Transmission Line Defect Detection Algorithm Based On Sample Enhancement And Deep Convolutional Neural Network Technology

As the core component of power transmission networks,regular inspections of power transmission lines are particularly important.With the development and application of intelligent inspection technology,the traditional manual inspection method has gradually been replaced due to its low efficiency,high risk factor and high resource consumption.However,due to the limitations of intelligent power line defect detection technology including limited applicable area,insufficient detection accuracy,and high detection cost,the current deepening application of inspection technology still has many obstacles.This thesis has carried out a systematic study on intelligent transmission line inspection based on sample enhancement technology and deep convolutional neural network,aimed at providing a feasible technical solution for solving the above limitations,improving the application of the intelligent inspection technology,and supporting the final realization of China’s smart grid detection.The main work of this research includes:(1)We build a typical defect set for inspection of transmission lines.In order to improve the target recognition breadth of the defect detection model,this thesis comprehensively analyzes,sorts out and classifies the typical defects of the transmission lines starting from the core components so as to shoot the original defect images of the transmission line through data collection technology in a targeted manner and further make up for the insufficient or missing limitations of some defect samples.On this basis,a small sample initial data set is obtained by selecting pictures with higher definition at different times,locations,and angles.In view of the insufficient number of samples in the initial data set,this thesis first uses the Labellmg software to manually label the data set,and then expands the labeled data.Finally,a total of 54,000 sheets containing 9 categories and 36 sub-categories and a typical defect set of transmission lines stored in the VOC2007 data set obtained.(2)Based on the deep convolutional neural network,we build and train a typical defect detection model for inspection of transmission lines.In order to reduce the detection cost and improve the detection accuracy,this thesis proposes a transmission line defect detection network that integrates single-stage and dual-stage.The network integrates RPN network,FPN algorithm and SSD algorithm,and mainly includes anchor frame refinement module and transmission connection structure target detection module,SEnet module.Among them,the anchor frame refinement module is based on VGG-16,which is mainly used to refine the anchor frame,including candidate frame position and size adjustment,negative candidate frame removal,etc.The transmission connection structure is used to output the low-level output of the anchor frame refinement module.Features are fused into high-level features and converted into input features of the target detection module.The target detection module is used to return to a more accurate object position.It is based on the output characteristics of the transmission connection structure and uses the refined anchor generated by the anchor frame refinement module and the box as the input.The SENet attention module is used to automatically obtain the importance of each feature channel improve useful features and suppress features that are not very useful for the current task.At the same time,this thesis also uses network pruning and knowledge distillation algorithms to compress the model lossless,so as to achieve the purpose of effectively reducing the amount of model parameters and calculations.In the model training,this study introduces training strategies such as variable learning rate,label smoothing,and hard case mining.The training results show that the use of the above strategies effectively improves the model training effect.(3)We conduct and analyze typical defect detection experiments for inspection of transmission lines.To verify the effectiveness of the proposed method,the self-built defect set is used to conduct targeted comparison experiments before and after sample enhancement,test network compression effect experiments,and defect detection network recognition experiments.Experimental results show that after sample enhancement,the detection network has a significant improvement in detection accuracy for a small number of defect types in the original sample.The original detection network is pruned through the NS pruning method and the knowledge distillation algorithm based on the attention mechanism is adopted.After the performance of the network is improved,on the one hand,the amount of parameters and calculations of the detection network are significantly reduced.On the other hand,the network performance is not reduced but slightly improved.Finally,after a horizontal comparison with Faster R-CNN and YOLO target detection networks,the transmission line defect detection network that integrates single-stage and dual-stage has achieved the best accuracy performance,with m AP reaching 74%.At the same time,in the detection of 36 typical defects of transmission lines,the AP value is above 60%.Therefore,it can be considered that this study has successfully constructed a detection network based on the end-to-end model to achieve high-precision,low-cost,and automatic detection of multiple types of defects in transmission lines.