The purpose of this study is to evaluate the influence of positive data contamination in training datasets on defect detection of surface-coated RC specimens using hitting sounds based on the Local Outlier Factor (LOF). Three learning approaches are investigated: supervised, unsupervised, and semi-supervised learning. In supervised learning, performance decreases as the contamination rate increases, but stable results can be obtained through appropriate parameter tuning. Unsupervised learning, which utilizes the distribution of LOF values, achieves comparable performance to supervised learning. Furthermore, semi-supervised learning, which reused data with low LOF values as training data, shows improved performance even under conditions with higher contamination rates and coating effects.
