Presently, railroad monitoring strategies focus on preventative maintenance by detecting wheel anomalies using wayside detection methods (e.g., wheel-impact load detection), and direct detection of track anomalies using onboard systems (e.g., track geometry vehicles). Both approaches are periodic, manual, and do not support real-time track damage detection. Recent research has focused on detecting damage from acceleration signals obtained onboard moving vehicles and identifying anomalies from derived structural dynamic properties. Though promising due to inherent scalability and cost efficiency, its main goal is to detect damage on the supporting infrastructure and has never before been tested for detecting rail crack damage. Among other reasons, a robust anomaly detection algorithm is missing to allow the industry to embrace an automated and more cost-effective monitoring technique. In this work, we leverage a lab-scale track and moving vehicle actuation system that is scaled with the assistance of industry experts, and comprises a vehicle instrumented with two onboard vertical accelerometers. Cracked rails are simulated by introducing discontinuities (longitudinally and transversely). Several types of feature extraction and dimensionality reduction techniques are employed to evaluate their ability to separate damaged and undamaged records. Inspired from previous work, this work tests the ability of existing data-driven damage detection algorithms to detect local damage by using a novel super modular, precise, and realistically scaled down version of a train-track system. The results of the damage sensitivity show that principal component analysis has the highest balanced combination of recall and true negative rate, compared to other techniques.
|