Gershwin Eurandom0606

Manufacturing Systems Design and Analysis
Past Successes and Future Research
Stanley B. Gershwin
Massachusetts Institute of Technology
gershwin@mit.edu
web.mit.edu/manuf-sys
EURANDOM, Eindhoven, The Netherlands

Performance analysis of manufacturing systems
June 19–20, 2006
c 2006 Stanley B. Gershwin. All rights reserved.

Copyright
Introduction
Summary of research in Manufacturing Systems

Design and Analysis
• Motivation
⋆ Economic
⋆ Technical
• Specific MIT research areas
• Successes
• Current and future research
Copyright 2
Economic
Motivation
• Frequent new product introductions.

⋆ Short product lifetimes.
⋆ Short process lifetimes.
This leads to frequent building, rebuilding, and reconfiguration
of manufacturing systems, and this is expensive and
time-consuming .
• Huge capital costs.
⋆ The time required to bring factory to optimal performance is
expensive .

Copyright 3
Economic
Motivation
Industry Needs
• Tools to predict performance of proposed factory

designs.
• Tools for optimal real-time management (control) of

factories.
• Manufacturing Systems Engineering professionals

who understand factories as complex systems.
⋆ They must have systems expertise and intuition.

Copyright 4
Economic
Motivation
Industry Needs
• These tools must be fast as well as accurate.

⋆ Optimization for system design requires many evaluations.
The more evaluations, the better the outcome can be.
⋆ Real-time scheduling in response to random events must be
fast. The factory cannot be idle, and should not be following
an obsolete schedule, while a new one is being constructed.
• Simulation is typically not fast enough. We use

mathematical modeling and analysis to develop our
tools.
Copyright 5
Technical Challenges
Motivation
Complexity
• Collections of things often have properties that are

unexpected functions of the properties of the things
collected.
• Some kinds of production (e.g., semiconductor

fabrication) have hundreds of operations, hundreds
of part types, and/or complicated material flows.
• In many practical situations, infinite buffers are not

good approximations for real buffers.
Copyright 6
Motivation
Randomness, Variability, Uncertainty
• Factories are full of random events:

⋆ machine failures
⋆ changes in orders
⋆ quality failures
⋆ human variability
• The economic environment is uncertain:
⋆ demand variations
⋆ supplier unreliability
⋆ changes in costs and prices
Factories must be built and operated to minimize the

damage that such phenomena can cause.
Copyright 7
Motivation
Quantity, Quality, and Variability
• Quantity – how much and when.

• Quality – how well.
General Statement: Variability is the enemy of

manufacturing.
Variability in processes reduces quality; variability in
event times decreases production rates and increases
inventories and lead times.

Copyright 8
Fundamental
Problems
• The first fundamental problem of manufacturing

systems engineering is: design the best possible
system that meets specified requirements.
• The second problem is: operate a system in the best
possible way to meet specified requirements.
The criterion for “best” and the requirements are

defined in terms of the performance measures and
financial considerations.

Copyright 9
Practical Challenges
Fundamental
Problems
The challenge is to predict the performance measures

of the system before it is built or operated. This is
difficult because:
• Most people lack good intuition for complex systems.
• Good computational tools are challenging to develop.
• The data needed is hard to obtain.

Copyright 10
State of the field
Fundamental
Problems
• Most factories are built with simple rules by people with
experience.
• Often factories are built and then modified.
• Widely-used quantitative tools are either simple and crude
⋆ spreadsheet estimates of capacity
or excessively detailed
⋆ simulation or real-time scheduling optimization.
This was tolerable in the past. However, such approaches are

now inadequate due to reduced factory lifetimes and intensifying
global competition.

Copyright 11
Methodologies
Research
Manufacturing systems research is concerned with the modeling

of systems for the purpose of computing quantity- and
quality-related performance measures. It makes heavy use of
• stochastic processes,
• approximation methods,
• nonlinear analysis and optimization,
• statistics,
• and other mathematical fields.

Copyright 12
Flow Line
Earlier Research
... or Flow shop , Transfer line , or Production line.
Machine Buffer
Traditionally used for high volume, low variety

production.

Copyright 13
Flow Line
Earlier Research
Reference
Manufacturing Systems Engineering

by Stanley B. Gershwin
http://home.comcast.net/~hierarchy/MSE/mse.html

Copyright 14
Flow Line
Earlier Research
Two-Machine Line
M1 B M2
• The machines are unreliable — they fail at random times and
are repaired at random times.
• As a consequence, machines cause each other to be forced
idle at random times for random durations. This reduces
production.
• Goal of analysis: calculate production rate, in-process
inventory.

