Professional Documents
Culture Documents
Asset-V1 TUMx+QPLS2x+2T2019+type@asset+block@Six Sigma Statistics With Minitab
Asset-V1 TUMx+QPLS2x+2T2019+type@asset+block@Six Sigma Statistics With Minitab
Six Sigma
Statistics with Minitab
Copyright
Dr. Reiner Hutwelker
[email protected]
Content
Analysis Strategies 3
Graphical Analysis 21
Literature 135
Analysis Strategies
In the DEFINE-Phase the critical output variables (Y= CtQ´s= Problems) will be identified …
Data Collection Variable Scale Scale Level Hypothesis Chart Perf.-Ind. Test
Yt1_Q_1_n Error Type nominal Pareto-Diagram
Yt1_T_1_n Cycle Time cardinal Histogram Sigma Level Correlation
Yt1_R_1_n Rating ordinal Mood-Median
XI1_n cardinal
XM1_n ordinal
XR1_n nominal
Relationship Hypothesis
Problem-Focus Choc-Cookie Van-Cookie
2. Taste bad Y1 =/ ≠ Y2 …
Taste bad top-down
(Differentiation) (Test: Chi2, t-Test/ ANOVA)
(Y1) (Y2) 5 x WHY …? BECAUSE …!
Y= f(x) 1. Δ= 1,52 grades
R2= 96,5% 1. Why? n. If …,
1. Cause-Level Because! then …!
(Causation by Trigger) weight of chocolate
too low (x)
3. Why? 2. If …,
Chef has too few Because! then …!
3. Cause-Level Quality of chocolate is
to nibble chocolate is a opportunities to test
(Causation)
a determinant of Cust.
basic need (x1.2.1.1) (eat) chocolate
Satisfaction (x1.2.2.1)
(x1.2.3.1)
… as well as quantitative statistical and qualitative logical reasoning to identify the Root-Causes
Dr. Reiner Hutwelker Lean and Six Sigma 6
DMAIC > Statistics >> Hints and typical errors in Minitab Technische Universität München
1.a
Dialog
2.a
1.a All Y data are in one column
1.a
1.b Y data are in more than one column
1.b
1.b 2.b
1.a All Y data in one column with a grouping x - or - 1.b Y data grouped in different columns
Dr. Reiner Hutwelker Lean and Six Sigma 8
DMAIC > Statistics >> Hints and typical errors in Minitab Technische Universität München
The type of data must be appropriate for the procedure. Minitab offers
different types of data, e.g.: a) numeric, b) text and c) date/ time. The type of
date in a column is indicated in the top row.
The data type can be changed, if necessary, with: Data/ Change Data Type
…
All columns of data, used for the same procedure, must have the same
length. If they do not have the same length, then enter a value in the
„shorter“ column, one position under the needed length. The empty cells will
then be filled with the Missing Value indicator. At least delete this value
again. The Missing Value indicator will remain and the columns have the
same length.
If you enter data manually, then you can change the direction for entering
data – in a column or in a row – by a click on the arrow.
…
Dr. Reiner Hutwelker Lean and Six Sigma 9
DMAIC > Statistics >> Measurement System Analysis (MSA) Technische Universität München
Measurement System
Analysis (MSA)
Long-Term- Short-Term-
Sampling Error Reliability Validity
Variation Variation
Unique relation of measured value and scale Reproducibility: error between operator
The Measurement System Analysis (MSA) helps to reduce the size of error
Dr. Reiner Hutwelker Lean and Six Sigma 11
DMAIC > Statistics >> Measurement System Analysis (MSA) Technische Universität München
Variation between
Variation overall = Outputs + Variation of Measurement System
Measured Values True Values Error
SS overall SS between Outputs SS Measurement System(Gage R&R)
consistency consistency
within each between
operator operator
Repeatability Reproducibility
SS Repeatability SS Reproducibility
Repeatability
same output l
l
l
l
same operator l l
l
l
l
l
l l
l
l
l
l l
Operator A
Operator A
Reproducibility
Operator B
2nd Trial
Operator B
Operator C
2. 1.
5. 7.
1.
Purpose
1. a. c.
Evaluates the agreement of subjective Nominal-/ Ordinal-Scale based ratings by multiple
Operators to determine how likely a Measurement System is to misclassify an Output 2. b.
Focus 3.
Example
Y Scale Level
1 nominal/ ordinal Dialog: Create Worksheet
x Scale Level a. Number of Determine the Operators, which should be included as appraisers into the study; if
./. appraisers possible select all involved Operators
./.
Alternative in Minitab Stat Menu
b. Number of trials Select at least 2 trials for each appraiser, good style is 3
Stat/ Quality Tools/ Attribute Agreement Analysis
To use the real names of the appraiser supports specific training, that may become
c. Name of appraiser
note … necessary; but aka-names might be better, starting a series of MSA´s
Assure a balanced mix of: 0/good and 1/bad items. A small percentage of items of one
type reduces the ability to assess R&R d. Number of items select at least 10 items (5 good/ 5 bad)
e. values for good/ to differentiate the good vs. bad items, attributive and numeric labels can be assigned (My
bad preference: 0= good; 1= bad)
classify the test items according to their attributes; the items will be randomly arranged in
f. classify items
the worksheet, to avoid sequence effects;
Example: Evaluate the R&R of good vs. bad Inputs or Outputs (like Decisions or M&M´s)
Dr. Reiner Hutwelker Lean and Six Sigma 16
DMAIC > Statistics >> Measurement System Analysis (MSA) >>> Attributive Agreement Analysis Technische Universität München
1.
a.
b.
c.
d.
e.
f.
b. Name of appraiser for the specific evaluation b. Column with trial numbers (in previously created Worksheet)
c. specific trials for specific appraiser c. Column with names of the test items (in previously created Worksheet)
d. item, the appraiser should evaluate in this evaluation d. Column with answers/ results of evaluations (entered in previously created Worksheet)
result of evaluation of appraiser; enter the coded answers here, e.g. 0 for good of 1 for
e. e. Column with the standard evaluations from experts (in previously created Worksheet)
bad
f. previously - by experts - defined standard evaluations of the test items f. selection of the coded value which was selected for the category "good"
Example: Evaluate the R&R of good vs. bad Inputs or Outputs (like Decisions or M&M´s)
Dr. Reiner Hutwelker Lean and Six Sigma 17
DMAIC > Statistics >> Measurement System Analysis (MSA) >>> Attributive Agreement Analysis Technische Universität München
1.
2.b Overall % Accuracy: If 240 appraisals match the standard, the accuracy rate is:
(240/300) x 100 = 80% Item 5 Item 9
Overall Error Rate: If 60 appraisals do not match the standard, the error rate is: 2.
2.c (60/300) x 100 = 20% Item 4 Item 10
Good rated Bad: If appraisers rate a Good item as Bad 30 times, the misclassification rate is: 0 15 30 45 60 0 15 30 45 60
2.d (30/150) x 100 = 20%
Bad rated Good: If appraisers rate a Bad item as Good 15 times, the misclassification rate is: 5.
2.e (15/150) x 100 = 10%
4.d 4.e
Appraiser Misclassification Rates
4.f
% 0 rated 1 % 1 rated 0 % Rated both ways
Rated both ways: If appraisers rate 15 items inconsistently across trials, the misclassification rate is:
2.f (15/150) x 100 = 10%
A A A
3.
Result: Report Card B B B
Mix of Items: the highest power to differentiate between appraisers results from a
1.
balanced mix of good (0) vs. bad (1) items. C C C
- % good rated good (= hit) 3.e Evaluation of Items: % bad rated good (= false alarms)
- % bad rated bad (= correct rejection) 4.d Evaluation of Appraiser: % good rated bad (= missed signals)
b. - Overall % Accuracy= % hit + % correct rejection 4.e Evaluation of Appraiser: % bad rated good (= false alarms)
Example: Evaluate the R&R of good vs. bad Inputs or Outputs (like Decisions or M&M´s)
Dr. Reiner Hutwelker Lean and Six Sigma 18
DMAIC > Statistics >> Measurement System Analysis (MSA) >>> Attributive Agreement Analysis Technische Universität München
1.
C
B
c. % Accuracy by Appraiser d. Comments
D
57,5%
Consider the following when assessing how the measurement system
0 25 50 75 100 C can be improved:
• Low Accuracy Rates: Low rates for some appraisers may indicate a
need for additional training for those appraisers. Low rates for all
b.2 % by Standard A 100,0
appraisers may indicate more systematic problems, such as poor
0 D operating definitions, poor training, or incorrect standards.
• High Misclassification Rates: May indicate that either too many 0
items are being rejected, or too many 1 items are being passed on to
1 1 the consumer (or both).
• High Percentage of Mixed Ratings: May indicate items in the study
B 80,0
0 25 50 75 100 were borderline cases between 0 and 1, thus very difficult to assess.
A
% by Trial
b.3 B
C 50,0
1
2 D 0,0
D
Comments:
d.
Summary and comments about results
Example: Evaluate the R&R of good vs. bad Inputs or Outputs (like Decisions or M&M´s)
Dr. Reiner Hutwelker Lean and Six Sigma 19
DMAIC > Statistics >> Measurement System Analysis (MSA) Technische Universität München
1 ok ok ok ok ok ok ok Y
2 defect defect defect defect defect defect defect Y
3 ok ok defect ok ok ok ok N
4 ok ok ok ok ok ok ok Y
5 ok ok defect ok ok ok ok N
6 ok ok ok ok ok ok ok Y
7 ok ok ok ok ok ok ok Y
8 defect defect defect defect defect defect defect Y
9 defect defect defect defect defect defect defect Y
10 ok ok ok ok ok ok ok Y
11 ok ok ok ok ok ok ok Y
12 defect defect defect defect defect defect defect Y
13 ok defect defect ok ok ok ok N
14 ok ok ok ok ok ok ok Y
15 ok ok ok ok ok ok ok Y
16 ok ok ok ok ok ok ok Y
17 defect defect defect defect defect defect defect Y
18 ok ok ok ok ok ok ok Y
19 ok ok ok ok ok ok ok Y
20 ok ok ok ok ok ok ok Y
21 defect defect defect defect defect defect defect Y
22 defect defect defect defect defect defect defect Y
23 defect defect defect defect defect defect defect Y
24 ok ok ok ok ok ok ok Y
25 ok defect defect defect defect defect defect N
26 defect defect defect defect defect defect defect Y
27 ok ok ok ok ok ok ok Y
28 ok ok ok ok ok ok ok Y
29 ok defect defect ok ok ok defect N
30 ok defect ok ok ok ok ok N
Appraiser Score: 24/30= 80% 29/30= 97% 28/30= 93% 80%
Repeatability Repeatability Repeatability Reproducibility
Graphical Analysis
Various data
Distribution of the frequency of results in categories of one - Assistant/ Graph. Analysis
2. Pareto-Diagram
variable
Difference 1 nominal
- Statistics/ Quality Tools/ Pareto Chart
formats are Pie Chart
accepted
Comparison of the differences between center (median, mean) - Assistant/ Graph. Analysis
ordinal/ Interval Plot;
3. Boxplot and dispersion (Quartile) of dependent variable (Y), grouped Difference 1
cardinal
>=1 nominal - Graph/ Box Plot Grouping
Multi-Vari-Chart
by an independent variable (x) (also available in t-Tests/ ANOVA)
ordinal/
ordinal/ cardinal - Assistant/ Graph. Analysis Grouping/ Various
5. Time Series Plot Chronological representation of one or more variables Difference 1
cardinal
1
time - Graph/ Time Series Plot Time Stamps
Control Chart
stamps
Scatterplot
- Assistant/ Graph. Analysis Grouping/
7. Scatter Plot Examine the relationship between two variables (Y= f(x)) Relationship 1 cardinal 1 cardinal
- Graph/ Scatter Plot Regression
(groups)/
Marginal Plot
Examine the relationships between multiple variables (multiple Grouping/ Multiple Scatter
Matrix Plot Relationship >=2 cardinal >=2 cardinal - Graph/ Matrix Chart
Scatter Plots within one chart) (Y= f(x)) Regression Plots
1. 5. 7.
3.
= Dot Plot
2.
4. Interall-Plot: Graph/ …
I-Chart, I-MR Chart, Xbar-R Chart, P Chart and U Chart are part of Control-Charts
Dr. Reiner Hutwelker Lean and Six Sigma 23
DMAIC > Statistics >> Graphical Analysis >>> Histogram (as part of the graphical Summary) Technische Universität München
Purpose a.
Examine the center, shape, and variability.
d. Descriptive Statistics
N 200
- Frequency distribution of values for one variable; Mean 3,22
StDev 1,4976
- Comparison of the data distributions with the ND curve Minimum 1
Hypothesis 0 1 2 3 4 5 6
Y_Cookie-T
4
Example: Lead Time / Diameter/ Cookie weight/ Scale Level: ordinal, cardinal
Dr. Reiner Hutwelker Lean and Six Sigma 24
DMAIC > Statistics >> Graphical Analysis >>> Pareto-Diagram Technische Universität München
Description 40
x-axis: Categorized values of a variable, e.g. dependent variable (Y)
y-axis: Probability of a variable
d.
Y_Problem-Frequency
Hypothesis 30
b.
Example 10
3. Box-Plot: Integrated plot of the center and the dispersion of (grouped) variables
Chart Boxplot of Y_Cookie_Weight_Chocolate; Y_Cookie_Weight_Vanilla
Summary Report
Box Plot
Distribution of Data by Group
Compare the center and the variability across samples. Identify any outliers.
Purpose
12
Comparison of the differences between center (median, mean) and dispersion (Quartile) of c.
dependent variable (Y), grouped by an independent variable (x)
Description 10 h.
x-axis: nominal scaled categories (Factor levels) b.
y-axis: dependent variable ( e.g. cycle time, errors per order) g.
a. f. } = topview
Data
Hypothesis 8 on distribution
e.
b.
There are (no) differences in the values of the dependent variable (Y) between
Difference
the factor levels (x)
6 j. d.
c.
Example 4
Y_Cookie_Weight_Chocolate Y_Cookie_Weight_Vanilla
k.
