Location and fault detection of catenary support components based on deep learning

Catenary support components (CSCs) are the most important devices that support the overhead line and messenger of catenary system in electric railway. The faults of CSCs can result in the poor state of catenary system, and directly influence the normal operation of trains. In order to efficiently locate and identify the faults of CSCs, the deep learning algorithms are tried to process the captured images of CSCs in this paper. First at all, a dataset of CSCs that contains 50k labeled instances with 12 categories is built. Second, some traditional location methods of CSCs are introduced, and four recent representative deep learning networks, namely Faster RCNN (VGG16 and ResNet101), YOLOv2 and SSD, are applied to locate 12 categories CSCs, simultaneously and separately. In order to find more suitable algorithms of deep learning, their location performances are compared and evaluated. For the location of single category of CSCs, these algorithms show good performance. However, the models that are adopted to simultaneously locate all categories have a poor location performance on small-scale components of CSCs. Third, aiming at the fault detection of CSCs, some common methods are presented and compared, and the deep learning algorithms are tried to detect the faults of CSCs. Finally, the issues of deep learning for location and fault detection of CSCs, especially for the simultaneous location of CSCs are proposed and discussed, and further prospects are given.