Testing the Accuracy of Machine Learning-based Crack Localization Methods using Damage Localization Coefficients

Abstract

Artificial intelligence is often used to assess the integrity of engineering structures. Many methods are available to assess different types of damage, but the correctness of the results is not proven until local inspection methods are applied. Therefore, there is a need to develop a tool that can estimate the accuracy of the assessment process through a supplementary intervention. In this paper, we propose and test a procedure to establish the accuracy of the damage assessment results, which is the follow-up of a normal damage assessment process. First, we assess the damage involving a method previously developed by the authors that consider the relative frequency shifts (RFS) for several bending vibration modes. The method has the support of artificial neural networks (ANN). Applying this method, we estimate the location and severity of the damage. Next, we apply a procedure that presumes first to calculate the modal curvatures and the resulting damage location coefficients (DLC) for this location. Then, we normalize the RFSs used in the assessment process and compare them with the DLCs derived analytically for the presumed damage location. Finally, we compare the DLCs with the normalized RFSs via the Euclidian distance. This comparison shows how accurately we assessed the damage location, the smaller the distance, the better the prediction. Applying this procedure as a follow-up of a standard damage detection process, we know the accuracy of the assessment prediction realized with a standard detection method. If the accuracy is unsatisfactory, we can use an ANN model that is trained with data from the supposedly defective area.