基于R-D SSD模型航空发动机安装工位检测算法

陈科山; 郝宇; 何泓波; 李坤龙

doi:10.13700/j.bh.1001-5965.2020.0321

基于R-D SSD模型航空发动机安装工位检测算法

doi: 10.13700/j.bh.1001-5965.2020.0321

陈科山^{1, 2, ,},
郝宇¹,
何泓波¹,
李坤龙¹

1.
北京交通大学机械与电子控制工程学院, 北京 100044
2.
北京交通大学载运工具先进制造与测控技术教育部重点实验室, 北京 100044

基金项目:

国家自然科学基金 51735009

详细信息

作者简介:
陈科山男, 博士, 教授, 硕士生导师。主要研究方向: 机器人设计与控制技术、计算机视觉与图像识别技术、智能交通与新能源技术等

郝宇男, 硕士研究生。主要研究方向: 计算机视觉与图像识别技术

通讯作者:
陈科山. E-mail: kshchen@bjtu.edu.cn

中图分类号: V239;TP181
计量
- 文章访问数: 589
- HTML全文浏览量: 48
- PDF下载量: 77
- 被引次数: 0
出版历程
- 收稿日期: 2020-07-06
- 录用日期: 2020-09-04
- 网络出版日期: 2021-04-20

Detection algorithm of aeroengine installation station based on R-D SSD model

CHEN Keshan^{1, 2
, ,},
HAO Yu¹,
HE Hongbo¹,
LI Kunlong¹

1.
School of Mechanical, Electronic and Control Engineering, Beijing Jiaotong University, Beijing 100044, China
2.
Key Laboratory of Advanced Manufacturing and Measurement and Control Technology of Transport Tools, Ministry of Education, Beijing Jiaotong University, Beijing 100044, China

Funds:

National Natural Science Foundation of China 51735009

More Information

Corresponding author: CHEN Keshan. E-mail: kshchen@bjtu.edu.cn

摘要

摘要:
为解决航空发动机在安装过程中大多实行人工安装、定位不精确等问题，在研究其自动化安装方法中，针对航空发动机安装工位的检测需求，提出了一种残差网络与膨胀卷积相融合的SSD改进算法(R-D SSD)。将经典SSD模型的主干网络VGG16替换为残差网络ResNet-101，并增加其输出特征图上的预选框数量，解决了原始算法对底层特征抓取能力不足的问题，进而弥补了对小目标检测效果较差的缺陷；利用膨胀卷积扩大网络的感受野，获取足够的安装工位边缘特征细节信息，在不改变网络结构的同时，保证了模型良好的实时性和对目标的检测精度。实验表明：对于小目标数据集和整个数据集，R-D SSD算法的平均检测精度较原始算法分别提高了8.6%和4.0%，可以满足航空发动机安装时平均检测精度不低于85%的要求。
- 安装工位 /
- 残差网络 /
- 膨胀卷积 /
- SSD模型 /
- 小目标检测 /
- 预选框数量
Abstract:
In order to solve the problems of manual installation and inaccurate positioning in the process of aeroengine installation, an improved SSD algorithm (ResNet-Dilated SSD, R-D SSD) is proposed to meet the detection requirements of aeroengine installation station in the research of its automatic installation method. The VGG16, the backbone network of classical SSD model, is replaced by the residual network ResNet-101 and the number of preselected boxes on output feature map is increased, which solves the problem that original algorithm has insufficient ability to grasp the underlying features, and thus results in poor detection effect of small target. The dilation convolution is used to expand the network's receptive field to obtain enough edge feature information of installation station, which ensures the good real-time performance of the model and the detection accuracy of the target without changing the network structure. The experimental results show that the average detection accuracy of the R-D SSD detection algorithm is 8.6% and 4.0% higher than that of original algorithm for the small target dataset and the whole dataset. It can meet the requirement that the average detection accuracy is not less than 85% when the aeroengine is installed.
- installation station /
- residual network /
- dilation convolution /
- SSD model /
- small target detection /
- number of preselected boxes

HTML全文

图 1 SSD安装工位检测效果

Figure 1. SSD detection effect of installation station

下载: 全尺寸图片幻灯片

图 2 R-D SSD网络结构

Figure 2. R-D SSD network structure

下载: 全尺寸图片幻灯片

图 3 残差学习单元

Figure 3. Residual learning unit

下载: 全尺寸图片幻灯片

图 4 增加预选框数量

Figure 4. Increasing the number of preselected boxes

下载: 全尺寸图片幻灯片

图 5 膨胀卷积

Figure 5. Dilation convolution

下载: 全尺寸图片幻灯片

图 6 小目标检测效果对比

Figure 6. Comparison of small target detection effect

下载: 全尺寸图片幻灯片

图 7 正常大小目标检测效果对比

Figure 7. Comparison of normal-size target detection effect

下载: 全尺寸图片幻灯片

图 8 混淆矩阵

Figure 8. Confusion matrix

下载: 全尺寸图片幻灯片

表 1 实验数据集划分

Table 1. Experimental dataset partition

类别	训练集/张	占比/%	测试集/张	占比/%
前安装工位	2 069	41	236	5
后安装工位	1 881	38	192	4
前安装工位+后安装工位	550	11	79	1

