基于深度学习的航空发动机内部损伤实时检测方法

何超; 陈果; 王雨薇

doi:10.13224/j.cnki.jasp.20210381

基于深度学习的航空发动机内部损伤实时检测方法

doi: 10.13224/j.cnki.jasp.20210381

1.
南京航空航天大学民航学院，南京 211106
2.
南京航空航天大学通用航空与飞行学院，南京 213300

基金项目: 国家科技重大专项（2017-Ⅳ-0008-0045，J2019-Ⅳ-004-0071）

详细信息

作者简介:
何超（1998-），男，硕士生，主要研究方向为深度学习。E-mail：agatsumazenitsu@nuaa.edu.cn

中图分类号: V263.6
计量
- 文章访问数: 909
- HTML浏览量: 698
- PDF量: 179
- 被引次数: 0
出版历程
- 收稿日期: 2021-07-20
- 网络出版日期: 2023-05-11

Real-time detection method of aero-engine internal damage based on deep learning

1.
College of Civil Aviation， Nanjing University of Aeronautics and Astronautics，Nanjing 211106，China
2.
College of General Aviation and Flight， Nanjing University of Aeronautics and Astronautics，Nanjing 213300，China

摘要

摘要:
针对航空发动机内部损伤实时检测问题，提出了基于YOLOv4框架下的目标识别网络模型，该网络模型具有检测精确度高，推理速度快的优点，实现了发动机内部损伤的实时检测。在具体实施过程中，该方法首先对不同损伤类型进行分类并对损伤位置进行标注，将图片与之对应的标注导入到改进网络中进行训练并得到对应的检测模型，最后，基于训练好的模型，对图片和视频流上的损伤进行实时检测。利用Pascal VOC（visual object classes）标准数据集与真实的航空发动机孔探图像数据集进行方法验证，结果表明所提出的目标识别网络在保证准确率的前提下每秒检测的帧率相比原目标识别网络提升了23.7%以上。为解决孔探损伤检测中人为因素导致的检测结果不准确与检测效率低下等问题提供了有效途径，具有很强的工程实用价值。
- 航空发动机损伤 /
- 孔探检测 /
- 深度学习 /
- 损伤检测 /
- YOLOv4网络模型
Abstract:
In view of the problem in real-time detection of aero-engine internal damage, an object detection network model based on the YOLOv4 framework was proposed. With advantages of high detection accuracy and fast reasoning speed, this network model realized real-time detection of engine internal damage. In specific implementation process, the method first classified different damage types and annotated the damage location, and then imported the image and its corresponding annotations into the improved network for training to obtain the corresponding detection model. Finally, based on the trained model, real-time detection of damage on pictures and video streams was performed. The Pascal VOC (visual object classes) standard data set and the real aero-engine borescope image data set were used to verify the method. The results showed that the proposed object detection network can improve the frame rate of detection per second more than 23.7% on the premise of ensuring accuracy compared with the original object detection network. The method could provide an effective way to solve the problems of inaccurate detection results and low detection efficiency caused by human factors in borescope damage detection, showing strong engineering practical value.
- aero-engine damage /
- borescope detection /
- deep learning /
- damage detection /
- YOLOv4 network model

