考虑仿真单元时耗差异的系统多保真度全局优化方法研究

The present optimization methods for computationally expensive simulation systems still have the following weak points: 1. The difference of time consumption of simulation units in a system is ignored, which leads to the waste of computing resources and the low efficiency; 2. Overdependence on expensive simulation analyses for large-scale exploration leads to huge computational cost and long optimization periods; 3. The present surrogate-based global optimization methods can hardly be used for the computationally expensive simulation system directly, and the capability of multi-point sampling is limited. Therefore, according to the situation that a simulation system includes multiple expensive simulation units and meanwhile their time consumptions are different, this project plans to focus on the research of the multi-fidelity surrogate-based global optimization method, partitioning the system and constructing a mathematical model for optimization. By fusing the data from the low and high fidelity models, the multi-fidelity surrogate model is constructed, which can coordinate reasonably the number of samples for different simulation units, make the computing resources utilized efficiently, and decrease the cost. Besides, in order to realize the global exploration and the improvement of computational efficiency, a global multi-point sampling method using space division will be developed, and a novel strategy that uses the scoring criterion to select promising samples will be presented. Finally, the proposed method will be used to realize the overall optimization design of a blended-wing-body underwater glider, and will be compared with other system optimization methods. The research findings will enrich the present optimization theory for complex systems, and meanwhile will be significant for engineering applications.

现有耗时仿真系统优化方法存在以下不足：1.忽视系统内各仿真单元时耗差异，导致并行计算资源浪费、计算效率低；2.大范围全局探索设计空间时，过于依赖耗时仿真分析，导致计算代价大、优化周期长；3.现有代理模型全局优化方法难以直接用于耗时仿真系统，且多点采样能力有限。为此，针对涉及多种耗时仿真单元，且各单元耗时程度不同的仿真系统，开展多保真度代理模型全局优化方法研究，区分各单元耗时程度，细致地划分系统，构建一个可并行管理的优化数学模型。通过利用每个耗时仿真单元的高、低保真度样本，融合出多保真度代理模型，同时协调分配各单元采样量，使计算资源充分利用，降低建模成本。发展基于空间划分的全局多点采样方法，结合一种新的评分策略选择候选点，实现大范围高效探索，提高计算精度与优化效率。最后，完成翼身融合水下滑翔机系统优化应用与方法对比验证。研究成果将丰富和完善现有复杂系统优化理论，同时具有重要的工程应用价值。

自主研发新型复杂机电产品时，协同仿真能够提高产品质量、降低设计成本，但同时将产生高昂的计算代价与复杂的仿真数据流，如想进一步实现系统优化十分困难，而现有方法存在不适用等问题：1.忽视系统内各仿真单元时耗差异，导致并行计算资源浪费、计算效率低；2.大范围全局探索设计空间时，过于依赖耗时仿真分析，导致计算代价大、优化周期长；3.现有代理模型全局优化方法难以直接用于耗时仿真系统，且多点采样能力有限；4.面对协同仿真系统设计优化时，存在高维度，多约束，多数据源，离散域等难点，现有方法难以有效求解。为此，本项目开展了考虑仿真单元时耗差异的系统多保真度代理模型全局优化方法研究，解决了系统优化设计时，计算代价大、难以实现全局探索，以及由于仿真单元时耗差异带来的并行计算不易实施的问题。此外，围绕协同仿真系统设计与优化存在的并行性差、计算耗时、耗时度不同、设计维度高、约束条件复杂、设计域不连续等难点，分别建立了多代理优化、多保真度优化、高维优化、约束优化、离散优化理论。具体地，提出了一种基于仿真单元时耗差异的系统优化建模方法；提出了一种基于数据挖掘的多保真度代理模型全局优化方法；提出了一种基于并行采样的多代理模型全局优化方法；提出了一种考虑计算耗时约束的元启发式全局优化方法；提出了一种考虑高维空间的代理辅助群智能优化方法；提出了一种考虑离散设计域的代理辅助全局优化方法；完成了代理辅助的优化方法在翼身融合水下滑翔机系统上的应用验证，实现翼身融合水下滑翔机的全系统优化，使得各项指标得到了极大提升。经过优化滑翔机升阻比对攻角的位置导数提升了11.8%，能源携带量提高了2.8%，滑翔角降低14.56%，单周期滑翔时间缩减了18.98%，滑翔机单个周期能耗降低了14.0%，单个周期航程增加了18.26%，单位航程能耗提升了36.9%，使得滑翔机整体航程提升了38.9%。对我国新概念水中兵器的正向设计研发，以及未来我国重大装备数字化设计转型奠定了坚实的基础。