The Manufacturing Knowledge Repository

Consolidating Knowledge to Enable Holistic

Process Knowledge Management in Manufacturing

Christoph Gröger, Holger Schwarz and Bernhard Mitschang

Institute of Parallel and Distributed Systems, University of Stuttgart, Universitätsstr. 38, 70569 Stuttgart, Germany

Keywords: Knowledge Integration, Manufacturing, Process Optimization.

Abstract: The manufacturing industry is faced with strong competition making the companies’ knowledge resources

and their systematic management a critical success factor. Yet, existing concepts for the management of

process knowledge in manufacturing are characterized by major shortcomings. Particularly, they are either

exclusively based on structured knowledge, e. g., formal rules, or on unstructured knowledge, such as doc-

uments, and they focus on isolated aspects of manufacturing processes. To address these issues, we present

the Manufacturing Knowledge Repository, a holistic repository that consolidates structured and unstruc-

tured process knowledge to facilitate knowledge management and process optimization in manufacturing.

First, we define requirements, especially the types of knowledge to be handled, e. g., data mining models

and text documents. On this basis, we develop a conceptual repository data model associating knowledge

items and process components such as machines and process steps. Furthermore, we discuss implementation

issues including storage architecture variants and finally present both an evaluation of the data model and a

proof of concept based on a prototypical implementation in a case example.

1 INTRODUCTION

Today, manufacturing companies are exposed to

intense competition due to globalization, high mar-

ket volatility and rapid technological changes (Mon-

auni and Foschiani, 2013). In addition, worldwide

homogenization and dissemination of production

technologies and materials diminish the competitive

potential of tangible assets. Thus, knowledge, that is

the intangible intellectual capital of a company,

becomes a critical source for competitive advantages

emphasizing the need for a systematic knowledge

management (Goossenaerts et al., 2005).

Existing knowledge management systems in

manufacturing mainly focus on product knowledge

and customer knowledge. For example, knowledge-

based engineering systems integrate computer aided

design (CAD) data and additional product

knowledge to enrich product models (Chapman and

Pinfold, 2001). Yet, there are only rudimentary con-

cepts for the management of process knowledge in

manufacturing.

Existing approaches are characterized by three

major shortcomings limiting process knowledge

management and continuous process improvement:

(i) they are either exclusively based on structured

knowledge, e. g., formal rules, or they only deal with

unstructured knowledge like documents; (ii) they

make use of tailored and application-specific data-

bases to store knowledge items; (iii) they focus on

isolated aspects of manufacturing processes, e. g.,

specific resources, or selected phases of the process

lifecycle, e. g., process planning. This leads to an

ineffective, costly and time consuming discovery,

application and sharing of manufacturing knowledge

(Economist Intelligence Unit, 2007). For example,

production supervisors typically have to access dif-

ferent isolated IT systems and paper-based docu-

ments to find failure reports and improvement sug-

gestions in order to manually correlate them with

additional process information like metrics.

To address these issues, we present the Manufac-

turing Knowledge Repository (MKR), a universal

holistic repository that consolidates structured and

unstructured process knowledge to facilitate

knowledge discovery, knowledge management and

knowledge-based process optimization in manufac-

turing (see Figure 1).

The remainder of this article is organized as fol-

lows: First, we structure related work with respect to

Gröger C., Schwarz H. and Mitschang B..

The Manufacturing Knowledge Repository - Consolidating Knowledge to Enable Holistic Process Knowledge Management in Manufacturing.

DOI: 10.5220/0004891200390051

In Proceedings of the 16th International Conference on Enterprise Information Systems (ICEIS-2014), pages 39-51

ISBN: 978-989-758-027-7

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: The Manufacturing Knowledge Repository.

process knowledge repositories in Section 2. Next,

we define contents and requirements for the MKR in

Section 3. This provides the basis for the conceptual

repository data model presented in Section 4. In

Section 5, we focus on implementation issues and

present a prototypical implementation. A qualitative

evaluation of the MKR and a technical proof of con-

cept based on a case example are described in Sec-

tion 6. Finally, we conclude in Section 7 and high-

light future work.

2 RELATED WORK: PROCESS

KNOWLEDGE REPOSITORIES

Process Knowledge repositories are databases for

integrating, structuring and storing process

knowledge (Davenport and Prusak, 2000). The latter

comprises all types of insights related to processes.

In this work, we focus on explicit knowledge in the

sense of contextualized data connected by patterns

and relations (Ackoff, 1989). We further subdivide

explicit knowledge in structured knowledge having a

predefined technical data structure, e. g., formal

rules and metrics, and unstructured knowledge, e. g.,

photos and documents.

With respect to related work, we distinguish be-

tween manufacturing-specific repositories for pro-

cess knowledge and concepts from the business

process and workflow context. Manufacturing-

specific approaches can be found as part of various

expert systems for process planning (Kiritsis, 1995),

(Giovannini et al., 2012). They make use of formal

rules and logics to support the generation of work

plans. These kinds of repositories are typically based

on structured knowledge and focus on process plan-

ning aspects. Besides process planning, there are

only rudimentary repository approaches focusing on

the other lifecycle phases, that is, process execution

and process analysis. The tools presented in (Fischer

et al., 2000) share a common knowledge repository

for process analysis in manufacturing. It integrates

structured knowledge for rule-based, case-based and

model-based reasoning to identify root causes of

production failures. In (Mazumdar et al., 2012), a

manufacturing knowledge repository is presented. It

integrates and annotates process-related documents,

e. g., failure and performance reports, using manu-

facturing-specific ontologies to support semantic

search capabilities for process execution and analy-

sis. All these approaches make use of application-

specific databases and are either exclusively based

on structured or on unstructured knowledge.

Regarding process knowledge repositories in the

business process context, process repositories with

semantic search capabilities, e. g., (Ma et al., 2007),

can be seen as initial approaches. Most similar to the

concept presented in this article is the work in (Nie-

dermann et al., 2011). The authors present a univer-

sal process knowledge repository that stores results

of workflow analyses, especially metrics and data

mining models. Yet, it focuses on structured

knowledge and cannot simply be applied to manu-

facturing as it is based on workflow standards, espe-

cially the Business Process Execution Language.

The MKR goes significantly beyond existing ap-

proaches by integrating various types of structured

and unstructured process knowledge in a universal

database to support different analytics- and

knowledge-driven applications across the entire

process lifecycle in manufacturing.

