Vehicle-borne Autonomous Railroad Bridge Impairment Detection Systems

Abstract

Timber railroad bridges have been exposed to increasingly large axle loadings accompanied by a steady increase in the amount of railcar traffic over the past 50 years. In addition to mechanical loading, there exists a number of environmental conditions that can deteriorate the timber in these bridges: e.g. insects and decay organisms. The primary inspection method is conducted on a bridge by bridge basis and involves visually examining individual components of the bridge and assessing the damage.

This research examines an automated impairment detection system positioned on a railcar capable of traversing multiple bridges along a track to aid in determining critical bridges that need to be inspected. The technology and techniques presented in this dissertation are envisioned as a potential enhancement to current visual evaluation methods by providing system-wide trending data for human decision makers.

The objective of the research is to develop technology that will autonomously detect structural impairments in timber railroad bridges using data gathered from rail vehicles that cross the bridges. This was accomplished by recording the behavior of a bridge and the motion of a railcar passing over bridge spans. Artificial neural networks, a type of pattern recognition technology, were used to determine relationships between the bridge and vehicle behaviors. The results of a finite element analysis were utilized to train the neural networks to recognize the patterns associated with the bridge and railcar motions. Five different impairment conditions, or simulated damage scenarios, were developed for the training process. This allowed the networks to recognize the patterns correlating the railcar and bridge data streams. Once the artificial neural networks were successfully trained, new vehicle motions from a field test were presented to the network and the corresponding bridge behavior was predicted. The neural networks were accurate in predicting the maximum chord deflection to within 0.1 inches in 72% tested chords with improved accuracy at faster speeds.