Research On Identification Method Of Pipeline Defect Magnetic Leakage Data Based On Deep Learning

Long distance transportation pipeline is an indispen sable means of transportation in petroleum,chemical,thermal and other industries,and has a vital impact on the national economy and ecological environment.Therefore,regular inspection and maintenance of the pipeline is very necessary.As one of the mainstream nondestructive testing methods for pipelines,the magnetic leakage internal testing technology provides an important guarantee for oil,natural gas and other energy transportation.The magnetic leakage information is collected through the detector in the pipeline,and the type and location information of defects in the pipeline can be judged by analyzing the magnetic leakage signal.In order to effectively evaluate the harm degree of defects,further quantitative identification of the specific size of defects is the focus of current research.However,at present,only through professional personnel and manual interpretation with experience,there are problems such as false detection and missing detection,which can no longer meet the needs of the project.Deep learning technology has developed rapidly in recent years and has shown amazing capabilities in many fields.Compared with traditional recognition methods,it has a better performance in intelligent identification of defect size of pipeline magnetic signals.This thesis introduces a convolutional neural network based on multi-task learning,which is implemented by deep learning framework Pytorch.The network model contains three subtasks and each task is composed of shallow convolutional neural networks,which are used to identify the three-dimensional dimensions of defects respectively.With the circumferential,axial and radial magnetic flux leakage data of pipeline defects as the input of the network model,the three sub-networks are trained to simultaneously predict the length,width and depth of the defects.The total loss function of the multi-task learning network varies according to the number of training rounds,and the relative difficulty of task training is also different.Using the uncertainty of homoscedasticity as the benchmark to set the weight of the loss function reasonably,the identification results are compared with the model without weight,and the identification accuracy is significantly improved,indicating that the performance of each task can be effectively balanced.Finally,different optimization algorithms and learning rate experiments were set to realize the adjustment and optimization of network super parameters,and further improve the accuracy of network recognition.Through the comparison experiment between the single task network and the multi-task network,the defect length identification error,width identification error and depth identification error of the optimized multi-task network for 15 defect samples are between 0-10 mm,0-20 mm and 0-6mm,respectively.The average error of single task network is reduced by 0.7144 mm,0.9081 mm and0.0358 mm,respectively.Experiments show that the convolutional neural network constructed by multitask learning has certain advantages in recognition accuracy.