An Automatic Defect Detection Method for Catenary Bracing Wire Components Using Deep Convolutional Neural Networks and Image Processing

Abstract

Bracing wire components, including messenger wire bases, bracing wire hooks, and bracing wires, are essential to make certain the stability of high-speed catenary support devices. The core process of the proposed detection method includes two stages, the first is that brace wire components are localized first, and the second stage is that the defects of the components are detected. In this article, a new defect detection method for catenary bracing wire components is proposed. First, Faster R-CNN, one of the advanced deep convolutional neural networks, is introduced to localize messenger wire bases and bracing wire hooks, accurately. And then, the regions of bracing wires are extracted according to the structure information among bracing wire components and bracing wires are localized by the Hough Transformation. Next, the installation defects of messenger wire bases are detected with image processing methods, and the looseness defects of bracing wires are detected by combining with the results detected by the Hough Transform. Experiment results prove that the proposed scheme can localize and identify the defects of bracing wire components, accurately.