Abstract:

Aiming at the problems of automatic extraction of magnetic memory signal features and quantitative identification of defect levels in oil and gas pipelines， a multimodal fusion model was proposed combining residual neural network and graph neural network（ResGNNet）. The original magnetic memory signals of defects of different depths on L245N pipeline steels were collected by metal magnetic memory detector. In order to realize automatic feature extraction， the complete information of the original magnetic memory signals was retained， and the relationship among samples was taken into account. The original signals were converted into a node graph by K nearest neighbor-dynamic time warping， and the original signals were converted into a 2D image by Gram angle field. The designed graph neural network and residual neural network may automatically extract the embedded feature vectors of 1D signals and 2D images respectively. The multimodal embedded feature vectors were fused， weighted and screened by multi-head self-attention mechanism， and then input into the Softmax classification module to complete the defect level identification. The model verification results show that the accuracy of quantitative identification of pipeline defect levels reaches 93%.

Key words: oil and gas pipeline, metal magnetic memory technology, defect level, graph neural network, residual neural network

摘要：

针对油气管道磁记忆信号特征自动提取及缺陷等级定量识别难题，提出一种结合残差神经网络和图神经网络的多模态融合模型即ResGNNet模型。采用金属磁记忆检测仪采集L245N管线钢不同深度缺陷的磁记忆原始信号。为实现特征的自动提取，保留原始磁记忆信号的完整信息并考虑样本之间的相互关系，利用K近邻-动态时间规整将原始信号转换成节点图，并利用格拉姆角场将原始信号转换成二维图像。设计的图神经网络、残差神经网络可分别自动提取一维信号和二维图像的嵌入特征向量。融合多模态嵌入特征向量经多头自注意力机制加权筛选后，输入Softmax分级模块，完成缺陷等级识别。模型验证结果表明，管道缺陷等级定量识别的准确率达到93%。

关键词: 油气管道, 金属磁记忆技术, 缺陷等级, 图神经网络, 残差神经网络