Artificial neural networks for non-destructive identification of the interlayer bonding between repair overlay and concrete substrate

Abstract The article presents the application of artificial neural networks (ANNs) for the non-destructive identification of the pull-off adhesion fb values between the repair overlay with variable thickness and the substrate in concrete surface-repaired elements. For this purpose, a large database was built on the basis of the tests of model concrete elements. The numerical analyses were performed using this data and ANNs with various learning algorithms. Based on these analyses, it was shown that the ANN with the Broyden-Fletcher-Goldfarb-Shanno learning algorithm, with thirty-one input parameters and twenty hidden neurons, is the most useful for identifying the interlayer pull-off adhesion in repaired concrete elements. The reliability of the presented application of ANNs was confirmed on the basis of carried out validation, using a part of the database not used in the learning and testing. The application's reliability was also confirmed on the basis of experimental verification carried out using the results of tests performed on an additional model element made exclusively for this purpose. This is an important and original issue presented in the article. Another novelty presented in the article is the application of ANNs for a much more difficult case, which is the identification of the pull-off adhesion fb value of the repair overlay of variable thickness from the repaired element and in a very wide range of identified pull-off adhesion fb within the range of 0.5–3.60 MPa. Moreover, the unique value of the article is the use for the first time of spatial and function related parameters to describe the concrete surface morphology of a repaired element. The investigation presented in the article has also confirmed the high usefulness of these parameters for identifying the value of pull-off adhesion fb.