Damage Detection in Steel Railway Bridges using Vibration Data and Machine Learning Approach

Abstract

Steel railway bridges are vital members of railway and transport infrastructures, worldwide. These bridges are exposed to various external effects, such as high levels of dynamic train loads as well as adverse environmental impacts that cause them damage during the time. In the rail industry, conventional physical and expensive inspections are carried out to monitor all types, extents, and degrees of damage. The effectiveness of the physical inspection, however, is always compromised by limited human resources, the delayed discovery of damages, lack of knowledge, and subjectivity of inspectors. In recent years, structural engineers and researchers have focused on vibration-based damage detection approaches to fill the existing gaps due to the low effectiveness of conventional railway bridge inspections. Generally, vibration-based damage detection approaches are categorized into two types; conventional and modern. The conventional approaches rely on vibration responses of structures e.g. derived by changes of natural frequencies and mode shapes due to damage while the modern approaches include data-driven methods that use signal processing, and then artificial intelligence such as machine learning techniques to formulate the relationship between the change of structural properties due to damages to finally diagnose and prognosticate them. This thesis focuses on the modern application i.e. machine learning techniques to detect damages in operational steel railway bridges using vibration data collected from practical field testing. As the operational bridges are real in-service structures and no actual damage can be applied to them, extensive efforts are made to develop close Finite Element (FE) models from the studied real bridges and to validate them with the extracted modal parameters of the bridge using the FE model validation techniques in the modal and transient dynamic analyses to ensure the incorporated damages into these FE models can capture any change of structural modal parameters due to damages with good approximation. In addition, the work completed in this thesis advances both fundamental understandings and applied knowledge on vibration-based structural damage or anomaly detection methods using machine learning applications such as Support Vector Machine (SVM) and k-Nearest Neighbors (kNN)s to more advanced deep learning approaches such as Convolutional Neural Networks (CNN)s and Siamese Convolutional Neural Networks (SCNN)s. Machine learning applications employ complicated signal processing to manually extract complex Damage Sensitive Features (DSF)s. Appropriate machine learning approaches to detect damages and advanced deep learning applications are developed in such a way that completely removes requirements for complicated signal processing or manual DSF extraction before detecting damages to increase the speed and effectiveness of the damage detection process. The novel development of the application of advanced generalized vibration-based deep learning for the anomaly detection of not only a single bridge but also a group of bridges with various lengths and configurations pushes the boundaries of structural damage detection techniques a step closer to a fully practical data-driven tool for the damage detection of a real bridge network using an integrated, centralized and all-in-one SCNN tool for a real-life Structural Health Monitoring (SHM) of the in-service strategic railway bridges.