Unsupervised Machine Learning for Robust Bridge Damage Detection: Full-Scale Experimental Validation

• Published in: Engineering structures Vol. 249; p. 113250
• Main Authors: Akintunde, Emmanuel, Eftekhar Azam, Saeed, Rageh, Ahmed, Linzell, Daniel G.
• Format: Journal Article
• Language: English
• Published: Kidlington Elsevier Ltd 15-12-2021; Elsevier BV
• Subjects: Bridge decks; Damage assessment; Damage detection; Field tests; Herbivores; Independent component analysis; Learning algorithms; Live loads; Machine learning; Mockups; Robustness; Singular value decomposition; Structural response; Unsupervised learning
• ISSN: 0141-0296; 1873-7323
• DOI: 10.1016/j.engstruct.2021.113250

•An unsupervised Machine Learning scheme is developed for bridge damage detection•Validation is carried out using full scale experiments, a crash test for inducing realistic damages•The method does not require measuring vehicle weights and speeds•It is robust to sensor noise & works sparse sensor networks and low sampling rates•This method detected crash induced damage – it was undetected by visual inspection
This study focused on the development of damage detection indices using unsupervised Machine Learning with data obtained from tests of a full-scale bridge deck mock-up. Measured structural response to different live loads at various damage levels was analyzed and a detection tool, termed a novelty index, was developed. Three levels of damage were induced into the structure: crash-induced damage to the concrete barrier; “sawcutting” the entire height of the barrier; and extending the sawcut through the deck. Bridge response to passage of vehicles at various speeds was recorded for the undamaged structure and after introduction of each damage level. The main novelty of this work is development of the novelty index, its application and validation against measured field data from a full-scale field test of a mock-up bridge subjected to varying levels of damage. Data was incorporated into health monitoring frameworks previously developed by the authors (1, 2) and the viability of using Singular Value Decomposition (SVD) and Independent Component Analysis (ICA) as damage sensitive features were explored. Novelty indices were subsequently developed using two approaches: Left Singular Vectors (LSVs) of measured response, commonly known as Proper Orthogonal Modes (POMs); and Independent Component Modes (ICMs). It was shown that a novelty index based on POMs was relatively insensitive to vehicle load variability and speed when compared to ICMs. It was also shown that novelty indices based on POMs could detect the lowest amount of induced damage, which was caused by a crash into a bridge barrier. Damage induced by the collision did not lead to visible barrier cracks; however, a quantitative assessment indicated damage levels like damage caused by a “sawcut” of the barrier. This study shows that the proposed method has the potential to be a robust damage detection tool; however, its application for detecting damage types other than the one studies herein needs to be further investigated due to statistical nature of SVD.