&sixadvanced statistics_doe11 實驗目的 : &13 doe 簡介 對 y 影響最大的變數為...
Post on 22-Dec-2015
236 views
TRANSCRIPT
&Six Advanced Statistics_DOE1 1
實驗目的 :
&13 DOE 簡介
對 y 影響最大的變數為何?
如何設定 x1, x2, …, xp使 y 值趨近最佳值?
如何設定 x1, x2, …, xp使 y 值得變異最小?
如何設定 x1, x2, …, xp
使不可控制因素 z1, z2, …, zp 之影響最小?
&Six Advanced Statistics_DOE1 2
An Example: Play Golf Objective: Lower score without much practicing. Response Variable: Score (per round) Possible Factors:
The type of driver used (oversized or regular-sized)
The type of ball used (balata or three-piece) Walking or riding in a golf cart Beverage Type (water or beer) Time (in the morning or afternoon) Weather (cool or hot, windy or calm) The type of golf shoe spike (metal or soft)
&Six Advanced Statistics_DOE1 3
一般實驗進行方式 Best-guess approach
No Good, Guess Again Switching the levels of one (perhaps two) factors for
the next test based on the outcome of the current test
Good Enough, Stop! On-factor-at-a-
time Selecting a baseline starting point Varying each factor over its range with the other
factors held constant at the baseline level Interactions ruin everything
&Six Advanced Statistics_DOE1 4
最佳因子水準組合為? Driver: Regular Mode of travel: Ride Beverage: Water
But what if…………………
Results of the one-factor-a-time strategy for the golf experiment
&Six Advanced Statistics_DOE1 5
The two-factor factorial design for the golf experiment (I)
&Six Advanced Statistics_DOE1 6
The two-factor factorial design for the golf experiment (II)
Ball Effect = ? Ball-Driver Interaction Effect = ?
&Six Advanced Statistics_DOE1 7
Other Designs for the Golf Experiment
Four-factor factorial design
Three-factor factorial design
&Six Advanced Statistics_DOE1 8
Other Designs for the Golf Experiment
Four-factor fractional factorial design
&Six Advanced Statistics_DOE1 9
實驗計劃法 (DOE) 在一個或連串的試驗中刻意地改變製程輸入參數值 ,
以便觀察並找出影響製程輸出變數之因素 .
應用 : 改進製程產出率 降低製程變異 , 改善產品品質 降低研發時間 降低總體成本 評估各種可行之設定值 評估各替代原料 確定影響產品特性之因素
&Six Advanced Statistics_DOE1 10
Example:Optimizing a Process
&Six Advanced Statistics_DOE1 11
基本原則 複製 (Replication)
隨機化 (Randomization)
區隔化 (Blocking)
增進實驗之精確度
估計自然誤差 中央極限定理
“Averaging out” the effects from uncontrollable variables
&Six Advanced Statistics_DOE1 12
DOE 之程序 問題之認知與陳述 選擇因子與其水準 選擇反應變數 選擇適當之實驗設計 執行實驗 資料分析 結論與建議
Follow-up run and confirmation test Iterative No more than 25% of available resources should be invested in t
he first experiment
&Six Advanced Statistics_DOE1 13
Notes
使用統計以外之專業知識 實驗之設計與分析應愈簡單愈好 實驗之統計分析結果與現實上之差異
成本 技術 時間
實驗通常是遞迴式的 前幾次實驗通常只是學習經驗而已
&Six Advanced Statistics_DOE1 14
二因子實驗設計 二因子無交互作用
&Six Advanced Statistics_DOE1 15
二因子有交互作用
&Six Advanced Statistics_DOE1 16
One-factor at a time 之方法
&Six Advanced Statistics_DOE1 17
&Six Advanced Statistics_DOE1 18
二因子實驗設計之模式
0 oneleast at :
, allfor ,0: 3.
0 oneleast at :
0: 2.
0 oneleast at :
0: 1.
:
1
0
1
210
1
210
ij
ij
j
b
i
a
H
jiH
H
H
H
H
TestingHypothesis
&Six Advanced Statistics_DOE1 19
Data Sheet
&Six Advanced Statistics_DOE1 20
ANOVA 表 – Two-Factor Factorial
&Six Advanced Statistics_DOE1 21
a
i
b
j
ijsubtotals abn
y
n
ySS
1 1
2...
2.
