摘 要
在導(dǎo)航、移動(dòng)通信、雷達(dá)、電子戰(zhàn)系統(tǒng)、聲納以及地震等諸多領(lǐng)域，測向都是一個(gè)熱點(diǎn)問題。測向算法也被稱為空間譜估計(jì)、波達(dá)方向估計(jì)、到達(dá)角估計(jì)或者方位估計(jì)。其實(shí)，所謂波到達(dá)角估計(jì)（DOA）的目標(biāo)就是從一系列接收的信號中（包括噪聲）估計(jì)出我們所感興趣信號的方位。
在過去的幾十年中，一些有效的高分辨率算法得到了很好的發(fā)展，如MUSIC算法、ESPRIT算法。然而，這些傳統(tǒng)的方法通常運(yùn)用線性代數(shù)的方法，需要計(jì)算大量的矩陣反演，進(jìn)而消耗大量的時(shí)間，因此，它們不能滿足實(shí)時(shí)性的要求。
隨著計(jì)算智能技術(shù)的飛速發(fā)展，人們開始研究通過學(xué)習(xí)大量的樣本來解決來波到達(dá)角估計(jì)問題，而神經(jīng)絡(luò)因其非線性映射及泛化能力無疑被人們認(rèn)為是解決這一問題的強(qiáng)有力工具。神經(jīng)網(wǎng)絡(luò)的優(yōu)點(diǎn)在于建模過程是采用訓(xùn)練樣本構(gòu)造神經(jīng)網(wǎng)絡(luò)，而不再是精確的數(shù)學(xué)方程式，在實(shí)際的應(yīng)用環(huán)境中，采集到的訓(xùn)練樣本可以將噪聲、信號模型、信噪比、傳輸通道等因素考慮進(jìn)去，而無需進(jìn)行特征值分解、譜峰搜索、并且能快速實(shí)現(xiàn)并行就算，有望應(yīng)用到實(shí)際工程。
本文的主要工作包括：
1. 研究了一種基于選擇性神經(jīng)網(wǎng)絡(luò)集成的單信號源來波到達(dá)角估計(jì)算法。文中通過粒子群優(yōu)化算法合理選擇組成神經(jīng)網(wǎng)絡(luò)集成的各個(gè)神經(jīng)網(wǎng)絡(luò)，使個(gè)體間保持較大的差異度，減小“多維共線性”和樣本噪聲的影響，以此建立單信號源的波到達(dá)角估計(jì)模型。仿真結(jié)果表明，該方法同BP神經(jīng)網(wǎng)絡(luò)、RBF神經(jīng)網(wǎng)絡(luò)、泛回歸神經(jīng)網(wǎng)絡(luò)、MUSIC算法相比在處理單信號源的來波到達(dá)角估計(jì)時(shí)具有更好的準(zhǔn)確性，進(jìn)而有望應(yīng)用在實(shí)際的定位系統(tǒng)中。
2. 研究了利用粒子群算法優(yōu)化BP神經(jīng)網(wǎng)絡(luò)來提高DOA估計(jì)性能。傳統(tǒng)的BP神經(jīng)網(wǎng)絡(luò)易陷入局部最優(yōu)，因此采用粒子群算法對網(wǎng)絡(luò)的權(quán)值和閾值進(jìn)行優(yōu)化，并將其應(yīng)用到來波到達(dá)角估計(jì)中。另外，本方法僅利用陣列協(xié)方差矩陣的第一行作為來波方位特征，與常用的協(xié)方差矩陣上三角特征相比，在不損失有效方位信息的基礎(chǔ)上使特征維數(shù)得到極大地降低。經(jīng)仿真實(shí)驗(yàn)證明：同經(jīng)典的RBF神經(jīng)網(wǎng)絡(luò)方法相比，基于本文方法的神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)更簡潔，泛化性能更好，來波方位估計(jì)精度更高。
3. 在不同的信噪比、陣元數(shù)目、快拍數(shù)以及信號角度間隔的變化下，比較和分析PSO-BP神經(jīng)網(wǎng)絡(luò)和經(jīng)典的RBF方法進(jìn)行DOA估計(jì)的性能。實(shí)驗(yàn)證明在同等的實(shí)驗(yàn)條件下，PSO-BP神經(jīng)網(wǎng)絡(luò)方法的性能明顯優(yōu)于RBF方法。
關(guān)鍵詞：來波到達(dá)角估計(jì)；粒子群算法；神經(jīng)網(wǎng)絡(luò)；精確性

Abstract
Direction finding is one of the major problems for many applications such as radar, navigation, mobile communications, electronic warfare systems, sonar and seismology. Direction finding algorithms have also been known as spectral estimation, direction-of-arrival (DOA) estimation, angle of arrival (AOA) estimation, or bearing estimation. In fact, the goal of DOA estimation algorithm is to estimate the direction of the signal of interest from a collection of noise ‘‘contaminated’’ set of received signals.
Direction finding have followed an evolutionary trend. In the previous decade, some powerful and high-resolution methods for DOA estimation such as MUSIC and ESPRIT have already been developed. However, these conventional methods usually consumed a lot of time, because they use the method of linear algebra as to require the calculation of a matrix inversion. Therefore, they were not able to meet real-time requirements.
With the rapid development of computational intelligence technology, people begin to research to solve the direction of arrival estimation problem by learning a lot of samples. Neural Network is considered to be a powerful tool to solve this problem as result of the nonlinear mapping and generalization ability. The advantage of this method lies that modeling process is using the training samples to structure neural network instead of accurate mathematical equations. In practice, the collected training samples can take the noise, signal noise ratio, transmission channel model, and other factors into account, without the need for eigen value decomposition, spectral peak searching, and calculations can be fast parallel implementation, which is expected to be applied to practical engineering.
The main contributions of this paper are as follows.
1. This thesis studied a method for the direction of arrival estimation about one signal using Selective Neural Network Ensemble (SNNE). Selective Neural Network Ensemble based on Particle Swarm Optimization (PSO) is proposed to solve the direction of arrival estimation of one signal. The basic idea of the method is to optimally select Neural Network to construct Neural Network Ensemble with the aid of PSO. This may maintain the diversity of Neural Network and decrease the effect of co-linearity and noise of sample. The computer simulation shows that the method is more excellent compared with BPNN, RBFNN, GRNN and MUSIC algorithms about the one signal of DOA estimation, which makes it feasible to carry out in practical interference location system.
2. Particle swarm optimization is used for optimization of BP neural network to improve the performance of direction of arrival estimation. Due to the fact that BP neural network is inclined to be trapped in local extreme, a novel network, particle swarm optimization based BP neural network, is proposed to solve the above shortcoming, it is applied to direction of arrival estimation for study. This thesis only presents using the first row of cor..