UBC Theses and Dissertations
A progressive learning framework, leveraging machine-learning knowledgeability, towards Composites 4.0 Crawford, Bryn
The manufacture of fibre-reinforced polymer composite aerospace structures is a highly complex task subjected to a stringent framework of process qualification and structural substantiation, in order to minimize the risks associated with the complexity along with underlying aleatoric and epistemic uncertainties. Industry 4.0 is an emerging set of technologies and tools that could enable better decision-making towards risk reduction in aerospace manufacturing, supported by data-driven models which have the potential to aid in reducing production risk and improving efficiency. This research develops an Industry 4.0-based modelling pipeline in which both (hypothetical) factory-level datasets and engineering design activities pertinent to industrial considerations are included, recognizing the need for a framework that both qualifies the use of such models and systems, as well as containing functionality for in-situ decision-making and flexibility. The research is delivered as a case study, focusing on the autoclave cure process for AS4/8552 carbon prepreg material, modeled using black-box artificial neural networks (ANNs) in the prognostic regime, and Bayesian Belief Networks (BBNs) in the diagnostic regime. Two sub-case studies are then investigated for the latter regime; the first using a naïve Bayes model with random dataset selection, while the second leverages a combined-query and inference-based model that explores data selection strategies to optimally reduce prediction variance. Surrogate and highly interpretable Logistic Rule Regression models, combined with fuzzy logic scoring, provided a new tool for assessing the quality of such factory-level datasets prior to training production ANN models. Further, a proposed model accuracy metric, composed of specificity (false positive prediction rate) as a global indicator, as well as a local indicator Decision Boundary Crispness Score (DBSC), has been introduced to quantify the decision maker confidence in the context of using black-box models such as ANNs during production. Results demonstrated significant interdependence between the different modules in the Industry 4.0 data and model pipelines, towards formulating global strategies for the design and evaluation of such complex manufacturing systems.
Item Citations and Data
Attribution-NonCommercial-ShareAlike 4.0 International