22年9月读的图书《试验设计方法与Design-Expert软件应用》,其中有基本的功能介绍,参见 《试验设计方法与Design-Expert软件应用》 part1 实验设计、软件、方差分析,这儿对照软件和软件的说明书,按照功能介绍和操作步骤,简单整理,加深对软件的理解。
先列几个重要问题,整理完本文回来写答案。
Q1: 四种实验设计之间的功能差异和计算逻辑差异,如何选择:Factorial, RSM, mixture, combined design –> 析因实验(全析因,部分析因),响应曲面,混料设计,复合设计;其中部分析因主要用于筛选实验;发现存在弯曲时补充实验做响应曲面分析,混料设计针对配方开发,复合设计是在配方开发时同时研究过程因子。
Q2: mixture和combine的差异是什么?何时用combined design,何时将combined分成独立的factorial和mixture?如果一个DOE本来可以用mixture,却实际使用了combined design,会有什么影响? –> 如果我理解没错,combined design的process和mixture之间不存在交互作用的研究,这是关键,只是分别用析因实验的分析方法。(2024-8-3)
Q3:Design expert可以分析三水平实验设计吗,比如三因子三水平实验设计。 –> 可以,但是最好还是先做两水平,通过中心点判断是否弯曲,再补充augmented design
Q4: 析因设计,几种分析方法的差异是什么,应该如何选择?对试验次数和分析精度的影响是什么。特别是标准两水平设计regular two-level, min-run characterize, min-run screen这三种方法之间的差异! –> 把最基础的regular two level弄明白即可,其他都是相应的特例应用。
Q5:筛选实验screen和刻画实验characterize之间的差异是什么? –> 前者是筛选实验,只看主效应,针对大于5个因子的情况,解析度达到4或以上即可,实验次数更少;后者是“改进实验”,同时研究主效应和二交互,所以要确保解析度为5或以上。
0. Design Expert 功能
Design Expert是纯粹的DOE软件,聚焦实验设计功能,直接选择特定的设计方法和操作步骤即可,没有Minitab中的通用统计功能。
- Standard designs 标准实验设计;最新版包含了第四种设计“combined 混合设计”
- Factorial design 析因
- Response Surface 响应曲面
- Mixture 混料
- combined design 复合实验
- Custom Designs 自定义设计
- optimal (combined)
- user -defined
- historical data
- simple sample

每一种实验设计,都包括三个步骤:
- design:软件对设计者的试验方案做出总结,对设计方案的好坏做出评估。
- analysis:方差分析,回归分析,(transform,effects排列,anova,diagnostics,model graphs)
- optimization:预测最优值,给出最优试验条件以便验证
1. Factorial Design part 析因设计
析因设计:几因子、几水平对响应的关系,不同的方法的细节差异是,介绍哪种程度的混杂(比如2交互被混杂,还是只能接受3交互被混杂,抑或只能接受4交互被混杂,还是完全不能混杂那就是全析因设计),另外还要看变量是连续型还是离散型变量;这样就细分成了很多析因设计。

Q: the difference between screen (resolution >=IV) and characterize (resolution>=V)
答案:核心是解析度的差异!筛选实验的目的是筛选因子,只需要保证主效应不被混杂即可,所以解析度必须大于3,最常用自然是4,这样二交互之间会混杂,但是不影响主效应。 而刻画实验更细致,会研究主效应和二交互,所以解析度必须大于4,也就是V=5或全析因。
刻画实验就像Minitab教程中的“改进实验”,目的是研究主效应和二交互作用。
A: Screen DOE only to screen main effects, don’t screen the interaction. Make sure the resolution should be larger than III, otherwise, main effects will be aliased with two interaction. Therefore, resolution is usually be IV or V, or full factorial DOE. if there are lots of factors, like 10, or 20 or more, screen DOE is firstly run to screen and find the key factors. characterize and RSM then be followed. Only when we believe there is no interaction, screening DOE with resolution III can be used.
A: characterize DOE is to analyze both main effects and two interactions. therefore, resolution should be no less than V.
汇总:
- regular two level 两水平析因实验
- Mini-run characterize, resolution = V, less runs
- Mini-run Screen , resolution = IV, much less runs
1.1 Regular Two-level
Design for 2 to 21 factors where each factor is set to 2 levels. Useful for estimating main effects and interactions. Fractional factorials can be used for screening many factors to find the significant few. The color coding represents the design resolution:
- Green (Characterization) = Res V or higher, ——推荐使用
- Yellow (Screening)= Res IV——根据实际情况考虑
- Red (Ruggedness testing)=Res IIl. ——慎用

