TY - JOUR
T1 - Low Cost and Highly Sensitive Automated Surface Defects Identification Method of Precision Castings Using Deep Learning
AU - Yu, Huipeng
AU - Kang, Maodong
AU - Ding, Chenyang
AU - Liu, Yahui
AU - Gao, Haiyan
AU - Wang, Jun
N1 - Publisher Copyright:
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.
PY - 2024/12
Y1 - 2024/12
N2 - The surface of superalloy precision castings might exhibit defects after forming, posing a significant risk to their service life, necessitating inspection during post-process. Radiographic inspection, with its extensive research in automation, can achieve efficient and accurate detection of defects. However, it is limited in surface defects detection due to limited sensitivity to non-volumetric defects and high cost. In contrast, fluorescent penetrant inspection (FPI) is highly efficient for surface defect inspection due to its low cost, high sensitivity, and speed. However, manual examination introduces variability in the results, impacting the consistency and reliability of the inspection process. Automation is needed to ensure consistency and reliability of inspection. The implementation of an automated defect identification system based on FPI using convolutional neural networks (CNNs) was systematically investigated. Among the CNN models tested, MobileNetV2 exhibited exceptional performance, achieving a remarkable recall rate of 0.992 and an accuracy of 0.992. Additionally, the effect of class imbalance on model performance was carefully examined. Furthermore, the features extracted by the model were visualized using Grad-CAM to reveal the attention of the CNN model to the fluorescent display features of defects. This study underscores the strong capability of deep learning architectures in identifying defects of precision casting components, paving the way for the automation of the entire FPI process.
AB - The surface of superalloy precision castings might exhibit defects after forming, posing a significant risk to their service life, necessitating inspection during post-process. Radiographic inspection, with its extensive research in automation, can achieve efficient and accurate detection of defects. However, it is limited in surface defects detection due to limited sensitivity to non-volumetric defects and high cost. In contrast, fluorescent penetrant inspection (FPI) is highly efficient for surface defect inspection due to its low cost, high sensitivity, and speed. However, manual examination introduces variability in the results, impacting the consistency and reliability of the inspection process. Automation is needed to ensure consistency and reliability of inspection. The implementation of an automated defect identification system based on FPI using convolutional neural networks (CNNs) was systematically investigated. Among the CNN models tested, MobileNetV2 exhibited exceptional performance, achieving a remarkable recall rate of 0.992 and an accuracy of 0.992. Additionally, the effect of class imbalance on model performance was carefully examined. Furthermore, the features extracted by the model were visualized using Grad-CAM to reveal the attention of the CNN model to the fluorescent display features of defects. This study underscores the strong capability of deep learning architectures in identifying defects of precision casting components, paving the way for the automation of the entire FPI process.
KW - Deep learning
KW - Fluorescent penetrant inspection
KW - Precision casting
KW - Surface defect
UR - http://www.scopus.com/inward/record.url?scp=85204779513&partnerID=8YFLogxK
U2 - 10.1007/s10921-024-01121-5
DO - 10.1007/s10921-024-01121-5
M3 - 文章
AN - SCOPUS:85204779513
SN - 0195-9298
VL - 43
JO - Journal of Nondestructive Evaluation
JF - Journal of Nondestructive Evaluation
IS - 4
M1 - 105
ER -