Ultrasonic Defect Characterization Using the Scattering Matrix: A Performance Comparison Study of Bayesian Inversion...

Area of Science:

• Ultrasonic Nondestructive Evaluation (NDE)
• Materials Science
• Signal Processing

Background:

• Accurate characterization of material defects using ultrasonic nondestructive evaluation (NDE) is crucial for quantitative assessment of defect type and geometry.
• Image-based defect characterization using ultrasonic arrays performs poorly for small defects comparable to the wavelength.
• Extracting the far-field scattering coefficient matrix offers an alternative approach for characterizing small defects.

Purpose of the Study:

• To investigate two distinct approaches for characterizing small surface-breaking notches using ultrasonic array data.
• To compare the performance of a Bayesian framework with a coherent noise model against a supervised machine learning (ML) schema.
• To evaluate the effectiveness of convolutional neural networks (CNNs) within the ML approach for defect characterization.

Main Methods:

• Development of a general coherent noise model for characterization within a Bayesian framework.
• Implementation of a supervised machine learning (ML) schema utilizing a scattering matrix database for training.
• Training and evaluation of ML models, specifically focusing on convolutional neural networks (CNNs).

Main Results:

• Convolutional neural networks (CNNs) demonstrated the highest characterization accuracy among the evaluated ML approaches.
• For favorably oriented notches, CNNs achieved characterization uncertainty comparable to the Bayesian approach.
• Both approaches showed varied performance for unfavorably oriented notches, with the ML approach exhibiting higher variance and lower biases.

Conclusions:

• The study highlights the potential of scattering matrix analysis combined with advanced ML techniques like CNNs for small defect characterization in NDE.
• CNNs offer a competitive alternative to traditional Bayesian methods, particularly for specific defect orientations.
• Further research is needed to optimize performance for challenging defect geometries and orientations.