PhD in Population-Based Indirect Damage Detection System for Railway Bridges

The UK’s railway bridge asset stock represents over 80% well-aged (>50 years) infrastructure, often carrying loads beyond their original design capacity, hence in urgent need of a reliable real-time damage identification system. The current practice of visual inspection of bridges can be subjective and exposes the workforce to hazardous job sites. In recent years, there have been significant efforts in instrumenting bridges and assessing the condition of the bridge using direct measurements. These methods are categorised as non-destructive testing techniques, but they can be costly considering the number of sensors required and the maintenance of the data acquisition system. Hence, the alternative of direct instrumentation of the structure, whilst effective, can be logistically expensive to implement for the entire network.

To address these challenges, the project aims to develop a novel, population-based indirect damage identification system, leveraging data collected on instrumented railway vehicles to autonomously assess bridge condition while passing over the structure at operational speed, providing a scalable and cost-effective alternative to traditional methods. The fundamental principle in indirect damage inspection is that damage causes changes in physical properties of the structure, which can lead to altering the vibration behaviour of the structure. The challenge in indirect damage inspection methods is to identify and extract these changes from the measurements recorded on the travelling vehicle while it is driving over a damaged bridge at operational speed.

Due to a lack of large, real-world datasets with ground truth labels, the application of data-driven approaches in the indirect damage identification context, while promising for network-level monitoring, has been largely underexplored. To this end, the project will explore the application of the next generation of deep learning algorithms, e.g. self-supervised learning techniques, particularly suited to infrastructure applications where labelled data is scarce, enabling models to learn from the data itself without relying on extensive human annotation.