受电弓故障的车载图像识别技术

针对列车在途中因受电弓发生故障而影响运行安全的问题，提出了一种受电弓故障的车载图像识别技术，以实时检测受电弓降弓、变形与毁坏，碳滑板异常磨耗与缺口，弓角变形与缺失故障；基于更快速的区域卷积神经网络(Faster R-CNN)目标检测框架设计了弓头图像定位目标检测模型，利用残差网络代替原有卷积网络，利用特征金字塔多尺度预测结构构建了候选区域推荐网络，以精准、快速地进行弓头定位和状态检侧；基于掩码区域卷积神经网络(Mask R-CNN)实例分割框架设计了弓头图像分割模型，并针对性地重新设计了检测头的网络结构与特征图尺寸，以适应受电弓的细长弯曲特征，从而准确、快速分割弓头图像；为了在分割后的二值图中更快速地识别与定位故障，根据受电弓结构尺寸和图像分割模型输出的位置坐标，制定了弓角与碳滑板故障的快速模板匹配策略，并在此基础上编制了详细的故障检测算法与程序。研究结果表明：在相应的数据集上，弓头图像定位目标检测模型的平均检测精度为0.944，平均每帧检测时间为0.029 s，弓头图像分割模型的平均分割精度为0.967，平均每帧检测时间为0.031 s，模板匹配的检测精度为0.985，平均每帧检测时间为0.005 s，故障检测算法的平均检测精度为0.966，平均每帧检测时间为0.051 s。由此可见，提出的检测算法具备了较高的可靠性和实时性。

In-vehicle image technology for identifying faults of pantograph

In view of the problem that the operation safety of the train was affected by pantograph faults, an in-vehicle image technology for identifying pantograph faults was proposed to detect the dropping, deformation and destruction of pantograph, the abnormal wear and notch of carbon contact strip, and deformation and loss faults of pantograph horns in real time. Based on the faster region-convolutional neural network (Faster R-CNN) target detection framework, a target detection model for locating the pantograph bow images was designed, and the residual network was used to replace the original convolutional network. The candidate region recommendation network was constructed by using the feature pyramid multi-scale prediction structure, so as to accurately and quickly locate the pantograph bow and detect the status. Based on the mask region-convolutional neural network (Mask R-CNN) instance segmentation framework, a pantograph bow image segmentation model was designed, and the network structure and feature map size of the detection head were redesigned to adapt to the slender and curved features of the pantograph, so as to accurately and quickly segment the pantograph bow image. In order to identify and locate faults more quickly in the segmented binary image, a rapid template matching strategy for the faults of pantograph horn and carbon contact strip was formulated according to the pantograph structure size and the position coordinates output by the image segmentation model. On this basis, detailed fault detection algorithms and procedures were compiled. Research results show that on the corresponding dataset, the average detection accuracy and average detection time per frame of the target detection model for positioning the pantograph bow images are 0.944 and 0.029 s, respectively. The average segmentation accuracy and average detection time per frame of the pantograph bow image segmentation model are 0.967 and 0.031 s, respectively. In addition, the detection accuracy and average detection time per frame of the template matching are 0.985 and 0.005 s, respectively. The average detection accuracy and average detection time per frame of the fault detection algorithm are 0.966 and 0.051 s, respectively. Thus, the proposed detection algorithm has high reliability and real-time performance.