Keywords: Non-destructive Inspection

Aging concrete infrastructure such as bridges and tunnels requires effective inspection to ensure safety and durability, particularly for detecting invisible internal cracks subjected to structural integrity. Impact-echo, which is one of non-destructive testing methods, is widely used but costly and time-consuming with relying on skilled and experienced analysis. This study integrates AI with impact-echo data to improve crack detection. Supervised deep learning using FFT-transformed signals enables accurate classification of multiple crack levels, including intact condition of concrete. However, data labeling for each existing structures is impractical, so an unsupervised approach using an auto-encoder is proposed to identify internal crack levels through anomaly-based indices without labeled data.