Design and implementation of an intelligent manufacturing execution system for semiconductor manufacturing industry

IEEE ISIE 2006, July 9-12, 2006,Montreal, Quebec,Canada

Design

and

Implementation of

an

Intelligent Manufacturing Execution

System for Semiconductor Manufacturing Industry

Ruey-Shun Chen, Yung-Shun Tsai, Chan-Chine Chang

Institute of Information

Management, National Chiao Tung University, Taiwan

rschen@iim.

nctu.

edu.

tw

Abstract-This paper proposed an intelligent MES by integrating manufacturing execution system, data warehouse, online analytical processing and data mining system. The data warehousefor this system isprovided by the massiveamountsof datagathered from aMES. Through some process ofintegrating MES with data warehouses, data warehouses with decision analysis, and decision analysis with datamining systems. A three-tiered web-basedsystematic framework has been established. The resultof this study is the integration of the MES and datamining system. According to experimental analysis, the product yield and the manufacturing cycle time all have been improved for manufacturing industry.

Keywords: Data warehouse; Data mining; Decision tree; Manufacturing execution system (MES)

I. INTRODUCTION

Inthe 1960's, enterprises began using computer to manage

information relating to daily business transactions, thus,

saving valuable manpower and increasing the accuracy of

information. According to Nonaka et al. [9], to gain a

competitive advantage, knowledgemustfirst begrasped. Inan

era of knowledge economy, knowledge will replace the

traditional factors of production and become the most

important ofresources. To achieve success, enterprises must

constantly create new forms ofknowledge, infuse new types

of knowledge into existing organizational systems, rapidly

absorb new types of technology and replace outdated

technology withnewproducts.As moreenterprises implement

MES, they accumulate massiveamountsof data. This research

will explore some fundamental problems to be faced by

enterprises inthe future. Thegoal of this research is tomake

MES wiser. Through the integration of some major subject

areas(MES, data warehouse, onlineanalytical processing, data

mining system) the massive amounts of data assets

accumulated

by

MES may be used to obtain information

valuabletoenterprises inmaking important decisions. Aspart

of this research, the framework for a data

mining

systemhas

been designed. This framework allows for the

integration

of

MES and data mining tasks. Moreover, in

accordance,

an

appropriate informationsystem is

practically planned

outand

established.

II. LITERATURE REVIEW

A.IntelligentManagement DecisionSystem

Intelligent management decision systems are the

combination of information

technology

and artificial

intelligence. Such systems allow organizational networks to

work toward the computerization of management systems,

which are made up of particular regulations and tasks.

Moreover, by putting to use such computerized systems,

assisting systems may carry outdatamanagement, information

analysis and anticipatory system monitoring in a convenient

manner. Consequently, by performing all types ofresponses

and management and decision-making, the computerized

systems may carry outtheirassigned tasks.

B. DataWarehouse

Inmon [11] believes that adata warehouse is an integrated, subject oriented, time variant and nonvolatile collection of data

in supportofmanagement'sdecisions.

Berson etal. [4]believes that datamartisadatastorethat is

a subsidiary ofa data warehouse ofintegrated data. The data

mart is directedat apartition of data that is created for theuse

ofadedicatedgroup ofusers. Adatamartmight, infact, bea

set of denormalized, summarized, or aggregated data. Online

analytical processing gathers together, sorts through and

analyzes the data stored in data warehouses, creating

substantial data. Various modes of data arepresented forusers

to access. This allows users to view data using various

perspectives and subjects. Through complex search processes

and the comparison of data, data reportingmayfacilitatemany

different levels ofanalysis.

C. DataMining

Groth [10] points out that data mining is the process of

finding trends and patterns in data. The objective of this

process isto sortthrough large quantities of data and discover

new information. The benefit of data mining is to turn this

newfoundknowledge into actionable results, suchasincreasing

a customer's likelihood to buy, or decreasing the number of

fraudulent claims. Berry and Linoff[7,8] point out that data

mining is the exploration and analysis, by automatic or

semiautomatic means, of

large quantities

of data in order to

discovermeaningfulpatternsand rules.

