ºìÐÓÊÓƵ

Topic

Intelligent Machine Learning-based Leakage Detection and Localization in Vacuum-assisted Composite Manufacturing

Department of Mechanical Engineering

Date & location

• Tuesday, March 25, 2025

• 8:00 A.M.

• Virtual Defence

Reviewers

Supervisory Committee

• Dr. Homayoun Najjaran, Department of Mechanical Engineering, University of Victoria (Supervisor)

• Dr. Colin Bradley, Department of Mechanical Engineering, UVic (Member)

• Dr. Brandon Haworth, Department of Computer Science, UVic (Outside Member) 

External Examiner

• Dr. Navid Zobeiry, Department of Materials Science & Engineering, University of Washington 

Chair of Oral Examination

• Dr. George Tzanetakis, Department of Computer Science, UVic

   

Abstract

Vacuum-assisted composite manufacturing methods, such as vacuum bag prepreg layup and vacuum-assisted resin transfer molding (VARTM), utilize atmospheric pressure as a uniform external force to consolidate and saturate fabric components. However, vacuum bag leakages can result in defects such as air bubbles, resin traps, voids, non-uniform surface finishes, and ultimately, inferior mechanical properties. Detecting and repairing these leakages before the autoclave curing stage is therefore essential.

The leakage localization method used in this study relies on volumetric flow rate measurements of air evacuation lines. Multiple air evacuation channels, known as vacuum ports, are strategically placed at different locations in the production layup. Each port is equipped with sensors capable of independently measuring the volumetric f low rates of air during the process. In the presence of a leakage, the measured flow rate values will not stabilize at zero because air continuously enters the vacuum bag through the leak. The flow rate values correlate with the location of the leak, the overall layup configuration, and the positions of the vacuum ports.

We introduced an intelligent machine learning-based framework for leakage detection and localization, designed to learn the complex relationships between flow rate values and leak locations. To generate sufficient training data, an electric circuit analogy was developed to simulate the vacuum process. This approach provides a fast and reliable alternative to complex analytical simulations and extensive physical experiments. The proposed method has been validated and compared across various experimental configurations, demonstrating its effectiveness.

Using the available and synthesized data, we employed various machine learning models, including regression models, a Grid neural network, a physics-informed Grid neural network, leakage classification models, and physical parameter training algorithms for leakage prediction. Our methods not only predict leakage locations with acceptable accuracy but also generalize well across different configurations. Additionally, we addressed challenges associated with complex, non-uniform layups featuring regions of varying permeability. For the first time, our framework also tackled scenarios involving multiple simultaneous leakages, successfully localizing all leaks on the layup.

Our results demonstrate significant advancements over state-of-the-art methods. These improvements go beyond higher prediction accuracy, focusing on enhanced generalizability across various layups, reduced data requirements for training, and the ability to tackle complex scenarios, such as non-uniform permeability and multiple leakages, which were previously unaddressed. Notably, the novel PI-GNN framework outperforms regression models in both generalizability and data efficiency. By integrating physical knowledge with data science, the PI-GNN framework establishes a robust foundation for addressing layups of varying sizes and geometries. Furthermore, our proposed physical parameter training algorithm effectively learns the permeability of different regions within the layup, enabling the development of a more accurate and robust simulation tool for model training.

Optimizing the placement of vacuum ports to improve leakage location prediction is another challenge addressed in our work. Each layup offers numerous possible configurations for positioning vacuum ports to enhance leakage localization. We tackled this optimization problem by maximizing flow rate variance among the vacuum ports. Given the problem’s large state space, a hierarchical optimization approach was employed to identify the optimal configuration. Experimental validation confirmed that optimizing the port configuration significantly reduces leakage prediction errors. Keywords: VARTM, Prepreg, Graph neural network, Leak detection, Machine learning, Leak localization, port placement, optimization.