Periodic Reporting for period 1 - AMONTRACK (Acoustic monitoring of railway track quality)
Reporting period: 2018-05-01 to 2019-06-30
The project AMONTRACK focused on the development and implementation of methods and tools for acoustic monitoring of railway tracks. For a well-functioning railway infrastructure, it is essential to detect faults in an early stage. The potential consequences of advanced track damage can be damage to rolling stock, high maintenance costs, or even the interruption of transport corridors. Consequently, the social costs due to damages to the track system can be high. At the same time, the railway transport system is a crucial contributor to achieving sustainability in transport and the reduction of CO2 emissions.
Track damage often starts with small irregularities on the rail surface. Over time these irregularities grow to severe faults and eventually to track failures. As the control of modern track infrastructure by humans walking along tracks is impractical, and the detection of early-stage defects by visual inspection is difficult, an automated way of detecting localized rail and track default is important. The possibility of using acoustic track monitoring for this purpose has been investigated in AMONTRACK.
The goal was to identify defects on the track in an automated way from acoustic on-board measurements on a specially equipped railway car. To reach this goal, the project focused on four research objectives:
1. Identification of acoustics signatures of localized rail and track faults in the measurement signals based on the type of fault and its severity.
2. Understanding of the influence of different track design parameters and the rail roughness on the measured axle box acceleration and the sound pressure over a bogie of the railway car.
3. Clarification of what information can be extracted from the measured data by going the inverse way from the measurements to the source.
4. Suggestion of an algorithm to detect faults and to extract information from the measured signals based on pattern recognition.
In the project, an algorithm to detect a rail surface defect called ‘squat’ from measured axle box acceleration on-board the train has been implemented. The algorithm is based on machine learning and has been tested in a full-scale experiment on a track with well-documented squats. The approach has shown to work well and was able to identify squats with sufficiently high accuracy. The results indicate that the proposed method is worthwhile to be developed further for use in the field to detect squats in an automated way, resulting in more efficient track monitoring.
Track damage often starts with small irregularities on the rail surface. Over time these irregularities grow to severe faults and eventually to track failures. As the control of modern track infrastructure by humans walking along tracks is impractical, and the detection of early-stage defects by visual inspection is difficult, an automated way of detecting localized rail and track default is important. The possibility of using acoustic track monitoring for this purpose has been investigated in AMONTRACK.
The goal was to identify defects on the track in an automated way from acoustic on-board measurements on a specially equipped railway car. To reach this goal, the project focused on four research objectives:
1. Identification of acoustics signatures of localized rail and track faults in the measurement signals based on the type of fault and its severity.
2. Understanding of the influence of different track design parameters and the rail roughness on the measured axle box acceleration and the sound pressure over a bogie of the railway car.
3. Clarification of what information can be extracted from the measured data by going the inverse way from the measurements to the source.
4. Suggestion of an algorithm to detect faults and to extract information from the measured signals based on pattern recognition.
In the project, an algorithm to detect a rail surface defect called ‘squat’ from measured axle box acceleration on-board the train has been implemented. The algorithm is based on machine learning and has been tested in a full-scale experiment on a track with well-documented squats. The approach has shown to work well and was able to identify squats with sufficiently high accuracy. The results indicate that the proposed method is worthwhile to be developed further for use in the field to detect squats in an automated way, resulting in more efficient track monitoring.
The work in the project mainly focused on the detection of squats, which are one of the most common rail faults. Squats are defects on the running surface of the rail that occur due to rolling contact fatigue. They are characterized by a crack on the railhead accompanied by a depression of the running surface and a widening of the running band. Squats lead to high impact forces that further promote crack growth along and across the rail and can eventually lead to rail failure. They also contribute to fatigue of the rolling stock.
A measurement car (SMW) of DB Systemtechnik, which records the sound pressure signal above one of its bogies and the corresponding vibrations at the axle boxes, was utilised to identify track faults.
For the identification of squats, the vibrations on the axle boxes of the SMW were evaluated both by measurements and simulations. For the latter, the advanced simulation tool WERAN of Applied Acoustics, Chalmers University of Technology, was adapted and extended to calculate the whole chain including excitation in the wheel/rail contact, the dynamics of wheelset and track covering the needed frequency range up to 2 kHz and the propagation of vibrations to the axle box. The simulation tool is suitable to investigate the influence of different track design parameters on the observed axle box vibrations. The simulation approach has been validated with measurements in the field.
Both measurements and simulation of the axle box vibrations due to rail/wheel roughness and due to squats have been studied to identify possible characteristics for the detection and – although not implemented here - for the quantification of squats. As the primary characterization quantity, spectrograms have been analyzed as input for a logistic regression model.
A detection algorithm based on the logistic regression model has been trained with both simulated and measured data and tested in a full-scale experiment on a track with well-documented squats. The approach achieved high hit rates for both severe and light squats.
A measurement car (SMW) of DB Systemtechnik, which records the sound pressure signal above one of its bogies and the corresponding vibrations at the axle boxes, was utilised to identify track faults.
For the identification of squats, the vibrations on the axle boxes of the SMW were evaluated both by measurements and simulations. For the latter, the advanced simulation tool WERAN of Applied Acoustics, Chalmers University of Technology, was adapted and extended to calculate the whole chain including excitation in the wheel/rail contact, the dynamics of wheelset and track covering the needed frequency range up to 2 kHz and the propagation of vibrations to the axle box. The simulation tool is suitable to investigate the influence of different track design parameters on the observed axle box vibrations. The simulation approach has been validated with measurements in the field.
Both measurements and simulation of the axle box vibrations due to rail/wheel roughness and due to squats have been studied to identify possible characteristics for the detection and – although not implemented here - for the quantification of squats. As the primary characterization quantity, spectrograms have been analyzed as input for a logistic regression model.
A detection algorithm based on the logistic regression model has been trained with both simulated and measured data and tested in a full-scale experiment on a track with well-documented squats. The approach achieved high hit rates for both severe and light squats.
The originality of the project stemmed from the combination of a unique measurement train and a unique numerical model for wheel/rail interaction, which generated new possibilities for acoustic condition monitoring of the track. The analysis of axle box acceleration with the accurate and efficient numerical model allowed identifying acoustic signatures of localized rail surface defects for arbitrary wheel/rail combinations with reasonable calculation effort. Training a machine learning algorithm with such signals was a new approach in the context of acoustic condition monitoring of the track. The resulting methodology allows detecting squats from the axle box acceleration of the measurement train and is easily adapted to railway tracks with different dynamic characteristics. Consequently, the project opened up for an application range of condition monitoring beyond the state of the art.