求。此外,一些自适应算法还用到了基于梯度的优化计算,在步长选择不当的情况下很容易陷入局部极值,使波束形成无法正确实现。而以神经网络为代表的计算智能技术具有大量的并行性、高度的容错性以及自学习、自适应和自组织等特点,恰好可以弥补盲波束形成算法的不足。 本文重点在引入计算智能技术实现声呐盲波束形成方面展开了研究工作。全文的主要工作概括如下: 第二章给出了声呐盲波束形成算法的基本模型和处理框架,探讨了计算智能技术实现盲波束形成的基本途径和基本方法。 第三章提出了两种神经网络实现的基于高阶累积量的声呐盲波束形成算法。一种是基于TH网络的盲波束形成算法。该算法把盲波束形成权向量的求解问题转化为神经网络参数设计问题,当网络在纳秒级时间内达到稳态时,即给出了所求权向量。而且该算法易于硬件在线实现;另一种是基于径向基函数网络的盲波束形成算法。通过改进的学习方法可大大提高了网络的学习性能,且对系统误差具有较强的稳健性。这两种算法有机地把高阶累积量可消除高斯噪声影响的特性与神经网络并行结构有机地结合在一起,提高了盲波束形成的效率。 第四章分析了基于信号循环平稳性的盲波束形成方法与性能,研究了一种互相关神经网络实现声呐盲波束形成的算法。该算法利用水声通信信号的循环平稳特性把波束形成权向量的求解问题转化为阵列接收信号互相关函数的奇异值分解问题;引入一种互相关神经网络求解阵列接收信号相关函数的奇异值,从而减小了运算的代价,可高效实现盲波束形成。该算法把信号的正交特性引入到网络初值的选取中,并依据网络的输入对权值的学习速率进行了有效的约束,进而更多还原

【Abstract】 Beamforming is used widely in sonar, radar, mobile communication and electronic countermeasure and reconnaissance, etc. Conventional (i.e.. non-blind) beamforming approaches require a reference signal or accurate knowledge of the array response vector corresponding to the desired signal. In practice, array model must have some errors unavoidably. Even though there is a very small error, system will degrade severely. It is more serious in underwater environment. Configuration and dimension of the array will change with the change of temperature and depth of seawater. As a result, beamforming can be degraded because actual array manifold is not same as the theoretical one. Because array correction must be done for many times and a lot of array manifold information is stored, and correction signal source is provided, it is very inconvenient for the project implement. Blind beamforming techniques in general exploit some signal property itself. It can decrease the model error by some algorithms. But it is necessary to compute the inverse and/or matrix eigen-decomposition of a matrix in course of implementing blind beamforming. With the increase of array element number, computing load for the algorithm will increase severely. So it is difficult to carry out the beamforming in real time. Furthermore, some adaptive algorithms use the optimization calculation based on the gradient. When the step-size parameter can not be selected correctly, it is easy to get the local extremum. Computing intelligence whose representative is neural networks has a lot of advantages, for example, parallelism, fault tolerance, self-learning, self-adapting and self-organizing. It can overcome the shortcoming of the blind beamforming algorithm.This dissertation is focused in the study of introducing the computing intelligence to carry outthe sonar blind beamforming. The content can be outlined as follows:In chapter 2, basic model and process frame of blind beamforming is presented. At the sametime, we also study the basic approach and method using the computing intelligence to carry outblind beamforming.In chapter 3, two kinds of sonar blind beamforming algorithms implemented by neural networks are presented based on higher-order cumulants. One is a based-TH neural networks blind beamforming algorithm. This algorithm transforms the question of solving the weights of beamformer into the one of parameter design of neural networks. When the neural network converges to a stable state within a circuit time constant, its output will approximate the weight vector of beamformer. Furthermore, this algorithm is easily implemented on-line by hardware. Another is based radial basis function neural networks (RBFNN) blind beamforming algorithm. Through developed learning method, learning performance is improved remarkably. And the algorithm is robust to system errors. This two kinds of methods combine the characteristic that higher-order cumulant can eliminate effect of Gauss noise with structural advantage that neuralnetwork can operate in parallel. So blind beamforming can be performed efficiently.In chapter 4, we discuss the blind beamforming method based on signal cyclostationary, and study a sonar blind beamforming algorithm based on a cross correlation neural network model. This method transforms the question of estimating beamforming weight vectors into the one of computing the SVD of the cross correlation matrix of array input signals and their frequency shift signals so as to decrease minish the computing load in the course of blind beamforming. Through using the character of signal orthogonalization in the course of selecting the initial values of the neural network and restricting the learning rate of the network weight according to the input of the neural network, we present a improved cross-coupled Hebbian learning rule to make the convergence of the network rapid. This method can provide an approach to perform beamforming in real time and restrain noise and interference.In chapter 5, we study the based eigen-structure sonar blind beamforming method. A parallel learning algorithm of the PCA (Principal Component Analysis) neural network is presents. In the algorithm, principal components are obtained simultaneously. Furthermore, the algorithm does not require recycled use of the data. Hence the time of the neural network convergence is shortened considerably. In addition, we also propose a algorithm to obtain the minimal eigen-value directly based on MCA(Minor Component Analysis) neural networks. And we extend the two kinds of algorithm to the complex field and combine the JADE (Joint Approximate Diagonalization of Eigen—Matrices) algorithm to perform the blind beamforming.In chapter 6, a new blind beamforming algorithm for multitargets is presented. Through defining a new cost function, we develop a blind beamforming algorithm which can separate non-Gaussian signals with arbitrary kurtosis. In addition, weight vector is estimated using complex coding genetic algorithm so as to improve the performance of the global convergence greatly.In chapter 7, we study the conventional beamforming and blind beamforming methods based on a vector sensor. Firstly, received signal model for the vector sensor array is given. Then, an optimizing weight vector formulation is given for beamforming of the vector sensor array ? Based on this, we define a fourth-order cumulant of received signals and blindly estimate the steering vector for the vector sensor array. In addition, we propose a steering vector blind estimation method for a pressure-velocity combination hydrophone array by developing the ESPRIT and combining PCA. At the same time, we study a kind of aerial sonar based on the vector sensor. Finally, we bring the dissertation to a close by providing some key concluding remarks of this work and an overview of current development and future trends in this field.更多还原