This paper presents the application of the wavelet finite element methods (WFEM) and neural network to crack parameters estimation and discusses the accuracy and efficiency of this method. The crack is presented by a rotational massless spring, and the natural frequencies for various crack parameters (location and depth) are obtained through WFEM. The neural network is then applied to establish the mapping relationship between the natural frequencies and the crack parameters, which uses feed-forward multiplayer neural networks trained by back-propagation, error-driven supervised training. With this trained neural network, the crack location and depth is estimated through using the measured natural frequencies as the input. The results of a cantilever beam experiment indicate that the estimating error of crack location is less than 3%, and the error of crack depth is less than 2%.