Slides accompanying 2.008x* video module on Variation and Quality, Prof. John Hart, MIT, 2016.
*Fundamentals of Manufacturing Processes on edX: https://www.edx.org/course/fundamentals-manufacturing-processes-mitx-2-008x
5. 2.008x
Quality: Conformity to
requirements or specifications.
In other words, the ability of a
product or service to consistently
meet customer needs.
Variation: A change in outcome
of a process.
Tolerance: Permissible limit of
variation of a process.
6. 2.008x
What are the measures of
Lego quality?
Drawing from Clipstone, C. J., Hahn, S., Sonnenberg, N., White, C., and Zhuk,
A., 2004, âRazor blade technology.â
Blade edge:
https://scienceofsharp.files.wordpress.com/2014/05/astra_stainless_x_05.jpg
and for Gillette razors?
8. 2.008x
Car body build variation: production launch
Figure 4 from Ceglarek D, Shi J. "Dimensional Variation Reduction forAutomotive Body Assembly."
Manufacturing Review Vol. 8, No. 2, 1995:139-154.
2 mm body project: http://www.atp.nist.gov/eao/gcr-709.htm
6 standard deviations from the mean: 3.4 defects per million!
9. 2.008x
Car body assembly hierarchy
Figure 5 from Ceglarek D, Shi J. "Dimensional Variation Reduction forAutomotive Body Assembly."
Manufacturing Review Vol. 8, No. 2, 1995:139-154.
10. 2.008x
What do we need to
know?
§ What the customer wants (i.e.
what is âgood qualityâ) and how
to relate this to our
specifications.
§ How to quantify variation
(statistically).
§ What causes process variation,
and how to minimize variation
as needed.
§ How to monitor variation and
maintain process control.
11. 2.008x
Agenda: Variation and
Quality
§ The normal distribution
§ Error stackup and simple fits
§ The lognormal distribution
§ Process sensitivity
§ Principles of measurement
§ Statistical process control
§ Conclusion
14. 2.008x
Hex nut thickness:
observations
§ What do we learn from the
distribution of values?
§ Would the values be different
if we measure freehand
versus on the bolt?
Why/not?
§ What is the meaning of the
variation we measured?
15. 2.008x
The normal distribution
Figure 36.3b, Kalpkjian and Schmid, Manufacturing Engineering and Technology
n â â
The histogram of x with n samples approaches the normal
distribution as
Denoted by
: mean (Ă shift)
: standard deviation (Ă flatness)
sometimes denoted s; e.g., 2s
= 2 standard deviations
2
2
( )
21
( )
2
x
x x
x
f x e Ï
ÏÏ
â
â
=
x
xÏ
x â N x,Ïx( )
Ïx
=
1
N
xi
â x( )
2
i=1
N
â
16. 2.008x
Normal probability density function (PDF)
f (x) =
1
2Ï s
â e
â
xâx( )
2
2s2
#
$
%
%
&
'
(
(
From https://en.wikipedia.org/wiki/Normal_distribution (public domain)
17. 2.008x
Cumulative distribution function (CDF)
( )
â„
â„
âŠ
â€
âą
âą
âŁ
⥠â
â
â
â
â
â
â
â â
â 2
2
2
2
1 s
xx
e
sÏ
From https://en.wikipedia.org/wiki/Normal_distribution (public domain)
18. 2.008x
Probability: { }
{ } 1)(
)(
==ââ€â€ââ
=â€â€
â«
â«
â
ââ
dxxfxP
dxxfbxaP
b
a
Normalized to âZ-scoresâ
{ } â«
â
=â€â€
â
=
2
1
2
2
21
2
1
z
z
dz
z
ezzzP
s
xx
z
Ï
b
z
P
x
f(x)
a
0
