GPU并行计算及其在飞行器设计中的应用

发布时间：2018-03-23 15:46

本文选题：粒子群优化　切入点：弹道优化　出处：《北京理工大学》2015年硕士论文

【摘要】：现代飞行器设计是一个典型的多学科设计优化的过程,广泛存在着复杂、费时的仿真分析模型,如:基于有限单元法的结构分析和气动分析,以及学科间的耦合,如:结构和气动的耦合,因此计算量大成为现代飞行器优化设计中一个显著的难题。众所周知,优化算法的效率很大程度上决定了整个优化设计的效率。作为一种性能较为优越的全局优化算法,粒子群优化(Particle Swarm Optimization,PSO)由于算法结构简单、具有全局搜索能力,在飞行器设计中被广为关注。但是,PSO通过在整个设计空间随机进行搜索而保证最优解的全局最优性,计算时间较长,在现代飞行器设计中应用十分受限。因此,迫切需要提高PSO算法的计算速度。现阶段从算法本身的层面对PSO进行改进,来实现加速的目的几乎已达到一个瓶颈。近年来,图像处理器(Graphics Processing Unit,GPU)因其具有良好的浮点计算能力、高并发度、以及相对廉价的特点,被广泛应用于通用计算领域,即GPGPU(General-Purpose Computing on Graphics Processing Units),在科研及工程领域皆具有巨大的潜力。2007年,NVIDIA推出统一计算架构(Compute Unified Device Architecture,CUDA)并行计算平台,大大推广了GPU的应用,GPU现已被广泛应用于流体力学、有限元仿真、分子动力学等领域。因此,本文首先针对PSO优化费时的问题,提出CUDA平台下基于GPU并行计算的PSO算法,充分利用PSO所具有的并行计算的基本构架,采用GPU对粒子群算法进行细粒度并行化,即将每个粒子的速度位置初始化、适应度估计及速度位置更新同时并行化,实现PSO算法的全面加速计算,大为缩减PSO的计算时间。其次,弹道优化作为飞行器总体优化设计中一个重要的学科,其所得结果的精度和效率对飞行器总体设计具有重要影响。因此,本文提出利用工程上广为应用的直接法——直接打靶法,来离散弹道优化问题,首先将最优控制问题转换为非线性规划问题,然后采用上述提出的并行PSO算法进行求解,将弹道仿真子程序所得的飞行距离作为适应度函数,大幅度缩短了计算时间,为工程及科研提供了新的弹道优化解决方案。此外,基于有限单元法的结构和气动分析在现代飞行器设计中广泛应用,然而计算量大一直是其存在的显著问题。因此,代理模型通常用于替代有限元仿真来进行气动和结构分析,但代理模型必然会引入误差,甚至会产生错误的估算结果,而且为了保证代理模型的精度也需要抽取大量的样本,计算量依然很大。因此,本文从并行计算的角度,提出采用GPU来加速有限元仿真分析(ABAQUS和FLUENT)过程,在避免代理模型引入误差的同时,大为提高了计算效率。通过诸多实例的仿真计算分析以及在飞行器设计中的应用,得出本文提出和研究的诸多基于GPU并行计算的加速策略是有效的,相比于传统的基于CPU(Central Processing Unit)串行计算的方式,能够获得非常可观的加速比,因此GPU并行计算在飞行器设计中具有巨大的应用潜力和广阔的应用前景。
[Abstract]:Modern aircraft design is a typical multidisciplinary design optimization process, there is a wide range of complex and time-consuming simulation models, such as: analysis and dynamic analysis of structure based on finite element method, and the coupling between the disciplines such as structural and aerodynamic coupling, so the computation has become a modern aircraft optimization a significant problem in the design. As everyone knows, the optimization efficiency largely determines the efficiency of the whole optimization design. As a kind of better performance of global optimization algorithm, particle swarm optimization (Particle Swarm Optimization, PSO) because the algorithm has the advantages of simple structure, has the global search ability in aircraft design, but has been widely concerned. To guarantee the optimal solution, the global optimality of PSO by random search in the design space, the computation time is long, ten points in limited application in modern aircraft design. Therefore, urgent To improve the calculation speed of the PSO algorithm. At this stage, PSO improved the algorithm itself in, to achieve the purpose of accelerating has almost reached a bottleneck. In recent years, the image processor (Graphics Processing Unit, GPU) computing ability because of its good floating point, and high degree, and relatively inexpensive, is widely used in the field of general-purpose computing, namely GPGPU (General-Purpose Computing on Graphics Processing Units), in scientific research and engineering field has great potential for.2007 years, NVIDIA launched a unified computing architecture (Compute Unified Device Architecture, CUDA) parallel computing platform, greatly extended the application of GPU, GPU has been widely used in fluid mechanics, finite element simulation, the field of molecular dynamics. Therefore, in this paper, PSO optimization time-consuming, GPU parallel PSO algorithm based on CUDA platform is proposed, taking full advantage of PSO has the parallel computing framework, using GPU on the particle swarm algorithm for fine-grained parallelism, speed and position of each particle is initialized, fitness and speed estimation and location update parallelization, PSO algorithm to achieve the overall acceleration of computation, for computing time reduced PSO. Secondly, as the aircraft trajectory optimization overall optimization is an important subject in the design, the efficiency and accuracy of the results has an important influence on the overall design of the vehicle. Therefore, this paper proposes the use of the wide engineering application of direct method -- direct shooting method to discrete trajectory optimization problem, the optimal control problem into a nonlinear programming problem, then using the proposed the parallel PSO algorithm to solve the trajectory simulation of the flight distance of the subroutine as fitness function, greatly shorten the calculation time, for engineering and Science The research provides a new solution scheme of trajectory optimization. In addition, analysis is widely used in modern aircraft design and dynamic structure based on finite element method, but the large amount of calculation has been significant problems. Therefore, the models are often used to make aerodynamic and structural analysis instead of finite element simulation model, but the agent must be the introduction of error, and even produce error estimation results, and in order to ensure the accuracy of surrogate models also need to extract a large number of samples, amount of calculation is still great. Therefore, this paper from the perspective of parallel computing, GPU is proposed to accelerate the simulation of finite element analysis (ABAQUS and FLUENT) in the process, to avoid the agent error model is introduced at the same time, to improve the computational efficiency. The simulation calculation and analysis of many examples and applications in aircraft design, it is proposed in this paper and the research of many parallel computing based on GPU The acceleration strategy is effective. Compared with the traditional way based on CPU (Central Processing Unit) serial computing, it can achieve a very substantial speedup. Therefore, GPU parallel computing has great potential and wide application prospects in aircraft design.

【学位授予单位】：北京理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：V221

【参考文献】