面向复杂环境的接触网关键设备缺陷检测方案  

作　　者：常占宁 韦宝泉[2] 张一平 

机构地区：[1]中国铁路兰州局集团有限公司供电部 [2]华东交通大学

出　　处：《电气化铁道》2023年第S01期109-113,共5页Electric Railway

摘　　要：针对小样本及复杂环境下接触网关键设备缺陷检测难等问题,提出一种融合深度卷积神经网络和卡尔曼滤波的图像检测方案。采用MobileNet构建模型骨干网络,有效降低了计算成本;融合柔性非极大值抑制算法解决目标部件遮挡问题,并将上下文感知ROI池化层取代原始池化层,维护了小尺寸零部件的原始结构;最终通过卡尔曼滤波对检测结果进行修正,有效提高检测精度。实验结果表明:本文所述方法能够对复杂接触网设备实现零部件的精确检测,与相同条件下的其他检测算法相比综合性能最佳。With regard to the problem that it is difficult to inspect the defects of key OCS equipment due to the minor samples and complicated environment,an image inspection scheme combining of depth convolutional neural network and Kalman filter is proposed.MobileNet is used to build the model backbone network through which the calculation cost is reduced effectively;integrating of flexible non-maximum suppression algorithms is used to solve the occlusion problem of target components,and the original pooling layer is replaced by the ROI pooling layer with context aware,maintaining the original structure of small-sized components;finally,Kalman filter is used to correct the inspection results and effectively improve the inspection accuracy.The results of experiment show that:the method described in the paper is able to realize the accurate inspection of fittings of complicated OCS equipment,and it has the best overall performance compared to other inspection algorithms under the same conditions.

关 键 词：接触网设备 深度学习 缺陷检测 

分 类 号：U226.8[交通运输工程—道路与铁道工程]