Research On Fault Detection Of Power Insulator

With Insulator is widely used as a basic component in overhead lines,its fault is the primary cause of power line accidents.As a general means of troubleshooting,regular power inspection is indispensable to maintain the safety and stability of power transmission.The traditional inspection method consumes time and energy.The UAV technology which loaded with image acquisition has become more and more intelligent to inspect transmission line faults,but the fault images need more efficient target detection algorithm and accurate location of fault points.The target detection algorithm based on Faster R-CNN deep learning is more efficient than the traditional algorithm,but it can only be used in the target detection with large data set,while the insulator defect image data is less.At the same time,the network structure of deep learning fault detection is also the bottleneck of its detection accuracy development.(1)To solve the problem of insulator data shortage,which can not be trained directly by deep learning algorithm,this paper proposes two innovative data expansion methods: first,based on the traditional expansion method,the mask data expansion method is used to remove background interference,and the scope of network learning features is accurate.Then,the improved counter generated network(DCGAN)is used to expand data and obtain diversity data samples are used to enrich data sets.Finally,the comparison experiment shows that the average accuracy(mAP)of the traditional method is 69.6%,that of the mask method is 87.5%,and that of the DCGAN method is 90.6%,which improves the detection performance by 30.2%.It is verified that the proposed expansion method makes the deep learning algorithm applied to small-scale data sets,and effectively improves the accuracy of fault detection.(2)In order to further improve the detection performance and achieve the optimal detection effect,the Faster R-CNN algorithm is optimized from three aspects: adjusting the pooling mode,adding de-convolution layer and adjusting the dimension,and finally the optimal network structure is determined.Among them,adjusting the weight of maximum pooling and average pooling,increasing the proportion of global features in network training,adding de-convolution,the convolution dimension can adjust the number of extracted features,combining with the data expansion method to train the network structure,finally determining the optimal improvement scheme,and increasing the map by 2.7%.The improved model improves the accuracy of deep learning algorithm for small-scale image detection and classification,and verifies the effectiveness of the proposed method.(3)A software interface based on Faster R-CNN algorithm is designed and developed,which can detect insulator fault easily through interactive interface.