Abstract: Due to defects of time-of-arrival localization that influenced by many factors, a neural network technique was used to predict localizations of the acoustic emission sources. In order to reduce numbers of input samples, the most important characteristic parameters of acoustic emission sources were put up and adopted techniques of principle component analysis (PCA), and the number of hidden units was determined by training the neural network using different numbers of hidden units. A BP network was designed by use of the additional momentum method and chosen initial threshold optimized. The network was used in an illustration, by comparing with results of actual damage localization, the results showed that a reasonable network structure and input parameters could determine accurately position of structural damage. In addition, the precision of localization was improved , computation became more efficiency and simpler.