lean six sigma dmaic process: common mistakes and misconceptions during data collection and analysis...
TRANSCRIPT
Lean Six Sigma
DMAIC process:Common Mistakes and Misconceptions
During Data Collection and Analysis
Hans Vanhaute04/08/2014
Goal of tonight’s presentation
Give you a few examples of common mistakes made during “Measure” phase of DMAIC projects.
Draw more widely applicable lessons and conclusions that may benefit you (so you don’t make the same mistakes).
Hopefully provide you with some interesting insights (and don’t put you to sleep).
DMAIC and “Projects”
“A problem scheduled for a solution.”
Management decides the problem is important enough to provide the resources it needs to get the problem solved.
“DMAIC Projects”
Eliminates a chronic problem that causes customer dissatisfaction, defects, costs of poor
quality, or other deficiencies in performance.
Six Sigma DMAIC Project
DEFINE MEASURE ANALYZE IMPROVE CONTROL
Very Data-Intensive
M – Measure
Define a high-level process map.
Define the measurement plan.
Test the measurement system (“Gauge Study”).
Collect the data to objectively establish current baseline.
Typical tools:
- Capability analysis
- Gage R&Rs
The DMAIC steps
Y = f (unknown Xs)
Initial Analysis
Capability Analysis Conundrums
Black Box Process
Unknown Xs
Cpk values that inaccurately
predict process performance
Non-normal data
Instability of the process over time
Case 1: Inherent non-normality of the process output.
Some physical, chemical, transactional processes will produce outcomes that “lean” one way:
- Time measurements- Values close to zero, but that are always positive (surface
roughness RMS…) - …
Process experts or careful analysis of the metric should be able to help with understanding.
Capability Analysis Conundrums
Good news: Capability Analysis of Non-Normal data is possible.Bad news: This situation doesn’t happen very often.
Example
Case 2: Problematic measurement systems
(we’ll come back to that one when we discuss GR&R…)
Capability Analysis Conundrums
Case 3: Failure to stratify the data.
Stratification is the separation of data into categories.It means to “break-up” the data to see what it tells you.
Its most frequent use is when diagnosing a problem and identifying which categories contribute to the
problem being solved. Microsoft Word
97 - 2003 Document
Capability Analysis Conundrums
This is the big one!
Stream 1
Stream 2
Stream 3
Stream 4
Cpk1
Cpk2
Cpk3
Cpk4
Cpk ???
Capability Analysis Conundrums
1 Cpk value?
2 Cpk values?
What is a Cpk value supposed to tell us?
Expected future performance of the process(es) assuming statistical stability
over time.
Capability Analysis Conundrums
Over-estimating variation of the process.(Why?)
Under-estimating process capability.
Leading to all sorts of non-value-added activity for your organization.
Recognize two of the four streams are main drivers of overall capability.Correct estimation of the two most important process capabilities.
Points to appropriate improvement activities.
Capability Analysis Conundrums
Cpk = 0.67
Cpk = 1.15
Cpk = 1.50 Cpk = 1.50
Cpk = 1.15
12010896847260483624121Index
Data
Time Series Plot
Prediction (stratified)
Prediction (not stratified)
Actual data (stratified)
Actual data (not stratified)
Capability Analysis Conundrums
Example
Problematic measurement systems:
2a: Limiting factors to “how well” you can measure something.
2b: I passed my GR&R but I’m still getting “weird” results.
2c: Time effects
Capability Analysis Conundrums
Case 2a: Limits to measurementsGame: Identify the dataset with the highest resolution.
Resolution:
a: The process or capability of making distinguishable the individual parts of an object, closely adjacent optical
images, or sources of light b: A measure of the sharpness of an image or of the fineness with which a device can produce or record
such an image.
Case 2a: Limits to measurements
9.9 10.2 9.6 9.910.5 9.9 9.0 9.911.4 10.2 9.3 10.29.0 9.6 10.2 9.610.2 10.5 10.2 9.610.5 9.9 10.2 9.910.2 11.1 10.5 10.810.2 10.8 9.3 10.210.5 9.3 9.9 9.99.3 9.9 10.8 11.110.8 9.6 10.8 10.89.9 9.3 10.2 9.39.9 10.5 9.9 9.99.9 9.9 10.2 9.6
9.6 10.2 9.6 9.610.2 9.6 9.0 9.611.4 10.2 9.6 10.29.0 9.6 10.2 9.610.2 10.2 10.2 9.610.2 9.6 10.2 10.210.2 10.8 10.2 10.810.2 10.8 9.0 10.210.2 9.6 9.6 9.69.0 9.6 10.8 10.810.8 9.6 10.8 10.89.6 9.0 10.2 9.610.2 10.2 9.6 9.69.6 9.6 10.2 9.6
Which dataset has the highest resolution?
? ?
Measurement Resolution: a: The process or capability of making distinguishable the individual parts of a dataset or closely adjacent
data points.b: A measure of the sharpness of a set of data or of the
fineness with which a measurement device can produce or record such a dataset.
Less resolution
Less resolution
Less resolution
Limiting factors to “how well” you can measure something.
Case 2a: Limits to measurements
9.939710.594411.4290 9.0401
9.910.511.4 9.0
9.610.211.4 9.0
101111 9
21
43
14
12
10
8
6
4
2
0
Frequency
Histogram (unlimited decimal places)
14
12
10
8
6
4
2
0
Frequency
Histogram (1-in-10 rule)
20
15
10
5
0
Frequency
Histogram (1-in-5)
35
30
25
20
15
10
5
0
Frequency
Histogram (1-in-3)
Less resolution
Less resolution
Less resolution
Limiting factors to “how well” you can measure something.