f (x) =
1
2Ï s
â e
â
xâx( )
2
2s2
#
$
%
%
&
'
(
(
19. 2.008x
Z-scores
z =
x â x
s
Z 0 0.02 0.04 0.06 0.08
-3 0.0013 0.0013 0.0012 0.0011 0.0010
-2.5 0.0062 0.0059 0.0055 0.0052 0.0049
-2 0.0228 0.0217 0.0207 0.0197 0.0188
-1.5 0.0668 0.0643 0.0618 0.0594 0.0571
-1 0.1587 0.1539 0.1492 0.1446 0.1401
-0.5 0.3085 0.3015 0.2946 0.2877 0.2810
0 0.5000 0.5080 0.5160 0.5239 0.5319
0z
P
20. 2.008x
Z-scores
z =
x â x
s
0z
P
Z 0 0.02 0.04 0.06 0.08
0 0.5000 0.5080 0.5160 0.5239 0.5319
0.5 0.6915 0.6985 0.7054 0.7123 0.7190
1 0.8413 0.8461 0.8508 0.8554 0.8599
1.5 0.9332 0.9357 0.9382 0.9406 0.9429
2 0.9772 0.9783 0.9793 0.9803 0.9812
2.5 0.9938 0.9941 0.9945 0.9948 0.9951
3 0.9987 0.9987 0.9988 0.9989 0.9990
22. 2.008x
Example: manipulating the normal
distribution
Car tires have a lifetime that can be
modeled using a normal distribution with a
mean of 80,000 km and a standard
deviation of 4,000 km.
Ă What fraction of tires can be expected
to wear out within ±4,000 miles of the
average?
23. 2.008x
Solution: how many wear out between 76,000
and 84,000 miles?
Ă Area under the curve between these points
z(1) â z(-1) = 0.8413 â 0.1587 = 0.6826
= 68% will wear out
0.8413
0.1587
+1.00-1.00
0.6826
24. 2.008x
Example: manipulating the normal
distribution
Car tires have a lifetime that can be
modeled using a normal distribution with a
mean of 80,000 km and a standard
deviation of 4,000 km.
Ă What fraction of tires can be expected
to wear out within ±4,000 miles of the
average?
Ă 68% will wear out
Ă What fraction of tires will wear out
between 70,000 km and 90,000 km?
25. 2.008x
Solution: failures within 70,000-90,000 miles
Ă % of tires that will wear out =
z(2.5) â z(-2.5) = 0.9938 â 0.0062 = .9876
Ă 98%
0.99380.0062
0 +2.5-2.5
0.9876
27. 2.008x
Measured variation: hex nuts
Single hex nut
Stack of two hex nuts
Mean = 5.58 mm
Stdev = 0.033
Mean = 11.15 mm
Stdev = 0.049
Stack thickness [mm]
Hex nut thickness [mm]
28. 2.008x
Modeling âstackupâ: superposition of random
variables
Proof: http://en.wikipedia.org/wiki/Sum_of_normally_distributed_random_variables
1 2y x x= ±
1 2y x x= ±
1 2
2 2
y x xÏ Ï Ï= +
( )Ï,1 xNx â ( )Ï,yNyâ
( )Ï,2 xNx â
1 2 3 n
In general, if we define a new random variable
y = c1x1 + c2x2 + c3x3 + c4x4 + âŠ
âą ci are constants
âą xi are independent random variables
It can be shown that: ”y = c1”1 + c2”2 + c3”3 + c4”4 + ...
Ïy
2 = c1
2Ï1
2 + c2
2Ï2
2 + c3
2Ï3
2 + c4
2Ï4
2
29. 2.008x
What is the the probability of a successful
assembly?
c = D â d
The new critical dimension
is the clearance (c):
c = D â d
Ïc
= ÏD
2
+Ïd
2
The distribution of
clearances is defined by: DD t±
dd t±
32. 2.008x
Lognormal distribution
Ă The logarithm of x is distributed normally
From http://en.wikipedia.org/wiki/Log-normal_distribution (public domain)
Probability density function (PDF)
Example: size distribution of particles in a powder, size distribution of
grains within a metal
N(ln x;”,Ï ) =
1
xÏ 2Ï
e
â
(ln xâ”)2
2Ï 2
” = ln
m
1+ v / m2
!