Comparison of compensation based on gender and profession
Y_Cookie_Weight Y_Cookie_Weight
Statistics _Choc _Van
N 100 100
Comparison of the weight of Chocolate vs. Vanilla" cookies Mean 8,1040 9,9665
StDev 1,0307 0,29468
Minimum 5 9,3374
Y Scale Level Maximum 12 10,590
1 ordinal/ cardinal
x Scale Level a. Box
>=1 nominal b. Whisker
Mintab Menu c. Outlier
- Assistant/ Graph. Analysis d. Smallest not extreme value (Minimum if there are no outliers)
- Graph/ Box Plot
(also available in t-Tests/ ANOVA)
e. 1. Quartile (= Q1: 25% cases below)
f. 2. Quartile (= Q2: 50% of cases below and above = Median)
Options
g. 3. Quartile (= Q3: 75% cases below)
Grouping h. Biggest not extreme value(Maximum if there are no outliers)
Alternative i. Interquartil-Range (IRQ) = 50%
Interval Plot; j. Mean
Multi-Vari-Chart k. Grouping of Results (Y) by Factors (x)
Example: Cycle Time duration / Cookie Weight/ Cookie Taste / Scale Level: ordinal, cardinal
Dr. Reiner Hutwelker Lean and Six Sigma 26
DMAIC > Statistics >> Graphical Analysis >>> Box-Plot Technische Universität München
Purpose 10,0
Comparison of the difference between averages and confidence intervals of a variable (Y), grouped by an
independent variable (x)
Description 9,5
Y_Cookie_Weight
x-axis: nominal/ ordinal or cardinal scaled categories (Factor levels)
y-axis: dependent variable ( e.g. cycle time, weight)
9,0
Hypothesis
There are (no) differences in the values of the dependent variable (Y) between the factor 8,5
Difference
levels (x)
a.
Example
8,0
d. [ b.
c.
Area of normal, i.e. expected values for parameters of medicine Y_Cookie_Weight_Choc Y_Cookie_Weight_Van
x_Cookie_Type
e.
Comparison of the consistency of "dark" vs. "light" cookies Individual standard deviations are used to calculate the intervals.
Y Scale Level
1 cardinal
x Scale Level a. Upper Confidence Limit for Mean
1 nominal b. Mean
Mintab Menu
c. Lower Confidence Limit for Mean
d. Confidence Interval CI (error margin of the Mean)
- Graph/ Interval Plot
:= with a 95% Confidence the Mean of the Population is expected
Options
within this Interval ( CI= xbar +/- 1,96* s/sqrt(N) )
e. Grouping of Results (Y) by Factors (x)
Grouping
Alternative
Boxplot
Example: Cycle Time duration / Cookie Weight/ Cookie Taste / Scale Level: cardinal
Dr. Reiner Hutwelker Lean and Six Sigma 27
DMAIC > Statistics >> Graphical Analysis >>> Time Series Plot Technische Universität München
5
Description
x-axis: Time Series (e.g. Days/ Hours/ order number) in chronological order
4
y-axis: dependent variable ( e.g. cycle time, errors per order)
Data
Hypothesis 3
There are (no) differences in the values of the dependent variable (Y) 2
Difference between the time points (x)
1
1 10 20 30 40 50 60 70 80 90 100
Example Index
Example: Lead Time over time/ Cookie Taste over time / Scale Level: cardinal
Dr. Reiner Hutwelker Lean and Six Sigma 28
DMAIC > Statistics >> Graphical Analysis >>> Multi-Vari Chart Technische Universität München
Y_Cookie_Taste
4
0
e. Vanilla; Butter Vanilla; Margarine
16
12
g.
f.
a. 8
0
Spelt c. Wheat
X2_Flour c.
Panel variables: X4_Cookie-Type, X3_Fat
Example: Screening of complex influence factors (x) on the Problem (Y)/ Scale Level: cardinal
Dr. Reiner Hutwelker Lean and Six Sigma 29
DMAIC > Statistics >> Graphical Analysis >>> Scatter-Plot Technische Universität München
Scatter Plot
Purpose e.
Examine the relationship between two variables
Description
x-axis: continuous independent variable (x)
y-axis: continous dependent variable ( e.g. cycle time, error per job)
f.
Hypothesis
c.
There is (no) ("the...the...") relationships between: the independent
Relationship
variable (x) and the dependent variable (Y) d.
b.
Example
Relationship between: baking time of the dough (x) and the brightness of the cookie (Y) a.
Y Scale Level
1 cardinal
x Scale Level a. X-Variable: Length of baking time (from 1- 30 min)
1 cardinal b. Y-Variable: Brightness of Cookies (from bright=0 to dark=10)
Mintab Menu c. Data points for each measured Cookie
- Assistant/ Graph. Analysis d. (Linear) Regression line: Y= b+ ax (+error)
- Graph/ Scatter Plot e. Descriptive Statistics/ Parameter
Options f. Type of relationship (Linear, Quadratic, …)
Grouping/ Regression
The scatter plot is the graphical display for the:
Alternative
- Correlation (rxy) and the
Scatterplot (groups)/ Marginal Plot - Regression Analysis (Y= b+ ax +e)
Example: Relationship Y= f(x) between speed (x) and mileage (Y)/ Scale Level: Cardinal
Dr. Reiner Hutwelker Lean and Six Sigma 30
DMAIC > Statistics >> Process Capability and Process Control Technische Universität München
Process control refers to the historical compliance with calculated control limits
▪ Focuses on single, obvious, systematic, ▪ Considers the position (e.g. Mean) and the scatter (z.B.
Standard Deviation) of the random variation in the
observations over time
results
▪ Sets the sequence of the results in relation to the
▪ Puts position and scatter of the results in relationship to
control limits (for example: X+/- 3s)
the specification limits and/ or a target value (:= Target)
▪ The control limits are calculated using the ▪ Specification limits are based on the requirements from
variation of the results and therefore mirror past „outside“ e.g. customers
results
▪ Depending on requirements and the range the following
▪ The following becomes relevant depending on become relevant:
the extent of the variation and range of values : - Upper specification limit (USL)
- Lower specification limit (LSL)
- Upper control limit (UCL)
- Target value
- Lower control limit (LCL)
▪ The level of the process capability reveals :
▪ Through the control limits and other tests more - the size of the deviation of the results from the
systematic influences are identified: specification limits and
- Outliers - how accurate the target values meet the goal
- Patterns and
▪ Tool: Process Capability Index (cp/ cpk, dpu, dpmo,
- Trends. Sigma-Level, …)
▪ Tool : Control Chart (xbar/s, …)
Process Capability refers to the compliance with externally defined specification limits
Dr. Reiner Hutwelker Lean and Six Sigma 32
DMAIC > Statistics >> Process Ability and Process Control Technische Universität München
If you treat a systematic variation like a random variation you loose the opportunity to…
▪ to eliminate a permanent and specific negative causes
▪ to use a specific positive trigger as an information of best practice for improvement
no
no
Process Capability
Details
From every day life we already have an idea of the specification limit spectrum, for example, the standard
„quarter of an hour“ with appointments, the cut off points with grades (must be better than Grade x), the
minimum requirement for elections (more than x%), the speed limit on a street and the experation date of a
product.
The specification limit defines a value on a scale that defines acceptable and not acceptable, basically it
separates problematic and not problematic result areas. Through the comparison of the product‘s results with
its specification limits it can be deduced if the product meets expectations.
100
50 150
120
0 50 100
The process capability is deduced from the comparison of the observed values with the specified
expectations. This comparison comes from the calculation of a process capability measurement.
The specification values separate the acceptable from the unacceptable values
Dr. Reiner Hutwelker Lean and Six Sigma 37
DMAIC > Statistics >> Process Capability Technische Universität München
Scale Level of the data of the relevant Product/ Service attribute (CtQ)
… the Scale Level of the product characteristics and the distribution of the measurements
Dr. Reiner Hutwelker Lean and Six Sigma 38
DMAIC > Statistics >> Process Capability Technische Universität München
Example Defective cookies/ baking sheet Defective/ Cookie Weight of the cookies within specification limits
Routes with traffic jam/ all routes Incorrect transfers/ Call Call times within specification limits
Traffic jams/ Route Delivery accuracy within specification limits
Objective
USL Tolerated Error-% Tolerated Defects/ Unit Tolerated value of the upper limit on the scale
LSL ./. ./. Tolerated value of the lower limit on the scale
Target Values ./. ./. Target value on the scale
Calculation
Basis all products (variable or fixed) Opportunities * Units Cp/ Pp= (USL-LSL)/ 6 s
Portion of defective Products/ Services Defects/ Unit Cpk/ Ppk= min((USL-Xbar); (Xbar-LSL))/ 3 s
Reference to the Accurate Units, e.g.: Accurate Units, e.g.: Accurate Units, e.g.:
formation of - per Hour - concrete product - per Hour
Subgroups (subsamples) - per Day - per Day
- per Shift - per Shift
- per Lot - per Lot
Performance-Measure Comparison with the target Comparison with the target Comparison with the target
Sigma-Level (Z-Bench) DPU-% Cp/ Pp; Cpk/ Ppk
Error-%; Yield-% Error-%; Yield-% Sigma-level (Z-Bench)
The Process Capability can be calculated for all Products and Services
Dr. Reiner Hutwelker Lean and Six Sigma 39
DMAIC > Statistics >> Process Capability Technische Universität München
Cookie Quality Weight 18g Goal: 20g 22g > 22g Cardinal/ Ordinal normal (cp/pp/ Z) binomial (%-Z) Customer
Consistency Chewy Crunchy doughy// crumbly Ordinal normal (cp/pp/ Z) binomial (%-Z) Customer
Time 24.12./ 14:00 24.12./ 15:30 24.12./ > 15:30 Cardinal normal (cp/pp/ Z) Customer
Used Resources Price yi,- €/ Cookie yj,- €/ Cookie Cardinal normal (cp/pp/ Z) Customer
▪ The weight of the Cookie can be measured with a scale using a Cardinal Ccale. Then if the weight< 5g (ko), 5 < Weight < 20g (ok), Weight > 20g (ko).
The suitable Process Capability Analysis for normal distributed data is: (normal) cp/pp/ Z
▪ The sweetness can attributely measured on a nominal scale, with : {low, medium, high = ok vs. neutral, salty, bitter = ko}.
The suitable Process Capability Analysis is: binomial (%-Z)
▪ The sweetness can also be measured by a rating on an Ordinal Scale, then for example z.B.: 0= neutral taste (ko),
1= low, 2= medium, 3= high sweetness (ok), 4= extreme sweet (ko). Process Capability for normal distributed data is: normal (cp/pp/ Z)
▪ If 10 ingredients belong to the composition of the Cookie and each missing ingrediant is a Problem (Opportunity for Defect), then the Defects would be the
missing ingrediants per Cookie: {0 = ok vs. 1-10 = ko}. The suitable Process Capability Analysis is: Poisson (DPU)
… based on the Scale Level of the tolerated characteristics and the definition of the Problem
Dr. Reiner Hutwelker Lean and Six Sigma 40
DMAIC > Statistics >> Process Capability Technische Universität München
3. 1. 2.
1. 2. 3. 4.
Weight (cardinal) Cookie (ok vs. ko) (nominal) Different Types of Defects in a Cookie (x Defects/ Unit) (nominal)
Taste (ordinal) Taste (ok vs. ko) (nominal) Different Types of Defects in a Call (x Defects/ Unit) (nominal)
Cycle Time (cardinal) Quantity/ Tin (ok vs. ko) (nominal) Traffic jams on different Highways (x / 1 km) (nominal)
Processes are affected by many influences – the Outputs can be checked by various analyses
Dr. Reiner Hutwelker Lean and Six Sigma 41
DMAIC > Statistics >> Process Capability >>> Binomial Technische Universität München
1.
Binomial Capability
Purpose a.
Checks whether a certain Process is capable to yield Products/ Services that meet Business/ Customer
Requirements.
Focus
Chance (p) that a selected Product/ Service is defective. The data collected are the number of defective
Products in individual subgroups, which is assumed to follow a binomial distribution with parameter p.
c.
Data
- Number of defects
- in a series of constant/ variable subgroups of a sample
Specification Limits
Result Dialog
- % defective and its 95% Confidence Interval (CI)
- ppm/ dpmo
- Process Z (Sigma-Level) Y-Variable with the summarized number of defects in a subgroup
a.
Probability Mass Function (defects/ day; defects/ batch; defects/ shift; defects/ sheet)
n= Number of Products
k= Number of Defects
p= prob(Defect) 1. Constant Size for all Subgroups: e.g. always 30 Cookies/ sheet
(which is related to the number of defects)
Example b.
2. Column of Subgroup Sizes: e.g. variable number of Cookies/ day
engine starts (yes/ no)
mobile network access (yes/ no) (which is related to the number of defects)
Cookie (ok/ ko)
Y Scale Level
Upper Specification Limit (USL) in %
1 nominal c.
(Example: 3%)
Transform Scale Values
not defective = 0 vs. defective= 1
Mintab Menu
- Assistant/ Capability Analysis/ Binomial Capability
- Stat/ Quality Tools/ Capability Analysis/ Binomial
1.
a. Stability
!
Stability is an important assumption of capability analysis. To determine whether your process is stable, examine the control charts on the
Diagnostic Report. Investigate out-of-control points and eliminate any special cause variation in your process before continuing with this Process Stability
analysis. Use the best chart to confirm that the process is stable.
b. Number of
Subgroups i You have 200 subgroups. For a capability analysis, this is usually enough to capture the different sources of process variation when
collected over a long enough period of time.
P Chart
** Laney P′ Chart (correction factor = 1,660)
a. b.
Expected The variation in your data does not match the expected variation (overdispersion or underdispersion), resulting in a P chart on the 0,50
b. !
Proportion
Variation Diagnostic Report that may not signal appropriately. Consider using the Laney P′ chart instead, which corrects the control limits to account
for this condition.
0,25
d. Amount
of Data i The 95% confidence interval for the % of defective items is (4,66; 5,36). If this interval is too wide for your application, you can gather more
data to increase the precision.
0,00
Excess variation results in control limits that are too narrow The Laney P′ chart corrects the control limits to account for the
for your data, which can cause an elevated false alarm rate. excess variation. The chart should signal appropriately.
Cumulative % Defective
As the points level out, the estimate of % defective becomes more reliable.
30
c.
