-
摘要:
针对航空发动机损伤检测深度学习模型在嵌入式设备上部署应用的实时性差、检测精度低等问题，提出一种轻量化航空发动机损伤检测模型FDG-YOLO。引入FasterNet网络重构YOLOv5的主干网络，解决主干网络参数量大的问题；通过深度可分离卷积改进YOLOv5颈部网络的普通卷积，减少颈部网络的冗余参数；以GSConv为基础构建GS C3结构，替换原模型的C3结构，增强模型的表达能力和感受野；在航空发动机损伤数据集进行实验验证。结果表明：与YOLOv5原模型相比，所提FDG-YOLO模型参数量降低了52.5%，浮点运算速度降低了66%，在嵌入式设备上的平均精度均值(mAP)达到89.6%，高于其他轻量化模型，帧率达到61帧/s，检测速度适配于发动机损伤图像采集速度，能够更好地满足航空发动机损伤检测的智能化应用需求。
-
关键词:
- 损伤检测 /
- FDG-YOLO模型 /
- 轻量化 /
- FasterNet网络 /
- 深度可分离卷积 /
- GSConv技术
Abstract:In response to the issues of poor real-time performance and low detection accuracy when deploying deep learning models for aero-engine damage detection on embedded devices, this paper introduces the FDG-YOLO lightweight model for aviation engine damage detection. Firstly, FasterNet was introduced to restructure the backbone network of YOLOv5, addressing the issue of large parameter count in the backbone network. Second, depth-wise separable convolutions were used to eliminate superfluous parameters in the neck network of YOLOv5 by improving ordinary convolutions. In order to improve the model's expressive power and receptive field, the original C3 structure was replaced with the GS C3 structure, which was built concurrently based on GSConv. Finally, experiments were conducted and validated on an aviation engine damage dataset. In the end, experiments were conducted and validated on an aero-engine damage dataset. The findings show that the FDG-YOLO model reduces the number of parameters by 52.5% and the giga floating-point operations per second by 66% when compared to the original model. On embedded devices, the mean average precision (mAP) reaches 89.6%, surpassing other lightweight models. The frames per second achieves 61, making the detection speed suitable for the engine damage image acquisition rate. It more effectively satisfies aero-engine damage detection's intelligent application criteria.
-
图 1 FDG-YOLO轻量化航空发动机损伤检测模型
Figure 1. FDG-YOLO lightweight aero-engine damage detection model
图 2 基于FasterNet的损伤检测模型主干网络改进
Figure 2. Improvement of backbone network of damage detection model based on FasterNet
图 4 基于DSConv的损伤检测模型CBS结构改进
Figure 4. Improvement of CBS structure in damage detection model based on DSConv
图 7 基于FDG-YOLO轻量化模型的航空发动机损伤检测方法流程
Figure 7. Workflow of aero-engine damage detection method based on FDG-YOLO lightweight model
表 1 数据集样本分布
Table 1. Sample distribution of dataset
幅 损伤类型 训练集 验证集 测试集 烧蚀 1360 343 125 凹痕 857 140 85 裂纹 827 136 85 材料缺失 1756 581 150 
下载: 导出CSV
表 2 消融实验结果
Table 2. Ablation experimental results
模型 参数量 mAP/% 浮点运算速度/109 s−1 帧率/(帧·s−1) 实验a 20 865 057 92.0 47.9 92 实验b 11 654 613 88.3 22.5 139 实验c 9 633 429 88.1 20.0 150 实验d 9 909 621 90.5 16.3 142
下载: 导出CSV
表 3 不同模型检测性能对比
Table 3. Comparison of detection performance of different models
模型 参数量 mAP/% 浮点运算速度/109 s−1 帧率/(帧·s−1) YOLOv3-tiny 8 673 622 86.7 12.9 159 YOLOv7-tiny 6 014 737 89.5 13.0 63 FDG-YOLO 9 909 621 90.5 16.3 142
下载: 导出CSV
表 4 嵌入式设备上不同模型检测性能对比
Table 4. Comparison of detection performance of different models on embedded devices
模型 mAP/% 帧率/(帧·s−1) YOLOv3-tiny 85.4 76 YOLOv5 86.7 75 FDG-YOLO 89.6 61
下载: 导出CSV
-
