A YOLOv8n-based fault detection method for electric multiple unit body structures
摘要
随着我国高速铁路的快速发展,动车组列车的安全运行对动车车身故障检测提出了更高要求.动车车身故障检测的核心难点集中于3大关键问题:一是低照度环境导致图像细节丢失,故障特征辨识度大幅降低;二是故障尺度跨度显著(从几像素的细微断裂到数百像素的大面积漏油),多尺度目标同步检测难度大;三是故障形态复杂、边缘特征模糊且与背景相似度高,易受管线、阴影等干扰导致漏检误检.传统人工巡检效率低、实时性差,现有的动车组运行故障图像检测系统(trouble of moving EMU detection system,TEDS)依赖传统模板匹配算法,在复杂场景下仍存在精度不足、漏检误检率高等瓶颈,难以满足智能化检测需求.针对上述痛点,提出一种基于深度学习的动车车身故障检测方法.通过对动车组车头异物、盖板打开、防松铁丝断裂、涂层脱落及漏油这5大类典型故障的视觉特征进行剖析,并经多组严谨对比实验验证后,最终选定高效单阶段目标检测模型YOLOv8n作为本研究的基础检测框架.针对实际检测场景中低照度环境导致的图像细节丢失、质量不佳问题,引入基于限制对比度自适应直方图均衡化(contrast limited adaptive histogram equalization,CLAHE)的图像增强方法,有效提升了输入图像的细节清晰度与视觉质量,为后续故障检测提供高质量数据支撑.为进一步突破检测性能瓶颈,针对上述故障目标存在的尺度差异显著、形态结构复杂、边缘特征模糊及与背景相似度高的共性问题,提出改进型网络结构 YOLOv8-DCNet:其中,多尺度扩张残差融合模块(C2f with dilation-wise residual,C2f_DWR)通过不同扩张率卷积核的特征捕捉与残差连接的信息传递,有效解决故障目标尺度差异大的难题;上下文与空间特征校准网络(context and spatial feature calibration network,CSFCN)则借助全局上下文关联建模与局部空间注意力权重分配,显著提高对边缘模糊特征的识别精度,强化复杂背景下故障目标的区分能力.最终模型在精确率、召回率、mAP50及mAP50-95上分别提高7.88%、6.62%、6.65%和11.53%.
With the rapid development of high-speed railways in China,the safety of Electric Multiple Units(EMU)has imposed more rigorous demands on car body fault detection.The core challenges in EMU fault detection are centered on three critical issues.Firstly,the low-illumination environments lead to the loss of image details,significantly reducing the discriminability of fault features.Secondly,the significant span in fault scales(ranging from minute fractures of a few pixels to large-scale oil leaks spanning hundreds of pixels)makes simultaneous multi-scale target detection difficult.And thirdly,the complex morphology,blurred edge features,and high similarity between faults and backgrounds,often compounded by interference from pipelines and shadows,lead to frequent missed detections and false alarms.Traditional manual inspections are inefficient and lack real-time capability,while the existing Trouble of moving EMU Detection System(TEDS)relies on conventional template-matching algorithms.The latter still faces bottlenecks such as insufficient precision and high error rates in complex scenarios,failing to meet the requirements for intelligent detection.To address these points,this study proposed a deep learning-based method for EMU car body fault detection.After analyzing the visual characteristics of five typical fault categories—foreign objects on the train head,abnormal opening of covers,fracture of anti-loose wires,coating peeling,and oil leakage—and conducting rigorous comparative experiments,the efficient one-stage object detection model YOLOv8n was selected as the foundational framework.To mitigate the loss of image detail and poor quality caused by low-illumination environments,a Contrast Limited Adaptive Histogram Equalization(CLAHE)-based image enhancement method was introduced,effectively improving the clarity and visual quality of input images and providing high-quality data support for subsequent detection.To further overcome performance bottlenecks associated with significant scale variance,complex structures,and blurred edges,an improved network structure,YOLOv8-DCNet,was proposed.Within this framework,the C2f with Dilation-wise Residual(C2f_DWR)module effectively addresses the challenge of large scale variance through feature capture using various dilation rates and information transmission via residual connections.Furthermore,the Context and Spatial Feature Calibration Network(CSFCN)significantly improves the recognition accuracy of blurred edge features and strengthens the ability to distinguish fault targets in complex backgrounds by utilizing global context correlation modeling and local spatial attention weight allocation.Experimental results demonstrate that the proposed model can achieve improvements in Precision,Recall,mAP50,and mAP50-95 by 7.88%,6.62%,6.65%,and 11.53%,respectively.