Research On Insulator Damage Detection Based On Improved Faster R-CNN Algorithm

Due to the rapid development of electric power transport in China’s power grids,higher requirements are made for the safety of grid insulators.In order to handle the increasingly complex changes and to guarantee safe power transport,the detection of the insulator damage needs to be improved.Generally,drones are used to inspect the grid in China,yet due to the complex and diverse characteristics of insulators and the limitations of the filming conditions,there are often missed and wrong inspections.In response to the current problems of the insulator detection,this paper presents an insulator damaged detection method based on the improved Faster R-CNN algorithm.In view of the low accuracy of the traditional Faster R-CNN algorithm for insulator damaged small targets detection,a deeper residual neural network,the Res Net-101 is used as the feature extraction network.It also introduces an improved feature pyramid to improve the semantic information of the damaged target in the feature map through the fusion of multi-level and multi-scale features.The original network parameter normalization method is varied and group normalization is applied instead of batch normalization to enhance the problem of poor batch normalization due to the small batch size of the Faster R-CNN algorithm.An improvement of the traditional Non-Maximum suppression method is introduced by using Soft-NMS to avoid incorrect screening of damaged features as a result of overlapping with insulator targets.A training strategy for online hard example mining is also introduced to enhance the training effect of hard examples in the network.This paper sets up its own insulator image dataset by image enhancement and other methods,on which the improved algorithm is trained before and after.A comparison of the detection effect of the algorithm before and after the improvement and the change of the Loss function during training shows that the detection effect of the algorithm after the improvement is significantly improved,in particular for the insulator damaged target.The results indicate that the detection effect of the damaged insulator target is increased by 37.45% and the detection effect of the insulator target is increased by 5.09%,which demonstrates the effectiveness of the improved method.