Gas pipeline defect detection based on improved deep learning approach

Zhang, Ting; Ma, Cong; Liu, Zhaoying; ur Rehman, Sadaqat; Li, Yujian; Saraee, Mohamad

Abstract

The working conditions of gas pipelines directly impact urban populations and factory operations. However, accurate and rapid detection of gas pipeline defects is challenging. To improve the accuracy of gas pipeline defect detection, we propose an improved RefineDet (Im-RefineDet) for gas pipeline defect detection, in which the improvement is carried out from the backbone network and the detection head. Specifically, to extract richer features, we design an improved CrossFormer as the backbone network. It first adopts a small convolutional cross-scale embedding layer to perform convolution, and then uses stripe window self-attention in vertical and horizontal directions in sequence to extract different features. In the detection head, we present a Double Attention Decouple Head (DADH) for classification and localization, enabling the model to perform independent optimization of the two branches. DADH employs spatial-aware and scale-aware attention to acquire multi-scale features, subsequently conducting classification and localization separately to derive final detection outcomes. Additionally, we apply channel pruning to the model to achieve a lightweight design, improving computational efficiency without significantly compromising detection performance. Experimental results, derived from an in-house developed gas pipeline defect image dataset, as well as two publicly available datasets — the NEU-DET dataset and the PCB dataset — demonstrate the effectiveness of the proposed Im-RefineDet. These results highlight its superior performance compared to state-of-the-art methods, further validating its robustness and adaptability across diverse scenarios. Specifically, the model achieves the mean Average Precision (mAP) of 92.6% on the gas pipeline defect image dataset, 77.8% on the NEU-DET dataset, and 99.2% on the PCB defect detection dataset.

Citation

Zhang, T., Ma, C., Liu, Z., ur Rehman, S., Li, Y., & Saraee, M. (2025). Gas pipeline defect detection based on improved deep learning approach. Expert systems with applications, 267, Article 126212. https://doi.org/10.1016/j.eswa.2024.126212

• Journal Article Type: Article
• Acceptance Date: Dec 16, 2024
• Publication Date: Apr 1, 2025
• Deposit Date: Dec 28, 2024
• Publicly Available Date: Jan 3, 2025
• Journal: Expert Systems with Applications
• Print ISSN: 0957-4174
• Publisher: Elsevier
• Peer Reviewed: Peer Reviewed
• Volume: 267
• Article Number: 126212
• DOI: https://doi.org/10.1016/j.eswa.2024.126212