基于AlexNet神经网络的手推式双轨探伤仪超声检测系统研究  

Research on Ultrasonic Detection System of Hand-pushed Dual-rail Ultrasonic Flaw Detector Based on Alexnet Neural Network

作　　者：王旭华 郑韵娴 安尚文 黄凤英 吕关仁 WANG Xuhua;ZHENG Yunxian;AN Shangwen;HUANG Fengying;LYU Guanren

机构地区：[1]中国铁路济南局集团有限公司工务部,山东济南250001 [2]中国铁道科学研究院集团有限公司金属及化学研究所,北京100081 [3]北京中铁科新材料技术有限公司,北京100081

出　　处：《铁道技术监督》2022年第4期19-24,共6页Railway Quality Control

基　　金：中国国家铁路集团有限公司科研开发计划重点课题(N2020G017)。

摘　　要：针对传统方法对钢轨各部位探伤需要采用不同的探伤设备,效率低,并且无法综合评价钢轨质量的问题,设计可以兼顾钢轨内部伤损(含垂向裂纹)和钢轨表面伤损检测的手推式双轨探伤仪。介绍手推式双轨探伤仪超声检测系统组成、工作原理、探伤方法和伤损自动识别方法。详细论述伤损自动识别方法的技术架构、基于无监督聚类的B显数据分割方法和AlexNet神经网络架构,并针对标准样轨4类典型钢轨伤损进行验证。验证结果表明,基于改进的无监督聚类与AlexNet神经网络自动识别算法识别准确率可达90%以上,钢轨伤损图像提取准确率达到98.53%。手推式双轨探伤仪超声检测系统可以同时检测2股钢轨,自动判别伤损,有效识别钢轨伤损特征,较好地解决了现场判读工作量大的问题。The traditional method uses different equipment for rail flaw detectors,which is inefficient and unable to evaluate comprehensive rail quality.Therefore,it has appeared that a hand-pushed dual-rail ultrasonic flaw detector that can take into account the flaw detection of rail surface and internal(vertical crack included).This paper intro⁃duces the composition,working principle,flaw detection method and flaw automatic identification method of the ultra⁃sonic detection system of the hand-pushed dual-rail ultrasonic flaw detector.The technical architecture of flaw auto⁃matic identification method,B-display data segmentation method based on unsupervised clustering and the architec⁃ture of AlexNet neural network are discussed in detail,and the four types of typical rail flaw of standard sample track are verified.The verification results show that the recognition accuracy of automatic recognition algorithm based on improved unsupervised clustering and AlexNet neural network can reach more than 90%,and the extraction accuracy of rail flaw image can reach 98.53%.The ultrasonic testing system of hand-pushed double-rail ultrasonic flaw detec⁃tor can detect two rails at the same time,automatically identify the flaw and effectively identify the characteristics of rail flaw,and better solve the problem of heavy workload of on-site interpretation.

关 键 词：钢轨伤损 超声波探伤 伤损自动识别 算法 神经网络 

分 类 号：U213.43[交通运输工程—道路与铁道工程]