Forming Ahead with Deep Learning for Composites Manufacturing

能够在快速周转时间内试验成形航空航天几何形状的速度支持解决方案是非常可取的。这与未来复合材料制造研究中心最近探索干纤维织物成型技术的活动是一致的，目的是提高生产率、产量和零件质量。数字孪生的引入在智能优化这些复合材料设计和制造过程中显示出了希望。为了利用这些技术，需要解决一个重要的分析挑战：处理捕获的车间数据质量的可变性。例如，可变的照明条件可能会使基于计算机视觉的分析复杂化。为了提高鲁棒性，需要获取具有代表性工厂条件的实验数据，并用计算机视觉模型进行训练。为此，该项目的目的是开发数字线程需求：(a)允许仪器测试平台试验快速执行，(b)跟踪和更新数字孪生中的变化，以及(c)根据规格提供缺陷的自动分析。这些贡献将减少表征成形反应所需的物理试验次数，从而节省成本并降低材料废品率，从而带来好处。目的:1。研究在工厂环境条件下捕捉预成型缺陷的传感解决方案。2. 探索深度学习技术来分析成形试验数据。3. 开发数字线程需求并理解数据处理流程。方法：本研究的新颖之处在于对捕获的数据进行自动化分析，以跟踪成形过程中预成形件的变化。这需要车间数据，如照片，来自复杂航空结构的成型和灌注试验。为了说明工厂环境对缺陷捕获的变化，介绍了不同的代表性照明配置。对控制参数和物料响应进行了采集。积累的传感器数据用于构建训练集，然后输入到数字孪生中。宏观缺陷，如面内波纹和面外褶皱，使用基于计算机视觉的方法自动表征。这样就可以自动填充特许数据库，并根据工程规范对预制件进行验证。

Rate-enabling solutions that allow the trialling of forming aerospace geometries in a rapid turnaround time are highly desirable. This is consistent with recent Future Composites Manufacturing Research Hub activities which explored dry-fibre fabric forming technologies, with the aim of improving production rates, volumes, and part quality. The introduction of digital twins has shown promise in intelligently optimising these composite design and manufacturing processes. To leverage these technologies, a significant analysis challenge needs to be addressed: handling the variability in quality of captured shopfloor data. For instance, variable lighting conditions might complicate a computer-vision based analysis. To improve the robustness, experimental data with representative factory conditions needs to be obtained and trained with computer vision models. To this end, the aim of this project is to develop the digital thread requirements that: (a) allows instrumented testbed trials to be quickly performed, (b) tracks and updates changes in a digital twin, and (c) provides automatic analysis of defects against specification. These contributions would provide benefit by lessening the number of physical trials needed to characterise forming responses - resulting in cost-savings and reduced material scrap rates.Objectives: 1. Investigate sensing solutions to capture preforming defects in conditions representative of a factory environment. 2. Explore deep learning techniques to analyse forming trial data. 3. Develop the digital thread requirements and generate understanding of the data process flow. Method: The novelty of this research is in the automation of analyses of captured data to track changes of preforms during forming processes. This requires shopfloor data, such as photographs, from forming and infusion trials of complex aerostructures. To account for variation in factory environment on defect capture, different representative lighting configurations are introduced. Acquisition of the control parameters and material response is also carried out. The accumulated sensor data is used to construct training sets before being fed into the digital twin. Macroscale defects, such as in-plane waviness and out-of-plane wrinkling, are automatically characterised using computer-vision based methods. This enables the concessions database to be automatically populated and the preform to be verified against the engineering specifications.