Automated Squat Detection in Railway Switches & Crossings Using Vibration Measurements and Machine Learning

Railway switches and crossings (S&Cs) play a vital role in the operational integrity of a railway network, and their malfunctioning poses substantial risks such as traffic disruptions, economic ramifications, and potential life-threatening accidents. Consequently, it becomes imperative to proactively identify potential defects in S&Cs at an early stage and maintain continuous monitoring of their condition to avoid adverse consequences. Among the various types of defects that may arise, a notable one is referred to as a squat, which manifests as a localized anomaly like a dent or an open pit on the rail surface. The thesis comprises a series of five studies aimed at investigating the automatic detection and quantification of squat defects in railway S&Cs. The initial three studies were conducted on a testbed, consisting of a full-scale S&C and a bogie, to assess the feasibility of identifying and evaluating the severity of squats. Vibrational measurements were obtained using accelerometers positioned at the point machine, while the bogie traversed the S&C with squats of varying sizes. The first study proposed a data processing method, combining vibration and speed data, to monitor the progression of squat size within the testbed. In the second study, a feature extraction method and an isolation forest algorithm were deployed to automatically generate anomaly scores, providing an estimation of the health condition of the S&C for condition monitoring purposes. Wavelet denoising, a fundamental technique for noise reduction in vibration data, was incorporated in this study. The third study introduced additional unsupervised machine-learning algorithms for the identification of distinct levels of squat defects within the testbed. Subsequently, the fourth study involved data collection from Notviken (Sweden) in a semi-controlled field test to verify the applicability of the method proposed in the second study beyond the confines of the testbed. Lastly, the fifth study utilised data obtained from real field tests with in-service trains at Gammelstad (Sweden) to further validate the feasibility of the utilisation of vibration data collected from rails close to the point machine for the detection of squats in S&Cs under real operating conditions. The findings derived from the conducted studies substantiated the feasibility of utilising vibrations collected from or near the point machine as a mean to detect and track the progression of the