PSI - Issue 78

Livio Pedone et al. / Procedia Structural Integrity 78 (2026) 1609–1616

1611

Fig. 1: Flowchart for vision-based residual capacity assessment of earthquake-damaged buildings (after Saquella et al. 2025).

Data collected in Step 1 are then processed using an ML-based algorithm ( Step 2 ). For the correct implementation of the framework, the ML-based tool should be able to provide, with good accuracy, information on: (i) the member typology (e.g., beams, columns, walls, non-structural components) and (ii) the level of the observed structural damage. The latter must also be consistent with the main classification criteria for earthquake-induced damage adopted at the national and international level. In this work, the observed damage severity is automatically evaluated from images using a CNN based on a VGG16 architecture, which has been trained on 5,000 RGB images. More details on this are given in the following Section 3. Moreover, in this phase, it is also crucial to evaluate the location of damaged structural components with respect to the structural skeleton. As an example, in the case of a reinforced concrete (RC) frame structure, the same level of damage can be more critical for a base column than for a column in the last story. In Step 3 , results provided by the CNN-based damage detection tool are used to update the expected nonlin ear response of damage components, according to state-of-the-art procedures available in literature and previously discussed. Di ff erent criteria can be adopted for the definition of these reduction factors without a ff ecting the ef fectiveness of the proposed methodology. Then, a subsequent nonlinear static (pushover) analysis is carried out to evaluate the structure’s force-displacement capacity curve in its damaged configuration. To this end, a simplified analytical / mechanical procedure can be adopted to allow for a rapid estimation of the seismic residual capacity of the structure without the need for more complex and time-consuming numerical (software-based) simulations (Pam panin, 2021). More specifically, this research proposes the use of the Simple Lateral Mechanism Analysis (SLaMA) method (NZSEE, 2017; Pampanin, 2017) and its recent extension for post-earthquake safety and loss assessment (Matteoni et al., 2023). A detailed description of the procedure can be found in the cited paper. Clearly, the global force-displacement capacity curve of the “damaged” structure is expected to show a reduction in terms of sti ff ness, strength, and ductility if compared to the “intact” (i.e., as-built) configuration. Finally, in Step 4 , the results of the pushover analysis are used to perform seismic response analysis through state of-the-art spectrum-based approaches. Moreover, safety and economic loss assessment can be carried out through simplified pushover-based procedures (Cosenza et al., 2018). The output of the framework can be used to support decision-making in post-earthquake scenarios, especially for re-occupancy and repair vs. demolition.

3. ML-based tool for structural damage level classification

The proposed framework for seismic residual capacity assessment relies on a CNN-based tool to classify visible damage to structural and non-structural components. The involved CNN model has been selected as the best perform ing among twelve di ff erent and well-known CNN models implemented by the author in recent research (Saquella et al., 2025), as a part of a wider PNRR (National Recovery and Resilience Plan) - National Research Centre (CN1) research project (Pampanin et al., 2025). The adopted methodology for training and testing the CNN model, as well as some relevant results, are briefly summarized in this section. The reader is referred to Saquella et al. (2025) for more information.