摘要： 针对流形学习方法用于声目标识别时易受噪声干扰的情况，提出一种加权邻域重构算法，采用加权迭代方式构造出带噪流形子曲面中最能反映该曲面变化趋势的曲线，通过拓展该曲线对带噪流形子曲面进行重构，利用新曲面计算低维嵌入. 该算法在去除噪声的同时，最大限度地保持了原流形曲面的变化趋势，是一种适用于声目标识别的算法. 在公开数据库和低空飞行目标实际数据中进行实验，结果表明在识别正确率及运行时间上，本文提出的算法相对于其他3 种对比算法均取得了较好的效果.

Abstract: Manifold learning methods are sensitive to noise especially in acoustic targets recognition. To deal with this problem, we present a novel manifold learning algorithm for noisy manifold, termed weighted neighborhood reconstruction (WNR). The algorithm builds a curve that can best reflect the trend of the noisy manifold sub-surface. The curve is extended to reconstruct the manifold sub-surface and calculate low dimensional embedding on the new surface. The proposed algorithm can minimize noise effects on manifold while keep the original surface trend. The algorithm is tested on public database and low attitude flying targets acoustic signal. Experiment results show that the proposed algorithm is robust against noise, and outperforms the other three methods cited in this paper.