Bartlett Test
Note that for calculating the confidence intervals, we assumed that the true variance 𝝈𝟐 is the same for
all observations and that the observations are independent.
● How can we check if this assumption is valid? Bartlett Test
H 0 : 12 = 22 = = m2 = y2 m = 2k (Total experiments)
H1 : at least one i2 2j , i j Sample size = n N = n1+n2+…+nm
2 𝑀 The value of M will be large if the sample variances si2 differ
𝜒𝑐𝑎𝑙 =
𝐶 greatly in magnitude, and will be zero if all the sample
where
𝑚 variances are exactly equal.
𝑀 = (𝑁 − 𝑚) ln(𝑠𝑝2 ) − (𝑛𝑖 − 1) ln(𝑠𝑖2 )
𝑖=1 We will reject H0 if c2cal is too large, i.e.,
𝑚
1 1 1
c cal c m2 −1,
𝐶 =1+ − 2
3(𝑚 − 1) 𝑛𝑖 − 1 𝑁−𝑚
𝑖=1
c m2 −1,
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
Statistical Design of Experiments and Data Analysis
Bartlett Test
Example: Bartlett Test
𝑀 1 1 1
2
𝜒𝜈=𝑚−1 = where, 𝑀 = (𝑁 − 𝑚) ln(𝑠𝑝2 ) − σ𝑚 2
𝑖=1 𝑛𝑖 − 1 ln(𝑠𝑖 ), and 𝐶 =1+ σ𝑚
𝑖=1 −
𝐶 3(𝑚−1) 𝑛𝑖 −1 𝑁−𝑚
Here, N = m =
The value of M will be large if the sample variances si2 differ
greatly in magnitude, and will be zero if all the sample
variances are exactly equal.
c c
S12 = 24.5, S22=21.78, S32=134.48, S42=242.0, 2 2
S52=3.92, S62=8.82, S72=33.62, S82=72.00 We will reject H0 if c2cal is too large, i.e., cal m −1,
2 [(𝑦𝑎1 −𝑦lj 1 )2 +(𝑦𝑏1 −𝑦lj 1 )2 +...............+(𝑦𝑎8 −𝑦lj 8 )2 +(𝑦𝑏8 −𝑦lj 8 )2 ]
𝑆𝑝 = (2−1)+...........+(2−1)
= 67.64
Yai Ybi Average
Test # X1 X2 X3
(kpsi) (kpsi) (kpsi)
1 -1 -1 -1 84 91 87.5
2 1 -1 -1 90.6 84 87.3
3 -1 1 -1 69.6 86 77.8
4 1 1 -1 76 98 87
5 -1 -1 1 77.7 80.5 79.1
6 1 -1 1 99.7 95.5 97.6
7 -1 1 1 82.7 74.5 78.6
8 1 1 1 93.7 81.7 87.7
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
Statistical Design of Experiments and Data Analysis
Example
The yield form, a certain chemical depends on
either the chemical formulation of the input materials or the mixer speed, or both.
A 2-level factorial design was run with three replicates and the yield data are shown below.
Find Main and Interaction Effects and their Confidence Intervals, and Significance using ANOVA
x1 x2 ya yb yc ഥ
𝒚
-1 -1 10 20 30 20
1 -1 40 30 50 40
-1 1 60 30 60 50
1 1 40 45 50 45
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
Statistical Design of Experiments and Data Analysis
x1 x2 ya yb yc ഥ
𝒚
-1 -1 10 20 30 20
1 -1 40 30 50 40
-1 1 60 30 60 50
1 1 40 45 50 45
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
Statistical Design of Experiments and Data Analysis
x1 x2 ya yb yc ഥ
𝒚
-1 -1 10 20 30 20
1 -1 40 30 50 40
-1 1 60 30 60 50
1 1 40 45 50 45
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
Statistical Design of Experiments and Data Analysis
Statistical Design of Experiments
ME 794 Instructors: Prof. Suhas Joshi, Prof. Soham Mujumdar (
[email protected])
Response Surface Methodology
Acknowledgement: Design and Analysis of Experiments by Montegamory. Some
of the course material has been adopted from similar courses taught previously by
Prof. Shiv Kapoor (UofI), and Prof. Suhas Joshi (IITB).
