GOALI/Collaborative Research: Curating Complex Data Sets for Machine Learning Applied to Flexible Assembly Design and Optimization

在高度竞争的市场中，例如汽车行业，产品开发时间、质量和成本都是至关重要的。成功实现这些目标需要快速探索设计替代方案。这项学术与工业联络资助机会（GOALI）项目的最终目标是利用人工智能推进汽车结构设计中数据驱动的设计空间探索。数据科学推进机器学习的发展高度依赖于满足算法训练和验证的某些技术目标的大型数据集的可用性。虽然训练数据集在社交网络、消费者偏好和金融领域广泛可用，但这些数据集需要人为地用于工程产品。该项目将为特定工程设计目标提供大量可选设计配置的数据集，这些目标与技术验证的性能指标相关。研究将集中在柔性装配设计的应用领域，这是一种广泛应用于汽车和家电行业的多阶段设计和制造工艺。一旦经过培训和验证，使用由此产生的深度学习工具不需要专门的专业知识。除了推动数据科学，这项工作的另一个影响将是复杂结构设计和分析的民主化，通过支持没有高等学位的个人做出的设计和制造决策。它还将使下一代工程师能够接受有关将先进的机器学习应用于工程设计和制造以及在产品开发中采用数据驱动工具的教育。该项目将研究数据管理特征（例如，数量、模态、粒度、异质性、平衡），同时考虑相关人工神经网络算法的应用领域和能力，包括卷积、循环、生成对抗网络、多层感知器和池化架构。为了生成所需的数据集，将制定符合管理标准的自动化模拟管道。结果将通过行业基准和工业合作伙伴（本田）的实验数据进行验证。将设计数学方法从仿真数据中提取关键性能参数。将设计其他方法来研究每个数据样本的抽象、分解和划分为适合通过联合人工神经网络并行处理的子集，或通过研究社区选择的分布式机器学习网络单独处理的子集。所有数据集将通过亚马逊云发布，供其他工程设计研究人员使用，以推进各自领域的设计科学。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

In highly competitive markets, such as the automotive sector, product development time, quality, and cost are all critical. Successfully meeting such goals requires rapid exploration of design alternatives. The ultimate goal of this Grant Opportunity for Academic Liaison with Industry (GOALI) project is to advance data-driven design space exploration in automotive structural design using Artificial Intelligence. The development of Data Science to advance Machine Learning is highly dependent on the availability of large data sets that meet certain technical goals for algorithm training and validation. While training data sets are widely available for social networks, consumer preferences, and finance, such data sets need to be artificially curated for engineered products. This project will produce large data sets of alternative design configurations for particular engineering design objectives interrelated with technologically verified performance metrics. The research will focus on the application domain of flexible assembly design, a multi-stage design and manufacturing process widely used in the automotive and appliance industries. No specialized expertise will be needed for using the resulting deep learning tools once they have been trained and validated. In addition to advancing Data Science, another impact of this work will be democratization of complex structural design and analysis by supporting design and manufacturing decisions made by individuals without advanced degrees. It will also enable the next generation of engineers to be educated about applying advancing Machine Learning to engineering design and manufacturing and adapting data-driven tools in product development. This project will investigate data curation characteristics (e.g., volume, modality, granularity, heterogeneity, balance) while simultaneously considering the application domain and capabilities of the related Artificial Neural Net algorithms, including convolution, recurrent, generative adversarial networks, multi-layer perceptrons, and pooling architectures. To generate the required data sets, an automated simulation pipeline will be formulated that meets curation criteria. The results will be verified through industry benchmarks and experimental data from the industrial partner (Honda). Mathematical methods will be devised to extract key performance parameters from the simulation data. Additional methods will be designed to investigate abstractions, decompositions, and partitions of each data sample into sub-sets suitable for processing in parallel through federated Artificial Neural Nets, or individually through distributed machine learning networks, as chosen by the research community. All data sets will be published through Amazon Cloud for use by other engineering design researchers to advance design science in their respective fields.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.