PSI - Issue 78

Alina Elena Eva et al. / Procedia Structural Integrity 78 (2026) 387–394

388

normal conditions by detecting anomalies in sensor data that may indicate changes in structural integrity. In recent years, research in SHM has introduced innovative approaches such as smart masonry, where structural elements also function as distributed sensors[F. Di Trapani et al. (2024); A.Meoni et al. (2021)]. This concept includes the use of damage-sensitive mortar and strain-sensitive bricks. This work focuses on smart bricks, which are piezoresistive clay bricks that generate electrical signals in response to mechanical strain[A.Meoni et al. (2020)]. When integrated into masonry, smart bricks can detect strain variations caused by crack formation or environmental changes, providing real-time insights into the structural condition. To make the most of smart bricks, we propose using Machine Learning (ML) and Artificial Intelligence (AI) techniques. These tools enable the development of intelligent damage classi fiers that can handle environmental variability and generalize across di ff erent structures. One of the main challenges in ML-based SHM is the limited availability of labeled data representing damaged conditions in real buildings. To address this, we use Finite Element Models (FEMs) to simulate di ff erent structural scenarios—both undamaged and damaged—allowing us to create a synthetic, labeled dataset. The classifier trained on simulated data acts as the source domain model. However, applying this model to real-world cases introduces di ff erences due to material variability, environmental factors, and construction uncertainties. To bridge this gap, we use Domain-Adversarial Neural Net works (DANN) [I.A. Ozdagli et al. (2021); V. Giglioni et al. (2025)], which help adapt the classifier to real-world data. This improves the model’s ability to detect structural anomalies even without labeled damage data from the monitored structure. This methodology represents an important step forward in SHM, combining the accuracy of numerical simulations with the adaptability of AI techniques to provide more e ff ective and reliable monitoring of masonry structures. Section 2 presents the proposed methodology and details the structure of the DANN network. Section 3 illustrates the case study, including the representative FEM model with various simulated damage scenarios and the development of the classifier. Additionally, an exploratory application is introduced to evaluate the modelling of more complex struc tures. This step aims to assess the potential of extracting meaningful information from smart bricks while significantly reducing computational e ff ort. This work proposes a methodology for monitoring masonry structures by integrating smart masonry and deep learning algorithms. One of the main challenges in this field is the limited knowledge of damage characteristics. To address this issue, a classifier has been developed to distinguish di ff erent damage states using labeled data from both healthy and damaged structures through supervised training. This classifier is capable of predicting the health status of masonry elements based on the analysis of strain variations measured by smart bricks. In the current phase of the research, the objective is to investigate how knowledge acquired from a numerical model of a masonry panel — rep resenting a simplified, controlled structural element — can be transferred to a di ff erent numerical model representing a panel embedded in a full masonry fac¸ade. Both models are used to simulate structural behavior and to generate data for training and testing SHM algorithms based on smart brick sensors. The use DANN is being explored as a possible approach to facilitate this knowledge transfer. DANNs are designed to train classifiers on labeled data from a source domain (the isolated panel) while promoting generalization to a target domain (the fac¸ade-integrated panel) by learn ing domain-invariant features. This investigation aims to assess whether such a method can e ff ectively bridge the gap between di ff erent numerical representations of masonry systems, in preparation for future applications to real struc tures. The objective of this work is therefore to apply domain adaptation through DANN to enable reliable damage classification on real masonry fac¸ades, using knowledge learned from simulated data on masonry panels. 2. Proposed Methodology 2.1. Problem definition

2.2. Domain Adversarial Neural Network

Given two domains, the source domain (Ds) and the target domain (Dt), where the feature spaces (Xs and Xt) and label spaces (Ys and Yt) are respectively assigned, domain adaptation is a process aimed at improving the target predictive function using the knowledge available in the source domain. The assumption is that the feature and label spaces are the same (Xs = Xt; Ys = Yt), but the probability distributions are di ff erent( p(X) and p(Y) are di ff erent;