下载: 导出CSV

表 2 小目标检测精度对比

Table 2. Comparison of small target detection accuracy

检测算法	检测对象	AP/%	mAP/%
SSD	前安装工位	67.2	71.9
SSD	后安装工位	76.6	71.9
R-D SSD	前安装工位	77.9	80.5
R-D SSD	后安装工位	83.1	80.5

下载: 导出CSV

表 3 正常大小目标检测精度对比

Table 3. Comparison of normal-size target detection accuracy

检测算法	mAP/%
SSD	86.7
R-D SSD	88.1

下载: 导出CSV

表 4 不同检测算法性能比较

Table 4. Comparison of performance among different detection algorithms

检测算法	主干网络	mAP/%	检测速度/(帧·s^-1)
Faster R-CNN	ResNet50	88.5	8.2
YOLOv3	Darknet-53	83.2	51.0
SSD	VGG16	82.9	52.6
R-D SSD	ResNet-101	86.9	40.5

下载: 导出CSV

参考文献(16)

[1]	GULA P, ULMA D, ZUREK K, et al. Challenges of turboprop engine installation on small aircraft[J]. Aircraft Engineering and Aerospace Technology, 2019, 91(7): 938-948. doi: 10.1108/AEAT-09-2017-0198
[2]	赵哲. 航空发动机数控安装架车的研制[D]. 大连: 大连理工大学, 2014: 14-20. ZHAO Z. Research on NC installation vehicle for aero-engine[D]. Dalian: Dalian University of Technology, 2014: 14-20(in Chinese).
[3]	GAO H, LEI H, ZHAO Z, et al. The large aero-engine NC installation method and its multi-axial position adjustment platform design[C]//2013 IEEE International Symposium on Assembly and Manufacturing. Piscataway: IEEE Press, 2013: 270-273.
[4]	李洋, 谢曦鹏. 飞机发动机数字化对接安装工艺研究[J]. 装备制造技术, 2013(9): 90-92. https://www.cnki.com.cn/Article/CJFDTOTAL-GXJX201309035.htm LI Y, XIE X P. Research on aeroengine digital assembly technology[J]. Equipment Manufacturing Technology, 2013(9): 90-92(in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-GXJX201309035.htm
[5]	高航, 宋强, 刘国, 等. 航空发动机整机数控安装多轴调姿方法及其应用[J]. 航空制造技术, 2017, 60(11): 16-21. https://www.cnki.com.cn/Article/CJFDTOTAL-HKGJ201711005.htm GAO H, SONG Q, LIU G, et al. Method and application for aeroengine overall unit NC installation with multi-axis attitude adjustment[J]. Aeronautical Manufacturing Technology, 2017, 60(11): 16-21(in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-HKGJ201711005.htm
[6]	LIU W, ANGUELOV D, ERHAN D, et al. SSD: Single shot multibox detector[C]//Proceedings of the European Conference on Computer Vision. Berlin: Springer, 2016: 21-37.
[7]	陈映雪, 丁文锐, 李红光, 等. 基于视频帧间运动估计的无人机图像车辆检测[J]. 北京航空航天大学学报, 2020, 46(3): 634-642. doi: 10.13700/j.bh.1001-5965.2019.0279 CHEN Y X, DING W R, LI H G, et al. Vehicle detection in UAV image based on video interframe motion estimation[J]. Journal of Beijing University of Aeronautics and Astronsutics, 2020, 46(3): 634-642(in Chinese). doi: 10.13700/j.bh.1001-5965.2019.0279
[8]	李慧. 飞机蒙皮图像的深度特征学习与损伤监测[D]. 北京: 北京邮电大学, 2019: 35-36. LI H. High-level feature learning and damage monitoring of aircraft surface images[D]. Beijing: Beijing University of Posts and Telecommunications, 2019: 35-36(in Chinese).
[9]	LIU S, DENG W. Very deep convolutional neural network based image classification using small training sample size[C]//3rd IAPR Asian Conference on Pattern Recognition. Kuala Lumpur: Institute of Electrical and Electronics Engineers Inc., 2015: 730-734.
[10]	黄豪杰, 段先华, 黄欣辰. 基于深度学习水果检测的研究与改进[J]. 计算机工程与应用, 2020, 56(3): 127-133. https://www.cnki.com.cn/Article/CJFDTOTAL-JSGG202003020.htm HUANG H J, DUAN X H, HUANG X C. Research and improvement of fruits detection based on deep learning[J]. Computer Engineering and Applications, 2020, 56(3): 127-133(in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-JSGG202003020.htm
[11]	HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2016: 770-778.
[12]	TIAN Y L, ZHANG Q S, REN Z L, et al. Multi-scale dilated convolution network based depth estimation in intelligent transportation systems[J]. IEEE Access, 2019, 7: 185179-185188. doi: 10.1109/ACCESS.2019.2960520
[13]	IOFFE S, SZEGEDY C. Batch normalization: Accelerating deep network training by reducing internal covariate shift[J]//International Conference on Machine Learning, 2015, 1: 448-456.
[14]	REN S, HE K, GIRSHICK R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149. http://ieeexplore.ieee.org/document/7485869
[15]	GIRSHICK R. Fast R-CNN[C]//Proceedings of the IEEE International Conference on Computer Vision. Piscataway: IEEE Press, 2015: 1440-1448.
[16]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: Unified, real-time object detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2016: 779-788.