3 REPOSITORY CONTENTS

AND REQUIREMENTS

The MKR integrates different kinds of process

knowledge, called insights, by associating them with

corresponding process components. Hence, the two

core building blocks of the MKR’s content are a

holistic process meta model as well as a catalogue of

different types of insights. The main requirements

for these building blocks are described in the follow-

ing and are used as a basis for the definition of the

data model in Section 4.

The holistic process meta model defines essential

components of discrete manufacturing processes,

e. g., process steps and resources, whereas it is inde-

pendent of a concrete industry in order to be univer-

sally applicable. It has to integrate both design-time

and a run-time perspective, that is, aspects of pro-

cess planning and execution, to provide a holistic

view. The design-time perspective comprises the

process model defining, e. g., the types of resources

needed, whereas the run-time perspective covers all

Process

Dashboard

Manufacturing Knowledge Repository

Failure

Report

18min

34%

46kg

Pic

Manufacturing

Processes

Process

Mining App

…

Knowledge Discovery and Management

Process

Optimization

Knowledge

Discovery

Knowledge Consolidation

ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems

aspects of the execution of the model, e. g., individ-

ual employed resources or occurred failures. There-

by, both a process view referring to the flow of pro-

cess steps including the routing of materials as well

a resource view referring to the detailed specifica-

tion and dependencies of resources like machines

have to be combined. Moreover, changes of process

models, that is, their evolution over time, have to be

traceable in order to support process optimization

purposes (Niedermann et al., 2011). It is important

to remark, that there is no need for a highly detailed

meta model like in computer aided planning sys-

tems. Instead, the meta model has to cover all major

components of manufacturing processes to associate

corresponding insights while remaining easy to un-

derstand for non-expert users in IT like production

supervisors. Finally, it has to be able to be imple-

mented in a database environment in order to use it

for repository storage.

Regarding insights, a huge variety of knowledge-

relevant objects exists in manufacturing ranging

from work instructions over failure reports to key

performance indicators. Thus, we analyzed insights

across the entire process life cycle from a technical

point of view differentiating structured and unstruc-

tured insights. We observed the following types of

structured insights:

 Metrics, e. g., lead time, aggregating quantitative

process attributes (Brown, 1996)

 Data mining models, e. g., decision trees or cluster

models, representing patterns and relationships of

process attributes (Han et al., 2012)

 Formal rules in terms of if-then relations, which

can be used for rule-based reasoning (Giarratano

and Riley, 2005) or as business rules (Morgan,

2002)

 Special process constructs, e. g., rework sequenc-

es, which refer to sets of process steps with certain

business semantics (Niedermann et al., 2011)

 Ontology concepts in terms of semantic annota-

tions using manufacturing-specific ontologies like

MASON (Lemaignan et al., 2006) to enable rea-

soning and semantic search capabilities

In addition, we identified the following types of

unstructured insights:

 Text referring to any kind of unstructured textual

data, e. g., emails or reports

 Images like photos, graphics or diagrams

 Aud io comprising any kind of sound recordings

 Videos

4 CONCEPTUAL REPOSITORY

DATA MODEL

The conceptual repository data model realizes the

contents and requirements discussed in Section 3

and comprises a holistic process meta model as well

as an insight model. In the following, we represent

both parts as class diagrams in the Unified Modeling

Language (UML) and describe their association.

4.1 Holistic Process Meta Model

The basis of the holistic process meta model is the

basic meta model described in (Gröger et al.,

2012a). The latter comprises a manufacturing pro-

cess meta model which takes a run-time perspective

on manufacturing processes and is designed for the

implementation in a data warehouse environment.

We refine and extend this meta model with respect

to design-time aspects in order to derive the holistic

process meta model. To this end, we analyze exist-

ing process-oriented manufacturing meta models,

especially (Erlach, 2011), (Zor et al., 2011), (Inter-

national Society of Automation, 2000), (Lemaignan

et al., 2006). Figure 2 shows the main components

of the resulting process meta model, which we de-

scribe in the following. For the sake of simplicity,

we omit many additional classes of the model, e. g.,

for spatial aspects of process steps, and do not detail

attributes.

4.1.1 Design-time Aspects

From a design-time point of view, that is, with re-

spect to process planning and design, a manufactur-

ing process in terms of a process model produces

one or more types of products. A product can be

described by features referring to informational

aggregations of product characteristics, like geomet-

ric or functional aspects (Shah and Mäntylä, 1995).

Features relevant for a certain process step are asso-

ciated with the latter to enable both feature-oriented

analysis across different manufacturing processes as

well as the association of feature-oriented insights,

especially rules for knowledge-based process plan-

ning.

A manufacturing process comprises several pro-

cess steps, that is, all steps necessary to produce the

specified product. In order to analyze the evolution

of a manufacturing process over time, different pro-

cess versions can be defined, which comprise indi-

vidual compositions of process steps. According to

(Erlach, 2011), (Zor et al., 2011), we differentiate

manufacturing steps, comprising the actual

TheManufacturingKnowledgeRepository-ConsolidatingKnowledgetoEnableHolisticProcessKnowledgeManagement

inManufacturing

Figure 2: Main components of the holistic process meta model.

manufacturing and assembly of parts, testing steps,

which refer to quality control activities in a process,

transporting steps, covering the movement of parts

between different steps, and warehousing steps,

referring to stock-keeping. Process steps are further

associated with three types of resource groups,

namely operating resource groups comprising ma-

chine groups and productions aid groups, as well as

employee groups. These groups define requirements

for the actual resources selected during process exe-

cution and control, e. g., specific machines, tools and

workers. Input material refers to products and parts

as external input of process steps, e. g., for assembly

operations. It defines necessary material properties

and amounts as described in the work plan.

Regarding the process flow, that is, the connec-

tion of process steps and the modeling of different

paths, we exclusively focus on the flow of material

as done in value stream design (Erlach, 2011). Thus,

we omit additional control flow aspects for the sake

of understandability. Moreover, we model the flow

of material using material gateways and refine the

concept in (Zor et al., 2011) as follows: Two process

steps are always connected by a material gateway.

The first and the last step of a process have no input

gateway or no output gateway respectively. Moreo-

ver, we differentiate five types of gateways: The

sequence gateway defines a simple sequential pass-

ing of material from one process step to the other.

