A vision-based unsupervised framework for structural damage localization in masonry minarets using slope-informed convolutional autoencoders

This study proposes a hybrid unsupervised framework for structural damage detection and localization by combining a Convolutional Autoencoder (CAE) and Slope-Based Autoencoder (SBAE) models. The approach was experimentally validated on a scaled masonry minaret model subjected to progressive damage scenarios. Structural response data were acquired using a vision-based monitoring system consisting of high-resolution video cameras and QR marker affixed to the structure. Each QR marker served as a virtual sensor, and its displacement over time was extracted using computer vision techniques. Displacement signals from four such virtual sensors were analyzed to capture the dynamic behavior of the structure. The CAE, trained solely on healthy-state data, identified anomalies through Mahalanobis distance applied to reconstruction errors, yielding high sensitivity to structural deviations with minimal false positives. In parallel, relative displacement trends between adjacent sensors were quantified via slope features and modeled with lightweight AEs to achieve spatially resolved damage localization. Experimental results revealed that the first damage scenario was localized near the top of the minaret (S1-S2), while the second scenario indicated severe damage progression toward the midsection (S2-S3), consistent with observed physical cracks. The combined method offers robust generalization, spatial accuracy, and interpretability, supporting its practical applicability for real-world structural health monitoring of slender masonry structures.