摘要
长期以来，由于受许多因素的影响，使得水声目标的分类已成为一个十分困难的问题。现在，随着人工神经网络技术的发展，众多的研究人员已经致力于基于人工神经网络的水声目标分类的研究．本文介绍了一种高阶神经网络即Pi-Sigma网络，研究了它的两种学习算法（基于梯度下降法和共轭梯度法的学习算法），并将Pi－Sigma网络用于水声目标辐射噪声的分类。和多层感知器（MLP）网相比，Pi-Sigma网络具有结构简单、收敛速度快及存储量少等优点。Pi－Sigma网络分类器的输入为一个常Q带通滤波器组作特征提取形成的特征向量。对不同类别的实际水声数据的分类结果表明取得了令人满意的分类正确率（达到或超过了95％）。
Abstract: For a long time, the classification of underwater acoustic objects has been a very difficult problem to be solved, because it is affected by many factors. Now, with the development of technique of artificial neural network (ANN), a lot of researchers have devoted to the classification subject using several of kinds of ANNs. In the paper, we first introduce the Pi-Sigma network (PSN), a higher order ANN, research it's two learning algorithms based respectively on the gradient descent method and the conjugate gradient method, and then use it for objects radiated-noise classification. Comparing with multiple layer perceptron (MLP) network, Pi-Sigma network has some advantages such as simple structure, rapid convergence speed and small storage needed and so on. The inputs of the Pi-Sigma net classifier are the feature vectors which are extracted by a constant Q bandpass filter bank. The classification results on the realistic data, which are noises radiated by different class objects, demonstrate the feasibility of the classification system consisted of a constant Q bandpass filter bank as a feature extractor and a Pi-Sigma net as a classifier, achieving a statisfactory classification accuracy: ＞= 95%.
出处
《声学学报》
EI
CSCD
北大核心
1997年第4期345-351,共7页
Acta Acustica