Copyright 15
Flow Line
Earlier Research
Two-Machine Line
Method of analysis:
1. Construct Markov chain model.
2. Determine steady-state probability distribution.
3. Calculate average production rate and average

inventory from steady-state probability distribution.

Copyright 16
Flow Line
Earlier Research
Two-Machine Line
n=0 n=1 n=2 n=3 n=4 n=5 n=6 n=7 n=8 n=9 n=10 n=11 n=12 n=13
=N−1 =N
(α1 ,α2)
(0,0)
(0,1)
(1,0)
(1,1)
key transitions
states out of transient states
transient out of non−transient states
non−transient to increasing buffer level
boundary to decreasing buffer level
internal unchanging buffer level

Copyright 17
Quantity
Earlier Research
M1 B M2
0.8
• We vary the buffer size N and P Infinite buffer

observe its effect on the
production rate P .
• Observation: the production
rate increases monotonically 0.7
up to a limit.
• However, increasing N
Zero buffer
increases inventory and other
costs.
0.6
20 40 60 80 100 120 140 160 180 200
N
Note: The upper limit is what would be estimated by a simple
capacity formula.

Copyright 18
Quantity
Earlier Research
Long Lines
• Problem: Longer lines have larger state spaces. The

state space grows exponentially in the length of the
line.
• Dealing with large state spaces is the fundamental
technical problem of manufacturing systems
analysis.
• Solution: Approximate one long line by many small
lines.
Copyright 19
Decomposition
Earlier Research
Long Lines
• Decomposition breaks up systems and then reunites them.

• Conceptually: put an observer in a buffer, and tell him that he is
in the buffer of a two-machine line.
• Question: What would the observer see, and how can he be
convinced he is in a two-machine line? Construct the
two-machine line. Construct all the two-machine lines.
Copyright 20
Decomposition
Earlier Research
Long Lines
• Consider an observer in Buffer Bi. Bi−1 Mi Bi Mi+1 Bi+1
⋆ Imagine the material flow process that the observer sees entering and
the material flow process that the observer sees leaving the buffer.
• We construct a two-machine line L(i)
⋆ ie, we find machines Mu(i) and Md(i) Mu (i) B(i) Md (i)
such that an observer in its buffer will see almost the same
processes.
• The parameters are chosen as functions of the parameters of
the long line and of the parameters of the other two-machine
lines.

Copyright 21
Decomposition
Earlier Research
Long Lines
• The number of equations is now linear in the length

of the line.
• An iterative algorithm works well.
• Results have been extensively compared with

simulation, and they are very accurate.

Copyright 22
Successes
Earlier Research
• A Hewlett Packard ink jet printer factory was built using MIT
techniques. The estimated economic impact was hundreds of
millions of US dollars.
• General Motors has used similar tools and claims a savings of

two billion US dollars.
• Peugeot is also using, and further developing, such methods

for designing automobile factories.
• These tools have also been used in frozen food and medical
device production.

Copyright 23
Decomposition
Earlier Research
Extensions
• Optimization: evaluation embedded in gradient search.

⋆ Primal Minimize buffer space subject to production rate
constraint.
⋆ Dual Maximize production rate subject to buffer space
constraint.
• Acyclic A/D (tree-structured) systems
• Single- and multiple-loop systems
⋆ Pallets and kanbans

Copyright 24
Optimal buffer space distribution
Example
• Design the buffers for a 20-machine production line.

• The machines have been selected, and the only
decision remaining is the amount of space to
allocate for in-process inventory.
• The goal is to determine the smallest amount of
in-process inventory space so that the line meets a
production rate target.

Copyright 25
Example
• The common operation time is one operation per

minute.
• The target production rate is .88 parts per minute.
Machine Reliability Parameters

Case MTTF MTTR MTTF MTTR P∞, parts/min
Most machines, min Machine 5, min
1 200 10.5 200 10.5 .95
2 200 10.5 100 10.5 .905
3 200 10.5 200 21 .905

Copyright 26
Example
First question: are buffers really needed?
LineProduction rate with no buffers,

parts per minute
Case 1 .487
Case 2 .475
Case 3 .475
Yes.

Copyright 27
Example
Solution
60
Bottleneck Case 1 −− All Machines Identical

50 Case 2 −− Machine 5 Bottleneck −− MTTF = 100
Case 3 −− Machine 5 Bottleneck −− MTTR = 21
Line Space
Buffer Size
40
30 Case 1 430
20
Case 2 485
Case 3 523
10
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Buffer

Copyright 28
Example
• Observation from studying buffer space allocation

problems:
⋆ Buffer space is needed most where buffer level
variability is greatest!

