Deep Learning Approach for Automated Railroad Ballast Condition Evaluation

Ballast has a significant impact on track performance, and the evaluation of ballast condition is crucial for safe railroad operations. This paper focuses on a Ballast Scanning Vehicle (BSV) recently developed for automating ballast inspection using a deep learning-based, computer vision approach. Traditional evaluation methods, e.g., visual inspection or ballast sampling followed by sieve analysis, are subjective and labor-intensive. Furthermore, ballast samples collected from a single location/depth may not be representative of accurately revealing variations of degradation and ballast condition along the track. In contrast, the BSV employs three image acquisition devices to continuously capture high-quality scans of ballast cut sections, enabling accurate and in-depth evaluation of continuous sections of the track. The deep-learning framework was trained to process acquired ballast scans, generating image-based metrics including percent degraded segments (PDS), fouling index (FI) estimates, and in-service ballast gradations. The accompanying user-friendly graphical interface integrates all data processing algorithms and provides comprehensive visualizations of results. Field data was collected using the BSV, from cut trenches opened using a ballast regulator, along the High Tonnage Loop (HTL) at the Transportation Technology Center (TTC) in the U.S. The FI and gradations from the BSV were compared to laboratory sieve analyses and Ground Penetrating Radar (GPR) data. Additional laboratory tests with various fouling conditions were conducted to validate the deep learning algorithms and clarify any differences between sieving results and algorithm estimates possibly attributed to sampling issues. This field deployment demonstrated that the BSV could accurately evaluate ballast conditions close to real time, thus making it a robust system for quantifying ballast degradation.