Research On Insulator Detection And Fault Recognition With Deep Learning

Electric power is one of the most fundamental energy source in China.As one of the main electric power equipment,the insulator is easy to be damaged due to exposure to the harsh natural environments for long term.The faults such as bending,abrasion,etc.bring threats to the stable operation of the electric power system.This dissertation focuses on localization of insulators and recognition of faults based on the strategy of localization before classification.An improved You Only Look Once(YOLO)v3 method is proposed in this dissertation for quickly locating insulators.Meanwhile,the lightweight multi-scale network is combined with the bottleneck layer for further improving localization speed.Besides,the K-means++ cluster algorithm is utilized to achieve high precision localization of an insulator with its priori knowledge.For avoiding the limitation of the previous improved YOLOv3,a multi-scale rotating area localization method is proposed for improvement of accuracy of insulator detection.With the priori knowledge of insulators,the Rotation Dense Feature Pyramid Networks(R-DFPN)is utilized to improve the anchor ratio,select the optimum combination of anchors and proposals,and introduce the Balanced L1 loss function for a trade-off between the hard and easy samples for improving the inference speed and detection accuracy.An insulator fault classification algorithm is proposed in this dissertation for two kinds of faults,including bending and abrasion.And the transfer learning and data enhancement strategy are utilized to solve the problem of inadequacy of fault samples.The insulator fault detection system is built on a cascade of the insulator localization and fault classification modules whose efficiency has been invalidated by the actual data tests.And the fault recognition accuracy achieves 87.76%.