The route gateway represents a diversion point in

the material flow, i. e., one out of several possible

subsequent process steps has to be chosen according

to a defined condition. As a counterpart, the select

gateway refers to a selection of one out of several

preceding process steps. The split gateway creates

parallel flows of material with a condition defining

how the material is split up. The join gateway again

joins parallel material flows.

4.1.2 Run-time Aspects

The run-time perspective focuses on the execution of

single instances of a manufacturing process which

are initiated by a production order. The latter de-

fines the customer as well as various order details

like batch size. Instantiation refers to process execu-

tion and control and comprises the detailed planning

of resources and materials. That is, individual ma-

chines, production aids and employees are selected

for process execution and are therefore associated to

a process step instance which in turn belongs to a

manufacturing process instance. Moreover, material

consumption associates the actual batch of input

material processed in a step instance.

In addition, there are elements which are not

modeled at design-time, especially failures, which

may occur during process execution, and the con-

sumption of operational material. The latter refers

Employee

Operating Resource Group

Machine Group Machine

Operating Resource

Process Version

Manufacturing Process Instance

Production Order

ProductFeature

Production AidProduction Aid Group

Transporting

Manufacturing

Employee Group

Manufacturing Process

Material Consumption

Batch

Input Material

Warehousing

Testing

Material Gateway Material Gateway Instance

Route Gateway

Split Gateway

Select Gateway

Join Gateway

OpMaterial Consumption

Sequence Gateway

Run-Time

Aspects

Design-Time

Aspects

1..*

Input

Output

0..1

1..*

0..1

Input

0..1

Output

Operational Material

Process Step Instance

Failure

Process Flow

RessourceMaterial

Product

Process Step

Exception

ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems

Figure 3: Main components of the insight model.

to external input material which is consumed in a

process step but does not become part of the product

itself, e. g., oil or electricity.

4.2 Insight Model

Figure 3 shows the main components of the reposi-

tory’s insight model. In the following, we focus on

metrics, data mining models and unstructured in-

sights as major types of insights in manufacturing.

In general, an insight is associated with a creator

referring to the employee who created the insight.

This enables the integration of the MKR with exist-

ing yellow page systems for community-based

knowledge management by linking the creator with

its entry in the yellow page system.

Metrics primarily comprise the actual value and

the unit of measurement, e. g., seconds or kilos.

Moreover, they are organized in general target di-

mensions of manufacturing, especially time, quality,

flexibility and cost (Kaushish, 2010). For example,

lead time and adherence to delivery dates are metrics

belonging to the time dimension. The calculation

defines the formula as well as the meaning of the

metric itself. Besides, we differentiate two types of

metrics: Target metrics define values in terms of

thresholds to be achieved during process execution,

e. g., the maximum lead time of a process, whereas

measured metrics comprise the actual recorded val-

ue. Thereby, measured metrics are associated with

one or more target metrics with the latter defining,

e. g., maximum or minimum values.

With respect to data mining models, we differen-

tiate six major types, namely regression models,

classification models, association models, time se-

ries and sequences. For a detailed description, we

refer to (Han et al., 2012). Each model is generated

by a certain algorithm, e. g., a classification tree can

be generated by the C4.5 algorithm, and algorithm-

specific parameters, e. g., whether tree pruning is

activated, are stored as well. Moreover, the input

data that is used as a source for the algorithm is

specified using a predicate filter which is evaluated

over the repository’s data. Further, the repository

allows to store application results of data mining

models, e. g., when a regression model is applied for

predicting a metric. Yet, we assume this to be useful

only in special cases, e. g., for compliance reasons.

Unstructured insights have no predefined compo-

nents and are thus generally descripted by a title and

a textual description.

4.3 Insight Association

In general, insights can be associated with all com-

ponents of the process meta model whereas one

insight can be associated with multiple components

and vice versa. In the following, we detail the asso-

ciation of metrics, data mining models and unstruc-

tured insights with respect to the major meta model

components for a process-oriented browsing of in-

sights, namely processes, process steps and re-

sources (see Table 1). These associations are to be

seen on a conceptual level independent of the im-

plementation, e. g., whether they may be enforced

using application logic or database constraints.

With respect to metrics, target metrics are solely

associated with design-time components like the

version of the manufacturing process and operating

resource groups as they define values to be achieved.

Measured metrics are generated during process exe-

cution, e. g., the actual lead time of a process in-

stance is measured. Thus, they are associated with

corresponding run-time components. However,

measured metrics may be aggregated over several

Regression Clustering

Association

Classification

Insight

Target

Audio

Special construct

Text

Image

Video

Measured

Time seriesSequence

Algorithm

Parameter

Input Data

Calculation

Target Dimension

Result

Unit

Rule

Creator

Value

Title

Description

Ontology Concept

Metric

Data Mining Model

Unstructured Insight

Insight

Detail

TheManufacturingKnowledgeRepository-ConsolidatingKnowledgetoEnableHolisticProcessKnowledgeManagement

inManufacturing

Table 1: Association of insights with process meta model

components.

run-time components representing values on the

design-time level as well, e. g., the average lead time

of a selected manufacturing process.

Data mining models describe patterns and rela-

tionships of a set of run-time elements, e. g., a clus-

tering of process instances of a selected manufactur-

ing process. Thus, they are solely associated with

design-time elements.

Unstructured insights are generally associated

with both design-time and run-time components like

certain machines or entire machine groups. Yet, with

respect to processes and process steps, unstructured

insights are solely associated with the corresponding

design-time components in order to reuse them

across all process instances and step instances.

5 IMPLEMENTATION ISSUES

In this section, we analyze the characteristics of the

data in the MKR and discuss different storage archi-

tectures. Moreover, we present a prototypical im-

plementation of the MKR.

5.1 Data Characteristics

A storage-oriented analysis of the conceptual data

model presented in Section 4 reveals several kinds

of data that have to be stored. In the following, we

characterize these different kinds of data as a basis

for the development of a storage architecture for the

MKR. Thereby, we focus on selected types of in-

sights, namely, metrics, data mining models and

unstructured insights, as major types of insights in

manufacturing process management. Thus, there are

four kinds of data to be stored:

 Data concerning the manufacturing process and

metrics: This comprises data related to all compo-

nents of the process meta model as well as on cor-

responding metrics. Thus, the data are well struc-

tured and can be very large in volume, especially

with respect to process instance data. Moreover,

they have to be efficiently accessed by analytical

applications, in particular data mining tools, in or-

der to generate data mining models.

