Semantic-Conditioned Diffusion Augmentation and Lock-Point Targeted Deep Ensemble for Seismic Damage Recognition of RC Columns

Rapid and accurate identification of post-earthquake damage states in reinforced concrete (RC) columns is essential for structural safety assessment and emergency response. However, manual inspection is constrained by expert experience and site conditions, while image-based methods face challenges of data scarcity, ambiguous damage features, and background disturbances. To address these issues, this study proposes a method based on semantic-conditioned diffusion augmentation and deep ensemble learning. According to the seismic damage evolution mechanism, the damage states are classified into cracking, spalling, rebar exposure, and rebar buckling. A damage-semantics-oriented augmentation strategy is developed to generate enhanced samples with explicit failure characteristics, followed by a multi-scale feature extraction and ensemble discrimination framework for automatic classification. Using only 182 original images and low-dose targeted semantic augmentation, the proposed three-model ensemble improves the overall accuracy from 80.77% to 82.42% under stratified five-fold cross-validation, with class-specific accuracies of 91.11%, 78.26%, 83.33%, and 76.74% for cracking, spalling, rebar exposure, and rebar buckling, respectively. Further analyses using confusion matrix, Precision/Recall/F1-score, confidence distribution, and model complexity experiments demonstrate favorable class balance, stable recognition of complex damage states, and satisfactory engineering interpretability. The proposed framework offers a promising solution for intelligent post-earthquake inspection and rapid component-level damage identification, providing reliable, confidence-informed inputs to support risk-based emergency response and resource allocation decisions.