% Defective
20
Number of Subgroups:
b. - too few observations (subgroups) indicate short term observation
> interpret Parameters of Process Capability with reservation
Control Charts to asses the stability of the Process
p-Chart for the Process
Expected Variation: Overdispersion and Underdispersion
- Variation in Subgroup Sizes can lead to Overdispersion and false alarms a. = Time Series Plot with Upper/ Lower Control Limits (UCL/ LCL)
c. - Intercorrelation of Subgroup results can lead to Underdispersion and missed = Outliers and Patterns indicate deviation from Stability
signals
Laney p-Chart for the Process
Amount of data: b. = corrected p-Chart due to overdispersion
d. - too few data/ small sample size might prevent Significance = Outliers and Patterns indicate deviation from Stability
> Collect more data
Chart with cumulative % Defective:
c. - an asymptotic progress of the line indicates the reliability of the
estimation of the % Defective Process Capability Parameter
1.
3%
Example: e. Comments
- p= 1,000 Acceptable % defective: 3%
Process Characterization:
2.
Poisson Capability
Purpose
a.
Checks whether a certain Process is capable to yield Products/ Services that meet Business/ Customer Requirements.
Focus
Product/ Service can have multiple defects and the number of defects on each item is counted. b.
Evaluation
Number of defects per unit. The data collected are the total number of defects in k
units contained in individual subgroups, which is assumed to follow a Poisson
distribution with an unknown mean number of defects per unit (u).
c.
Data
- Number of defects in a Number of units * opportunity for defect
- in a series of constant/ variable subgroups of a sample
Specification Limits
Result
Dialog
- Defects per Unit (DPU and its 95% Confidence Interval (CI)
- Yield (Probability of producing a unit without defects)
Y-Variable with the summarized number of defects in a subgroup
Probability Mass Function a.
(Opportunities * Defects/ sheet)
lambda= average number of events per interval
e= 2.71828 (Euler's number, base of natural logarithm)
k= Number of defects
k! = factorial of k
1. Constant Size for all Subgroups: e.g. always 50 Cookies/ sheet (which is
Example related to the number of defects)
b.
aircraft (1.000.000 Opportunities for Defect) -> 1 aircraft= 1.000.000 units, e.g. compared with 2. Column of Subgroup Sizes: e.g. variable number of Cookies/ sheet (which is
Chewing Gum (5 Opportunities for Defect) -> 1 Chewing Gum= 5 units related to the number of defects)
Cookie (10 Opportunities for Defect) -> 1 Cookie= 10 units
Y Scale Level
1 nominal Upper Specification Limit (USL) in DPU
c.
Transform Scale Values (Example: 0,1)
Opportunity not defective = 0 vs. defective= 1, count number of defect Opportunities
Mintab Menu
- Assistant/ Capability Analysis/ Poisson Capability
- Stat/ Quality Tools/ Capability Analysis/ Poisson
Capability for the overall statement: Unit (Opportunity * No. of Product/ Service) is: ok vs. ko
Dr. Reiner Hutwelker Lean and Six Sigma 45
DMAIC > Statistics >> Process Capability >>> Poisson Technische Universität München
2.
a. Stability
!
Stability is an important assumption of capability analysis. To determine whether your process is stable, examine the control charts on the
Diagnostic Report. Investigate out-of-control points and eliminate any special cause variation in your process before continuing with this Process Stability
analysis. Use the best chart to confirm that the process is stable.
b. Number of
Subgroups i You have 100 subgroups. For a capability analysis, this is usually enough to capture the different sources of process variation when
collected over a long enough period of time.
2
U Chart
** Laney U′ Chart (correction factor = 0,179)
a. b.
Amount The 95% confidence interval for the number of defects per unit is (0,48; 0,52). If this interval is too wide for your application, you can gather
d. of Data i more data to increase the precision.
0
Excess variation results in control limits that are too narrow The Laney U′ chart corrects the control limits to account for the
for your data, which can cause an elevated false alarm rate. excess variation. The chart should signal appropriately.
Cumulative DPU
As the points level out, the estimate of DPU becomes more reliable.
2,0
c.
Number of Subgroups:
b. - too few observations (subgroups) indicate short term observation
> interpret Parameters of Process Capability with reservation
Control Charts to asses the stability of the Process
u-Chart for the Process
Expected Variation: Overdispersion and Underdispersion
- Variation in Subgroup Sizes can lead to Overdispersion and false alarms a. = Time Series Plot with Upper/ Lower Control Limits (UCL/ LCL)
c. - Intercorrelation of Subgroup results can lead to Underdispersion and missed = Outliers and Patterns indicate deviation from Stability
signals
Laney u-Chart for the Process
Amount of data: b. = corrected u-Chart due to overdispersion
d. - too few data/ small sample size might prevent Significance = Outliers and Patterns indicate deviation from Stability
> Collect more data
Chart with cumulative Defects per Unit (DPU):
c. - an asymptotic progress of the line indicates the reliability of the
estimation of the DPU Process Capability Parameter
2.
Example: e. Comments
- Confirmation of H0 • The process DPU is not significantly less than the maximum
acceptable level (p > 0,05).
- DPU is not <= 3% (USL) • The chance of producing a unit with no defects is 60,7%.
Process Characterization:
Comment:
e.
- Summary of results and additional hints
3.
LSL USL
Centering-Index Cpk relates the distance of the
center (mean) to the nearest Specification Limit to
the dispersion of the results -5 s -3 s -1s 1s 3s 5s
x
… Cpk/ Ppk relates the Center of the Results to the nearest Specification Limit
Dr. Reiner Hutwelker Lean and Six Sigma 48
DMAIC > Statistics >> Process Capability >>> Cpk/ Ppk/ Z Technische Universität München
3.
Purpose a.
Checks whether a certain Process is capable to yield Products/ Services that meet Business/ Customer
Requirements.
Description
Product/ Service has a single, cardinal scaled attribute, with an Upper and/ or Lower Specification Limit (as b.
well as a target vale on the scale), which should be met.
Evaluation c.
Mean and Standard Deviation of the a) Subgroups of the sample (Within) and b) the whole sample (Overall)
is related to the Specification Limits/ Target to calculate a) the actual (Cp/ Cpk) and b) the potential (Pp/ d.
Ppk) Capability.
Data
- Values of the attribute of the Product/ Service
- in a series of constant/ variable subgroups of a sample e.
Specification Limits
- Upper Specification Limit (USL)
- Lower Specification Limit (LSL)
- Target
Result
- Pp/ Ppk (actual Performance)
- Cp/ Cpk (potential Performance)
- Z-Bench/ % out of Spec./ PPM (DPMO)
f.
Probability Mass Function
m= Mean of the Population Dialog
s= Standard Deviation of the Population
- Complete: if data in sample are collected from a continuous time series
s2= Variance of the Population a. - Snapshot: data are arbitrary collected and arranged (data are seldom ND!)
Example
- Data in one column, Subgroups in another column with Subgroup Id´s
Cycle Time b. - Data in multiple columns, with one column for each Subgroup
Viscosity
Weight of Cookies within Specification Limits (9g < Cookie Weight < 11g) c. Y-Variable with single data for each Product/ Service
Y Scale Level
1 cardinal - Constant Size for all Subgroups: e.g. always 30 Cookies/ sheet
d. - Column with Subgroup ID´s
Transform Scale Values
Data of attribute must be normal distributed; if not: a) transform data or b) use alternative Capability Index - Lower Specification Limit (LSL
e. - Upper Specification Limit (USL) in DPU
Mintab Menu
- Assistant/ Capability Analysis/ Capability Analysis Cpk, Ppk, Z
- Stat/ Quality Tools/ Capability Analysis/ Normal f. Target Value (optional; typically between LSL and USL)
Capability for the overall statement: Unit (Opportunity * No. of Product/ Service) is: ok vs. ko
Dr. Reiner Hutwelker Lean and Six Sigma 49
DMAIC > Statistics >> Process Capability >>> Cpk/ Ppk/ Z Technische Universität München
3.
a. Stability
!
Stability is an important assumption of capability analysis. To determine whether your process is stable, examine the control charts on the
Diagnostic Report. Investigate out-of-control points and eliminate any special cause variation in your process before continuing with this
a.
11
Xbar-S Chart
**
Confirm that the process is stable.
analysis.
b. Number of
i You only have 10 subgroups. For a capability analysis, it is generally recommended that you collect at least 25 subgroups over a long 10
Mean
Subgroups enough period of time to capture the different sources of process variation.
Normality Your data failed the normality test. A Box-Cox transformation will not correct the problem. Get help to determine next steps because the 9
b. ! capability estimates may be inaccurate.
Amount The total number of observations is 100 or more. The capability estimates should be reasonably precise. b.
of Data
d. 1,6
StDev
0,8
0,0
1 2 3 4 5 6 7 8 9 10
Normality Plot
The points should be close to the line.
c.
Normality Test
(Anderson-Darling)
Results Fail
P-value < 0,005
Report Chart with statements about:
Stability of the Process:
a. - Outliers and Patterns in the data indicate low Stability of the Process
> interpret Parameters of Process Capability with reservation
Number of Subgroups:
b. - too few observations (subgroups) indicate short term observation
> interpret Parameters of Process Capability with reservation Control Charts to asses the stability of the Process
Expected Variation: Overdispersion and Underdispersion Xbar-Chart for the Process
- Variation in Subgroup Sizes can lead to Overdispersion and false alarms a. = Time Series Plot with Mean of Subgroups and Upper/ Lower Control Limits (UCL/ LCL)
c. = Outliers and Patterns indicate deviation from Stability
- Intercorrelation of Subgroup results can lead to Underdispersion and missed
signals
S-Chart for the Process
Amount of data: b. Time Series Plot with Standard Dedivation (SD) of Subgroups and UCL LCL for SD
d. - too few data/ small sample size might prevent Significance = Outliers and Patterns indicate deviation from Stability
> Collect more data
3.
Z potential = Z-Bench potential = Sigma-Level for the within Capability based on Cpk, i.e. c.
the within standard deviation= 0,65 (see next slide) d. Process Characterization
Does the process mean differ from 10?
Mean 9,8333
0 0,05 0,1 > 0,5 Standard deviation (overall) 1,1987
3.
c.
Capability Statistics
Process Characterization
Actual (overall)
Pp 0,28
Ppk 0,24
Description of the Sample Z.Bench 0,24
- Total N= 300 % Out of spec (observed) 32,69
% Out of spec (expected) 40,40
- Subgroup Size= 30 PPM (DPMO) (observed) 326923
b. (- Number of Subgroups= 10) PPM (DPMO) (expected) 404016
Potential (within)
Cp 0,37
Parameter of the Sample: Cpk 0,32
Z.Bench 0,60
- Mean % Out of spec (expected) 27,30
PPM (DPMO) (expected) 273007
- Standard Deviation (overall)
4,5 6,0 7,5 9,0 10,5 12,0
- Standard Deviation (within)
μ
USL μ
μ
μ
LSL
Cp = 1,33 Cp = 1,33 Cp = 1,33 Cp = 1,33
Cpk = 1,33 Cpk = 1,0 Cpk = 0,0 Cpk = -0,33
If the Process is centered, then Cpk = Cp, in other cases Cpk < Cp
Dr. Reiner Hutwelker Source: SiemensLean
AG, Kennzahlen zur Prozessqualität
and Six Sigma 53
DMAIC > Statistics >> Process Capability >>> Cpk/ Ppk/ Z Technische Universität München
si st
sj sj
sj sj
USL USL
Cp good (2) Cp insufficient (6)
Cpk good (2) Cpk insufficient (6)
Pp good (2) Pp insufficient (6)
Ppk good (2) Ppk insufficient (6)
LSL LSL
USL USL
USL USL
Cp very good (1) Cp satisfactory (3)
Cpk good (2) Cpk sufficient (4)
Pp satisfactory (3) Pp deficient (5)
Ppk satisfactory (3) Ppk deficient 5)
LSL LSL
USL USL
Cp good (2) Cp good (2)
Cpk deficient (5) Cpk satisfactory (3)
Pp insufficient (6) Pp deficient (5)
Ppk insufficient (6) Ppk deficient (6)
LSL LSL
Process Control
Details
K out of K+1 > 1 standard deviations** from the center line (same side)
6 early notification if something is amiss
5. (:= warning limit; UWL= upper warning limit, LWL= lower warning limit)
K points in a row within 1 standard deviation** of the center line (either side)
6. 7 early notification of a series of similar results on the same level of deviation
Cause: consistant negative influence (?)
7. 8
K points in a row > 1 standard deviation** from center line (either side)
early notification of a series of similar results on the same level of deviation
Cause: consistant negative influence (?)
8. ** applies to the different dispersion parameter of the choosen Chart, like MR, R, …
2 2 2 4
2
2
2
2 _
4
3
3
X=3,92
TEST 2. 7 points in a row on same side of center line.
3
0
3
3
3 4 Test Failed at points: 8; 9; 10; 11; 12; 13; 14
LCL=-2,43
1
1 4 7 10 13 16 19 22 25 28 31 TEST 3. 5 points in a row all increasing or all decreasing.
Observation Test Failed at points: 15; 16; 17; 18; 19; 20
1
6
2 2 2 2 2 2
4
__
MR=2,39
2 Test Results for MR Chart of Y_Choc-Cookie-Taste first Week
2 2 2 2 2 2 2 2 2 2
0 2 2
LCL=0
1 4 7 10 13 16 19 22 25 28 31 TEST 1. One point more than 3,00 standard deviations from center line.
Observation
Test Failed at points: 2
Signals in the Ouptuts indicate, that the Process is not
TEST 2. 7 points in a row on same side of center line.
under control. Before the Control Limits can reliably be
Test Failed at points: 9; 10; 11; 12; 13; 14; 15; 16; 17; 18; 19; 20; 27; 28;
interpreted, the Signals have to be eliminated. 29; 30; 31; 32
To do so, every single Signal needs to be inspected by
analysing the corresponding Product/ Service for the Root * WARNING * If graph is updated with new data, the results above may
Cause of the Signal. no longer be correct.
The Root Cause needs to be eliminated because a
controlled Process shows only random fluctuation and is
therefore normal distributed.
Then apply the Control Chart again and check it for Signals.
