A generative framework for structural crack discovery under limited sensing

Characterizing internal cracks in structures using sparse sensor data remains a significant challenge due to limited spatial resolution and incomplete field coverage. In this study, we propose SparseGen-Damage (SG-Damage), a novel generative AI-based framework that reconstructs full-field strain information and identifies internal cracks using only sparse sensing data. The framework adopts a perceive–reconstruct–identify paradigm, integrating a Long Short-Term Memory (LSTM) network for handling variable-length sparse inputs, a Conditional Generative Adversarial Network (CGAN) for generating high-resolution strain fields, and a Feature Pyramid Network (FPN) for multi-scale crack localization and quantification. Experimental validation on aluminum alloy plates with internal straight cracks shows that SparseGen-Damage can accurately estimate crack length, orientation, and position using a limited number of strain sensors. The method achieves crack length prediction errors under 4 % and orientation errors below 2°, demonstrating both robustness and precision. This work presents a novel approach to sparse sensing in structural health monitoring (SHM), offering fast inference, reduced sensor deployment, and strong extensibility to diverse damage types, with promising potential for cost-effective, high-resolution, and real-time structural diagnostics.