International journal of computer integrated manufacturing , tập 24, số 5, 2011

develop a systematic and comprehensive definition of dynamic bottlenecks of the production networks based on both the TOC and the Bottleneck Oriented LogisticAnalysis BOLA.. Targeting the

The CIRP-Sponsored International Conference of

Digital Enterprise Technology (DET2009) was held

on 14–16 December 2009 at The University of Hong

Kong This international conference series of Digital

Enterprise Technology aims to provide a forum for

academia and industrialists to disseminate, to all

branches industries and businesses, information and

knowledge on the most recent and relevant

innova-tions, theories and practices in electronic business and

digital enterprise technology This special issue is based

on contributions carefully reviewed and extended from

the Proceedings of DET2009 on the topic of Enterprise

Informatics Enterprise Informatics deals with the

optimal use of enterprise-level information and

knowl-edge to support decision-making processes and/or

day-to-day operations Eleven papers were eventually

selected out of over 130 DET2009 contributions

They cover a set of issues concerning some important

research and application developments of enterprise

informatics technologies, forming the broad basis of

research agenda to identify and explore the ways in

which users and their actions can be facilitated

The issue begins with a paper entitled ‘Planning

towards enhanced adaptability in digital manufacturing’

by Wang The paper presents an integrated approach

for developing a web-based system with enhanced

adaptability, including distributed process planning,

real-time monitoring and remote machining This

approach is enabled by a Wise-ShopFloor (Web-based

integrated sensor-driven e-ShopFloor) framework

tar-geting distributed yet collaborative manufacturing

environments It allows end-users to plan and control

distant manufacturing operations based on runtime

information from shop floors An example of

dis-tributed process planning for remote machining is

chosen to demonstrate the effectiveness of the

approach towards web-based digital manufacturing

Bottlenecks, as the key ingredients for improving

the performances of the production networks, have

been profoundly studied However, the major

defini-tions of bottlenecks are derived in terms of the

throughput and based on the Theory of Constraints

(TOC) In the paper ‘Modelling dynamic bottlenecks in

production networks’, Scholz-Reiter et al develop a

systematic and comprehensive definition of dynamic

bottlenecks of the production networks based on both

the TOC and the Bottleneck Oriented LogisticAnalysis (BOLA) Distinguishing from the traditionalview at the bottlenecks, the defined dynamic bottle-necks are modelled by means of discrete simulationusing practical data, aimed at visualising them in theproduction network By applying the logistic operatingcurves (LOCs), the practical application of theproposed research and its procedures is discussed indetail

Multi-material products are getting increasinglypopular in recent years However, no systems havebeen developed to support the design of suchproducts In the paper ‘A haptic-based part decom-position method for multi-material product design’,Chen et al propose a haptic-based method fordesigning multi-material products This approachconsists of three main parts: haptic painting/marking,boundary smoothing/fitting and volume decomposi-tion The haptic device provides an intuitive userinterface for quick volume mark up in a multi-material product by direct mesh painting Based onthe boundaries of painted regions, volume decom-position can be done automatically when needed Thenumerous iterations of volume mark up and decom-position in the early stage of multi-material productdesign can now be made easy and effective with theproposed method

The paper ‘Optimal service selection and tion for service-oriented manufacturing network’, byHuang et al presents the service management ofservice-oriented manufacturing network (SOMN) Itconsiders the key problem, the optimisation of serviceselection and composition, to realise the integrationand optimisation of services in an open environmentwhich contains large amounts randomicity and un-certainty The integrated performance evaluationmetrics for SOMN is described, which combines thekey performance indicators of services from business,service and implementation level The performanceevaluation model is brought forward to analyse thelocal and global performance An uncertainty andgenetic algorithm-based method is developed to realisethe optimisation of service selection and composition

composi-in an effective and efficient way

Zhang et al introduce ‘Real-time work-in-progressmanagement for smart object enabled ubiquitous shop

ISSN 0951-192X print/ISSN 1362-3052 online

Ó 2011 Taylor & Francis

DOI: 10.1080/0951192X.2011.568841

floor environment’ This paper proposes a

work-in-progress (WIP) management framework for a

ubiqui-tous manufacturing (UM) environment Under the

framework, two types of services and a WIPA

(work-in-progress agent) are designed and developed To

implement the integration of heterogeneous Enterprise

Information Systems (EISs), the wipML

(work-in-progress markup language) is established based on

some important standards such as ISA 95 and

B2MML During production execution, real-time

visibility explorers are provided for operators and

supervisors to reflect the real-time situation of current

manufacturing environment The presented framework

is studied and demonstrated using a near real-life

simplified shop floor that consists of typical

manufac-turing objects

In view of the seamless integration of products,

production systems and business processes, the paper

‘Framework for extended digital manufacturing systems’

by Nylund et al deals with their work on building up a

framework of EDMS – Extended Digital

Manufactur-ing Systems A reference model of EDMS is presented

that consists of manufacturing entities with different

roles and similar structures EDMS provides an

integrated environment for products, production

systems and business processes The theoretical

appli-cation areas of the past, present and future and the

process from ideas to innovative solutions are

de-scribed A real-life example of an intelligent

manufac-turing environment is introduced to demonstrate

manufacturing development, operation, performance

monitoring and measurement, etc of the EDMS

framework

Production service system (PnSS) is a new

business mode where a manufacturer obtains

man-ufacturing resources in the form of continuous

production services instead of resource entities In

the paper ‘Analytical target cascading enabled optimal

configuration platform for production service systems’,

Qu et al focus on the configuration platform of the

PnSS business mode A systematic PnSS

configura-tion methodology and the enabling platform are

developed based on a newly extended Analytical

Target Cascading (ATC) method As ATC

accom-modates heterogeneous sub-system integration and

multi-level problem solving, the methodology is able

to address the typical challenges that a practical

component service PnSS configuration process

nor-mally faces, such as distributed decision rights,

uncertain decision structure and short decision

period

In the paper ‘A hierarchical deployment of

distributed PLM system in collaborative product

devel-opment’, Chu et al propose a novel methodology of

distributed Product Lifecycle Management (PLM)platform deployment It incorporates informationtechnologies to support collaborative product devel-opment in a global enterprise with multiple distributedsites The operational guidelines are provided fordetermining the hardware configuration integratedfrom the analysis of the organisational, data, contentand application views of the requirements of theenterprise A customised deployment plan that reflectsthe actual user needs is then generated to construct acost-effective collaborative PLM platform Finally, acase study is included to demonstrate the feasibility ofthe proposed methodology

One of the most difficult problems in mobile robotnavigation is the acurate estimation of the robot’sposition and orientation In the paper entitled ‘Highaccuracy mobile robot positioning using external largevolume metrology instruments’, Wang et al present amethod of accurately controlling the position of amobile robot using an external Large Volume Metro-logy (LVM) instrument such as the laser tracker, anavigation algorithm, and a low cost robot, arepeatability of 5 mm could be achieved over a volume

of 30 m radius In addition, a surface digitisation scan

of a wind turbine blade section is also demonstrated,illustrating possible applications of the proposedmethod for manufacturing processes

The ‘near-zero inventory production’ and time delivery’ is becoming a powerful solution toreduce the inventory cost and improve the productionefficiency and benefits for enterprises The paper ‘ARFID-based optimal material delivery for digital plantproduction’ by Zhou et al uses Radio FrequencyIdentification (RFID) technology to develop a real-time optimal material delivery method A mathema-tical model for dynamically obtaining the optimalroutes for forklifts is established by taking theminimal travel distance as the objective Then, anAnt Colony Optimisation (ACO)-based real-timeoptimum route planning algorithm is designed tosolve the material delivery problem The feasibility ofpresented model and algorithm is validated by a casestudy

‘real-Evolvable systems has been developed and tested as

a next-generation production system paradigm sinceits inception in 2002 The paper entitled ‘Evolvablesystems: an approach to self-X production’ by Onori

et al presents current developments and applications

of Evolvable Product Systems (EPS) It has beenpointed out that the essence of evoluability resides notonly in the ability of system components to adapt tothe changing operational conditions, but also in theevolution of these components over time such thatprocesses may become self-X, where X stands for one

comments to the papers The guest editors are grateful

to Springer, the publisher of DET2009 Proceedings,

for giving permission to extend the papers for this

special issue Finally, the guest editors would like to

express their thanks to Professor Stephen Newman

(Editor-in-Chief) and the Journal Office for their

advice and support that made this special issue project

a success

Professor George Q HuangDepartment of Industrial and Manufacturing

Systems Engineering,The University of Hong Kong, Hong Kong, PR China

gqhuang@hku.hk

Guangdong University of Technology, Guangdong,

PR Chinaquting@gdut.edu.cnProfessor Paul G MaropoulosDepartment of Mechanical

EngineeringUniversity of Bath, UKP.G.Maropoulos@bath.ac.uk

Planning towards enhanced adaptability in digital manufacturing

Lihui Wang*

Virtual Systems Research Centre, University of Sko¨vde, Sweden(Received 25 March 2010; final version received 5 July 2010)This paper presents an integrated approach for developing a web-based system with enhanced adaptability,including distributed process planning, real-time monitoring and remote machining The objective is to develop anew methodology and relevant processing algorithms for enhancing adaptability in digital manufacturing Thisapproach is enabled by a Wise-ShopFloor (Web-based integrated sensor-driven e-ShopFloor) framework targetingdistributed yet collaborative manufacturing environments Utilising the latest Java technologies (Java 3D and JavaServlet) for system implementation, it allows end-users to plan and control distant manufacturing operations based

on runtime information from shop floors Details on the principle of the Wise-ShopFloor framework, systemarchitecture, and a prototype system are reported in this paper An example of distributed process planning forremote machining is chosen as a case study to demonstrate the effectiveness of this approach toward web-baseddigital manufacturing

Keywords: process planning; web-based monitoring; remote machining; digital manufacturing

1 Introduction

In the last decade, digital manufacturing has emerged

as the norm of manufacturing in a computer- and/or

web-based environment This is largely due to the

global business decentralisation and outsourcing,

where the large-scale operations are better tested

digitally than physically To stay competitive in the

global market, companies with distributed operations

are demanding a new way of effective collaborations

between themselves and service providers Among

many factors, flexibility, timeliness and adaptability

are identified as the major characteristics in this

research to bring dynamism to manufacturing

Targeting the manufacturing dynamism in a

distributed environment, this research introduces a

Wise-ShopFloor (Web-based integrated sensor-driven

e-ShopFloor) framework for distributed process

plan-ning, dynamic scheduling, real-time monitoring, and

remote control This approach is supported by sensors,

function blocks, as well as Java and Web technologies

The Wise-ShopFloor is designed to use the popular

B/S (browser/server) architecture, as well as VCM

(view-control-model) and publish-subscribe design

patterns for effective information sharing during

decentralised planning and control

The rest of the paper is organised as follows In

Section 2, enabling technologies including Web,

Inter-net, Java 3D and Java servlets are introduced based on

a brief literature review It is followed by a description

of the Wise-ShopFloor framework in Section 3 Details

on adaptive and distributed process planning are ented in Section 4, which leads to web-based real-timemonitoring and control documented in Section 5 Acase study using planning results for web-basedremote machining are described in Section 6 Finally,contributions are summarised in Section 7

pres-2 Enabling technologiesWith the growing manufacturing decentralisation,products and services are distributed everywhere andsourced anywhere along supply chains Product designand fabrication have shifted rapidly from intra-corporation to global networks How to coordinatemanufacturing activities and keep them under control

is a challenging issue Flexibility, timeliness andadaptability of manufacturing operations are theessential requirements for digital manufacturing insuch a dynamic environment Fortunately, the Webinfrastructure today is mature enough to form adistributed manufacturing network through browser-server inter-connections In addition to the Webtechnology, Java has brought about a fundamentalchange in the way that applications are designed anddeployed With Java, the browser paradigm hasemerged as a compelling way to produce collaborativeapplications over the Web Examples include WebCA-DET (Caldwell and Rodgers 1998) for collaborative

