An Automated Defect Detection Approach for Catenary Rod-Insulator Textured Surfaces Using Unsupervised Learning

This article aims to present a defect detection approach for catenary rod-insulators based on unsupervised learning. The idea of the proposed detection method is that 1) the regions of catenary insulator pieces are separated first; 2) then, the real samples of the segmented catenary insulator pieces are reconstructed as the foreground, and through the difference between the segmented and reconstructed images, the defect regions are separated as the background; and 3) the defect levels are evaluated according to the separated defect regions. The separation effectiveness of the foreground and background in the second stage is the most vital key. Therefore, a reconstruction and classification convolutional autoencoder network (RCCAEN) is built. Compared with standard autoencoders (AEs), the proposed convolutional AE network can improve the system real time and the ambiguity of the edges of the recovered images. Moreover, it can avoid the interference from the incorrectly segmented components by incorporating a small classification network after the latent support vector. Besides, the dropout layer is added behind the input images to enhance the robustness of the proposed network to noise. In this article, first, the regions of catenary insulator pieces are segmented by combining Mask regions with CNN features (R-CNN) and curving fitting method. And then, the defect regions are extracted by the proposed network RCCAEN. Finally, the defect levels are judged by the indexes of the density-based spatial clustering of applications with noise. The experimental results show that the proposed approach can accurately detect the defects of catenary insulator severe pollution.