PSI - Issue 84

Prajwal Giri et al. / Procedia Structural Integrity 84 (2026) 1119–1126

1120

Keywords: Damage sensitive features; Ensemble learning; Finite element model; Prestressed reinforced bridges; Structural health monitoring.

1. Introduction Prestressed reinforced concrete bridges constitute a critical component of modern transportation infrastructure; however, their structural performance strongly depends on the integrity of the prestressing system and the sti ff ness of key load-bearing components. Deterioration mechanisms such as prestress loss and localized sti ff ness degradation (Sconocchia et al., 2024) typically evolve gradually and induce only subtle variations in the essentially linear global structural response, making early-stage damage particularly di ffi cult to detect using conventional inspection techniques. This fundamental challenge has motivated the development of advanced structural health monitoring (SHM) frameworks capable of extracting reliable damage-sensitive information from measured responses for timely and accurate condition assessment. Recent advances in machine learning (ML) have significantly improved the analysis of structural response data, enabling more reliable damage classification and condition assessment of prestressed concrete bridges. Among these approaches, ensemble learning (EL) has emerged as a powerful strategy because it integrates multiple models to enhance robustness, predictive accuracy, and generalization capability (Zhou, 2012). Bagging-based EL algorithms, such as Random Forest (RF) and Extra Trees, train individual models independently on random subsets of the data and combine their outputs through voting or averaging. By contrast, boosting-based methods, including AdaBoost, Gradient Boosted Decision Trees (GBDT), XGBoost, and LightGBM, construct models sequentially, with each learner focusing on previously misclassified samples to improve performance. In addition, classifiers such as K Nearest Neighbors (KNN), Support Vector Machines (SVM), and Multilayer Perceptrons (MLP) can also be incorporated as base models within bagging-based ensemble frameworks (Giri et al., 2025). EL has been widely applied in SHM for damage detection and condition assessment. Zhou et al. (2013, 2014) pioneered RF-based approaches for multi-damage detection and feature elimination, highlighting the value of ensemble strategies for improving classification performance and efficiency. Building on this, Lu et al. (2020) demonstrated that SVR and RF provided superior performance in predicting the natural frequencies of the Tamar Bridge under varying environmental conditions. Li and Song (2022) expanded the scope of EL by proposing a six model framework comprising RF, Extra Trees, AdaBoost, GBDT, XGBoost, and LightGBM for steel bridge deck defect prediction, and reported that XGBoost achieved the highest accuracy (94.9%), with superstructure condition identified as the dominant predictor. For seismic damage classification, Mangalathu et al. (2019) found RF to be the most effective model. Likewise, Yaghoubzadehfard et al. (2024) combined RF with interferometric radar measurements and finite element modelling (FEM) for reliable damage localization and severity estimation. More recently, Gautam et al. (2024) applied soft-voting ensembles to post-earthquake bridge damage classification, where RF and MLP achieved the best predictive performance. Despite these advances, the application of EL techniques to structural health assessment and damage classification in prestressed reinforced concrete bridges remains limited. Moreover, most existing studies have focused on a narrow set of feature types or relatively simple binary classification problems. In such settings, relying on a single classifier can introduce model specific bias and may limit the ability to capture the underlying complexity of structural response patterns. To address these gaps, this study proposes a weighted ensemble framework that integrates four complementary base learners, namely MLP, KNN, SVM, and RF, to improve robustness, classification accuracy, and generalization capability. The framework is validated on a full-scale prestressed bridge deck using distinct simulated damage states generated from a calibrated numerical model. 2. Proposed Ensemble Learning Framework for Damage Classification The proposed EL framework (Figure 1) integrates four complementary base learners, MLP, KNN, SVM, and RF, to improve robustness, accuracy, and generalization in multi-class damage classification of prestressed concrete bridges. All base learners are trained independently on the same dataset to preserve diversity in their decision boundaries and exploit complementary modeling capabilities. After training, each base learner produces probabilistic