A Computer Vision-Based Structural Health Monitoring Framework: Feature-Mining of Damage for Predictive Numerical Simulations

For the infrastructure in U.S., the structural health monitoring (SHM) community has generalized a strategy with a primary focus of accurately monitoring in-situ behavior to assess performance, detecting damage, and determining structural condition. At the core of this strategy is the need to identify and quantify damage, but also to predict the implications of this damage on the structural system. However, for scenarios where this damage is not visible, this challenge becomes amplified due to the potential for structural failure in the presence of this unknown risk. For these types of structures, a more local strategy is needed, one that is able to provide in-situ information about the current condition state of performance of an individual structure in the absence of previous baseline performance data. This need drives our research question; can internal damage be effectively identified using only limited surface observations obtained from image-based sensing techniques as a non-contact full-field approach?
Recent advancements in camera technology, optical sensors, and image-processing algorithms have made optically-based and non-contact measurement techniques such as photogrammetry and 3D Digital Image Correlation (3D-DIC) appealing methods for non-destructive evaluation (NDE) and SHM. Conventional sensors (e.g. accelerometers, strain gages, string potentiometers, LVDTs) provide results only at a discrete number of points. Moreover, these sensors need wiring, can be time-consuming to install, may require additional instrumentation (e.g., power amplifiers, data acquisition), and are difficult to implement on large-sized structures without interfering with their functionality or may require instrumentation having a large number of data channels. On the contrary, optical techniques can provide accurate quantitative information about full-field displacement, strain and geometry of a structure without contact or interfering with the structure’s functionality.
This dissertation centers around recovering unseen damage within a structural system using limited, but full-field surface deformation measurements. The proposed approach leverages full-field surface deformation measurements of structural elements derived using 3D-DIC coupled within a structural optimization process to search for and identify the presence of invisible damage. The idea initiates from preliminary work that has proven successful in identifying constitutive properties implied for quantifying damage from material distribution in structural specimens. While this preliminary work was promising, the concept needed further research to extend the framework towards a more robust approach that can be used for in-situ assessment of in-service structural systems. The research herein centers on a laboratory scale investigation of structural components, which exhibits variability in its constitutive properties that are typically uncertain within existing structures and is also vulnerable to internal damage and/or heterogeneous material distributions that are unseen from the surface.
First, full-field sensing measurements from 3D-DIC was applied to update a finite element model (FEM) of a full-size I-shaped steel beam under flexural loading. A hybrid optimization algorithm consisting of a gradient-based and a genetic algorithm (GA) optimizer was introduced to attain and optimize the structural unknowns including constitutive properties and boundary condition assessments. The updated model was illustrated to generate improved estimations of the response through comparisons with ground truth measurements acquired from discrete sensors. Second, based on the previous fact that constitutive properties can be resolved accordingly using St-Id using full-field sensing, the framework was extended to identify regions with internal defects in steel specimens. This work employed a hybrid algorithm combining a GA and a limited-memory Broyden-Fletcher-Goldfarb-Shanno algorithm (LBFGS-B) to execute the optimization problem. While the method showed promise in detecting the existence and vicinity of the defect, recovering the 3D shape of the defect was not possible.
Most recently, the work went one step further and a new approach using full-field surface measurements coupled with topology optimization was proposed to localize and reconstruct the 3D shape of unseen subsurface defects. Thus it aimed to expand on the work and to demonstrate that unlike a limited set of discrete sensing data points or global dynamic properties, the rich data from full-field image-based measurements can enable the identification of a more detailed picture of the internal defects. The main contributions of this dissertation can thus be summarized as: 1) Unlike NDE/T techniques which depend upon specialized sensing equipment (e.g. radars or radiation-based scanners, etc.), the introduced method solely applied digital cameras coupled with structural mechanics to deduce subsurface conditions. 2) The proposed method leverages the rich full-filed response data from DIC to enable the reconstruction of the 3D shape of damage, representing an advancement over current practice which has been limited primarily to identification and basic localization.