Abstract: In order to determine the type, range and size of concrete defects, the paper gave description of wavelet packets analysis of received ultrasonic waves, by finding the character of different frequency parts and reconstructing the decomposing coefficients to get the energy proportion of different frequency parts. Based on this, a characteristic vector was constructed, which was the input character of neural network. After being trained, the vector was used to identify and judge the defect. The results showed that this method was akle to give a highly accurate measurement not only for the range and size of defects, but also for the type of defects.