 Data concerning data mining models: These data

have to allow for a universal representation of data

mining models as well as associated parameters in

order to exchange them with external data mining

tools for model evaluation and application.

 Data concerning unstructured insights: These data

are semi-structured or unstructured and may com-

prise large volumes of multimedia data and text.

The latter should be searchable whereas the former

is primarily stored for manual exchange by the us-

er.

 Data concerning associations: These data are

structured and refer to the association of insights

and components of the process meta model as out-

lined in Section 4.3. These data have to facilitate a

flexible association, even if insights and meta

model components are stored in different systems.

5.2 Storage Architectures

With respect to the above data characteristics, rela-

tional database technology constitutes the starting

point of a storage architecture for the MKR to store

data on processes and metrics in a multidimensional

warehouse structure. This mature technology is suit-

able here because it handles huge amounts of struc-

tured data in a scalable and universally accessible

way.

Regarding data mining models, there are two ma-

jor references for their specification and exchange:

The Predictive Model Markup Language (PMML)

(Data Mining Group, 2013) is an XML-based format

to specify data mining models in a semi-structured

and vendor-independent way. Besides, the Common

Warehouse Meta Model (CWM) (Poole et al., 2003)

and its data mining package define a general meta

model for data mining models. Both approaches

define the structure of the actual mining model, e. g.,

a decision tree, as well as parameters used to gener-

ate it, e. g., pruning settings. Yet, in contrast to the

CWM data mining model, PMML is supported by a

wide range of commercial and open source data

mining tools and thus represents the first choice to

store data mining models in a semi-structured format

in the MKR.

Hence, semi- and unstructured data on unstruc-

tured insights and data mining models have to be

stored and associated with structured data on pro-

cesses and metrics in the MKR. As mentioned, rela-

tional database technology is suitable to store these

data on processes and metrics. Taking this into ac-

count, there are two major storage architecture vari-

ants for the MKR:

x /

○

/ - Insights fully/partially/not associated

Insights /

Meta Model Components

Target

Metrics

Measured

Metrics

Data

Mining

Models

Un-

str uctu re d

Insights

Design-Time Components

xxxx

Run-Time Components

-x-

○

ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems

 In the relational-only architecture, additional fea-

tures of relational database management systems

are exploited to store semi- and unstructured data

together with structured data on processes and

metrics in the relational database. That is, PMML

files are stored as XML data and binary large ob-

jects (BLOB) and character large objects (CLOB)

are used to store unstructured insights. Associa-

tions are directly realized as foreign key relation-

ships between tuples representing insights and

process components in the database. Moreover,

full-text search capabilities of relational database

management systems are employed to make use of

text in unstructured insights like PDF documents.

 In the extended architecture, semi- and unstruc-

tured data are stored separately from the relational

database in a Content Management System (CMS)

(Kampffmeyer, 2007). A CMS allows for the cen-

tral storage and management of content items,

which are accessed by object identifiers. The latter

are used to realize the association of insights and

process components based on mapping tables.

These tables combine primary keys of process

components with object identifiers of correspond-

ing insights.

For a comparison of these architecture variants, we

refer to three major criteria: the handling of insights,

the realization of associations between insights and

process components as well as maintenance issues

(see Table 2).

In view of the handling of insights, the extended

architecture profits from advanced functions of a

CMS. Apart from simple full-text search capabilities

as in the relational-only architecture, a CMS typical-

ly provides text recognition functions as well as

versioning concepts for content items. Moreover, it

allows for a workflow-oriented handling of content

items and thus eases the reuse and sharing of in-

sights in workflow-based processes (Kampffmeyer,

2007).

Regarding the realization of associations be-

tween insights and process components, the relation-

al-only architecture allows for a simple implementa-

tion using foreign key constraints. In contrast, the

extended architecture requires additional efforts to

ensure consistency of associations, e. g., to make

sure that all affected associations are deleted if a

corresponding content item in the CMS is removed.

With respect to maintenance issues, the relation-

al-only architecture reduces maintenance efforts as

existing database procedures, e. g., for backup and

recovery, can be seamlessly applied to data on in-

sights. In contrast, the extended architecture requires

the maintenance of two separate storage systems.

To conclude, we opt for the relational-only archi-

tecture to implement the MKR as it eases the associ-

ation of insights and process components and reduc-

es maintenance efforts.

Table 2: Comparison of architecture variants.

5.3 Prototypical Implementation

Our prototypical implementation is based on the

work of (Vetlugin, 2012) and is carried out as part of

our Advanced Manufacturing Analytics platform for

the data-driven optimization of manufacturing pro-

cesses. The platform comprises data mining use

cases for continuous process improvement (Gröger

et al., 2012b) and makes use of a manufacturing-

specific process warehouse, the Manufacturing

Warehouse (Gröger et al., 2012a).

The technical architecture of our prototype is

based on the relational-only architecture discussed

above and is shown in Figure 4. We implemented a

simplified version of the MKR’s conceptual data

model as a relational schema in an IBM DB2 data-

base. To this end, we extended the schema of the

Manufacturing Warehouse with respect to the holis-

tic process meta model and selected insights. The

schema is oriented towards a relational snowflake

schema to realize a multidimensional structure of the

holistic process meta model with minimum redun-

dancy in the dimension tables. The schema compris-

es process step instances as central facts with met-

rics like lead time of a process step. Process details,

e. g., the process the step belongs to, as well as em-

ployed resources in a step are treated as dimensions.

This structure enables a multidimensional analysis

of process execution data in the MKR, e. g., using

Online Analytical Processing (OLAP).

Moreover, we defined an API implemented in

Java which comprises two services: Navigation ena-

bles both browsing of the MKR’s contents, e. g., to

view insights associated with a certain process step,

and uploading of new contents, e. g., photos. Root

cause analysis is a data mining use case described in

(Gröger et al., 2012b) and focuses on the analysis of

metric deviations using decision trees, e. g., for pro-

duction supervisors identifying influence factors for

high lead times as shown in Figure 4.