21
43
Case 2a: Limits to measurements
99
95
90
80
70
60
50
40
30
20
10
5
1
Perc
ent
1-in-3 rule
99
95
90
80
70
60
50
40
30
20
10
5
1
Perc
ent
1-in-5 rule
99
95
90
80
70
60
50
40
30
20
10
5
1
Perc
ent
1-in-10 rule
99
95
90
80
70
60
50
40
30
20
10
5
1
Perc
ent
Unlimited decimal places
Less resolution
Less resolution
Less resolution
Limiting factors to “how well” you can measure something.
S = 1.000 S = 1.005 (0.5% over)
S = 1.040 (4% over) S = 1.150 (15% over)
21
43
Case 2a: Limits to measurements
Limiting factors to “how well” you can measure something:
Case 2a: Limits to measurements
Why??“Always done it that way, never given it any thought”.Focus on “meeting specs” not on controlling process.
“Always” round to x decimal places.Nobody told me how many decimals were needed
…
The old “1 in 10” rule of thumb seems to make sense.Resolution must be at least 1/10th of data rangeResolution must be at least 1/10th of spec range
Case 2b: “Weird” StuffI passed my GR&R but I’m still getting “weird” results.
Distribution of Measurements
Distribution of measurement variability
GR&R 101:
P/TV ratio expresses the total measurement variability as a percentage of the total historical process variation.Here P/TV ~ 14%
“Metrics”
Case 2b: “Weird” StuffI passed my GR&R but I’m still getting “weird” results.
Distribution of measurement error
GR&R 101: “Metrics”
P/T expresses the total measurement variability as a percentage of the tolerance width of the process:Here P/T ~ 12.5%
Spec. limits
Case 2b: “Weird” StuffI passed my GR&R but I’m still getting “weird” results.
GR&R 101: “Metrics”
P/TV P/T
Very good <10% <10%
Marginal 10 – 30% 10 – 30%
Needs Improvement > 30% > 30%
Simple, right? Not so fast…
Case 2b: “Weird” StuffI passed my GR&R but I’m still getting “weird” results.
Part-to-PartReprodRepeatGage R&R
100
50
0
Per
cent
% Contribution% Study Var
% Tolerance
0.8
0.4
0.0
Sam
ple
Ran
ge
_R=0.2333
UCL=0.7624
LCL=0
1 2 3
50
40
30Sam
ple
Mea
n
1 2 3
UCL=35.65
LCL=34.77
__X=35.21
10987654321
50
40
30
Parts
321
50
40
30
Tester
10 9 8 7 6 5 4 3 2 1
50
40
30
Parts
Ave
rage
1
2
3
Tester
Gage name: NaOH titrationDate of study: 11/25/2008
Reported by: J im SmithTolerance: 0.01Misc:
Components of Variation
R Chart by Tester
Xbar Chart by Tester
Result by Parts
Result by Tester
Tester * Parts Interaction
NaOH Titration
R chart by operator:Points inside control limits indicate that operator is consistent between repeat measurements made on same sample (GOOD)
Points outside control limits indicate that operator is not consistent between repeat measurements made on same sample (BAD)
Case 2b: “Weird” StuffI passed my GR&R but I’m still getting “weird” results.
Example
P/T = 22%
P/TV = 16%
Case 2b: “Weird” StuffI passed my GR&R but I’m still getting “weird” results.
ExampleP/T = 70%P/TV = 40%
P/T = 10%P/TV = 6%
Case 2b: “Weird” StuffI passed my GR&R but I’m still getting “weird” results.
Example
So… what caused this?
Camera
Lens
Ring Light
Pin Tip Position
Case 2c: Time EffectsThe speed of Information is finite.
Information can come from different distances.
Case 2c: Time EffectsThe speed of Information is finite.
Information can come from different distances.
Moon: 1.2 light-seconds awaySun: 8 light-minutes awayMars: 12.5 light-minutes awayPluto: 5.5 light-hours awayProxima Centauri: 4.2 light-years away
Case 2c: Time EffectsThe speed of Information is finite.
Information can come from different distances.
Just because you “observe” (measure / see) several events “at the same time”, doesn’t mean they all occur(red) at the same time.
Case 2c: Time Effects
Example
Arranged by order of occurrenceArranged by order of observation
Case 2c: Time Effects
What can you do?
Collect the data as close as possible to the origin of the event you are observing.
“Traceability” of the events you are observing.
“De-convolution” of the data.
In mathematics, de-convolution is an algorithm-based process used to reverse the effects of convolution on
recorded data
Blind reliance on some index value (Cpk, Cp, P/T, P/TV,…) to tell you what is going on might get you in trouble.Always:- Make sure you understand how the index is calculated- Use the approach fully, not half-way- Verify that all assumptions were met Data stratification opportunities abound. Identify them early on in your project.
A few simple rules of thumb will quickly help you determine if you have a chance of having a good measurement system.
So… What did we learn?
Further analysis of the Gage R&R data can provide you with some great insights into and improvement opportunities for your measurement process.
Data has a finite speed. Being aware of this and planning for it during your measure phase will help keep you on the right track.
So… What did we learn?
Parting Thoughts
My organization doesn’t use Six Sigma, do these insights benefit me
as well?
Thank You
Questions?