Abstract

A Cyber-Additive Manufacturing Network (CAMNet) integrates connected additive manufacturing processes with advanced data analytics as computation services to support personalized product realization. However, highly personalized product designs (e.g., geometries) in CAMNet limit the sample size for each design, which may lead to unsatisfactory accuracy for computation services, e.g., a low prediction accuracy for quality modeling. Motivated by the modeling challenge, we proposed a data-driven model called family learning to jointly model similar-but-non-identical products as family members by quantifying the shared information among these products in the CAMNet. Specifically, the amount of shared information for each product is estimated by optimizing a similarity generation model based on design factors, which directly improve the prediction accuracy for the family learning model. The advantages of the proposed method are illustrated by both simulations and a real case study of the selective laser melting process. This family learning method can be broadly applied to data-driven modeling in a network with similar-but-non-identical connected systems.

摘要

网络增材制造网络 (CAMNet) 将连接的增材制造流程与作为计算服务的高级数据分析相集成，以支持个性化产品的实现。然而，CAMNet 中高度个性化的产品设计（例如几何形状）限制了每个设计的样本大小，这可能导致计算服务的准确性不令人满意，例如质量建模的预测准确性低。受建模挑战的启发，我们提出了一种称为家庭学习的数据驱动模型通过量化 CAMNet 中这些产品之间的共享信息，将相似但不相同的产品联合建模为家庭成员。具体来说，通过优化基于设计因素的相似性生成模型来估计每个产品的共享信息量，直接提高了家庭学习模型的预测精度。通过模拟和选择性激光熔化过程的真实案例研究说明了所提出方法的优点。这种家庭学习方法可以广泛应用于具有相似但不相同的连接系统的网络中的数据驱动建模。