基于深度学习的超声导波识别轨头疲劳裂纹深度的研究  

Research on Ultrasonic Guided Wave Identification of Rail Head Fatigue Crack Depth Based on Deep Learning

作　　者：曾伟 孙杰辉 邓长寿 罗东云[1] 蔡苗苗[1] 姚飞艳 ZENG Wei;SUN Jiehui;DENG Changshou;LUO Dongyun;CAI Miaomiao;YAO Feiyan(School of Electronic and Information Engineering,Jiu Jiang University,Jiu Jiang,Jiangxi 332005,China)

机构地区：[1]九江学院电子信息工程学院,九江332005

出　　处：《自动化与仪器仪表》2024年第6期246-251,共6页Automation & Instrumentation

基　　金：国家自然科学基金资助项目(62061021,62161014,62341308);江西省自然科学基金联合资助项目(20202BABL201024);江西省教育厅科技项目(GJJ2201906)。

摘　　要：为了保障高速列车的安全运行,迫切需要一种能够自动化和智能化检测列车钢轨损伤的无损检测技术。因此,利用深度学习方法和超声导波技术相结合的方法对钢轨轨头疲劳损伤深度进行定量评估。首先搭建基于超声导波的钢轨轨头疲劳损伤检测系统,得到不同深度的钢轨轨头疲劳损伤与超声导波相互作用的透射波信号,然后通过小波变换获得小波时频图,同时增加三种高斯白噪声(高斯白噪声为0.1、0.3和0.5),将小波时频图分成8∶2的训练集和测试集,最后采用GoogLeNet、Mobilenetv1、Mobilenetv2、Mobilenetv3四种深度学习模型对钢轨疲劳损伤深度进行分类,分别比较四种模型的准确率、召回率、精确率及F1得分等模型性能。实验结果表明,Mobilenetv3深度模型与超声导波技术结合用于钢轨轨头疲劳损伤深度定量检测具有98%的识别准确率。本研究工作将深度学习模型与超声导波技术相结合用于识别钢轨疲劳损伤深度检测的可行性和可靠性提供基础。In order to ensure the safe operation of high-speed trains,there is an urgent need for a non-destructive testing technology that can automatically and intelligently detect rail damage in trains.Therefore,this article uses a combination of deep learning methods and ultrasonic guided wave technology to quantitatively evaluate the depth of fatigue damage to rail heads.Firstly,a rail head fatigue damage detection system based on ultrasonic guided waves is constructed to obtain transmission wave signals of rail head fatigue damage at different depths and the interaction between ultrasonic guided waves.Then,wavelet time-frequency maps are obtained through wavelet transform,and three types of Gaussian white noise are added(Gaussian white noise is 0.1,0.3,and 0.5).The wavelet time-frequency maps are divided into 8:2 training and testing sets.Finally,four deep learning models,GoogLeNet,Mobilenetv1,Mobilenetv2,and Mobilenetv3,are used to classify the depth of rail fatigue damage.The accuracy,recall,accuracy,and F1 score of the four models are compared.The experimental results show that the combination of Mobilenetv3 depth model and ultrasonic guided wave technology for quantitative detection of fatigue damage depth in rail heads has a recognition accuracy of 98%.This research provides a basis for the feasibility and reliability of combining deep learning models with ultrasonic guided wave technology for identifying deep detection of fatigue damage in steel rails.

关 键 词：钢轨轨头疲劳损伤 超声导波 小波变换 深度学习模型 定量检测 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置]