Autonomous train wheel damage detection using advanced deep learning

AECOM的业务和全球专家网络的主要部分集中在提供运输和货运铁路系统的解决方案。在铁路运输系统中，一个主要目标是在初始阶段准确地检测火车轮损坏，以防止灾难性故障。车轮缺陷的类型很多，但主要有三种类型：车轮扁疤、粗糙度和不圆度。有一些现有的方法基于振动测量，使用接触传感器来检测这些主要损坏类型中的任何一种。然而，这些传统的方法成本高，只能检测一种类型的损伤，因为它们不能准确地检测和识别所有主要的火车轮损伤类型。此外，很难确认所收集的数据实际上表明了损坏，而不是传感系统故障、噪声信号或这些的组合，并且需要亲自检查传感系统和车轮。这些感觉系统故障在加拿大特别普遍，因为恶劣的环境条件。在该项目中，申请人和AECOM希望开发一种使用计算机视觉和高级人工智能的新的自主火车轮损伤检测方法（即，深度学习），其基于申请人在使用高级深度学习的结构损伤检测中的先前成就。为了检测主要的车轮损坏，将使用相机和更快的区域卷积神经网络（FasterR-CNN）。为了降低监测和计算成本，并简化损伤检测系统的维护，传感器系统（摄像机）将连接到轨道上，而不是每个轨道车，以测量其振动水平，从而检测损伤。该系统将大幅降低监测成本，提高轮损监测系统的可靠性，从而提高现有铁路运输系统的可持续性.这种新的自主系统也将作为工业界和学术界开发更先进的损伤检测系统的模板。

A major portion of AECOM's business and global network of experts is focused on delivering solutions intransit and freight rail systems. In railway transportation systems, one major aim is to accurately detect trainwheel damage in the initial stages in order to prevent catastrophic failures. There are many types of wheeldefects, but there are three different types of major damage that are wheel flat, roughness, andout-of-roundness. There are some existing approaches based on vibration measurements using contact sensorsto detect any one of these major damage types. However, these traditional methods are costly and can detectonly one type of damage, because they are inaccurate to detect and identify all of the major train wheel damagetypes. Moreover, it is difficult to confirm that the collected data actually indicates damage rather than sensorysystem malfunction, noisy signals, or a combination of these, and requires that sensing systems and wheels bechecked in person. These sensory system malfunctions are particularly prevalent in Canada because of harshenvironmental conditions. In this project, the applicant and AECOM want to develop a new autonomous trainwheel damage detection method using computer-vision and advanced artificial intelligence (i.e., deep learning)based on the applicant's previous achievements in structural damage detection using advanced deep learning. Inorder to detect major wheel damage, a camera and faster regional-convolutional neural network (FasterR-CNN) will be used. In order to reduce monitoring and computational cost and easy maintenance of thedamage detection system, the sensor system (camera) will be attached to rails instead of each rail car tomeasure their vibration levels and thereby detect damage. This system will drastically reduce the cost ofmonitoring and improve the reliability of wheel damage monitoring systems to improve the sustainability ofexisting railway transportation systems. This new autonomous system will also serve as a template for industryand academia in the development of more advanced damage detection system.