Leveraging Segment Anything Model (SAM) for Weld Defect Detection in Industrial Ultrasonic B-Scan Images.

Area of Science:

• Engineering
• Artificial Intelligence
• Materials Science

Background:

• Automated ultrasonic testing (AUT) is crucial for oil and gas infrastructure integrity.
• Manual analysis of AUT data is time-consuming and requires specialized expertise.
• Artificial intelligence (AI) offers potential for automating AUT data interpretation, but reliability is a challenge.

Purpose of the Study:

• To develop an AI-assisted tool for weld defect detection in ultrasonic B-scan images using the Segment Anything Model (SAM).
• To evaluate the effectiveness of integrating B-scan image context into SAM for improved defect identification.
• To assess fine-tuning techniques for enhancing AI model performance in pipeline weld inspection.

Main Methods:

• Utilized a proprietary dataset of B-scan images from automated girth weld inspections.
• Applied the Segment Anything Model (SAM) with a promptable process to analyze ultrasonic images.
• Implemented vanilla and low-rank adaptation (LoRA) fine-tuning on SAM's image and mask decoders.
• Kept the SAM prompt encoder unchanged during fine-tuning.

Main Results:

• The AI-assisted tool successfully detected lack of fusion (LOF) defects in pipeline welds.
• Integration of B-scan context into SAM improved defect detection capabilities.
• Fine-tuning techniques, particularly LoRA, demonstrated enhanced performance over vanilla approaches.
• The developed method showed improved results compared to previous studies on the same dataset.

Conclusions:

• The AI-assisted tool leveraging SAM shows significant promise for automating weld defect detection in AUT data.
• Fine-tuning SAM with relevant industrial data and context is effective for improving performance.
• This approach offers a more reliable and efficient alternative to manual analysis for infrastructure inspection.