Manufacturing Process Control - MIT OpenCourseWare · PDF fileManufacturing 1 Control of...
Transcript of Manufacturing Process Control - MIT OpenCourseWare · PDF fileManufacturing 1 Control of...
MIT OpenCourseWare _________http://ocw.mit.edu___
2.830J / 6.780J / ESD.63J Control of Manufacturing Processes (SMA 6303)Spring 2008
For information about citing these materials or our Terms of Use, visit: ________________http://ocw.mit.edu/terms.
1Manufacturing
Control of Manufacturing Processes
Subject 2.830/6.780/ESD.63Spring 2008Lecture #4
Probability Models of Manufacturing Processes
February 14, 2008
2Manufacturing
Note: Reading Assignment
• May & Spanos– Read Chapter 4
• Montgomery– Skim/consult Chapters 2 & 3 if need additional
explanations or examples beyond May & Spanos
3Manufacturing
Turning Process
4Manufacturing
CNC Turning
0.6970
0.6980
0.6990
0.7000
0.7010
0.7020
Run Number
Shift changes
• Randomness + Deterministic Changes
random or unknown
Δα
Observations from Experiments
5Manufacturing
CNC Data
0.6235
0.6237
0.6239
0.6241
0.6243
0.6245
0.6247
0.6249
0.6251
0.6253
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
OuterMiddleInner
Operating Points:High Feed=1Low Feed = 2
6Manufacturing
Brake Bending of Sheet
M M
ε
σ
Output:
Angle
7Manufacturing
Bending Process
Springback
8Manufacturing
Angle changes with depthΔY ⇒ Δu
• Clear Input-Output Effects (Deterministic)• Also Randomness as well
Observations from Bending Process
20
30
40
50
60
70
80
90
Aluminum 0.3 In
Steel 0.3 In
Aluminum 0.6 In
Steel 0.6 In
9Manufacturing
Observations from Injection Molding
Run Chart for Injection Molded Part
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
Number of Run
Width (mm)Average
Holding Time = 5 secInjection Press = 40%
Holding Time = 10 secInjection Press = 40%
Holding Time = 5 secInjection Press = 60%
Holding Time = 10 secInjection Press = 60%
2/3/05 50
QuickTime™ and aGraphics decompressor
are needed to see this picture.
10Manufacturing
Observations from Data
• Clearly some measurement “noise”?
20
30
40
50
60
70
80
90
shift 1 shift 2
shift 3
11Manufacturing
Observations from Data
• Systematic/traceable “operator error”Sheet Shearing
0.85
0.87
0.89
0.91
0.93
0.95
0.97
0.99
1.01
shift change
shift changealuminum
steel
steel
12Manufacturing
Inputs
A Random Process + A Deterministic Process
How Model to Distinguish these Effects?
Process
Disturbances (Reducible)
Irreducible Disturbances
Outputs +
"Noise"
13Manufacturing
• Consider the Output-only, “Black Box” view of the Run Chart
• How do we characterize the process?– Using Y(t) only
• WHY do we characterize the process– Using Y(t) only?
Process Y(t)
Random Processes
14Manufacturing
The Why
• Did output really change?• Did the input cause the change?• If not, why did the output vary?• How confident are we of these answers?
• Can we model the randomness?
Process
15Manufacturing
Background Needed• Theory of Random Processes and Random
Variables• Use of Sample Statistics Based on
Measurements– SPC basis– DOE: use of experimental I/O data– Feedback control with random disturbances
16Manufacturing
How to Describe Randomness?
