Data-driven manufacturing in the automated lay-up of composites

Abstract

The quality of components manufactured by Automated Fibre Placement (AFP) depends on a high number of process variables and their interactions. These interactions can be complex and subject to rapid change, and therefore traditional modelling approaches are challenging to deploy in industry. As a result, multiple development components are required in order to define optimal tape paths for new AFP components, incurring substantial costs.

A potential solution to this problem was explored through data-driven modelling, including machine learning, describing quality trends through the analysis of data captured during the AFP process. Data-driven manufacturing is beginning to revolutionise the wider manufacturing industry through the launch of Industry 4.0. However, deploying such a methodology within the composites industry remains a significant challenge. This thesis contains an investigation into the feasibility and potential benefits of such an approach, when applied to the AFP process.

Methods of capturing AFP process data were analysed, where key process variables were identified using a combination of literature and process experience. Roller nip-point temperature was selected as a variable for further study, through the capture of a key parameter, effective emissivity. It was shown that temperature measurement accuracy can be improved between 4% to 12% when compared with the nominal unity value.

Measurement of these variables was used to train machine learning models, in order to accurately predict an output process variable, local preform thickness, chosen due to its demonstrated impact on downstream defects. These models were able to achieve a high accuracy of R2 = 0.941 during the manufacture of a complex 3D component.

Moving towards closing the process planning-to-manufacture loop, these models predicted regions of higher preform thickness using the chosen component. Process planning was then adjusted in these regions, using process expert knowledge, and thickness was improved in these regions by approximately 10%.

Date of Award	3 Oct 2023
Original language	English
Awarding Institution	University of Bristol
Sponsors	National Composites Centre & Engineering and Physical Sciences Research Council
Supervisor	Carwyn Ward (Supervisor), Byung Chul (Eric) Kim (Supervisor) & Amit Visrolia (Supervisor)