Abstract: A new fault diagnosis algorithm can be used to resolve the problem of fault diagnosis with prior knowledge more effectively. Taking the prior sample point as the center, using inner product to judge sample data similarity, it carries on the cluster analysis. It makes a super-plane intersect a sphere in the characteristics of the space, obtains a spherical covering area, thus transforms the neural network training question as the set of points cover question. Based on constructive neural networks, the algorithm’s characteristic is that the sample data of fault can be handled directly. Because the cover center is determined, it constructs out the least element hidden layer network structure. This new algorithm can reduce the long training time and learning complexity of traditional neural networks. Computer simulation results confirm the effectiveness of the algorithm.

Key words: constructive neural networks, fault diagnosis, covering, algorithm

摘要： 提出一种新的故障诊断方法，以便更加有效地解决具有先验知识的故障分类问题。以先验样本点为中心，利用内积判断样本数据的相似度，从而进行聚类分析，在特征空间里作超平面与球面相交，得到一个球面覆盖领域，从而将神经网络训练问题转化为点集的覆盖问题。该算法以构造型神经网络为基础，其特点是直接对故障样本数据进行处理，由于覆盖中心确定，该算法构造出的是隐层元最少的网络结构，有效地克服了传统神经网络训练时间长、学习复杂的问题。计算机仿真实验结果证实了该算法的有效性。

关键词: 构造型神经网络, 故障诊断, 覆盖, 算法