Diagnosis and Recognition of Pipeline Damage Defects Based on Improved Convolutional Neural Network

Abstract

Oil and gas resources are important strategic materials and economic resources. With the increase of pipeline mechanical failure caused by external stress changes have occurred frequently. Among them, mechanical failure and corrosion failure can cause pipeline rupture, explosion, property loss, and casualties. Casualties caused by pipeline damage account for 22% to 67% of total casualties. Therefore, it is particularly important to conduct safety inspections on pipelines. At present, there are many pipeline damage detection technologies. Just like pipeline magnetic flux leakage (MFL) detection technology is widely used, which has the advantages of fast speed and high efficiency and is suitable for defect detection of ferromagnetic metal material of pipeline. However, most pipelines are located in harsh outdoor environments, and there are various corrosion and mechanical damage defects, including axial damage and circumferential damage, resulting in large differences in magnetic leakage signals. Meanwhile, there are various interferenoes in the signal, so the current methods have low accuracy in the diagnosis and prediction of defects, and the recognition accuracy is maintained at about 86%. So, the neural network algorithm is designed to identify and analyze the inspection signal. Through the use of a multi-layer neural network for training and learning, the recognition accuracy and reliability of pipeline defects are improved. The main contributions are as follows: Firstly, the mechanism of MFL detection of pipeline defects is introduced. Through theoretical analysis, it is known that the change of lift value in the device will affect the change of MFL intensity, and an experimental device is built to complete pipeline defect detection and signal collection. Secondly, VMD algorithm is used to complete the preprocessing of pipeline magnetic flux leakage signal, and the processed data is smooth and stable without over envelope and under envelope and so on. Finally, an improved convolutional neural network structure (I-CNN) is studied. And the Dropout layer is inserted into the network structure of the convolutional neural network to optimize the algorithm. The simulation results show that the recognition accuracy of I-CNN algorithm is 99.19%, improved by 15.7%. Meanwhile, the test recognition of the I-CNN algorithm is performed on 300 sets of data signals, and the recognition accuracy is lower than the training accuracy, with a recognition test accuracy of 97.38%, but also higher than that of the CNN algorithm. Here, the method not only reduces the labor intensity and human error of manual labeling but also improves the accuracy of the network and the generalization ability of recognition learning.