Purpose The purpose of this study is to propose an automated pipeline integrity monitoring framework. Thermography is presented as the primary inspection and monitoring tool at critical locations in gas pipelines. Design/Methodology A custom laboratory-scale experimental setup was developed to induce commonly observed defects on the pipeline surface, such as dents, cracks, holes, and scratches. A dataset of 1,451 annotated samples was created by capturing the thermographic images corresponding to the different defects. A two-stage deep learning framework was designed. In the first stage, a Residual Network (ResNet-18) was implemented for defect classification. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to visualize and interpret the image regions that influenced the model’s classification. In the second stage, the You Only Look Once (YOLO) model was applied to detect and locate defects. Findings ResNet-18 combined with Grad-CAM provided accurate classification and interpretability. Grad-CAM highlighted the defect zones effectively. YOLO demonstrated a strong ability to detect and localize surface-level defects through bounding boxes and confidence scores. Some misclassification occurred between visually similar defects, such as dents and cracks, or cracks and scratches. Thermographic imaging produced stronger and clearer fault signatures compared with conventional camera images. Originality/Value This work combines thermography with deep learning for pipeline monitoring. The results from the lab-scale setup validate the feasibility of the approach as a proof of concept for using thermography as a reliable monitoring tool. The study highlights the potential to extend this framework for industrial-scale pipelines by integrating larger datasets, diverse operating conditions, and real-time implementation.
Muthalif et al. (Thu,) studied this question.