Abstract: In order to improve the generalization of a neural network model, input frequency is divided into sub-bands. These sub-bands enhance information compactness, and those which cover the full frequency ensure ergodicity. Higher compactness and ergodicity help to improve the generalization of the neural network. Frequency decomposition is performed by particle swarm optimization with a blindfold feature. Because particle swarm optimization and the usual neural network algorithm require an iterative calculation, they take a long time for execution. An extreme learning machine neural network with a one-time iteration is time efficient. Simulation results show that the generalization and accuracy of the neural network model are higher and can satisfy the demands of general engineering applications.