【作者】 叶志坚；

【导师】 陈建政；

【作者基本信息】 西南交通大学 ， 交通运输（专业学位）， 2024， 硕士

【摘要】 随着我国城市轨道交通规模不断发展,铁路运营的速度和里程也在的不断增加,列车运行时的线路维保也愈发重要。随着铁路运营里程的不断增加,传统轨道维保手段已逐渐不满足高效率且高精度的检测需求。传统的轨道检测无法对钢轨三维几何状态和表面损伤情况进行可视化识别。因此,钢轨三维廓形重建及表面损伤识别对于轨道长期运营的维护保养及监测具有重要意义。本文应用双线结构光和惯性测量单元,针对钢轨廓形进行三维可视化重建,并采用点云神经网络算法对钢轨表面损伤进行识别。主要研究工作如下:(1)基于霍夫变换的钢轨断面全轮廓融合及畸变矫正研究。轨道检测中,受到结构光传感器安装角度影响,采集轮廓均为钢轨单侧轮廓。为了实现钢轨全轮廓高精度合成,首先采用多结构光联合标定的方法对传感器进行标定,同时搭建静态标定试验台,分析了结构光传感器标定精度。针对检测过程中的振动影响,采用Hough检测进行动态全轮廓融合。对于检测梁摇头及点头运动所导致的钢轨廓形畸变问题,基于霍夫圆检测和双线结构光构造的钢轨虚拟三维体,对钢轨廓形进行三维空间矫正。(2)基于多传感器融合的钢轨三维廓形重建方法研究。利用多体动力学仿真软件建立检测车辆模型,仿真求得检测梁各惯性传感器的输出。应用钢轨二维模型及轨道随机激励模型建立虚拟钢轨三维模型,并构建检测梁刚体运动方程,通过仿真获取各结构光传感器的点云输出。基于卡尔曼滤波融合惯性测量单元和结构光传感器获取的运动参数,降低检测车辆振动对结构光传感器的影响,对钢轨三维廓形进行重建。同时,对结构光传感器和惯性测量单元分别添加不同程度的噪声,分析卡尔曼滤波中不同传感器噪声对钢轨三维重建精度的影响。结果表明,利用多传感器融合算法可有效地降低检测振动对廓形重建的影响,不同传感器噪声干扰下钢轨横向和垂向最大重建误差分别为0.2488 mm和0.2789 mm,满足工程检测的需求。搭建了钢轨三维重建实验平台,验证了钢轨三维重建算法的可行性和可靠性。(3)基于点云语义分割神经网络的钢轨表面损伤识别。对重建后的钢轨三维点云进行语义分割识别,针对不同类型的钢轨表面损伤进行了仿真建模和神经网络识别验证。同时,针对传统点云语义分割神经网络Point Net++在局部细节特征提取表现不足的问题,针对性地提出了适用于钢轨表面细微损伤识别的残差Point Net++语义分割网络,并在仿真和实测中进行了验证。结果表明,改进后的钢轨损伤识别网络的总体精度和平均交并比分别为98.91%和96.28%,并能有效识别不同损伤区域和深度,满足工程实际需求。更多还原

【Abstract】 With the continuous development of urban rail transport scale in China,the operating speed and mileage of railway vehicles are also increasing,and line maintenance has become more and more important.With the increasing operation mileage of the railway,the traditional rail maintenance methods have failed to meet the high-efficiency and high-precision inspection demand.The traditional rail inspection technology cannot visually identify the three-dimensional geometric state and surface damage of rails.Therefore,the 3D profile reconstruction and surface damage identification of rails are of great significance for the maintenance and monitoring of long-term operation.In this thesis,two-line structured light and an inertial measurement unit are employed to reconstruct the rail profile for 3D visualization,and a point-cloud neural network algorithm is utilized to identify the rail surface damage.The main research work is as follows:(1)Research on full profile fusion and distortion correction of rail sections based on the Hough transform.During rail inspection,influenced by the installation angle of structured light sensors,all the collected profiles are single side of the rail.To achieve high-precision synthesis of the full profile of the rail,the sensor is calibrated using the joint calibration method of multiple structured lights.Simultaneously,a static calibration test bench is built to analyse the calibration accuracy of the structured light sensor.For the influence of vibration in the detection process,Hough detection is used for dynamic full profile fusion.For the rail profile distortion caused by the transverse swing and nodding motion of the detection beam,the rail profile is corrected in three-dimensional space based on the Hough circle detection and the virtual three-dimensional body of the rail constructed by two-line structured light.(2)Research on rail three-dimensional profile reconstruction method based on multi-sensor fusion.Multi-body dynamics simulation software is used to establish the detection vehicle model,and the output of each inertial sensor of the detection beam is obtained by simulation.The two-dimensional model of the rail and the random excitation model of the track are applied to establish the virtual rail three-dimensional model.The rigid body motion equation of the detection beam is also constructed,and the point cloud output of each structural light sensor is obtained through simulation.Based on the Kalman filter fusion of the motion parameters obtained by the inertial measurement unit and the structured light sensor,reducing the impact of vehicle vibration on the structured light sensor and reconstructing the three-dimensional profile of the rail.Simultaneously,different degrees of noise are added to the structured light sensor and the inertial measurement unit to analyze the effect of different sensor noise in Kalman filtering on the 3D reconstruction accuracy of the rail.The results demonstrate that the multi-sensor fusion algorithm can effectively reduce the influence of detection vibration on profile reconstruction,and the maximum reconstruction errors of rail transverse and vertical under the noise interference of different sensors are 0.2488 mm and0.2789 mm respectively,meeting the needs of engineering inspection.An experimental platform for three-dimensional reconstruction of rails is constructed to verify the feasibility and reliability of the three-dimensional reconstruction algorithm of rails.(3)Rail surface damage recognition based on point cloud semantic segmentation neural network.The reconstructed three-dimensional point cloud of the rail is semantically segmented and identified,and simulation modeling and neural network identification verification are conducted for different types of rail surface damage.Additionally,to address the problem of insufficient performance of the traditional point cloud semantic segmentation neural network Point Net++in local detail feature extraction,the residual Point Net++semantic segmentation network for the identification of rail surface subtle damage is proposed and verified in simulation and real measurement.The results show that the overall accuracy and average intersection and merger ratio of the improved rail damage recognition network are98.91%and 96.28%,respectively.It can effectively identify different damage areas and depths,meeting the practical needs of engineering.更多还原