PSI - Issue 84

Vincenzo Mario Di Mucci et al. / Procedia Structural Integrity 84 (2026) 521–528

523

variability. All images were annotated by structural engineering experts according to visual features associated with each severity class: superficial orange-brown rust for Low, reddish-brown rust flakes for Medium, and dark brown coarse corrosion for High, aligned with the definitions provided in the Italian technical guidelines. Preprocessing steps were implemented to improve the CNN’s sensitivity to corrosion-specific patterns. Images were transformed into alternative color spaces (HSV, Lab) to emphasize rust-related chromatic features, followed by patch extraction using a 224×224 sliding window. Data augmentation techniques, including rotations and flips, were applied to balance class representation and increase robustness to varying visual conditions. The framework is based on a customized CNN architecture (BriCANet) enhanced with channel attention modules, which emphasize corrosion-related features (e.g., rust texture and surface discontinuities) while minimizing the influence of background noise. As demonstrated in the previous study by Di Mucci et al. (2025a), BriCANet was tested on a dataset split into 70% training and 30% testing, showing its effectiveness in classifying corrosion severity in images of exposed bars in RC members. The model outperformed standard CNN architectures, such as ResNet50 and InceptionV3 (Fig. 1).

Fig. 1. Comparison of confusion matrices (from left to right): BriCANet, InceptionV3, and ResNet50.

2.2. Mapping visual severity classes to probabilistic Mass Loss parameters To link image‐based corrosion classification to structural performance assessment, the predicted severity classes are converted into probabilistic reinforcement mass–loss parameters, Q corr . In accordance with CONTECVET guidelines (Fagerlund, 2001), the initiation of cracking generally corresponds to Q corr >5% , whereas concrete spalling typically develops beyond Q corr >10% . Because BriCANet works on reinforcement already exposed, the proposed framework adopts a minimum threshold of Q corr >10% , reflecting a post–spalling condition indicative of visible corrosion. Severity classes are subsequently parameterized following the Illinois Bridge Inspection Standards (Illinois Department of Transportation, 2015), defining the corresponding mass–loss intervals as: Low (10–20%), Medium (20–30%), and High (30–50%). Each class is modelled as a uniform probability distribution across its respective interval, allowing the degradation state to be incorporated stochastically within the structural analysis.

3. Structural modelling and fragility analysis 3.1. Corrosion–induced material degradation

The mechanical property variability of the bridge pier was incorporated through a probabilistic sampling strategy. Gaussian distributions derived from the study by Zelaschi et al. (2016) were adopted for the concrete compressive strength ( f c ), steel yield strength ( f y ), and elastic modulus ( E s ). The shear modulus of the neoprene bearings ( G neo ) was modeled as a random variable following the ranges reported by Cardone (2014). Corrosion effects were explicitly incorporated into the mechanical properties of the reinforcing steel. The reduction of the effective bar area and the associated degradation in yield strength and ultimate strain were computed following the