改进的RBF网络训练方法在故障诊断中的应用

目前已有的几种RBF网络训练方法对于含有随机噪声的复杂样本训练速度过慢且分类性能不稳定,依据相对熵最小原理,提出了一种改进的RBF网络训练方法--输出-输入聚类法.利用此方法对旋转机械故障样本进行训练,并与其它方法进行了比较,结果表明,此训练方法用时短,网络结构简单,受噪声影响小.将所创建网络应用于故障诊断,实例表明,此方法训练的网络诊断结果准确,在故障诊断中具有良好的应用前景.

An Improved Training Method of RBF Network and its Application to Fault Diagnosis

The key to the training of a radial basis function(RBF) network will determine the parameters of hidden layers of the network.There are a number of training methods of RBF networks,but they have the shortcomings in that the training speeds are too slow and the ability to classify is unstable,particularly for such complicated sampling data as with random noise.To overcome these shortcomings,an improved training method of RBF networks,the method of output-input cluster based on the minimum entropy theory is presented in the paper.The sample data of a rotary machine indicates that the training time by using the method is shorter;the network structure is simpler and the influence of random noise is less than that by using other methods.A fault diagnosis example illustrates the excellent performance of the algorithms.