基于深度学习的航空发动机涡轮叶片自动射线检测技术研究

Automatic radiographic testing for aeroengine turbine blades based on deep learning

关键词:  航空发动机  涡轮叶片  深度学习  缺陷检测  射线检测  射线图像

Key words:  Aeroengine  Turbine blade  Deep learning  Defect detection  Radiographic testing  X-ray images

摘要:

一直以来，航空发动机涡轮叶片的射线检测依靠检验员人工评片。为避免经验差异、眼睛疲劳、标准理解等人为因素影响，有效改善传统射线检测费时费力、效率低下等问题，针对航空发动机涡轮叶片射线图像，基于YOLOv4模型提出了一种双主干特征融合的缺陷自动检测算法（DBFF-YOLOv4）；通过设计包含所有特征映射的新型连接结构搭建缺陷检测颈部网络，建立了适用于涡轮叶片射线图像的缺陷自动检测模型；针对每个缺陷，采用9次裁剪、旋转和亮度增减的图像数据增强方法扩充样本数据，在此基础上进行了模型训练与测试。结果表明，针对完整涡轮叶片，建立的缺陷检测模型在0.5的置信度阈值下可获得96.7%的平均查准率和91.87%的平均查全率，优于通用目标检测算法YOLOv4模型。9次缺陷裁剪、旋转和亮度增减的图像数据增强方法能够显著提高模型的缺陷检测精度（平均精度分别得到了59.19%和2.53%的提升）。该研究为涡轮叶片自动射线检测提供了一种新方法。

Abstract:

Radiographic testing for aeroengine turbine blades usually depends on artificial detection. To avoid the influence of various artificial factors such as experience difference， eye fatigue and standard understanding， and to solve the problem of high cost， time consuming and low efficiency， a defect detection algorithm named DBFF-YOLOv4 was proposed for aeroengine turbine blade X-ray images by employing two backbones to extract hierarchical defect features based on YOLOv4. A novel concatenation form containing all feature maps was designed as the neck of defect detection framework. An automatic defect detection model for turbine blade X-ray images was established. Nine cropping cycles for one defect， flipping， brightness increasing and decreasing were applied for expansion of training samples and data augmentation. Finally， an automatic defect detection model was trained and test based on these defect samples. The results show that the defect detection model， which obtained 96.7% average precision and 91.87% average recall within the score threshold of 0.5 for complete turbine blade， outperformed others built by using the common object detection algorithm YOLOv4 directly. In addition， cropping nine times and data augmentation methods can significantly improve the defect detection accuracy of the model （mean average precision increased by 59.19% and 2.53% respectively）. This study provides a new method of automatic radiographic testing for turbine blades.