• Look at a Frequency Histogram of the Data• Estimates likelihood of certain ranges
occurring:
– Pr(y1 < Y <y2)
– “Probability that a random variable Y falls between the limits y1 and y2”
17Manufacturing
Example: Thermoforming Histogram (2000 data)
0
2
4
6
8
10
12
1.48 1.482 1.484 1.486 1.488 1.49 1.492 1.494 1.496 1.498 1.5 1.502
Thermoforming Run Chart
1.475
1.48
1.485
1.49
1.495
1.5
1.505
Run Number
Part Diameter
Shift Changes
18Manufacturing
How to Describe Continuous Randomness
• Process outputs Y are continuous variables• The Probability of Y(t) taking on any specific value for a
continuumProb(Y(t) = y*) = 0
• Must use instead a Cumulative Probability FunctionPr(Y(t)<y*)
– Look at Cumulative Frequency
19Manufacturing
Cumulative Frequency
0
10
20
30
40
50
60
1.48 1.482 1.484 1.486 1.488 1.49 1.492 1.494 1.496 1.498 1.5
Pr(Y<1.49)=14/50
= 0.28
0
2
4
6
8
10
12
1.48 1.482 1.484 1.486 1.488 1.49 1.492 1.494 1.496 1.498 1.5 1.502 A Continuous equivalent?
20Manufacturing
• Probability Function: (P(x))
• Probability Density Function pdf(x) = dP/dxx
P
x
p(x)
Continuous Equivalents
0
10
20
30
40
50
60
1.48 1.482 1.484 1.486 1.488 1.49 1.492 1.494 1.496 1.498 1.5
0
2
4
6
8
10
12
1.48 1.482 1.484 1.486 1.488 1.49 1.492 1.494 1.496 1.498 1.5 1.502
21Manufacturing
Process Outputs as a Random Variable
• The Histogram suggests a pdf– Parent or underlying behavior “sampled” by the
process• Standard Forms (There are Many)
– e.g. The Uniform and Normal pdf’s
0.1
0.2
0.3
0.4
-4 -3 -2 -1 0 1 2 3 4
Normal Probability Density Function for σ2 = 1, μ=0
x
p(x)
22Manufacturing
Analysis of Histograms
• Is there a consistent pattern?• Is an underlying “parent” distribution suggested?
23Manufacturing
Histogram for CNC Turning
0
5
10
15
20
25
0.697 0.698 0.699 0.7 0.701 0.702
Bin Dimension (in)
CNC Turning
0.6970
0.6980
0.6990
0.7000
0.7010
0.70201 3 5 7 9
11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47
Run Number
Shift changes
24Manufacturing
0
2
4
6
8
10
12
14
134 134.5 135 135.5 136 136.5 137 137.5 138 138.5 139 139.5 140 140.5 141 141.5 142
0.3 in depth only
Histogram for Bending(MIT 2002 data)
Bending Data Sorted by Depth
80
90
100
110
120
130
140
150
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97 101
105
109
113
117
121
125
129
133
Ang
le (D
eg.)
Steel0.6 in depth
Aluminum0.6 in depth
Steel0.3 in depth
Aluminum0.3in depth
Dashed Lines are at group changes
25Manufacturing
0
2
4
6
8
10
12
14
16
18
98 98.5 99 99.5 100 100.5 101 101.5 102 102.5 103 103.5 104 104.5
0.6 in depth only
Histogram for Bending(MIT 2002 data)
Bending Data Sorted by Depth
80
90
100
110
120
130
140
150
1 5 9
13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97
101
105
109
113
117
121
125
129
133
Ang
le (D
eg.)
Steel0.6 in depth
Aluminum0.6 in depth
Steel0.3 in depth
Aluminum0.3in depth
Dashed Lines are at group changes
Mean Shift?
26Manufacturing
Consider: No Intentional Changes (Δu = 0)
• Shearing during shift 1(‘02) , aluminum only
0
10
20
30
40
50
60
0.91 0.915 0.92 0.925 0.93 0.935 0.94 0.945 0.95
27Manufacturing
Consider: No Effective Changes (∂Y/∂u= 0)• Injection Molding Entire MIT Run (2002)
0
5
10
15
20
25
1.875 1.8775 1.88 1.8825 1.885 1.8875 1.89
1.865
1.87
1.875
1.88
1.885
1.89
1.895
Shift Change
28Manufacturing
Injection Molding (S’2003)
0
5
10
15
20
25
30
35
40
65.4 65.5 65.5 65.6 65.6 65.7 65.7 65.8 65.8 65.9 65.9 66 66
Fre
qu
en
cy
65.2
65.4
65.6
65.8
66
66.2
66.4
1 8 15 22 29 36 43 50 57 64 71 78 85 92 99 106 113 120 127
29Manufacturing
Conclusion?