[1] 官颂, 曲培树, 董文会, 等. 航空发动机智能内窥故障诊断[J]. 辽宁工程技术大学学报(自然科学版), 2009, 28(6): 981-984.GUAN S, QU P S, DONG W H, et al. Intelligent endoscope fault diagnose of aero engine[J]. Journal of Liaoning Technical University (Natural Science), 2009, 28(6): 981-984(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(15): 96-98.XIAO B R. Application of borescope inspection technology in aero-engine maintenance[J]. China High Technology, 2019(15): 96-98 (in Chinese). [3] 刘文凯, 陈宸, 袁传金. 航空发动机维护中孔探检测的运用[J]. 科技资讯, 2016, 14(7): 43-44.LIU W K, CHEN C, YUAN C J. Application of borescope inspection in aero-engine maintenance[J]. Science and Technology Information, 2016, 14(7): 43-44 (in Chinese). [4] 于辉. 孔探技术及其在航空发动机维修中的应用[J]. 航空制造技术, 2005, 48(6): 94-96.YU H. Borescope and its application in the maintenance of aeroengine[J]. Aeronautical Manufacturing Technology, 2005, 48(6): 94-96(in Chinese). [5] HE W F, LI C Z, NIE X F, et al. Recognition and detection of aero-engine blade damage based on improved cascade mask R-CNN[J]. Applied Optics, 2021, 60(17): 5124-5133. [6] LIU Y X, WU D B, LIANG J W, et al. Aeroengine blade surface defect detection system based on improved faster RCNN[J]. International Journal of Intelligent Systems, 2023, 2023(1): 1992415. [7] 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. [8] 李彬, 汪诚, 丁相玉, 等. 改进YOLOv4的表面缺陷检测算法[J]. 北京航空航天大学学报, 2023, 49(3): 710-717.LI B, WANG C, DING X Y, et al. Surface defect detection algorithm based on improved YOLOv4[J]. Journal of Beijing University of Aeronautics and Astronautics, 2023, 49(3): 710-717(in Chinese). [9] 刘树东, 刘业辉, 孙叶美, 等. 基于倒置残差注意力的无人机航拍图像小目标检测[J]. 北京航空航天大学学报, 2023, 49(3): 514-524.LIU S D, LIU Y H, SUN Y M, et al. Small object detection in UAV aerial images based on inverted residual attention[J]. Journal of Beijing University of Aeronautics and Astronautics, 2023, 49(3): 514-524(in Chinese). [10] 蔡舒妤, 闫子砚, 师利中. 基于改进YOLOv4的航空发动机损伤检测方法[J]. 现代制造工程, 2023(2): 99-108.CAI S Y, YAN Z Y, SHI L Z. Aeroengine damage detection method based on improved YOLOv4[J]. Modern Manufacturing Engineering, 2023(2): 99-108(in Chinese). [11] 张雯婷. 基于嵌入式图像目标检测系统的研究[D]. 沈阳: 沈阳航空航天大学, 2019: 45-59.ZHANG W T. Research on object detection system for embedded platforms[D]. Shenyang: Shenyang Aerospace University, 2019: 45-59(in Chinese). [12] HE K M, ZHANG X Y, REN S Q, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[C]//Proceedings of the Computer Vision-ECCV. Berlin: Springer, 2014: 346-361. [13] CHEN J R, KAO S H, HE H, et al. Run, don’t walk: chasing higher FLOPS for faster neural networks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2023: 12021-12031. [14] CHOLLET F. Xception: deep learning with depthwise separable convolutions[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2017: 1800-1807. [15] LI H, LI J, WEI H, et al. Slim-neck by GSConv: a better design paradigm of detector architectures for autonomous vehicles[EB/OL]. (2022-06-06)[2024-01-10]. https://arxiv.org/abs/2206.02424. [16] REDMON J, FARHADI A. YOLO9000: better, faster, stronger. [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2017: 6517-6525. [17] WANG C Y, BOCHKOVSKIY A, LIAO H M. YOLOv7: trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2023: 7464-7475. -
下载:
点击查看大图
计量
- 文章访问数: 496
- HTML全文浏览量: 125
- PDF下载量: 13
- 被引次数: 0