*Email: lihui.wang@his.se

Vol 24, No 5, May 2011, 378–390

ISSN 0951-192X print/ISSN 1362-3052 online

Ó 2011 Taylor & Francis

DOI: 10.1080/0951192X.2010.506657

application server must engage users in a 3D graphical

interaction in addition to the dialogue-like data

sharing, because remote users need visual aids to

coordinate their efforts in a digital environment

Today, digital manufacturing tops the wish list for

many manufacturers Unfortunately, most of the

manufacturing equipment of today does not have the

built-in capability to transmit and receive data Few of

the available Web-based systems are designed for

shop-floor monitoring and control or for advanced factory

automation Some related systems listed below are

limited in their functionality and platform requirements

The latest Cimplicity (GE Fanuc 2008) allows users to

view a factory’s operational processes through an

XML-based WebView screen, including all alerts on every

Cimplicity system The FactoryFlow of Unigraphics

Solutions (USA) can provide off-line factory layout

planning, material handling, and simulation

(Waurzy-niak 2001) By most estimates, the number of CNC

machines capable of linking to the Internet is less than

10% of the installed base, according to Waurzyniak

(2001) Seeking the opportunity in linking CNC

machines with the Internet, MDSI (USA) uses

Open-CNC(MDSI 2008) to automatically collect and publish

machine and process data on a network In 1999, Hitachi

Seiki (Japan) introduced FlexLink to its

turning/machi-ning centres (http://www.flexlink.com) Since 1998,

Mazak has operated its high-tech Cyber Factory concept

(Mazak 1998) in Japan The fully networkable Mazatrol

Fusion controlallows Mazak machines to communicate

over wireless factory networks for real-time machine

tool monitoring and diagnostics In addition,

Japan-based Mori Seiki introduced a CAPS-NET system that

polls machine tools on Ethernet at settable increments,

usually five-second or longer, for engineers to get

updates on machine tools’ run-time status in production

(Mori Seiki 2009) To bring legacy machine tools with

serial ports on-line, e-Manufacturing Networks Inc

(now Memex Automation, Canada) introduced its ION

Universal Interface and CORTEX Gateway (Memex

2009) to help the old systems go online

Despite various accomplishments, the available

systems mentioned above are either for off-line

simulation or for monitoring only Most systems

require a specific application to be installed instead

of using a standard user interface such as a web

digital manufacturing Web and Java technologies arealso adopted as enabling technologies for systemimplementation Details on how the two technologiescan work together are explained below

3 Wise-ShopFloor frameworkThe Wise-ShopFloor framework has been designed toprovide users with a web-based and sensor-drivenintuitive environment where distributed process plan-ning, dynamic scheduling, real-time monitoring andremote control are undertaken Within the framework,each machine should become an information node and

be a valuable resource in the information network Adirect connection to sensors and machine controllers isused to continuously monitor, track, compare, andanalyse production parameters Instead of cameraimages (usually large in data size), a physical device

of interest (e.g a milling machine) can be represented

by a Java 3D model with behavioural control nodesembedded Once downloaded from its applicationserver, the 3D model is rendered by the local CPUand can work on behalf of its remote counterpartshowing real behaviour for visualisation at a clientside It remains alive by connecting with the physicaldevice through low-volume message passing (sensordata) In addition to motion data, other sensor dataincluding temperature, vibration and force can also betransmitted via network and shown in colours orcontour lines on the 3D model for machine conditionmonitoring As the 3D model is driven by the sensorsand rendered locally for visualisation, there is no needfor transmitting camera images over the Internet Thelargely reduced network traffic makes real-time mon-itoring and remote control practical for dispersedusers It also enables users to make accurate decisions

in a timely manner, and to ensure that machines areoperating within defined expectations Being able toplan and control shop-floor operations from anywhereyet at any time, collaboratively, is what this research isaiming at Figure 1 illustrates the architecture of theWise-ShopFloor framework

The framework is designed in B/S architectureusing VCM design pattern with built-in secure sessioncontrol The mid-tier application server handles majorsecurity concerns, such as session control, session

registration, sensor data collection and distribution,

planning and scheduling, as well as real device

manipulation A central Session Manager is to look

after the issues of user authentication, session

syn-chronisation, and sensitive data logging All initial

transactions need to go through the Session Manager

for access authorisation In a multi-client environment,

different users may require different sets of data or

logic for different tasks For example, in the case of

monitoring, it is not efficient to have multiple users

who share the same model talking with the same device

at the same time Publish-subscribe design pattern is

adopted to collect and distribute sensor data at the

right time to the right user, efficiently As a server-side

module, the Signal Collector is responsible for sensor

data collection from networked physical devices The

collected data are then passed to another server-side

module Signal Publisher who in turn multicasts the

sensor data to the registered subscribers (clients)

through applet-servlet communication A Registrar is

designed to maintain a list of subscribers with the

requested sensor data when the subscribers have

selected appropriate machines for monitoring,

includ-ing IP address and port number of each subscriber,

along with the chosen machine A Java 3D model can

thus communicate indirectly with sensors no matter

where the client is HTTP streaming is chosen as the

communication protocol between server and clients

Although the global behaviours of a Java 3D

model are controlled by the server based on real-time

sensor signals, users still have the flexibility of viewingthe model from different perspectives at a client side

In order to control a device, an authorised user cansend control commands to the application serverwhich in turn manipulates the physical device TheWise-ShopFloor framework provides an alternative ofcamera-based monitoring through Java 3D models.Nevertheless, an off-the-shelf web-ready camera caneasily be switched on remotely to capture anunpredictable (un-modelled) scene for diagnosticpurposes

4 Distributed process planning4.1 Architecture designFigure 2 shows the detailed architecture of a newCAPP (computer-aided process planning) module Inthe current Wise-ShopFloor, the process planning isdedicated to machining operations and is realised by atwo-layer structure of shop-level Supervisory Planningand machine-level Operation Planning A process plangenerally consists of two parts: generic data (machin-ing method, machining sequence, and machiningstrategy) and machine-specific data (tool data, cuttingparameters, and tool paths) Such a two-layer structurecan separate generic data from machine-specific ones.Since resources, knowledge, and decisions are bothlogically and geographically distributed, such a processplanning approach is also named Distributed ProcessPlanning(DPP) (Wang et al 2003)

Figure 1 Wise-ShopFloor framework

The supervisory planning focuses on product data

analysis, machining feature (m-feature) parsing, setup

planning, machining process sequencing, and machine

selection, while the operation planning considers jig/

fixture selection and detailed working steps for every

machining operations, including cutter selection,

cut-ting parameters assignment, tool path planning, and

control code generation At the supervisory planning

stage, the decisions made are generic and applicable to

all machines Process optimisation is performed at the

operation planning stage when specific resources

(machine, tool and fixture) are known and within a

relatively small search space for better adaptability

Job assignment and dispatching to the best available

machines are dealt with by a separate dynamic

scheduling module

4.2 Process sequencing

One critical task in process planning is machining

sequence generation Since a part design can be

decomposed into basic m-features (such as hole, slot,

pocket, etc.), the task of machining process sequencing

is literally treated as the task of putting m-features into

proper setups and in the right sequence, which is called

m-sequencing in DPP A generic process plan, as a

result of m-sequencing, only consists of

machine-neutral information in the form of machining

se-quence, including both critical (with datum references

and other manufacturing constraints) and non-critical

machining operations Some of the non-critical

sequences are presented in parallel whose specific

sequence will be determined by a CNC controllerduring operation planning Before an m-feature can bemachined, it must be grouped into a setup for the ease

of fixturing The basic idea of feature grouping is todetermine a primary locating direction of a setup, andgroup the appropriate m-features into the setupaccording to their pre-defined tool access directions.This process is repeated for a secondary locatingdirection and so on until all the m-features areproperly grouped

Here, a primary locating direction is the surfacenormal V* of the primary locating surface (LS) It can

be determined by the following equations:

Tmax are the maximum values of A* and T* of allcandidate locating surfaces A generalised accuracygrade T can be obtained by applying the algorithmsdescribed in Boerma and Kals (1989) and Rong et al.(1997) The surface f(A*, T*) satisfying the second half

of Equation (1) is chosen as the primary locatingFigure 2 Architecture of distributed process planning

surface, whose surface normal can be derived by the

derivations with respect to x, y and z, respectively

Based on V*, those m-features whose tool access

directions T*EMF are opposite to V* are grouped into

To be generic, the setups at this stage are planned

for 3-axis machines only A setup merging is handled

by the Execution Control module for 4-axis or 5-axis

machines, if needed, after a specific CNC machine is

selected (Wang et al 2006)

4.3 Function block design

Function block specification (IEC 61499-1, 2005) is an

IEC standard for distributed process measurement and

control, particularly for PLC control A function block

(FB) is a reusable functional module based on an

explicit event-driven model, and provides for data flow

and finite state automata based control It is relevant

to DPP and CNC control in machining data

encapsu-lation and process plan execution In the DPP, we use

function blocks to address the manufacturing

uncer-tainty through resource-driven algorithms embedded

in each function block The event-driven model (or

resource-driven algorithms) of a function block gives a

CNC machine more intelligence and autonomy to

make decisions on how to adapt a generic process plan

to match the actual machine capacity and dynamics It

also enables dynamic task scheduling, execution

control, and process monitoring

Three basic FB types are defined in the DPP: (1)

machining feature FB (MF-FB), (2) event switch FB

(ES-FB), and (3) service interface FB (SI-FB) Figure

3(a) depicts a typical 4-Side Pocket MF-FB A basic

FB like this can have multiple outputs and can

maintain its unique internal state, meaning that it

can generate alternative outputs even if the same

inputs are applied The fact is of vital importance for

adaptive cutting condition modification, after the FB

has been dispatched to a machine, by changing the

internal hidden state of the FB For example, the same

4-Side Pocket MF-FB can be used for both roughing

and finishing at the same machine but with different

cutting parameters and/or tool paths, by adjusting the

internal state of the FB to fine-tune the algorithms in

use Such behaviour is controlled by a finite state

machine, whose operation can be represented by an

ECC (execution control chart) as shown in Figure 3(b)

The START state is an initial idle state ready for

receiving event inputs EI_INI (an incoming event

requesting initialisation) triggers the state transition

from START to INI for FB initialisation, and when

the state INI is active, the algorithm ALG_INI is beingexecuted Upon its completion, ALG_INI will trigger

an event output EO_INI indicating the success of theinitialisation

Similarly, for state transitions to RUN, UPDATEand MON (status monitoring), different algorithmsALG_RUN (MF-FB execution), ALG_UPDATE(cutting condition update), and ALG_MON (MF-FBmonitoring) are triggered, correspondingly An event

