Mitigating class imbalance in deep learning-based multi-class structural damage recognition using an informatics-oriented data augmentation framework

초록

Deep learning has advanced automated structural damage recognition; however, real-world datasets remain severely imbalanced, with cracks dominating while critical defects such as spalling, efflorescence, and leakage are underrepresented. This scarcity leads to model bias and significantly reduces generalization and reliability. To address this issue, this paper proposes an informatics-oriented data augmentation framework that employs high-fidelity virtual environments to systematically balance multi-class structural damage data. The framework integrates class-targeted balancing strategies with customized multilabel annotation to control the generation and distribution of augmented samples, ensuring balanced representation and eliminating labor-intensive manual labeling. Experimental evaluation demonstrates substantial performance improvements for minority classes: achieving F1-scores up to 92.95% for leakage and 87.16% for spalling in segmentation tasks, representing relative improvements of 26.72% and 15.56%, respectively, over the real-only baseline, while maintaining stable accuracy for majority classes. Comparative results further show that algorithmic balancing methods such as Focal Loss struggle under severe imbalance, confirming that informatics-driven data augmentation provides a reliable, scalable, and methodologically essential foundation for robust deep learning in structural health monitoring and smart infrastructure inspection.

키워드