中文摘要：

本文研究了一种集相空间重构的几何特征和盒数维为一体的计算复杂度较低的非线性特征提取方案．首先利用基于统计学方法的C—C方法提取嵌入维和时延，根据舰船辐射噪声的嵌入维确定维数范围，在维数范围内采用时延法进行相空间重构，将舰船辐射噪声时间序列扩展到高维空间，在高维空间中计算时间子序列的相关系数，并从相关系数中寻找阈值范围提取具有区分性相似性重复度，将相似性重复度进行升序重排画相似性重复度曲线，提取相似性重复度平均值这一几何特征，然后从盒子维的数学分析出发研究盒子维数的简化工程计算模型，三类舰船辐射噪声的盒子维分布图验证盒子维特征具有一定的区分性，最后设计后传播（BP）神经网络分类器，BP神经网络采用两层结构的批处理模式．推导特征的类间距离和类内距离，并通过实际海上数据仿真，计算结果表明结合特征的类间距离较大，类内距离较小，验证结合特征提取方法的有效性，三类目标的识别率表明该方法具有一定的水声目标识别能力．

英文摘要：

The mechanism of ship radiated noise is correlated with dynamics system, and the analyzing algorithm of nonlinear time sequence is an effective approach to the signal generated by dynamics system based on the phase space recon struction technique. An effective and low complexity nonlinear feature extraction algorithm is researched in the paper. The embedded dimension and the time delay are extracted through C C method based on the statistical method,the embedded dimension range of ship radiated noise is determined according to the embedded dimension, the phase space is reconstructed in the range of embedded dimension,and the correlation coefficient is computed between the time subsequence in the high di- mension space. Then the similar sequence repeatability is extracted from the correlation coefficient under the threshold condition, and the similar sequence repeatability can differ with each other. The similar sequence repeatability is ascending sorted and the curve of the similar sequence repeatability is depicted. The geometric characteristic of the mean of the similar repeatability is extracted. And then the simple engineering computation model of box-dimension is proposed through its mathematic analysis,and the distribution graph of three kinds of ship radiated noise shows the distinguish ability of the bo~ dimension characteristic. In the end,a back-propagation neural network classifier is designed, the structure of the classifier is two-layer and the processing model is batching model. The formulas of the distance between the classes and the inne-class distance are deduced,and the simulation result of sea trial data show bigger distance between classes and smaller inne-class distance. The algorithm combines the geometric feature based on phase space reconstruction technique with box-dimension feature, ar/d the results show the good recognition ability and prospective application of the proposed algorithm.