Effect of Mislabeled Data on Judgement Results for Re-bar Corrosion by Impact Sound Based on Neural Network

Abstract

The purpose of this study is to examine the effect of mislabeled data in the training data on the judgment results for reinforcement corrosion by the impact sounds of a steel ball colliding based on a neural network. For this purpose, the impact sounds of RC specimens with different corrosion levels were recorded, and the effects of contaminating with data in which corrosion has progressed beyond the target corrosion level into the positive training data were examined. As a result, it was found that the true positive rate decreased as the contamination rate increased when mislabeled data in the judgement the corrosion level of 1% was included. In addition, in the judgement of the corrosion level of 3%, the true positive rate tends to reduce when mislabeled data is included, but it was clarified that it is less affected by contaminating with the mislabeled data than the judgement of the corrosion level of 1%.