Relational-only

Architecture

Extende d

Architecture

Handling of Insights

Realization of Associations

Maintenance

TheManufacturingKnowledgeRepository-ConsolidatingKnowledgetoEnableHolisticProcessKnowledgeManagement

inManufacturing

Figure 4: Technical architecture of the prototype.

Decision trees are stored as insights in the MKR

whereas RapidMiner is used as an open source data

mining tool to derive the decision trees. These API

services can be accessed by applications using the

MKR.

6 EVALUATION

In this section, we provide a qualitative evaluation of

the MKR as well as a technical proof of concept. We

first evaluate the holistic process meta model and the

insight model including MKR’s support for analysis

and insight generation. Next, we show that the MKR

satisfies the whole range of process-oriented infor-

mation needs and thus enables a holistic process

management. Finally, for a technical proof of con-

cept, we employ our prototype of the MKR in an

exemplary case in the automotive industry.

6.1 Evaluation of the Holistic Process

Meta Model

To evaluate the holistic process meta model (see

Section 4.1), we analyze its universal applicability

by showing that it covers all types of discrete manu-

facturing processes. In addition, we compare it with

existing manufacturing meta models and show that it

provides a sound basis for insight association and

universal process representation in the MKR.

6.1.1 Evaluation of Universal Applicability

According to (Buzacott et al., 2013), there are three

general types of processes in discrete manufacturing,

namely mass manufacturing processes, series manu-

facturing processes and one-piece manufacturing

processes (see Figure 5). These types differ in their

scale of production, their organization and their

market orientation. In the following, we briefly de-

scribe these types and analyze how the holistic pro-

cess meta model covers them.

Figure 5: Types of discrete manufacturing processes.

Mass manufacturing processes focus on the produc-

tion of large quantities of highly standardized prod-

ucts. The organization follows a flow shop produc-

tion concept with a high degree of automation. That

is, workers are mainly in charge of controlling ma-

chines which are sequentially connected by auto-

mated transportation steps. Production is decoupled

from demand by a make-to-stock approach. To rep-

resent a flow shop production, the holistic process

meta models allows for the modelling of a flow of

manufacturing steps and transportation steps con-

nected by sequence gateways. Moreover, various

operating resources can be modeled to represent the

high degree of automation. Besides, warehousing

steps can be employed to represent make-to-stock

aspects. Thus, the holistic process meta model sup-

ports the modelling of mass manufacturing process.

One-piece manufacturing processes focus on the

production of single customized products. These

processes are organized according to a job shop

production layout with a high degree of flexibility

and only partial automation. There is a significant

amount of manual work whereas functionally similar

workplaces are grouped together in the factory.

Routing between these workplaces is complex as it

may vary for each product. One-piece manufacturing

follows the make-to-order principle with no signifi-

cant stock keeping. The holistic process meta model

allows for the representation of a job shop produc-

tion layout by modelling various material gateways

to represent flexible routings between manufacturing

steps and transportation steps. Moreover, a produc-

tion order models individual orders of customers

whereas features allow to represent the customiza-

tion of the ordered product. Hence, one-piece manu-

facturing processes can be represented by the holis-

tic process meta model, as well.

Series manufacturing processes focus on the

production of different but related products in prede-

fined lot sizes and represent a hybrid form between

one-piece manufacturing and mass manufacturing.

These processes are based on a combination of flow

shop production and job shop production depending

Process Dashboard

API

RapidMiner

Data Mining

Manufacturing Knowledge Repository

Navigation Root Cause Analysis

Performance Metrics (avg.)

Lead Time: 23min

Reject Rate: 11%

Failure Reports

Root Cause Analysis of Lead Times (LT)

• 18/12/12: Tool 18 broken

• 10/11/12: Machine 2 slow

• 05/11/12: Oil leakage

Utilization: 84%

LT too

high

Step_3_Batch_ID

Step_8_Tool_Age

LT too

high

>= 4

< 4

!= B7 = B7

Mass

Manufacturing

Processes

Series

Manufacturing

Processes

One-Piece

Manufacturing

Processes

• Flow shop production

• High automation

• Make-to-stock

Mixture

• Job shop production

• Partial automation

• Make-to-order

ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems

on the lot size. Thereby, a middle to high degree of

automation and temporary stock keeping are typical.

As stated above, the holistic process meta model

allows to represent both flow shop production and

job shop production as well as a combination using

additional material gateways. Besides, production

orders, products and features can be modelled to

represent series information, e. g., the number of

goods and the product variant to be produced. Thus,

series manufacturing is covered by the holistic meta

model, too.

To sum up, the holistic process meta model sup-

ports the modelling of all general types of discrete

manufacturing processes. Individual manufacturing

processes in industry practice can be seen as derived

or hybrid forms of these types (Buzacott et al., 2013)

and are thus supported by the meta model, as well.

This confirms the universal applicability of the ho-

listic process meta model and thus the generality of

the MKR for discrete manufacturing.

6.1.2 Comparison of Meta Models

To evaluate the holistic process meta model with

respect to existing process meta models in manufac-

turing, we did a qualitative comparison against the

requirements defined in Section 3. For the compari-

son (see Table 3), we chose the ISA-95-1 process

meta model (International Society of Automation,

2000) as it represents a common standard imple-

mented in manufacturing execution systems. More-

over, we selected the Virtual Factory (VF) Data

Model (Terkaj et al., 2012) as it integrates a wide

range of industrial and scientific manufacturing meta

models. In addition, we chose the value stream de-

sign (VSD) model (Erlach, 2011) because value

stream design is a typical method used to document

manufacturing processes.

All these models are universally applicable for

discrete manufacturing processes without focusing

on specific branches or industries. Yet, only our

holistic model integrates design-time aspects and

run-time aspects, that is, information on process

planning and process execution. We consider this an

important point for a holistic knowledge manage-

ment because the combined analysis of process exe-

cution information, e. g., resulting from machine

data or occurred failures, as well as process planning

information enables the generation of novel insights

(Kemper et al., 2013). For instance, a data-mining-

based root cause analysis of metric deviations as

described in Section 5.3 can reveal new knowledge

for process improvement. Neither the ISA-95-1

model, nor the VF model nor the VSD model sup-

port the explicit modelling of process execution

information.

Table 3: Comparison of process meta models.

Regarding the combination of a process view and a

resource view, our meta model as well as the VSD

model allow for the modelling of process flow as-

pects using gateways as well as basic resource in-

formation, e. g., on machines and production aids.