Purpose: Scale
# Control Chart Variables and Grouping Y Mintab Menu Option
Identify Signals in a sequence of chronological collected ... Level
... Single values in relation to other single values.
... Amount of errors in one subgroup compared to the Amount of errors in other subgroups. each value of the variable represents the number of
defective Units in a subgroup, with a dot for each
Assistant/ Control
4 p-Chart subgroup nominal
Charts/ P Chart
pbar-Chart: deviation of the percentage of defective Units of every subgroup (p_Subgroup) from - Number of values per subgroup: > 5;
the average deviation of defective Units of all subgroups (pbar) - Size of the subgroup: constant or variable
... amount of Defects per Unit in relation to the average Defects per Unit. each value of the variable represents the number of
Defects per Unit, with a dot for each Opprtunities
Assistant/ Control
5 u-Chart ubar-Chart: deviation of the Defects per Unit (DPU) from the average amount of Defects per Unit subgroup nominal
Charts/ U Chart
(ubar, specifically: DPUbar) - Number of Opprtunities: > 5
- Number of Opportunities can be: constant or variable
** for the type of identified Signals (:= systematic results) refer to slide:
Examples for identified Signals in the I-MR Chart
…
Dr. Reiner Hutwelker Lean and Six Sigma 61
DMAIC > Statistics >> Process Capability >>> Control Charts Technische Universität München
1. 2. 3. 4. 5.
…
Dr. Reiner Hutwelker Lean and Six Sigma 62
DMAIC > Statistics >> Process Capability >>> Control Charts >>>> Options Technische Universität München
Two options for the Calculation of the Control Limits and the Center Line in the Control Charts
a. Estimate Control Limits and Center Line or b. use known, i.e. historical data
Dr. Reiner Hutwelker Lean and Six Sigma 63
DMAIC > Statistics >> Process Capability >>> Control Charts >>>> Reports Technische Universität München
The Report Card and the Stability Report are very similar for all Control Charts
I-MR Chart of Y_Choc-Cookie-Taste first Week I-MR Chart of Y_Choc-Cookie-Taste first Week
Report Card Stability Report
Check Status Description a.1 Look for these patterns: c.
a. Stability
!
The process mean and variation may not be stable. 5 (15,6%) points are out of control on the I chart. 1 (3,2%) point is out of control on the
MR chart, which may affect the validity of the control limits on the I chart. You may see 0,7% out-of-control points on the I chart and 0,9%
Global Trend Cyclical
Individual Value
out-of-control points on the MR chart by chance, even when the process is stable. You should investigate out-of-control points and omit
those with special causes from the calculations.
b.
Normality If the data are nonnormal, you may see an increased number of false alarms. Because fewer than 2 points are outside the control limits on
the I chart, the normality test is not needed.
b. Amount You may not have enough data to estimate precise control limits. At least 100 data points should be included in the calculations.
Shifts Drifts
of Data !
a.2
d. Correlated
Data
If the data are correlated, you may see an increased number of false alarms. Because fewer than 2 data points are outside the control limits
on the I chart, the correlation test is not needed.
Moving Range
Alternative This chart is intended to monitor process control. If your primary objective is to explore your data or compare your process before and
e. Charts i after a change, use the Graphical Analysis Control Charts or the Before/After Control Charts.
Oscillation Mixture
Assess the stability of the mean and variation of your process and look for patterns that
can help you distinguish between common and special causes. Typically, a process that
exhibits only common causes has a constant mean and constant variability. However, global
trends or cyclical patterns may also be common causes. Other patterns, such as shifts and
drifts, may be special causes.
Excessive Out
of Control
Normality:
- If the data are not normal distributed, then a
b. > Box-Cox Transformation might help to align the data, especially for the Xbar-S Stability Report
Chart. Unfortunately, also after this transformation the data are still not normally The Charts show the:
distributed. a.1: Deviation of the data from the Center (Line of the overall Mean)
Amount of Data: a.2: Variability between adjacent values (I/MR) or Dispersion within Subgroups
- too few observations might lead to too narrow corridor between the Control Limits a. (Xbar-R, Xbar-S)
c. and a wrong position of the Center Line
> a Sample Size of at least 100 values is recommended Signal are indicated by:
Correlated Data:
- correlated values within a time series might lead, especially for extreme values, to Explanation of indicated Signals by the underlying test
d. false alarms, i.e. a sequence of Outliers which depend on the cause of the correlation b. (see also slide: Control Charts > Tests)
> identify the cause of the correlation, i.e. the dependency between values
Alternative Charts: Further patterns (Signals), beside the tested and identified Signals
e. c. > Identify Root Causes of this Patterns, if they occur
- hints about alternatives to the selected Control Chart
1.
I/ MR
a.
Purpose:
Identify Signals in a sequence of chronological collected ...
... Single values in relation to other single values.
b.
Data Typecardinal
Dialog
each value of the variable is represented as one number and one dot in the
Relation Data Column:
chart
a. Y- Variable with single values, observed in a consecutive time series
Recommended for usage if N< 50, otherwise charts are blacked with dots
Example
Monitoring the accessibility by phone/ day
Monitoring fuel consumption/ 100km
Determination of Control Limits and Center Line:
b. see slide: Two options for the Calculation of the Control Limits and the Center Line
Monitoring the amount of Cookies/ baking sheet (backing time/ backing sheet)
1.
Yes No ∑ 𝑥
15,6% Xbar 𝑥̅ =
𝑛
Individual and Moving Range Charts
f. 𝑈𝐶𝐿 = 𝑥̅ + 2,66 ∗ 𝑀𝑅 ∑
a.
Investigate any out-of-control points.
𝑥 −𝑥
10 UCL=10,04
UCL/ LCL 𝑀𝑅 =
e. 𝐿𝐶𝐿 = 𝑥̅ − 2,66 ∗ 𝑀𝑅 𝑛−1
Individual Value
5 _
X=4,08
0
b. MR-
Parameter for the Dispersion of the individual values and the corresponding Control Limits
LCL=-1,88 Chart
b. 10
UCL=7,32 ∑ 𝑥 −𝑥
Moving Range
5
e.
MR 𝑀𝑅 =
__
𝑛−1
MR=2,24
0 LCL=0
1 4 7 10 13 16 19 22 25 28 31
f. 𝑈𝐶𝐿 = 3,267 ∗ 𝑀𝑅
N: 32 Mean: 4,0833 StDev(within): 1,9870 StDev(overall): 2,1659
UCL/ LCL
f. Control limits are estimated using the StDev(within).
𝐿𝐶𝐿 = 0
Result: Comments (2/2) Result: Comments (1/2)
Dots in Chart
Process stability statement:
e. Each dot in both Charts represents a single value of the data. If the Charts are blacked out c. - displays the percentage of values out of control
due to a too large number of dots, then choose the Xbar-R (Grouping with <= 8 values) or
- the coloured bar indicates, whether the Process is under (yes) or out of control (no)
the Xbar-S Chart (Grouping with > 8 values)
note:
Comments:
f. Difference between Procedures in the Assistant and the Stats menu: d.
Summary and comments about results
- the Minitab Assistant uses the Standard Deviation (within) to determine the Control Limits
2.
Xbar-R Chart: monitor cardinal scaled and grouped values (n_Subgroup <=8)
Control Chart
Xbar/ R a.
Purpose:
Identify Signals in a sequence of chronological collected ...
... Subgroups in relation to other subgroups (N_Subgroup <=8).
Xbar Chart: deviation of the Mean of the subgroup (Xbar_Subgroup) from the overall Mean
(Xbar_Sample)
R-Card: Deviation of the Range per subgroup (R) from overalll average Range (Rbar)
b.
Variables and Grouping
Yvalues summarized in subgroups (Size <=8)
Data Typecardinal
each value of the variable is pooled with adjacent values to a subgroup, with a
dot for each subgroup
Relation
- Number of values per subgroup: <= 8;
- Size of the subgroup: constant or variable
Dialog
Data Column:
Example - Y-Variable with single values, observed in a consecutive time series
Monitoring the accessibility by phone/ day (Subgroup: week) - Data will be pooled in Subgroups, size for Subgroups can be <= 8, e.g. one value for each
Monitoring the jogging time for a specific route (Subgroup: week) a. day of the week
Definition of Subgroup:
Monitoring the weight of the Cookies (Subgroup: tin with 8 Cookies) see slide: Grouped data can be arranged in two alternative ways
Assistant/ Control Charts/ Xbar-R Chart Determination of Control Limits and Center Line:
b. see slide: Two options for the Calculation of the Control Limits and the Center Line
2.
Xbar-R Chart: monitor cardinal scaled and grouped values (n_Subgroup <=8)
Yes No ∑ 𝑥̅ ∑ 𝑥
70,0% Xbar 𝑥̿ = 𝑥̅ =
𝑚 𝑛
Xbar and R Charts
Investigate any out-of-control subgroups.
f. 𝑈𝐶𝐿 = 𝑥̅ + 𝐴 ∗ 𝑅 𝑅=𝑥 −𝑥
a. 10 UCL/ LCL ∑ 𝑅 −𝑅
e. 𝐿𝐶𝐿 = 𝑥̅ − 𝐴 ∗ 𝑅 𝑅=
UCL=7,22
__ 𝑛
Mean
X=6,1
(A2 is more stable and is based on the normal distribution )
5 LCL=4,98
b. R-Chart Parameter for the Dispersion of the grouped values and the corresponding Control Limits
0
b.
e.
10 ∑ 𝑅 −𝑅
Rbar 𝑅=
Range
5 UCL=5,59 𝑛
_
R=3
0 LCL=0,41
1 2 3 4 5 6 7 8 9 10
f. 𝑈𝐶𝐿 = 𝐷 ∗ 𝑅
Subgroups: 10 Mean: 6,1 StDev(within): 1,0537 StDev(overall): 4,0365
UCL/ LCL
f. Control limits are estimated using the StDev(within). 𝐿𝐶𝐿 = 𝐷 ∗ 𝑅
(D3/ D4 are two variables, based on the Normal Distribution to correct the Control Limits)
Result: Comments (2/2) Result: Comments (1/2)
Process stability statement:
The dots in both Charts represent the subgroups with values of the data. If the Charts are c.
e. blacked out with a too large number of dots, then the Xbar-S Chart (Grouping with > 8 dots) - displays the percentage of values out of control
- the coloured bar indicates, whether the Process is under (yes) or out of control (no)
note: Comments:
f. d.
Difference between Procedures in the Assistant and the Stats menu: Summary and comments about results
- the Minitab Assistant uses the Standard Deviation (within) to determine the Control Limits
3.
Xbar-S Chart: monitor cardinal scaled and grouped values (n_Subgroup >8)
Control Chart
Xbar/ S a.
Purpose:
Identify Signals in a sequence of chronological collected ...
... Subgroups in relation to other subgroups (N_Subgroup >8).
Xbar Chart: deviation of the Mean of the subgroup (Xbar_Subgroup) from the overall Mean
(Xbar_Sample)
R-Chart: Deviation of the Standard Deviations per subgroup (S) from the overall Standard
Deviation (Squer)
b.
Variables and Grouping
Yvalues summarized in subgroups (Size >8)
Data Typecardinal
each value of the variable is pooled with adjacent values to a subgroup, with a
dot for each subgroup
Relation
- Number of values per subgroup: > 8;
- Size of the subgroup: constant or variable
Dialog
Example Data Column:
- Y-Variable with single values, observed in a consecutive time series
Monitoring inbound calls (Subgroup: 30 min.) - Data will be pooled in Subgroups, size for Subgroups can be > 8, e.g. all Cookies of a day
Monitoring of daily expenses (Subgroup: Month) a. pooled to a subgroup each
Monitoring the weight of the Cookies of Cookies per Sheet (Subgroup: Sheet with about 30 Definition of Subgroup:
Cookies) see slide: Grouped data can be arranged in two alternative ways
Assistant/ Control Charts/ Xbar-S Chart Determination of Control Limits and Center Line:
b. see slide: Two options for the Calculation of the Control Limits and the Center Line
3.
Xbar-S Chart: monitor cardinal scaled and grouped values (n_Subgroup >8)
Yes No
Xbar
∑ 𝑥̅ ∑ 𝑥
𝑥̿ = 𝑥̅ =
55,6%
𝑚 𝑛
Xbar and S Charts
Investigate any out-of-control subgroups.
f. 𝑈𝐶𝐿 = 𝑥̅ + 3 ∗ s
a. 14
UCL/ LCL ∑ 𝑥 − 𝑥̅
𝑠=
e. UCL=12,407 𝐿𝐶𝐿 = 𝑥̅ − 3 ∗ s 𝑛−1
12
Mean
__
X=11,328
10
LCL=10,250 b. S-Chart Parameter for the Dispersion of the grouped values and the corresponding Control Limits
b.
3
e.
UCL=2,971
∑ 𝑠 ∑ 𝑥 − 𝑥̅
_ S-bar 𝑠̅ = 𝑠=
StDev
S=2,201
2
LCL=1,431
𝑚 𝑛−1
1
1 2 3 4 5 6 7 8 9
f. UCL= 𝐵 ∗ 𝑆̅
Subgroups: 9 Mean: 11,328 StDev(within): 2,2158 StDev(overall): 2,2016
UCL/ LCL
f. Control limits are estimated using the StDev(within).
𝐿𝐶𝐿 = 𝐵 ∗ 𝑆̅
(B3/ B4 are two variables, based on the Normal Distribution to correct the Control Limits)
Result: Comments (2/2) Result: Comments (1/2)
Process stability statement:
The dots in both Charts represent the subgroups with values of the data. If the Charts are
e. blacked out with a too large number of dots, then combine more values to a group or c. - displays the percentage of values out of control
narrow the time interval
- the coloured bar indicates, whether the Process is under (yes) or out of control (no)
note:
Comments:
f. Both Procedures in the Assistant and the Stats menu use the Standard Deviation (within) d.
Summary and comments about results
to determine the Control Limits
4.
P Chart: monitor nominal scaled and grouped defective Units (n_Subgroup >5)
Control Chart
p-Chart
a.
Purpose:
Identify Signals in a sequence of chronological collected ...
... Amount of errors in one subgroup compared to the Amount of errors in other subgroups.
Data Typenominal
4.
P Chart: monitor nominal scaled and grouped defective Units (n_Subgroup >5)
P Chart
Investigate any out-of-control subgroups.
