Research On Algorithms Of Damper Detection By Convolutional Neural Networks

The accurate classification and location of dampers is a vital technology of transmission line operation and maintenance.Due to the variable Angle of view and the shooting distance of UAV,images of dampers have a large scale change and occlusion phenomenon,which lead to the low accuracy of the detection for dampers by the existing methods.Therefore,this paper puts forward a damper detection algorithm based on convolution neural networks.From the construction of the damper data set and the design of multi-scale and occlusion damper detection algorithm,the research on the variable scale occlusion damper object detection of the transmission line is carried out.This research has significant application value for maintaining the stable operation of transmission lines.The main research work of this paper is as follows:Aiming at the detection task of transmission line dampers,the data set of transmission line damper detection TLDD is constructed.The data set has 3261 pictures,covering a variety of scales and different degrees of shielding.According to the experimental results on the TLDD data set,the existing object detection algorithms have low detection accuracy,which is challenging to meet the requirements of transmission line operation and maintenance.Aiming at the problem of the high miss detection rate of the small-scale dampers,this paper proposes Adaptive Feature Pyramid Networks(AFPN).AFPN mainly includes Adaptive Feature Fusion(AFF)and Residual Feature Augmentation(RFA).AFF module calculates the adaptive weight factor to integrates the features of different scales in a more scientific way.RFA adaptively pools the deep features to obtain the features of different scales,and makes up for the loss of semantic information of the deep feature through the fusion of residual branches.The experimental results show that AFPN proposed in this paper can effectively improve the detection results of dampers.Compared with FPN,the detection accuracy of AFPN is increased by 2.1%,and the missed detection rate of AFPN is reduced by 2.4%.Aiming at the problem of the low detection accuracy caused by occlusion in the occlusion damper detection,this paper proposes Supervised Attention Region Convolutional Neural Networks(SA RCNN).This method includes Box Supervised Attention Module(BSAM),One Proposal Multiple Predictions(OPMP),and improved non-maximum suppression algorithm(Group NMS).BSAM uses the actual box of dampers as the monitoring information to guide the learning of attention mechanism,enhance the feature of the damper,and improve the positioning accuracy of dampers.OPMP predicts multiple instances for a single region proposal which reduces the difficulty of detection of severe occlusion dampers,and the Group NMS algorithm combined with the OPMP module can not suppress each other for multiple predicted instances in the same region proposal,so that the detection results of dampers with severe occlusion can be retained.The experimental results on the damper data set TLDD show that compared with the existing Adaptive NMS and other occlusion detection algorithms,the detection accuracy of SA-RCNN is increased by 1.5%,and the missed detection rate of AFPN is reduced by 1.2%.The experimental results show that the proposed method can effectively improve the detection accuracy of multi-scale occlusion dampers,reduce the missing detection rate of dampers and provide a new technical way for the detection of transmission line dampers.