Defect intelligent detection for pipeline girth welds based on improved YOLOv5 model

Girth welds are critical components of oil and gas pipelines, significantly contributing to the safety of natural gas transportation. Currently, X-ray digital radiography (DR) technology has become indispensable for quality inspection of girth welds. However, existing manual visual inspection methods struggle with achieving high-precision automatic identification of various sizes and types of weld defect. To address this challenge, this paper proposes a novel and convenient defect intelligent detection method for DR images based on deep learning approach. Firstly, we implement the image augmentation and enhancement strategies to construct a high-quality DR image dataset. Inspired by the experienced workers' visual inspection mechanism, we develop an improved YOLOv5 network architecture, incorporating the C2F module to learn the feature distributions of multiple typical defects. This model also employs multi-scale feature fusion techniques to effectively identify varying defect sizes and types. Additionally, visualisation technique is integrated to enhance interpretability and ease further inspections. Finally, the detection effect of the improved model is objectively evaluated through experiments such as ablation experiments and a comparison of mainstream methods. The results show that our improved YOLOv5 achieves a highest mAP value of 93.7% for detecting various defects, which can effectively improve inspection efficiency and promote the development of X-ray automation detection.