"
##
$
%
&& Ï = ln 1+ v / m2
( )
m, v = mean and variance of raw data
Cumulative distribution function (CDF)
33. 2.008x
Lognormal distribution: metal powder for 3D printing
GE fuel nozzle: http://www.gereports.com/post/116402870270/the-faa-cleared-the-first-3d-printed-part-to-fly/
SEM image: http://advancedpowders.com/our-plasma-atomized-powders/products/ti-6al-4v-titanium-alloy-powder/#15-45_m
Ti6Al4V
Specification: 15-45 um
Selective Laser Melting (SLM)
35. 2.008x
How hex nuts are made
Excerpt from: https://www.youtube.com/watch?v=MR6q_nXH2IQ
36. 2.008x
What can cause
process variation?
§ The process: inherent capability;
change of settings.
§ Material: raw material variation,
defects.
§ Equipment: tool wear, equipment
needs maintenance/calibration
§ Operator: procedure, fatigue,
distraction, etc.
§ Environment: temperature,
humidity, vibration, etc.
§ Measurement: Capability of
measurement tool; change of
performance (Ă calibration needed)
§ âŠ
37. 2.008x
Climb milling (first cut) versus
conventional milling (second cut)
6061-T6 Aluminum with ÂŒâ
endmill
Spindle Speed: 4000 rpm
Feed: 20.0 in/min
Depth of cut: 0.400â
Width of cut: 0.070â
38. 2.008x
Example: climb versus conventional milling
Expected width of material .610â
Conventional cut width (red):
Top edge .609â, Bottom of cut .611â
Climb cut width (green):
Top edge .612â, Bottom of cut .619â
Conventional
§ Chip from thin à thick
§ Lower forces but rougher
surface
Climb
§ Chip from thick à thin
§ Higher forces but
smoother surface
41. 2.008x
Injection molding: varying process parameters
Note the mean shifts compared to the variation
40.60
40.65
40.70
40.75
40.80
40.85
40.90
40.95
41.00
0 10 20 30 40 50 60
WidthofPart(mm)
Number of Run
Run Chart for Injection Molded Part
Width (mm)
Average
Holding Time = 5 sec
Injection Press = 40%
Holding Time = 10 sec
Injection Press = 40%
Holding Time = 5 sec
Injection Press = 60%
Holding Time = 10 sec
Injection Press = 60%
Part%radius%[mm]
Run%number
Hold = 5 sec
Pressure = 40% of max
Hold = 5 sec
P = 40% max
Hold = 10 sec
P = 40% max
Hold = 5 sec
P = 60% max
Hold = 5 sec
P = 40% max
42. 2.008x
Ă Systematic (âspecial causeâ)
variation: influences of process
parameters or external
disturbances that can be isolated
and possibly predicted or removed.
Ă Random (âcommon causeâ)
variation: caused by uncontrollable
factors that result in a steady but
random distribution of output
around the average of the data. In
other words, this is the ânoiseâ of
the system.
43. 2.008x
A general model of process variation
Process
Input (u) Output (Y)
Disturbances, such as:
§ Equipment performance changes
§ Material property changes
§ Temperature fluctuations
Control inputs (process
parameter settings)
Sensitivity
Disturbance (α)
ÎY =
âY
âα
Îα +
âY
âu
Îu
44. 2.008x
Some example sensitivities
(if all other parameters are held constant)
Injection molding
§ Relationship between molecular weight of polymer
(determines viscosity) and accuracy (final part
dimension compared to mold)
§ Relationship between injection pressure and accuracy
Machining
§ Relationship between depth of cut and surface
roughness (= spatial frequency of tool marks)
§ Relationship between tool life (sharpness) and accuracy
(= workpiece deformation via higher force and
temperature rise)
45. 2.008x
All together, this determines the amount of variation, and thus
a reasonable tolerance that can be specified!
