复杂产品基于深度代理模型的高效仿真与优化方法

To solve the problem of efficient simulation and multi-objective optimization of complex products for simulation based design, based on the big simulation data accumulated in the past time of its development, deep surrogate model based approach is to be presented for efficient design simulation and optimization of complex products, in which, deep model is constructed from big simulation data by self-coding learning and inner and outer loop based progressive and iterative training, deep surrogate model is obtained with the deep model and a few data through a close-loop of design points planning based on its uncertainty analysis and the deep model updating based transfer learning and the data instantly obtained from current product design simulation, and then, multi-objective optimal points can be efficiently obtained from the deep surrogate model by introducing the reference point according the working condition of the product, constructing composite fitness function for multi-objective evolutionary optimization to enable the optimal population converges to the Pareto frontier and additionally distributed uniformly. . Finally, a prototype system of deep surrogate model based complex products simulation and optimization approach is to be developed on the above algorithms, and is validated with the control parameters optimization for engine. Such approach is very meaningful on theory and engineering applications as it can relieve the technique bottleneck, and enables the problem of accurate simulated model construction and efficient multi-objective optimization can be solved in simulation optimization, and then efficiently supports the optimal design of complex products.

针对复杂产品仿真与多目标优化效率问题，基于以往产品研发过程中积累的大量仿真数据，拟研究复杂产品基于深度代理模型的高效仿真与优化方法，通过自编码器学习及内外环两层递阶式迭代训练，构建具有强泛化能力深度模型，以准确刻画多设计变量与多响应输出间的高阶非线性关系；通过不确定性分析、闭环式规划与仿真，以及深度模型迁移学习更新，尽量以较少仿真次数，获取复杂产品较为准确的深度代理模型；通过引入参考设计点，优化构造复合适应度函数，以及多目标进化优化，使进化种群有向快速收敛到Pareto优化前端，且具有较好多样性分布。以此为基础，研制复杂产品基于深度代理模型的高效仿真与优化原型系统，并通过发动机多种工况控制参数优化整定进行应用验证。本项目研究有望突破现有方法所面临的技术瓶颈，解决复杂产品仿真与多目标优化效率问题，为复杂产品仿真分析与优化提供理论方法及支撑软件，具有重要理论研究及工程应用意义。

项目针对复杂产品仿真优化中的效率、稳定性和适应性等瓶颈问题，提出了基于深度代理模型的高效仿真优化方法。基于已有的大量仿真数据，结合深度多层神经网络准确表征能力以及超参数自适应选取策略，获取稳定且泛化能力强的深度代理模型；并基于少量仿真数据规划方法，设计主动闭环迁移学习的深度代理模型，从而稳定获取新产品高泛化能力深度代理模型；在此基础上，运用谱聚类发掘解的结构特征，基于该结构特征和决策者给定的设计参考点集，设定概率因子引导多种不同的变异算子产生新解，使得新解更加多样化且自动带有决策者偏好信息，结合MOEA/D-DP与深度代理模型，实现高效仿真优化。通过以上的研究，对伸缩臂叉车的设计优化进行应用验证，结果显示项目提出的深度代理模型优化方法可以少量仿真数据完成研究对象的仿真优化，有效提高了仿真优化效率。项目为复杂产品设计提供了理论方法及支撑算法，具有重要理论及工程应用意义。