Chen [12]pointsoutthat the workprocess of datamining is

composed

of

eight primary

tasks. The

eight primary

tasks are

task relevant

data, background knowledge, problem

statement,

kinds of knowledge to be mined, data mining algorithm,

modelsorknowledgepatternsmined, interestingness, anduser.

The goal of data mining is to extract valuable and new

information from existing data. Datamining is the process of

discovering interesting

patterns

in databases that are useful in

the following major functions: classification, clustering, estimation, prediction, affinity grouping, description, etc.

III. MESANDPROBLEMDEFINITION A. The MES

A semiconductor manufacturing entity includes integration of the processing equipment with all of the supportingsystems

forproduct andprocessspecification, production planning and

scheduling, and material handling and tracking. IC fabrication

isa verycomplex andcapital-intensive manufacturingprocess.

Presently, high productivity and a quick response to

customers are essential for most producers. Reducing the

processing time and inventories are objectives in managing

production systems[3]. Customerstodayare moredemanding:

they want to buy high quality, low cost, high performance products configured with just the required features. Semiconductor companies arebeing forcedtomovefromhigh

volumeproduction of commodityparts to lowvolume, flexible and leaner production of application-specific parts. The IC manufacturers' goals are toreduce costs, cutproduction time,

improve quality, increase asset utilization and guarantee

on-time delivery. An available tool to assist semiconductor manufacturers in achieving theseobjectives isaManufacturing

Execution System. Advanced Manufacturing Research (AMR) organization has defined MES as "an integrated architecture

for plant wide information management that groups

applications and functions around a central common database

usedtoshareproduct andprocessdataamongtheapplications" [14].

The MESA International definition of MES is: "MES delivers information that enables the optimization of production activities from order launch to finished goods. Using current and accurate data, MES guides, initiates, responds to, and reports onplant activities as theyoccur. The

resulting rapidresponsetochanging conditions, coupled witha

focus onreducing non value-added activities, drives effective

plant operations and processes. MES improves the return on

operationalassets aswellas on-time delivery, inventoryturns,

gross margin, and cash flow performance. MES provides

mission-critical information about production activities across

the enterprise and supply chain via bi-directional communications." The functional model of MES is illustrated inFigure 1. [13]

Figure 1. The Functional Model of MES

B. ProblemsofMES

The advancement of automation technologies has eased the operations of manufacturing and material handling. For example, an automated manufacturing cell may consist a

computer numerical control for part processing, an articulate

robot for loading and unloading, two conveyors for

transporting arrival and departureparts, and severalsensorsfor

signal capture. Those automated equipments are then

harmonized by a centralized controller, usually a

programmable logic controller.

Unlike traditional production activity control or shop floor

control systemswhichtryto stabilize the production activities

on the shop floor and focus largely onthe scheduling and/or

dispatching functions, MES focuses mainlyonmonitoring and

summarizing the status ofoperational systems (i.e., the status

ofproduct quality, equipment, materials, andmanpower onthe

shop floor). A MES is defined as an on-line integrated

computerized system that is the accumulation of the methods and tools usedtoaccomplish production [15].

However a single MES is not enough. The manufacturing

execution system should be conducted to integrate physical operational systems, so a fully automated and integrated

manufacturing management environment can be developed.

And it isnecessaryto develop intelligent MESto integrate the

information of distributedoperational systemsand discover the newfound knowledge. MES is unable to carry out acomplete

process ofanalysis, allowing massive amounts of data stored

within the system to become more meaningful. Moreover,

MES is unableto satisfy the needs of high-level management

personnel in policy-making decisions,to speedupthe creation

ofnewstrategy.

IV. DESIGNANINTELLIGENT MES

A. The Overall Description of Proposed Framework for IntelligentMES

Based upon the data mining system, the design primarily

focuses uponthe following problems: the integration of MES and data warehouses, the integration of data warehouses and decision-making analysis, the integration of decision-making analysis and data mining systems, and the set up of a data mining engine.

B. The Frameworkfor IntegratingMES and data warehouses

With advancements in data warehouse technology, six distinct models have been developed in the establishment of data warehouses. This research utilizes thetoptobottom model

to design a data warehouse system. The database of MES

serves as the origin of data, by extracting and transforming

data, an integral and unified data warehouse system may be designed. Data marts and data warehouses have an one sided

relationship, in which data fromadata warehouse flows into a

datamart.

