Research on a novel RBF neural network and its application in fault diagnosis

The limitations of Back Propagation (BP) neural network (NN) are analyzed. The model, transfer function and training process of Radial Basis Function (RBF) NN are introduced in details. The contrasts between BP NN and RBF NN on the mathematical explanation, the function of the neurons in hidden layer and the training speed are made. To avoid the architecture design only with experiences, a RBF NN is proposed by the way that the nodes in hidden layer are automatically chosen. Meanwhile, the average square error is minimal. The training speed of this RBF NN is greatly improved, while the approximation is still accurate. The RBF NN is used for fault diagnosis of refrigeration system in a sailer, which illustrates the effectiveness of the network designed.