Articles

Automated Track Change Detection Technology for Enhanced Railroad Safety Assessment

Authors: Federal Railroad Administration

Abstract: This report documents the successful use of 3D laser scanning, Deep Convolutional Neural Networks (DCNNs), and change detection technology to reliably detect and classify a wide variety of track components and conditions that influence the safety of train operations, and to report changes in these features over time with high precision. This technology advances the state-of-the-art in automated track inspection, going beyond the simple pass/fail assessments typical of current inspection approaches. During the test program, conducted between April 2019 and October 2020, it detected a wide range of both small and large changes related to elastic fasteners, spikes, joint bar gaps, joint bar bolting, crosstie skew, ballast level, and ballast fouling.

Link to article HERE

Deep Learning for Railroad Inspection – Phase 2

Authors: Richard Fox Ivey, Mario Talbot, John Laurent (Pavemetrics)

Abstract: This paper builds on prior work (Deep Learning for Railroad Inspection – Phase 1) to develop a Deep Neural Network that can automatically identify key railway components as a step in the process of automating rail inspection in an effort to overcome the limitations of traditional methods. This new study adds the identification of new railway components (Tie Plates) as well as the automated assessment of their condition.

Link to article HERE

Deep Learning for Railroad Inspection – Phase 1

Authors: Richard Fox Ivey, Mario Talbot, John Laurent (Pavemetrics)

Abstract: Railway networks around the world are an important part of the transportation network and represent billions of dollars of investment. Poorly maintained networks negatively impact asset longevity, schedule performance and pose a serious threat to safety. In order to safeguard against these risks, Railroads typically inspect 100% of their mainline network at least annually and key locations even more frequently. Railroad inspection has traditionally been a manual process with inspectors walking the track or driving slowly in a high-rail vehicle to visually spot problems. This practice is very costly, time consuming, impacts schedule performance (due to the need for track possession), and puts staff at risk. While there have been some recent attempts to modernize the inspection process through the adoption of machine-vision technologies, these technologies are often still reliant on human inspectors manually reviewing images in order to spot defects. Manual review of images suffers from many of the same problems as manual inspections do: it is time consuming, subjective as opposed to being objective, and requires significant amounts of labor. This paper will explore a new approach which makes use of Deep Learning algorithms, specifically a Deep Neural Network, to automatically inspect images and has the potential to overcome these limitations.

Link to article HERE

Laser Triangulation for Track Change and Defect Detection

Authors: Federal Railroad Administration

Abstract: This report documents the successful demonstration of automated change detection on railroad track. Pavemetrics Systems Inc. performed this research under contract with the Federal Railroad Administration between March and December 2017. The project successfully demonstrated the ability of its Laser Rail Inspection System (LRAIL) to detect changes in fasteners, anchors, spikes, ties, joints, and ballast—as well as record rail stamping information on Amtrak’s Harrisburg line.

Link to article HERE

Extended Field Trials of LRAIL for Automated Track Change Detection

Authors: Federal Railroad Administration

Abstract: This report details the deployment of Pavemetrics’ Laser Rail Inspection System, “LRAIL,” for the purposes of automated change detection. The project was conducted between September 2018 and December 2019 at filed locations on Amtrak property and at Pavemetrics’ offices in Quebec, Canada.The project involved a combination of field sensor data acquisition, deliberate manual changes in the field, office algorithm development, algorithm testing and validation, and system performance reporting. The extended field trial proved successful. Repeatability, mean, and standard deviation of change measurements were determined and noise floors for each measured parameter were established.

Link to article HERE