‘1’ means a state transition is always true That is tosay, the state will transit back to the START state and

be ready for receiving the next event input

While MF-FBs define the functional relationships

of events, data and algorithms for each machiningfeatures fabrication, their combination can form acomposite FB representing a setup It may consist ofmore than one basic FB with partially sequencedconnections via events and data The event flow amongMF-FBs determines their unique machining sequence.Figure 4(a) shows a composite FB, where the eventflow among the three MF-FBs is facilitated at run-time

by an event switch FB (ES-FB) For instance, if asequence of ‘342’ is given, the ES-FB will fire eventsaccordingly to the appropriate MF-FBs for featurefabrications in the order of 3!4!2 It thus addsFigure 3 A basic machining feature function block

In addition to the MF-FBs and ES-FB, a service

interface FB (SI-FB) is defined and used, as shown in

Figure 5, to facilitate execution control of MF-FBs It

also enables machining process monitoring during FB

execution In the DPP, all MF-FBs are grouped into

setups before being dispatched to appropriate

ma-chines An SI-FB is plugged to each setup with the

following duties: (1) collects runtime execution status

of an MF-FB including FB ID, cutting parameters,

and job completion rate; (2) collects machining status

(cutting force, cutting heat, and vibration, etc.) if made

available; and (3) reports any unexpected situations to

the DPP, e.g security alarm and tool breakage, etc

Similar to other FB types, an SI-FB has five

embedded algorithms for requesting and reporting

execution status (ES), machining status (MS), and

The specific algorithms of each function block type areimplemented in Java More details on FB design andimplementation will be reported separately

5 FB monitoring and control5.1 System configurationFigure 6 illustrates a typical configuration for web-based machining, where a 5-axis milling machine ishooked up to the network for remote monitoring andmachining The machine is equipped with a PC-basedcontroller that serves as a gateway between itself andthe application server TCP (transmission controlprotocol) is adopted for communication between themachine and the server, whereas HTTP streaming isused for data sharing from the server to the remoteusers While the former is better for hardwareprotection with handshaking, the latter is firewall-transparent and suitable for web-based applications.This configuration allows a remote user to monitor theabsolute and relative motions of all axes as well as tocontrol the spindle speed and feed rate for off-sitemachining

5.2 Sensor data collection

As mentioned earlier, Wise-ShopFloor implements thePublish-Subscribe design pattern An end-user sub-scribes to information pertaining to a specific machine,leaving an open connection to receive events When anew event for that machine is posted, it is publishedonly to those users who have subscribed to it In theWise-ShopFloor, this communication is handled by amodification of the Pushlet (van den Broecke 2000).Figure 7 shows the event and data flows between end-users (clients) and real machines

A client communicates with the Pushlet directly forsubscription The Pushlet leaves the connection to theclient open after receiving the request, allowing data to

be streamed without reopening a connection for eachupdate When the publisher has new data to thestream, the Pushlet will then ensure that only the mostrecent information is used for monitoring ThePushlet also provides a Postlet, used by clients topost events to the Publisher This is equivalent tosending a command for machine control When a userFigure 4 ES-FB for parallel m-features sequencing

wishes to control a machine, he/she must seek

permission first from the application server before

entering into the control mode At any given time, only

one user can be granted the control authority for

manipulating a given machine to prevent possible

control conflicts from happening

On the real machine side, the data collection is

slightly different There are many different types of

machines usually with different types of

controllers The Pushlet package provides an

adap-ter, the Event Pull Source, which can be extended to

obtain data from a required source (real machine)

Events are pulled from an Event Pull Source at a

regular interval, which can be set to a desired

increment to approximately replicate real-time

monitoring

In the collection of sensor data from machines, theserver containing the Pushlet acts in fact as a client ofthe device controllers, establishing a socket connectionand working with the provided interface of eachcontroller The concrete implementation of the EventPull Source is one adapter between the interface of adevice controller and the interface to the Pushlet.However, the communication to the real machine must

be bi-directional to achieve control, even though theEvent Pull Source communicates in only a singledirection – from the machine to the application server

A device controller is not able to interpret the Pushletevent, and thus will not be a client of the Pushlet.Another Pushlet adapter, the Machine Adaptor, isrequired to take information from the Postlet (i.e fromthe client), and send it to a device controller in theFigure 5 SI-FB for execution control and monitoring

Figure 6 Configuration for web-based machining

required format As the Pushlet does not provide this

functionality, the Wise-ShopFloor uses a wrapper for

the Postlet, which determines whether data is destined

for the publisher or the machine, and thus directs it

appropriately

5.3 Data packet format

As shown in Figure 6, sensor data collection from the

milling machine is accomplished over the TCP

connection using a series of 12 floating numbers (of

relative and absolute positions of 5 axes, feed rate FR

and spindle speed SS) and one long integer (of a

control word CW) that form one complete data

packet In the current implementation, a typical data

packet is defined as follows

The control word is reserved to indicate the status

of the machine, including its operation mode (manual/

auto/jogging), coordinate system, and axis status, etc

Similar to a real CNC controller, the data packet

provides both relative and absolute positions of the

five axes that are used for joints transformation and

3D model rendering for off-site monitoring and

machine control

5.4 Java 3D enabled visualisation

For the sake of network bandwidth conservation,

Java 3D is chosen for geometric modelling of the

milling machine, as an alternative of camera-based

solutions Java 3D is a fourth-generation 3D API

(Barrilleaux 2001) What sets a fourth-generation

API apart from its predecessors is the use of graph architecture for organising 3D objects in thevirtual world Enabled by the scene-graph, Java 3Dprovides an abstract, interactive imaging model forbehaviour control of 3D objects Different fromother scene graph-based systems, a Java 3D scenegraph is an acyclic graph The connections betweenJava 3D nodes are always forming a direct relation-ship: parent to child

scene-The 5-axis machine shown in Figure 6 requireslinear motion control of X, Y, and Z axes, as well asrotary motion control of B and C (around Y and Zaxes, respectively) A combined rotary stage havingtwo rotary motions is mounted on top of an X-table,whereas the spindle head of the machine provides theother two linear motions along Y and Z axes Figure 8illustrates the Java 3D scene graph model of themachine

The scene graph contains a complete description ofthe entire scene It includes the geometries, theattributes, and the viewing information needed torender the scene from a particular point of view AllJava 3D scene graphs must connect to a VirtualUniverse object to be displayed, which provides thegrounding for the entire scene A scene graph itself,however, starts with BranchGroup (BG) nodes ABranchGroup node serves as the root of a sub-graph,

or branch graph, of the scene graph The formGroupnodes inside of a branch graph specify theposition, the orientation, and the scale of the geometricobjects in the virtual universe Each geometric objectconsists of a Geometry object, an Appearance object, orboth The Geometry object describes the geometricshape of a 3D object The Appearance object describesthe appearance of the geometry (colour, texture, andmaterial reflection characteristics, etc.) The behaviour

Trans-of the machine is controlled by Behaviour nodes, which

is subject to sensor data and is specific The results of sensor data processing can beembedded into the codes for remote monitoring Once

implementation-Figure 7 Event/data flow between clients and machines

1 2 3 4 5 6 7 8 9 10 11 12 13

Relative

position

Absoluteposition

FR SS CW

applied to a TransformGroup node, the defined

behaviour control affects all the descending nodes In

our case, the five motions (X-Table, Rotary Stage-1,

Rotary Stage-2, Spindle Head, and Spindle) are

controlled by their corresponding behaviour control

nodes, for on-line monitoring/control and off-line

simulation As the Java 3D model is connected with

its physical counterpart through the control nodes by

low-volume message passing (real-time sensor signals

and control commands), it becomes possible to

remotely machine a part on a real machine through

the Wise-ShopFloor, where the physical security is

addressed separately

5.5 Web-based machining

Web-based rapid machining is possible by sending

proper NC control commands through the

applet-servlet (or Cyber Controller–Control Commander–

Machine) communication as shown in Figure 1 In

order to remotely machine a part (e.g rapid tooling/

prototyping), user authentication and authorisation

must be accomplished for the user who demands this

operation This is done by setting a bit in the control

word in a data packet that is sent to the user If the

client has requested the control right and the bit is

set, a message will appear on the screen notifying the

user that he/she is now in control of the machine

For the purpose of remote machining, a control

word, similar to CW in the monitoring data packet,

is sent back to the machine controller, augmented by

a text string containing lines of an NC program

Thus not just manual control can be exercised

off-site, but a complete NC program generated by the

DPP can be remotely executed For example, thefollowing NC line tells the machine controller toproceed from the current position to the next,Figure 8 Scene graph model of a 5-axis milling machine

Figure 9 A test part and its machining sequence

incrementally by (20, 730, 10) in linear rapid

traverse mode At the same time, it sets the spindle

speed to 3000 rpm with flood coolant

G0 Xþ 20 Y  30 Z þ 10 S3000 M8

Most web-based systems rely on camera-based

monitoring to guide remote operations Compared

with one 8-bit VGA camera image of 640 6 480

(307,200 bytes), the data packet size of

Wise-Shop-Floor is only 52 bytes – a significant reduction suitable

for web-based real-time applications This comparison

is within the context of data streaming for remote

monitoring Although sensory and image information

are different, they are treated as digits when

trans-mitted Therefore, the efficiency of data transmission

depends on their data packet size Moreover, in the

case of real-time data streaming, limited

pre-processing is done for video images owing to thetime constraint

6 Case study

A test part shown in Figure 9(a) is chosen for the casestudy After applying defined reasoning rules of DPP(Wang et al 2006), the 14 m-features are grouped intotwo setups, each of which consists of two or morepartially sequenced m-features as shown in Figure 9(b).While each m-feature is mapped to an FB, a setupforms a composite FB Figure 10 shows the composite

FB for Setup-2, where the needed NC code can begenerated at runtime by the MF-FBs

In the Wise-ShopFloor, the adaptive process planshown in Figure 10 can be dispatched to a millingmachine for rapid fabrication utilising the real-timemonitoring and control functions discussed above

Figure 11 User interface for web-based remote machining

Control mode.

