Domain Adapted Deep-Learning for Improved Ultrasonic Crack Characterization Using Limited Experimental Data.

Area of Science:

• Materials Science
• Non-Destructive Evaluation
• Artificial Intelligence

Background:

• Deep learning offers automated and accurate ultrasonic crack characterization.
• Simulating data is crucial for nondestructive evaluation (NDE) but introduces discrepancies between simulated and real-world data.
• Applying models trained on simulated data to experimental data leads to inaccuracies due to distribution shifts.

Purpose of the Study:

• To address the challenge of applying deep learning models trained on simulated ultrasonic data to experimental measurements.
• To investigate the effectiveness of domain adaptation (DA) techniques in improving crack characterization accuracy.
• To evaluate different DA methods for sizing surface-breaking defects in in-line pipe inspection using a convolutional neural network (CNN).

Main Methods:

• A CNN was employed for predicting the depth of surface-breaking defects.
• Three domain adaptation methods were compared against two non-DA baseline methods.
• Performance was evaluated by sizing 15 experimental notches with varying lengths (1-5 mm) and angles (up to 20°).
• Experimental training sets were incrementally built from 1 to 15 notches.

Main Results:

• The adversarial domain adaptation method demonstrated superior performance in utilizing limited experimental data.
• With only three experimental notches, the adversarial DA method achieved a root-mean-square error (RMSE) of 0.5 ± 0.037 mm.
• In contrast, methods using only experimental data yielded an RMSE of 1.5 ± 0.13 mm, and methods using only simulated data resulted in an RMSE of 0.64 ± 0.044 mm.

Conclusions:

• Adversarial domain adaptation is highly effective for ultrasonic crack characterization with limited experimental data.
• DA significantly enhances the accuracy of deep learning models in NDE applications, bridging the gap between simulated and experimental data.
• This approach offers a robust solution for accurate defect sizing in real-world inspection scenarios.