Predicting local material thickness from steady-state ultrasonic wavefield measurements using a convolutional neural...

Area of Science:

• Materials Science and Engineering
• Non-Destructive Evaluation (NDE)
• Ultrasonic Testing
• Artificial Intelligence in Engineering

Background:

• Acoustic Steady-State Excitation Spatial Spectroscopy (ASSESS) is a full-field ultrasonic NDE technique for defect characterization in plate-like structures.
• Traditional ASSESS data processing relies on complex wavenumber domain analysis.
• Accurate local thickness prediction is crucial for reliable defect detection using ASSESS.

Purpose of the Study:

• To investigate the use of a Convolutional Neural Network (CNN) for direct local plate thickness prediction from ASSESS data.
• To evaluate the defect detection accuracy of CNN-based thickness predictions.
• To assess the CNN's ability to generalize to complex wavefield regions and experimental data.

Main Methods:

• A CNN was trained using simulated ASSESS data.
• The CNN directly predicts local plate thickness at each pixel of the wavefield measurement.
• Performance was evaluated based on defect detection accuracy for varying defect sizes and thickness reductions.

Main Results:

• CNN-based thickness prediction demonstrated improved defect detection accuracy for larger defects and greater thickness reductions.
• The CNN accurately predicted thickness in regions with complex or unknown Lamb wave dispersion relations.
• The CNN showed successful generalizability to experimental ASSESS data despite a fully simulated training dataset.

Conclusions:

• CNNs offer a viable alternative to traditional wavenumber analysis for ASSESS data processing.
• Direct thickness prediction using CNNs enhances defect localization and characterization capabilities.
• Simulated data training enables robust CNN performance on experimental ultrasonic NDE data.