Intelligent identification of girth welds defects in pipelines using neural networks with attention modules

Abstract

Girth weld defects (crack, lack of penetration, lack of fusion, and edge nibbling) can cause pipeline cracking failure accidents. Internal magnetic flux leakage (MFL) detection can successfully identify pipeline defects, while the intelligent identification of MFL signals based on deep learning can promote the accurate determination of pipeline girth weld defects. Although the YOLOv5 model can effectively identify abnormal image objects, it exhibits no attention preference during the feature extraction process, proving insufficient for small objects. This study targeted minor defects in the girth weld of the pipeline and used the Convolutional Block Attention Module (CBAM) to optimize the YOLOv5 network model structure, increasing detection network attention preference toward extracting small-target defect signals. The CBAM+YOLOv5 model improved the detection accuracy of the MFL signal of the girth weld in the pipeline from 89.33% to 98.11% and correctly identified and classified the MFL signal of the pipeline girth weld with 85% confidence, with minor anomalies. The CBAM+YOLOv5 model effectively improved the identification accuracy of the girth weld defect signal in the pipeline, providing technical support for safety grade assessment and excavation verification.