Adversarial Reconstruction Based on Tighter Oriented Localization for Catenary Insulator Defect Detection in High-Speed Railways
Junping Zhong, Zhigang Liu, Cheng Yang, Hongrui Wang, Shibin Gao, Alfredo Núñez
Abstract
The catenary insulator maintains electrical insulation between catenary and ground. Its defects may happen due to the long-term impact from vehicle and environment. At present, the research of defect detection for catenary insulator faces several challenges. 1) Localization accuracy is low, which causes the localized object to be incomplete or/and merge with unnecessary background. 2) Horizontal localization brings inevitable unnecessary information because horizontal box cannot fit well with the shape of insulator. 3) Supervised learning models for defects recognition are unreliable as the available defect samples are insufficient to train models well. To address these issues, this article proposes a novel two-stage defect detection method. In the localization stage, a novel localization network called TOL-Framework is constructed to reduce the background and realize tighter oriented localization. Compared with general basic framework Faster R-CNN, the TOL-Framework cascades a regression module inside basic framework and adds an external postprocess network, which is adversarially trained by standard insulators to refine the localization. These two novel steps greatly improve the oriented localization accuracy. In the defect detection stage, an adversarial reconstruction model that is trained only using normal samples is proposed to evaluate the defect states. A comparison with other methods is conducted using a dataset collected from a 60km section of the Changsha-Zhuzhou railway line in China. The results show the proposed method has the highest localization accuracy, and is effective for insulator defect detection.