Session: Research Posters

Paper Number: 173470

Detecting the Undetectable: Methods to Detect Bvid in Wind Turbine Blade Samples 

A significant application for glass fiber reinforced polymers (GFRPs) lies in wind turbine blades. In the blade’s lifetime, they suffer impact damage during transportation, installation, and operation. A significant issue regarding impact damage is a specific type of damage called barely visible impact damage (BVID), which is damage resulting from an impact at the surface of the composite material, with the majority lying beneath the surface in the form of matrix cracks and delamination. This makes traditional visual inspections challenging, as BVID is difficult to see with the naked eye. If this damage is not detected in the early stages, it will propagate through the blade and result in catastrophic failure, leading to increased expenses and significant safety concerns. Nondestructive techniques to detect this damage have been developed using techniques such as X-ray and LiDAR. However, these methods can be expensive, require disassembly of the blade, increase inspection times, and have low accuracy. To combat these issues, infrared thermography (IRT) techniques are proposed. IRT is considerably more cost-effective and can be implemented in applications like drone inspections. As the blade heats, the defects will heat at a different rate, which is then captured by IRT. However, IRT still has limitations like low resolution and false shadows. Additionally, manually classifying these images can be extremely time-consuming and inaccurate. To mitigate the limitations in manual classification and increase accuracy, deep learning models can be implemented, which have been shown to be effective in IRT applications. Despite their indicated effectiveness, two questions must still be answered: (1) Which model is most appropriate? And (2) How much data is enough? To answer these questions and determine the most effective model and the associated parameters, a sensitivity analysis is necessary. To address this gap, the research question “Can a pre-trained model detect the BVID damage with higher performance than the non-pretrained model with the help of sensitivity analysis for an infrared image dataset of wind turbine blade samples collected experimentally?” was developed. To answer this question, two objectives were established: (a) to develop a framework to classify the collected IR image dataset over various environmental conditions and (b) to conduct a comparative analysis for the performance of multiple deep learning models with various-sized datasets to determine data sufficiency. The two types of deep learning models considered were a convolutional neural network (CNN) and transfer learning models, namely, InceptionV3 and VGG16. The models were run on increments of 100 to 1008 images and ran for 10 iterations over each dataset to capture any variability in results. Results indicated that transfer learning models reached a higher accuracy with a lower number of images, and the sensitivity analysis provided crucial insights for determining data sufficiency. 

Presenting Author: Hannah Jones University of Arkansas at Little Rock

Presenting Author Biography: Hannah is a first-year M.S. Mechanical Engineering student at the University of Arkansas at Little Rock and a researcher in the Material Intelligence and Prognostics Laboratory under Dr. Shweta Dabetwar. She began researching in her junior year of her undergraduate studies and has continued her work into her master's. She has a particular interest in composite materials and aerospace applications, having had previous experience in composites research and as a NASA intern. She is also currently an AAUW Selected Professions Fellow, which supports her studies during the 2025-2026 academic year.

Authors:

Hannah Jones University of Arkansas at Little Rock
Shweta Dabetwar University of Arkansas at Little Rock
Aditi Barua University of Arkansas at Little Rock