Dynamic response prediction of heritage masonry buildings by dual-phase hybrid regression modeling under partial monitoring parameters

Effective dynamic monitoring of heritage masonry buildings depends on reliable data from multi-sensor monitoring systems. Machine learning-based response prediction offers an intelligent solution to practical limitations of in-situ measurements. However, existing predictive models struggle to make reliable predictions in the presence of incomplete data. This study proposes a novel dual-phase residual-augmented regression (DPRAR) method for predicting long-term modal frequencies of masonry buildings by integrating random forests (RF) with a deep regression-based neural network (DRNN). Initially, the RF predicts responses from measured data to extract residuals between observed and predicted values, which serve as latent information. Subsequently, these residuals, combined with measured data, form an enhanced dataset to train the DRNN for final predictions. The main contributions include integrating statistical and deep learning regressors and innovatively using residuals to address missing unmeasured factors. Validation on a heritage masonry building shows that DPRAR substantially improves dynamic behavior prediction despite limited environmental measurements. • Predicting modal frequencies of heritage masonry structures with limited environmental/operational data. • Proposing a dual-phase regression method via advanced machine learning algorithms. • Integrating statistical and deep regressors in a residual-augmented hybrid algorithm. • Employing residuals of random forests to reveal unmeasured environmental/operational factors. • Validating the proposed method on a real-world heritage masonry building.