• When there are no input effect (no Δu or ∂Y/∂u) a consistent histogram pattern can emerge
• How do we use knowledge of this pattern?– Predict behavior– Set limits on “normal” behavior
• Define analytical probability density functions
30Manufacturing
Underlying or “Parent” Probability• A model of the “true”, continuous behavior of the
random process• Can be thought of as a continuous random variable
obeying a set of rules (the probability function)• We can only glimpse into these rules by sampling the
random variable (i.e. the process output)• Underlying process can have
– Continuous values (e.g. geometry)– Discrete values (e.g. defect occurrence)
31Manufacturing
Continuous Probability Functions
• Recall Probability Function (Cumulative)P(x) = Prob(Y(t)<x)
• Define pdf = p(x) = dP/dx
Thus :P(x) = pdf (x)dx
−∞
x
∫
x
p(x)x
P
32Manufacturing
Use of the pdf : Expectation
E{x(t)} = expected value of x(t)
E{x(t)} = x(t) pdf (x,t )dx-∞
∞
∫μ(t) = E{x(t)} mean value of x
Note that pdf and μ (or any other expected value)can be functions of time.
In general, they may be non-stationary.
33Manufacturing
pdf(x,t) = pdf(x) = p(x) : stationary pdf
• For a stationary process μx is a constant
Stationary Processes
E{x} = x p(x)dx−∞
∞
∫ = μx
μx : theoretical or " true" mean
34Manufacturing
• For a stationary process σx is a constant
Stationary Processes
σ x2 = E{x 2} − μx
2
E{(x − μx )2} = (x − μx )2 p(x)dx−∞
∞
∫ = σ x2
= "true" variance
= mean square - square of mean
35Manufacturing
The Uniform Distribution
r x
p(x)
1/r
p(x) =1r
⇒ x1 < x < x2
p(x) = 0 ⇒ x < x1 x > x2
x1 x2
36Manufacturing
The Normal Distribution
p(x) =1
σ 2πe
− 12
x − μσ
⎛ ⎝ ⎜
⎞ ⎠ ⎟
2
z =x −
z
μσ
“Standard normal”
μz=0σz=1
0
0.1
0.2
0.3
0.4
-4 -3 -2 -1 0 1 2 3 4
37Manufacturing
-4 -3 -2 -1 0 1 2 3 4
P(z)
Cumulative Distribution
0
0.2
0.4
0.6
0.8
1
-4 -3 -2 -1 0 1 2 3 4
z
38Manufacturing
Properties of the Normal pdf
• Symmetric about mean• Only two parameters:
μ and σ
• Superposition Applies:– sum of normal random variables has a normal
distribution
p(x) =1
σ 2πe
− 12
x − μσ
⎛ ⎝ ⎜
⎞ ⎠ ⎟
2
39Manufacturing
Superposition of Random Variables
If we define a variabley = c1x1 + c2x2 + c3x3 + c4x4 + ...
• ci are constants • xi are independent random variables
μy = c1μ1 + c2μ2 + c3μ3 + c4μ4 + ...
σy2 = c1
2σι2 + c2
2σ22 + c3
2σ32 + c4
2σ42
From expectation operation, for any pdf.
40Manufacturing
Use of the PDF: Confidence Intervals
• How likely are certain values of the random variable?
