Efficient Model Assisted Probability of Detection Estimations in Eddy Current NDT with ACA-SVD Based Forward Solver

Yang Bao, Minxuan Xu, Jiahao Qiu, Jiming Song2022 · DOI

TL;DR

This paper introduces a novel, efficient 3D eddy current non-destructive evaluation (ECNDE) forward solver that significantly speeds up the process of estimating the Probability of Detection (PoD) for flaws. By combining Adaptive Cross Approximation (ACA) with Singular Value Decomposition (SVD), the method drastically reduces computational resources needed for simulations, making it a powerful tool for NDT reliability studies.

What Problem Does This Paper Address?

Estimating the Probability of Detection (PoD) is crucial for understanding the reliability of non-destructive testing (NDT) systems. PoD quantifies the likelihood of detecting a flaw of a certain size. Traditionally, PoD estimation relies heavily on experimental data, which is time-consuming and expensive to collect in sufficient quantities. Model-assisted PoD (MAPoD) approaches aim to bridge this gap by using simulation models to generate the necessary data. However, existing simulation models, particularly those based on the Finite Element Method (FEM) or even standard Boundary Element Method (BEM), can be computationally intensive, especially for complex 3D scenarios and the large number of simulations required for robust PoD analysis. This paper addresses the need for a faster and more efficient simulation tool for MAPoD studies in eddy current NDT.

Methods

The core of this paper is the development of an ACA-SVD based BEM forward solver. Here's a breakdown of the methodology:

  • Boundary Element Method (BEM): The paper utilizes BEM for its advantage in discretizing only the surface of the object, leading to fewer unknowns compared to FEM. The Stratton-Chu formulation is employed for its ability to capture both tangential and normal components of electromagnetic fields.
  • Adaptive Cross Approximation (ACA): To combat the O(N²) complexity of standard BEM, ACA is used to approximate the dense impedance matrix. ACA compresses low-rank submatrices by representing them as a product of two smaller matrices (U and V).
  • Singular Value Decomposition (SVD): While ACA is effective, the resulting U and V matrices can still contain redundancies. SVD is applied as a post-compression technique to further optimize these matrices, removing redundancies and reducing storage requirements. This combination, ACA-SVD, is the novel contribution.
  • Model-Assisted PoD (MAPoD): The developed solver is integrated into a MAPoD framework. The simulation model generates flaw features (e.g., maximum impedance variations) based on input parameters, which are then used for PoD calculations.

Datasets

The paper validates its proposed solver using several benchmark cases:

  • Coil C27 above Conducting Plate B2: This case involves a coil with a finite cross-section placed above a conducting plate. Data from analytical, semi-analytical, and experimental sources are used for comparison.
  • Coil C5 above Conducting Plate B1: Another scenario with a different coil and plate configuration, also validated against existing methods.
  • TEAM Workshop Benchmark Problem 15: This case involves a coil scanning a thick plate with a surface slot. The solver's predictions are compared with experimental results from this established benchmark.

No new datasets were introduced; rather, existing experimental and analytical data were used to validate the proposed simulation model.

Key Insights and Ideas

  • Synergy of ACA and SVD: The combination of ACA for initial compression and SVD for further optimization of low-rank submatrices in the BEM impedance matrix is a key innovation. This dual approach significantly enhances both memory and computational efficiency.
  • Efficiency for PoD: The primary motivation is to accelerate PoD studies. The ACA-SVD BEM solver drastically reduces the time and memory needed to generate the large number of simulations required for accurate PoD estimation, making MAPoD more practical.
  • Robustness: The solver's predictions show excellent agreement with analytical, semi-analytical, and experimental results across various benchmark cases, demonstrating its accuracy and reliability.
  • Handling Uncertainties: The paper demonstrates how the solver can incorporate uncertain parameters, such as liftoff distance, to generate realistic flaw response data for PoD analysis.

Key Results

  • Significant Savings: The ACA-SVD based BEM solver offers substantial savings in memory and CPU time compared to traditional BEM and even ACA-based BEM. For instance, one benchmark showed savings of up to 75.1% in memory and 9.01% CPU time per iteration compared to ACA-based BEM.
  • High Accuracy: The predicted impedance variations from the ACA-SVD solver closely match experimental and analytical results, with relative differences generally below 1%.
  • PoD Analysis: The paper successfully applies the solver to estimate PoD for a rectangular surface flaw. Key PoD metrics (a50, A90, a90_95) were calculated, demonstrating the solver's utility in practical NDT reliability assessment.

Limitations and Future Work

  • Computational Burden for Multiple Uncertainties: The authors acknowledge that while the ACA-SVD solver is efficient, calculating a large number of flaw responses when dealing with multiple uncertain parameters (e.g., flaw length, height, width, probe location simultaneously) might still be computationally demanding.
  • Metamodel/Surrogate Model: For future work, the authors suggest the development of metamodels or surrogate models to further alleviate the computational burden when faced with a high number of uncertain parameters. This would build upon the fast solver developed in this paper.

Summary updated

BibTeX
@misc{bao2022efficient,
  author = {Yang Bao and Minxuan Xu and Jiahao Qiu and Jiming Song},
  title = {{Efficient Model Assisted Probability of Detection Estimations in Eddy Current NDT with ACA-SVD Based Forward Solver}},
  year = {2022},
  doi = {10.3390/s22197625},
}
Figure