Development of track component health indices using image-based railway track inspection data

The primary role of the US Department of Transportation (USDOT) Federal Railroad Administration (FRA) is ensuring the safe operation of railway rolling stock and infrastructure by way of regulatory oversight. FRA regulations require US railroads to conduct visual track inspections as often as twice per week depending on a specific track segment’s FRA track class, which also governs maximum train operating speed. Such inspections are often subjective due to the inherent limitations of human visual inspection and cognition. Additionally, human visual inspections require some level of risk given the need for inspectors to be on track while also consuming valuable network capacity. As a result, and the desire to collect objective data to improve both safety and maintenance planning, railroads are pursuing new means and methods to assess track condition and evaluate track component health. This paper presents a numerical method to define track component health using field data collected on the High Tonnage Loop (HTL) at the Transportation Technology Center (TTC) in Pueblo, Colorado, USA. Line scan laser and image data of the track were captured using a 3D Laser Triangulation system and were subsequently processed using Deep Convolutional Neural Networks (DCNNs). The track heath quantification method proposed establishes benchmarks that were developed based on the understanding of railway track mechanics, high axle load (HAL) railroad engineering instructions, and FRA regulations. The novel metrics presented are referred to as Track Component Heath Indices (TCHIs) and are quantitative values that objectively assess track condition and provide a means to monitor condition change with time and tonnage. These data can be used in conjunction with traditional track geometry and other forms of track heath data (e.g. GPR and rail profile) to more holistically assess the condition of the track structure and its components and ultimately predict its future state.