一种面向管道缺陷检测的漏磁数据增强方法

MFL Data Augmentation Method for Pipeline Defect Detection

大规模标注数据集是深度学习模型应用的前提。由于油气管道的特殊性,短时间内较难建立具有良好标注信息的大规模数据集。文中提出一种基于漏磁场物理学模型的漏磁数据增强方法。首先,利用三维磁偶极子模型生成各种尺寸缺陷漏磁场法向分量、切向分量和周向分量的仿真数据,并与在役管道检测数据进行随机融合,实现缺陷样本数据的扩充。最后,基于只包含仿真缺陷的样本,对YOLOX-Tiny、-S和-M 3个版本分别进行了100轮训练,再利用真实缺陷样本进行测试。实验结果表明:所扩充的样本数据能够使模型收敛,其中Tiny和S版本的平均精度可稳定在40%左右,间接地说明了所扩充的样本与真实管道缺陷具有一定的相似性,所提数据增强方法具有一定的有效性。

Large-scale labeled dataset is the prerequisite for the application of deep learning models.Due to the particularity of oil and gas pipelines,it is hard to collect well-labeled defect sample and establish a large-scale dataset.Therefore,a data aug-mentation method based on the physics model of leakage magnetic field was proposed in this paper.Firstly,the magnetic dipole model was used to generate normal,tangential and circumferential MFL data of different defect size,and the generated simulation data was fused to the in-service pipeline inspection data on a random location so as to augment the dataset.At last,the Tiny,S and M version of YOLOX object detection network was trained for 100 epochs with the samples that only contain the simulation defects and was tested by the samples collected from in-service pipeline inspection data that just contain real defects.The experiment re-sults show that the expanded sample data can make the model converge,the AP performance of the Tiny and S version models can reach about 40%,which indirectly proves that the augmented dataset is similar to the real samples and the proposed augmentation method is valid in some degree.