基于RGB-D信息三维曲波体特征分析的电气化铁路接触网故障智能识别理论

The catenary faults have became the primary trouble in the electric railway.The current catenary detection system is hard to recognize the faults automaticlly by analysing the two two-dimensional gray image date because of overlap problem.It influences the railway safety enormously. This research performs the catenary three-dimensional detection intelligently by using the RGB-D data collected from catenary system. The research utilizes the three-dimensional curvelet transformation to decompose the three-dimensional catenary depth data first. The three-dimensional model can be reconstructed by analyzing the body feature and extracting special characteristic. The model surface can be modified accroding to the RGB information.Finally the catenary components can be distinguished and the faults can be recognized by analysing the curvelet decomposed data. The effectiveness can be proved in the laboratory and real railway tests. This detection theory utilizes the RGB-D data advantages of including the three-dimension coordinate and color information simultaneously.It overcomes the problems of traditional complex algorithm,the global information lose and noise sensitivity. This research is helpful for reinforcing the electric railway safety by achieving the catenary faults recognition intelligently.

接触网故障已成为电气化铁路牵引供电系统中最主要的故障源。现有接触网非接触式故障检测系统主要基于二维灰度数据分析，重叠和遮蔽问题造成检测系统难以进行自动故障识别，严重影响了电气化铁路的安全运营。课题将针对采集的接触网设备RGB-D数据进行接触网三维化故障智能检测研究。利用稀疏分解理论中的三维曲波理论对接触网三维深度数据进行“体”特征分析，从全局语义上对数据进行特征提取和三维模型重构；通过分析RGB信息与三维模型表面相关性对接触网模型进行修正；利用接触网设备的三维曲波系数体特征对设备进行识别和故障诊断；最终通过实验室和线路实验验证算法的有效性。该方法充分利用了RGB-D数据在空间上包含设备三维坐标、在色域上比灰度数据增加更多信息的优势，克服现有构建三维模型算法流程多且复杂、基于表面邻域分析缺乏全局性、易受噪声干扰等问题，可望实现有效的接触网故障智能检测，确保电气化铁路供电系统设备安全可靠运行。

项目针对电气化铁路接触网检测进行研究，利用三维RGB-D设备在移动检测平台上对接触网进行实时检测，最终获得接触网几何参数、故障识别结果。项目主要取得了如下成果：. 1）课题优化设计了检测平台的设备安装，视角控制。. 2）利用条件滤波、背景分割等预处理手段对含噪声点云进行了数据清洗。利用NARF、SAC-IA、ICP、曲面重建等过程实现了接触网点云的三维模型重构。. 3）实现了基于法相一致性分析的三维点云数据故障识别，实现了基于单点激光深度数据测距的改进PSO-GPF法的接触线导高、拉出值检测，实现了基于BIM模型与多点云数据协同的Local-RANSAC的接触网支撑结构、接触线几何参数检测识别。. 4）通过深度学习理论，利用Faster-RCNN、SSD对采集的RGB图像进行定位识别，并对接触网等电位线散股等故障进行了自动识别。.实验表明，利用三维点云对接触网设备进行监控。能够较好的实现接触网几何参数和故障检测结果。三维点云检测解决了平面二维检测两个主要问题。. 1）克服了接触网转换柱位置的多支支撑装置的相互遮挡造成识别困难问题。. 2）二维识别由于边缘检测造成的识别结果误差较大，三维点云由于利用了模型体特征，检测结果受数据污染干扰较小，更容易实现高精度检测。. 项目研究为接触网安全监控提供了有效的新手段。