Transfer learning framework for post-seismic anomaly detection in historic masonry structures under varying environmental conditions

Historical masonry structures present unique challenges for structural health monitoring due to their brittle behaviour, material heterogeneity, and extreme sensitivity to environmental variations, which can often mask earthquake-induced damage. This study proposes an unsupervised deep transfer learning framework based on a Feature-wise Linear Modulation (FiLM)-conditioned convolutional autoencoder designed for post-seismic diagnostics. The methodology utilises dominant singular values extracted via singular value decomposition as compact spectral features, which are explicitly coupled with temperature channels within a multivariate latent space. To address data scarcity, the framework transfers learned “healthy manifolds” from a data-rich source domain (the Consoli Palace) to a data-scarce target domain (the San Pietro Bell Tower). A grid-search optimisation strategy is employed to identify critical transferable layers, while a multi-objective training pipeline autotunes distinct loss terms, ensuring balanced learning across diverse operational states. Validation results from the 2016 Central Italy seismic sequence demonstrate that the transfer-learning model significantly outperforms a scratch-trained baseline, achieving higher balanced accuracy. The framework effectively distinguishes cumulative seismic damage from seasonal thermal noise, providing a robust and scalable tool for the rapid safety evaluation of architectural heritage.