Thesis
Deep learning for ultrasonic non-destructive evaluation of aerospace composites
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2025
- Thesis identifier
- T17224
- Person Identifier (Local)
- 202166078
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- Carbon fibre reinforced polymer (CFRP) composites are critical materials in aerospace components and are seeing ever increasing demand. Ensuring the integrity of these composites is critical and Non-Destructive Evaluation (NDE) is widely used to provide crucial insights about the component. While robotic sensor delivery has automated the physical aspect of sensor placement, the analysis of the resulting data remains a manual and labour-intensive task. This manual analysis is not only time-consuming but also susceptible to human error, thereby limiting manufacturing capabilities. Recent advancements in machine learning (ML) and deep learning (DL) offer potential solutions to automate this analysis process. This thesis investigates the application of DL methods to analyse Ultrasonic Testing (UT) data from CFRP composites, one of the most common NDE techniques used for aerospace components. The initial phase of research addressed the challenge of data scarcity for training DL models by utilising synthetic datasets. Various methods for generating synthetic UT data were explored, with a deep generate method resulting in a trained classifier with a 24.2% improvement in defect detection accuracy when compared to the same classifier trained on simulated data. This helped to bridge the gap between simulated and experimental training data. Building on this foundation, the research then focused on the automatic analysis of volumetric UT data. Initially targeting defect detection, the study progressed to volumetric segmentation. For defect detection, a novel architecture was developed, achieving a 22.6% increase in classification accuracy compared to established architectures. A fully supervised model was then employed to train a 3D U-Net for segmentation, which performed well in sizing and localising defects within the training distribution. However, the model's performance declined when tested on out-of-distribution samples. The final phase of research explored self-supervised learning to train a model for defect segmentation, reframing the problem as one of outlier prediction. This approach eliminates the need for defective examples during training and significantly enhances generalisability. The method was also tested on an industrial component and scan setup, demonstrating promising performance and applicability in real-world scenarios. This thesis presents promising methods to automating the analysis of UT data from CFRP composites, highlighting the potential of DL methods to improve accuracy, reduce human error, and enhance manufacturing processes in the aerospace industry.
- Advisor / supervisor
- Mohseni, Ehsan
- MacLeod, Charles
- Pierce, Gareth
- Resource Type
- DOI
Relations
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
PDF of Thesis T17224 | 2025-03-05 | Public | Download |