Research On Power Line Security Detection Algorithm Based On Cascaded Deep Learning

With the upgrading of national industry and the acceleration of smart grid construction,the detection of overhead transmission lines is an important part of the smart grid,and its detection quality is related to the intelligence degree of smart grid.With the development of UAV,robot industry,the precipitation of related technologies,high efficiency and safety and other performance of UAV has gradually become the main tool for power line detection.The detection tool is used to collect data on elevated power lines,the image data is detected by human eye to realize offline detection power lines,but manual detection has a low detection speed,low detection accuracy,detection accuracy is unstable,resulting in a certaion amount of missed detection errors.The rapid development of artificial intelligence has exceeded the precision of human eye detection in image classification,target detection and image segmentation.Compared with the traditional inspection method,the digital image processing of deep learning has the advantages of low cost,high precision and high detection stability.This dissertation first introduces the knowledge of deep learning,analyzes the power inspection data,briefly explains the detection difficulties faced in the detection: the background is complex and diverse,the pixel area of the detection object is too small,the scale of the detection object changes greatly,the focus of the detection object is not clear.In order to ensure the detection accuracy of bird’s nest,insulator,shock-proof hammer,pin and so on,a multi-stage target detection method is proposed,the improvement of Faster RCNN is used to achieve the detection of global critical areas in the first stage,the improvement of the Retina Net network is used to achieve the key area pins,insulators missing detection in the second stage,through the improvement of Ghost Net network for the classification of shock hammers,Thus in the power line detection experiment,the insulator positioning accuracy is 98.4%,the detection accuracy is 99.1%,the bird’s nest detection accuracy is89.2%,the positioning accuracy of the shock hammer is 97.6%,the detection accuracy is99.7%,the pin positioning accuracy is 86.1%,the detection accuracy is 96.7%,and the accuracy remains stable in the actual use of the field.The main work of this article is:1)Detection scheme and multi-level target area regression constraint method are proposed.This paper analyzes the factors that affect the accuracy of the detection object in the power line detection data,and puts forward a method of power line detection based on cascading deep learning.This data of power line labeling is analyzed,and a multi-level target area regression constraint method is proposed.The aspect ratio and scale of the detection box are redesigned by means of K-means based unsupervised clustering.In order to realize the Faster RCNN network in different feature layers using different Anchor sizes,aspect ratio,reducing the difficulty of RPN regression in the network.2)A key area detection algorithm for power lines based on improving the Faster RCNN network is designed.According to the detection problem analyzed above,a Faster RCNN network based on SCNet network is proposed,by using the multi-layer features in the feature pyramid for border regression prediction,the dissertation get good prediction results in a smaller size shock hammer,the feature pyramid network contains the second layer of feature layer to the sixth layer of feature layer,thereby reducing the loss of small target information,and in order to increase the sensory field of the same layer of data,the third layer of the feature layer can be deformed convolution operation.The efficient channel attention mechanism of ECANet is used in the classification network of RPN and regression network,which reduces the global channel mixing to achieve local channel mixing,enhances the aggregation ability of network characteristics,and finally uses Soft-NMS to handle the prediction box,thus improving the stability of network training,reducing the leakage of objects with overlapping frames,and using Ro I Align feature pooling to reduce computational errors.3)An insulator and pin defect detection algorithm based on improving the Retina Net network is designed.The key areas of insulators and pins detected in the target detection above are secontested to determine the location of defects,mainly using the improved Retina Net algorithm,which takes Rep VGG network as the main network for the target detection,while adding ASPP and CBAM channel spatial attention mechanism in the network,in order to achieving the expansion of the network’s sensory wild and enhance the network’s feature extraction capacity,improving detection accuracy while reducing the network detection time.4)A shock hammer classification algorithm based on the Ghost Net network is designed.This dissertation analyzes the image characteristics of shock hammer after cropping,and adopts Ghost Net as the basic network of shock hammer detection algorithm based on time and performance considerations.By increasing the mixed depth reflow,channel spatial attention mechanism module,and residual Ghost module,the information transmission ability of the model is enhanced,the size of the sensory field is increased,and the classification accuracy of the network is improved.At the end the classification mechanism of the network is verified by Grad-CAM.