【作者】 陈琪锋；

【导师】 戴金海；

【作者基本信息】 中国人民解放军国防科学技术大学 ， 航空宇航科学与技术， 2003， 博士

【摘要】 论文以航天装备体系及其它复杂飞行器系统的多学科设计优化(Multidisciplinary Design Optimization，简称MDO)为应用背景，针对航天装备卫星系统设计中存在大量离散和整数设计变量、设计空间非凸和不连通、具有多个局部最优点等特点和现有MDO方法过程难以处理离散和整数变量、对复杂设计空间非常敏感、易于陷入局部最优、不能有效处理多目标等缺陷，采用协同进化的思想，系统地进行基于分解协调的MDO方法研究。 给出了一种多学科设计优化问题的分解与协调方式，这种分解方式同时适用于单目标和多目标问题。证明了用这种分解方式分解后的问题与MDO原始问题的全局最优解(或全局Pareto最优解)的等价性，以及这种分解方式保留了MDO原始问题的所有局部最优解(或局部Pareto最优解)。 采用协同进化算法求解按上述方式分解后的问题，给出了协同进化MDO算法的基本过程，并分别构造了合作协同进化MDO算法和分布式协同进化MDO算法。采用NASA兰利研究中心MDO分部提供的四个MDO测试问题进行测试的结果表明，协同进化MDO算法在搜索全局最优解的能力方面明显优于其它MDO方法，其中分布式协同进化MDO算法比合作协同进化MDO算法具有明显的优势。 研究了异步并行的分布式协同进化MDO算法，并基于CORBA进行了网络分布式实现。不但实现了各学科的并行设计优化，而且同一学科内部也实现了并行优化，从而能大大加快设计进程，同时具有很好的灵活性、可伸缩性和容错性。测试表明异步并行的分布式协同进化MDO算法与单机顺序执行的分布式协同进化MDO算法的收敛性能相当。 给出了一种分布式的多目标进化算法，它是用分布式进化方式对Deb的NSGA-Ⅱ的改进，用五个困难的多目标优化测试问题对该算法的测试表明这种改进明显的提高了算法的性能。与其它多目标进化算法求解结果的比较显示了分布式多目标进化算法具有很强的求解多目标优化问题的能力。此外，分布式多目标进化算法还非常适合在网络环境的多台微机上异步并行实现。在分布式多目标进化算法研究的基础上，将非优超排序和排挤的多目标处理机制引入分布式协同进化MDO算法中，形成了多目标的分布式协同进化MDO算法。 将以上方法应用于某导弹的总体参数优化设计，同时也是对算法的进一步测试。计算结果表明：协同进化MDO算法求解该问题是有效的，其中分布式协同进化MDO算法优于合作协同进化MDO算法；异步并行的分布式协同进化MDO算法在保证收敛性能的同时大大加快了优化进程；多目标的分布式协同进化MDO算法仅一次运行就很好的逼近了问题的整个Pareto最优前沿，比用约束法求解Pareto最优集节省了大量计算开销，而且通过网络多台微机的分布式并行执行大大缩短了搜索时间。 针对不能直观确定星座结构的复杂卫星星座设计问题，采用一种可变维数优化的染色体表示和一组适合卫星星座设计的专用进化算子，设计了一种区域覆盖卫星星座结构与参数同时优化的进化算法，用一个虚拟的海洋监视卫星星座设计实例对算法进行了检验，结果表明该方法十分有效。 研究了一个具有离散／连续混合变量、需进行结构和参数同时优化的海洋监视卫星星座国防科学技术大学研究生院学位论文系统的多学科设计优化问题，作为算法在航天装备体系多学科设计优化中应用的示例。分析了卫星星座系统设计所涉及的学科及它们之间的祸合关系，特别是卫星设计中各分系统之间的祸合关系。建立了该海洋监视卫星星座系统的覆盖分析模型、卫星设计分析模型、卫星成本分析模型和发射费用分析模型。应用分布式协同进化MDO算法进行优化，实现了星座设计优化和卫星设计优化在自治基础上的充分协同，获得了比传统的串行设计方法的优化结果更好的解。 论文研究初步形成了一套适用于各种类型设计变量、复杂设计空间，可进行单学科、多学科、单目标、多目标设计优化，以异步并行方式在分布式计算机网络环境下运行，能保持各学科设计优化在自治基础上充分协同搜索系统整体最优，能以较大概率收敛到系统全局最优解的分布式协同进化MDO算法体系，为我国航夭装备体系及其它复杂工程系统的多学科设计优化提供了有效、实用的理论与方法基础。更多还原

