基于深度学习方法的钢轨多裂缝导波检测  

Detection of Multiple Rail Cracks Using Guided Waves Based on Deep Learning Methods

作　　者：刘平心 孛爱 陈家安 赵春宇[1] 黄震宇[1] LIU Pingxin;BO Ai;CHEN Jiaan;ZHAO Chunyu;HUANG Zhenyu(School of Electronic Information and Electrical Engineering,Shanghai Jiao Tong University,Shanghai 200241,China)

机构地区：[1]上海交通大学电子信息与电气工程学院,上海200241

出　　处：《铁道学报》2025年第2期145-151,共7页Journal of the China Railway Society

基　　金：国家自然科学基金(50905107)。

摘　　要：沿钢轨长距离传播的导波遇到钢轨裂缝和内部缺陷可产生反射波,适用于在线检测钢轨裂缝。钢轨可能存在多个裂缝,且端面也会产生反射,从而造成导波的多重反射现象,严重影响导波检测钢轨裂缝的准确性。提出一种能够精确检测钢轨多个裂缝的位置和损伤程度的深度学习算法,其特点是引入注意力机制,增强反射波内的多重反射时序特征,加快深度学习模型的收敛速度。基于实验验证的钢轨有限元模型,在钢轨不同位置生成不同损伤程度的裂缝,模拟导波检测数据用于训练深度学习模型。在真实钢轨上制造不同深度的人工裂缝,测试导波数据用于验证深度学习模型有效性。验证结果表明,在轨头、轨腰和轨底的多裂缝条件下,增加注意力机制可提高裂缝位置和损伤程度预测精度至99%。Guided waves propagating along the rails can generate reflected waves when encountering rail cracks and internal defects,making them suitable for online detection of rail cracks.However,the presence of multiple cracks in the rails and the occurrence of reflections from end faces result in the phenomenon of multiple wave reflections,severely affecting the accuracy of crack detection using guided waves.In this paper,a deep learning algorithm was proposed to detect the positions of multiple rail cracks and identify the extent of damage.The algorithm introduced an attention mechanism to enhance the temporal features of multiple wave reflections within the packet of reflected waves,thereby accelerating the convergence of the deep learning model.A finite element model of rail segments,verified through experimental validation,was utilized to generate guided wave detection data with different crack positions and sizes for training the deep learning model.Finally,artificial cracks of varying depths were created on actual rails to test the validity of the deep learning model through the analysis of guided wave data.The results indicate that,under the presence of multiple cracks in the rail head,rail web,and rail foot,incorporating an attention mechanism can improve the prediction accuracy of crack location and damage severity up to 99%.

关 键 词：钢轨裂缝 导波检测 深度学习 注意力机制 

分 类 号：U213.43[交通运输工程—道路与铁道工程]