Discriminant pursuit selects time-frequency atoms as discriminant features that differentiate signals of different classes. Nonlinear distance measure was proposed to replace the common Euclid distance measure in Fisher's class separability criterion, which was used as discriminant criterion in discriminant pursuit. The presented method emphasizes minimizing the inner distance of the same class mainly. The coefficient vectors obtained by discriminant pursuit represent the time-frequency discriminant features of each class and were feed to the multiplayer perceptron neural network. The results show that the classification based on the proposed method perform best in identifying ultrasonic testing signal with accuracy rate 96-100% and very low mean squared error