Although the Wise-ShopFloor prototype provides

an alternative of camera-based monitoring, an

off-the-shelf web-ready camera can easily be switched on

remotely at any time to capture un-modelled scenes for

trouble-shooting

7 Discussions and conclusions

This paper presents a novel approach toward

web-based digital manufacturing, including distributed

process planning, real-time monitoring and remote

control, whereas scheduling is handled separately by a

third-party system Within the Wise-ShopFloor

frame-work, a prototype has been designed in

view-control-model architecture and developed using

publish-sub-scribe design pattern for sensor data collection and

distribution For process planning, our approach is to

separate machine-specific data from generic ones using

a two-layer structure of Supervisory Planning and

Operation Planning A generic process plan has been

embedded into function blocks with built-in

algo-rithms for machine-level decision-making A

planning-machining case study demonstrates its feasibility and

shows promise of this approach in a distributed digital

manufacturing environment By closing the loop of

information flow from machine status monitoring

through planning to machine control, an adaptive

and well informed decision making becomes

possible Thus, planning and control can be performed

right up to the point when there is a change on a

shop floor

In order to deal with system complexity,

Wise-ShopFloor decomposes a complex problem to simple

ones and then solves them individually The

monitor-ing module of the Wise-ShopFloor provides runtime

information from shop floors in a holistic way, while

the planning and control modules target specific

components on specific shop floors Supported by the

runtime information, an engineer can perform

multiple tasks, or a task can be performed by multiple

engineers, in a distributed environment

Regarding resources, there exist two types of

resource allocation The first is resource allocation

for end users, and the second is that for tasks The first

type is a resource scheduling issue that is done by

process planning, and (3) a sensor-driven approach forweb-based real-time monitoring and machine control.The Wise-ShopFloor is operation independent It can

be applied to machining, or extended for welding andassembly operations However, the planning algo-rithms and control logics (some of them are embeddedinto function blocks) are operation specific Differentalgorithms and logics are required when the Wise-ShopFloor is applied to different operations, as is oftenthe case on real shop floors

The future work of Wise-ShopFloor researchincludes further improvement of system security, in-cluding byte-code verification, user permissions, andsecurity policies In addition, other issues such asdigital rights management, data encryption, confiden-tiality protection, and shop equipment protection arealso planned

AcknowledgementThis paper has been presented in DET2009 conference held

at the University of Hong Kong during 14–16 December

2009 and was included in the conference proceedingspublished by Springer

ReferencesAzevedo, A., Toscano, C., Sousa, J.P., and Soares, A.L.,

2004 An advanced agent-based order planning systemfor dynamic networked enterprises Production Planningand Control, 15 (2), 133–144

Barrilleaux, J., 2001 3D user interfaces with Java 3D.Greenwich, CT, USA: Manning Publications Co.Boerma, J.R and Kals, H.J.J., 1989 Fixture design withFIXES: the automated selection of positioning, clampingand support features for prismatic parts Annals of CIRP,

GE Fanuc, 2008 [online] Available from: http://www.gefanuc.com/as_en/products_solutions/hmi_scada/index.html

IEC 61499-1, 2005 Function Blocks – Part 1: Architecture,International Electrotechnical Commission, Switzerland.Mazak, 1998 [online] Available from: http://www.mazak.jp/english/cyber/outline/ index.html

MDSI, 2008 [online] Available from: http://www.mdsi2.

Rong, Y., Liu, X., Zhou, J., and Wen, A., 1997

Computer-aided setup planning and fixture design International

Journal of Intelligent Automation and Soft Computing, 3

(3), 191–206

Smith, C.S and Wright, P.K., 1996 CyberCut: a World

Wide Web based design-to-fabrication tool Journal of

Manufacturing Systems, 15 (6), 432–442

van den Broecke, J., 2000 Pushlets, Part 1: send events fromServlets to DHTML client browsers JavaWorld, March.Wang, L., Cai, N., Feng, H.-Y., and Liu, Z., 2006 Enrichedmachining feature based reasoning for generic machiningprocess sequencing International Journal of ProductionResearch, 44 (8), 1479–1501

Wang, L., Feng, H.-Y., and Cai, N., 2003 Architecturedesign for distributed process planning Journal ofManufacturing Systems, 22 (2), 99–115

Waurzyniak, P., 2001 Electronic intelligence in ing Manufacturing Engineering, (3), 44–67

Bremen Institute of Industrial Technology and Applied Work Science at the University of Bremen (BIBA), Hochschulring 20,

28359, Bremen, Germany;bSchool of Engineering and Science, Jacobs University, Campus Ring 1, 28759 Bremen, Germany;

c

International Graduate School for Dynamics in Logistics (IGS), Hochschulring 20, 28359 Bremen, Germany

(Received 3 March 2010; final version received 21 June 2010)Bottlenecks, as the key ingredients for improving the performances of the production networks, have beenprofoundly studied However, the major definitions of bottlenecks are derived in terms of the throughput and based

on the theory of constraints (TOC) Moreover, before the specific measures can be applied on them, it is notstraightforward to localise dynamic bottlenecks due to their complex dynamic characteristics Distinguishing fromthe traditional view at the bottlenecks, this article therefore develops a systematic and comprehensive definition ofdynamic bottlenecks of the production networks based on both the TOC and the bottleneck-oriented logisticanalysis Afterwards, the defined dynamic bottlenecks are modelled by means of discrete simulation using practicaldata, aiming at visualising them in the production network By applying the logistic operating curves, the practicalapplication of the proposed research and its procedures is discussed as well

Keywords: dynamic bottlenecks; production networks; discrete simulation, TOC, BOLA, LOCs

Introduction

In the past decades, a considerable amount of literature

has been published on bottlenecks The generalizability

of the researches on this issue follows the theory of

constraints (TOC) presented by Goldratt (1990) The

approach developed from Optimised Production

Tech-nology (OPT) that is now more commonly known as the

Drum-Buffer-Rope approach In the view of Goldratt

(1993), the goal of a manufacturing organisation is to

make money To make as much money as possible,

manufacturing enterprises must strive to increase

throughput while minimising operating expense and

inventory Because the latter two cannot be reduced

endlessly (approach to zero), the main avenue or

objective of developing a manufacturing system is

maximising the throughput Thus, Goldratt described

the throughput limitation of a manufacturing system as

bottleneck and attempted to maximise the system

throughput by locating bottleneck (e.g bottleneck

machine) and utilising it as much as possible

Although Goldratt mentioned that there are also

some competitive objectives (e.g high due-date

per-formance, short quoted throughput time) might be

thought instead of the high throughput, maximising

throughout is only taken into account in his research

because the contributions of the competitive objectives

to making money is intangibles and impossible to be

accounted by the financial systems (Goldratt and Fox1986) However, to survive in the globalised market,nowadays manufacturing enterprises have to distin-guish themselves from their competitors by not onlyproducing quality products at high throughput butalso especially with a superior logistic performance(e.g high delivery reliability and short delivery time) It

is thus far from enough to describe the bottleneck ofthe manufacturing system only in terms of thethroughput Accordingly, there is a great necessity toidentify bottlenecks from a logistics perspective and toimplement specific measures on them, as so to improvelogistic performance and reduce logistic cost

Therefore, the bottleneck-oriented logistic analysis(BOLA), as an ongoing control method designed forlogistically evaluating and improving productionprocess, is developed by Wiendahl and Nyhuis(1993), Wiendahl and Mo¨ller (1995), Wiendahl andNyhuis (1998) and Nyhuis and Wiendahl (1999) Based

on the BOLA, Windt (2001) concentrated on tions of different bottlenecks types and developed abottlenecks-oriented subcontracting (BOS) method forproduction networks In the BOLA, based on thesystematic analysis of manufacturing process data, thelogistic objective-oriented bottleneck workstations,which including throughput time, schedule adherence,utilisation and work in process (WIP) orientedbottleneck workstation, can be localised in the

defini-*Corresponding author Email: liu@biba.uni-bremen.de

ISSN 0951-192X print/ISSN 1362-3052 online

Ó 2011 Taylor & Francis

DOI: 10.1080/0951192X.2010.511655

observation period By using the logistic operating

curves (LOCs) (Nyhuis and Wiendahl 1999, Nyhuis

and Vogell 2006), the logistic potential of the

bottle-neck workstation can be quantified through the logistic

positioning (Nyhuis 2007) The appropriate measures

can then be applied to the localised bottleneck

work-station to develop its logistic potential In this way, the

performance of the manufacturing system can be

enhanced by developing the existing logistic potential

of the bottleneck workstation For example to achieve

the short delivery time, the throughput time-oriented

bottleneck workstation is localised, and the target

throughput time of this workstation is derived at first

The measures (e.g temporarily increasing capacity)

can afterwards be applied to the bottleneck

work-station until its target throughput time is reached

However, the BOLA and BOS as well as the OPT

also suffers a number of limitations Among them, the

common and prominent one is that the investigated

bottlenecks are derived based on the statistical analysis

of manufacturing process data, thus they are called as

static bottlenecks for a certain observation period

Because of the sequence of random events (e.g

equipment failures, variations in demand patterns,

unsatisfied raw material delivery) as well as the gradual

change of the manufacturing systems, the bottlenecks

are not static but dynamic, which stay not still but

rather ‘shift’ or ‘wander’ in both time and space,

so-called bottlenecks shifting or wandering (Stephen and

Arnold 1995, Za¨pfel and Piekarz 1998) Consequently,

implementing the measures on the static bottlenecks

does not always yield ideal results The most likely

causes of this problem are on one hand, the measures

are sometimes applied to the mislocated bottlenecks in

the whole observation period; on the other hand,

implementing the measures on bottlenecks leads to

bottlenecks shifting, which again causes that the

measures are applied to the mislocated bottlenecks

Because consideration of dynamic bottlenecks

plays a decisive role at all the levels of

decision-makings in production practice (e.g production

planning, production scheduling and production

con-trol) and in decision-making for strategic investments

(e.g new facility purchases), there has been an

increasing amount of literature on identifying the

dynamic bottleneck (e.g Chiang and Kuo 2000, Pollett

2000, Delp et al 2003, Roser and Nakano 2003, Wang

et al 2005, Yan 2006, Yan et al 2009) Nonetheless,

these researches are barley focused on detecting

the dynamic bottleneck, which is defined in terms of

the throughput under the philosophy of the TOC, the

other competitive objective-oriented dynamic

bottle-necks have not been investigated thoroughly

To cope with the discussed limitations of previous

studies, this article develops a systematic and

comprehensive definition of dynamic bottlenecks in theproduction networks with regards to multiple objectives[i.e throughput time, relative lateness, schedule relia-bility, WIP, utilisation loss (UL), throughput] andproduction resources (i.e workstation and productionsegment) By modelling the predefined dynamic bottle-necks based on a practical case, they are concretelyvisualised and their complex dynamic characteristics aswell as clear differences from static bottlenecks aredemonstrated Aiming at supporting improvement ofexisting production processes with minimal efforts, apractical application of this study and its recommendedprocedures are developed to present the specific causes

of problems in the form of cause-and-effect ships Distinguishing from the existing approaches, theproposed method expands the application range fromone-time achievement of a single objective to simulta-neous accomplishment of multiple objectives for net-worked manufacturing enterprises

relation-The rest of this article is organised as follows relation-Thefollowing section systematically defines the dynamicbottlenecks of production networks After introducing

a discrete simulation model established by usingpractical data from a German hanger manufacturer,the dynamic bottlenecks are visualised and theircomplex dynamics are presented in ‘Modelling dy-namic bottlenecks’ section ‘Application’ section dis-cusses the main concept and proposed procedures forpractical application of configuring and modelling thedynamic bottlenecks Finally, major advantages anddisadvantages of the proposed research are sum-marised in the last section

