Micro-Leakage Image Recognition Method for Internal Detection in Small, Buried Gas Pipelines.

Area of Science:

• Pipeline integrity
• Artificial intelligence in engineering
• Non-destructive testing

Background:

• Internal detection of DN100 buried gas pipelines faces challenges with limited and difficult-to-identify microleakage image samples.
• Existing methods struggle with sample diversity and accurate recognition of small-scale leaks.

Purpose of the Study:

• To propose a novel recognition method for microleakage images captured by internal pipeline detection robots.
• To enhance the diversity and identification accuracy of microleakage image samples.

Main Methods:

• Employed non-generative data augmentation to expand existing microleakage images.
• Utilized Deep Convolutional Wasserstein Generative Adversarial Networks (DCWGANs) for generating diverse microleakage image features.
• Integrated a bi-directional feature pyramid network (BiFPN) with YOLOv5 for improved deep feature retention.
• Developed a small target detection layer within YOLOv5 to preserve shallow feature information for small leak recognition.

Main Results:

• Achieved a precision of 95.04% for microleak identification.
• Obtained a recall rate of 94.86%.
• Reached a mean Average Precision (mAP) value of 96.31%.
• Successfully identified leaks as small as 1 mm.

Conclusions:

• The proposed method significantly improves the accuracy and reliability of microleak detection in gas pipelines.
• The combination of generative adversarial networks and enhanced YOLOv5 architecture effectively addresses the challenge of limited and small-scale defect samples.
• This approach offers a promising solution for enhancing the safety and maintenance of buried gas pipeline infrastructure.