Research On Transmission Line Small Target Location And Fault Detection Algorithm Based On FPN

With the continuous development of artificial intelligence technology and intelligent lines,a single means of inspection has been challenging to meet the power grid’s needs.Unmanned aerial vehicles(UAV),helicopters and satellite inspection technology combined with live inspection robots is gradually applied to power grid operation and maintenance,forming an integrated three-dimensional inspection system of space and sky in the future.As an essential fastener of power tower,pins account for about one-third of all fault types.Because pins are small targets and have many similar components,the fault detection of pins is not well performed.Therefore,there is an urgent need to solve the detection of small objects such as pins.In this thesis,data annotation is performed on the UAV line inspection image obtained from aerial photography for deep learning.Then the feature pyramid network(FPN)is applied to test the pin data.We found that there are some deficiencies in distinguishing between faulty pins and normal pins.We proposed two improvement strategies by analyzing the exsiting problems of FPN.First,the feature map obtained by direct scaling is supplemented with the detailed features of the small target through the feature fusion method.Then,an attention mechanism is introduced to enhance the local information.Finally,the two schemes are combined to improve the pin detection accuracy based on FPN.The main contributions of this paper are as follows:(1)In order to solve the problem of insufficient feature information caused by small pin targets,this paper proposes a high-resolution feature pooling method based on feature fusion.The proposed method achieves high-resolution features through feature fusion and bilinear interpolation,which enhances the feature representation of pins.Experimental results show that the proposed model can effectively enrich the feature information of small targets.Compared with the FPN,the accuracy of fault pin detection is increased by 9.82%.The average accuracy of this method is 81.88%,which is 9.39%higher than the FPN.(2)In order to solve the problem that pin faults are challenging to distinguish,this paper designs a spatial attention mechanism based on space maximization pooling.This method can capture the global information on the channel of the feature map and fuse the weight information of different channels to improve the classification performance of FPN.Experimental results show that this method can effectively solve little distinction between good and bad compared with FPN.The average accuracy of the proposed method is 82.44%,which is 10.3%higher than the FPN.Finally,this paper combines the two improved methods and proposes a pin fault detection framework based on FPN.The average accuracy of pin detection is 83.77%,which is 11.63%higher than the original FPN.Compared with FPN,the recognition accuracy of the detection framework has improved by 12.87%.The recall has increased by 10.57%.Meanwhile,the average accuracy of the fault pin has reached 79.49%,which is an increase of 13.69% compared with FPN.