When the process is âunder controlâ:
If tolerances are too tight:
§ Extra cost (slower rate)
§ More process steps (e.g.
finishing)
§ Lots of scrap (rejects)
§ Manufacturer âno quoteâ
(unreasonable expectations)
ÎY =
âY
âα
Îα +
âY
âu
Îu
Figure 13.30 from Ashby, Material Selection in Mechanical Design
47. 2.008x
Ă Where must the
Resolution be on
this chart?
True (exact) value
Repeatability
Accuracy
Probabilitydensity
48. 2.008x
Accuracy = âthe ability to tell the truthâ
Ă Difference between the measured and true value
Repeatability = âthe ability to tell the same story many timesâ
Ă Difference between consecutive measurements intended to be
identical
Resolution = âthe ability to tell the differenceâ
Ă Minimum increment that can be measured
A. Slocum, Precision Machine Design
50. 2.008x
Mitutoyo high performance micrometer
§ A highly rigid frame and high-performance constant-force
(7-9 N) mechanism enable more stable measurement*
*Patent pending in Japan, the United States of America, the European Union, and
China.
§ Body heat transferred to the instrument is reduced by a
(removable) heat shield, minimizing the error caused by
thermal expansion of the frame when performing
handheld measurements.
http://ecatalog.mitutoyo.com/MDH-Micrometer-High-Accuracy-Sub-Micron-Digimatic-Micrometer-C1816.aspx
Range = 0-25 mm
Resolution = 0.0001 mm (0.1 micron)
Accuracy = 0.0005 mm (0.1 micron)
Flatness: 0.3 micron (across âjawsâ)
Parallelism: 0.6 micron
52. 2.008x
Robot-mounted 3D scanner (Creaform)
â70 micron accuracy over the âsize of a pickup truckâ Ă correcting for low
robot accuracy by imaging dots on the sphere
At IMTS 2014
56. 2.008x
Monitoring a process: CONTROL CHARTS
invented by Walter A. Shewhart (Bell Labs, 1920âs)
§ Needed to improve reliability of telephone transmission systems
§ Stressed the need to eliminate all but âcommon causeâ variation, and
minimize this variation
Ă âa process under surveillance by periodic sampling maintains a constant
level of variability over timeâ
0.990
0.995
1.000
1.005
1.010
0 10 20 30 40 50 60 70 80 90 100
Run number
Average(of10samples)Diameter
Upper control limit
Lower control limit
Step disturbance
66.3%
95.5%
99.7%
58. 2.008x
What might be going on here?
Ă âa process under surveillance by periodic
sampling maintains a constant level of
variability over timeâ
UCL
CL
LCL
0 10 20 30 40 50
57
60
63UCL
CL
LCL
0 10 20 30 40 50
5.0
5.6
6.2
? ?
59. 2.008x
Basic types of control charts
Average chart: plot of mean values
of each sample , centered around
the grand average (mean of all
samples)
Range chart: plot of range of each
sample (max - min), centered
around the average range.
Ă Why do we need both charts?
Figure 36.5 from "Manufacturing Engineering & Technology (7th Edition)" by Kalpakjian,
Schmid. (c) Upper Saddle River; Pearson Publishing (2014).
Control charts are constructed from measurements of samples (each with n parts)
from the population (N, all parts manufactured).
60. 2.008x
Reveals shift
Process mean is
shifting upward
Does not reveal
shift
When the mean shifts:
Sampling
Distribution
x-Chart
R-chart
UCL
LCL
UCL
LCL
61. 2.008x
Does not reveal
increase
Process variability
is increasing
Reveals increase
When the mean shifts:
Sampling
Distribution
x-Chart
R-chart
UCL
LCL
UCL
LCL
62. 2.008x
How do we choose the sample size (n) and frequency
of sampling?
