Research And Application Of Pipeline Defect Detection Method Based On Deep Learning

Resource transportation through pipelines is one of the modes of transportation,and now it has become the fifth largest transportation industry after railway,highway,waterway and air transportation.With the acceleration of social industrialization and the advancement of energy structure,the use of pipelines has greatly increased,which also makes the inspection and maintenance of pipelines more and more complicated.In order to alleviate the hidden dangers of pipeline safety,it is very important to effectively detect and analyze pipeline defects.With the continuous development of computer vision,the use of machine learning and image processing for detection and recognition has become the mainstream,but there are problems such as poor detection effect,slow detection speed,low precision and lack of robustness.Therefore,how to detect pipeline defects accurately and effectively has become a research hotspot.This thesis proposes a pipeline defect detection method based on deep learning.First,the defects in the pipeline image are classified and identified through the target detection network,and then the extracted defect area image is segmented through the semantic segmentation network.Finally,the pipeline defect detection system is designed and implemented.The main work of this thesis is as follows:1.An improved infrared defect detection network FIT-YOLOv5 is proposed.Based on YOLOv5 network,CBAM attention module is first introduced into the convolution layer to enhance the feature extraction capability of the backbone network;Add a small target detection layer to the Neck network,and improve the CSP module into a lightweight Ghost module to enhance the multi-scale feature fusion of the network and reduce the network computation;optimize the Io U loss function to EIo U to improve the prediction accuracy of the anchor frame,and accelerate the convergence of the network.The experimental results indicate that the m AP of FIT-YOLOv5 network increases from 92.1% to 93.8% compared with YOLOv5,and the model size is only 10.6M.2.An improved defect semantic segmentation network ESM-UNet is proposed.For the defect area image,improving the encoding part of UNet network to the Bneck module of Mobile Netv3,which reduces the network computation and improves the defect segmentation rate of the network.The Bneck＿MA module based on the mixed attention mechanism is designed to replace the Bneck module,strengthen the feature learning ability of the network,and improve the segmentation accuracy of the network.The experimental results indicate that the m Io U value and m PA value of ESM-UNet network for the segmentation of three types of defect targets reach 83.5% and 88.1%,which are improved compared with UNet network,and the edge contour of the segmented defect area is closer to the real value.3.Design and implement pipeline defect detection system.The main functions of the system include pipeline inspection,equipment monitoring and data management.The system is deployed on the cloud server to complete the tasks of image defect detection and data storage.After testing,the detection accuracy of the pipeline defect detection system is about 90%,the average detection time is 1.58 s,and all functions meet the actual requirements.