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
Goal of RSM
● So far, the focus of the design of experiments was ‘factor screening’ – which factors strongly affect the
process, which factors are less important, how the factors interact ..
● After screening, we now shift our focus to optimization – which factor level combinations give us
maximum (e.g. yield) or minimum (e.g. cost), or target result.
● The objective of Response Surface Methods (RSM) is optimization, finding the best set of factor levels to
achieve some goal.
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
ME 794 57
Example
Suppose, yield (𝑦) of a chemical process depends on temperature (𝑥1 ) and pressure (𝑥2 ). The chemical
engineer would like to find out which levels of temperature and pressure give the maximum yield.
One may write, 𝒚 = 𝒇 𝒙𝟏 , 𝒙𝟐 + 𝝐
Where ′ϵ′ is the error/noise observed in response ‘y’
The expected value of the response ‘y’ will be E(y) = 𝒇
𝒇 𝒙𝒙𝟏𝟏,,𝒙𝒙𝟐𝟐 𝒇 𝒙𝟏 , 𝒙𝟐
One could show this graphically,
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
ME 794 58
Sequential Process
‘RSM’ is sequential procedure
● In most problems, the exact relationship between the response variable and the independent variables is
unknown
● Therefore, the first step in RSM is to find a suitable approximation of the true functional relationship between
response and independent variables.
● Typically, the approximations are in the form of low-order polynomials in some region of independent variables
For example, if response (y) is well modeled by linear function of independent variables 𝑥1 , 𝑥2 , 𝑥3 , … , 𝑥𝑘 , then we
can write the approximate function as ‘first order model’
𝑦 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + … . +𝛽𝑘 𝑥𝑘 + ϵ
If there is curvature/non-linearity in the system, we must use polynomial of 2nd or higher degree,
For example, second degree model : 𝑦 = 𝛽0 + σ𝑘𝑖=0 𝛽𝑖 𝑥𝑖 + σ𝑘𝑖=0 𝛽𝑖𝑖 𝑥𝑖2 + ΣΣ 𝛽𝑖𝑗 𝑥𝑖 𝑥𝑗 + 𝜖
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
ME 794 59
● In real-problems, it is unlikely that these polynomials will provide reasonable
approximation of the true functional relationship over the ENTIRE range of
independent variables, but they work quite well for a relatively small region
● The coefficients in the RSM models (model parameters) are estimated using
least square method (least square fitting)
● The response surface analysis is then performed on the fitted surface
● The model parameters can be obtained more effectively if proper experimental
designs are used to collect the data (responses). Designs for fitting the response
surfaces are called response surface designs.
● Often we start at a point that is far from optimum such as the existing
operating conditions. If the region is linear, we use first order model.
● We then take the shortest and most efficient path towards the optimum
● As we near the optimum, there may be non-linearities, so we can employ
higher
NOTE: order
You do NOT modelsto share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
have permission
ME 794 60
Method of Steepest Ascent
● If we want to find maximum response, then we will be ‘climbing the hill’, if
we want to minimize the response, we will be ‘descending into a valley’
● We then take the shortest and most efficient path towards the optimum
● ‘Method of steepest ascent’ is a procedure of moving sequentially along the
path of steepest ascent, i.e., direction of the maximum increase in the
response.
● If minimization is desired, we follow the ‘method steepest descent’
● If we use first order model,
Then, the contours of y will be a set of parallel lines
So the path of steepest ascent will be along a line perpendicular to the contours
from center of the region
The actual step-size will be dependent on other practical considerations
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
ME 794 61
Example
Time, Temp
Region of (30, 40) minutes of time, and (150, 160) F temperature was
explored and responses were collected.
Note the experimental design is 22 factorial design augmented by five
center points. 5 replications at the center point [35, 155] allow
estimation of error as well as help us determine adequacy of linear
(first-order) model
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
ME 794 62
22 Factorial Analysis
Can we find which terms are important?
What will be the first-order model?
Will a first-order model be appropriate?
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
ME 794 63
ANOVA
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
ME 794 64
ANOVA
NOTE: You do NOT have permission to share this file or any of its contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
ME 794 65
‘Climbing the hill’
New model
NOTE: You do NOTneeds to betoemployed
have permission around
share this file or any of its [85, 175]
contents with anyone else, and/or upload it on internet or any of the platforms where it can be accessed by others.
ME 794 66
