A Dual-Domain Deep Convolutional Variational Autoencoder Framework for Unsupervised Structural Health Monitoring Using Vision-Based Vibration Analysis

This study introduces a novel unsupervised machine learning approach for structural health monitoring (SHM), employing a dual-domain deep convolutional variational autoencoder (DD-CVAE). Conventional SHM techniques often require expensive sensor networks, extensive manual calibration, and labeled datasets, limiting real-world scalability and real-time application. To overcome these limitations, the proposed DD-CVAE integrates customized reconstruction losses in both the time and frequency domains (using fast Fourier transform), capturing amplitude deviations and subtle spectral shifts indicative of structural anomalies. The vibration responses required for the analysis are obtained using vision-based vibration monitoring, specifically employing the optical-flow method to accurately determine structural displacements. The DD-CVAE utilizes one-dimensional convolutional encoding to generate a compact latent representation, regularized using Kullback–Leibler divergence to facilitate robust probabilistic anomaly detection. Validation through experimental datasets collected from a controlled steel-frame laboratory setup demonstrates the effectiveness of DD-CVAE in accurately reconstructing healthy signals while sensitively detecting anomalies through elevated reconstruction errors in damaged states. Results indicate significant advantages over traditional single-domain methodologies, confirming DD-CVAE’s potential for scalable, automated, and efficient structural integrity assessments.