Get documents about "Microsoft PowerPoint - UKKDD-2007-Jenny-talk
W
Description
Microsoft PowerPoint - UKKDD-2007-Jenny-talk
Document Sample
Data Mining Applications in
Manufacturing
Dr Jenny Harding
Senior Lecturer
Wolfson School of Mechanical &
Manufacturing Engineering,
Loughborough University
Identification of Knowledge - Context
Intelligent
Support
Systems
Problem Areas
Additional Facilities?
igns
w Des
New
Prod
uct Ne
Exis
t
Prod ing
g
Existin s Extra Resources?
ucts De sign
New Technology?
Tools:
New
Information Modelling Business
Objectives
Process Modelling
Redesigned Processes?
Artificial Intelligence
Simulation How do we find and use the
Data Mining??? “Right” Knowledge?
Sources of Manufacturing Knowledge
Production Processes
People
Expertise and Experience
Processes
(Operational Data)
Facts or Folklore? Under Exploited?
Background History
Why Data Mining?
What is the “Right” Knowledge?
Are “experts” the only sources?
Is there any substance in folklore?
Or is the knowledge hidden
somewhere else as well?
Reported Applications of DM in Manufacturing
Harding, J A, Shahbaz, M, Srinivas
8 and Kusiak, A, 2006, "Data Mining in
7 Manufacturing: A Review", American
6 Society of Mechanical Engineers
No. of Papers
8
5 (ASME): Journal of Manufacturing
7 4 Science and Engineering.
4 7
2 4
6 3
1 43 5
1
1 1
1 12
3
3 2 23 3 3
3 2
1 1 232
11 2
1
1 1
12 1
1
1 1 121 2 2 2
11
1 11 1
1 2 1 Earliest - Fault diagnosis (1987)
2 2
1987
1 1
11 0
1989
1 1
1 1
1991
1
1993
1995
Ma
1997
E g
Eng
1999
D ec
a nu
2001
Ma
Ma
e is
Substantial increase in
2003
L y
La y
uf a
inee
Sho
2005
i io
Ma
Ma
inte
h p
act
on S
o ut
t n
t ur
r
rin
pF
teri
u in
nan
(manufacturing) data mining
S up
r al
gD
gD
De s
i g
l o
lo o
nce
ppo
l Pr
si gn
rC
esig
i n
Sy s
ort
s n
ro p
Con
y te
publications in last 2 years
t
per
n
n tr
ems
rtie
ro l
o
s
es
s
Reported Applications of DM in Manufacturing (2)
Summary:
Use of data mining applications in Manufacturing is
relatively new
To date – most of the research done has focused on
semi-conductor industry (fault detection and
productivity improvement)
To date – most of the reported research uses decision
trees, clustering and neural networks
Challenges for DM in Manufacturing
Why slow adoption?
? Investment – very wide range of tools and techniques
and not obvious in most applications which will produce
best results and rewards (DM preparation costs are high).
? Availability and quality of data – substantial data
cleaning and pre-processing required.
? Combinations of expertise needed – Data Mining
Expertise + Expertise in Processes / Technologies
under consideration
? Inadequate Exploitation of Resulting Knowledge -
“one-off” applications addressing specific problems
Challenges for DM in Manufacturing (2)
Why slow adoption?
? Uncertainty and risk - there are no guarantees that
useful knowledge will be discovered. In highly
competitive situations high risks may be unacceptable
? Complexity and diversity - extremely difficult to
devise generic data mining processes even for
particular types of manufacturing problems
Case Studies
Experience from literature and
industrial case studies shows……….
An Example
ne
ss aa ab ac ad ae af
ic k
Th
Height
Height
Height
Height
aa
ab
ac
ad
Width
A simple sample
product
ac
ab
aa
Data Cleaning
• Very important! (But very time-consuming)
•Iterative Process
Process 1 Data
Data Data consolidated
Process 2 Error-free
Cleaning
Module table
Data
Process n
General Classifications Possible Causes
Identical Records Breakdowns / Restarts
Some Duplication in Records Rework cycles
Confusing Records Operator Error
Missing Values in Records Operator Training
etc Etc., etc.
