Deep learning interprets measurements of micro damage

Guided ultrasonic waves help researchers detect defects in industrial equipment, such as planes and high-speed rails, before accidents occur. However, conventional linear ultrasonic techniques aren’t sensitive enough to resolve early-stage micro damages in materials. Nonlinear ultrasonic techniques (NUT), which examine the nonlinear interaction between a material and an ultrasonic wave, are gaining popularity due to their sensitivity to these micro damages.

But the application of NUT is impeded by the convoluted relationship between these nonlinear ultrasonic signals and the multi-dimensional features of micro damages. This relationship is too complex to understand with physics-based models, so Liu et al. proposed an approach that uses deep learning instead. Their approach, called nonlinearity-aware discrete wavelet transform-bidirectional long short-term memory (DWT-BiLSTM) network, can establish the relationship between nonlinear ultrasonic signals and micro damage characteristics.

This nonlinearity-aware deep learning network was able to extract damage features from a robust dataset of ultrasonic signals containing both size and location information. Specifically, the framework could accurately predict the length and location of closed cracks, an important type of micro damage.

“Our work paves a promising and practical way to promote the transformation of NUT from qualitative analysis to accurate and efficient quantitative prediction,” said author Lishuai Liu.

The authors believe their approach could be easily extended to other types of micro damages. This would allow researchers to obtain comprehensive information about micro damages using NUT without creating a physics-based model.

“The next step is to apply data-driven NUT to predict the remaining useful life of degrading equipment,” said author Yanxun Xiang.

Source: “Deep learning-based solvability of underdetermined inverse problems in nonlinear ultrasonic characterization of micro damages,” by Lishuai Liu, Di Sun, Yanxun Xiang, and Fu-Zhen Xuan, Journal of Applied Physics (2022). The article can be accessed at https://doi.org/10.1063/5.0107205 .