When integrating a MES and a data warehouse, it is

commonto encountertheproblem ofinconsistent, incomplete

andduplicate data. Therefore,theintegrationof MES and data warehouses involves the collection of different types of data

from theiroriginalsources.This data isplacedinadatastaging

area where it undergoes such processes as the cleaning,

pruning, combination and removal of duplicates. Next,the data is stored withina presentation server. Atthis point, users can

carryoutsearch tasks.

Even ifa data warehouse with all the potentially relevant

data is available, it will often benecessary to pre-process the

data before they can be analyzed [1]. The procedures for

integrating MES and data warehouses include eight distinct

steps.Thesestepsareexplained below:

(I)Collection: After gathering primary data, needed data is copied intoadatastagingareafor furtherprocessing.

(2)Transformation: This includes revising of data accuracy

and the removal fromstorageof unneeded data.

(3)Loadinganindexing: Transformed data is saved inadata

martand indexed.

(4)Quality control check: Assuring the quality of data. (5)Announcement orpublication: The preparatorywork for

the official on-line installation of thesystem.

(6)Renovation: The continuous revision ofout of date and inaccurate data.

(7)Search: Data search servicesareprovided.

(8)Checks and preparations: Assuring the safety of the data warehousetoavoidpotential damages.

C. The Framework for Integrating Data Warehouses and

Decision-Making Analysis

There are somesteps tobuild the framework for integrating

data warehouses anddecision-making analysis.

(1) Designingaschema for data warehouse: Whendesigninga

data warehouse system, the following schemas can be

utilized: star schema, snowflake schema, and fact constellation schema. This research utilizes the starschema indesigning the schema for data warehouse. This schema is based uponthe manufacturing fact table, a time dimension

table, an area dimension table, a product dimension table,

Figure 3. Designingastarschema for data warehouse

(2) Designingafact table: The real data thatweneed isplaced

inthe fact table. The data in this tablecannotbealtered;we may only add new information. Moreover, this table

includes an indexkey related to a dimension table. When

designing a fact table, several factors must be taken into

consideration.

(3)Designing the dimension table: Data of dimension table is used as a reference to fact table data. When necessary, complex descriptions can be divided into several small

parts.

(4)Designingamultidimensional data model: When analyzing

data, multiple dimensions arebrought togetheras onepoint

of consideration. Thisprocess is calledamultidimensional

data model [6]. Data warehouse systemsmayincludemany

data cubes. Each data cubemaybe theproduct of different

dimension and fact tables. The OLAP operations in data cubes include rollup, drilldown, slice, dice, and pivot. A data cube may be an N-dimensional data model [16]. In

ordertoprovidean evenwiderrangeof searchcapabilities,

this research uses the four dimensions of time, area,

product and quality to construct a four-dimensional data

cube model.

D. The Framework for Integrating Decision-making Analysis

and DataMining System

After completing the construction of a data cube, it is

possible to integrate decision-making analysis and the data mining system (as shown in Figure 4) [12]. The goals of integration are to allow OLAP analysis results to supply the

knowledge base within the data miningsystem,thusproviding analysis information tothe datamining system and creatinga

point of reference for data mining tasks. OLAP technology is able to blend together people's observations and intelligence within the datamining system, thus improving the speed and depthatwhich data is excavated.Furthermore, the intelligence discoveredby the data miningsystem acts as aguide in OLAP

analysis tasks, increasing the depth of analysis. As a result,

information left unearthedby the OLAP, is extremely complex and delicate innature.

Figure 4. The framework for integrating decision-making analysis and data miningsystem

E. DesigningaDataMining System

Datamining technology can be divided between traditional

and refined technologies. Statistical analysis is the representative characteristic of traditional technology. As for refined data mining technologies, all types of artificial intelligence areputtouse.Morecommonly usedtypesinclude:

decision tree, neural network, genetic algorithm, fuzzy logic, rules induction, etc. Decision tree induction is one of the

widely used tools for data mining [5]. This paper's algorithm used decisiontreemethod for data classification andprediction. Thealgorithm actsasthe nucleus of the datamining engine in

classifying data hidden within the database and in anticipating information. The operational process and knowledge rules

excavated by data mining can be displayed through visual

InterfaIeforusere±nqu~3~es

T

I

T

I

|Onl1in ec;lt mlninS3 =:O A- eng,2ine

OfE-infDdata3 minirig|

.K

MtES syisret andl therItimbiniai S;ystem)

figure. This allows users to operate data mining more easily

andtounderstand the meaning represented by excavated data.