1.2 Mini-run characterize (Resolution V )
Minimum-Run Resolution V Characterization Design
Design for 6 to 50 factors where each factor is set to 2 levels. Resolution V designs will allow estimation of main effects. Two-factor interactions will only be aliased with three-factor and higher interactions. Excellent designs to reduce the number of runs and still obtain clean results.
Q:和regular two level方法 选择解析度=V的选项,有什么区别?
A:比regular two level的试验次数更少。
regular two level只能执行2^k个试验次数(4,8,16,32,64, etc.) ,比如6因子两水平,只有8,16,32和64次实验。而选择min-run characterize,就可以执行解析度=V的试验次数,只需要22次,介于16和32之间。
1.3 Mini-run Screen (Resolution IV)
备注:解析度=IV,所以同样因子数量下,run times比解析度=V的characterize design多。
Minimum-Run Resolution IV Screening Design
Design for 5 to 50 factors where each factor is set to 2 levels. Resolution IV designs will allow estimation of main effects. Two-factor interactions will be aliased with other two-factor and higher interactions. Good designs to reduce the number of runs if interactions are unlikely.
Note:和min-run characterize的问题类似,只不过min-run screen的解析度=IV,不需要分析二交互作用,所以试验次数更少,
Q:假设5因子2水平的实验,分别选择regular, min-run characterize, and min-run screen,差异是什么?
A:选regular two level,在5 factor下可以分别选择8次(V=3),16次(V=4),和32次(V=5,全析因)试验次数; 选min-run characterize,factor最低是6次,所以不能做5 factor; 选min-run screen, 解析度还是4,试验次数是12次;就比regular的16次少一些。这样看来,factor越多,后两种方法越有优势,如果只是五六个因子,那就老老实实用regular two level吧
总结以上Factorial – Randomized 中的三种DoE方法的差异。
- regular two-level可以用到更少的factors(改进和优化实验),而runs只能是2的几次方;
- min-run characterize 和min-run screen锁定了resolution,而且最少factor分别是6和5次(factor再少就不存在解析度为5和4了),后两种方法的优势是,保证同样的V,试验次数更少,自然screen比characterize更少,而且少很多。
- regular two -level是标准的两水平析因设计; 而min-run characterize和min-run screen主要针对很多factor(大于5个,多多益善)的情况,一个是改进(刻画),一个是更粗略的筛选实验。
resolution | factors | runs | |
regular two levels | 3, 4, 5, etc. (max=factors) | 2,3,4,…, to 21 | 2^k: 4,8,16,32,64,…, 512 |
min-run characterize | 5 | 6 to 50 | 22, 30, 38, 46… |
min-run screen | 4 | 5 to 50 | 12, 14, 16, 18 , (min runs +2) |
factor=4 | factor=5 | factor=6 | factor=9 | factor=12 | |
regular two levels | run= 8, 16 | run=8,16,32 | run=8,16,32,64 | ||
min-run characterize | / | / | run=22 | run=46 | run=80 |
min-run screen | / | run=12 | run=14 | run=20 | run=26 |
1.4 Multilevel categoric / General Factorial 多水平离散因子分析
Also known as “general factorial”, this is a design for 1 to 12 factors where each factor may have a different number of levels. factors are treated as categoric.
Nominal:level之间没有好坏、高低之分,比如性别,国籍等。
Ordinal: level 之间有好坏、高低之分,比如满意度,比如触感评分,比如收入梯度。

Categoric Factor is Numeric
If numeric factors are entered as categoric factors predictions are limited to the tested settings. Polynomial models, generated by response surface designs, allow for interpolated predictions.
Categoric models can be more complex than necessary to fit the response surface. A smaller design for a less complex model can be found when numbers are treated as numeric factors. Use the discrete type to limit the design to fixed settings when the response surface design is built.
1.5 Optical (custom) 优化实验,只针对Categoric factor?
A flexible design structure to accommodate custom models, categoric factors, and irregular (constrained) regions. Runs are determined by a selection criterion chosen during the build.

1.6 miscellaneous 其他设计(细节略)
Irregular Res V: 两水平,解析度5,只分析主效应和二交互,试验次数少于“1.1 regular two level” (why?)
Plackett-Burman (PB): 2-47个factor,两水平,不考虑二交互,是一种 特殊的筛选实验
Taguchi OA: 田口设计,暂略
1.7 Split-Plot 区组化(细节略)
针对HTC (hard to change)因子,进行区组化实验设计。
软件的help部分,有一些案例和数据分析:

2022-10-9 overnight 软件设置截图
2023-1-15 重新梳理1.1到1.3的factorial design的内容,整理Q&A和方法差异,自己列表格做对比。
2024-8-4 接着最近回顾DoE的机会,重读一年前多的博文草稿,回答一开始的问题;发布