The growing adoption of Carbon Fibre Reinforced Plastic (CFRP) composites in
safety-critical structures, such as aircraft fuselages and wind turbine blades, requires
thorough inspection process to ensure material integrity and prevent catastrophic
failures. These inspections are conducted using Non-Destructive Evaluation (NDE), a
collective term for methods that assess the quality of materials without causing
damage. Among these methods, Ultrasonic Testing (UT) stands out as a preferred
inspection technique in the aerospace industry. The field of NDE has experienced
significant advancements with the introduction of advanced sensor technologies and
robotic manipulators, which have automated and accelerated data acquisition
processes. However, data analysis and interpretation remain predominantly manual,
making the process time-consuming and prone to human error, thus creating a
bottleneck in manufacturing. Recent advancements in Artificial Intelligence (AI)
present new opportunities to automate these tasks.
This thesis explores the application of AI techniques to analyse UT datasets obtained
from reference CFRP samples representative of those used in the aerospace industry.
The initial research focused on evaluating the performance of supervised AI methods,
specifically object detection models, in defect detection tasks using ultrasonic C-scan
images, which represent the top cross-sectional view of the inspected materials. As a
baseline, both a traditional signal thresholding technique and an enhanced statistical
thresholding method, based on theoretical mathematical distributions fitted to the
observed data, were examined. The primary contribution of this work is the
demonstration of the superior performance of AI models in this context over
thresholding methods. Additionally, supervised training was conducted exclusively on
simulated data, thereby addressing the data scarcity challenge.
Building on this, an unsupervised AI method in the form of anomaly detection was
explored. This approach addressed the scarcity of datasets containing defective
indications and the challenges of relying on large-scale simulations, which require
significant computational resources and extensive manual effort to generate ground
truths for supervised training. A two-step workflow was developed, comprising an
automated signal gating method based on unsupervised clustering and an autoencoder model serving as an anomaly detector applied to ultrasonic B-scans (which represent
cross-sectional images of the material). The key advantages of this method include a
streamlined development process focused on the use of pristine data (in this thesis, the
term pristine refers to samples that contain no intentional or unintentional
manufacturing defects that are detectable using the ultrasonic inspection setup
employed, within the limits of its resolution), which is more readily available, and the
elimination of ground truth generation requirements. This workflow was successfully
applied to samples with both uniform thickness and complex geometries. Additionally,
this research investigated the impact of human factors on AI results and highlighted
the challenges posed by inconsistent data quality during scans.
The final stage of this research focused on strategies for integrating the developed AI
models into NDE data analysis, driven by the recent evolution towards NDE 4.0 (a
concept that combines digitalisation, automation, and connectivity to modernise NDE)
focused on enabling fully automated systems. Despite significant research in this field,
implementation strategies are often underexplored, and clearly defined automation
levels achievable with AI remain lacking. To address these gaps, four levels of data
analysis automation using AI were defined and evaluated. Additionally, the
synchronous use of multiple AI models, each designed to process distinct views of the
ultrasonic data, enabled cross-validation between models. This approach enhanced
trust in the automated system while offering mechanisms to mitigate potential
performance degradation of NDE operators using such systems. Furthermore, this
strategy aligns with NDE 4.0 objectives, transitioning operators into supervisory roles
while delegating repetitive tasks to the AI systems. The developed methods were
evaluated in a case study involving complex geometry samples, demonstrating their
effectiveness for potential industrial applications.
Overall, this thesis proposes methods to assist in automating NDE data analysis
processes for UT of CFRP composites, with the primary goal of enhancing accuracy
and reducing analysis time to address the current bottleneck in aerospace
manufacturing.
| Date of Award | 15 Sept 2025 |
|---|
| Original language | English |
|---|
| Awarding Institution | - University Of Strathclyde
|
|---|
| Sponsors | EPSRC (Engineering and Physical Sciences Research Council) |
|---|
| Supervisor | Ehsan Mohseni (Supervisor), Gareth Pierce (Supervisor) & Gordon Dobie (Supervisor) |
|---|