Session: 15-03-02: Machine Learning and Deep Learning in Safety, Reliability, and Maintenance II

Paper Number: 167166

Optimization of Experimental Infrared Image Dataset for the Identification of Barely Visible Damage in Wind Turbine Blade Samples Using Transfer Learning 

One of the significant applications for composite materials, specifically, glass fiber reinforced polymers (GFRPs), is in wind turbine blades. Wind turbine blades are manufactured for at least a lifespan of 20 years. In this time, they suffer a significant amount of impact damage, which is not easily visible with the naked eye or with traditional visual inspection techniques. This type of damage is known as barely visible impact damage (BVID). Similar damage is observed in the aerospace and aviation industry. When an impact occurs on the surface of the material, damage to the blade may not be detected during a visual inspection, as the damage can lie beneath the surface in the form of matrix cracks and delamination. When damage goes undetected for a prolonged period, the damage can propagate through the material and lead to catastrophic failure of the blade much before the intended lifespan. This catastrophic failure can lead to severe economic losses and even loss of life. Efforts to detect this damage have been explored using nondestructive techniques such as Lamb wave signals, LiDAR, and X-rays. However, these methods are not applicable for large components such as whole wind turbine blades. These techniques become inapplicable for on-site inspection due to the need to disassemble the blade, the high cost and time required for inspection, and the low accuracy. Infrared thermography (IRT) can be implemented as an effective alternative to these techniques. Using IRT, the defects within the material can become more prominent and be detected. However, IRT has a lot of limitations, such as low resolution, fish-eye effect, and shadows due to collecting images in varying environmental conditions. In addition, manual classification of damage can be erroneous and time-consuming. These problems can be solved using artificial intelligence (AI) algorithms. Deep learning algorithms were observed to be effective even with the limitations posed by IR images. However, two important questions remain: (1) Which algorithm is most appropriate? And (2) How much data is enough? To determine the appropriate model and data sufficiency for the models, a sensitivity analysis is crucial for each application. To fulfill the research gap, the research question developed for this paper is, “Can a pre-trained algorithm detect the BVID damage with higher performance than the non-pretrained algorithm with the help of sensitivity analysis for an infrared image dataset of wind turbine blade samples collected experimentally?” To answer this question, two objectives were developed: (a) to create a damage classification framework for collected IR image datasets in varying environmental conditions to identify BVID and (b) to compare the performance of multiple deep learning algorithms with varying sizes of training datasets to determine data sufficiency. In this objective, statistical analysis was conducted to create a performance matrix over multiple iterations of each dataset. The algorithms under consideration were a non-pretrained algorithm, namely, a convolutional neural network (CNN), and two pre-trained algorithms, namely, VGG16 and InceptionV3. These algorithms were trained and tested using multiple datasets containing experimentally collected IR images from 100 to 1398 images with an increment of 100 images per dataset. Each dataset was used for 10 iterations to quantify the uncertainty of the resulting parameters. For each of the iterations, validation accuracy, validation loss, and training loss were recorded, along with the precision, recall, and f-1 scores. The results showed that the transfer learning models outperformed the CNN model significantly, reaching a higher accuracy at a lower number of images. The sensitivity analysis provided a framework for data sufficiency, and with the help of statistical analysis, the best-performing algorithm was identified.

Presenting Author: Hannah Jones University of Arkansas at Little Rock

Presenting Author Biography: Hannah Jones is a Master's student at the University of Arkansas at Little Rock in the Mechanical Engineering program.

Authors:

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