• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2024, Vol. 60 ›› Issue (8): 1-10.doi: 10.3901/JME.2024.08.001

• 仪器科学与技术 • 上一篇    下一篇

扫码分享

Cr/Zr涂层环向压缩表面裂纹识别研究

郑杰1, 权伟1, 刘洋1, 卢学民1, 刘晓红2, 岳雅楠2, 周腾2, 蔡振兵2   

  1. 1. 西南交通大学电气工程学院 成都 611756;
    2. 西南交通大学机械工程学院 成都 610031
  • 收稿日期:2023-04-11 修回日期:2023-10-25 出版日期:2024-04-20 发布日期:2024-06-17
  • 作者简介:郑杰,男,1996年出生。主要研究方向为模式识别、机器学习、人工智能。E-mail:867682404@qq.com;权伟(通信作者),男,1982年出生,博士(后),副教授,博士研究生导师。主要研究方向为模式识别与人工智能、智能信息处理。E-mail:wquan@swjtu.edu.cn
  • 基金资助:
    国家自然科学基金(52277127)、四川省科技创新人才(2021JDRC0012)、国家轨道交通电气化与自动化工程技术研究中心开放课题(NEEC-2019-B06)和西南交大学科交叉基础研究重点(2682021ZTPY089)资助项目。

Research on Surface Cracks Identification of Cr/Zr Coating in Circumferential Compression

ZHENG Jie1, QUAN Wei1, LIU Yang1, LU Xuemin1, LIU Xiaohong2, YUE Yanan2, ZHOU Teng2, CAI Zhenbing2   

  1. 1. School of Electrical Engineering, Southwest Jiaotong University, Chengdu 611756;
    2. School of Mechanical Engineering, Southwest Jiaotong University, Chengdu 610031
  • Received:2023-04-11 Revised:2023-10-25 Online:2024-04-20 Published:2024-06-17

摘要: 针对Cr/Zr涂层表面裂纹图像存在的视野暗、噪声高及人工识别效率低等问题,建立Cr涂层表面裂纹自动精确识别与关键参数提取的技术和方法,并开展相关分析。在Cr/Zr涂层包壳管上进行环向压缩试验,采集裂纹图像,对采集到的图像数据设计了一种Cr涂层表面裂纹识别与分析算法—CSCRA算法,该算法基于U2-Net神经网络对裂纹进行语义分割,从而获得Cr涂层表面裂纹的位置信息,然后设计了一种中轴线提取算法并结合连通域分析和距离计算对识别出的裂纹进行统计分析(包括数量、密度、长宽等)。相比于传统的边缘检测算法、阈值分割算法,设计的CSCRA算法能够准确有效识别多种形态的Cr涂层表面裂纹,自动完成识别裂纹的数量、密度、长宽等信息分析,得到了裂纹数量、密度、宽度信息与环向压缩形变量的对应关系。CSCRA算法能够解决传统识别算法连续性差、噪声高的问题。从裂纹数量、密度、宽度与环向压缩试验形变量之间的联系中,发现Cr涂层表面裂纹的数量和宽度随着环向压缩形变量的增大呈现出增大趋势,但两者在不同阶段增加速率不一致。

关键词: Cr/Zr涂层, 环向压缩, 裂纹识别, CSCRA算法, 神经网络

Abstract: Objective to realize the accurate recognition and analysis of Cr coating surface cracks, aiming at the crack image on the surface of Cr/Zr coating has problems such as dark vision, high noise and low manual recognition efficiency. we present Cr coating surface cracks recognition and analysis framework (CSCRA) for crack detecting and crack analyzing in crack images. The crack images are produced by the circumferential compression experiment on zirconium alloy Cr-coated cladding tube. Our approach, based on the U2-Net neural network, obtain each crack location in the Cr coated surface crack images by the semantic segmentation. Then, we design a novel algorithm to analyze the four important indicators, two of them are the amount of cracks and the density of cracks in whole crack images, others are length and width of each crack. The algorithm combines the connected domain analyze and distance calculation with the central axis extraction algorithm. Experimental results on the Cr coating surface crack images show that CSCRA can detect multiform crack images accurately and effectively, meanwhile, the amount of the cracks and the density of cracks in one crack image calculated by our method are almost consistent with the real data. Due to the detailed crack segmentation output, the two indicators, length and width, achieve highly accuracy. Compared to the edge detection method or the threshold segmentation method, CSCRA has much better segmentation accuracy even with high noise and low continuity crack images. While analyzing the relationship between the circumferential compression experiment deformation and the four indicators, we find that both the crack numbers and width show an increasing trend as the amount of hoop compression increases, but are different in increase rate at different stages.

Key words: Cr/Zr coating, circumferential compression, crack identification, CSCRA algorithm, neural network

中图分类号: