Abstract: Ultrasonic TOFD method takes advantages in testing thick weld, but the qualitative recognition of testing results now still relies on inspectors, therefore the testing results estimation is affected by personnel experience. By contrast, automatic recognition technology can reduce human factors impact on the testing results analysis. However, in order to implement the automatic recognition, feature extraction is the key point. In this paper, by controlling the welding specification, the specimens with porosity, slag, cracks, incomplete fusion and lack of penetration defect were prepared. Then two-dimensional entropy features in frequency domain of TOFD signal, which were spectral entropy and center frequency of spectral, for above mentioned kinds of defects were analyzed. The results show that the two-dimensional entropy in frequency domain can separate the five types of defect signal. It could provide effective characteristics for automatic identification of defect.