𝑝̅ ∗ 1 − 𝑝̅
0,16
f.
UCL/ LCL 𝑈𝐶𝐿 = 𝑝̅ + 𝐿𝐶𝐿 = 0
a. 𝑛
0,12 UCL=0,1185
b. Variation
not defined
Proportion
0,08 chart
_
0,04
e. P=0,0385
0,00 LCL=0
note: Comments:
f. The variable Upper Control Limit results from the variable Subgroup size, because the d.
Control Limits depend on a) the % Defective and b) the size of each Subgroup Summary and comments about results
5.
u-Chart
a.
Purpose:
Identify Signals in a sequence of chronological collected ...
... amount of Defects per Unit in relation to the average Defects per Unit.
ubar-Chart: deviation of the Defects per Unit (DPU) from the average amount of Defects per
Unit (ubar, specifically: DPUbar)
b.
Data Typenominal
each value of the variable represents the number of Defects per Unit, with a dot
Dialog
for each Opprtunities subgroup
Relation Data Column:
- Number of Opprtunities: > 5
- Y-Variable with the number of Defects/ Unit, observed in a consecutive time series, e.g.
- Number of Opportunities can be: constant or variable
the number of defects per Cookie
a.
Example Definition of Subgroup: Opportunites for defect, which can be
- constant: e.g. always 10 Opportunites for Defect in the same Product/ Service
Monitoring of permission changes per call - variable: e.g. different "complex" Cookies measured, with a variable number of Opps.
Monitoring of traffic jams per highway section
Example: monitor the Defects per Unit for Cookies with 10 Opportunities for Defect per Cookie
Dr. Reiner Hutwelker Lean and Six Sigma 73
DMAIC > Statistics >> Process Capability >>> Control Charts >>>> U Chart Technische Universität München
5.
0% > 5%
out-of-control subgroups by chance, even when the process is
stable. 𝐷𝑒𝑓𝑒𝑐𝑡𝑠
∑
u-bar 𝑈𝑛𝑖𝑡
Yes No
𝑢=
1,0%
𝑛
U Chart
Investigate any out-of-control subgroups. 𝑢
USL/ LSL 𝑈𝐶𝐿 = 𝑢 + 𝐿𝐶𝐿 = 0
𝑛
0,4
a. f. UCL=0,3585
0,3
e.
Defects per Unit
0,1 _
U=0,0838
0,0 LCL=0
1 11 21 31 41 51 61 71 81 91
Subgroup
Number of subgroups: 100 Total units: 990 Defects per Unit (DPU): 0,08
Subgroup size: 10 Total defects: 83 PPM (DPMO): 83838
note: Comments:
f. The stable Upper Control Limit results from the constant number of Opportunities; with a d.
variable number of Opportunities the Upper Control Limit becomes variable respective Summary and comments about results
Example: monitor the Defects per Unit for Cookies with 10 Opportunities for Defect per Cookie
Dr. Reiner Hutwelker Lean and Six Sigma 74
DMAIC > Statistics >> Process Capability >>> Control Charts >>>> Implementation Technische Universität München
Select Process
no yes
Data
normal distributed ?
no yes
Transform data (Box-Cox) or Signals identified ?
Select other Control Chart
no yes
Freeze Control Chart, i.e. save: Root-Cause known ?
Center Line and Control Limits
Statistical Tests
Introduction
H0 HA
modeled observation
There is no relationship between x and Y according to: There is a relationship between x and Y according to:
- If (x), then (Y). - If (x), then (Y).
- The (x), the (Y). - The (x), the (Y).
Relationship
Example
There is a/no relationship between x and Y according to:
- If the temperature of oven is (too) high (x), then the Cookie is burnt (Y).
- The higher the temperature of the oven (x), the darker the Cookie (Y).
statistical formulation
rxY = 0 rxY ≠ 0
modeling of the
observations modeled observation
Example
There is a/ no Difference
- in: the weight of Cookies (Y)
- between: Types of Cookies (x) (e.g. Vanilla vs. Chocolate vs. ...)
statistical formulation
Yi = Yj Yi ≠ Yj
… and formally split into the Hypothesis H0 vs. HA for their statistical examination
Dr. Reiner Hutwelker Lean and Six Sigma 77
DMAIC > Statistics >> Hypothesis >>> Analysis Strategy Technische Universität München
Y
Data in 2 Levels Data in > 2 Levels Data Rank Ordered Data discrete or continuous
(Nominal-Scale) (Nominal-Scale) (Ordinal-Scale) (Cardinal-Scale)
Relationship Hypothesis Relationship Hypothesis Difference Hypothesis Difference Hypothesis
Data in 2 Levels
(Nominal-Scale) Chi-Square-Test Chi-Square-Test Wilcoxon-Mann-Whitney-Test t-Test
β-error α-error
gust cat jumps oncar ball shot person jostled Earth quake
Every decisions has at least two outcomes and every outcome can be right and wrong
Dr. Reiner Hutwelker Lean and Six Sigma 79
DMAIC > Statistics >> Hypothesis >>> Analysis Strategy Technische Universität München
H0 HA
H0 20% 5%
- 1σ μ0 + 1σ attribute
- 1σ μA + 1σ
… the variability of attributes, effect size, alpha-error, beta-error and sample size
Dr. Reiner Hutwelker Lean and Six Sigma 81
DMAIC > Statistics >> Hypothesis >>> Analysis Strategy Technische Universität München
The purpose of each Statistical Test is to support the decision between H0 and HA, i.e. to support the
decision whether a Difference or a Relationship in the Population is given or not.
These factors mutually influence each other and we have to specify the optimal tradeoff for our decision. If
the “rule of the thumb” values for the different purposes are accepted for the
- alpha-error (1%, 5%, 10%) and
- beta-error (20%) and its inverse value (Power= 1-beta= 80%)
then the tradeoff remains between Sample Size and the degree of Difference/ Relationship that has a
practical relevance/ value. Thus in practice, the Sample Size is determined by the tradeoff between:
Chi-Square Tests evaluate the Observed Frequencies of at least two categories in relation to Expected Frequecies in these at
least two categories. The Expected Frequecies assume a H0 with uniform distributions in the categories.
(O − )
2
χ = i Ei
2 k
Chi-Square Test Statistic:
E
i =1 i
If the Difference between O_bserved and E_xpected frequencies is small, than (O-E)2 is small, Chi-Square is small, meaning:
- H0 confirmed
If the Difference between O_bserved and E_xpected frequencies is high, than (O-E)2 is high, Chi-Square is high, meaning:
- H0 to be rejected
Preferences for O_bserved E_xpected (O-E)2/ E
(O-E)2
Cookie-Types frequencies Frequencies (Contributions to Chi2)
Vanilla 30 25 25 1
The larger the Difference between:
O_bserved and E_xpected, Chocolate 25 25 0 0
i.e. the larger the Chi-Square Value,
Cocos 40 25 225 9
the lower the probability, that the O_bserved
frequencies are collected from a Population,
Muffins 5 25 400 16
where H0 is true.
Summe 100 100 650 Chi-Square= 26
(1. Develop a Test Statistic) 2. Calculate the value of the Test Statistic and …
Dr. Reiner Hutwelker Lean and Six Sigma 83
DMAIC > Statistics >> Hypothesis >>> Analysis Strategy Technische Universität München
Density
probability p ≤ alpha 0,10
0,05
0,05
0,00
0 7,815
Example for Cookies: X
Chi-Square= 26,
df= 3 Distribution Plot
Chi-Square; df=3
p= 0,00000954. 0,25
Density
HA: There is a Difference in: the Preference (Y)
between: Cookie-Types. 0,10
0,05
0,00000954
0,00
0 26
X
3. Calculate the probability of the value of the Test Statistic and 4. compare it with alpha
Dr. Reiner Hutwelker Lean and Six Sigma 84
DMAIC > Statistics >> Hypothesis >>> Analysis Strategy Technische Universität München
The balanced optimum between Sample Size and detectable Difference/ Relationship …
If you want to evaluate, whether the Quality of the Cookies meets the target of: < 3% (ca. 1 defect/ Sheet)
and if you could easily collect data about the status of 30 produced Cookies of one sheet, then you could
calculate the detectable % of Defects, which are necessary to indicate that the target has been exceeded:
need 0,8
Power= 1-beta Assumptions
= to be calculated α 0,05
specified 0,6
= Target (3%)
Power
0,4
13,35%
0,2
0,0
0,05 0,10 0,15 0,20 0,25
Comparison p
If the Difference/ Relationship, that can at least be detected is too large for your purposes …
If you promised your Customer to refund the money, if there are: >= 10% Defects, then you could recalculate,
how big the sample needs to be, to detect at least 10% of Defects, if they are given in the Population:
specified 0,6
= Target (3%)
Power
0,4
10%
0,2
0,0
0,050 0,075 0,100 0,125 0,150 0,175
Comparison p
… then start a 2nd trial with the Difference/ Relationship which at least must be detected
Dr. Reiner Hutwelker Lean and Six Sigma 86
DMAIC > Statistics Technische Universität München
Statistical Tests
Details
nominal
1-Sample % Compare the percentual amount of a variable (Y_Sample) with a (counted - Input: single %-value (not a data
a.1 Difference 1 ./. ./. Stat/ Basic Statistics/ 1 Proportion
Defective target value (Y_Target) > discrete column)
cardinal)
nominal
Chi-Square Compare the relative frequencies of the categories of a variable (Y), (counted - Input: enter values in Table or get Data Stat > Tables > Chi-Square Goodness-
a.2 Difference 1 ./. ./.
Goodness-of-Fit in relation to a) their expected values or b) specific target values (YT) > discrete from Worksheet of-Fit Test (One Variable)
cardinal)
Compare the Means of two independent variables with each other (Y1 Y-Variable: normal distributed; Sample
b.4 2-Sample t-Test Difference 1 cardinal 1 nominal Stat/ Basic Statistics/ 2-Sample t
vs Y2) Size N: > 20
Compare the Means of two dependent/ matched variables with each Y-Variable: normal distributed; Sample
b.5 Paired t-Test Difference 1 cardinal 1 nominal Stat/ Basic Statistics/ Paired t
other (Ya vs. Ya´ ) Size N: > 20
2-Sample
Compare the Standard Deviations of two independent variables (Y1 Y-Variable: normal distributed; Sample
b.6 Standard Difference 1 cardinal 1 nominal Stat/ Basic Statistics/ 2 Variances
vs. Y2) Size N: > 20
Deviation
nominal
Chi-Square Test Compare the percentual amounts of Y in respect to the factorial levels (counted > Number of Factor Levels for Xi and Xj Stat/ Tables/ Chi-Square-Test for
c.9 Difference 1 2 nominal
for Association of 2 categorial Variables (Xi, Xj) discrete can vary from: 3 - 6. Association
cardinal)
a. b. c.
3. 1. 2. 4. 5. 6. 7. 8. 9.
3. 1. 2.
4. 5. 6.
7. 8. 9.
1.
1-Sample % Defective
1.
Purpose
2.
Compare the percentual amount of a variable (Y_Sample) with a target value (Y_Target) 3.
Hypothesis
Example
The votes for a specific political party compared with the electoral threshold (e.g. 5%)
6.
The portion of defect Cookies in one package
Y Scale Level
1 nominal (counted -> discrete cardinal) 7.
x Scale Level
Dialog
./. ./.
1. Name Enter the name of your variable Y
Alternative in Minitab Stat Menu
2. No. of Items tested Data input: all units of the sample (integer)
Stat/ Basic Statistics/ 1 Proportion
3. Defective Units Data input: defective Units of the sample (integer)
note …
4. Target Maximum-/ Target value or proportion of errors that is acceptable (integer as %)
Input: single %-value (not a data column)
5. Direction of Test % defective > target value (one-sided significance test)
Example: compare the % defective Cookies baked per day with a target
Dr. Reiner Hutwelker Lean and Six Sigma 93
DMAIC > Statistics >> Hypothesis >>> Tests >>>> 1-Sample % Defective Technische Universität München
1.
4.
Comments Details for the significance test: Interval Chart with confidence intervals for the
• Test: You can conclude that the % defective is greater than 3% at
3. Chart Parameter of the variable, that includes the value of the Target (H0) or does not include
the 0,05 level of significance. (HA)
3. 90% CI for % Defective • CI: Quantifies the uncertainty associated with estimating the %
Is the entire interval above the target? defective from sample data. You can be 90% confident that the true
3 % defective is between 4,45% and 18,79%, and 95% confident that 4. Comments Summary and comments about results
it is greater than 4,45%.
5. Power & Sample Power of the Sample Size to detects the practical relevant difference (if it is present in
Size the population)
Example: the critical, to be identified difference of 7% if present (3% Defects:= Target
value, 10% Defects:= Customer gets money back -> Difference =7%), can be identified
with the probability of 80,8% with the Sample Size of 60 (Rule of the thumb: necessary
5 10 15 20
Power= 80% = 1-beta)
6. Power & Sample - Relations between Power and Sample Size for different levels of Power
Size - actual Power and Sample Size of the Test
6.
5.
Example: compare the % defective Cookies baked per day with a target for defects
Dr. Reiner Hutwelker Lean and Six Sigma 94
DMAIC > Statistics >> Hypothesis >>> Tests >>>> Chi-Square Goodness-of-Fit Technische Universität München
2.
Chi-Square Goodness-of-Fit 1.
Purpose
Compare the relative frequencies of the categories of a variable (Y), in relation to a) their
expected values or b) specific target values (YT) 6.
Hypothesis
2.
There is a/ no Difference in: percentual amounts of Y and Target-
Difference
Value/ Expected-Value between: categories of Y 3. 4. 5.
Example
2.
Outcome Comparison Chart a) Description of the Sample values, Target values and deviation for each category
Compare the sample and target percents. 2. Statistics
b) 95% Confidence Intervals (CI) for Sample Values (%)
Percents
Sample
Chocolate Target
a) Sample % and Target %
3.a 4. 3. Chart b) Contribution to the Chi-Square Value (= Test Value) by Category , i.e.
Comments the higher the value/ longer the bar, the more contributes the effect of this category to the
Vanilla
• Test: You can conclude that the process percents differ from the target
Significance of the result
percents at the 0,05 level of significance.
