According to the state that low reliability of flaws identification at present, this paper put forward a fusion model combining by three BP networks and D-S evidence theory and applied data fusion technology into the field of ultrasonic flaws identification. Aimed at the characteristics of non-stationary ultrasonic echo-signals of flaws, wavelet packet transform, empirical module decomposition and wavelet transform were applied to feature extraction. Experiment shows that the method mentioned above has the power of effective identification, and measurement accuracy could reach 96%.