Managing Railway Bridges Crossing Waterways through a Machine Learning-Based Maintenance Policy

Abstract

Recently, more frequent and severe natural hazards that are caused by climate change have posed a great threat to the safety of transport systems worldwide. To enhance bridges’ resilience to natural hazards, this paper proposes a new maintenance policy that is based on machine learning (ML) for managing bridges that cross waterways in France. Two ML models, for example, random forest (RF) and extreme gradient boosting (XGBoost) classifiers, are tested on bridges in France and Japan to investigate the model’s practicality and robustness. Data from these bridges has never been seen by the model before; however, it is in the same range as the original data set. To verify the test results on the unseen data, predictions from the French cases are compared with engineering judgment, and they are in agreement (95% between the senior engineer and the XGBoost model). When comparing the Japanese case test results with the Japanese guideline’s scoring table (ST), predictions are not as accurate as in the French cases. This might be caused by the different data distribution between the two countries and the lower threshold for high scour risk cases in the Japanese guidelines. Based on the results of the original and unseen data sets, application scenarios are suggested for each model. Finally, to facilitate the use of the proposed model, a friendly web application was demonstrated to reduce computational complexity. The outcome of this paper could help to identify bridges that are vulnerable to scour in an effective yet intelligent way, which will, in the end, ensure the safety of the rail network. In addition, it could provide insights to other countries’ transport agencies who want to develop their ML-based maintenance policy. Scour has been identified as one of the major causes of bridge failures worldwide. This refers to a process of erosion or removal of material from the riverbed. Several factors could contribute to the occurrence of bridge scour, such as high-velocity water flow, sediment transport, channel alignment, and sediment composition. A practical process for identifying the railway bridges that are vulnerable to scour has not been developed in France. To address this issue, this paper provides a novel maintenance policy that is based on machine learning (ML). The proposed ML models are tested on bridges in France and Japan. Data from these cases has never been seen before; however, it is in the same range as the original data set. The predictions from the unseen data are compared with the engineering judgment and Japanese practical guideline, the scoring table (ST). Based on the results, application scenarios are suggested for the models. This paper aims to improve the inspection and maintenance process at the French National Railway Company [Société Nationale des Chemins de fer Français (SNCF)]. A user-friendly web application is built at SNCF to ensure the accessibility of the research outcome.