• For a “Standard Normal” Distribution:
z = (x − μ)σ
N(0,1)μ = 0
σ =1
z = 1 ⇒ x = 1σz = 2 ⇒ x = 2σz = 3 ⇒ x = 3σ
0
0.1
0.2
0.3
0.4
-4 -3 -2 -1 0 1 2 3 4
41Manufacturing
Confidence Intervals
P(-1≤ z ≤ 1) = P(z ≤1) - P(z ≤-1) = 0.841 - (1-0.841) = 0.682(+ 1σ)
P(-2≤ z ≤ 2) = P(z ≤2) - P(z ≤-2) = 0.977 - (1-0.977) = 0.954(+ 2σ)
P(-3≤ z ≤ 3) = P(z ≤3) - P(z ≤-3) = 0.998 - (1-0.998) = 0.997(+ 3σ)
P(z) tabulated (e.g. p. 752 of Montgomery)
42Manufacturing
0.1
0.2
0.3
0.4
-4 -3 -2 -1 0 1 2 3 4
±σ
±2 σ
±3 σ
z
Confidence Intervals
68%95%
99.7%
• Probability that |x| >μ+3σ = 3/1000
43Manufacturing
Is the Process “Normal” ?
• Is the underlying distribution really normal?– Look at histogram– Look at curve fit to histogram– Look at % of data in 1, 2 and 3 σ bands
• Confidence Intervals– Probability (or qq) plots (see Mont. 3-3.7)– Look at “kurtosis”
• Measure of deviation from normal
44Manufacturing
Kurtosis: Deviation from Normal
k = E(x − μ x )4
σ 4
k =n(n + 1)
(n − 1)(n − 2)(n − 3)xi − x
s⎛ ⎝ ⎜
⎞ ⎠ ⎟ ∑
4⎡
⎣ ⎢ ⎢
⎤
⎦ ⎥ ⎥
−3(n − 1)2
(n − 2)(n − 3)
Or for sampled data:
k=1 - normalk>1 more “peaked”k<1 more “flat”
45Manufacturing
The Central Limit Theorem
– If x1, x2 ,x3 ...xN … are N independent observations of a random variable with “moments” μx and σ2
x,
– The distribution of the sum of all the samples will tend toward normal.
46Manufacturing
Example: Uniformly Distributed Data
Sum of 100 sets of 1000 points each
35 40 45 50 55 60 650
50
100
150
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
10
20
30
40
50
60
70
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
10
20
30
40
50
60
70
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
10
20
30
40
50
60
70
x1
x2
x100
+
+ . . .
y = xii =1
100
∑
47Manufacturing
Sampling: Using Measurements (Data) to Model the Random Process
• In general p(x) is unknown• Data can suggest form of p(x)
– e.g.. uniform, normal, weibull, etc.• Data can be used to estimate parameters of distributions
– e.g. μ and σ for normal distribution - p(x) = p(x, μ ,σ)
• How to Estimate– Sample Statistics
• Uncertainty in Estimates– Sample Statistic pdf’s
p(x) =1
σ 2πe
− 12
x − μσ
⎛ ⎝ ⎜
⎞ ⎠ ⎟
2
48Manufacturing
Sample Statistics
x =1n
x( j)j =1
n
∑ : Average or Sample Mean
S2 =1
n − 1
x( j) = samples of x(t ) taken n times
(x( j)j =1
n
∑ − x )2 : Sample Variance
S =1
n − 1(x( j)
j =1
n
∑ − x )2 : Sample Std.Dev.
49Manufacturing
Sample Mean Uncertainty
• If all xi come from a distribution with μx and σ2x, and
we divide the sum by n:
x =1n
xii =1
n
∑
μx = μx σ x2 =
1n
σ x2 or σ x =
1n
σ x
x = c1x1 + c2 x2 + c3x3 + K cn xn
ci =1n
Then: and
50Manufacturing
Conclusions
• All Physical Processes Have a Degree of Natural Randomness
• We can Model this Behavior using Probability Distribution Functions
• We can Calibrate and Evaluate the Quality of this Model from Measurement Data