Self-supervised learning-aided ultrasonic testing for overcoming long-tail problems in stress-strain curve prediction

Area of Science:

• Materials Science
• Non-Destructive Testing
• Machine Learning

Background:

• Deep learning (DL) in ultrasonic testing faces challenges with the long-tail problem (LTP), where rare defective samples degrade performance.
• Accurate prediction of material properties like stress-strain curves is crucial for quality control.

Purpose of the Study:

• To address the LTP in ultrasonic-based stress-strain curve prediction using a novel self-supervised learning (SSL) framework.
• To evaluate the effectiveness of the Value Imputation and Mask Estimation (VIME) approach in improving DL model performance on rare samples.

Main Methods:

• Implemented a VIME-based SSL framework for ultrasonic data.
• Utilized 816 aluminum alloy samples, including low yield strength (YS) cases triggering LTP.
• Analyzed frequency-domain signals, fundamental and second harmonic components, attenuation, and nonlinearity.

Main Results:

• The VIME-SSL framework significantly reduced the mean absolute percentage error (MAPE) for LTP data (from 26% to 21%) compared to the baseline model.
• The baseline model showed a substantial performance drop on LTP data (10% MAPE for non-LTP vs. 26% for LTP).
• Frequency-domain features, particularly fundamental and second harmonic components, proved effective for VIME-SSL in handling LTP.

Conclusions:

• VIME-based SSL is a promising approach for enhancing DL-based ultrasonic testing, especially for rare or anomalous material samples.
• The framework offers improved accuracy and robustness in predicting stress-strain curves, overcoming limitations posed by the long-tail problem.
• Understanding signal components like harmonics is key to optimizing SSL for challenging ultrasonic testing scenarios.