Intelligent detection method for debonding and voids in concrete-filled steel/aluminum tubular structures based on impact acoustics and unsupervised learning

Abstract

Debonding and voids between concrete-filled steel/aluminum tubes and the internal concrete are recognized as critical defects that can significantly compromise structural integrity, load-bearing capacity, and service life. Impact-acoustics-based methods offer operational simplicity and low cost, yet most current approaches rely on manual tapping, making them highly dependent on operator skill, poorly generalized, and low in accuracy and automation, which limits large-scale engineering application. To address these limitations, firstly, this study proposes an impact-acoustics autoencoder framework that leverages the reconstruction error of tapping sound spectrograms as a primary indicator for defect identification. Power spectral density peak frequency and wavelet packet energy ratio are used to automatically label normal data samples, converting a semi-supervised autoencoder into a fully unsupervised model. Secondly, an anomaly threshold determination method based on exceedance theory is developed to enhance automation. Furthermore, a channel self-attention mechanism is embedded in the convolutional autoencoder to strengthen key feature extraction, thereby improving detection accuracy and robustness. Thirdly, an automatic crawling and tapping robot is developed and validated on an actual bridge. Experimental results show that the proposed method significantly outperforms traditional manual techniques in accuracy and recall, achieves performance close to supervised approaches, detects void regions with over 96 % accuracy, and produces results highly consistent with ultrasonic phased array imaging. In addition, the method maintains similarly high recognition accuracy in concrete-filled aluminum tubes. This method demonstrates strong generalization, high precision, and adaptive capability, making it particularly suitable for integration with intelligent robotic platforms for fully automated detection in practical engineering projects.