External Fault Identification Method of High Voltage Transmission Line Based on Partial Supervision Convolutional Neural Network

In recent years, deep learning algorithm, i.e., artificial neural network with multiple layer structure, is increasingly being used in intelligent and automatic condition monitoring of electric power equipment, such as instant equipment classification and fault location. With optimized annotation for the image dataset and parameter tuning for the network training process, an average diagnosis accuracy higher than 85.0% and detection time within few hundred microseconds can be reached for deep learning. However, the accuracy of deep learning methods for fault detection heavily relies on the size of image dataset. In general, tens of thousands of on-site inspection images are required for the training of neural network and all the images needs to be labeled with careful annotations. This requirement makes it expensive to annotate new fault categories and has restricted instance segmentation of fault detection to less than 10 well-annotated classes. In this paper, a new partially supervised convolutional neural network, together with a novel weight transfer function, is proposed for automatic location of defects from on-site inspection images of high voltage transmission line, e.g., the identification of wildfire, mechanical invasion, tree barrier and ageing of metal fittings. With improved training paradigm, this new type of convolutional neural network allows training instance segmentation models on a large set of fault categories all of which have simple box annotation, by only a small fraction of which have complex mask annotation. These contributions reduce more than 70% workload for annotation of inspection images, and an average recognition speed of approximately 0.3s and accuracy of 93.7% for automatic segmentation and recognition of defects on transmission lines even when the fault categories is scaled to more than 10 for the same dataset.