Dynamic bottlenecks in production networksFrom the logistic point of view, manufacturing enter-prises, especially in the make-to-order (MTO) environ-ment (Kuroda and Takeda 1998), strive to accomplishthe different internal logistic objectives (Figure 1) in themanufacturing area, so as to achieve high logisticperformance or reduce logistic cost As noted already,the dynamic bottlenecks should be thus identified withrespect to not only the high throughput but also thecompetitive objectives (i.e internal logistic objectives).Because output lateness can be indicated by both relativelateness and schedule reliability, we only focus on theinternal logistic objectives, except for the output lateness.Therefore, the dynamic bottlenecks in this article arerespectively defined regarding to throughput time,relative lateness, schedule reliability, WIP and utilisation

as well as the system throughput (i.e output rate).Furthermore, when the production network, inwhich various products must be processed on multipleworkstations located in different production segments,

is considered as a whole, any components or

must be derived from the definition of bottleneck

workstations, we will first define the bottleneck

work-stations as follows

Bottleneck workstations

Throughput time bottleneck

To obtain short delivery time, manufacturing

enter-prises usually strive to reduce the throughput times of

workstations As shown in Figure 2, the sum of all the

order throughput times corresponds to the sum of the

operation throughput times of each workstation

[Equation (1)] (Nyhuis and Wiendahl 2003),

where TTPorder,kis the order throughput time for the

order k (hr); TTPiis the throughput time per operation

(hr); m is the number of orders; j is the number of

workstation; n is the number of accomplished

opera-tions per workstation and w is the number of

workstations

The relative proportion of throughput time (TTPrp)

thus directly describes the degree to which the

individual workstations contribute to the order

is the number of workstations

When this calculation is completed for each of theworkstations they can be ranked according to whichones measures for reducing the throughput time should

be primarily implemented on Therefore, the put time bottleneck workstation (BNTTPws) can bedescribed as the workstation j with the maximumTTPrpas following:

through-BNTTPWS¼ fjjTTPrp;j

¼ max ðTTPrp;1;TTPrp;2; ;TTPrp;wÞg: ð3Þ

Relative lateness bottleneck

In the manufacturing area, the output schedulelateness of an operation can be caused by bothinput lateness and the relative lateness The latter isthe result of the difference between the actualthroughput time and the target throughput time asshown in Figure 3

Therefore, the relative lateness can be used toidentify whether the output schedule situation wor-sened or improved compared with the input situation.The sum of all the order relative lateness corresponds

to the sum of the operation relative lateness of eachworkstation, as described:

Figure 1 Logistic objectives (Lo¨dding 2005)

Figure 2 Order throughput time and workstationthroughput time

where RLorder,kis the relative lateness for the order k

(hr); RLiis the relative lateness per operation (hr); m is

the number of orders; j is the number of workstation; n

is the number of accomplished operations per

work-station and w is the number of workwork-stations

Consequently, the relative proportion of relative

lateness (RLrp) describes the degree to which the

individual workstations contribute to the order relative

lateness and can be defined as (Windt 2001):

RLrp¼

Pn j¼1jRLij

Pw j¼1

Pn i¼1jRLij 100 ð5Þwhere RLrp is the relative proportion of relative

lateness (%); RLiis the relative lateness per operation

(hr); n is the number of accomplished operations and w

is the number of workstations

When the relative proportion of relative lateness

for each workstation is available, all the workstations

can be ranked according to the RLrpvalues And, the

relative lateness bottleneck workstation (BNRLWS)

can be described as the workstation j with the

maximum RLrpas follows:

BNRLWS¼ fjjRLrp;j

¼ max ðRLrp;1;RLrp;2; ;RLrp;wÞg: ð6Þ

Schedule reliability (ReS) bottleneck

In the production area, high levels of schedule

reliability of the workstations are a prerequisite for

accomplishing high level of delivery reliability

(Lo¨dding et al 2002) Under the conditions that

throughput time is normally distributed and orders

are completed following the dispatching sequence rule

first-in-first-out (FIFO), the schedule reliability of the

workstation can be calculated like the probabilities of anormal distributed random variable WIP(t) (Yu 2001).The calculation is based on the distribution functionf(u) for the standard deviation (Papula 1994) and iswritten as (Adapted from Yu 2001):

Re SðWIPðtÞÞ ¼ f UB TIOmðWIPðtÞÞ

is lower bound (SCD); UB is upper bound (SCD);TIOm(WIP(t)) is the mean inter-operation time (SCD)and TIOs(t) is the standard deviation of interoperationtime (SCD)

The schedule reliability bottleneck workstation(BNRe SWS) can therefore be determined as theworkstation j with the minimum schedule reliability

WIPðtÞ ¼ IWIP  INðtÞ  OUTðtÞ ð9Þwhere WIP(t)is the WIP level of workstation (hr);IWIP is the initial WIP level; IN(t) is the input(cumulative work content of the incoming operations)(hr) and OUT(t) is the output (cumulative workcontent of the outgoing operations) (hr)

So, the WIP bottleneck workstation (BNWIPWS)can be determined as the workstation j with themaximum WIP:

BNWIPWS¼ fjjWIPj

¼ max ðWIP1;WIP2; ;WIPwÞg: ð10ÞFigure 3 Relative lateness for an operation (Nyhuis and

Wiendahl 2003)

definition cannot be used.

In a high-wage country such as Germany, the

available capacity mainly depends on not equipment

but operator capacity, the latter can be used to express

the utilisation by introducing the relative work in

process (WIPrel), which reflects the scope of the

available operator capacity and can be measured

comparably quickly with changes in the load situation

According to the normalised LOCs, there will not be

significant UL under the condition of WIPrel 4 250%

(Nyhuis and Wiendahl 2008) The UL bottleneck

workstation (BNULWS) can be therefore determined

as the workstation j with the maximum UL:

BNULWS¼ fjjULj¼ max ðUL1;UL2; ;ULwÞg:

ð11Þ

In Equation (11), the UL of each workstation is

greater than zero and can be derived based on the

WIPrel(t) using Equation (12):

ULðtÞ ¼ 2:5  WIPrelðtÞ: ð12Þ

The WIPrel(t) level of each workstation can be

calculated by applying Equations (13) and (14)

according to Nyhuis and Wiendahl (2008):

WIPrelðtÞ ¼ WIPmðtÞ

WIPIminðtÞ ð13Þwhere WIPrel(t) is relative WIP level (%); WIPm(t) is

mean WIP level (hr) and WIPImin is ideal minimum

WIP level (hr)

WIPIminðtÞ ¼

Pn i¼1ðWCi WCiÞ

Pn i¼1WCi

ð14Þ

where WIPImin is the ideal minimum WIP level (hr);

WCiis the work content of operation per lot (hr); n is

the number of accomplished operations

Throughput bottleneck

Apart from the internal logistic objectives,

manufac-turing enterprises generally strive to maximise the

system throughput (i.e output rate) so as to make

time belongs to the first category When measuring theaverage waiting time, the machine with the longestaverage waiting time is considered to be the through-put bottleneck (see e.g Pollett 2000) Regarding to theLittle’s law, measurement of average queue length isalso within this category This method is suitable foranalysing production networks with unlimited inter-medial buffers For systems containing only limitedbuffers and systems without buffers, it is not a suitablechoice In the second category, the throughput bottle-neck is detected by measuring average utilisation(workload) and the machine with the largest busy-to-idle ratio is considered as the throughput (see e.g Lawand Kelton 1991) As more than one machine mayhave a similar workload, the difference between theutilisations of the machines may be very small.Although this method is easy to automate, it mayresult in multiple bottlenecks On the other hand, it isalso difficult to identify the dynamic bottlenecks asdiscussed in last section Another way to identify thethroughput bottleneck is to find the machine whosethroughput mostly affects the overall system through-put, i.e measuring the sensitivity (see e.g Chiang andKuo 2000) The sensitivity of the system performance

to the perturbation of machine parameters is used asthe measurement Except for the earlier describedmethods, measuring the active duration was developed

by Roser and Nakano (2003) When measuring theactive duration, the machine with the longest averageactive period is recognised as the bottleneck The activestate of machine is different from traditional busyconcept All activities towards improving the systemthroughput including repair and service states areactive states With simulation results in a productionline, they argued that the proposed method can moreaccurately detect the dynamic bottleneck based on thesensitivity definition

These throughput bottleneck identification ods have also some disadvantages, either too complexfor the practical applications (measuring the sensitivityand active duration) or the difficulty in identifying thedynamic bottleneck (e.g measuring the averageutilisation) By contrast, the method of measuringqueue length is easy to implement and can effectivelyidentify the dynamic bottleneck if the system containsunlimited size buffers This method is thus applied to

meth-our study, and the workstation with the longest queue

(i.e the highest WIP level) is defined as the throughput

bottleneck as we discussed in ‘Work in process

bottleneck’ section

Bottleneck production segments

Apart from the bottleneck workstations, in the

production network the bottleneck production

seg-ments, which consist of several workstations, can also

be determined on the basis of the definition of the

bottleneck workstations Because the modelling of the

dynamic bottlenecks is based on a practical case, in this

article, we only investigate the production network in

which the individual production segments consist of

parallel workstations

In the case that each production segment of the

production network consists of several parallel

work-stations, in which incoming orders are processed

following the uniform distribution, (i.e the parallel

workstations in a production segment have the same

possibility to produce incoming orders), the parallel

workstations are less dependent on each other and

make the equal contribution to the bottlenecking

degree of the production segment Therefore, the

bottleneck indicators of production segments can be

determined by the mean values of the bottleneck

workstations’ indicators by using Equations (15)–(19),

in which the variable wps represents the number of

workstation in a production segment

TTPrp;ps¼

Pwpsj¼1TTPrp;j

wps

ð15Þ

RLrp;ps¼

Pwpsj¼1RLrp;j

wps

After the bottleneck indicators of individual

production segments are derived, the bottleneck

production segments can be respectively determined

as the production segment i as follows:

¼ min ðReSps;1;ReSps;2; ;ReSps;nÞg ð22Þ

BNWIPps¼ BNTTps¼ fijWIPps;i

¼ max ðWIPps;1;WIPps;2; ;WIPps;nÞg ð23ÞBNULps¼ fijULps;i

(1) Production order data order number lot size start, begin and end of target date actual start, begin and end date routing plan

(2) Operation data workstation number operation number operation sequence number work contents

start, begin and end of target date actual start, begin and end date processing time and setup time(3) Workstation data

production segment number workstation number

shift calendarThe simulation model describes a partial productionnetwork of a German hanger manufacturer Theproduction network consists of four production seg-ments that are located at different places and representdifferent processing steps including manual turning(PS78), CNC turning (PS79), drilling (PS77) and CNCcenter (PS81) Each production segment works in thegiven shift calendar and consists of several parallelworkstations (Table 1)

The four production segments are connectedthrough the complex material flows including not

following the uniform distribution function within aproduction segment, and processed following theFIFO rule at each workstation Besides, regardingthe impacts of the schedule tolerance and the meanvalue of due date distribution on schedule reliability(Nyhuis and Vogell 2006), the schedule tolerance andthe mean value of due date distribution are respectivelyset at + 0.7 SCD and 0.8 SCD, the both of which arederived based on the statistical analysis of the averageplanned inter-operation times.