Example Thoughts during Cleaning
Should Reworked records be included
or ignored?
Can Missing values be replaced or
should the whole record be rejected?
Can domain experts of more detailed
files be consulted about replacing the
abnormal data?
All the replaced values need to be
documented for future consultation.
Data Transformation
Upper Limit Uout
Upper
H-Upper
M-Upper
S-Upper
Nominal
S-Lower
M-Lower
H-Lower
Lower Limit Lower
Lout 01 02 03 04 05 06 07 08 09 10 11
Products - split σ σ σ
Process variables
tolerance range for
dimension. Normal distributions?
Equal divisions??? Equal divisions???
Data Mining
A complete set of dimensional / manufacturing
data for each product needs to be collected and
transformed to make each record
Regression analysis used first – are there any
simple relationships that can be exploited?
Several data mining approaches tried
Association Rules the Apriori Algorithm has
been most successful
Data Mining
In Manufacturing - Data mining decisions should
be made in the context of the physical realities of
the data that is being examined.
In very highly controlled manufacturing
processes, unusual combinations of results
(representing problems) may be rare.
Support level kept very low to reduce risk of
missing unusual but important combinations of
manufacturing outputs
Association Rule
Used in mostly Market Basket Analysis
Apriori Property:
“Any subset of a large itemset must be large”
Rules are in the form of
A Ł B or A Ł (B & C) etc.
e.g. Chips Ł Coke
•Each rule is based on certain Support and
Confidence level.
Rule Quality
Need to check Quality of the generated
rules – use Support, Confidence, Statistical
methods AND check with domain experts
that the rules make sense in the
Manufacturing Context.
Support: Support is the frequency of the occurrence of
items in a transaction or record.
Confidence: Confidence is the certainty of the
discovered rule.
Example Rules
Types of Manufacturing Rules that may be identified
include:
Product Design Errors or Constraints
Process Errors or Constraints
From Product Data - Rules of the form
Good Dimension output Ł Good Dimension output
Bad Dimension output Ł Bad Dimension output
Good Dimension output Ł Bad Dimension output
From Product and Process Data
Process variable values that produce Good Product Results
Process variable values that produce Poor Product Results
Example Rules
•Rules that indicate one dimension being “nominal”
results in another dimension being “nominal” are
reassuring – but generally not very interesting
•Rules that indicate one dimension being “nominal”
results in another dimension being out of tolerance or
away from “nominal” are more interesting.
Possible Causes include:-
Design Error
Process Error
Process Constraints
Results
• Data Mining can be used to extract
Manufacturing process limitations.
• Data Mining can be used to discover design
constraints and then help to improve the product
design, (if manufacturing process is acceptable).
• Data Mining can be used to improve the output
quality by controlling the manufacturing process
variables.
Future Considerations……….
Data Mining projects tend to be “one-off”
applications to solve a particular problem
But how do we make sure the acquired
knowledge is made use of when it has been
discovered?
Greater benefits would be gained if they were
part of an ongoing Knowledge Reuse
programme
Data Mining Network
Local Site / Process n-1 Local Site / Process 2
Local Data Local Data
Mining Mining
Engines Engines
Local Pool of Local Pool of
Knowledge Knowledge
Local Databases and Files Local Databases and Files
Main Data
Warehouse
Local Pool of
Knowledge
Main Pool of
Local Site / Process 1
Local Site / Process n Knowledge
Local Data
Integration & Local Data
Mining DM Engine Mining
Engines Engines
Local Pool of
Knowledge
Local Databases and Files Local Databases and Files
Any Questions
Thank You
J.A.Harding@lboro.ac.uk