V. PRACTICAL IMPLEMENTATIONANDOPERATIONOF INTELLIGENT MES

A. The Information Infrastructure ofthe Intelligent MES

With the systematic information infrastructure provided by

the worldwide web, users may use the Internet to interact in

more convenient manner. This three-tier information infrastructure of the intelligent MES is illustrated in Figure 5.

Webserver

Clie:t

T

Figure 5. The information infrastructure ofthe intelligent MES

B. Operational Procedures ofthe Intelligent MES

The operational procedures ofthe intelligent MESareshown

inFigure 6. The detailed explanations of the procedures are as

follows:

(1)Userentry system: Users operaterelated procedures of the dataminingsystemthrough the IntranetorInternet.

(2)Loading the data of the MES into the data warehouse

system: Based on the requirements of the manufacturing

subject of the data warehouse, the manufacturing data tables, quality data tables, product data tables of the MES will be considered data source of the entries. Afterwards,

through the clearing, collation, and transformation of the data, theyarethen entered into the data warehousesystem.

(3)Establish schema of data warehouse: The data warehouse

system is built according to the manufacturing subject, and thestarschema is established. Themanufacturing fact table is at its center, with the related quality dimension table and theproduct dimension table.

(4)Establish OLAP decision-making analysis: According to

themanufacturing fact table, quality dimension table, and product dimension table, the operation of multidimensional data cube is simulated using ROLAP method.

(5)Select source of data and attribute for data mining: The

data in the data warehouse system and the results of OLAPoperations can be sources of data for datamining.

The mission model editor of the datamining engine can

be usedtohelpuserselect thesourceof data and attribute.

(6)Select algorithm and functions of data mining: The data mining algorithm provided by the calculation database include decision tree, neural network, genetic algorithm,

and market basket analysis; data mining functions include classification, clustering, prediction, and affinity grouping. This system uses decision tree as the algorithm, and it usesclassification and predictionasthe function.

(7)Executing data mining system: Mission processor is the core of the data mining. It uses a target-oriented

processing system to execute data mining and acquire the wanted results.

(8)Interpretation, evaluation, and exhibition of results: The results acquired from data mining system are usually someabstract data. Consequently, thesystemusesthe rule

based knowledge presentation method of the expert systemandcomplements it with aweb-based framework

to express and interpret data mining results to help user

understand the resultsgained.

Figure 6. Operational procedures of the intelligent MES.

C. PracticalOperation ofthe Data Mining

High quality data from the electronics industry database is usedby this researchtoanalyze categories of quality factors in themanufacturingprocess.

Table 1.Training data from the quality database.

Nbo Heat-Up CbblDbvm TrarimssibhSpeed Dr*hgfirne Qirnhty Jnfomnitibn

I(CCec CC/s ntm sec 1 >3 -3 70 80 NG 2 A3 3 70 s0 OK 3 >3 >-3 90 100 OK 4 2 3 90 t00 NG 5 < >-3 90 120 HG 6 >3 >-3 70 100 OK 7 23 -3 70 s OK 8 < >-3 9 80 OK 9 123 ;-3 90 120 OK I0 2-3 >-3 770 1 20 OK I <2 >3 90 120 NG 1 2 <2 >-3 70 100 OK 1 3 <2 >-3 90 NG 1 4 <2 -3 7 0 80 OK 15 <2 -3 70 100 OK 16 23 >3 70 80 HG 17 2-3 <*3 90 OK 18 >3 3 70 80 NG 19 I <2 >-3 70 so OK 20 >3 3 90 100 NG 21 >3 -3 9 100t NG 1000 >3

-3

90 120 OK Iv1E server

In thisresearch, werandomly select 1000records fromMES in 2004asthetraining data.Preparepreviously classified

trainingdata,asshowninTable 1.

