华南理工大学学报(自然科学版) ›› 2005, Vol. 33 ›› Issue (4): 5-9.

• • 上一篇    下一篇

基于神经网络的超声检测缺陷表征

罗雄彪 陈铁群   

  1. 华南理工大学 机械工程学院,广东 广州 510640
  • 收稿日期:2004-01-09 出版日期:2005-04-25 发布日期:2005-04-25
  • 通信作者: 罗雄彪(1979-)男,硕士生,主要从事超声缺陷检测及信号处理的研究 E-mail:lxbxl_0207@ sohu.com
  • 作者简介:罗雄彪(1979-)男,硕士生,主要从事超声缺陷检测及信号处理的研究
  • 基金资助:

    广东省科技计划资助项目(2004A1 1303001)

Neural Network-Based Characterization of Flaws Tested by Ultrasoni

Luo Xiong-biao  Chen Tie-qun   

  1. College of Mechanical Engineering,South China Univ.of Tech.,Guangzhou 510640,Guangdong,China
  • Received:2004-01-09 Online:2005-04-25 Published:2005-04-25
  • Contact: Luo Xiong—biao(born in 1979),male,graduate student,mainly researches on the inspection of flaws tested by ultrasonic an d signal processing. E-mail:lxbxl_0207@ sohu.com
  • About author:Luo Xiong—biao(born in 1979),male,graduate student,mainly researches on the inspection of flaws tested by ultrasonic an d signal processing.
  • Supported by:

    Suppofled by the Project of Science and Technology of Guangdong Province(2004A1 1 303001)

摘要: 提出了一种基于神经网络的缺陷表征方法.该方法采用Fischer线性判别分析对表征缺陷的时域信号的波形参数进行选择,并将这些参数作为神经网络的输入对智能缺陷表征系统进行训练,用概率神经网络和BP神经网络分别对缺陷的类型和大小进行识别.对135种人造焊接缺陷(裂纹、夹杂和气孔)的试验结果
表明,文中方法对辨识缺陷表征信息和提高缺陷识别率非常有效.

关键词: 超声检测, 缺陷表征, 无损评价, 神经网络

Abstract:

This paper proposes a method for flaw characterization on the basis of neural networks.In this me-thod ,a selection of the shape parameters defining the pulse-echo envelope reflected from a flaw is carried out by Fischer linear discriminant analysis.The selected parameters are then used as the inputs of neural networks to train the propo sed intelligent flaw characterization system. Moreover,probabilistic neural networks and back pmpagation neural networks are respe ctively adopted to determine the sizes and numbers of flaws.Experimental results for 1 35 systematic weld flaws(crack,slag and porosity)indicate that the proposed method is effective in the flaw characterization with great classification rate.

Key words: ultrasonic testing, flaw characterization, nondestructive evaluation, neural network