Copyright 29
Acyclic A/D
Earlier Research
Acyclic A/D (tree-structured) systems
• Straightforward extension of line equations and algorithm

Copyright 30
Single-loop systems
Earlier Research
B1 M2 B2 M3 B3
M1 M4
B6 M6 B5 M5 B4
• Finite buffers (0 ≤ ni(t) ≤ Ni).

P
• Fixed population invariant: i ni (t) = N .
• Motivation:
⋆ Limited pallets/fixtures.
⋆ CONWIP (or hybrid) flow control.

Copyright 31
Loops
Earlier Research
• The invariant creates complications for the

decomposition.
• Multiple-loop systems have one invariant per loop.
• More loops create more opportunities for control.

Copyright 32
Loops
Earlier Research
Raw Material
Raw Material FG
Raw Material
Raw Material CONWIP loop
Machine
Buffer
Material flow
Token (information) flow
Demand

Copyright 33
Multiple part types and reentrant flow
Current
Research

Copyright 34
Current
Research
Type 1
Type 2
Type 1 Type 2

Copyright 35
Current
Research
• We assume strict priority.

⋆ This assumption will be extended.
• Two-part type decomposition has been done.
⋆ Extension to three part types is next;
⋆ extension to more than three should be easy.
• Reentrant flow is modeled as an extension to
multiple-part types.
⋆ Part reenters system as a new part type.
⋆ Parts with fewer remaining operations can be given priority.

Copyright 36
Quality/Quantity Modeling
Current
Research Quality Dynamics
• Definition: How the quality of a machine changes over time.

• The quality literature distinguishes between common causes
and special causes . (Other terms are also used.)
⋆ Common cause: successive failures are equally likely,
regardless of past history.
GGGGGBGGGBGGGGGGGBGGBGGGGBBGGGGGGGG.....
⋆ Special cause: something happens to the machine, and
failures become much more likely.
GGGGBGGGGGBGGGGGGGGBBBBBBGBBBBGBBGB.....
• We use this concept to extend quantity models.

Copyright 37
Current
Research Simplest model
UP/Good UP/Bad
Versions:
• The Good state has 100%

yield and the Bad state has
0% yield.
• The Good state has high yield

and the Bad state has low DOWN
yield.

Copyright 38
Current
Research Simplest model
The relationship between quality dynamics and statistical process

control:
G B
Note: The operator does not know when the machine is in the
bad state until it has been detected.

Copyright 39
Current
Research Inspection in long lines
• The transition from UP/BAD to DOWN is signaled from a

downstream inspection.
• The detection of the failure can only occur when the first bad
part reaches the inspection station.
• Thus the production rate of good parts depends on how much
inventory there is between the machine and the inspection.

Copyright 40
Current
Research Inspection in long lines
To analyze this system by decomposition, we must

• analyze two-machine lines with multiple up- and
down-states, and
• relate the transition rate from UP/BAD to DOWN to
the amount of inventory between the operation and
the inspection.

Copyright 41
Current
Research Separation of Operation and Inspection
Opinions:
• Quantity-oriented people tend to assume that
increasing a buffer increases the production rate.
• Quality-oriented people tend to assume that

increasing a buffer decreases the production rate of
good items.
• However, we have found that the picture is not so

simple.
Copyright 42
Current
Research M1 B M2
Beneficial Buffer
0.8
Effective Production Rate
0.75
0.7
0.65
0 5 10 15 20 25 30 35 40 45 50
Buffer Size
Effective production rate = production rate of good parts.

Copyright 43
Current
Research M1 B M2
1.5
Harmful Buffer
0.5
0
0 5 10 15 20 25 30 35 40 45 50
Buffer Size

Copyright 44
Current
Research M1 B M2
0.4
Mixed-Benefit Buffer
0.395
0.39
0.385
0.38
0.375
0.37
0 5 10 15 20 25 30 35 40 45 50
Buffer Size

Copyright 45
Current
Research Inspections
How many inspections should there be? And where?
• Intuition: more inspection improves quality.
• Reality: increasing inspection can actually reduce

quality, if it is not done well.

Copyright 46
Current
• We simulated a 15-machine, 14-buffer line.

• All machines and buffers were identical.
• We looked at all possible combinations of inspection stations in
which all operations were inspected.
⋆ Example: Inspection stations just after Machines 6, 9, 13, and 15.
⋆ The first inspection looks at the results from Machines 1 – 6; the second
looks at results from Machines 7 – 9; the third from 10 – 13; and the last
from 14 and 15.
⋆ There is always one inspection after Machine 15.
• A total of 214=16,384 cases were simulated.

