Lightweight Neural Network for Component Recognition and Fault Detection of Large-Scale Transmission Line

Abstract

Recognition and fault detection of transmission line components is a basic problem that puzzles the development of smart grid. However, due to the limitation of current datasets, it is difficult to carry out large-scale model training and mobile reasoning based on artificial intelligence technology. Although the current standard datasets for the classification of components in transmission line scenarios contain many different types of objects, they are all smaller datasets based on a certain category. This paper constructs and annotates 10 226 standard image Power Database which contain 6 component categories (including 3 fault component categories). In addition, aiming at the problem that the general target detection algorithm infers too slowly or can not run in the real-time inspection process of transmission line mobile terminal equipment, a model MSFF-KCD (Multi-scale feature fusion in key component detection) is presented, which combines the multi-scale feature fusion method with the detection of transmission line key components. Experiments on real-time component identification and fault detection by deploying the model to mobile devices show that the proposed method establishes a new performance bound and this mothed suitable for mobile terminal detection of key components of transmission lines. We also compared the models of recognition and detection for large and small targets on Power Database, which provided enlightenment for the identification and fault detection of power components.