Abstract: To improve the recognition rate of sucker rod's defects, the composite characteristics including both the characteristics of wavelet packet energy and peak-to-peak values in time domain were applied to the recognition in combination with the SVM based on small samples. The separability of the composite characteristics was proved better than that of the characteristics of single wavelet packet energy and the former can enhance the effectiveness of recognition to a certain extent by introducing the separability criterion based on the distance between classes. On the other hand, the pattern recognition of the sucker rod's defects was carried out with the one-to-one data classified SVM using lots of data, and the results revealed that the separability of composite characteristics is better than that of single wavelet packet energy, with smaller errors due to the generalization of defect recognition.