Explainable AI Models for Interpretable SHM in Aging Infrastructure

Structural Health Monitoring (SHM) plays a vital role in ensuring the safety and longevity of aging infrastructure. However, many AI-based SHM systems remain black-box in nature, limiting trust and adoption in critical applications. This study presents a comparative investigation of three explainable artificial intelligence (XAI) models—SHAP-enhanced XGBoost, LIME-integrated Random Forest, and an attention-based bidirectional LSTM network—for interpretable damage detection and prognosis in deteriorating civil structures. Each model is evaluated on a curated dataset representing vibration, strain, and acoustic emission signals from aging bridges and industrial steel frames. The results demonstrate that while all three models achieve high classification accuracy, their explainability varies significantly in terms of feature attribution, temporal interpretability, and computational cost. The SHAP-based XGBoost model offers robust global and local interpretability with moderate computational demands. LIME provides intuitive, instance-level explanations but suffers from scalability issues. The attention-based deep learning model, though computationally intensive, excels in capturing temporal dependencies and highlighting critical sequence patterns. The findings underscore the necessity of selecting XAI models not only based on predictive accuracy but also on interpretability tailored to engineering decision-making. This work contributes to the growing field of trustworthy SHM by advancing interpretable AI techniques suitable for deployment in safety-critical infrastructure monitoring.

کلیدواژه ها:

Explainable Artificial Intelligence (XAI) ، Structural Health Monitoring (SHM) ، Aging Infrastructure ، SHAP ، LIME ، Attention Mechanism ، Interpretable Machine Learning ، Damage Detection ، Time-Series Analysis ، Civil Engineering Applications