Using machine learning for crack detection in beams under limited experimental data: The role of sliding window augmentation

Abstract

In this study, the crack detection problem in steel cantilever beams with rectangular cross-section has been addressed under extremely limited experimental data conditions. Starting with only 16 experimental configurations (4 beam systems × 4 crack depth levels: 0–3 mm), the effectiveness of Sliding Window-based data augmentation strategy has been systematically investigated. Under baseline conditions, classical machine learning models showed moderate performance: Random Forest achieved 62.5% accuracy in crack depth classification, while XGBoost obtained 25.0% accuracy. Deep learning models could not be trained at all with 16 samples. Sliding window application transformed 16 samples into 304 windows, making model training possible and improving performance dramatically. XGBoost achieved perfect crack depth classification (100% accuracy) and 0.0744 MAE in dimensionless crack location prediction. LSTM became the best deep learning model with 0.1202 MAE in location prediction. Statistical analysis confirmed substantial improvements with large effect sizes (Cohen’s d up to 1.73 for depth classification). The study shows that high-accuracy crack detection is possible with limited experimental data through appropriate data augmentation strategies, proposing a cost-reducing solution for structural health monitoring applications where data collection is expensive and time-consuming.