Anomaly detection using novel unsupervised autoencoders for vibration-based monitoring of masonry minarets
Structures, cilt.90, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 90
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.istruc.2026.112440
- Dergi Adı: Structures
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus
- Anahtar Kelimeler: Anomaly detection, Autoencoder, Deep learning, Masonry minaret, Shake table testing, Structural health monitoring
- Karadeniz Teknik Üniversitesi Adresli: Evet
Özet
While deep learning-based autoencoders show promise for automated structural health monitoring (SHM) of masonry minarets, a systematic performance comparison of architectural designs is conspicuously absent from the literature. This study addresses this gap by systematically comparing six autoencoder architectures for vibration-based anomaly detection: Convolutional Autoencoder (ConvAE), Linear Bottleneck Autoencoder (AE), Strong Convolutional Autoencoder (StrongConvAE), Temporal Convolutional Network Autoencoder (TCNAE), Denoising Autoencoder (Denoise-AE), and Fast Residual Dilated Autoencoder (FastResDilatedAE). The architectures were evaluated using data from shake table experiments on a scaled physical model. Signals from six accelerometers were fused to construct the dataset, which consists of 17,343 analysis windows with a severe 18.3:1 class imbalance. Results demonstrate that exceptional detection performance can be achieved through either global integration mechanisms or efficient multi-scale feature extraction. Critically, the findings also show that severe class imbalance can be considerably mitigated for the current dataset by models with sufficient discriminative capacity. This work makes three principal contributions: it presents the first comprehensive architectural comparison for masonry minarets; it quantitatively demonstrates the impact of class imbalance on performance metrics; and it establishes evidence-based guidelines for practical architecture selection. For practical deployment, FastResDilatedAE is identified as the optimal choice, offering the best accuracy-efficiency trade-off. These contributions represent a significant advance toward reliable, automated SHM systems for cultural heritage preservation.