Visualising dynamic bottlenecksFigures 6–10 present that the dynamic bottlenecks shift

in time (x axis) and space (y axis) in a randomlyselected observation period (from 1 April 2003 to 30June 2003) The horizontal length of lines present thetime duration of the production segments and work-stations being bottlenecks The vertical changes of linesshow that the bottlenecks jump from one workstationand production segment to another

From the earlier figures, we can see that thedynamic bottlenecks dramatically shift in the produc-tion network, especially the throughput (or WIP)bottlenecks (Figure 9) For example, as shown inFigure 6, workstation 37703 was the throughput timebottleneck until 10 June, and then workstation 37701,

37703 and 37704 turned into the bottleneck alternately,and production segment 77 was always the throughputtime bottleneck in the entire observation period

In contrast, the static throughput time bottlenecks,both workstation and production segment, are derived

by using Equations (2), (3), (15) and (20) and based onstatistical simulation results collected in the observa-tion period (from 1 April 2003 to 30 June 2003) As can

be seen from Table 2, the workstation 37701 andproduction segment 77, respectively, have the max-imum relative proportion of throughput time and thuswere identified as the static bottlenecks By comparingthe locations of static and dynamic throughput timebottlenecks, a clear difference between the static anddynamic bottlenecks can be observed It indicates thesignificance of investigating and modelling the dy-namic bottlenecks as well

Moreover, the different objective-oriented necks might be overlapped For example, production

bottle-Table 1 Production facilities in the production network

segment 77 was throughput time, relative lateness,

throughput and WIP bottlenecks from 10 to 30 June

(Figures 6, 7 and 9) Most interestingly, the dynamic

bottlenecks’ shifting is in a total chaotic state and

might not be explained using fundamental knowledge

obtained from investigations on relative steady

pro-cesses For example, it is well known that for both

production segment and workstation, the high WIP

level inevitably results in the long throughput time in a

long-term observation period Hence, the throughput

time bottlenecks are supposed to be consistent with the

WIP bottlenecks, and the both bottlenecks should

synchronously shift to the same production resources

Nevertheless, the controversial phenomenon is

demon-strated by comparing the locations of the throughput

time and WIP bottlenecks as shown in Figures 6 and 9

In the observation period, the throughput time

bottleneck production segment is production segment

77 In contrast, the WIP bottleneck productionsegment jumps among the four production segmentsuntil about end of May and then locates at theproduction segment 77 As to the bottleneck work-station, workstation 37703 is the throughput timebottleneck until 10 June, but the WIP bottleneckworkstation jumps among all the workstations in thesame observation period

This phenomenon might result from complexdynamics of dynamic bottlenecks, which also couldlead to difficulties in practical applications of proposedresearch To overcome this problem, effective faultdetection and isolation approaches could be devel-oped For example, the complex dynamics of dynamicbottleneck was first demonstrated experimentally byScholz-Reiter et al (2009) In their study, a faultFigure 5 Diverse characteristics of incoming production orders

Figure 6 Throughput time bottlenecks Figure 7. Relative lateness bottlenecks.

detection and isolation approach was developed with

regards to the throughput time bottleneck workstation

To verify its effectiveness, the identified dynamic

bottleneck is considered in a release control process

The main idea is that aiming at decreasing the

throughput time of whole production system, the

developed release control mechanism offers the

custo-mer orders, which are not about to be processed by the

bottleneck, the priority to release at first, i.e the

throughput bottleneck workstation always refuse to

produce more until another workstation becomes the

bottleneck The simulation results indicates that the

release decisions made based on the identified dynamicbottleneck, combing with the application of the faultdetection and isolation approach, are able to reducethe throughput time of production system

ApplicationThe proposed research is mainly developed to supportthe improvement of existing production processes bylocalising and presenting the specific causes of pro-blems in the form of cause-and-effect relationships.Primary application objects could include middle- orlarge-scale production networks, which conduct man-ufacturing activities based on customer specificationsand there are frequent changes to the product design.Because of the complexity of these production net-works, it is not convenient to modify the existingmanufacturing processes and approaches or redesign awhole set of new ones Nonetheless, constant marketpressures push them to do so to continuously improvelogistic performance and/or reduce logistic cost It thusbecomes a crucial issue that achieving the goal withminimal efforts, and the proposed application of thisstudy comes forward Because of the diversity ofproduction networks caused by diverse products,different marketing strategies and changing customerdemands, software of production planning and control(PPC) systems, etc the proposed application should beadopted according to different circumstances inpractice In this article, we only focus on a funda-mental application of configuring and modellingdynamic bottlenecks with regards to the investigatedhanger manufacturer

Figure 8 Schedule reliability bottlenecks

Figure 9 Work in process and throughput bottlenecks

Figure 10 UL bottlenecks

To apply the proposed research to practice, the

manufacturing process data must be continuously

obtained by building interface between eMplant and

PPC systems, e.g the American Standard Code for

Information Interchange interface, which, as a

stan-dard interface, is installed in most PPC systems and

eMplant Moreover, employees need to be well

educated with regards to fundamental and systematic

theories of production logistics and PPC as well as

production processing designing, so as to assure an

effective and efficient application On this basis, the

objectives-oriented dynamic bottlenecks can be

loca-lised in real time, and their existing logistic potentials

for improvement can be derived with the integration of

the synchronous logistic positioning, as shown in

Figure 11 Afterwards, the appropriate measures can

be determined and applied to the determined dynamic

bottlenecks so that the production network

perfor-mance can be improved by developing the logistic

potentials of the bottlenecks For the practical

application, it is strongly recommended that the

following specific steps are followed

Choosing bottlenecks

First and foremost, the choice of bottlenecks has to be

oriented on the concrete analysis of the enterprise’s

goals Generally, an enterprise establishes not a single

but multiple goals (i.e short throughput time, low

relative lateness, high schedule reliability, low WIP

level, low unitisation loss and high output rate)

according to actual situation The bottlenecks neck workstations and bottleneck production seg-ments) can be then chosen in terms of the focusedobjectives Moreover, those objectives are alwaysconnected with each other and to some extent mightcontradict one another (Gutenberg 1951) In the casethat a bottleneck is identified as the bottleneck relevant

(bottle-to multi-objectives (i.e overlapped bottleneck), it isnecessary to rank the priority of each objective at first

so that the prioritised objective-oriented measures can

be chosen If both of them have the existing logisticpotentials for improvement, the corresponding mea-sures for shortening throughput time (e.g shifting loadlocation) and reducing relative lateness (e.g remainingthe WIP at the planned level) should be applied to thebottlenecks, respectively In the case that the bottle-necks are overlapped as described in ‘Visualisingdynamic bottlenecks’ section, the measures only forshortening throughput time are supposed to be applied

to the overlapped bottleneck production segment 77.For example, as noted already, the throughputbottleneck and WIP bottleneck are defined as theidentical production resources (i.e both workstationsand work systems) When a manufacturer establishesthe low WIP level and high output as the goals, aiming

Table 2 Static throughput time bottlenecks

Workstation no

Sum ofthroughputtimes (hr)

Relative proportion

of throughputtime (%)

Productionsegment no

at, on one hand, making money as much as possible,

on the other hand, saving the production costs

(removing production waste) to increase the value of

end products, both of the goals must be prioritised at

first The priority of the two goals directly determines

the corresponding measures If the high throughput is

established as the principal goal, the measures for fully

utilising the bottleneck [e.g remaining the buffer

inventory of the bottleneck at a safety WIP level, like

the Starvation Avoidance (Glassey and Resende 1988)]

should be used On the contrary, the WIP level of the

bottleneck should be reduced as much as possible so as

to reduce WIP of the whole production network, as

long as the bottleneck resources not interrupt the

material flow of the whole production network

Logistic positioning

After choosing the bottlenecks relevant to the existing

goals, the logistic positioning must be conducted on all

of the bottlenecks before the specific measures can

determined In the logistic positioning, the target

values for the logistic objectives are determined based

on the current manufacturing situations By using the

LOCs (Figure 12), it becomes clear whether or not the

target values are consistent and achievable, or if it is

necessary to develop additional logistic potential For

example if the actual throughput time level of the

throughput time bottleneck workstation is higher than

the target level, it implies that this bottleneck

work-station has the potential for reducing throughput time,

e.g by increasing its capacity or shifting load location

Otherwise, it is necessary to develop the new additionallogistic potential (e.g by facilitating overlappedmanufacturing)

Apart from the logistic positing in terms ofthroughput time, relative lateness and schedule relia-bility, it is also necessary to independently conduct thelogistic positioning regarding the utilisation cost andthe WIP cost Although this task cannot be immedi-ately accomplished by using the production costoperating curve, the both types of costs can be depicted

by decomposing the total costs according to Jainczyk(1994), Großklaus (1996) and Kerner (2002) Becausethe total costs, as the function of output rate and WIP,

Figure 12 Trends of the WIP Dependent TTP, RL, Res,Costs and Output Rate (Nyhuis and Wiendahl 2008).Figure 11 Practical application of modelling dynamic bottlenecks

are determined by the sum of production cost,

processing cost, WIP cost and setup cost, each of the

four type of costs can be descript as the function of

WIP after the output rate operating curve is derived

When the utilisation cost is approximated to the

difference between the total costs and the WIP cost,

the utilisation cost and the WIP cost can be

independently depicted by the curves, which can be

further applied in the logistic positioning

As noted already, the logistic positioning must be

conducted on not only the bottleneck workstations but

also the bottleneck production segments However, the

LOCs can only be applied in positioning the individual

workstations, the quantifying the logistic potentials of

the bottleneck production segments is hitherto

im-possible except for using the Manufacturing System

Operating Curves (Schneider 2004), which however

can be applied in the logistic positioning only

regarding to throughput time and output rate

Here, it should be noted that the main concept of

the proposed application, similar to the TOCs and the

BOLAs, is to improve the performance of a whole (i.e

a production network) by improving the performance

of its parts (i.e bottlenecks) However, according to

the dialectical whole and part (Bertell 2003), the

individual workstations can be considered as not

only the parts of the production network but also the

parts of the relevant production segment when we

consider a production network and a production

segment as a whole, respectively Thus, the logistic

potentials of a bottleneck production segment can be

developed by implementing the measures on its

bottleneck workstation until the desired performance

of the production network is reached Accordingly, it is

only need to position the bottleneck workstation of

bottleneck production segment, instead of directly

positioning the bottleneck production segment For

example production segment 77 was throughput time

bottleneck from 31 May to 10 June (Figure 6)

Meanwhile, workstation 37701, which with the

max-imum TTPrpwithin production segment 77 and can be

identified as the bottleneck workstation of bottleneck

production segment 77 (Figure 13) Therefore, the

bottleneck workstation of bottleneck production

seg-ment (i.e workstation 37701) as well as the bottleneck

workstation of production network (i.e workstation

37703 as shown in Figure 6) is able to be positioned by

using the LOCs

Determining appropriate measures

After choosing bottlenecks and logistic positioning, a

wide range of possible measures can be developed and

implemented on the bottleneck workstations of

pro-duction network and the bottleneck workstations of

bottleneck production segments For example, thepossible measures to reduce throughput time mightinclude (1) production process designing (e.g increas-ing capacity flexibility, implementing new manufactur-ing technologies); (2) production planning (e.g lotsizing, scheduling, loading and capacity planning); (3)production control (e.g increasing capacity, shiftingworkload, reducing or harmonising work content,changing processing sequences etc.); and (4) subcon-tracting (e.g buying extra production capacity frompartner companies) [the details can be found in Windt(2001)]