【Abstract】 The background of this dissertation is the multidisciplinary design optimization (MDO) of the space equipment system and other complex flying vehicle systems. In the design optimization of space equipment satellite systems, there are lots of discrete and integer design variables, the design space is nonconvex and even disjointed, and has multimodality. Unfortunately, current MDO procedures or strategies have difficulty to deal with discrete or integer design variables, they are very sensitive to complex design space, have propensity to converge to local optima near the starting point, and can not handle multiple objectives effectively. To overcome these difficulties, this dissertation adopt the idea of coevolution to systematically develop new multidisciplinary design optimization methods based on decomposition and coordination.A new way of decomposing and coordinating MDO problems was given. It is suitable for both single objective and multiobjective MDO problems. Proofs were made that global optimum (or global Pareto optimum) of the decomposed problem and the original MDO problem are equivalent, and the decomposed problem retains all local optima (or local Pareto optima) of the original MDO problem.By solving the decomposed problem using coevolutionary algorithms, the revolutionary MDO algorithms were formed. The basic procedure of coevolutionary MDO algorithms was given and two coevolutionary MDO algorithms were constructed, i.e. cooperate coevolutionary MDO algorithm and distributed coevolutionary MDO algorithm. Test result on four MDO test problems from the MDO Test Suite of NASA Langley Research Center MDO Branch shows that coevolutionary MDO algorithms are evidently better than other MDO methods in searching global optimal solutions, and distributed coevolutionary MDO algorithm has obvious advantage than cooperate coevolutionary MDO algorithm.An asynchronous parallel distributed coevolutionary MDO algorithm was proposed and was implemented in network distributed environment using CORBA. Not only inter disciplinary design optimization was parallelized, but also inner disciplinary design optimization was executed in parallel, thus the optimization procedure can be speeded up greatly. The asynchronous parallel distributed coevolutionary MDO algorithm also has good flexibility, scalability and fault tolerance. Test result shows that convergence performance of the asynchronous parallel version of distributed coevolutionary MDO algorithm is similar to the sequential version.By introducing distributed evolution to Deb’s NSGA-II, a distributed multiobjective evolutionary algorithm was given. It was tested on five difficult multiobjective optimization test problems, and the result shows that distributed evolution does improve the performance of the algorithm. Comparing with the result of other multiobjective evolutionary algorithms reveals that the algorithm is powerful in solving multiobjective optimization problems. Furthermore, thealgorithm is very suitable for asynchronous parallel implementation in network distributed environment. On the basis of distributed multiobjective evolutionary algorithm, the nondominated sorting and crowding multiobjective handling mechanism was introduced to distributed coevolutionary MDO algorithm, and the multiobjective distributed coevolutionary MDO algorithm was formed.Above methods were applied to a missile design problem. It was also a further test to those methods. Computing result shows that: coevolutionary MDO algorithms are effective on this problem; distributed coevolutionary MDO algorithm is better than cooperate coevolutionary MDO algorithm; asynchronous parallel version of distributed coevolutionary MDO algorithm speeds up the optimization procedure greatly while maintains good convergence performance; multiobjective distributed coevolutionary MDO algorithm approximates the whole Pareto optimal front well in only one single run, saves much computing cost than constraint method to obtain Pareto optimal set, and greatly shortens search time by distri更多还原