Condition modelling of railway drainage pipes

Performance of drainage asset systems can have substantial impact on the structural and operational integrity of railway tracks. It is therefore important to ensure that the various components of the drainage system are well-maintained. To this end, decision makers in the railway industry have been moving towards predictive, risk-informed drainage asset management. The approach aims to optimise the allocation of the limited time and financial resources for maintenance works. To achieve this more research is required to develop predictive condition models for railway drainage assets.

This paper describes the development of data-driven condition prediction models using drainage pipe asset records. The models were tested for both structural and service condition prediction. Nine input factors were considered in the prediction models. Significance of the factors was evaluated using Connection Weight Analysis. Four Machine Learning (ML) algorithms, namely, Neural Networks, Decision Trees, Bagged Trees, and K-Nearest Neighbour, were compared based on their condition prediction performance for pipe drainage assets. The models were developed and tested using field data collected from the UK owner of rail assets, Network Rail. The results demonstrate that Bagged Trees performed best on a balanced dataset with 87% overall accuracy for structural condition prediction and 72% accuracy for service condition prediction. It was found that pipe length, previous condition, years since previous condition and maintenance are the most significant factors in predictive performance.