Date of Award

8-21-2025

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Systems Engineering

First Advisor

Shweta Dabetwar

Abstract

Composite materials offer a high strength-to-weight ratio, making them ideal for large structures like wind turbine blades. However, they are prone to subsurface defects such as delamination, matrix cracking, and debonding, often caused by impact and worsened by fatigue. These defects can significantly compromise the structural integrity of composite structures like wind turbine blades. In particular, wind turbine blades frequently experience barely visible impact damage (BVID), which can further initiate subsurface defects and lead to catastrophic failure during operation. Because these defects are hidden beneath the surface, advanced inspection technologies are required for reliable detection. Infrared thermography (IRT), a non-destructive testing (NDT) technique, enables non-contact, real-time monitoring by detecting temperature variations on the surface that reveal subsurface anomalies. For this reason, IRT is especially effective for inspecting large structures such as wind turbine blades. However, few studies address the challenges of real-time infrared (IR) data acquisition for subsurface defect detection in such applications. Furthermore, IR images often suffer from issues like low contrast and fisheye effect, limiting the availability of high-quality data. Artificial Intelligence (AI), particularly deep learning (DL) models like convolutional neural networks (CNNs), with transfer learning capabilities, can help overcome these limitations. The CNN models perform well even with relatively small datasets due to transfer learning. Although few studies used advanced CNN models with lab-scale IRT data, they often lack justification for the optimum number of training images and epochs required for efficient model training. Moreover, few studies used IR images collected under uncertain environmental conditions for subsurface defect and BVID detection. Lab-scale images fail to fully capture the environmental variability which can make subsurface defects less visible and hence difficult to detect. There are also fewer studies that address the inherent uncertainties of CNN models which hinder the reproducibility of the results and generalization of the AI application to real world data. These uncertainties enhances when the CNN models are implemented with real-world IRT data due to environmental variability, human error etc. This study aims to address these gaps by investigating the research question: Can mask R-CNN automatically localize and identify subsurface defects in composite materials using real-world NDT data? To answer this, the study is guided by two specific aims: 1) To develop and evaluate a Mask R-CNN framework for subsurface defect localization and detection using a publicly available IR dataset, 2) To validate and generalize the framework using an experimental IR dataset collected from wind turbine blade samples under uncertain environmental conditions. To achieve Aim 1, the mask R-CNN model was trained on a public IR dataset of carbon fiber reinforced polymer (CFRP) composites containing artificial Teflon inserts to simulate delamination. A detailed framework was developed to determine the optimum number of training images and epochs through sensitivity analyses. Performance was assessed using mean average precision (mAP), loss, and defect detection accuracy across two test datasets. For the image sensitivity analysis, the model was trained for 30 epoch on 15 custom datasets with varying number of images from 200-900. Each training session was repeated for 10 iterations for repeatability and reproducibility of the results. These iterations were also used for statistical analysis that accounted for inherent variability in Mask R-CNN. The image sensitivity analysis showed that training mAP stabilized at around 81% and training loss at approximately 0.16 when the model was trained for 700 or more images. Interestingly, while the mAP remained steady, accuracy slightly declined as the dataset size increased. To investigate this further, a separate sensitivity analysis was conducted by varying the number of training epochs. For the epoch sensitivity analysis, the model was trained on the largest dataset (900 images) with epochs ranging from 30 to 130 in increments of 20. The mAP reached 94% by epoch 90 and remained steady beyond that point. Accuracy improved by 12% and 9% for the two test datasets, respectively. To fulfill Aim 2, the optimized framework was tested on an IR dataset of a wind turbine blade made of glass fiber reinforced polymer (GFRP) with BVID defects. The IR images were collected under uncertain environmental conditions. The number of training images and epochs were selected based on the findings from Aim 1. The model achieved multi-class BVID detection with an accuracy of approximately 90% and mAP of 85%. Similar to specific aim 1, the training was iterated 10 times to account for the variability in the model’s performance due to inherent uncertainties. These iterations were used for detailed statistical analysis. Thew maximum entropy principle was implemented to address the variability in the model’s performance in the repeated training iterations. This probabilistic approach provided deeper insights into the variability of model performance across training iterations and allowed for estimation of confidence bounds in the evaluation metrics. In summary, this study contributes to the advancement of non-destructive testing by developing an automated framework for defect detection in composite structures. The proposed method offers a comprehensive solution for condition monitoring and defect detection in components used in renewable energy and aerospace systems through IR image analysis.

Download

Share

COinS