A Machine Learning and Computer Vision Framework for Damage Characterization and Structural Behavior Prediction

Abstract

This work focuses on using computer vision to relate surface (and limited subsurface) damage observations to quantitative damage and load levels in structural components. In particular, image processing and machine learning regression techniques have been used to build predictive models capable of estimating internal loads (e.g., shear and moment) and damage states in RC beams, slabs, and panels based on surface crack pattern images. The predictive models have been generated and tested using image data sets obtained from earlier published studies, which together provide about 1,500 crack pattern images captured from about 170 individual RC beam, slab, and panel tests across a range of load and damage levels. Working with these existing image data sets, various textural and geometric attributes of surface crack patterns have been defined and evaluated with respect to their efectiveness in building useful estimation models. Relatively simple crack representations have been used, consistent with the varying nature of the images available in the earlier studies, but also with an eye toward potential field applications in which image capture and segmentation quality could be limited. The fundamental state quantification tasks range from the relatively simple (e.g., given an image showing damage, generate from that image a predicted load level in terms of percentage of capacity) to relatively advanced (e.g., estimating principal stresses in a panel). In addition to investigating the basic feasibility of the approach, these studies also identify and evaluate a range of strategies, algorithms, and parameters that affect the accuracy of the estimations, and these are discussed, as well. In general terms, the results show that the predictive models based on surface crack image data can work well across a wide range of geometries, loadings, concrete strengths, and reinforcement details. Size effects can be accounted for by including specimen physical dimensions in the feature sets used for model training, and fundamental design relations can be used to develop useful non-dimensional prediction parameters.