HTML

图 1 航空发动机典型损伤图片

Figure 1. Typical damage picture of aero-engine

下载: 全尺寸图片幻灯片

图 2 航空发动机损伤检测模型训练流程

Figure 2. Aero-engine damage detection model training flowchart

下载: 全尺寸图片幻灯片

图 3 OSA模块

Figure 3. OSA module

下载: 全尺寸图片幻灯片

图 4 改进后的OSA模块

Figure 4. Improved OSA module

下载: 全尺寸图片幻灯片

图 5 YOLOv4-VoVNetv2-39网络结构图

Figure 5. YOLOv4-VoVNetv2-39 network structure diagram

下载: 全尺寸图片幻灯片

图 6 $ c $与$\,\rho （b,{b^{{\text{gt}}}}）$的示意图

Figure 6. Schematic diagram of $ c $ and $\,\rho （b,{b^{{\text{gt}}}}）$

下载: 全尺寸图片幻灯片

图 7 不同种类的平均精确度对比

Figure 7. Comparison of AP of different classes

下载: 全尺寸图片幻灯片

图 8 部分Pascal VOC测试集检测结果

Figure 8. Part of Pascal VOC test data set detection results

下载: 全尺寸图片幻灯片

图 9 部分发动机孔探损伤检测结果

Figure 9. Part of aero-engine borescope damage detection results

下载: 全尺寸图片幻灯片

表 1 VoVNet-39网络结构

Table 1. VoVNet-39 network structure

阶段	VoVNet-39	输出
初始阶段1	3×3卷积块, 64通道数, s=2 3×3卷积块, 64通道数, s=1 3×3卷积块, 128通道数, s=1	208×208×128
OSA模块阶段2	[3×3卷积块, 128通道数, ×5堆叠 1×1卷积块,256通道数]×1	104×104×256
OSA模块阶段3	[3×3卷积块, 160通道数, ×5堆叠1×1卷积块, 512通道数]×1	52×52×512
OSA模块阶段4	[3×3卷积块,192通道数, ×5堆叠1×1卷积块,768通道数]×2	26×26×768
OSA模块阶段5	[3×3卷积块, 224通道数, ×5堆叠 1×1卷积块,1024通道数]×2	13×13×1024

下载: 导出CSV

表 2 网络参数与模型大小比较

Table 2. Comparison of network parameters and model size

模型	参数量	模型大小/MB
YOLOv4网络	63, 312, 477	248.6
YOLOv4-VoVNetv2-39	59, 747, 965	230.9

下载: 导出CSV

表 3 两种模型的推理速度对比

Table 3. Comparison of FPS of the two models

模型	平均每张图片推理时间/s	FPS/帧
YOLOv4	0.02528	39.6
YOLOv4-VoVNetv2-39	0.02035	49.2

下载: 导出CSV

表 4 两种模型的检测平均精确度、速度对比

Table 4. Comparison of detection average precision and FPS of the two models

模型	A_p/%			m_AP/%	平均每张图片推理时间/s	FPS/帧
模型	掉块	裂纹	腐蚀	m_AP/%	平均每张图片推理时间/s	FPS/帧
YOLOv4	98.32	97.86	83.58	93.25	0.02534	39.5
YOLOv4-VoVNetv2-39	97.95	99.09	85.36	94.13	0.02049	48.8

下载: 导出CSV

参考文献(21)

