Abstract: The technique has a broad prospect of applications in nondestructive testing and evaluation. However, low recognition rate and bad robustness have been faced with at present in ultrasonic 3D target recognition. Aiming to these problems, this paper proposed an effective method. The object inner artificial discontinuities were as the ultrasonic targets, and a few characteristic parameters were extracted by means of dealing with the pulse ultrasonic echo signals of the ultrasonic targets, such as relative energy, relative amplitude, relative frequency domain bandwidth, relative coefficient of kurtosis, relative coefficient of variation, relative envelope area, relative coefficient of skewness and relative FWHM of frequency spectrum, then, FLDA was utilized to fuse the characteristic parameters and integration features were discovered, finally, BP neural network was used to train and recognize integration features. The recognition rate of three kinds of ultrasonic targets, which were object interior rectangular groove, horizontal hole and flat-bottom hole, achieved 93.3%, 93.3% and 100% respectively. The experimental result shows that the method is quite effective, having inhibitory ability to noise, not sensitive to test condition and robust. It can provide theoretical and technical reference to ultrasonic 3D target recognition.