• Outcome Table: Shows which process percents differ from their target
percent. Consider the size and direction of the difference to determine if
4. Comments Summary and comments about results
it has practical implications.
• Outcome Comparison Chart: Compare the relative size of the sample
Cocos
and target percents for each outcome.
Almond
3.b 2.b
3.
1-Sample t-Test
Purpose 1.
Compare the Mean of a variable (Y) with a target value (YT)
Hypothesis
2.
There is a/ no Difference in: the level of Values (Y) between: Sample
Difference
and Target
Example 3.
Comparison of Cycle Times with a Target
Y Scale Level 5.
1 cardinal
x Scale Level
./. ./.
Alternative in Minitab Stat Menu Dialog
1. Data Column Variable with the measured values
Stat/ Basic Statistics/ 1-Sample t
2. Target Value Target Value, which should be compared with the Mean of the Variable
note …
Mean of Sample > Target Value (one-sided significance test)
Y-Variable: normal distributed; Sample Size N: > 20
3. Type of Test Mean of Sample < Target Value (one-sided significance test)
3.
6. 7.
4.
2-Sample t-Test
1.
Purpose
Compare the Means of two independent variables with each other (Y1 vs Y2)
Hypothesis 2.
There is a/ no Difference in: Mean between: Variables (Y_pre vs.
Difference
Y_post)
Example
Y Scale Level
1 cardinal
x Scale Level
1 nominal
Dialog
Alternative in Minitab Stat Menu Arrangement of Data: see slide: Grouped data can be arranged in two alternative ways
1. Sample Data
two cells with the measured data of both variables
Stat/ Basic Statistics/ 2-Sample t
Mean of Sample Y_1 > Mean of Sample Y_2 (one-sided significance test)
note … 2. Type of Test Mean of Sample Y_1 < Mean of Sample Y_2 (one-sided significance test)
Y-Variable: normal distributed; Sample Size N: > 20 Mean of Sample Y_1 ≠ Mean of Sample Y_2 (two-sided significance test)
4. Power Critical Difference between Samples, that need to be at least discriminable, if given
4.
6. 7.
5.
Paired t-Test
Purpose 1.
Compare the Means of two dependent/ matched variables with each other (Ya vs. Ya´ )
Hypothesis
Difference There is a/ no Difference in: Mean between: Variables (Ya vs. Ya´) 2.
Example
4. Power Critical Difference between Samples, that need to be at least discriminable, if given
5.
5. 6.
6.
2-Sample Standard Deviation Test: Compare the Standard Deviations of two Variables
Test
Purpose
Compare the Standard Deviations of two independent variables (Y1 vs. Y2)
1.
Hypothesis
Example
2.
Compare the Standard Deviations of Cycle Time between Experts vs. Beginners
Comparison of the variation of Six Sigma Competence before vs. after Training
3.
Comparison of the Distribution of Chocolate Pieces for 2 different stirring durations
Y Scale Level
1 cardinal 4.
x Scale Level
1 nominal
Alternative in Minitab Stat Menu
Dialog
Stat/ Basic Statistics/ 2 Variances Arrangement of Data: see slide: Grouped data can be arranged in two alternative ways
1. Sample Data
two cells with the measured data of both variables
note … Mean of Sample Y_1 > Mean of Sample Y_2 (one-sided significance test)
Y-Variable: normal distributed; Sample Size N: > 20 2. Type of Test St. Dev. of Sample Y_1 < St. Dev. of Sample Y_2 (one-sided significance test)
St. Dev. of Sample Y_1 ≠ St. Dev. of Sample Y_2 (two-sided significance test)
4. Power Critical Difference between Samples, that need to be at least discriminable, if given
Example: Compare the Distribution of Chocolate Pieces for 2 different stirring durations
Dr. Reiner Hutwelker Lean and Six Sigma 103
DMAIC > Statistics >> Hypothesis >>> Tests >>>> 2-Sample Standard Deviation Test Technische Universität München
6.
2-Sample Standard Deviation Test: Compare the Standard Deviations of two Variables
Sample Size, Mean, Standard Deviation and 95% Confidence Interval (CI) for the
2. Statistics
Standard Deviation
3. Standard Deviations Comparison Chart
Red indicates the standard deviations differ.
Y_Choc_Dis_1
5. Comments
Standard Deviation with Confidence Intervals for: Y1 vs.Y2 (Test: H0: Confidence Intervals
• Test: You can conclude that the standard deviation of Y_Choc_Dis_1 3. Interval Chart
Y_Choc_Dis_2 is greater than Y_Choc_Dis_2 at the 0,05 level of significance. intersect ; HA: CI do not intersect)
• Comparison Chart: Red intervals indicate that the standard deviations
1 2 3 4 differ. Consider the size of the difference to determine if it has practical
implications.
• Distribution of Data: Compare the spread of the samples. Look for
4. Histogram Distributions of the Values Y1 and Y2
unusual data before interpreting the results of the test.
4. Distribution of Data
Compare the spread of the samples. 5. Comments Summary and comments about results
Y_Choc_Dis_1
Power (%) for critical Difference between Mean and Target, that needs at least to be
identified
6. Power
Y_Choc_Dis_2
The difference of >=50% can with the current sample size (N= 50) be identified with a
probability of 96,6% (Rule of the Thumb: Power >= 80%)
Power (%) for the to be identified critical difference between the Mean and the Target
Value for different Sample Sizes (N)
6 8 10 12 14 16 18 20 7. Power
Example: the critical difference of 50 can be detected, if given, with a probability of 80%, if
the Sample Size= 27
6. 7.
Example: Compare the Distribution of Chocolate Pieces for 2 different stirring durations
Dr. Reiner Hutwelker Lean and Six Sigma 104
DMAIC > Statistics >> Hypothesis >>> Tests >>>> ANOVA Technische Universität München
7.
One-Way ANOVA
1.
Purpose
Compare the differences in the Means of a dependent variable (Y) in respect to factorial
scaled independent variable (x)
2.
Hypothesis
Example
3.
Compare the Cycle Time of more than 2 Processes
Y Scale Level
1 cardinal
x Scale Level
Dialog
1 nominal / ordinal
Arrangement of Data: see slide: Grouped data can be arranged in two alternative ways
Alternative in Minitab Stat Menu 1. Sample Data
two cells with the measured data of both variables
2. Y-Column Y (Cardinal Scaled Results)
Stat/ ANOVA/ One-Way or other
2. X-Column X (Nominal Scaled Factorial Levels)
note …
Y-Variable: normal distributed; Sample Size N: > 20; N of all Factorial Levels must be the 3. Alpha-Level Significance level for the test
same 4. Power Critical Difference between Samples, that need to be at least discriminable, if given
7.
2. Comparison of a) Table indicates sign. Differences between every xi and every other Factor Level of x
Means Comparison Chart Means b) Interval Plot indicates sign. Differences by red/ non overlapping CI´s for x´s
2.b Red intervals that do not overlap differ. 3. Comments
• Test: You can conclude that there are differences among the means
at the 0,05 level of significance. 3. Comments Summary and comments about results
Choc • Comparison Chart: Look for red comparison intervals that do not
overlap to identify means that differ from each other. Consider the size
of the differences to determine if they have practical implications. 4. Histogram Distribution of Y (for the different Factor Levels of x)
5. Time Series Plot Time Series Plot of Y with Outliers (for the different Factor Levels of x)
Vanilla
4. 5. 6.
7.
7.
8.
Chi-Square % Defective Test: Compare the percentual amounts Y of different factor levels of x
Test
Chi-Square % Defective
1.
Purpose
Compare the percentual amounts Y of different factor levels of one attribute x (e.g. % 2.a 2.b
defective vs. not; % sold vs. not)
Hypothesis
Y Scale Level 5.
1 nominal (counted > discrete cardinal)
x Scale Level Dialog
1 nominal 1. Test item name Name of the Analysis
Alternative in Minitab Stat Menu
2.a X variable name Name of x
note … 2.c Factor Levels Factor Levels/ Categories of X; Data can be imported from Worksheet or entered manually
Test does not have to be about defects, but about interesting portions in any other
3.a Total Number Total Number of Units on each Factor Level/ Category x
attribute X. - Number of Factor Levels of Xi can vary from: 3 - 12.
3.b Number of
Number of Units with interesting attribute (e.g. defects) on each Factor Level/ Category x
Defects
4. Alpha-Level Significance level for the test
5. Power Critical Difference between Samples, that need to be at least discriminable, if given
8.
Chi-Square % Defective Test: Compare the percentual amounts Y of different factor levels of x
Chi-Square % Defective Test for Cookie_Defec by Cookie_Type Results
Summary Report
Do the % defectives differ?
1. Do the % defectives differ? 2.a Which % defectives differ? The bar of the chart indicates:
0 0,05 0,1 > 0,5 # Cookie_Type Differs from - orange line: actual p-value of the Significance Test
1. Significance Test
1 Vanilla 4 - dark blue sector: 0% > alpha <= 5%
Yes No 2 Choc 4
P < 0,001 3 Cocos
- light blue sector: 5% > alpha <= 10%
Differences among the % defectives are significant (p < 0,05). 4 Muffin 1 2 accept H0 , if p> alpha, e.g. There are no differences
accept HA, if p<= alpha, e.g. There is a difference
a) Table indicates sign. Differences between every xi and every other Factor Level of x
2. Comparison of %
2.b % Defectives Comparison Chart
Red intervals that do not overlap differ.
3. Comments
b) Interval Plot indicates sign. Differences by red/ non overlapping CI´s for x´s
• Test: You can conclude that there are differences among the %
defectives at the 0,05 level of significance. 3. Comments Summary and comments about results
Vanilla • Comparison Chart: Look for red comparison intervals that do not
overlap to identify % defectives that differ from each other. Consider
oberseved and expected frequencies, (from the deviations the Chi^2 statistic is
the size of the differences to determine if they have practical 4. Statistics
implications. calculated)
Number of tested Units, defective Units, % defect Units and Confidendence Interval (CI)
Choc
5. Statistics
for % Defectives
Power (%) for critical Difference between Mean and Target, that needs at least to be
identified
The difference of >= 10% can with the current varying sample sizes (10 .. 1000) be
Cocos
identified with a probability of 97,7 - 100% (Rule of the Thumb: Power >= 80%)
6. Power
Power (%) for the to be identified critical difference between the Mean and the Target
Value for different Sample Sizes (N)
Muffin
Example: the critical difference of 10% can be detected, if given, with a probability of
0 20 40 60 80
80%, if the Sample Size>= 26 in each Category/ Factor Level
5.
4.
6.
9.
Chi-2 Test for Association: Compare the frequencies (Y) of 2 concatenated Variables (Xi, Xj)
Test
Compare the Salary (Y) in respect to Profession (Xi) and Country (Xj)
Compare the number of car accidents (Y) by car manufacturer (Xi) and Level of
Expertise of Driver (Xj)
Compare the amount of sold Cookies (Y), differentiated by Type (Xi) and Continent (Xj)
7.
Y Scale Level
1 nominal (counted > discrete cardinal) Dialog
x Scale Level
1. Sample Data The data can be a) imported from the Worksheet or b) entered directly into the data table
2 nominal
2. Orientation of the
Alternative in Minitab Stat Menu table The table can be inverted. This switching of variables in the table influences their
orientation in the result charts, but does not affect the resulks/ significance test
Stat/ Tables/ Chi-Square-Test for Association
3. Name descriptive name for Y/ X
note …
4. Number of
3-6 different categories are possible
Number of Factor Levels for Xi and Xj can vary from: 3 - 6. Categories
4. Number of
3-6 different categories are possible
Categories
6. Data Table Data imported from the Worksheet through Drop-Down-Lists or entered directly
Example: Compare the amount of sold Cookies (Y), differentiated by Type (X1) and Continent (X2)
Dr. Reiner Hutwelker Lean and Six Sigma 109
DMAIC > Statistics >> Hypothesis >>> Tests >>>> Chi-Square Test for Association Technische Universität München
9.
Chi-2 Test for Association: Compare the frequencies (Y) of 2 concatenated Variables (Xi, Xj)
Chi-Square Test for Association: Cookie_Type by Cookie-Type Results
Summary Report
Do the percentage profiles differ?
1. Do the percentage profiles differ? 4. Comments
0 0,05 0,1 > 0,5 • Test: You can conclude that there are differences among the The bar of the chart indicates:
outcome percentage profiles at the 0,05 level of significance. - orange line: actual p-value of the Significance Test
Yes No • Percentage Profiles Chart: Use to compare the profile for each value 1. Significance Test
P < 0,001
of Cookie-Type and the average profile. - dark blue sector: 0% > alpha <= 5%
• % Difference Chart: Look for long bars to identify outcomes with the
Differences among the outcome percentage profiles are significant (p < greatest % difference between observed and expected counts.
- light blue sector: 5% > alpha <= 10%
0,05). You can conclude there is an association between Cookie_Type and accept H0 , if p> alpha, e.g. There are no differences
Cookie-Type.
accept HA, if p<= alpha, e.g. There is a difference
Percentual distribution of Y in Xi/ Xj pairings. The average describes the profile of variable
2. Percentage-Profile
Percentage Profiles Chart
over all Xi categories. The deviations of Xi on the different Levels of Xj shows the specific
Chart
2. Compare the profiles. Xi/ Xj profile. The deviations are explicitly depicted in the % Difference chart.
Average 25% Vanilla
25% Choc
% Difference between Observed and
25%
25%
Cocos
Muffin
3. Expected Counts
The difference chart shows the differences in Xi and Xj between expected and observed
Vanilla
Africa 0% Africa
Choc
3. % Difference Chart values. The larger the difference, positive or negative, the greater the contribution to the
50%
1%
49%
Cocos Chi^2-Test-Value and thus to the significance of the results.
Muffin
America
America 25%
25%
25%
25% 4. Comments Summary and comments about results
Asia
Asia 10%
Number of observed vs. expected values for each Xi/Xj combination, as well as violations
20%
30%
5. Statistics
40% Australia
of conditions of the Chi2-Test
Australia 40%
30%
20%
10% Europe
Europe 50%
0%
49% -100% -50% 0% 50% 100%
1%
Positive: Occur more frequently than expected
0% 15% 30% 45% 60% Negative: Occur less frequently than expected
5.