&Six Advanced Statistics_DOE1 22
Example
&Six Advanced Statistics_DOE1 23
&Six Advanced Statistics_DOE1 24
決策模式: 因為 F0(Primer Types) = 28.63 > F0.05,2,12 = 3.89
F0(Application Methods) = 61.38 > F0.05,1,12 = 4.75
所以此二因子對黏著力皆有顯著影響。 但 F0(Interaction) = 1.5 < F0.05,2,12 = 3.89 ,所以此二因子的交互作
用對黏著力無明顯之影響。
&Six Advanced Statistics_DOE1 25
Another Example
&Six Advanced Statistics_DOE1 26
ANOVA Results
Computer Output (Model Adequacy Checking)
&Six Advanced Statistics_DOE1 27
Multiple Comparisons
&Six Advanced Statistics_DOE1 28
&6 2k 因子階層設計 k 個因子,每個因子 2 個水準 (+,-) ,共 2k次實驗(當
n = 1 時)。
在因子數不多的狀況下,常用於實驗初期,來了解因子對反應變數之可能影響。
只能看出因子對反應變數之線性作用 (linear effect) ,無法預估高階曲面作用。
&Six Advanced Statistics_DOE1 29
The 22 Factorial Design
&Six Advanced Statistics_DOE1 30
The Calculation (I)
Effects
Contrast
baabn
ababn
AB
ababn
baabn
B
baabn
ababn
A
12
11
2
1
12
11
2
1
12
11
2
1
12
1
1
1
kA
AB
B
A
n
ContrastA
baabContrast
ababContrast
baabContrast
&Six Advanced Statistics_DOE1 31
Sum of Square Errors
The Calculation (II)
ABBATE
ijkijkT
kAB
AB
kB
B
kA
A
SSSSSSSSSSn
yyyySS
n
ContrastSS
n
ContrastSS
n
Contrast
n
baabSS
4
2
2
24
1
222
2
2
22
&Six Advanced Statistics_DOE1 32
The AVONA Table
&Six Advanced Statistics_DOE1 33
The Regression Model Since the effect of AB is not significant, the
regression model would be
22110 xxy
The estimates of these coefficients are
2
2
2
^
^
1
^
0
B
A
Effect
Effect
y
Why?
&Six Advanced Statistics_DOE1 34
The Residuals
Residuals at x1=1 and x2 = -1
Computer Output
&Six Advanced Statistics_DOE1 35
23 因子階層設計
&Six Advanced Statistics_DOE1 36
23 因子階層設計 _ 符號表
&Six Advanced Statistics_DOE1 37
計算 各因子之 Contrast = (Sign) (Treatment Combination)
ContrastA = -(1)+a-b+ab-c+ac-bc+abc ContrastB = -(1)-a+b+ab-c-ac+bc+abc ContrastAB =
各因子之效用 Effect = Contrast / (n 2k-1)
各因子之 Sum of Square = Contrast / (n 2k) SSA = {-(1)+a-b+ab-c+ac-bc+abc}2 / 8n
&Six Advanced Statistics_DOE1 38
Example for 23 Design
A: 速度 B: 切割深度 C: 切刀角度
&Six Advanced Statistics_DOE1 39
計算 _Example 各因子之效用
各因子之 Sum of Square SSA = {-(1)+a-b+ab-c+ac-bc+abc}2 / 8n = (27)2 / (82) = 45.5625
&Six Advanced Statistics_DOE1 40
ANOVA 表 _Example
&Six Advanced Statistics_DOE1 41
Example for 24 Design
&Six Advanced Statistics_DOE1 42
24 因子階層設計 _ 符號表
請完成
&Six Advanced Statistics_DOE1 43
ANOVA 表 _Example
&Six Advanced Statistics_DOE1 44
AD 交互作用與迴歸函數
2/
2/
2/
3
2
1
3210
之效用為
之效用為
之效用為
AD
D
A
ADDAy
&Six Advanced Statistics_DOE1 45
2k Design with Center Points
增加預估曲線作用之能力
不破壞設計之平衡性 (Balanced Design)
只需增加少數幾個實驗
cF
cFcFquadraticpure nn
yynnSS
)(
&Six Advanced Statistics_DOE1 46
Example
&Six Advanced Statistics_DOE1 47
ANOVA 表 _Example
&Six Advanced Statistics_DOE1 48
2k 因子實驗之區隔化與混雜化 22 factorial design with Blocking
&Six Advanced Statistics_DOE1 49
The ANOVA
&Six Advanced Statistics_DOE1 50
Confounding( 混雜化 ) 受限於資源(時間、金錢、人力等),無法再每一個
區隔皆有完整的因子實驗。
Confounding is a design technique for arranging a complete factorial design in blocks, where the block size is smaller than the number of treatment combinations in one replicate.
It causes information about certain treatment effects (usually high-order interactions) to be indistinguishable from, or confounded with, blocks.
&Six Advanced Statistics_DOE1 51
Confounding in 2 blocks
Effect AB confounded with block.
&Six Advanced Statistics_DOE1 52
The 23 Design Confounded(I)
&Six Advanced Statistics_DOE1 53
The 23 Design Confounded(II)
The Degrees of Freedom
&Six Advanced Statistics_DOE1 54
The 23 Design Confounded(III) Replications => n = 4.
&Six Advanced Statistics_DOE1 55
The 23 Design Confounded(IV)
Replications => n = 4.
&Six Advanced Statistics_DOE1 56
An Example
&Six Advanced Statistics_DOE1 57
The ANOVA
&Six Advanced Statistics_DOE1 58
Confounding in 4 blocks
ADE and BCE are confounded, in addition, ABCD is also confounded. Why?
Suggested Blocking => pp. 298, Table 7-8.