• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2021, Vol. 57 ›› Issue (12): 136-144.doi: 10.3901/JME.2021.12.136

• 仪器科学与仪器 • 上一篇    下一篇

扫码分享

基于神经网络和支持向量机的导波弯管腐蚀损伤程度辨识研究

周澄1, 邓菲1, 刘尧2, 刘秀成2, 陈洪磊2, 刘增华2   

  1. 1. 上海应用技术大学电气与电子工程学院 上海 200235;
    2. 北京工业大学机械工程与应用电子技术学院 北京 100124
  • 收稿日期:2020-08-29 修回日期:2020-12-06 出版日期:2021-08-31 发布日期:2021-08-31
  • 通讯作者: 邓菲(通信作者),女,1977年出生,博士,副教授。主要研究方向为现代检测技术与方法、信号处理技术等。E-mail:2606897447@qq.com
  • 作者简介:周澄,女,1994年出生。主要研究方向为轨道交通安全检测。E-mail:522769104@qq.com
  • 基金资助:
    国家自然科学基金(11202137)、上海应用技术大学协同创新基金专项(XTCX2018-11)和上海市联盟计划—难题招标专项(2019025)资助项目

Identification of Corrosion Damage Degree of Guided Wave Bend Pipe Based on Neural Network and Support Vector Machine

ZHOU Cheng1, DENG Fei1, LIU Yao2, LIU Xiucheng2, CHEN Honglei2, LIU Zenghua2   

  1. 1. College of Electrical and Electronic Engineering, Shanghai Institute of Technology, Shanghai 200235;
    2. College of Mechanical Engineering and Applied Electronics Technology, Beijing University of Technology, Beijing 100124
  • Received:2020-08-29 Revised:2020-12-06 Online:2021-08-31 Published:2021-08-31

摘要: 利用导波的远场检测优势和机器学习模型,开展管道弯曲处腐蚀损伤程度智能辨识方法研究。在普通碳质钢管弯头处加工不同程度的腐蚀缺陷,按腐蚀程度分为10个等级。采用自激自收和一激一收两种信号激励接收方式,在管道中激励具有非频散的T (0,1)型超声导波,采集得到不同腐蚀程度缺陷对应的导波检测信号。在时域和频域对检测信号进行分析,提取多个导波信号特征值用于表征损伤程度,并通过BP神经网络和支持向量机两种分类模型对数据进行训练分析得到缺陷损伤辨识模型,实现弯曲处腐蚀程度的准确辨识。研究中分析两种模型的超参数对缺陷辨识模型精度的影响,对比研究两种模型对弯管腐蚀损伤辨识的性能。试验结果表明,两种损伤辨识模型对不同激励接收模式下的导波检测信号均能取得较优的分类效果。相较于BP神经网络,支持向量机在小样本条件下对弯管腐蚀损伤具有更好的辨识效果。

关键词: 超声导波, 弯管腐蚀, 损伤特征值, BP神经网络, 支持向量机

Abstract: The far-field detection advantage of guided wave and machine learning models are used to carry out the research on the intelligent identification methods for corrosion damage in pipeline bends. Different levels of corrosion defects are processed at the bends of ordinary carbon steel pipes elbow, which are divided into 10 grades according to the degree of corrosion. Two kinds of signal excitation receiving methods, self excitation receiving and one excitation receiving, are used to excite T(0,1) ultrasonic guided wave in the pipeline, and the guided wave detection signals corresponding to defects with different corrosion degrees are acquired. Characteristic values of multiple guided wave signals are extracted to represent the damage degree by analyzing the features of detection signal in time domain and frequency domain. Two classification models, BP neural network and support vector machine, are used to train and analyze the data to obtain the defect damage identification model, which can accurately identify the corrosion degree of the bending part. In the research, the influence of the super parameters of the two models on the accuracy of the defect identification model is analyzed, and the performance of the two models on the identification of the corrosion damage of the bend is compared. The experimental results show that the two damage identification models can achieve better classification effect for guided wave detection signals under different excitation and reception modes. Compared with BP neural network, support vector machine has better identification effect for rail crack damage under the condition of small sample.

Key words: ultrasonic guided waves, elbow corrosion, damage eigenvalue, bp neural network, support vector machine

中图分类号: