Research On Pipeline Defect Detection Method Based On Deep Learning

Pipelines serve as a vital transportation infrastructure for energy resources,including oil and natural gas.However,with the ever-increasing pipeline length and project complexity,the potential risks associated with their operation also rise.Ensuring the safety of oil and gas pipelines,therefore,becomes critical to maintaining the uninterrupted transportation of energy resources in China.However,because the transportation pipeline is very long,the data collected is very large.Pipeline defect detection using the traditional manual interpretation method often faces challenges such as low efficiency,poor accuracy,missed inspections,and false inspections.However,with the significant advancements in computer vision in recent years,this thesis aims to investigate a depth learning-based approach for detecting defects in pipelines.The primary focus of this research encompasses the following:(1)Aiming at the problem of missing pipeline defect data set,this thesis proposes a method of graying and image enhancement of magnetic flux leakage signal to make data set.Initially,the magnetic flux leakage signal is processed using image processing techniques to produce a curve,a pseudo-color image,and a grayscale image.Based on the specific characteristics of pipeline defects and their background in the images,the most appropriate grayscale image is selected as the original image.Subsequently,in order to augment the defect image and enhance the diversity of the pipeline defect detection dataset,we propose various image enhancement techniques,such as flipping and rotating the defect image,as well as applying noise reduction algorithms.Finally,the image is labeled according to the pipeline defect features,and a pipeline defect detection data set with complete features is produced.(2)Aiming at the problems of high complexity of detection network,lack of defect recognition ability and weak generalization in the current pipeline defect detection field,this thesis studies and improves on the basis of YOLOv4 in the target detection algorithm and the characteristics of pipeline defects.First of all,due to the large amount of data of pipeline defects,the efficiency of detection is required to be very high.Therefore,the lightweight improvement of the backbone network CSPMarket53 of YOLOv4 is adopted to replace it with Mobile Netv2 network,thus reducing the complexity of the model and improving the reasoning speed;Secondly,in view of the insufficient detection ability of YOLOv4 network for small targets in the data set,it is proposed to add a small target detection layer in the fusion network PANET to improve the feature extraction ability of small targets;Then in order to increase the receptive field and enrich the pipeline defect feature information,the enhanced receptive field module RFB is introduced into the network;Finally,in order to obtain the key features of pipeline defects in the image and reduce the impact of background noise,CBAM attention mechanism is introduced into the network.The experiment shows that the improved YOLOv4 network model can quickly and effectively detect pipeline defects.(3)Aiming at the problems of insufficient small defect feature information in the current pipeline defect data set and high fusion of defect and background information,this thesis proposes to combine the super-resolution reconstruction network EDSR with the improved YOLOv4 detection network.First,in order to improve the detection efficiency,the pipeline defect is detected first,the low-quality detection image is extracted for super-resolution reconstruction,and then sent back to the detection sub-network;Then,by analyzing the proportion and size distribution of small targets in the data set,as well as the confidence distribution of small targets in the detection results,the basis for judging low quality detection is obtained;Finally,in order to reduce the computational complexity of the hyper-segmentation network and improve the reasoning speed,the low-quality image is overlapped and segmented,and the hyper-segmentation network EDSR is sent to enrich the defect feature information,reduce the fusion degree of the defect and the background,and then the target frame is returned to the original image after detection.The experimental results show that the fusion of super-resolution reconstruction network EDSR and the improved YOLOv4 pipeline defect detection algorithm effectively improves the recall rate of small defect targets and further improves the pipeline defect detection accuracy.This thesis contains 72 figures,6 tables and 83 references.