Speaker
Description
Variability in environmental and operational conditions significantly affects the structural responses of bridges, making accurate assessment of these influences essential for effective structural health monitoring (SHM). In many cases, it is impractical or impossible to capture all relevant environmental and operational variables, limiting the effectiveness of standard supervised approaches to mitigate such effects. This study introduces a hybrid methodology that leverages the synergy between supervised and unsupervised learning to address both measured and unmeasured variability in bridge displacement responses extracted from synthetic aperture radar (SAR) satellite imagery. The proposed framework begins with a supervised regression-based artificial neural network (SR-ANN), which integrates available environmental data (e.g., recorded temperature) with SAR-extracted displacements. The SR-ANN’s performance is evaluated using the R-squared (R²) metric between original and predicted responses. If a high R² value (close to 1) is achieved, indicating a strong correlation with the measured temperature, the process concludes. Otherwise, an unsupervised reconstruction-based neural network (UR-ANN) is employed using only response data to reconstruct the displacements. Environmental and operational effects are mitigated by computing the residuals between original and recon-structed responses, serving as normalized outputs. This approach is validated using data from long-span bridges. Results confirm the effectiveness and practical value of the proposed technique in reducing the impact of both observed and latent variability.
