基于时频判别特征的焊缝超声检测缺陷分类

Weld defect classification in ultrasonic testing basing on timefrequency discriminant features

关键词:  超声检测  判别追踪  时频判别特征  概率神经网络

Key words:  ultrasonic test  discriminant pursuit  time-frequency discriminant feature  probabilistic neural networks

摘要:

根据超声检测信号的瞬变特性,针对焊缝检测的缺陷分类问题,提出用判别追踪算法提取缺陷信号的局部时频判别特征,并结合概率神经网络实现了焊缝超声检测信号的缺陷分类。在提取时频判别特征时,提出考虑新选原子与已选原子的相关性的判别基提取方案,以降低特征之间的冗余,使提取出的特征能更有效地鉴别不同类别的缺陷。用该方法对一电子束焊缝试块中的缺陷进行了分类,结果表明,时频判别特征适合超声信号的缺陷分类,并能有效地抑制晶粒噪声的影响,考虑判别原子间相关性后可获得更高的分类正确率。

Abstract:

According to transient property of ultrasonic signal, the discriminant pursuit method was proposed to extract local time-frequency features of defect signal and the features were fed to a probabilistic neural networks to classify the defects.During extracting features, the correlation between the incoming atom and the atoms selected before was considered to reduce the redundance among the selected atoms so that the extracted features discriminated different class of signals effectively.Finally, the defects of an electronic welded joint were classified by proposed approach, and the experimental results show that time-frequency discriminant features are appropriate for defects classification in ultrasonic testing, and can suppress the effect of grain noise.In addition, the higher accuracy can be reached if considering the correlation of the selected atoms.