Example: Compare the amount of sold Cookies (Y), differentiated by Type (X1) and Continent (X2)
Dr. Reiner Hutwelker Lean and Six Sigma 110
DMAIC > Statistics >> Hypothesis >>> Tests >>>> General Linear Model Technische Universität München
Many parametric Statistical Tests are based on the General Linear Model:
Examples for GLM based Tests are: Correlation, Regression, ANOVA, Factor-Analysis, Discriminant Analysis
Dr. Reiner Hutwelker Lean and Six Sigma 111
DMAIC > Statistics >> Hypothesis >>> Tests >>>> General Linear Model Technische Universität München
The predictions from the general linear model (GLM) are more accurate, …
y= b + ax (+ e) y= b + ax + e
R-Sq(adj) 100,0%
y= b + ax + e R-Sq(adj) 83,8%
5 5
Y_Choc-Cookie-Taste_1
Y_Choc-Cookie-Taste_1
4
4
3
3
2
1
1
0
2 4 6 8 10 12 0 2 4 6 8 10 12
x_01_Chocolate-Weight x_01_Chocolate-Weight
… the smaller the variability in the collection of values around the regression line (=Residuals).
Dr. Reiner Hutwelker Lean and Six Sigma 112
DMAIC > Statistics >> Hypothesis >>> Tests >>>> General Linear Model Technische Universität München
}
5
residual yi - yi
Y_Choc-Cookie-Taste_1
4
yi - y
estimated yi - y
3 yquer (3,5)
2 SST = SSR + SSE
y= b + ax + e R2= SSR/ SST
1
0
Regression Line
0 2 4 6 8 10 12
x_01_Chocolate-Weight
… the Determinmation with R2= 84% offers a good basis for forecasts and improvements
Dr. Reiner Hutwelker Lean and Six Sigma 113
DMAIC > Statistics >> Hypothesis >>> Tests >>>> General Linear Model Technische Universität München
The positions of the xY datapoints give a first impression of the degree of Relationship between x and Y.
The Correlation-Coefficient: rxY numerically indicates the degree of Relationship between x and Y.
rxY can vary between -1 and +1,
- 1 := perfect negative linear Relationship,
0 := no Relationsdhip
+ 1 := perfect positive linear relationship. -1 ≤ rxy ≤ 1
The Regression Analysis calculates the influences of multiple independent variables (Xi, Xp) on the
dependent variable (Y) at the same time.
r = 0,85 r = 0,5 r = 0,1 r = - 0,8
Value of “r“ Definition
. ... . ....... . ... .
... . . .. . .. . ... .
.
< r <_ 1
0,7 _ Strong positive correlation
. ..
..
.
... .
.
. . . .. . .
.
.
0,3 _< r _ Medium positive correlation
.
< 0,7
.. .
. . ... .
.
.
..
- 0,3 _< r _ ... .
. ..
. .
..
- 0,7 _< r _
.
... .
.. . . .
.
.
.
- 1 _< r _
< - 0,7 strong negative correlation
.
..
10.
Regression
Purpose 1.
Analyse the relationship between (multiple) x and Y
2.
Hypothesis
3.
Relationship There is a/ no Relationship between: Influence (x) and: Result (Y)
Example
Relationship between baking time (x) and browning degree of the Cookie (Y)
3.
Y Scale Level
1 cardinal
x Scale Level
n cardinal Dialog
Alternative in Minitab Stat Menu Y := cardinal scaled variable of Results
1. Sample Data
x := cardinal scaled variable of Influence
If data are collected in time order then the time dependency of the Residuals can be
Stat/ Regression 2. Order of Data
shown, if present
Based on the type of assumed Relationship between x and Y the type of model can be
note … 3. Type of Model
selected, that will be tested
Choose for me: Minitab chooses the model with the best adjustments according to the
X- and Y-Variables: normal distributed; Sample Size N: > 15 a)
collected data
b) Linear: Influence is represented as x in the Model to predict Y
Example: Relationship between baking time (x) and browning degree of Cookie (Y)
Dr. Reiner Hutwelker Lean and Six Sigma 115
DMAIC > Statistics >> Hypothesis >>> Tests >>>> Regression Technische Universität München
10.
20
1.a 13
13,5
13,122
13,616
(11,936; 14,308)
(12,432; 14,801) Clusters Unequal Variation
14 14,111 (12,927; 15,295)
14,5 14,606 (13,422; 15,789)
15 15,100 (13,916; 16,284) -1
18
Y_Browning_of_Cookie
14
18,5
19
18,562
19,057
(17,372; 19,752)
(17,864; 20,249) 2. Residuals vs Observation Order
Look for nonrandom patterns and large residuals.
19,5 19,551 (18,357; 20,746) Large Residuals Cyclical
20 20,046 (18,849; 21,243)
12 20,5 20,540 (19,341; 21,740) 1
21 21,035 (19,832; 22,238)
10 1.b
0
8 Trend Shifts
10 12 14 16 18 20 22
-1
x_Baking_Time
1 10 20 30 40 50 60 70 80 90 100 110
To obtain additional predicted values, right-click the graph and use the crosshairs tool.
Results Results
Plot shows the Residuals, i.e. the deviation of the data points from the Regression Line,
1. Prediction Plot Scatter Plot for the Variables x and Y 1. Residuals vs. Fitted
along the scale of x, i.e. from small to large values of x
Plot shows the Residuals, i.e. the deviation of the data points from the Regression Line,
1. Residuals vs.
a) Regression Line along the time order of the collected data, as given in the Worksheet, i.e. from first to last
Observation Order
collected data of x
95% Prediction Intervall (PI), i.e. the interval in which the predicted value of Y will be with a Signals as different patterns which show, that the Residuals are systematically influenced;
b) 3. Signals
95% confidence, for a given x try to identfy the Root-Causes of these patterns and eliminate them;
2. Statistics Values for x, the predicted Y and ist 95% Prediction Interval (PI)
Example: Relationship between baking time (x) and browning degree of Cookie (Y)
Dr. Reiner Hutwelker Lean and Six Sigma 116
DMAIC > Statistics >> Hypothesis >>> Tests >>>> Regression Technische Universität München
10.
Y_Browning_of_Cookie
P < 0,001
18
Y_Browning_of_Cookie
14
2. % of variation explained by the model 10
12 0% 100%
10,0 12,5 15,0 17,5 20,0
Low High x_Baking_Time
10
R-sq = 96,58%
96,58% of the variation in Y_Browning_of_Cookie can be explained by the
5. Comments
10 12 14 16 18 20 regression model. The fitted equation for the linear model that describes the
x_Baking_Time relationship between Y and X is:
Y = 0,2632 + 0,9891 X
If the model fits the data well, this equation can be used to predict
Y_Browning_of_Cookie for a value of x_Baking_Time, or find the
Statistics
Selected Model
Linear
Alternative Model
Quadratic 3. settings for x_Baking_Time that correspond to a desired value or
2. R-squared (adjusted) 96,54% 96,51% -1
Correlation between Y and X
0 1
range of values for Y_Browning_of_Cookie.
P-value, model 0,000* 0,000* Perfect Negative No correlation Perfect Positive A statistically significant relationship does not imply that X causes Y.
P-value, linear term 0,000* 0,000*
P-value, quadratic term — 0,848
Residual standard deviation 0,594 0,597 0,98
The positive correlation (r = 0,98) indicates that when x_Baking_Time
increases, Y_Browning_of_Cookie also tends to increase.
* Statistically significant (p < 0,05)
Results
Is there a Relationship between X and Y?
Results The bar of the chart indicates:
- orange line: actual p-value of the Significance Test
1. Significance Test
1. Fitted Line Plot Scatter Plot for the Variables x and Y - dark blue sector: 0% > alpha <= 5%
- light blue sector: 5% > alpha <= 10%
accept H0 , if p> alpha, e.g. There are no differences
a) Regression Equation for the Prediction of Y by the values of x accept HA, if p<= alpha, e.g. There is a difference
2. % variation The determination coefficient (R-squared/ R2) shows the percentage of variation of Y that
Large Residuals (if the Root-Cause of these deviating values is known, then they might be explained can be explained by the variation of x. The remaining percentage is the error portion.
b)
exluded from the calculation)
Level of Correlation between x and Y, which can be positive or negative. (The correlation
2. Statistics for the - R2 (adjusted): Percentual degree of variation of Y explained by x 3. Correlation
rxy in this univariate Model (only one X) is the squareroot of R2.)
tested Models - p-values for the tested Models
4. Scatter Plot Scatter Plot for X and Y with Regression Line and the Regression Equation
Example: Relationship between baking time (x) and browning degree of Cookie (Y)
Dr. Reiner Hutwelker Lean and Six Sigma 117
DMAIC > Statistics >> Design of Experiments (DoE) Technische Universität München
Design of Experiments
(DoE)
The DoE:
identifies Variables (X) with the highest impact (Main Effects) on the result of the Outputs
(Y) and thus can be seen as Root-Causes of the Variation
identifies interdependencies (Interactions) between different levels of at least two
Influences (X) on the result of the Outputs (Y)
quantifies and thus predicts, how and to which degree the variation of Influences (X) affect
the result of the Outputs (Y)
identifies the specific adjustments for the Influences (X) to optimize the results of the
Outputs (Y) in direction of a target
SigmaCopter AG
Situation
Legal conditions:
It is not allowed to change the basic design of the SigmaCopter (e.g. „paper planes“ are not admitted)
Budget for material and tests is limited to 2.500.000 €.
Each Prototype can be tested in repeated measurements but:
It is not allowed to modify a configured and tested Prototypes to serve as another Prototype
Rotor
Allowed Tools: short long
Fuselage Fuselage
scissors, glue stick, timer straight cut straight cut
yes
Clip
big
Zeit: 90 min.
no
Size
Example
yes
small
Clip
no
Dr. Reiner Hutwelker Lean and Six Sigma Exercise 121
DMAIC > Statistics >> Design of Experiments (DoE) Technische Universität München
** Source: Minitab 17
Dr. Reiner Hutwelker Lean and Six Sigma Exercise 122
DMAIC > Statistics >> Design of Experiments (DoE) >>> SigmaCopter (Size: big; …) Technische Universität München
SigmaCopter SigmaCopter
A C
B D
A C
B D
C C
SigmaCopter
SigmaCopter
A A
B D B D
C C
SigmaCopter
SigmaCopter
A A
B D B D
C C
SigmaCopter
SigmaCopter
A A
B D B D
C C
SigmaCopter
SigmaCopter
A A
B D B D
Dr. Reiner Hutwelker Lean and Six Sigma Exercise 124
DMAIC > Statistics >> Design of Experiments (DoE) Technische Universität München
1.
2.
1.
DoE Screening Experiment: Identify the important Influences x for the Result Y
Design of Experiments (DoE)
Screening Design
Purpose 1.
Evaluate 6-15 probably relevant Factors with 2 categorical (nominal) or continuous
(cardinal) levels each (= Influences x) on one (or more) attributes of the Output (Y), to 2.
identify the important x for the succeeding Modelling Design.
Focus
3.
3.a 3.b 3.c
An experiment based on a Screening Design is the most efficient
method to identify the relevant x in a large number of x´s. Screening
Efficiency Designs are typically of resolution III or IV (fractional designs) which
allow to identify significant main effects of many factors with an
efficient number of runs without considering interaction effects.
Example
Identify important influences of Inputs, Methods and Resources (x) on attributes of (Y)
Identify important influences of the layout (x) of the sigmaCopter on its flight duration (Y)
4.
Y Scale Level
1 cardinal Dialog: Create Screening Design (Worksheet)
x Scale Level 1. Name of the Y-Variable (in the Worksheet) (arbitrary)
6-15 nominal or cardinal, slit in 2 levels each
2. Number of Factors/ Influences x (in the Worksheet)
Alternative in Minitab Stat Menu
3 Specification of Factors/ Influences x
Stat/ DoE/ Factorial 3.a Name of the x-Factors (arbitrary)
Type of the Factor:
note … 3.b - Categorical (= nominal) (a categorical variable cannot be handled as continuous)
- Continuous (=cardinal) (a continuous variable can be handled as categorical)
Identify the probably relevant Influences (x) in a preceding Brainstorming / Graphical Range of the Factors/ Influences (x) that should be investigated:
Analysis. Then specify the probably relevant range of the 2 Levels (low vs. high) of each 3.c - Low: lower corner point for each Factor
- High: upper corner point for each Factor
selected Influence/ Factor (x).
Number of Runs (determines the sample size; the higher the number of Runs, the higher
4.
the Power, the smaller the Differences, that could be detected, if present)
Example: Evaluate six probably relevant design features (x) for the flight duration (Y)
Dr. Reiner Hutwelker Lean and Six Sigma 126
DMAIC > Statistics >> Design of Experiments (DoE) >>> Screening Experiment Technische Universität München
1.
DoE Screening Experiment: Identify the important Influences x for the Result Y
1. Create Screening Design
Report Card
Check Status Description
Randomization
i When you create a screening design, Minitab automatically randomizes the order of the experimental runs. Randomization balances
the effect of uncontrollable conditions, such as changes to materials or personnel, and reduces the chance that these conditions will
bias the results. When you conduct the screening experiment, make sure you perform the runs in random order as specified in the
1.a worksheet.
Next Steps
i To complete the screening process:
1. Complete all pre-experiment activities. For more information, view the Pre-Experiment Checklist.
2. Run your experiment in the order specified in the worksheet and collect the response data.
1.b 3. Enter the response data in column C11.
4. Fit the screening model.
5. Identify the critical few factors (5 or fewer) to include in the modeling design.
2.a Y
Main effect: Describes how the response (Y) changes if you
change the setting of one factor (X).
Design Information X
Response Y_Fl_time
Base design 6 factors, 12 runs
Total runs 24
2.b Detection Ability
What effect sizes can you detect with this 24-run design?
< 40% 60% Power 80% 100%
Column with the point type. If you create a 2-level design, Minitab names this
C3 (CenterPt or Effect Size (Shift in the Mean)
column CenterPt. If you create a Plackett-Burman or general full factorial design, Minitab Small Moderate Large
PtType)
names this column PtType. The codes are: 0 is a center point run and 1 is a corner point.
