Machine Learning, Probabilistic and Mathematical Models for Damage Recognition in Structural Systems
MetadataShow full item record
An increasing percentage of building and bridge structures across United States are exceeding their design life. Ensuring the structural integrity of such structures demands health monitoring strategies. Over the past years, various structural health monitoring approaches have been developed for detection of the intensity and the location of damages in various structures such as building and bridges. Advancements in structural health monitoring rely on development of sensors, and development of mathematical models that can interpret the sensor data to meaningful information about the health of the structure. As an example, cables of cable-stayed bridges play a critical role in cable-stayed bridges by transmitting the forces from the bridge deck to the pylons. Hence, assurance of integrity of the cables during the design life of the bridge is inevitable. In this research, two mathematical models were developed to detect and quantify damage in cable-stayed structures. The first model uses the distributed measurement of strains along the bridge deck to detect the cables that have totally or partially lost their tensile force. The fundamental principle employed in formulating the method is the interrelationship between the individual cable forces and the bending moment along the bridge span. The efficiency of the methodology was evaluated through both numerical simulations and experimentations on a reduced-scale cable-stayed bridge. The second proposed model utilizes point-style sensors to estimate the deck element shear forces adjacent to the supports. An analytical approach was developed to quantify the damage in the cables using the shear forces. The formulations are based on a recursive optimization technique, in which model updating is employed to account for the changes in the cable stiffness as a result of damage. In addition to mathematical models for damage recognition in cable-stayed bridges, the current research proposes a hybrid approach based on machine learning models to determine damages in the structural elements of building structures. The implemented machine learning methods in this thesis are Support Vector Machines, Neural Networks and Gaussian Naïve Bayes. The efficiency of the proposed approach was evaluated through numerical simulations and actual experimentations on a six-story heritage timber-masonry building.
Structural Health Monitoring