U-Net driven semantic segmentation for detection and quantification of cracks on gas turbine blade tips

• Introduces a U-Net-based semantic segmentation model for automated crack detection in gas turbine blade tips, addressing the limitations of manual inspection methods. • To address the limited availability of cracked images, whole turbine blade tip images were divided into 128 × 128-pixel patches and trained using data augmentation, enhancing the model’s robustness and generalization. • Achieves approximately 85 pixels, demonstrating strong potential for real world application in aviation maintenance. • Automated Crack Detection: Introduces a U-Net-based semantic segmentation model for automated crack detection in gas turbine blade tips, addressing the limitations of manual inspection methods. • High Performance Metrics: Achieves approximately 85% recall and precision in detecting cracked pixels, demonstrating strong potential for real-world application in aviation maintenance. Crack detection and quantification on gas turbine blades is crucial for component validation during the development phase and for operational efficiency in service, as unexpected cracks can compromise blade integrity and lead to early engine removals. Gas turbine blades operate under extreme thermal and mechanical stresses, making them particularly susceptible to crack formation. At the same time deterministic predictions of crack formation are subject to high uncertainty in material data and actual loading conditions. Accurate detection and quantification of cracks, therefore, is essential for the validation and calibration of life predictions in order to prevent in-service failures, to extend component lifespan, and to reduce maintenance costs. This study introduces a U-Net based semantic segmentation model designed to automate crack detection on turbine blade tips. The model was trained on a dataset of 210 surface images with and without evidence of cracks, each divided into 128 × 128 pixel patches. Data augmentation techniques were applied to address the class imbalance between cracked and non-cracked pixels. The U-Net architecture, optimized with a Dice loss function, achieved a validation IoU of 0.7557, along with approximately 85% recall and precision in identifying cracked pixels. The pixel-based accuracy of the model primarily affects the quantification of cracks rather than their identification. A sliding window pipeline was implemented to extend the model’s applicability, enabling segmentation of entire blade tip images for comprehensive crack localization. While the model may occasionally miss low-contrast cracks, it holds potential as a supplementary tool for manual inspection as part of the life prediction validation. By providing automated crack localization and quantification, the model can assist in analyzing crack characteristics relative to engine operating conditions.