PSI - Issue 78

Francesco Mariani et al. / Procedia Structural Integrity 78 (2026) 875–882

878

Fig. 2. View of the case study bridge

superstructure is supported by cast-in-place reinforced concrete piers and abutments, with heights ranging from 7.5 m to 18.5 m, following the natural slope. The bridge was constructed using the balanced cantilever method, with conti nuity ensured by four vertically prestressed internal joints located at spans 3, 5, 7, and 9. These joints are designed to accommodate horizontal thermal displacements while maintaining structural integrity between adjacent cantilevers. Each joint includes two sets of sliding supports, an upper and a lower one. Additionally, 30 prestressed Dywidag bars are installed at the lower supports to maintain contact under positive bending moments. Visual inspections revealed significant vertical deflections, particularly in the spans incorporating the internal joints. These deflections resulted in noticeable cusps, interrupting the longitudinal continuity of the superstructure. In contrast, the spans without joints ex hibited upward displacements. This abnormal behavior prompted the development of a calibrated finite element (FE) model, validated using both static and dynamic experimental data. The FE model serves two main purposes: (i) to investigate the root causes of the observed deformations, and (ii) to support the development of predictive algorithms for early damage detection. A total of 62 measurement points, representing a simulated sensor network, are strategically distributed to capture both dynamic (e.g., modal frequencies) and static (e.g., rotations) structural responses. Damage scenarios are dis cretized to reflect plausible post-earthquake structural conditions, with 11 potential damage locations along the girder near the piers and 9 additional locations at the pier bases. Each segment is assigned independent material properties to simulate variations in sti ff ness. The concrete elastic modulus, selected as the damage-sensitive parameter, is modeled in each segment using a log-normal distribution with a coe ffi cient of variation of 0.20, centered around a design value of 32,860 MPa for the girder concrete and 31,246 MPa for the pier concrete. To explore a broad range of degrada tion scenarios, a dataset of approximately 2,500 sti ff ness configurations is generated using Latin Hypercube Sampling (LHS). For each configuration, structural analyses are performed to simulate the corresponding sensor responses. This dataset is then used to train the BNN, which learns to predict the elastic modulus of each girder segment based on sensor data. The trained BNN enables both the localization and quantification of structural damage by identifying reductions in sti ff ness with probabilistic confidence. 4. Finite Element Model: damage representation and key sensitive parameters