Fluorescent penetrant inspection system for minor defects in jet-engine blade surface

Fluorescent penetrant inspection (FPI) represents a novel approach that efficiently inspects minor defects on the blade surface. However, it presents several challenges, including being sensitive to the environment, relying heavily on experiential knowledge, difficulty quantifying defect characteristics, and a lack of traceability of results. This study proposes an artificial intelligence (AI)-based detection framework to detect minor defects on jet-engine blade surfaces through fluorescent penetrant. A FPI system and a high-precision automation control method are first constructed, and a comprehensive database of typical surface defects found in the blade manufacturing process is obtained through the detection system. The feature extraction and fusion structure of the You Only Look Once (YOLO) v8 algorithm is then redesigned for enhancing its multi-scale feature detection capability, enabling efficient detection of blade surface defects. Rather than using the traditional Task-Aligned Assigner (TAA) strategy for minor defects, a Normalized Gaussian Wasserstein Distance (NWD) label assignment is applied to enhance bounding box similarity measurement. Ablation and comparative experiments are finally conducted for validation. The experimental outcomes highlight the effectiveness of the intelligent defect detection system in capturing and detecting minor defects on blade surfaces, utilizing the proposed YOLOv8-RN. Specifically, the mean average precision (mAP) of YOLOv8-RN across all categories reaches 0.933 at an intersection over union (IoU) of 0.5 for four typical defects: crack, cold lap, inclusion, and cavity. Moreover, YOLOv8-RN demonstrates enhancements of 11.7% in mAP, 13.9% in F1-score, and 8% in frames per second (FPS) in contrast to the original YOLOv8. When benchmarked against other mainstream algorithms, YOLOv8-RN achieves a relatively fast speed and the highest accuracy in the same processing speed range. It also shows a 23.9% and 14.3% increase in recall values compared to YOLOv10 and YOLOv11, respectively, making it more suitable for jet-engine blade fluorescent defect detection.