【作者】 孙涛；

【导师】 赵立军；

【作者基本信息】 哈尔滨工业大学 ， 机械工程（专业学位）， 2025， 硕士

【摘要】 随着高速铁路运行时间的增加,其承载主体桥梁表面极易出现各种表观损伤,若不及时发现将严重削弱桥梁设计性能,因此开展桥梁表观损伤检测具有重大意义。本研究针对传统人工在夜间巡检作业危险且检测技术对环境因素敏感等问题,开展基于机械臂搭载检测设备对相应桥体结构巡检的研究工作,旨在设计一套铁路桥梁巡检系统,通过合理规划机械臂展开过程并结合基于深度学习的检测技术,实现桥梁待检测部位的表观损伤检测。本文首先根据总体需求设计指标完成了桥梁检测机械臂的整体机械结构设计和视觉检测设备设计。针对桥梁底部检测这一工况,结合实际工程情况,提出采用人工标准化示教对桥梁检测机械臂进行展开,针对常规方法无法有效求解冗余机械臂逆运动学解问题,使用了一种改进的灰狼算法进行求解,实现了高精度以及快速求解目标,同时,采用随机点进行了仿真和实验,验证了算法的性能,使用蒙特卡洛方法对机械臂的工作空间进行了分析,验证了满足待检测部位的覆盖检测。针对基于深度学习的桥梁表观损伤检测算法在实际应用中受到的桥梁表观损伤数据集参差不齐问题,使用实际采集的桥梁表面损伤图像和数据扩充手段DCGAN以及Copy-Pause技术作为数据来源,并考虑采集表面的多样性、表观损伤类别的多样性、表观损伤尺寸的多样性、采集视角的多样性和采集条件的多样性,构建了大规模桥梁表观损伤数据集BD（Bridge Damage）,共包括24380张桥梁表观损伤图片,涵盖7类常见的桥梁表观损伤类别。为了缓解正常表观类图像过多导致直接用于表观损伤检测造成的误检问题和无效耗时问题,本文设计了一个分级检测的解决方案,首先设计了一种表观损伤多分类算法,在这一步将输入的图像划分为正常类和损伤类,预测出的正常类图像直接剔除,保留损伤类图像,在自建数据集和公开数据集上验证了算法在保证实时性的同时取得了较好的预测效果。对于预测出的损伤类图像,进一步提出了一种基于Transformer桥梁表观损伤检测网络,将图像中的损伤位置进行检测识别,结合自建数据集和公开数据集验证了算法在提高推理速度的基础上大大提高了检测精度。为确保机械臂可以顺利到达桥梁待检测部位,进行了避障规划的研究。对于此大型机械臂,使用ROS工具建立其在虚拟环境下的模型,采用Move It构建障碍物识别和规划策略,通过仿真验证了可行性。在路径规划实验中,首先搭建了桥梁检测机械臂的硬件系统和软件系统,之后根据机械臂到达待检测部位的姿态使用Move It生成路径点并结合PVT插补算法作为机械臂的轨迹规划算法,实现了末端检测设备快速到达检测部位,总体规划时间为920s。在桥梁表观损伤多分类和损伤检测实验中,通过对梁拱立面、桥梁支座和桥梁底部的分区域采集和检测验证了算法的有效性,损伤二分类准确率达100%,多分类结果达到94.2%,单张图片推理时间为0.0075s,准确率为95.4%,召回率为98.8%,m AP₅₀达到92.4%,单张图片的检测时间为0.009s。更多还原

【Abstract】 With the increase in the operation time of high-speed railways,various apparent damages are very likely to occur on the surface of the main bridges.If not discovered in time,the design performance of the bridges will be seriously weakened.Therefore,it is of great significance to carry out bridge apparent damage detection.In view of the problems that traditional manual inspections at night are dangerous and the detection technology is sensitive to environmental factors,this study conducts research on the inspection of the corresponding bridge structure based on the detection equipment carried by the robotic arm.The purpose is to design a railway bridge inspection system.By rationally planning the deployment process of the robotic arm and combining the detection technology based on deep learning,the apparent damage detection of the bridge to be inspected can be realized.This paper first completed the overall mechanical structure design and visual inspection equipment design of the bridge inspection robot arm according to the overall demand design indicators.In view of the working condition of bridge bottom inspection,combined with the actual engineering situation,it is proposed to use manual standardized teaching to expand the bridge inspection robot arm.As the conventional method cannot effectively solve the inverse kinematics solution of the redundant robot arm,an improved gray wolf algorithm is used to solve it,achieving high precision and fast solution goals.At the same time,random points are used for simulation and experiments to verify the performance of the algorithm.The Monte Carlo method is used to analyze the workspace of the robot arm to verify that it meets the coverage detection of the parts to be inspected.In order to solve the problem of uneven bridge surface damage datasets in practical applications of deep learning-based bridge surface damage detection algorithms,a large-scale bridge surface damage dataset BD（Bridge Damage）was constructed using actual collected bridge surface damage images and data expansion methods DCGAN and Copy-Pause technology as data sources.Considering the diversity of collected surfaces,the diversity of surface damage categories,the diversity of surface damage sizes,the diversity of collection perspectives and the diversity of collection conditions,the dataset includes 24,380 bridge surface damage images,covering 7 common bridge surface damage categories.In order to alleviate the problem of false detection and ineffective time-consuming caused by too many normal appearance images directly used for appearance damage detection,this paper designs a hierarchical detection solution.First,a multi-classification algorithm for appearance damage is designed.In this step,the input image is divided into normal and damaged classes.The predicted normal class images are directly eliminated,and the damaged class images are retained.The algorithm is verified on self-built data sets and public data sets to achieve good prediction results while ensuring real-time performance.For the predicted damage class images,a Transformer-based bridge appearance damage detection network is further proposed to detect and identify the damage location in the image.The combination of self-built data sets and public data sets verifies that the algorithm greatly improves the detection accuracy while improving the reasoning speed.In order to ensure that the robot arm can smoothly reach the part of the bridge to be inspected,obstacle avoidance planning was studied.For this large robot arm,the ROS tool was used to build its model in a virtual environment,and Move It was used to build obstacle recognition and planning strategies,and the feasibility was verified through simulation.In the path planning experiment,the hardware system and software system of the bridge inspection robot arm were first built.Then,according to the posture of the robot arm reaching the part to be inspected,Move It was used to generate path points and combined with the PVT interpolation algorithm as the trajectory planning algorithm of the robot arm,so that the end detection equipment can quickly reach the inspection part,and the overall planning time is 920s.In the bridge apparent damage multi-classification and damage detection experiment,the effectiveness of the algorithm was verified by collecting and detecting the beam arch facade,bridge support and bridge bottom in different regions.The accuracy of the damage binary classification reached 100%,and the multi-classification result reached 94.2%.The reasoning time for a single image was 0.0075s,the accuracy was 95.4%,the recall was 98.8%,the m AP₅₀ reached92.4%,and the detection time for a single image was 0.009s.更多还原