基于深度迁移学习的多可信度复杂产品设计优化方法研究

在现代复杂产品设计中，基于物理行为模型和科学计算的仿真计算已逐渐发展成不可或缺的手段，它可大幅降低实验成本，缩短产品开发周期。但是随着计算机技术的高速发展，仿真精度虽然得到极大提升，但随之而来的是计算成本呈指数级增长，极大限制了数值仿真技术在设计优化中的应用。在此背景下，基于近似模型的设计优化方法应运而生。它通过对计算昂贵的仿真分析模型或物理实验进行近似替代，有效降低了设计成本。其中，融合高/低可信度模型的多可信度近似建模技术，能有效平衡近似模型预测性能和建模成本之间的矛盾，更是表现出了巨大潜力。本课题拟对基于多可信度近似模型的复杂产品设计优化理论展开研究，提出一种基于深度迁移学习的多可信度近似建模方法及考虑多因素平衡的全局优化策略，以提升近似模型的预测精度，改善复杂产品设计质量，同时大幅度降低计算成本，为我国高端电子制造、精密测量、智能无人系统的关键部件研发提供新的理论与方法。

In modern complex product design, simulation technology based on physical behavior models and scientific calculations has gradually developed into an indispensable means. It can greatly reduce the high cost of experiments and shorten the product development cycle. However, with the rapid development of computer technology, although simulation accuracy has been greatly improved, the cost of simulation analysis has also increased exponentially, which greatly limits the application of numerical simulation technology in design optimization. In this context, design optimization methods based on surrogate models have emerged. It effectively reduces the design cost by using the surrogate to replace computationally time-consuming simulation analysis or physical experiment. Among them, the multi-fidelity modeling technology which fused the information both from high and low-fidelity models, shows great potential in effectively balancing the prediction performance of the surrogate model and the modeling cost. In this project we plan to conduct a research on multi-fidelity design optimization for complex products, and propose a multi-fidelity modeling method based on deep transfer learning, and a multi-criteria balanced global optimization strategy to improve the prediction accuracy of the surrogate model, improve the design quality, while greatly reduce the computational costs. It provides new theoretical and methodological support for the development of key components of our country's high-end electronic manufacturing, precision measurement, and intelligent unmanned system.