基于归一化流模型的机车车侧部件检测研究  

Detection Research for Locomotive Key Components Via Difference Detection Algorithm

作　　者：黄洪 谢茂华 姚毅甡 HUANG Hong;XIE Maohua;YAO Yisheng(Liuzhou Locomotive Depot,China Railway Nanning Group Co.,Ltd.,Liuzhou 545007 Guangxi,China)

机构地区：[1]中国铁路南宁局集团有限公司柳州机务段,广西柳州545007

出　　处：《铁道机车车辆》2025年第S1期51-57,共7页Railway Locomotive & Car

摘　　要：随着重载铁路技术的快速发展,机车关键部件的检测与维护变得尤为重要。机车车底及车侧由于环境复杂,存在较大检测难度。针对这一问题,设计1种基于归一化流模型的检测方法,利用通道注意力和空间注意力机制对特征进行深度处理,使模型更关注待检测区域,将多层特征图合并,从而对分层处理的特征进行深度交叉融合。在MVTec AD数据集上的试验结果表明,性能优于其他同类算法,与同类型模型Fastflow和CS-flow相比,所提方法的AUROC达到了98.6%。在机车车侧数据集上,所提方法达到了98.44%的检出率和1.6%的误报率,通过对机车车侧部件异常的定位和检测,可以有效地确保机车的安全。With the rapid development of heavy haul railway technology,the inspection and maintenance of locomotive key components become particularly important.Because of the complex environment,it is difficult to detect the bottom and side of the locomotive.To solve this problem,this paper designs a detection method based on normalized flow model,which uses channel attention and spatial attention mechanism to deeply process features,so that the model pays more attention to the region to be detected,and finally combines multi-layer feature maps to deeply cross-fuse the features processed by layers.The experimental results on the MVTec AD dataset show that the proposed method outperforms other similar algorithms,with an AUROC of 98.6%compared to similar models such as Fastflow and CS-flow.On the locomotive side dataset,the proposed method achieved a detection rate of 98.44%and a false alarm rate of 1.6%.By locating and detecting anomalies in locomotive side components,it can effectively ensure the safety of locomotives.

关 键 词：机车关键部件检测 归一化流模型 深度学习 

分 类 号：U260.7[机械工程—车辆工程]