However, detailed specifications of resources, e. g.,

regarding maintenance requirements, are not cov-

ered by these models. In contrast, the ISA-95-1

model as well as the VF model provide an additional

resource view with details on all types of resources.

The tracking of the process evolution is fully

supported by our meta model and the VF model.

Both models include versioning concepts and keep

track of the adaption of process models. This is im-

portant to support continuous knowledge-driven

process improvement by reusing insights and evalu-

ating their improvement impact over time. The VSD

model does not focus on tracking process adaptions

and the ISA-95-1 model only supports versioning of

selected parts of the model without tracking changes

of the entire process model over time.

The model simplicity refers to the number of el-

ements and the structure of the model with respect to

the comprehensibility for the end user. We consider

model simplicity an important factor as it reduces

the barriers for the collection and reuse of insights

by the user which stimulates knowledge manage-

ment. The VF model is comparatively complex as it

integrates various meta models, e. g., on products,

processes and resources, and comprises several ab-

straction layers. Similarly, the ISA-95-1 model

comprises multiple generic definitions on processes

and resources, e. g., abstract resource requirements

are matched with actual resource capabilities. In

contrast, the VSD model is designed for a simple

modelling of manufacturing processes with a core

Holistic

Process

Meta

Model

ISA-

95-1

Model

Virtual

Factory

Data

Model

Value

Stream

Design

Model

Universal Applicability in

Discrete Manufacturing

++++

Integration of Design-Time

and Run-Time Aspects

+---

Combination of Process

View and Ressource View

-++-

Support for Process

Evolution

+-+-

Model Simplicity for Insight

Association

+- -+

TheManufacturingKnowledgeRepository-ConsolidatingKnowledgetoEnableHolisticProcessKnowledgeManagement

inManufacturing

list of process elements. Our holistic meta model

refines and extends these elements without referring

to generic views or definitions.

All in all, the qualitative comparison reveals that

only the holistic process meta model fully supports

the integration of design-time and run-time aspects

and provides both model simplicity and support for

process evolution. Although the resource view is

only basically represented, a coarse-grained associa-

tion of resource-related insights is possible with the

holistic process meta model. Hence, it provides a

sound basis for insight association and universal

process representation in the MKR.

6.2 Evaluation of the Insight Model

and MKR’s Analysis Support

In the following, we evaluate the insight model (see

Section 4.2) in combination with the MKR’s analy-

sis support and show that the MKR provides a com-

prehensive basis for the generation and storage of

analysis results and insights of major data analytics

systems (see Figure 6).

According to (Kemper et al., 2010), there are

four general types of data analytics for knowledge

generation in business intelligence: free data explo-

ration, OLAP, reporting and model-based analytics.

The analysis results of these systems represent in-

sights and thus have to be covered by the insight

model to store them in the MKR. Moreover, the

MKR as a whole should support the use of these

data analytics for the generation of new insights.

Free data exploration refers to the direct search

and browsing of insights in the MKR by the user.

There is no direct generation of new analysis results.

Free data exploration rather provides the basis for

further analytics by identifying needs for new anal-

yses, e. g., failure reports which require further root

cause analyses. The MKR fully supports free data

exploration by navigation features (see Section 5.3).

OLAP comprises the multidimensional analysis

of metric-oriented information (Pendse and Creeth,

1995). Metrics represent analysis facts and dimen-

sions constitute views on these facts, e. g., analyzing

the lead time of certain process steps. The MKR

fully supports process-oriented OLAP analyses be-

cause (1) the insight model defines metrics and their

relationships and (2) the MKR makes use of a multi-

dimensional warehouse structure with these metrics

as facts and elements of the process meta model as

dimensions (see Section 5.3).

Reporting systems focus on the textual, graphical

or diagram-oriented documentation of metric-

information in reports. Reports constitute semi-

structured or unstructured text documents and are

thus covered as text insights in the insights model.

Moreover, the multidimensional structure of the

MKR with metrics as central facts supports the use

of reporting systems, which are typically employed

on multidimensional data warehouses.

Figure 6: Types of data analytics and support by the MKR.

Model-based analytics comprise data mining ap-

proaches (Han et al., 2012) and expert systems

(Giarratano and Riley, 2005). The former refer to the

broad range of data mining techniques and models,

e. g., clustering and classification. The latter mainly

comprise case-based, model-based and rule-based

approaches. With respect to the insight model, the

six major types of data mining models are covered

explicitly and further types may be added flexibly by

inheritance. Formal rules are supported by the in-

sight model, as well. In contrast, formal cases and

formal models are not directly supported by the

insight model due to their heterogeneity in different

case-based und model-based applications. That is,

they have to be represented as unstructured textual

insights in order to incorporate them in the MKR.

Considering the generation of data mining models,

the MKR, with its multidimensional structure, fully

supports the use of data mining tools like

RapidMiner (see Section 5.3). Yet, the use of specif-

ic case-based, rule-based or model-based applica-

tions may require application-specific adaptions of

the MKR’s data structure due to missing standards

for expert systems.

To sum up, the MKR supports both the genera-

tion and storage of analysis results from reporting

and OLAP applications as well data mining systems

and includes free data exploration. The key enablers

are the insight model in combination with a multi-

dimensional structure based on the process meta

model which comprises process model data and

process execution data for analysis purposes. With

respect to expert systems, formal rules, models and

cases can be stored in the MKR. Yet, the use of the

MKR as a data basis for expert systems to generate

Types of Data

Analytics

Free Data

Exploration

Online Analytical

Processing

(OLAP)

Reporting

Model-based

Analytics

Fully supported by the MKR

Partially supported by the MKR

ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems

new insights requires additional application-specific

adaptions.

6.3 Evaluation of the MKR for

Knowledge Management

On the basis of the above evaluation of the holistic

process meta model and the insight model, we show

that the realization of the models in the MKR ena-

bles a holistic process knowledge management by

satisfying the whole range of process-oriented in-

formation needs in manufacturing.

In our previous work (Gröger et al., 2013), we

identified four general types of process-oriented

information, namely process context, process per-

formance, process documentation and process com-

munication. In the following, we describe these

information needs and analyze how the MKR satis-

fies them. Table 4 shows for each information need

whether it is satisfied using insights or meta model

components of the MKR.

