基于两层迭代聚类算法的RBFNN及在发电机诊断中的应用

首先分析了传统的基于K均值聚类算法径向基函数神经网络(RBFNN)的缺点，即需要人为确定隐含层神经元数量，并且不同的初始化方法有不同的聚类结果和学习误差。然后提出了一种新的RBFNN算法——两层迭代聚类算法，能根据样本的分布情况自动计算RBFNN隐含层神经元数量、中心向量和宽度。将实测的MJF-30-6型发电机正常运行、转子励磁绕组故障运行和定子绕组故障时定转子径向振动信号作为学习样本，运用文中所述算法与传统的RBFNN算法比较，结果表明，基于两层迭代聚类算法的RBFNN具有较小的学习误差。

RBFNN BASED ON TWO LEVELS ITERATION CLUSTER ALGORITHM AND ITS APPLICATION IN GENERATOR FAULT DIAGNOSIS

Radial basis function neural network (RBFNN) is one of many kinds of artificial neural networks (ANNs) with wide application. A new adaptive RBFNN algorithm namely two levels iteration cluster algorithm is put forward, which can calculate automatically parameters of RBFNN in terms of samples' distribution, and overcomes conventional K-mean value algorithm's shortcomings that the neuron number of hidden layer must be given in advance, and the different initialization method often has different cluster result and diagnosis precision. Using vibration data acquired on the spot of MJF-30-6 type generator in three conditions of normal operation, rotor excitation winding short circuit and stator winding fault as learning samples of RBFNN, the results of verification show the method has less learning error and higher diagnosis precision than the conventional methods.