Rail squat detection using hybrid processing of axle box acceleration measurements

• Hybrid method for detecting rail squats using axle box acceleration measurements • Hybrid signal processing method addresses existing limitations and enhances performance • Efficiently identifies rail squats under various background noise levels • Attains over 96% data compression, making it highly suitable for edge computing • Validated through field measurements, identifying previously undetected defects Rail squat is the most common type of Rail Surface Spot Irregularity (RSSI) that comes into direct contact with the wheel surface and is distinguished for its relatively short length. As these defects progress within the rail, they compromise operational safety and accelerate wear on train and track components. Early detection of squats at marginal stages offers significant advantages for preventive maintenance, as minor squats can be resurfaced, avoiding costly rail replacements. This study introduces a squat detection method that identifies the location and a proxy of the size of squats along a track using Axle Box Acceleration (ABA) measurements collected from in-service trains. The method adapts a hybrid signal processing approach that combines Wavelet Packet Analysis (WPA) and the Hilbert-Huang Transform (HHT). The proposed technique leverages the strengths of each method, mitigating their individual limitations to improve overall performance. Initially, the algorithm is implemented in a numerically simulated environment, which accounts for various levels of background noise contamination in the measurements. This simulation ensures the algorithm's robustness and effectiveness before its application to real-world data. Subsequently, the method is validated through field measurements. To this end, ABA measurements collected from an instrumented railcar under regular passenger service operations were obtained. The algorithm was implemented and successfully identified multiple squats and other types of RSSI on each rail of the track, as corroborated independently by the operator. Furthermore, the algorithm achieves remarkable data compression of the recorded signals, by over 96%, all while retaining essential information needed for squat identification.