 Process context refers to the structure and the

status of the overall process and its underlying re-

sources, e. g., machines, as well as the goods to

be produced. The MKR’s process meta model

comprises all information relevant for the process

context: Process steps and material gateways rep-

resent the structure and employees and operating

resources provide information on process re-

sources both from a design-time and a run-time

point of view. Information on the product and its

features is available, as well.

 Process performance alludes to information about

the effectiveness and efficiency of the process

and its resources. All information relevant for

process performance is provided by insights, es-

pecially metrics and data mining models, as well

as information about material consumption in the

meta model.

 Process documentation refers to information to

support the execution of a process, e. g., work in-

structions, as well as information for process im-

provement, especially improvement suggestions.

Process documentation can be represented as spe-

cial kinds of unstructured insights which may

comprise text, audio or video supported by the

MKR.

 Process communication covers information ex-

changed between employees participating in the

process, especially text, video or audio messages.

These can be treated as corresponding insights

and are therefore supported by the MKR, too.

Table 4: Information needs satisfied by the MKR.

To conclude, the MKR satisfies the whole range of

process-oriented information needs in manufacturing

and thus enables a holistic knowledge management.

The MKR consolidates knowledge across the entire

process lifecycle and facilitates sharing amongst

various target groups of users. Moreover, the MKR

enables the cross-correlation of different types of

knowledge like failure reports, metrics and data

mining models to support the discovery of new in-

sights for process improvement.

6.4 Case Example and Technical Proof

of Concept

The technical proof of concept is based on the appli-

cation of the prototype of the MKR (see Section 5.3)

in an exemplary case in the automotive industry, that

is, the mass production of steel springs for car mo-

tors as described in (Erlach, 2011). The manufactur-

ing process consists of several sequential steps for

winding, tempering and shot peening of springs and

involves various machines like winding machines.

For our technical proof of concept, we modelled

the manufacturing process according to the holistic

process meta model whereas we used the typical

model constructs to represent mass manufacturing

processes as described in Section 6.1.1. On this ba-

sis, we identified attributes of resources and process

steps, e. g., winding speed of winding machines, and

generated corresponding process model and process

execution data to populate the MKR with instance

data. Thereby, we generated data on 100.000 execu-

tions of the manufacturing process and calculated

metric values, e. g., for lead times and quality rates.

With respect to insights, we did several root

cause analyses on lead times using process execution

data in the MKR and deduced corresponding deci-

sion trees as data mining models which were stored

in the MKR. Moreover, we stored exemplary ma-

chine manuals, photos and reports as JPEG and PDF

files representing unstructured insights in the MKR.

Considering an application on top of the MKR,

we implemented a knowledge-based process dash-

board on an Android tablet pc addressing both

Process

Meta Model

Insights

Process Context

Process Performance

Process Documentation

Process Communication

TheManufacturingKnowledgeRepository-ConsolidatingKnowledgetoEnableHolisticProcessKnowledgeManagement

inManufacturing

workers on the factory shop floor and production

supervisors (see Figure 4). The dashboard is based

on our requirements analysis described in (Gröger et

al., 2013) and represents an application using the

MKR and its API to provide mobile access to differ-

ent kinds of process knowledge in different applica-

tion scenarios. For instance, workers can get infor-

mation on best practices and work instructions as

well as upload photos and reports of manufacturing

failures. Besides, production supervisors can corre-

late metrics and failure reports and execute root

cause analyses.

Based on our test system (Windows Server 2008

R2, Core i7-2620M@2,7 GHz, 8 GB RAM) and

data on 100.000 process instances, the MKR proved

to provide acceptable system performance for inter-

active usage in typical application scenarios of the

dashboard described in (Gröger et al., 2013).

This technical proof of concept demonstrates the

fundamental feasibility and applicability of the

MKR combined with suitable applications like the

dashboard. The MKR proved to provide the facilities

for insight generation, storage and reuse based on

data of a realistic manufacturing process.

7 CONCLUSION

AND FUTURE WORK

In this article, we introduced the Manufacturing

Knowledge Repository, a holistic repository facili-

tating process knowledge management in manufac-

turing. It consolidates structured and unstructured

knowledge, e. g., metrics, data mining models and

text documents, and can be used by various applica-

tions. We presented the conceptual data model in-

cluding a holistic process meta model and an insight

model and discussed different storage architectures.

We did a qualitative evaluation of the data models

and presented a technical proof of concept based on

a prototypical implementation in a case example.

With respect to future work, our goal is to im-

plement an alternative storage architecture for the

MKR and to investigate novel analytics on top of the

MKR. That is, on the one hand, we are going to

implement the extended architecture introduced in

this article. This architecture seams promising to us

as it exploits the functionality of a Content Man-

agement System for the workflow-based reuse and

distribution of insights in business processes. On the

other hand, we are going to examine novel analytics

which combine structured and unstructured

knowledge to generate new insights, e. g., combin-

ing data mining on process execution data and text

analytics on unstructured text documents.

ACKNOWLEDGEMENTS

The authors would like to thank the German Re-

search Foundation (DFG) for financial support of

this project as part of the Graduate School of Excel-

lence advanced Manufacturing Engineering

(GSaME) at the University of Stuttgart.

REFERENCES

Ackoff, R. L. (1989), “From Data to Wisdom”, Journal of

Applied Systems Analysis, Vol. 16, pp. 3–9.

Brown, M. G. (1996), Keeping score. Using the right

metrics to drive world-class performance, American

Management Association, New York.

Buzacott, J. A., Corsten, H., Gössinger, R. and Schneider,

H. M. (2013), Production planning and control. Ba-

sics and concepts, Oldenbourg, München.

Chapman, C. B. and Pinfold, M. (2001), “The application

of a knowledge based engineering approach to the rap-

id design and analysis of an automotive structure”,

Advances in Engineering Software, Vol. 32 No. 12,

pp. 903–912.

Data Mining Group (2013), “PMML 4.1 Standard”,

http://www.dmg.org/v4-1/GeneralStructure.html (ac-

cessed 06.02.13).

Davenport, T. H. and Prusak, L. (2000), Working

knowledge. How organizations manage what they

know, Harvard Business School Press, Boston.

Economist Intelligence Unit (2007), Knowledge manage-

ment in manufacturing, London.

