Intelligent Aging Diagnosis of Conductor in Smart Grid Using Label-Distribution Deep Convolutional Neural Networks

Quantitatively aging diagnosis of conductor surface remains critical challenging in fault diagnosis of smart high-voltage electricity grid. Inspired by the facial age estimation in computer vision, this work proposes a label-distribution deep convolutional neural networks (CNNs) model, which includes an AlexNet-based deep convolution network and a designed loss embedded with Gaussian label distribution. The aging diagnosis problem of conductor morphology is transformed into a multiclassification problem. The proposed model is improved via a weakly labeled training dataset and a designed loss function (combination of entropy loss, cross-entropy loss, and Kullback-Leibler divergence loss). Compared with four frequently used CNN-based classifiers, the proposed classifier on the collected dataset achieves a better performance. In addition, the influence of parameters and types of label distribution on classification accuracy is also investigated. Here, a promising technique is presented for the aging estimation of aged conductor with high accuracy when the images of conductor surface are available.