[1]	易中辉,李世林,李子铭. 图像识别在航空发动机维修中的应用[J]. 中国科技信息,2020,32(23): 27-28. YI Zhonghui,LI Shilin,LI Zimin. The application of image recognition in aero-engine maintenance[J]. China Science and Technology Information,2020,32(23): 27-28. (in Chinese)
[2]	肖柏荣. 航空发动机维护中孔探检测技术的应用[J]. 中国高新科技,2019,3(15): 96-98. doi: 10.13535/j.cnki.10-1507/n.2019.15.19 XIAO Bairong. Application of borescope detection technology in aero-engine maintenance[J]. China High-Tech,2019,3(15): 96-98. (in Chinese) doi: 10.13535/j.cnki.10-1507/n.2019.15.19
[3]	王晓兵. 基于孔探检测的民航发动机维修间隔优化研究[D]. 天津: 中国民航大学, 2018. WANG Xiaobing. Research on civil aviation engine maintenance interval optimization based on borescope detection[D]. Tianjin: Civil Aviation University of China, 2018. (in Chinese)
[4]	余佳琛. 民航发动机电子内窥镜孔探检查的管理[J]. 电子世界,2017,39(24): 190-191. doi: 10.19353/j.cnki.dzsj.2017.24.103 YU Jiachen. The management of the borescope detection of civil aviation engine electronic endoscope[J]. Electronics World,2017,39(24): 190-191. (in Chinese) doi: 10.19353/j.cnki.dzsj.2017.24.103
[5]	胡静,徐拓. 基于孔探技术的航空发动机常见损伤及案例分析[J]. 装备制造技术,2017,45(11): 183-185. doi: 10.3969/j.issn.1672-545X.2017.11.059 HU Jing,XU Tuo. The common damage and case analysis of aero-engine based on borescope technology[J]. Equipment Manufacturing Technology,2017,45(11): 183-185. (in Chinese) doi: 10.3969/j.issn.1672-545X.2017.11.059
[6]	杨晓军,王瑛琦,刘智刚. 航空发动机涡轮叶片损伤分析[J]. 机械工程与自动化,2017,46(3): 203-205. doi: 10.3969/j.issn.1672-6413.2017.03.087 YANG Xiaojun,WANG Yingqi,LIU Zhigang. Analysis of damaged blades of aero-engine turbine[J]. Mechanical Engineering & Automation,2017,46(3): 203-205. (in Chinese) doi: 10.3969/j.issn.1672-6413.2017.03.087
[7]	刘斌江,孙科. 孔探技术在航空发动机维修中的应用[J]. 国防制造技术,2016,8(3): 67-69. doi: 10.3969/j.issn.1674-5574.2016.03.020 LIU Binjiang,SUN Ke. Application of hole exploring technology in aero-engine maintenance[J]. Defense Manufacturing Technology,2016,8(3): 67-69. (in Chinese) doi: 10.3969/j.issn.1674-5574.2016.03.020
[8]	樊玮,段博坤,黄睿,等. 基于风格迁移的交互式航空发动机孔探图像扩展方法[J]. 计算机应用,2020,40(12): 3631-3636. doi: 10.11772/j.issn.1001-9081.2020040585 FAN Wei,DUAN Bokun,HUANG Rui,et al. Interactive augmentation method for aircraft engine borescope inspection images based on style transfer[J]. Journal of Computer Applications,2020,40(12): 3631-3636. (in Chinese) doi: 10.11772/j.issn.1001-9081.2020040585
[9]	张栋善,赵成. 航空发动机维修中孔探技术的应用分析[J]. 电子制作,2019,26(12): 98-99. doi: 10.3969/j.issn.1006-5059.2019.12.040 ZHANG Dongshan,ZHAO Cheng. Application analysis of hole exploring technology in aero-engine maintenance[J]. Practical Electronics,2019,26(12): 98-99. (in Chinese) doi: 10.3969/j.issn.1006-5059.2019.12.040
[10]	李华,陈果,陈新波,等. 航空发动机内部裂纹自动测量方法研究[J]. 计算机工程与应用,2016,52(11): 233-237. doi: 10.3778/j.issn.1002-8331.1407-0423 LI Hua,CHEN Guo,CHEN Xinbo,et al. Study on automatic measurement method for aero-engine inner damage crack[J]. Computer Engineering and Applications,2016,52(11): 233-237. (in Chinese) doi: 10.3778/j.issn.1002-8331.1407-0423
[11]	旷可嘉. 深度学习及其在航空发动机缺陷检测中的应用研究[D]. 广州: 华南理工大学, 2017. KUANG Kejia. Research on deep learning and its application on the defects detection for aero-engine[D]. Guangzhou: South China University of Technology, 2017. (in Chinese)
[12]	赵烨. 基于卷积神经网络的叶片损伤识别方法研究[D]. 天津: 中国民航大学, 2019. ZHAO Ye. Research on blade damage identification method based on convolution neural network[D]. Tianjin: Civil Aviation University of China, 2019. (in Chinese)
[13]	李龙浦. 基于孔探数据的航空发动机叶片损伤识别研究[D]. 天津: 中国民航大学, 2020. LI Longpu. Research on damage identification of aero-engine blades based on borescope data[D]. Tianjin: Civil Aviation University of China, 2020. (in Chinese)
[14]	HUANG G,LIU Z,PLEISS G,et al. Convolutional networks with dense connectivity[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence,2019,44(12): 8704-8716.
[15]	LEE Y, HWANG J, LEE S, et al. An energy and gpu-computation efficient backbone network for real-time object detection[C]//CVF Conference on Computer Vision and Pattern Recognition Workshops. Piscataway: IEEE, 2019: 752-760.
[16]	HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]//Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2016: 770-778.
[17]	LEE Y, PARK J. Centermask: real-time anchor-free instance segmentation[C]//CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2020: 13906-13915.
[18]	LIU S, QI L, QIN H, et al. Path aggregation network for instance segmentation[C]//Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2018: 8759-8768.
[19]	BOCHKOVSKIY A, WANG C Y, LIAO H Y M. Yolov4: optimal speed and accuracy of object detection[EB/OL]. [2022-12-21]. https://arxiv.org/abs/2004.10934.
[20]	HE K,ZHANG X,REN S,et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence,2015,37(9): 1904-1916. doi: 10.1109/TPAMI.2015.2389824
[21]	ZHENG Z,WANG P,REN D,et al. Enhancing geometric factors in model learning and inference for object detection and instance segmentation[J]. IEEE Transactions on Cybernetics,2021,52(8): 8574-8586.