1.School of Mechanical, Electrical and Vehicle Engineering, Beijing University of Civil Engineering and Architecture, Beijing 100044, China
2.Beijing Key Laboratory of Performance Guarantee on Urban Rail Transit Vehicles, Beijing 100044, China
Show less
文章历史+
Received
Published
2024-02-19
2025-01-01
Issue Date
2026-07-13
PDF (2842K)
摘要
针对轨道扣件车载实时定量检测中检测的速度与精度难以平衡以及设备空间有限的问题,提出一种基于边缘AI计算的改进YOLOv8(You Only Look Once version 8)轨道扣件定量检测方法,并进行了部署和测试。首先,为实现轻量化部署,优化YOLOv8网络结构,嵌入移动神经网络V3(MobileNetV3)轻量化网络,引入压缩激励(SE)注意力机制,并重构颈部网络,加入可变形卷积;其次,结合推理引擎加速方法优化重构的网络模型,并将其部署在Jetson AGX Xavier边缘AI计算设备上;最后,对分割结果进行像素级提取,引入最小外接矩阵,定量分析扣件的断裂程度和偏转角度,并优化检测结果。结果表明:改进后的轻量化网络参数数量减少了22%;在边缘AI计算设备上的帧率相对原YOLOv8模型提升了80%,达到58帧·s-1;通过定量分析结果对模型进行修正,平均精度达到了97.0%,满足检测车辆所需的最低45帧·s⁻¹检测要求,实现了轨道扣件的定量化实时检测。
Abstract
The aim of this study is to address the challenges of balancing detection speed and accuracy of rail fasteners in real-time on-board quantitative detection, as well as the limitations posed by equipment space. The proposed solution is an enhanced YOLOv8 quantitative detection method for rail fasteners, utilising edge AI computing. This method has been tested and deployed. Firstly, to achieve lightweight deployment, the YOLOv8 network structure is optimised by embedding the MobileNetV3 lightweight network, introducing the Squeeze-and-Excitation (SE) attention mechanism, and reconstructing the neck network to incorporate deformable convolution. Secondly, the reconstructed network model is further optimised using the Jetpack-TensorRT acceleration method and deployed on a Jetson AGX Xavier edge AI computing device. Finally, the segmentation results are extracted at the pixel level, and a minimum outer join matrix is introduced to quantify the fracture degree and deflection angle of the fasteners, thereby optimising the detection results. The results demonstrate that the number of parameters of the enhanced lightweight network is reduced by 22%; the frame rate on the edge AI computing device is enhanced by 80% relative to the original YOLOv8 model, reaching 58 frames per second (fps). By refining the model based on quantitative analysis results, the average accuracy reaches 97.0%, which meets the minimum detection requirement of 45 fps for vehicles. Thus, quantitative real-time detection for rail fasteners is achieved.
近年来,随着计算机技术的不断发展,基于深度学习的轨道损伤检测方法逐渐成为研究热点。裴莹玲等[6]利用改进Faster-RCNN[7](Regions with Convolutional Neural Networks)对扣件状态进行检测,并采用阿尔法交并集作为目标回归损失函数提高扣件的回归精度。许贵阳等[8]提出一种基于改进快速卷积神经网络的方法对轨道板裂缝进行检测,该改进方法精简网络模型,采用非极大值抑制算法,改善轨道板裂缝检测的重叠状况,提高裂缝检测效果。高嘉琳等[9]利用改进YOLOv4(You Only Look Once version 4),在主干网络的第2个残差块嵌入卷积结构和YOLO头部结构,增加输出端,以增强对扣件的检测。白堂博等[10]利用最小外接矩形法改进掩膜卷积神经网络的输出层,以提高偏移扣件检测的准确率。吴送英等[11]提出一种改进的YOLOv5s(You Only Look Once version 5s)方法实现对昏暗、遮挡、杂物干扰和模糊复杂背景下铁路扣件外观缺陷形态的准确检测。
YEYangsheng, CAIDegou, ZHANGQianli, et al. Current Status and Development Trend of Design Method for High-Speed Railway Subgrade Bed Structure [J]. China Railway Science, 2021, 42 (3): 1-12. in Chinese
[3]
胡亚东.我国铁路重载运输技术体系的现状与发展[J].中国铁道科学,2015,36(2):1-10.
