2026 Volume 7 Issue 3 Pages 606-617
This study aims to establish a highly accurate non-destructive method for detecting internal cracks in concrete structures for infrastructure maintenance. In the proposed method, a concrete structure is subjected to forced vibration using an exciter, and internal cracks are identified by analyzing the frequency characteristics of the resulting elastic waves through a machine-learning-based approach. In recent years, methods that transform inspection images into bottleneck feature vectors using pre-trained convolutional neural networks and perform anomaly detection in the resulting feature space have demonstrated excellent performance in industrial visual inspection. Although this approach is expected to be effective for elastic-wave-based inspection, an effective method has not yet been established for transforming elastic-wave frequency characteristics into feature vectors. Therefore, this study proposes a method for transforming elastic-wave frequency characteristics into feature vectors and investigates the effectiveness of feature-space-based anomaly detection. The experimental results showed high detection accuracy, confirming the effectiveness of the proposed method.