Open Collections

A first step toward intelligent forming of fabric composites : adapting basic industry 4.0 concepts

University of British Columbia

Industry 4.0 is becoming a next-generation R&D area in the advanced composites manufacturing sector, to fully automate fabrication processes and increase the production quality and agility. Forming woven fabric reinforced composites on to complex 3D shapes is particularly of interest under this R&D trend, as these composites have become materials of choice in high-performance aerospace, energy, and automotive structures. Defects during the forming of long-fiber fabrics can affect the mechanical properties of the product. A vision under Industry 4.0 would be to develop AI-based tools that can predict, detect, and mitigate such defect during manufacturing. This thesis presents a prototype example of intelligent fabric composite manufacturing by conducting a series of hemisphere draping trials on a thermoplastic fiberglass/polypropylene weave, followed by the digitization of quality check. An image acquisition system was designed and integrated into a robotic arm to facilitate data collection. The quality function for each product was defined based on the height and area of formed wrinkles, as well as the sphericity and roundness of the formed hemisphere. By gathering data from around 70 samples and over 1200 images from different forming set-ups, machine learning algorithms were trained and consequently, the quality class of each product was predicted. Furthermore, to optimize the product quality, machine learning tools recommend the desired input variable of forming. Overall, a higher number of clamps and a higher number of reinforcement layers, with a higher and uniformly distributed torque magnitudes across clamps yielded the least visible wrinkles on the product. As another imaginary scenario in a smart factory, an on-line quality check tool was developed and tested. Namely, when the forming process is finished, a smart image acquisition system scans the part and checks the quality of different formed areas of the part and informs the AI tool of the ‘wrinkle’ or ‘no wrinkle’ continuously. It was found that the K-nearest neighbors (KNN) and Support Vector Machine (SVM) machine learning models could detect the wrinkles with an error rate of less than 5%, regardless of the background noise in the images such as an external object, marks, etc.

Text

2020-06-01

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/74627

Mechanical Engineering

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/