After calculated by decision tree algorithm, training data

will bepartitionedinaccordance with the selectedtestattribute.

Thecompleted decisiontreeis illustratedinFigure7.

in~~ ~ ~ ~ oldw te .0 t(yine ImIds

3Cse 70cm/i 8s OK A

>

X -3°C/se 9c mXi 120-e OK

3C Xl1 NG-3°C/sc 9 nOKlOOsec/s e 7C-i.120-OK3

N

/deA -3C/sec 70=/,mi lOOsec OK

Fligu 7 The desio treeiorfdor-_draiingdddas

towhe down time- -sed tme tsped hm

>-3C 120sec NG >-3'C 80se OK 70rCm.i. 8so-c NG100Xmi lOse OK C--3C120Osc NG --3Cl100sec OK 9 mX n IO-e NG 70c..in 80-se OK

>-3' OsecM NG >-3C lOOsec OK 70Cm/mi. 120-e NG 70CmGm.i. 100sec OK

<--3°C lOOs-c IN-G Node ED Noledi r Ne lde E'

No&l A

Figure.

7Thedecisiontreefor

training

data

D.AnalysisofDataMiningPractical Operation

The decisiontreemakesknowledge ruleseasy tounderstand,

thus, the decision tree, selected as

algorithm,

serves as the

nucleus of the data mining. Any specific information, which

has already beenclassified, canbe easily tracedalong the

path

from the rootnodetothe leaf node by using the decisiontree

model. Inthisway, knowledge rules for classificationmaybe

established. When the goal of data

mining

is to

classify

or

anticipate the results of data and create easy to

comprehend

rules, the decision tree algorithm is best suited to act as the

nucleus of the datamining [8].

The knowledge rules possessed by the decision tree

described above allow for the convenient

gathering

of

information, by tracing this information

along

the

path

from

root node to leaf node. In this research,

knowledge

rules are

expressed through the rule based

knowledge presentation

method of the expert system, thus,

revealing

the

knowledge

rulespossessedby the decisiontree.

Whether or not a decision tree can

successfully

classify

a

new setof datacanbe evaluatedthrough the data

mining

error

rate. Test datacanbe usedto evaluate the leaf nodeerror rate.

The evaluation of decision tree errorrate can be divided into

the

following

twosteps[7]:

(1)Evaluate

theerror rate of each node: Calculation formula is " the leaf node error rate= the amount of

incorrectly

classified data within the leafnode/ the total amount of

data in the leaf node". The leaf node error rate is as

follows:A=0.2, B=0,C=0.2, D=0,E=0

(2) Evaluate the totalerror rateof the decisiontree: The total

error rateofthe decisiontreeis the

weighted

totalamount,

calculated as shown below: The total error rate of the

decisiontree=0.091

E.BenefitAnalysis

This research uses the data mining system, with the vast

amountof

quality

data from the

manufacturing

processas

basis,

to

produce

accurate

quality

information and

knowledge

rules.

Moreover, based on the acquired knowledge rules, is able to

understand the characteristics and attributes of the important

causesofquality,allowing ittostay ontop of thequality issue.

By strengthening the training of quality assurance personnel,

adding extraheaters for the increase oftemperature, installing

adjustable fans for the decrease oftemperature,andcontrolling

thespeed ofconveyors,these methodsimprove product quality,

lowermanufacturing costs, andturn raise the overall business

performance. The benefits analysis of this improvement is as

follows:

(l)Using the intelligent MES to dig up similar category

quality information. Consequently, the average improvement

rate onthe productyield is5.87%., asshowninFigure8.

hIoAr.lt Jar Feb Iri Ap. DiAy Jr Jul Aug Sep |Olct TNov D ec- Averge U-np-ovd 92.8 94.3 94.7 93.6 93.1 94.3 92.1 95.2 93.1 |93.6 |94.5 |94.5 |93.8'2

,lc dae° }l

Imp-oed 99.5 98.9 99.2 99.1 99. 1 99.8 99.2 99.7 99.1 |99.3 |99.4 |99.6 |99.33 yieldratt }l

hmp-org 7.22 4.88 4.75 5.88s 6.44 5.83 7.71 4.73 6.44 6.09 |5.19 |5.40 |5. S7

| I)lagl: en:ll s}~~~f~itlcrease1-ateof-pr:odut,eyietll

105l

| -r Feb hIa -Apr hIs J-r Ju1 Au-g Sep Oct 1oo De-c)

Figure. 8.Diagram of increasedrateofproduct yield

(2)UsingintelligentMES toreduce therateofmanufacturing

cycle timeto23.49%,asshownin Figure9.

