Session: 06-02: Machine Learning and Statistical Methods in NDE II

Paper Number: 147604

147604 - Acoustic Emission Source Localization on a Laminated Veneer Lumber Plate by Probabilistic Machine Learning 

Abstract: 

Acoustic emission testing is a widely used tool to locate sources both in laboratories and on-sites, which is unavailable for other Nondestructive testing methods. Unlike the isotropic materials AE localization, where wave velocity can be simplified as a constant and free of propagation angle thus the analytical solutions exist, AE localization in anisotropic materials is much more complicated due to angle dependent velocity and heterogeneity. Moreover, for natural materials, like wood, the knots and defects make the AE source localization even more intractable. The Laminated Veneer Lumber (LVL) is an engineered wood material composed of thin pieces of wood from solid wood. This work investigated the 2D AE source localization methods on a LVL plate with four AE sensors mounted. Multiple pencil break tests were conducted on a grid drawn on a LVL plate. The differences of time of arrival (dTOAs) were computed using the AIC (Akaike information criterion) function to avoid the noise interference on determining the onset of the waveform. A dataset composed of the dTOAs and the coordinates the source was constructed for later analysis. Student-t process and Gaussian process with Bayesian considerations were all tried to locate the sources. Due to heterogeneity, nonuniformity of the velocity profile and internal flaws, the velocity profile method cannot handle the source localization well. In contrast, the results of source localization by Student-t process and Gaussian process were reasonable and accurate. This paper proposed a framework of using AE system to localize the AE sources in anisotropic materials with multiple sensors. Besides, the existing methods for 2D AE source localization were evaluated and the Gaussian process with Bayesian considerations was proposed. The training process was discussed, and prediction results were thoroughly evaluated. Root mean square errors in different sensor pair selection scheme were computed. Results showed that the Gaussian process and Student-t process with Bayesian considerations is quite capable of handling AE localization on wood.

Presenting Author: Xiangdong He University of Utah

Presenting Author Biography: Currently, Xiangdong He is a Ph.D. candidate in Civil Engineering with focus on structural Engineering. His research mainly revolves around applying probabilistic machine learning tools to aid acoustic emission testing and evaluation research. His Ph.D. study focused on structural health monitoring on engineered wood structures by probabilistic machine learning and acoustic emission.

Authors:

Xiangdong He University of Utah
Xuan Zhu University of Utah