Toward Automated Field Ballast Condition Evaluation: Algorithm Development Using a Vision Transformer Framework

Abstract

Ballast degradation in the track substructure may cause poor drainage, settlement, and reduced lateral stability that may affect safety, daily operations, and long-term maintenance of a railroad system. Extreme levels of degradation in the ballast may result in service interruptions because of safety concerns. Therefore, field ballast condition evaluation is deemed crucial. Current state-of-the-practice methods for evaluating ballast condition primarily rely on subjective visual inspection, labor-intensive sampling, and laboratory sieve analyses of collected field samples. A network-level condition assessment of ballast and track substructure is commonly performed using ground-penetrating radar. For site-specific and detailed geotechnical analyses, development of a reliable, accurate, and cost-effective technique for ballast condition evaluation is urgently needed. This paper presents an innovative approach to accomplish image-based ballast condition evaluation based on deep learning techniques. A ballast image dataset (library) is established by collecting images from various railroad sites and laboratory setups of ballast piles. A vision transformer-based segmentation framework is implemented and trained with the established dataset and employed to serve as the image segmentation kernel to relate the image-based Percent Degraded Segments (PDS) with the ground-truth Fouling Index (FI). Based on the presented research findings, the proposed approach for field ballast condition evaluation will serve as the core data analysis component of an automated ballast scanning vehicle to conduct field ballast inspection, which is being developed to serve as an efficient and reliable system for the evaluation of ballast condition in ballasted railroad tracks.