Master of Science
Civil and Environmental Engineering
Moncef L. Nehdi
Applications of Machine Learning (ML) algorithms in Structural Health Monitoring (SHM) have recently become of great interest owing to their superior ability to detect damage in engineering structures. ML algorithms used in this domain are classified into two major subfields: vibration-based and image-based SHM. Traditional condition survey techniques based on visual inspection have been the most widely used for monitoring concrete structures in service. Inspectors visually evaluate defects based on experience and engineering judgment. However, this process is subjective, time-consuming, and hampered by difficult access to numerous parts of complex structures. Accordingly, the present study proposes a nearly automated inspection model based on image processing, signal processing, and deep learning for detecting defects and identifying damage locations in typically inaccessible areas of concrete structures. The work conducted in this thesis achieved excellent damage localization and classification performance and could offer a nearly automated inspection platform for the colossal backlog of ageing civil engineering structures.
Summary for Lay Audience
Diagnosing damage is civil engineering structures and infrastructures has been getting increasing attention due to the very large portfolio of ageing civil assets and concerns related to its serviceability and safety. Until now, visual inspection has been the most used method to assess structural damage. However, in many cases, it is difficult and unsafe to access parts of such infrastructure (e.g., massive offshore bridge, large dam, tall building, etc.). At the same time, assessing damage can vary from one operator to the other depending on expertise and personal judgment. In this research, this subjectivity is mitigated using advanced statistical and probabilistic approaches such as artificial intelligence combined with image processing techniques to localize structural damage, quantify it, and even predict its type and degree of severity. This is done by implementing algorithms based on datasets of images for both damaged and intact structures. Then, depending on whether the structure is cracked or not, a quantification algorithm is developed to measure the width, length, and angle of orientation of cracks.
Nevertheless, in many cases, bridges, buildings, and other structures have collapsed without presenting any warning signs, for instance, via loss of the stiffness of key structural elements due to inner degradation that cannot be detected by visual inspection at the surface of the structure.
For this reason, a global technique based on signal processing is needed. When a random excitation is applied to a building, and its acceleration signals are measured, then damage features from the signal can be automatically extracted. Accordingly, the position of damage can be determined. The contribution of this thesis in this area is part of a larger effort to minimize and substitute to the subjective human operator in inspection and rehabilitation protocols. This study could, with further developmental work, optimize the service lifecycle, minimize maintenance costs, and mitigate failure risks for the lifetime of a civil infrastructure asset. Eventually, this research aims at making vital structures highly durable and long-lasting in Canada and worldwide. It might be very costly to erect new buildings and bridges, but we could give more life to the old ones at a lower cost.
Flah, Majdi, "Classification, Localization, and Quantification of Structural Damage in Concrete Structures using Convolutional Neural Networks" (2020). Electronic Thesis and Dissertation Repository. 7188.
Available for download on Friday, August 20, 2021