Explainable Artificial Intelligence for Infrastructure Condition Assessment: A Comprehensive Review

Abstract

This paper presents a comprehensive review of Explainable Artificial Intelligence (XAI) techniques applied to infrastructure condition assessment, with particular emphasis on their role in improving the interpretability, reliability, and practical applicability of machine learning (ML) models. The integration of XAI methods, including SHAP (SHapley Additive exPlanations), LIME (Local Interpretable Model-agnostic Explanations), and Grad-CAM (Gradient-weighted Class Activation Mapping), with advanced ML and deep learning models addresses the critical challenge of model interpretability in safety-critical infrastructure monitoring. The reviewed studies demonstrate that XAI-enhanced approaches can achieve detection accuracies of approximately 92–96% while simultaneously providing meaningful explanations of model predictions. These capabilities enable engineers, infrastructure managers, and regulatory authorities to understand the underlying factors influencing predictions and identify critical deterioration patterns more effectively. In comparison, conventional black-box ML models generally report accuracies in the range of 85–92% but provide limited insight into their decision-making processes. Beyond predictive performance, XAI contributes to increased stakeholder confidence, improved transparency, enhanced decision support, and greater accountability in infrastructure management. The findings further indicate that interpretable AI frameworks can support proactive maintenance, risk-based prioritization, and more informed resource allocation. Overall, integrating XAI with infrastructure condition assessment represents a promising pathway toward reliable, transparent, and practically deployable intelligent monitoring systems