PSI - Issue 78

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

392

Fig. 5. Label classifier rappresentation with all features

3.3. Damage classifier

To apply a DANN to structural health monitoring, a damage classifier was developed based on strain variations measured in the masonry panel. These variations, denoted as ∆ ε , were calculated with respect to the undamaged con figuration and used as input features. The variation between the initial (healthy) and final (damaged) states represents the key indicator for classifying the damage level, as it directly reflects how the structure has responded to seismic action. To account for realistic environmental influences such as temperature and humidity, a 10% Gaussian noise was introduced to the dataset A.Meoni et al. (2025) resulting in a total of 200 samples. All input features were normalized within the [0,1] range to eliminate scale-related biases during training. The classifier was designed to recognize four distinct structural conditions: a healthy, undamaged state (h), flexural damage (d1), shear damage (d2), and compres sive damage (d3) (see Fig. 3). The classification model demonstrated strong performance in distinguishing between these damage scenarios, validating its role in supporting the DANN training process. The label classifier, as constructed, was able to correctly identify the four classes. This can be observed both in the feature space—where di ff erent colors represent di ff erent health conditions of the panel—and in the confusion matrix (see Fig. 5). Specifically, the classifier was trained using 70% of the data, while the remaining 30% was used for testing. As an extension of the proposed methodology, a full-scale finite element model of a masonry facade was devel oped to evaluate the generalization capabilities of the DANN neural network across structurally di ff erent yet physically consistent domains. The geometry of the facade is based on a real-scale masonry construction built just outside the Department of Civil, Chemical, Environmental and Materials Engineering (DICAM) at the University of Bologna. Designed for long-term structural health monitoring with Smart Brick sensors, this test structure is continuously ex posed to real environmental conditions, providing an ideal reference for physically grounded numerical modeling. The facade measures 4725 × 4670 mm and shares the same Gothic bond pattern, material properties, and smart brick sen sor configuration as the panel. However, it di ff ers in overall geometry and loading conditions: while the panel used for training was loaded via imposed top displacements, the facade experiences seismic loading through distributed inertial forces along its height. In the finite element model developed in Abaqus, the seismic action was simulated by applying a horizontal gravity-type load with a vertically varying distribution along the height of the structure. Specifically, the load magnitude was set to 0.3g at the base, increasing linearly to 1.2g at the top. This variation aims to realistically reproduce the inertia forces induced by seismic acceleration. To balance modeling fidelity and computational cost, a preliminar hybrid simulation strategy was adopted. The sensor-instrumented region was micromodeled with detailed representation of bricks and mortar, using the same setup as the panel. In the rest of the structure, a macromodel- 3.4. Exploratory application: hybrid modelling of masonry facade