基于深度风格迁移合成数据的扣件异常状态检测

Detection of Fastener Abnormal State Based on Deep Style Migration Synthetic Data

扣件系统是轨道结构中的关键基础部件,其服役状态直接影响铁路列车的正常运行。现有基于机器视觉检测方案受样本不均的影响易漏检或错检。提出一种基于图像风格迁移学习的板式无砟轨道扣件系统异常状态检测算法。该算法基于实际场景进行铁路轨道结构的正向建模,按照设计阶段的约束关系参数化生成几何构造完全相同的BIM模型,通过轻量化物理引擎渲染场景细节,并随机部署不同类型状态的扣件系统,输出虚拟巡检图像。在此基础上,利用循环对抗生成网络将虚拟巡检图像进行真实化的风格迁移,得到高仿真且正负样本均衡的合成数据集。最后利用该数据集对典型的深度目标检测网络进行充分训练,实现对板式无砟轨道结构扣件异常状态的精确化检测。实验结果表明:经过风格迁移得到的合成数据集充分训练的Faster R-CNN网络具有最佳的板式无砟轨道结构扣件异常状态检测精度,对正常、断裂、缺失和位移4种类型状态的扣件检测MAP值为94.91%,相比真实数据集高出5.39%,比虚拟BIM数据集高出2.37%。其中缺失状态扣件检测精度提升最大,比真实数据集提高10.13%,实现了扣件状态识别精度的有效提升。

The service status of the fastener system,a key basic component in the track structure,directly affects the normal operation of railroad trains.Existing machine vision-based inspection schemes are prone to missed or false detection due to uneven samples.In this paper,an algorithm was proposed to detect the abnormal state of the fastener system of slab ballastless track based on image style migration learning.By performing forward modeling of railroad track structures based on actual scenes,the algorithm generated BIM models with identical geometric parameters according to the constraint relations in the design phase,rendering details through a lightweight physics engine,and randomly deploying fastener systems with different types of status to output virtual inspection images.On this basis,the virtual inspection images were realistically migrated using recurrent adversarial generation network to obtain a highly simulated synthetic dataset with balanced positive and negative samples.Finally,this dataset was used to fully train the deep target detection network to achieve accurate detection of the abnormal state of the fasteners of slab ballastless track.The experimental results show that the Faster R-CNN network fully trained from the synthetic dataset obtained by style migration has the best accuracy in the detection of the abnormal state of the fasteners of the slab ballastless track,with the MAP value of 94.91%for fastener detection for four types of normal,fractured,missing and displaced states,5.39%higher than the real dataset and 2.37%higher than the virtual BIM dataset.Among them,the detection accuracy of missing fasteners is improved the most,with a 10.13%increase compared to the real dataset,achieving an effective improvement of fastener state recognition accuracy.