Copyright 47
Current
Range of Good Production Rates for Different Numbers of Inspection Stations
(15 five-state machines, 14 buffers, information feedback only)
No inspection: 0.10040
0.375
0.365
Best locations
0.355
Good Production Rate
0.345
0.335
0.325
Worst locations
0.315
0.305
0.295
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Number of Inspection Stations

Copyright 48
Current
Research Possible Extensions
• Design of inspection policies using Bayesian

analysis.
• Analysis of multiple-state machines.
• Analysis of lines with multiple-state machines.
• Control of material flow in systems with quality
issues.

Copyright 49
Analytic evaluation of long lines
Current
Research
Ubiquitous inspection:
I I I I I I I I
Single remote inspection of a single machine:

I
Single remote inspection of multiple machines:

Copyright 50
Machine quality dynamics
Current
Research
• The three-state machine model is much too simple.

• One extension is
Quality
UP/Good UP/Bad Repair
DOWN G DOWN B
• ... but even this leaves out important features.

Copyright 51
Machine quality dynamics
Current
Research
• Another extension is
Distribution
of measured
parameter
Gk G k−1 G6 G5 G4 G3 G2 G1 B
Quality
Repair
Dk D k−1 D6 D5 D4 D3 D2 D1 D
• This allows more general wear or aging models.

Copyright 52
Conclusions
• Manufacturing systems engineering is valuable

economically and intellectually.
• There are many interesting and important untreated

problem areas.

Copyright 53

Gershwin Eurandom0606

Caricato da

Informazioni sul documento

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Gershwin Eurandom0606

Caricato da

Copyright:

Formati disponibili

Manufacturing Systems Design and Analysis

Past Successes and Future Research

EURANDOM, Eindhoven, The Netherlands

c 2006 Stanley B. Gershwin. All rights reserved.

Summary of research in Manufacturing Systems

• Frequent new product introductions.

c 2006 Stanley B. Gershwin. All rights reserved.

• Tools to predict performance of proposed factory

• Tools for optimal real-time management (control) of

• Manufacturing Systems Engineering professionals

c 2006 Stanley B. Gershwin. All rights reserved.

• These tools must be fast as well as accurate.

• Simulation is typically not fast enough. We use

• Collections of things often have properties that are

• Some kinds of production (e.g., semiconductor

• In many practical situations, infinite buffers are not

• Factories are full of random events:

Factories must be built and operated to minimize the

• Quantity – how much and when.

General Statement: Variability is the enemy of

c 2006 Stanley B. Gershwin. All rights reserved.

• The first fundamental problem of manufacturing

The criterion for “best” and the requirements are

c 2006 Stanley B. Gershwin. All rights reserved.

The challenge is to predict the performance measures

• Good computational tools are challenging to develop.

• The data needed is hard to obtain.

c 2006 Stanley B. Gershwin. All rights reserved.

This was tolerable in the past. However, such approaches are

c 2006 Stanley B. Gershwin. All rights reserved.

Manufacturing systems research is concerned with the modeling

c 2006 Stanley B. Gershwin. All rights reserved.

... or Flow shop , Transfer line , or Production line.

Traditionally used for high volume, low variety

c 2006 Stanley B. Gershwin. All rights reserved.

Manufacturing Systems Engineering

c 2006 Stanley B. Gershwin. All rights reserved.

c 2006 Stanley B. Gershwin. All rights reserved.

1. Construct Markov chain model.

2. Determine steady-state probability distribution.

3. Calculate average production rate and average

c 2006 Stanley B. Gershwin. All rights reserved.

states out of transient states

transient out of non−transient states

non−transient to increasing buffer level

boundary to decreasing buffer level

internal unchanging buffer level

c 2006 Stanley B. Gershwin. All rights reserved.

• We vary the buffer size N and P Infinite buffer

c 2006 Stanley B. Gershwin. All rights reserved.

• Problem: Longer lines have larger state spaces. The

• Decomposition breaks up systems and then reunites them.

• Consider an observer in Buffer Bi. Bi−1 Mi Bi Mi+1 Bi+1

c 2006 Stanley B. Gershwin. All rights reserved.

• The number of equations is now linear in the length

• An iterative algorithm works well.

• Results have been extensively compared with

c 2006 Stanley B. Gershwin. All rights reserved.

• General Motors has used similar tools and claims a savings of

• Peugeot is also using, and further developing, such methods

c 2006 Stanley B. Gershwin. All rights reserved.

• Optimization: evaluation embedded in gradient search.

c 2006 Stanley B. Gershwin. All rights reserved.

• Design the buffers for a 20-machine production line.