Session: 06-01: Machine Learning and Statistical Methods in NDE I

Paper Number: 136845

136845 - Lamb Wave Anomaly Detection by Ensembling Spatial and Wavenumber Domains 

Abstract: 

Guided wave field anomaly detection has proven to be a feasible tool for nondestructive evaluation and structural health monitoring. These anomaly detection methods are often based on the fact that waves in damaged regions and undamaged regions propagate differently.  Yet, most anomaly detection approaches have two setbacks. First, they provide limited characterization of the anomaly / defect / damage.  Such a capability not only enables quantifying the differences in wave propagation in damaged regions but also provides trustworthiness in the result. Second, most existing anomaly detection methods are performed over a single domain, resulting poor damage discrimination under certain conditions. Therefore, by analyzing wave propagation in multiple domains, we should be able to achieve both better characterization and more robust detection.

We present two methods for analyzing wave propagation in the spatial domain and wavenumber domain to detect and locate damaged regions. To validate our approach, we generate simulated guided wave fields with 2000 frames with a sampling frequency of 1MHz. In this simulated wave propagation, wave speed is 5337.6 m/s and there is 10% variation at two square-shaped regions, representing corrosion-like damage. In spatial domain, we utilize pixel intensity change in each time frame to detect anomalies. In the wavenumber domain, we cluster wave modes in polar coordinates. We find that the spatial method detects all damaged regions but contains significant amounts of noise. The wavenumber domain shows less noise but fails to detect both regions simultaneously. By combining, or ensembling, our detection methods in both domains, we can achieve more robust damage detection. We employ the intersection over union (IOU) to quantitatively assess this approach. This metric is computed as the intersection of the true and predicted regions divided by their union. As there are two damaged regions, we measure the IOU score for each region and then average the scores. When using the spatial domain only, the IOU score is 47%. After ensembling our methods, the score raises to 78.50%.

Presenting Author: Woohyun Eum University of Florida

Presenting Author Biography: Woohyun Eum is a Ph.D. student in the Department of Electrical and Computer Engineering at the University of Florida.

Authors:

Woohyun Eum University of Florida
G. Austin Simon University of Florida
Charlie Tran University of Florida
Joel B. Harley University of Florida