Detection and quantification of delamination in concrete via time-lapse thermography with machine learning

This study developed a framework to automatically extract sub-surface defects from time-lapse thermography (TLT) images of reinforced concrete bridge components. Traditional approaches for processing TLT data typically require manual interventions that are not easily scaled to a large network of concrete bridges. A backbone of robust algorithms for detecting and analyzing deep sub-surface defects in concrete is needed to support condition assessment of concrete structures such as bridges. The current study leverages advances in adaptive signal and image processing to develop a fully automated TLT data processing pipeline that is capable of efficiently detecting defects at different depths in concrete. The methodology decomposes raw TLT datasets into narrow band time-frequency domains via a multiscale data analysis approach called a Wavelet Transform. The resulting decomposed modes are mined to extract defect information using thermal contrast enhancement routines. An objective measure of effectiveness based on signal-to-noise ratio was developed and used to compare the current framework with traditional approaches for processing TLT data. Active contour models were also designed to automatically extract the boundary location and geometric properties of the sub-surface defects. The results of this study show that the detection of deeper defects (3 in. and beyond) can be improved by analyzing the time-frequency response of surface temperature variations over a period of time. Compared to traditional lock-in algorithms and conventional infrared thermography images, the proposed framework is more effective at removing noisy information and produces images with greater contrast between intact and defective areas of concrete. Furthermore, a new process has been established to predict depths of delamination in reinforced concrete bridge components. For previous works, traditional approaches were adopted to quantify depths in active thermography, which mainly depend on estimated models as a function of time, frequency, phase contrast, material properties of specimens. This work deals with the passive thermography that is affected by several environmental parameters such as solar heating, daytime or nighttime, wind speed, clouds, shadow. The current work has employed the Machine Learning (ML) technology to estimate defect depths in concrete block. Features, such as phases, amplitudes, frequencies, have been extracted by utilizing the Fast Fourier Transform (FFT) in a stage of analysis. Furthermore, additional subfeatures, minor features, have been added to the ML analysis, for instance average and/or subtraction values between the maxima and minima features, to attain an acceptable learning performance. Support vector machine (SVM) and k-Nearest Neighbor (KNN) classifiers have been trained by using crossvalidation with different folds and hold validations. The predicted models have achieved an improved accuracy in estimating delamination depths in the concrete specimens with a good agreement.Includes bibliographical references