Probabilistic Approach to Evaluate Need for Rail Remediation Using Large Data Sets
Academic Institution: University of Edinburgh
Industry Partner: Loram Ltd
Academic Supervisor: Prof Michael Forde
PhD Student: Konstantin Popov
Summary:
As railway track beds age, they deteriorate due to dynamic loads from traffic, as well as wear and tear. If they do not perform at an optimal level, significant train delays and cancellations can lead to large financial penalties for the operators and an overall poor passenger experience. The line of interest in this study accommodates high-speed trains, which can result in very large penalties. Major maintenance activities require special machinery and must be planned months and potentially years in advance. Evaluating track quality is a challenging process, which requires a lot of information and expertise. Data in the form of mainly track geometry, but also GPR (radar) scans and other geophysical surveys is periodically collected to assess track performance. Maintenance works in the form of tamping are often carried out excessively in track segments of high quality which will in turn exhaust the ballast and reduce its lifespan. Track defects are usually exhibited in highly localised regions that require local tamping. By combining the various data sets collected by rail operators and applying machine learning techniques to analyse them individual defects can be detected. This allows maintenance to be targeted at eliminating local defects, rather than covering high quality track.
A reduction in tamping will result in significant cost savings in both maintenance and renewal activities, which make up a majority of Network Rail’s annual budget. A single tamping shift costs approximately £25,000 and maintenance costs across the UK amounted to over £1.8 billion in 2020-2021. Renewal works took up over £3.9 billion of the budget in that period. The two figures made up 56% of Network Rail’s annual budget. Reducing track tamping will lower maintenance cost, but also further extend track life. As such, renewals will be less frequently required. A reduction of 10%-20% of the costs associated with maintenance and renewals can lead to a significant overall reduction in annual expenditure. Machine learning algorithms have been developed for anomaly detection and clustering purposes, so that individual defects can be identified and a further root-cause investigation can be performed. This will help understand track behaviour better and perform targeted maintenance to stabilise track quality and reduce overall costs.
This project studies the long-term behaviour of high-speed track in the UK, by applying machine learning techniques to big data sets. Track geometry will deteriorate over time, which affects a train’s response and can potentially lead to derailments. To prevent this and provide high quality rides, maintenance is performed regularly. Track tamping is a process which induces vibrations in the ballast in order to adjust its geometry to a desired state. However, the vibrations cause fine ballast particles to break off, which will then begin to infiltrate voids. This is known as fouling and over time it reduces track bearing capacity, leading to excessive deformation. Thus, over-tamping a track is detrimental in the long-term and needs to be minimized whenever possible.
Key Results/Outcomes:
· Artificial Neural Network (ANN) developed to assess high-speed track maintenance efficiency.
· The track is found to be dominated by individual defects, which are best addressed using local maintenance works.
· An anomaly detection tool (autoencoder) was used to identify poorly performing regions of track and follow the progression of defects.
· It has been established that while tamping usually corrects track geometry, it does not always stabilise the conditions and repeat faults can occur shortly after. This may require further tamping on a short timescale.
· The majority of defects were found to be in close proximity to special track assets, such as switches and crossings and under-track utilities. Such regions require actions other than tamping in order to improve long-term quality.
· Using algorithms such as the ones developed in this study, maintenance approach can be refined, which will lead to a reduction in associated costs and extended track life – less frequent renewal.
· Further case studies have been performed to confirm applicability of these techniques to other tracks exhibiting different behaviour. Results indicate that models trained on the high-speed track can be applied to lower-speed lines as well without any adjusting or re-training.
Conference posters:
· Pilot Study of High Speed Rail Track Monitoring using AI and Machine Learning - poster presentation at the Transportation Research Board annual conference, January 2021.
· Big-data driven assessment of railway track and maintenance efficiency using Artificial Neural Networks – paper accepted and published in the Journal of Construction and Building Materials, August 2022.
· Rail Track Monitoring using AI and Machine Learning – paper presented at the International Conference on Trends on Construction in the Post-Digital Era, September 2022.
· Railway track performance enhancement using machine learning techniques – progressed to the final round of the three-minute thesis competition at the Transportation Research Board annual conference, January 2023.
· Using Artificial Intelligence & Machine Learning to reduce ballasted track maintenance costs – presentation at the Permanent Way Institution (PWI) monthly meeting, February 2023.
· Data-driven track geometry fault localisation using unsupervised machine learning – paper submitted to the journal of Construction and Building Materials, February 2023. Currently under review.
Collaborations
· The project has included close collaboration with Network Rail and Network Rail High Speed Ltd. Regular meetings have been held to exchange information and better understand the needs of the industry. Data has been acquired from three separate railway tracks in the UK. In September 2021, I undertook a placement at Network Rail for 4 weeks, where I gained more knowledge of the track by speaking with key engineers in the office. In October 2021, I attended a workshop hosted by SNCF and Sol Solutions in Paris, France along with Network Rail to discuss and learn more about track bed investigations being done in the UK.
· Throughout the entirety of the project the University of Edinburgh has benefited from collaborative work with Prof Carlton Ho from the University of Massachusetts, Amherst. Prof Ho has an extensive background working with the railway industry in the USA and has provided valuable insight regarding track behaviour. Our work together has resulted in several publications.
· The University of Edinburgh has recently collaborated with the University of Birmingham as well. New data describing track behaviour was collected to further aid the research done at Edinburgh. The work is currently on-going.
Contact Information
Konstantin Popov
PhD student and teaching assistant (IEE), University of Edinburgh
Professor Michael Forde,
Chair of Civil Engineering Construction, University of Edinburgh.