Nevertheless, it should be noticed that not all of thepossible measures but only few of them are preferable

in the practice, and even the favour measures are notalways preferred because of the changes of manufac-turing situation and customers’ demand Therefore, todetermine the appropriate measures, employees firstrequire a systematic introduction into the fundamen-tals of production processing designing and PPC,preferably in conjunction with training, in whichconcrete examples from the company can be observed

Summary

In this article, we developed a systematic andcomprehensive definition of the dynamic bottlenecksfrom a logistics perspective By the means of modelling

of the dynamic bottlenecks, not only the dynamicbottlenecks are visualised, but also their dynamiccharacteristics are demonstrated Moreover, a practicalguidance for utilising the proposed research was alsoprovided In contrast to the BOLA, the proposedapplication concepts are able to not only overcome themain drawback of the application of the BOLA butalso expand the application range from one-timeachievement of a single logistic objective to simulta-neous trade-off multiple logistic objectives and fromthe individual production systems to the productionnetworks

Figure 13 WIP trend of bottleneck production segment’sbottleneck workstation

sures and evaluating system performance according to

the current manufacturing situation and market’s

demand Future work will focus on realising the

proposed application within the context of production

control Aiming at trading-off multiple logistic

objec-tives, a hybrid production control system, named as a

Dynamic Bottleneck-oriented Manufacturing Control,

will be developed on the basis of the modelling of the

dynamic bottlenecks

Acknowledgements

This work was supported by the International Graduate

School (IGS) for dynamics in logistics as well as the

Intelligent Production and Logistic Systems department at

Bremen University This article was presented in the

DET2009 conference held at the University of Hong Kong

during 14–16 December 2009 and included in the conference

proceedings published by Springer

References

Bertell, O., 2003 Dance of the dialectic: steps in Marx’s

method Urbana, IL: University of Illinois Press, 36–51

Delp, D., Si, J.Y., and Wang, H., 2003 A dynamic system

regulation measure for increasing effective capacity: the

X- factor theory In: IEEE Advanced Manufacturing

Conference, Munich, Germany, 81–88

Chiang, S.Y and Kuo, C.T., 2000 DT-bottlenecks in serial

production lines: theory and application IEEE

Transc-tions on Robotics and Automation, 16, 567–580

Großklaus, A., 1996 Ablauforientierte Productionslogistik –

eine modellbasierte Analyse, Dt University - Verlag,

Gabler – Verlag, Wiesbaden

Goldratt, E.M and Fox, R.E., 1986 The race

Croton-on-Hudson, NY: North River Press, 18–72

Goldratt, E.M., 1990 Theory of constraints

Croton-on-Hudson, NY: North River Press

Goldratt, E.M., 1993 The goal 2nd revised ed Great

Barrington, MA: North River Press

Gutenberg, E., 1951 Grundlagen der Betriebswirtschaftslehre,

Band 1: Die Produktion Berlin, Heidelberg, NY:

Springer-Verlag

Glassey, C.R and Resende, M.G.C., 1988 Closed-loop

job release control for VLSI circuit manufacturing

IEEE Transactions on Semiconductor Manufacturing,

11, 36–46

Jainczyk, M., 1994 Logistikorientierte Bewertung von

Invesstitionsvorhaben im Productionsbereich Munich,

Germany: PSK-Fachtagung, gfmt

Kuroda, M and Takeda, K., 1998 General structure and

characteristics of quick response production system

Computer and Industrial Engineering, 35, 395–398

Nyhuis, P and Wiendahl, H.-P., 1999 LogistischeKennlinien: Grundlagen, Werkzeuge und Anwendungen.1st ed Berlin, Heidelberg, NY: Springer-Verlag, 191–238

Nyhuis, P and Wiendahl, H.-P., 2003 LogistischeKennlinien: Grundlagen, Werkzeuge und Anwendungen.2nd ed Berlin, Heidelberg, NY: Springer-Verlag, 199.Nyhuis, P and Vogell, M., 2006 Adaptation oflogistic operating curves to one-piece flow processes.Journal of Productivity and Performance, 55, 284–299

Nyhuis, P., 2007 Practical applications of logistic operatingcurves Annals of the CIRP, 56, 483–486

Nyhuis, P and Wiendahl, H.-P., 2008 Fundamentals ofproduction logistics: theory, tools and applications Berlin,Heidelberg, NY: Springer, 59–178

Papula, L., 1994 Mathematik fu¨r Ingenieure undNaturwissenschafttler Braunschweig, Wiesbaden: View-eg-Verlag

Pollett, P.K., 2000 Modelling congestion in closed queuingnetworks International Transactions in Operations Re-search, 7, 319–330

Roser, C and Nakano, M., 2003 Simulation test bed formanufacturing analysis: comparison of bottleneck detec-tion methods for AGV systems In: 2003 WinterSimulation Conference, New Orleans, Louisiana, USA,1192–1198

Stephen, R and Arnold, H., 1995 Economic analysis ofproduction bottlenecks Mathematical Problems in En-gineering, 1, 341–363

Schneider, M., 2004 Logistische nien Du¨sseldorf, Germany: Fortschritt-Berichte, VDI.Scholz-Reiter, B., Windt, K., and Liu, H., 2009 Apreliminary investigation on a bottleneck concept inmanufacturing control In: Second international con-ference on dynamics in logistics, Bremen, Germany:Springer-Verlag

Fertigungsbereichskennli-Windt, K., 2001 Engpaßorientierte Fremdvergabe in tionsnetzwerken Du¨sseldorf, Germany: Fortschritt-Berichte VDI, 59–65

Produk-Wiendahl, H.-P and Nyhuis, P., 1993 Die logistischeBetribskennlinie-ein neuer Ansatz zur Beherrschung derProductionslogistik RKW-Handbuch Logistik RKWEdition Berlin: Erich Schmidt Verlag

Wiendahl, H.-P., 1995 Load-oriented manufacturing control.Hanover, Germany: Springer-Verlag

Wiendahl, H.-P and Mo¨ller, J., 1995 EngpaßorienttierteLogistikanalyse auf der Basis von Betriebskennlien IoManagement Zeitschrift, 64, 27–33

Wiendahl, H.-P and Nyhuis, P., 1998 EngpaßorienttierteLogistikanalyse-Mothoden zur kurzfristigen Leistungstei-gerung in Productionsprozessen Munich, Germany:Transfer Centrum GmbH

Wang, Y., Zhao, Q., and Zheng, D., 2005 Bottlenecks inproduction networks: an overview Journal of SystemsScience and Systems Engineering, 14, 347–363

Yu, K.-W., 2001 Terminkennlinie – Eine

Beschreibungs-methodik fu¨r die Terminabweichung im

Produktions-prozess Du¨sseldorf, Germany: Fortschritt-Berichte,

71

Yan, H.S., An, Y.W., and Shi, W.W., 2009 A new

bottleneck detecting approach to productivity

improve-ment of knowledgeable manufacturing system Journal of

Intelligent Manufacturing,

doi:10.1007/s10845–009-0244-3

Yan, H.S., 2006 A new complicated-knowledge tion approach based on knowledge meshes IEEETransactions on Knowledge and Data Engineering, 18,47–62

representa-Za¨pfel, G and Piekarz, B., 1998 Regelkreisbasiertes supplychain controlling In: H Wildemann, ed Innovationen inder Produktionswirtschaft: Produkte, Prozesse, Planungund Steuerung Mu¨nchen, Germany: TCW Transfer-Centrum-Verlag, 63–65

Department of Mechanical Engineering, The University of Hong Kong, Hong Kong, PR China

(Received 9 March 2010; final version received 5 July 2010)Multi-material products are getting increasingly popular in recent years Yet no systems have been developed tosupport the design of such products In this article, a haptic-based method for designing multi-material products isproposed The proposed method consists of three main parts: haptic painting/marking, boundary smoothing/fitting,and volume decomposition A prototype system based on the three parts has been implemented using a haptic inputdevice The haptic device provides an intuitive user interface for quick volume mark-up in a multi-material product

by direct mesh painting Each painted region represents a material volume whose boundary can be automaticallytraced The boundary points are then used as control points for Catmull-Rom spline fitting To constrain a spline tothe original mesh surface, the spline is projected onto the mesh Now, using the boundary splines, volumedecomposition can be done automatically when needed Each decomposed volume can be assigned a differentmaterial or colour The numerous iterations of volume mark-up and decomposition in the early stage of multi-material product design can now be made easy and effective using the proposed method

Keywords: multi-material product design; haptic painting; boundary smoothing

1 Introduction

In recent years, multi-material products are becoming

more and more popular in the market place because of

their superiority to traditional homogenous

counter-parts in the following aspects (Goodship and Love

2002, Kromm et al 2003, Li and Gupta 2004):

Better Safety: for a product that needs a grip or

handle, a soft material part provides improved

grip in dry and wet environments, vibration

damping and electric insulation

Ergonomics: increased comfort and visual

pleasure

Enhanced product performance: water-resistant

seals, sound absorption, electric insulation and

vibration damping

Material optimisation: material composition can

be optimised to comply with the targeted

applications

Aesthetics: multiple colours, soft-touch effects

and over-moulded intricate shapes present

at-tractive decorative effects

Reduced cost: due to the in-mould assembly and

in-mould decoration features of multi-material

moulding technologies, less factory floor space

and shorter manufacturing cycle time can be

achieved due to reduced assembly processes

A multi-material product is normally manufactured

by a single moulding setup, e.g over-moulding Byremoving the need for assembly processes or perhaps afurther finishing stage, multi-material over-moulding canoffer a considerable unit cost reduction Consider forexample the case of a toothbrush (Figure 1(a)) Applica-tion of multi-material moulding of this product hastransformed this simple and standardised design Theproper combinations of hard and soft touch materialsand a myriad of multiple colour combinations have filledsupermarkets shelves all over the world Over-mouldingtechnologies and the resulting design freedom have againhelped the dominance of international brands in tooth-brush business In the future, product design composed ofhybrids of many materials such as plastics, metals andceramics could be manufactured as a single component,

in which the individual material properties are optimised.This kind of manufacturing technology advancementdictates the need of new product design methods.The emergency of multi-material manufacturingtechnologies demands a more flexible and intuitivedesign process There were frequent curiosities abouthow a multi-material product is designed For in-stance, for samples shown in Figure 1: Is the individualmaterial lump designed separately and assembledafterwards, or the whole part designed first and