Erlach, K. (2011), Value stream design. The way to lean

factory, Springer, Berlin.

Fischer, U., Stokic, D. and Beckmann, T. (2000), “Man-

agement of Corporate Knowledge for Process Im-

provement in Manufacturing Companies”, in Stanford-

Smith, B. and Kidd, P.T. (Eds.), E-business: Key Is-

sues, Applications and Technologies, IOS Press, Am-

sterdam, pp. 721–727.

Giarratano, J. C. and Riley, G. (2005), Expert systems.

Principles and programming, 4th ed., Thomson

Course Technology, Boston.

Giovannini, A., Aubry, A., Panetto, H., Dassisti, M. and

Haouzi, E. H. (2012), “Knowledge-Based System for

Manufacturing Sustainability”, in Proc. of the IFAC

Symposium on Information Control Problems in Man-

ufacturing (INCOM) 2012, Elsevier, pp. 1333–1338.

Goossenaerts, J., Arai, E., Shirase, K., Mills, J. J. and

Kimura, F. (2005), “Enhancing Knowledge and Skill

Chains in Manufacturing and Engineering”, in Arai,

E., Goossenaerts, J., Kimura, F. and Shirase, K. (Eds.),

Knowledge and skill chains in engineering and manu-

facturing, Springer, London, pp. 1–10.

ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems

Gröger, C., Schlaudraff, J., Niedermann, F. and

Mitschang, B. (2012a), “Warehousing Manufacturing

Data. A Holistic Process Warehouse for Advanced

Manufacturing Analytics”, in Proc. of the Int. Confer-

ence on Data Warehousing and Knowledge Discovery

(DaWaK) 2012, Springer, Berlin, pp. 142–155.

Gröger, C., Niedermann, F. and Mitschang, B. (2012b),

“Data Mining-driven Manufacturing Process Optimi-

zation”, in Proc. of the World Congress on Engineer-

ing (WCE) 2012, Newswood, Hong Kong, pp. 1475–

1481.

Gröger, C., Hillmann, H., Hahn, F., Mitschang, B. and

Westkämper, E. (2013), “The Operational Process

Dashboard for Manufacturing”, in Proc. of the CIRP

Conference on Manufacturing Systems (CMS) 2013,

Elsevier, pp. 205–210.

Han, J., Kamber, M. and Pei, J. (2012), Data Mining.

Concepts and Techniques, 3rd ed., Morgan Kaufmann,

Waltham.

International Society of Automation (ISA) (2000), Enter-

prise-Control System Integration. Part 1: Models and

Terminology, Standard ISA-95-1.

Kampffmeyer, U. (2007), ECM - Enterprise Content

Management, Project Consult, Hamburg.

Kaushish, J. P. (2010), Manufacturing processes, 2nd ed.,

Prentice Hall India, New Delhi.

Kemper, H.-G., Baars, H. and Lasi, H. (2013), “An Inte-

grated Business Intelligence Framework. Closing the

Gap Between IT Support for Management and for

Production”, in Rausch, P., Sheta, A.F. and Ayesh, A.

(Eds.), Business Intelligence and Performance Man-

agement, Springer, London, pp. 13–26.

Kemper, H.-G., Baars, H. and Mehanna, W. (2010), Busi-

ness Intelligence, 3rd ed., Vieweg, Wiesbaden.

Kiritsis, D. (1995), “A review of knowledge-based expert

systems for process planning. Methods and problems”,

International Journal of Advanced Manufacturing

Technology, Vol. 10 No. 4, pp. 240–262.

Lemaignan, S., Siadat, A., Dantan, J.-Y. and Semenenko,

A. (2006), “MASON: A Proposal For An Ontology Of

Manufacturing Domain”, in Proc. of the IEEE Work-

shop on Distributed Intelligent Systems (DIS) 2006,

IEEE, Los Alamitos, pp. 195–200.

Ma, Z., Wetzstein, B., Anicic, D., Heymans, S. and Ley-

mann, F. (2007), “Semantic Business Process Reposi-

tory”, in Proc. of the Workshop on Semantic Business

Process and Product Lifecycle Management (SBPM)

2007, CEUR Proceedings, Innsbruck, pp. 92–100.

Mazumdar, S., Varga, A., Lanfranchi, V., Petrelli, D. and

Ciravegna, F. (2012), “A Knowledge Dashboard for

Manufacturing Industries”, in Proc. of the European

Semantic Web Conference Worksops (ESWC) 2011,

Springer, Berlin, pp. 112–124.

Monauni, M. and Foschiani, S. (2013), “Agility in Produc-

tion Networks - Classification, Design and Configura-

tion”, in Proc. of. the Int. Conference on Production

Research (ICPR) 2013.

Morgan, T. (2002), Business rules and information sys-

tems. Aligning IT with business goals, Addison-

Wesley, Boston.

Niedermann, F., Schwarz, H. and Mitschang, B. (2011),

“Managing Insights - A Repository for Process Ana-

lytics, Optimization and Decision Support”, in Proc.

of the Int. Conference on Knowledge Management and

Information Sharing (KMIS) 2011, SciTePress, Paris,

pp. 424–429.

Pendse, N. and Creeth, R. (1995), The OLAP Report.

Succeeding with on-line analytical processing. Volume

1, Business Intelligence.

Poole, J., Chang, D., Tolbert, D. and Mellor, D. (2003),

Common warehouse metamodel. Wiley, New York.

Shah, J.J. and Mäntylä, M. (1995), Parametric and feature

based CAD/CAM. Concepts, techniques, applications,

John Wiley & Sons, New York.

Terkaj, W., Pedrielli, G. and Sacco, M. (2012), “Virtual

Factory Data Model”, in Proc. of the Int. Workshop on

Ontology and Semantic Web for Manufacturing

(OSEMA) 2012, CEUR, Aachen, pp. 29–43.

Vetlugin, A. (2012), A process insight repository support-

ing process optimization, Master Thesis, University of

Stuttgart.

Zor, S., Leymann, F. and Schumm, D. (2011), “A Pro-

posal of BPMN Extensions for the Manufacturing

Domain”, in Proc. of the CIRP Conference on Manu-

facturing Systems (CMS) 2011, pp. 1–6.

TheManufacturingKnowledgeRepository-ConsolidatingKnowledgetoEnableHolisticProcessKnowledgeManagement

inManufacturing