| China’s electric power development is still in the rising period,the power line erection area is still expanding,the length of transmission lines is constantly extending,and the power system has higher and higher requirements for safe and stable operation.Glass insulator is one of the most important power equipment on high-voltage transmission line.It is easy to explode when exposed to natural environment for a long time,which is a threat to the stable operation of power system.Regular inspection of transmission lines and timely detection of abnormal insulators are the necessary premise to ensure the safe and stable operation of power system.With the key breakthrough of UAV technology,it has been widely used in various industries,and also plays a very important role in the inspection of insulator status.Aiming at the problems of low efficiency and poor robustness of traditional insulator detection algorithm in UAV image,an improved Faster-RCNN algorithm is proposed to detect the insulator in the image,train the original Faster-RCNN,cluster the real box of the training set by K-means,optimize the parameters of the original model and train it again.The experimental results show that taking the real frame of insulator training data set as the initial value of anchor box can improve the detection performance of the algorithm.Aiming at the data set containing insulator explosion,an improved YOLOv3 insulator explosion detection algorithm YOLOv3＿In＿Res is proposed.This method changes the network structure to improves YOLOv3 in three aspects: adding the Inception-Resnet module,improving the loss function according to the characteristics of the data set,using the K-means algorithm for the original model to choose initial parameters.The experimental results show that,compared with the original YOLOv3 algorithm under the self-made insulator data set,the algorithm YOLOv3＿In＿Res proposed in this paper before parameter optimization without K-means has improved the accuracy rate of 5.55%,the recall rate of 7.26%,and the m AP value increased by 6.82% than the original YOLOv3;More over,by K-means algorithm clustering the data set and adjusting the parameters,the accuracy rate was further increased by 0.74%,the recall rate was increased by 0.23%,and the m AP value was increased by 0.34%. |
