中文摘要：

根据非稳态超声检测信号的特点，将小波包变换用于缺陷信号的降噪及特征提取问题的研究，并利用类别可分性判据和RBF神经网络分别对特征值提取结果进行评价。引入了平均阈值的概念，在此基础上研究了小波包降噪效果。提出了以选取小波包分解频带的能量作为缺陷信号特征值的方法。实际焊接缺陷的实验结果表明，小波包降噪效果明显；在特征数据得以压缩的同时，分类的可分性较高。

英文摘要：

It contributes to the application of wavelet packet transform in denoising and feature extraction of nonstationary ultrasonic flaw signals, and the sort separability criterion and RBF neural network are respectively used for evaluating the validity of feature classification. Mean threshold is introduced on which wavelet packet denoising is studied. The energy of the frequency domain selected based on wavelet packet decomposition is taken as the feature information. The experimental results over welding flaw signals demonstrate the effectiveness of the proposed schemes.