Predicting Track Geometry Data using Multimodal Methods with Neural Network Models

Asset management and maintenance of rail track assets often rely on data collected using remote condition monitoring approaches. Through wear due to freight or passenger trains passing over the track, different faults can develop that can lead to serious safety and performance risks. Observing the wear on the track requires a variety of methods to monitor the track adequately, as signs of faults can depend on their type. Some faults may need multiple sources of data from different monitoring approaches to accurately identify the fault and the level of damage. Being able to identify faults in advance is a predictive maintenance goal of railway infrastructure providers, such as Network Rail, due to the potential to avoid paying compensation to train operators from unplanned maintenance affecting train schedules (Xie et al., 2020). Track maintenance engineers work to accomplish this by utilising different modes of data, including digital imagery, track geometry data, and other types of data depending on the scenario (García et al., 2009). Inspired by recent developments in computer vision, this paper examines the potential viability of incorporating multimodal approaches for prediction of track geometry, using high-definition photographs and historical track geometry readings. The performance of different neural network architectures is compared in the context of track condition monitoring data. This results in predictions of future track geometry data which include historical readings to aid in prediction. Further work would aim to continue developing models to predict dates and classifications of future faults and failures directly.

Keywords: Track geometry prediction, Neural networks, Resilience engineering.