Health Monitoring of Structures and Damage Detection Using Vision-based Methods and Artificial Intelligence
A significant portion of the maintenance cost of the bridges in the U.S. is due to the high cost of structural health monitoring. Currently, bridge health is assessed by visual inspection which can be user-biased, difficult to conduct, and sometimes unreliable. Traditional contact sensors, such as strain gages, displacement sensors, and potentiometers, are also being used for this purpose. These monitoring methods, however, require access to the measurement locations under bridges, which may not be possible or require traffic interruptions for system installation. The main objective of this dissertation is to develop novel vision-based system that addresses the issues with the current health monitoring techniques and can be utilized as an alternative approach to provide quantitative bridge structural health metrics. A high resolution three-dimensional Digital Image Correlation (3D DIC) system was developed to monitor deformations of structures, from small-scale coupons to bridges in the field. The software for the system was coded in the National Instruments LabVIEW programming environment. The capabilities of the system are demonstrated within the context of bridge monitoring under load testing. This study illustrates the capability of this system to monitor movement of any target selected on the surface of a bridge to a resolution on the order of 1/40th of millimeter (1/1,000th inch) even when the cameras are over 30 meters (100ft) away. Significant advantages of the new system include its relative low cost, ease of operation, ability to monitor any point over a large field of view in a single setup, and high accuracy necessary for assessing structural capacity from deformation measurements. The developed 3D DIC system was applied in diagnostic load testing of several bridges in the state of Texas. Different types of bridges including flat slab concrete bridges, multi girder steel bridges, and concrete culverts were load tested and the distributions of live load throughout the bridge deck and among girders were investigated and compared with approximate distribution factors provided by AASHTO Standard Specifications and AASHTO LRFD. Load testing results provided higher load rating outcomes than previously calculated, and therefore indicated that standard assessment methodologies for existing bridges may be too conservative in some instances. Outcomes of the work provided justification for removing load postings for several bridges. In addition to diagnostic deformation measurements, images recorded from the DIC system were used to advance vision-based health monitoring of infrastructure, with particular focus on crack monitoring using two vision-based techniques. The first method uses DIC measurements and finite element method to monitor crack growth and measure crack width of structures at different stages of stages of loading. The second method focuses on locating cracks on concrete surfaces in an automated manner. A fully convolutional neural network (FCN) is implemented to develop a model that detects cracks at pixel level. The developed network is trained on images of cracked concrete surface of structural elements such as beams and columns during load testing. The developed crack segmentation method has accuracy, precision, and recall values of 93.96%, 99.8% and 78.97%, respectively.