[4]
HUYadong. Current Status and Development Trend of Technology System for Railway Heavy Haul Transport in China [J]. China Railway Science, 2015, 36 (2): 1-10. in Chinese
[5]
HUANGJ Y, LIUJ H, GONGH, et al. A Comprehensive Review of Loosening Detection Methods for Threaded Fasteners [J]. Mechanical Systems and Signal Processing, 2022, 168: 108652.
[6]
ZHANZ K, SUNH, YUX D, et al. Wireless Rail Fastener Looseness Detection Based on MEMS Accelerometer and Vibration Entropy [J]. IEEE Sensors Journal, 2020, 20 (6): 3226-3234.
[7]
LIUC, ZHANGL H, RENT Q, et al. Detection of Rail Fastener Conditions Using Time-Frequency Entropy Based on Orthogonal Empirical Mode Decomposition [J]. Applied Mechanics and Materials, 2013, 333/334/335: 1708-1712.
PEIYingling, LUOHui, ZHANGShihui, et al. High-Speed Rail Fastener Detection Algorithm Based on Improved Faster R-CNN [J]. Journal of East China Jiaotong University, 2023, 40 (1): 75-81. in Chinese
[10]
RENS Q, HEK M, GIRSHICKR, 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.
XUGuiyang, ZHANGShiquan, BAITangbo. Crack Detection Method of CRTSⅡ Track Slab Based on Faster R-CNN Improvement [J]. China Railway Science, 2023, 44 (1): 106-113. in Chinese
GAOJialin, BAITangbo, YAODechen, et al. Detection of Track Fastener Based on Improved YOLOv4 Algorithm [J]. Science Technology and Engineering, 2022, 22 (7): 2872-2877. in Chinese
BAITangbo, GAOJialin, YANGJianwei, et al. Research on Railway Fastener State Inspection Method Based on Improved Mask R-CNN [J]. Railway Engineering, 2021, 61 (6): 140-143. in Chinese
WUSongying, LIULinya, JIANGJiaming, et al. Detection Algorithm of Improved YOLOv5s of Railway Fastener under Complex Background [J]. China Railway Science, 2023, 44 (3): 53-63. in Chinese
[19]
HOWARDA, SANDLERM, CHENB, et al. Searching for MobileNetV3 [C]// 2019 IEEE/CVF International Conference on Computer Vision (ICCV). Seoul, Korea (South). Danvers: IEEE, 2019: 1314-1324.
[20]
HUJ, SHENL, ALBANIES, et al. Squeeze-and-Excitation Networks [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42 (8): 2011-2023.
[21]
ZHUX Z, HUH, LINS, et al. Deformable ConvNets V2: More Deformable, Better Results [C]// 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seoul, Korea (South). Danvers: IEEE, 2019: 9300-9308.
[22]
JUB S, HYEON-JUNC, JEONGG M. YOLO Model FPS Enhancement Method for Determining Human Facial Expression Based on NVIDIA Jetson TX1 [J]. The Journal of Korea Institute of Information, Electronics, and Communication Technology, 2019, 12 (5): 467-474.
[23]
LIY T, FANQ S, HUANGH S, et al. A Modified YOLOv8 Detection Network for UAV Aerial Image Recognition [J]. Drones, 2023, 7 (5): 304.
WANGWeidong, ZHANGChenlei, HUWenbo, et al. Fine-Grained Measurement of Ballastless Track Slab Cracks Based on Improved Faster R-CNN and Orthogonal Projection [J]. China Railway Science, 2023, 44 (6): 46-56. in Chinese
[26]
MAN N, ZHANGX Y, ZHENGH T, et al. Shufflenet v2: Practical Guidelines for Efficient CNN Architecture Design [C]// Computer Vision – ECCV 2018. Lecture Notes in Computer Science, vol 11218. Munich, Germany. Cham: Springer International Publishing, 2018: 122-138.
[27]
HANK, XIAOA, WUE H, et al. Transformer in Transformer [J]. Advances in Neural Information Processing Systems, 2021, 34: 15908-15919.
[28]
XIAOX, ZHANGD Y, HUG W, et al. CNN-MHSA: a Convolutional Neural Network and Multi-Head Self-Attention Combined Approach for Detecting Phishing Websites [J]. Neural Networks, 2020, 125: 303-312.