Column with the blocking variable. When the design is not blocked, Minitab sets all
C4 (Blocks)
column values to 1.
< 1 std dev shift 1-2 std dev shift 2+ std dev shift
C5 - Cn Columns with the Factor-Level-Combinations for the experimental runs. Result: Report Card and Summary Report
1. Report Card Information about: a) Randomization StdOrder into RunOrder and b) next steps
empty Column for the measured Results of each experimental run (Factor-Level- Information about: a) the specified Design and b) Power of the experiment, i.e.
C11 2. Summary Report
Combination) probabilities to detect differences of a certain size, i.e. portions of standard deviations
Example: Evaluate six probably relevant design features (x) for the flight duration (Y)
Dr. Reiner Hutwelker Lean and Six Sigma 127
DMAIC > Statistics >> Design of Experiments (DoE) >>> Screening Experiment Technische Universität München
1.
DoE Screening Experiment: Identify the important Influences x for the Result Y
Fit Screening Model for Y_Fl_time Fit Screening Model for Y_Fl_time
Report Card Diagnostic Report
Check Status Description
Look for these patterns: 3.
1. Unusual Data
!
One data point has a large residual and is not well fit by the model. This point is marked in red on the Diagnostic Report and is in row
3 of the worksheet. Because unusual data can have a strong influence on the results, try to identify the cause for its unusual nature.
1. Residuals vs Fitted Values
Look for nonrandom patterns and large residuals.
Correct any data entry or measurement errors. Consider performing trials associated with special causes again and redoing the 0,2
Large Residuals Unequal Variation
analysis.
2. Randomization
i When you create a designed experiment, Minitab automatically randomizes the order of the experimental runs. Randomization
balances the effect of uncontrollable conditions, such as changes to materials or personnel, and reduces the chance that these
0,1
conditions will bias the results. If you did not perform the runs in random order, consider repeating the experiment.
0,0
3. Next Steps Your screening experiment identified 5 critical factors, which are represented by the blue bars in the Pareto chart. You can now use
those factors in a modeling design to create a predictive model for the response. When you set the factor levels in the modeling
design, it is common practice to set them closer together than in the screening design. This can increase the chances of identifying
optimal settings for the critical factors.
-0,1
-0,2
3,0 3,5 4,0 4,5 5,0 5,5 6,0
0,1
0,0
Trend Shifts
-0,1
-0,2
Example: Evaluate six probably relevant design features (x) for the flight duration (Y)
Dr. Reiner Hutwelker Lean and Six Sigma 128
DMAIC > Statistics >> Design of Experiments (DoE) >>> Screening Experiment Technische Universität München
1.
DoE Screening Experiment: Identify the important Influences x for the Result Y
Fit Screening Model for Y_Fl_time Fit Screening Model for Y_Fl_time
Effects Report Summary Report
1. 1. 2.
Main Effects Plot for Y_Fl_time Pareto Chart of Effects Design Information
Describes how changes to a single factor affect the mean of Y_Fl_time. Factors with longer bars have more influence on Y_Fl_time.
Base design 6 factors, 12 runs
x_Weight_g x_Size_mm x_Rotor_Lgth x_Fuselage_D x_Rotor_Des x_Fuselage_L Total runs 24
5,2
x_Weight_g
1.a
Mean of Y_Fl_time
4,8
1.c x_Size_mm
x_Rotor_Lgth
1.a
4,4 x_Fuselage_D
1.b x_Rotor_Des
4,0
80 120 125 250 short long straight cut straight cut short long
1.b
x_Fuselage_L
A gray background represents a factor that was removed from the model because it is not statistically significant.
Comments
4.
0,00
1.c0,25 0,50
Effect
0,75 1,00
You can conclude that 5 of the factors in your model are significant at
The red line is the effect size at the 0,10 level of significance. the 0,10 level of significance.
Gray bars represent non-significant factors that were removed from
the model. The blue bars in the Pareto chart represent the significant factors that
are included in the model. Evaluate the size of the effects to
determine whether they have practical implications.
R-sq = 98,48%
98,48% of the variation in Y_Fl_time can be explained by the model.
Example: Evaluate six probably relevant design features (x) for the flight duration (Y)
Dr. Reiner Hutwelker Lean and Six Sigma 129
DMAIC > Statistics >> Design of Experiments (DoE) >>> Modeling Experiment Technische Universität München
2.
DoE Modeling Experiment: Identify Main-Effects, Interactions of x and optimize the Result Y
Design of Experiments (DoE)
Modeling Design
Purpose 1.
Evaluate 2-5 significant Factors (x) on the Result (Y), identified in the previous Screening
2.
Design, to find those Factor-Level combinations to a) maximize Y, b) minimize Y or c)
achieve a target value of Y.
Focus
An experiment based on a Modeling Design is the most effective
method to identify the optimal settings of a small number of x´s. 3.
Effectivity Modeling Designs are typically full-factorial designs with as many
replicates, which are necessary to identify Effect Sizes of a certain 4.
degree.
Example
Optimize attributes of (Y) based on identified influences of Inputs, Methods and
Resources (x)
Optimize your fitness (Y) based on the identified influences on your lifestyle (x)
4.a 4.b 4.c
Identify the Levels of the design features (x) for the optimal flight duration (Y) 5.
Y Scale Level Dialog: Create Modeling Design (Worksheet)
1 cardinal 1. Name of the Y-Variable (in the Worksheet)
x Scale Level 2. Goal for Y: a) maximize, b) minimize Response or c) achieve a target value for Y
2-5 nominal or cardinal, split in 2 levels each 3. Number of Factors/ Influences x (in the Worksheet)
Alternative in Minitab Stat Menu 4. Specification of Factors/ Influences x
4.a Name of the x-Factors (arbitrary)
Stat/ DoE/ Factorial
Type of the Factor:
note … 4.b - Categorical (= nominal) (a categorical variable cannot be handled as continuous)
- Continuous (=cardinal) (a continuous variable can be handled as categorical)
Range of the Factors/ Influences (x) that should be investigated:
Verify the selected range of the 2 Levels (low vs. high) of each selected Influence/ 4.c - Low: lower corner point for each Factor
- High: upper corner point for each Factor
Factor (x).
Number of Replicates (multiple experimental runs with the same factor settings (levels)
5.
which increase the precision of the model)
Example: Identify the Levels of the design features (x) for the optimal flight duration (Y)
Dr. Reiner Hutwelker Lean and Six Sigma 130
DMAIC > Statistics >> Design of Experiments (DoE) >>> Modeling Experiment Technische Universität München
2.
DoE Modeling Experiment: Identify Main-Effects, Interactions of x and optimize the Result Y
1. Create Modeling Design
Report Card
Check Status Description
Randomization When you create a modeling design, Minitab automatically randomizes the order of the experimental runs. Randomization balances
i the effect of uncontrollable conditions, such as changes to materials or personnel, and reduces the chance that these conditions will
bias the results. When you conduct the modeling experiment, make sure you perform the runs in random order as specified in the
1.a worksheet.
Design Information Interaction: Describes how the response (Y) changes if you
Y
Response Y_Fl_time change the settings of two factors (X).
Goal Maximize
X
Base design
Replicates
5 factors, 16 runs
1 2.b Detection Ability
What effect sizes can you detect with this 1-replicate design?
Center points 16
Total runs 32 < 40% 60% Power 80% 100%
C5 - Cn Columns with the Factor-Level-Combinations for the experimental runs. Result: Report Card and Summary Report
1. Report Card Information about: a) Randomization StdOrder into RunOrder and b) next steps
empty Column for the measured Results of each experimental run (Factor-Level- Information about: a) the specified Design and b) Power of the experiment, i.e.
C10 2. Summary Report
Combination) probabilities to detect differences of a certain size, i.e. portions of standard deviations
Example: Identify the Levels of the design features (x) for the optimal flight duration (Y)
Dr. Reiner Hutwelker Lean and Six Sigma 131
DMAIC > Statistics >> Design of Experiments (DoE) >>> Modeling Experiment Technische Universität München
2.
DoE Modeling Experiment: Identify Main-Effects, Interactions of x and optimize the Result Y
Fit Linear Model for Y_Fl_time Fit Linear Model for Y_Fl_time
Report Card Diagnostic Report
Check Status Description
Residuals vs Fitted Values
Look for these patterns:
3.
1. Unusual Data
!
One data point has a large residual and is not well fit by the model. This point is marked in red on the Diagnostic Report and is in row
11 of the worksheet. Because unusual data can have a strong influence on the results, try to identify the cause for its unusual nature.
1. Look for nonrandom patterns and large residuals.
Correct any data entry or measurement errors. Consider performing trials associated with special causes again and redoing the 0,2
Large Residuals Unequal Variation
analysis.
2. Randomization
i When you create a designed experiment, Minitab automatically randomizes the order of the experimental runs. Randomization
balances the effect of uncontrollable conditions, such as changes to materials or personnel, and reduces the chance that these
0,1
conditions will bias the results. If you did not perform the runs in random order, consider repeating the experiment.
0,0
3. Curvature
i Minitab did not detect any evidence of curvature in your data. When curvature exists, the average response at the center points is
either higher or lower than the average response at the corner (cube) points. A linear model may adequately describe the relationship
between the response and the factors.
-0,1
4. Next Steps Evaluate the optimal solutions in the Summary Report and the Prediction and Optimization Report, which show factor settings that
optimize Y_Fl_time. The Prediction and Optimization Report also shows alternative solutions that are nearly optimal. If the settings
from the optimal solution or one of the alternative solutions are adequate, you should perform 20-30 confirmation runs using those
settings to verify the solution. If the solutions do not meet your goals, you may need to run another experiment using different factor -0,2
settings. If necessary, get help to determine the appropriate next steps. 4,0 4,5 5,0 5,5 6,0 6,5 7,0
0,1
0,0
Trend Shifts
-0,1
-0,2
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
4. Next Steps Summary of the Results of the Modeling Experiment and advices for validating the Result.
Example: Identify the Levels of the design features (x) for the optimal flight duration (Y)
Dr. Reiner Hutwelker Lean and Six Sigma 132
DMAIC > Statistics >> Design of Experiments (DoE) >>> Modeling Experiment Technische Universität München
2.
DoE Modeling Experiment: Identify Main-Effects, Interactions of x and optimize the Result Y
Fit Linear Model for Y_Fl_time
Prediction and Optimization Report
1.
1. Goal: Maximize Y_Fl_time Solution: Optimal Settings
Predicted Y 7,00188 A: x_Size_mm 250 D: x_Rotor_Des straight
95% PI (6,7507; 7,2531) B: x_Weight_g 120 E: x_Fuselage_D cut
C: x_Rotor_Lgth long
1.a
2. Settings and Sensitivity for Optimal Solution
1.c
Blue lines show optimal settings. Black lines or symbols show the predicted Y at different settings.
x_Size_mm x_Weight_g x_Rotor_Lgth x_Rotor_Des x_Fuselage_D
7
6 1.b
5
150 20
0
25
0 80 100 1 20 sho
rt
lon
g
str
a igh
t
cu
t
str
a igh
t
cu
t
1.
A B C D E Predicted Y
250 120 long cut cut 6,9275
250 120 long straight straight 6,74125
250 120 short straight cut 6,46125
250 80 long straight cut 6,33688
250 80 long cut straight 6,24688
3. Predicted Y Predicted values for Y for all Design Points, i.e. all Factor-Level combinations (x)
4. Alternative Ranking of alternative solutions to the optimal solution, which might have advantages, not
Solutions investigated in the experiments (e.g. costs)
Example: Identify the Levels of the design features (x) for the optimal flight duration (Y)
Dr. Reiner Hutwelker Lean and Six Sigma 133
DMAIC > Statistics >> Design of Experiments (DoE) >>> Modeling Experiment Technische Universität München
2.
DoE Modeling Experiment: Identify Main-Effects, Interactions of x and optimize the Result Y
Fit Linear Model for Y_Fl_time Fit Linear Model for Y_Fl_time
1. Effects Report Summary Report
Interaction Plots for Y_Fl_time 1. Pareto Chart of Standardized Effects
2.
Design Information
1.a x_Size_mm * x_Weight_g
Describes how the mean of Y_Fl_time changes if you change the settings of two factors.
x_Weight_g
Terms with longer bars have more influence on Y_Fl_time.
Base design 5 factors, 16 runs
80
6 120 A Factor Name Replicates 1
C A x_Size_mm Center points 16
5 B x_Weight_g Total runs 32
B
C x_Rotor_Lgth
4
x_Size_mm * x_Rotor_Lgth x_Weight_g * x_Rotor_Lgth
E 1.a D x_Rotor_Des
BE
3.
x_Rotor_Lgth E x_Fuselage_D
short D
6
Mean of Y_Fl_time
long
BD Optimal Factor Settings Predicted Y
AE
5
4
CD 1.b x_Size_mm
x_Weight_g
250
120
7,00188
Your goal is to maximize Y_Fl_time. Using the optimal settings for the
Main Effects Plots for Y_Fl_time factors included in the model, the predicted value of Y_Fl_time is
Describes how changes to a single factor affect the mean of Y_Fl_time. 4. 7,00188.
1.b If there is an interaction between factors, use the interaction plots to determine the optimal factor settings. % of variation explained by the model
The model explains 98,33% of the variation in Y_Fl_time.
x_Size_mm x_Weight_g x_Rotor_Lgth x_Rotor_Des x_Fuselage_D 0% 100%
6
R-sq = 98,33%
5 98,33% of the variation in Y_Fl_time can be explained by the model.
4
125,0 250,0 80 120 short long straight cut straight cut
A gray background represents a term that was removed from the model because it is not statistically significant.
Example: Identify the Levels of the design features (x) for the optimal flight duration (Y)
Dr. Reiner Hutwelker Lean and Six Sigma 134
DMAIC > Statistics >> Literature Technische Universität München
Literature
Breyfogle, Forrest W. (2003): Implementing Six Sigma; John Wiley & Sons; 2 edition
Montgomery, Douglas C. (2012): Design and Analysis of Experiments; John Wiley & Sons; 8 edition
Rother, M. & Shook, John (1999): Learning to see; The Lean Enterprise Institute; Spi edition
end of course
Copyright
Dr. Reiner Hutwelker
Kurztrum 31
A-8966 Aich-Assach
[email protected]