§ Likelihood of unexpected disturbances
§ Importance (cost) of defects
§ Cost of measurement
Ă Typically based on experience and knowledge of the above
(sometimes trial and error)
How do we define the control limits (LCL, UCL)?
§ Based on pre-tabulated statistics of sample variation versus
sample size
63. 2.008x
Calculating the control limits
Average chart
Grand average:
Control limits:
Range chart
Average range:
Control limits:
Figure 36.5 from "Manufacturing Engineering & Technology (7th Edition)" by Kalpakjian,
Schmid. (c) Upper Saddle River; Pearson Publishing (2014).
LCL = X â A2
R
UCL = X + A2
R
LCL = D3
R
UCL = D4
R
R =
Ri
i=1
N
â
N
X =
X i
i=1
N
â
N
64. 2.008x
Factors for calculating control limits
Ă These constants are for a 3-sigma approach, i.e., control limits are
placed at +/- 3 standard deviations from the estimated process mean
Table 36.2 from "Manufacturing Engineering & Technology (7th Edition)" by
Kalpakjian, Schmid. (c) Upper Saddle River; Pearson Publishing (2014).
Average chart
Grand average:
Control limits:
Range chart
Average range:
Control limits:
LCL = X â A2
R
UCL = X + A2
R
LCL = D3
R
UCL = D4
R
R =
Ri
i=1
N
â
N
X =
X i
i=1
N
â
N
65. 2.008x
Process control vs. capability
Ă Even if a process is in control (i.e., constant mean and variation), it
may not be capable (i.e., giving what we want as set by the
specifications a.k.a. the tolerances)
Upper control limit
(UCL)
Lower control limit
(LCL)
In Control and Capable
(Variation from common cause reduced)
In Control but not Capable
(Variation from common causes excessive)
Lower specification
limit(LSL) Upper specification
limit(USL)
66. 2.008x
Control limits vs. tolerances (specification
limits)
Control limits are:
§ Based on process mean and variability.
§ Dependent on the sampling parameters.
Ă Thus, control limits are a characteristic of the process
and measurement method.
Tolerances (specification limits) are:
§ Based on functional considerations.
§ Used to establish a partâs conformability to the design
intent.
Ă Thus, we must have a formal method of comparison.
67. 2.008x
Process capability: compares process
variation to tolerances
use whichever is smaller,
because Ă
Cp
=
USLâ LSL
6Ïx
Cpk
=
USLâ”x
3Ïx
Cpk
=
”x
â LSL
3Ïx
or
General rule: Cp should be at least 1.33
LSL, USL = tolerance limits
Ïx = process stdev
LSL USL
LSL USL
9.80 10.00 10.05 10.20 (mm)
Design
Intent
True process
70. 2.008x
Recommended values of process capability
Ă How do we really judge whatâs good enough?
Knowledge of the âcostâ of defects in our product, thereby
defining a âquality loss functionâ (beyond scope today).
Recommended process
capability for two-sided
specifications
Defects (parts out of
spec) per million
operations
Existing (stable) process 1.33 63
New process 1.50 8
Existing process, safety-
critical
1.50 8
New process, safety-critical 1.67 1
Six-sigma quality 2.00 0.002
72. 2.008x
The big picture
âPilotâ production
This is a control chart
Design for
Manufacturing (DFM)
$$
Does not
conform
Conforms
(good!)
Change design?
Modify process
(know what to do)
75. 2.008x
Reflection: learning objectives
§ Recognize how process tolerances are defined and
variation is monitored, and how a manufacturing process
is established to control variation.
§ Be fluent with manipulation of normally distributed
dimensions, combinations of dimensions (e.g., to predict
fits, lifetimes, etc.).
§ Understand how process physics influence statistical
outcomes (e.g., mean, variation). What are the sensitive
parameters, and how can the variation be addressed?
§ Understand accuracy, repeatability, resolution; assess
the suitability of a measurement technique to monitor a
process.
§ Know how to construct and interpret control charts and
evaluate process capability.