Abstract: In order to enhance the extrapolation capacity of neural networks in the case of bigger noise and lesser samples,an array structural model of wavelet neural networks and its learning algorithm are proposed by using the structural risk minimization principle of statistic learning theory.The construction of the array structural model is based on the frequency distribution of wavelet basis units,and the learning strategy is based on the multi-resolution approximation approach and synthetic risk analysis.The method takes full advantage of wavelet neural networks,such as solid theory basis,explicit practical sense,simple algorithm realization and strong adaptability,etc.To demonstrate the effectiveness of the method for nonlinear system identification,some simulation results and application examples are presented,and a new way for the engineering application of the statistic learning theory is also proposed.