Probabilistic machine learning aided acoustic emission source localization in mode-I fracture test of multi-layer engineered wood
Xiangdong He, Peng Zhang, Keping Zhang, Yuning Wu, Xuan Zhu · Mechanical Systems and Signal Processing · 2026
The mode-I fracture test is an important technique to evaluate materials’ fracture toughness and behaviors, where tracking the crack front is essential. This is usually done by visual inspection or computer vision, which could be problematic due to the large process zone and the difficulty of determining the onset of material cracking. Source localization via acoustic emission (AE) offers an alternative solution to this challenge. Locating AE sources in homogeneous materials is intuitively straightforward by leveraging phase differences from the measured time of arrival and potentially angle- or path-dependent velocity models. But it becomes complicated and intractable once the material heterogeneity comes into play. This is especially true for multilayer engineered wood, like laminated veneer lumber (LVL), where natural and manufacturing defects (knots, voids, discontinuity, and non-uniform textures) impose great challenges for 2D AE source localization. To cope with this problem, this study proposes probabilistic machine learning models, namely the Student-t process regression and the Gaussian process regression, for the 2D AE source localization on two LVL samples. To enhance prediction performance, both the multi-output Student-t and Gaussian process regression are developed. These probabilistic machine learning models are first trained and verified via pencil lead break tests and then applied to locate the crack tips in two mode-I fracture tests. Meanwhile, the Gaussian mixture models, along with Laplacian scores, are considered to identify AE events related to crack tip locations. The results suggest that the Student-t process can outperform the Gaussian process both in pencil lead break tests and mode-I fracture tests. The multi-output models can further enhance the prediction performance compared with the single-output processes. The multi-output Student-t process and the multi-output Gaussian process can support similar prediction performance when the degree of freedom of the Student-t process is large enough, and therefore, both are recommended for AE source localization in complex heterogeneous media like engineered wood.