Explainable Boosting Machine for Structural Health Assessment: An Interpretable Approach to Data-Driven Structural Assessment

Abstract

Machine learning models used in structural health monitoring often act as "black boxes," offering predictions without justifying their logic. This lack of transparency undermines trust in safety-critical infrastructure assessments. To solve this, we propose the Explainable Boosting Machine, an interpretable method that explicitly links input variables (e.g., sensor data, and structural parameters) to predictions, enabling engineers to validate results against engineering principles. Real-world structural health monitoring and assessment struggles with sparse data, structural complexity, and hidden biases. Explainable Boosting Machine addresses these challenges by prioritizing transparency and physically meaningful insights. We apply it to predict the shear load- carrying capacity as a percentage of the ultimate load, based on the maximum diagonal crack widths observed on the surface of reinforced concrete beams—a critical metric for shear failure risk. Our results show that the model achieves an RMSE of 10.40% on the test dataset while identifying the influence of key predictors (e.g., beam depth, shear and skin reinforcement ratios). For instance, the model reveals that, for the same maximum diagonal crack width observed in two beams, a structure with a larger depth is farther from failure compared to the one with a smaller depth, enabling engineers to audit model logic and enhance structural assessment. This work advances trustworthy AI in structural health monitoring by bridging data-driven innovation and engineering accountability. Interpretability of explainable boosting machine ensures models remain consistent with physical laws, actionable for decision-making, and adaptable to realworld constraints. We advocate for machine learning frameworks that prioritize transparency as rigorously as predictive performance.