Automatic Insulator Detection for Power Line Using Aerial Images Powered by Convolutional Neural Networks

Abstract

The inspection and fault diagnosis of power transmission line via aerial images is becoming widely used in smart grids and electrical power systems. However, the accuracy of detecting objects with complicated backgrounds still remains an important challenge in the defect detection system. In this article, a deep-learning based method is presented to realize the automation of insulator fault detection. In the proposed approach, the insulator localization network based on Single Shot Multibox Detector (SSD), a one-stage detecting backbone, is utilized to extract multi-level features and make predictions in order to determine the insulator situation in the aerial images. Additionally, Densely Connected Convolutional Networks (DenseNet) is newly introduced to strengthen the insulator detecting system’s classification ability. To address the data scarcity problem, a two-stage data augmenting strategy is also applied. The first stage is mainly based on the image combination, while the second stage includes random affine transformation, Gaussian blur, brightness and contrast transformation, and salt-and-pepper noise. The insulator detection network and the classification network precisions of the proposed method have reached 0.95 and 0.98 using the text dataset, respectively. The results have shown that the proposed method can ensure robustness and accuracy during insulator detection.