Aiming at the shortcoming of determining the of hidden nodes of LSRBF neural networks based on the imputting information of training data, in paper, the modified monotone index method based on training data with the whole inputting and outputting information is presented to determine the number of hidden nodes of LSRBF neural networks.

In this paper,the new method is presented to determine the number of hidden nodes of RBF neural networks,it makes use of the whole inputting and outputting information of training samples to establish similar matrix of samples,then to determine the number of hidden nodes of RBF neural networks by maximal matrix element method.

This paper taking one-hidden layer combined BP network model used in the fault diagnosis of power transformer as a exapmle and on the based of training flow process chart, points out that the number of hidden nodes, initial weight values, training error, the max training times and the order of training samples have different influence on the BP network generalization.