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

Abstract

Addressing the long-tail problem (LTP) is critical when applying deep learning (DL) to ultrasonic testing, as defective samples often lead to poor testing performance. This study addresses the LTP in stress–strain curve prediction using ultrasound by applying a Value Imputation and Mask Estimation (VIME)–based self-supervised learning (SSL) framework. Using 816 aluminum alloy samples, including low yield strength (YS) cases (100–200 MPa) that trigger LTP, the baseline model performed well overall but degraded sharply on LTP data (mean absolute percentage error (MAPE): 10% for non-LTP vs. 26% for LTP). VIME-SSL reduced the MAPE to 9.4% and 21%, respectively, with greater relative improvement for LTP cases. Notably, frequency-domain signals containing fundamental and second harmonic components were found to be especially effective for VIME-SSL in addressing the LTP. This finding was substantiated by separate ultrasonic measurements of attenuation and nonlinearity. Overall, this study demonstrates VIME-SSL as a promising approach for improving DL-based ultrasonic testing on rare or anomalous samples.