Magnetic flux leakage defect size estimation method based on physics-informed neural network.

Area of Science:

• Engineering
• Materials Science
• Computer Science

Background:

• Magnetic flux leakage (MFL) is a key non-destructive testing method for pipeline integrity.
• Current MFL defect quantification relies heavily on data-driven approaches, potentially overlooking physical principles.
• Machine learning advancements offer new possibilities for MFL defect size estimation.

Purpose of the Study:

• To propose a novel physics-informed neural network (PINN) for accurate MFL defect size estimation.
• To integrate physical constraints, specifically the magnetic dipole model, into the neural network training process.
• To evaluate the performance of the PINN against purely data-driven methods.

Main Methods:

• Development of a physics-informed neural network incorporating magnetic dipole model constraints.
• Training the neural network using synthetic MFL data from virtual pipeline defect testing.
• Comparative analysis of the PINN against data-driven neural networks and support vector machines.

Main Results:

• The physics-informed approach demonstrated improved predictive accuracy in MFL defect size estimation.
• The method effectively mitigated physical violations often associated with purely data-driven models.
• Validation using synthetic data confirmed the efficacy of the proposed physics-informed strategy.

Conclusions:

• Physics-informed machine learning offers a robust framework for MFL defect size estimation.
• Integrating physical knowledge enhances the reliability and accuracy of non-destructive testing.
• The proposed PINN method provides a valuable advancement for pipeline integrity assessment.