基于仿真实验的灵敏度分析方法及工具研究

发布时间：2018-11-04 14:25

【摘要】：与传统的实验相比，仿真实验具有因子数多、各因子对应的水平数多以及指标多等特点。如果对每个因子都使用定量灵敏度分析，获取各因子的不确定性对仿真输出不确定性的影响，实验开销过大，不切实际。而且Pareto定理或20-80法则认为，只有少量因子对仿真输出起到重要作用。针对这一问题，本文结合仿真实验的特点，从灵敏度分析的角度出发，开展了以下几个方面研究。首先，针对仿真实验的特点，实验分析人员需要先筛选出重要因子，，再定量分析这些重要因子的灵敏度系数。本文基于这个思想，提出了灵敏度分析过程模型；研究了典型的全局灵敏度分析方法，并对这些方法的特点进行了分析比较；结合灵敏度分析方法和实验设计的特点，提出了基于混合推理的灵敏度分析智能化技术，指导实验分析人员合理高效地完成灵敏度分析工作。其次，由于因子筛选和定量分析是灵敏度分析过程的核心，实验分析人员只有在了解筛选实验设计方法的基础上才能更高效地筛选因子。同时，定性或定量灵敏度分析需要通过采样产生样本数据。因此，本文一方面研究了顺序分支法及其衍生的方法、分式析因设计等筛选实验设计方法，另一方面研究了蒙特卡罗法、拉丁超立方抽样等采样实验设计方法。最后，由于灵敏度分析过程比较复杂、工作量大，这给实验分析人员的工作带来困难。为了提高灵敏度分析的效率和准确性，本文设计并实现了灵敏度分析辅助工具，协助实验分析人员分析灵敏度。该工具采用实验项目分阶段的管理模式和基于混合推理的智能化设计模式，并在实际灵敏度分析问题中得到应用，验证了工具的有效性。
[Abstract]:Compared with the traditional experiment, the simulation experiment has many characteristics, such as the number of factors, the number of levels corresponding to each factor and the number of indexes. If the quantitative sensitivity analysis is used for each factor and the influence of the uncertainty of each factor on the uncertainty of simulation output is obtained, the experiment cost is too large and impractical. Moreover, only a few factors play an important role in the simulation output, according to Pareto's theorem or the 20-80 rule. According to the characteristics of simulation experiments, the following aspects are studied in this paper from the point of view of sensitivity analysis. Firstly, according to the characteristics of the simulation experiment, the experimental analysts need to screen out the important factors first and then quantitatively analyze the sensitivity coefficients of these important factors. Based on this idea, the process model of sensitivity analysis is proposed, the typical global sensitivity analysis methods are studied, and the characteristics of these methods are analyzed and compared. According to the characteristics of sensitivity analysis method and experimental design, an intelligent sensitivity analysis technique based on hybrid reasoning is proposed to guide the experimental analysts to accomplish sensitivity analysis reasonably and efficiently. Secondly, because factor screening and quantitative analysis are the core of sensitivity analysis, experimental analysts can screen factors more efficiently only on the basis of understanding the design methods of screening experiments. At the same time, qualitative or quantitative sensitivity analysis needs to produce sample data by sampling. Therefore, on the one hand, the sequential branching method and its derivation, the fractional factorial design and other screening experimental design methods are studied, on the other hand, the Monte Carlo method, Latin hypercube sampling and other sampling experimental design methods are studied. Finally, the process of sensitivity analysis is complicated and the workload is heavy, which brings difficulties to the work of the experimental analysts. In order to improve the efficiency and accuracy of sensitivity analysis, an assistant tool for sensitivity analysis is designed and implemented in this paper. The tool adopts the experimental project management model and the intelligent design pattern based on mixed reasoning, and is applied to the practical sensitivity analysis, which verifies the effectiveness of the tool.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.9;TP302.7

【相似文献】