Research On Quantitative Method For Pipeline Defect Recognition Based On Machine Learning

Pipeline is the most important way to transport oil and gas energy with its advantages of long laying distance and large transmission capacity,However,the special environment of buried pipelines and the corrosiveness of the transported medium make it very easy to have defects such as metal loss,deformation and depression.Regular testing of pipelines can reduce the risk of disasters and accidents and ensure safe operation of pipelines.Magnetic Flux Leakage is the most widely used in-line inspection method for oil and gas pipelines,using an in-pipe detector to saturate the pipe with magnetization,a magnetic circuit is formed with a magnetic yoke,a steel brush,a permanent magnet and a pipe,and detection is completed by a Hall sensor sensitive to leakage flux at the defect.At present,off-line manual discrimination is still used to analyze pipeline magnetic leakage detection data,which is highly susceptible to the subjective experience of data analysts,and the data analysis efficiency is low,making it difficult to meet the demand of the actual pipeline’s great data volume detection.And,accurate quantification of defects is an important prerequisite for assessing the extent of pipeline damage.Therefore,an efficient,accurate and quantifiable intelligent identification method can significantly improve the efficiency of pipeline defect data analysis,reduce the influence of subjective factors and guarantee the reliability of detection,which is of great significance to the intelligent construction of pipeline integrity management.In this thesis,an intelligent processing method of defect leakage magnetic data based on machine learning is proposed for the problem of large amount of leakage magnetic data and difficult quantification of defect size for long-distance oil and gas pipeline inspection.Starting from the characteristics of leakage magnetic detection signal,16 characteristic parameters related to defect size are defined by analyzing the spatial leakage magnetic field distribution of defects of different sizes,and combined with the manual production of defects and the identification of the position of the defect item signal in the detection data of the drawing experiment,the automatic identification and extraction of the characteristic parameters of the defect signal is realized,which greatly simplifies the data level and completes the establishment of the characteristic data set of pipeline defects;Pre-processing of the data set from the matrix dimension,analysis of the characteristics of the leakage field for defects of different sizes and establishment of the correspondence between data labels and defect sizes;Design a quantitative mathematical model of neural network with convolutional layer,pooling layer,fully connected layer and output layer,the Re Lu activation function,MSE loss function and Adam adaptive learning rate algorithm are used to optimize the training weights in the model to minimize the error and output the training optimal parameters,realize intelligent identification and dimensional quantification of pipe defects.Build an experimental platform for intelligent quantification of defects,and conduct experiments for quantification of defect sizes,validate the effectiveness of the proposed intelligent identification quantification method for magnetic leakage data.The experimental results show that the proposed method has good quantification effect on the defect size of pipes with small diameter,small data dimension and large aspect ratio,and the quantification error of the size is below 3%.For pipelines with large caliber,large data dimensions,and close aspect ratios,the length and depth of defects have good quantification effects.The quantization error for length is below 5%,and the quantization error for defect depth is below 8%.Meet the error needs of practical industrial inspection.The method can provide a reference for the engineering application of defect quantification for pipeline leakage magnetic detection.