-hh Jan T Feb T - Ap rT. J1T h4- Ju T A g Sep T 1 T D TA-g Uri-mp-o d r,6 7 2 6 8 9 5 8 66 7 6 8 7 7 2 7 5 l6 3

|I-pmpong at%}|2500|2500 |33822| 26,09|1034 |2727|2388 27 94 |2254|1364 |2500 2213 3 |2349

| ~~~~~ID)ia.W-amlofloxve-ledmlamffitchutiriigc-ycle tinlle

I S~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~4 -T.TF. p,Aly A, -T,l ,,g Sly 0t,IT.v),

| NI~~~~~~onth

Figure9.Diagram of loweredmanufacturing cycle time.

VI. CONCLUSION

Through the actual establishment of data

mining

and

intelligent MES, and the use of these systems on the

manufacturingprocesses of the electronics industry for

quality

improvement, theexperience has verified the

following

results:

* The basic framework ofestablishing

intelligent

MESwill

allowcorporate

computerized

applications,

which willnot

be limitedto the level of data

processing,

but it can also

aggressively

work towards an

information-based,

knowledge-based, and

intelligent

form of management

informationsystem.

*

Using

the

intelligent

MES to

dig

up similar category

quality information.

Consequently,

the average

improvementrate ontheproduct yield is5.87%.

*

Using

intelligent

MEStoreduce therateof

manufacturing

cycle timeto23.49%.

[1] A. Feelders, H. Daniels & M. Holsheimer, "Methodological and practical aspectsof data mining",Information & Management, vol. 37 (5), 2000, pp. 271-281.

[2]Indranil Bose & Radha K. Mahapatra, "Business data mining- a machine learning perspective", Information & Management, vol. 39 (3), 2001, pp. 211-225.

[3] Ruey-Shun Chen & Kun-Yung Lu, "A case study in the design of BTO/CTO shop floor control system", Information & Management, vol. 41(1), 2003, pp. 25-37.

[4] Alex Berson, Stephen Smith, Kurt Thearling, "Building data mining

applicationsfor CRM",McGraw-Hill,2000.

[5] Yonghong Peng, "Intelligent condition monitoring using fuzzy inductive learning", Journal of Intelligent Manufacturing, vol. 15(3), 2004, pp. 373-380.

[6] N. Gorla, "Features to consider in a data warehousing system", CommunicationsofThe ACM, vol.46(11),2003, pp. 11 1-1 15.

[7] Michael JA.Berry, Gordon S.Linoff, "Data mining techniques: for marketing, sales, and customer support", John Wiley & Sons, 1997. [8] Michael J.A.Berry, Gordon S.Linoff,"Mastering data mining: the art &

scienceof customerrelationshipmanagement", JohnWiley&Sons, 1999. [9] Peter F. Drucker, Ikujiro Nonaka, David A. Garvin, "Harvard business

review onknowledge management", Harvard Business School Press, 1998. [10] Robert Groth," Datamining: building competitive advantage",

Prentice-Hall, Inc, 2000.

[12] Zhengxin Chen, "Data mining and uncertain reasoning: an integrated

approach",JohnWiley&Sons, 2001.

[11] W. H. Inmon,"Building the data warehouse", John Wiley & Son, 2002.

[13]MESAInternational, "MESA International White Paper Number6:MES Explained:AHigh Level Vision" (Pittsburgh, PA: MESA International), 1997.

[14]Samanich, N. J., "Understand your requirements before choosing an MES",Manufacturing Systems, January, 1993.

[15] McClellan, M., "Applying Manufacturing Execution System", St. Lucie Press,Florida, 1997.

[16] StephenR. Gardner, "Buildingthe datawarehouse", Communicationsof