Automatic Training Data Selection for Autoencoder-Based Acoustic Defect Detection Robust against Class-Imbalance

This study proposes an automatic training data selection method for Autoencoder-based defect detection in hammering inspection, designed to address the severe class imbalance between normal and defective sounds. The proposed method employs a physically grounded indicator, Acoustic Energy per Impact, to automatically select and collect only the sound data considered normal. An Autoencoder is then trained exclusively on the collected normal sounds to identify defects based on reconstruction errors. To evaluate the effectiveness of the proposed method, experiments were conducted using con- crete specimens with cracks. The results demonstrate that the proposed method achieves higher defect detection performance than a conventional approach, even under highly imbalanced class conditions.