*Corresponding author Email: yhchen@hku.hk

ISSN 0951-192X print/ISSN 1362-3052 online

Ó 2011 Taylor & Francis

DOI: 10.1080/0951192X.2010.511656

decomposed afterwards? These questions are natural

as traditional design process requires the design of

individual parts first and assemble them to form an

integral product This process has obvious drawbacks

in designing the multi-materials products as in Figure 1,

because in designing each individual material lump, the

designer has no global view of the product itself If the

design and decomposition process is used, current

computer-aided design (CAD) software does not

support easy and quick part mark-up and

decomposi-tion To the best knowledge of the authors, there is so

far no report as to the easy and effective design of

multi-material products The main purpose of this

research is therefore to propose a practical and efficient

method for computer-aided multi-materials product

design

For a soft-touch multi-material product, it usually

contains two main parts: a rigid part providing the

required stiffness and strength, and a soft part offering

better aesthetics and improved tactile or compliant

properties The combined rigid part and soft part can

also optimise some mechanical properties of a product

such as stiffness As a result, the final products appear

more attractive to potential consumers

In the product development period, the use of

different materials in different areas of an object allows

designers to fulfil various technical and aestheticrequirements in a single object This expands thedesign space and allows designers to realise conceptsthat are not possible with single material objects.However, traditional modelling software provideslimited functions to support this kind of process Infact, different material parts still need to be treatedseparately like an assembly Due to the differentaesthetical and functional requirements, the combina-tion of materials usually does not appear in regularshape, which may largely increase the difficulty andwork load especially to a designer who is not an expert

of the CAD software For example, a simple material toothbrush design (Figure 2(a)) may needindividual consideration of four different parts Tosolve this problem, several methods were developedparticularly for multi-material part design

multi-An early attempt to multi-material part design wasreported by Kumar and Dutta (1997) by defining newmodelling operations and computer representation.The operations are limited to simple addition indimension to the solid model A systematic designmethod for multi-material product was also proposedfrom material combination to prototype manufactur-ing (Li and Gupta 2004) To overcome the restriction

in the design of mechanical components with regular

Figure 1 Sample designs of multi-material products: (a) A multi-material toothbrush; (b) A multi-material water-proof camera;(c) A multi-material razor

Figure 2 A multi-material product designed as an assembly: (a) A multi-material toothbrush assembly; (b) A waterproofcamera assembly; (c) A razor assembly

colour Because a haptic interface is used for mark-up

material regions, the system will be easy to learn and

use It is possible to evaluate many different mark-ups

before actual volume decomposition is done

A multi-material product normally has different

colours for its individual material volume It is

therefore intuitive to mark-up the individual lumps

using painting Painting is a very common tool for

computer graphics and has been well studied for

painting on 2D surfaces However, it is

time-consum-ing for a designer to paint on complex 3D surfaces

using traditional CAD software and mouse input

The rapid development of computer haptics

pro-vides an opportunity to exploit the potential of

painting in the multi-material product design process

Using 3D haptic painting, any idea about material

mark-up can be brought to the screen in a few minutes

for visual evaluation Designers are free to try various

combinations of mark-ups on a 3D part until the best

one is found Since different colour is used to represent

different material lumps in the product design, the

model needs to be decomposed according to the

painted regions To achieve a better decomposition,

an efficient algorithm to trace and smooth the

boundary curves of the painted regions has been

developed The boundary curves are used to generate

cutting surfaces for volume decomposition

2 Related works

There were only preliminary works on geometric

algorithms for automated design of simple multi-shot

moulds and related part (Kumar and Dutta 1997, Li

and Gupta 2004) To the best knowledge of the

authors, there is so far no report as to the effective

design method of multi-material products Currently

available software tools for product design cannot

assist the design of multi-material product efficiently

This article proposes a haptic-based top-down design

process First, the part geometry of a design can be

created in a traditional CAD system or using haptic

sculpting techniques To specify different material

regions intuitively, the proposed method uses

inter-active haptic painting (Johnson et al 1999, Gregory

et al 2002) The design is now visually evaluated in

terms of both colour assignment and the location of

surface to cut the part into individual lumps Now thedesigner can assign material to each of the lumps.Computer-aided painting has been researchedextensively in previous studies Hanrahan and Haeberli(1990) proposed a 3D painting system, using a mouse

to control the painting brush to achieve the WYG (What You See Is What You Get) fashion In

WYSI-1994, Massie and his colleagues developed the ToM–haptic interface (Massie and Salisbury 1994),which is used in our system Since then, lots of workhas been done to solve the dynamics of colliding bodies

PHAN-in haptic force modellPHAN-ing PHAN-includPHAN-ing the penalty-basedmethod (Mark et al 1996), which may suffer from astrong force discontinuity and constraint-based ‘god-object’ method (Zilles and Salisbury 1995), in whichthe movement of the god-object was constrained to theobject’s surface Some recent 3D painting systemsbased on haptic interface have been reported (Johnson

et al 1999, Gregory et al 2002) However, in thesesystems, texture mapping and parameterisation remain

a big problem to the developer The Chameleonpainting system (Igarashi and Cosgrove 2001) auto-matically builds a texture map and parameterisationduring interactive painting instead of traditionalpredefined UV-mapping Adams et al (Adam et al.2004) provide an alternative solution by representingthe object surface as collections of point samples Allthe previous painting methods can be treated as forvery precise and smooth paining, and therefore goodfor the detailed design In this article, haptic painting isimplemented directly on the surface mesh Eventhough the painting effect is not as good and accurate,

it is nevertheless good enough for early visualevaluation Since accuracy is not a major concern inthe early design stage, some optimisation techniquescan be applied to smooth the painted regions andboundaries

Each painted region is bounded by a boundary,which is fit to a curve Curves can be very useful in thesurface segmentation with possible application in thereconstruction and parameterisation of complex sur-face There are articles that are especially dedicated tocompute smooth curves on the triangular meshes.Krishnamurthy and Levoy (1996) proposed a way tobuild piecewise linear curves through a sequence ofpicked vertices Bonneau and Hahmann (2003)

presented an algorithm to obtain smooth curves by

iterative execution of geometric and topological

optimisation steps In our proposed system, the

reconstruction of parametric cutting surface can be

achieved by fitting smooth polylines, which represent

the boundaries of the patches Various techniques in

surface fitting have been developed through the years

Welch and Witkin (1982) described a method for

preserving a set of geometric constraints to achieve an

interactive sculpting on a free-form B-spline surface,

while Hoppe et al (Hoppee et al 1994) took advantage

of the subdivision surfaces for fitting point clouds of

arbitrary topology In this article, boundary polylines

are fit to curves and constrained to the part mesh

surface

3 Haptic painting for material or colour mark-up

Haptic painting is used to mark-up a volume that has a

different material or colour at locations of a product

conceptualised by a designer By using haptic device in

our proposed system, painting directly on the object

model can be achieved with force feedback to the user

Compared with a mouse input, a haptic device like

PHANToM (SensAble 2010) can provide users a

natural feeling of painting on a real object

In the proposed research, we choose to paint colour

directly on a mesh model (Figure 3) Direct mesh

painting greatly facilitates subsequent boundary

tra-cing and smoothing process by eliminating the

complicated parameterisations from 2D texture to

3D boundary curves Furthermore, due to reduced

computation, our painting system can easily handle

painting on complex models without latency in both

visual display and force feedback

The virtual brush used to paint on a mesh model is

directly controlled by a six degree-of-freedom (DOF)

input device PHANToM (SensAble 2010) In order to

display the colour difference after painting, colourinformation for every painted triangle is stored

In the 3D painting system, it is crucial that the user

be able to place colour on the surface mesh easily andaccurately Two different tools are implemented forusers to paint on a mesh model: a sphere tool (Figure4(a)) enabling large area brushing and a sharp stylustool (Figure 4(b)) dealing with single triangle painting.Usually the sphere tool is applied to gain a rough andfast paint, while the stylus tool is used to handle theboundary or small features For easy painting opera-tion, zoom and rotation functions are provided Whilethe painting is done through a haptic device, rotationand zoom of a model are controlled by the mouse

In brush paint, the sphere is directly controlled bythe haptic device and its centre point (x0, y0, z0) istracked Triangles with all three vertices (xp, yp, zp)inside the sphere (as governed by Equation (1)) will bepainted immediately when a collision is detectedbetween the sphere and the mesh surface (Figure4(a)) The radius of the sphere Rtool is changeable sothat various painting efficiency can be achieved.ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

(1) The triangle DABC containing the tool contactpoint is first decomposed into three sub-triangles DADE, DDCE and DDBC according

to a triangle subdivision algorithm

(2) Since the tool contact point is inside DDBC,DDBC is divided into four sub-trianglesDDGH, DGBF, DHFC and DGFH

(3) Further check shows that the tool contact point

is inside DGFH

(4) Recursive subdivision is applied to DGFH.(5) The subdivision process is repeated until therequirement (d 5 Dtool) is met (where d is themaximum edge length of the triangle and Dtool

is the diameter of the painting tool)

In Figure 6(a), the tool sphere is large and thepainting speed can be very fast, while in Figure 6(b) thepainting speed is slowed down due to a small radius ofthe tool Yet, by using a smaller painting tool sphere,more dedicated features can be painted

Figure 3 Sample painting on a pumpkin

Using a sharp stylus tool, it is easy for the user to

constrain the painting to a single triangle (Figure 4(b))

The sharp tip is considered as a point; the collision

between a surface and a point is checked, and the

triangle in collision is marked for subdivision and

painting

A virtual wall (Massie and Salisbury 1994) is

applied to all triangles of the model in order to

prevent users from penetrating the surface when

painting A large coefficient K is used to simulate the

force F from the virtual wall in Equation (2), where

Xwall represents the point position on the polygon

surface corresponding to the real avatar position Xp

is indispensable, that is, boundary tracing of paintedregions

Because colours are painted directly on a triangularmesh, the boundary tracing process turns into a searchand sort problem By iteratively searching and check-ing every vertex inside a painted region, we can get allthe boundary points and link them together according

to the sequence to obtain a preliminary boundarypolyline as in Figure 8(a)

Owing to the complexity and randomness of themesh model, there may be some sharp edges onthe boundary, which may affect the smoothness of theoverall boundary curve To solve the problem, aboundary smoothing process is provided for thedesigner to automatically or semi-automaticallysmooth the boundaries By choosing the automaticapproach, our tracing function will quickly filter theboundary with single sharp corners by adding orremoving a triangle as in Figure 8(b)

In automatic boundary smoothing, convex cornerssuch as in Figure 9(a) where all the three vertices V1,

V3, V2form a painted triangle can be detected To dealwith convex corners, the angle b formed by the threevertices V1V3V2is calculated If b 5 1358, the corner

is treated as a sharp corner and removed by linking V1,

V2to construct a new boundary segment and deleting

V3from the boundary list The colour of triangle V1,

V3, V2is reverted to its previous colour If b 1358,the corner is considered a smooth corner and keptunchanged The actual selection of a threshold valuefor angle b is up to the user Now consider the caseslike the one in Figure 9(b) where three consecutiveFigure 4 Illustrations of painting tools: (a) s-sphere tool; (b) a stylus tool

Figure 5 Painting with a small tool (stylus)

boundary points V1, V3, V2 form an un-painted

triangle Such triangles are called concave triangle If

a 5 1358, the colour of the triangle V1, V3, V2 is

changed to the painted colour, and point V3 is

removed from the boundary point list Otherwise, the

corner is treated as a smooth corner and no action is

taken

It is noticed that the above boundary smoothing

operation is restricted to handle single triangle only As

a result, it may not work efficiently on larger corners

especially when the model is in high resolution with

thousands of small triangles Thus, a semi-automatic

filtering operation is also provided In this operation,

the user can mark-up a portion of the boundary that

he/she wishes to smooth; boundary points that are

outside the marked line are removed as in Figure 8(c)

Now the boundary points can be approximated by

a curve Because of the good performance and high

efficiency in producing smooth curves, Hermite curve is

used to interpolate the boundary points The following

is the calculation of the Hermite curve (Figure 10):

375M ¼

375ð3Þwhere P1is the start point of the curve; T1, the tangent

value that shows how the curve leaves the start point;

P2, the end point of the curve; T2, the tangent valuethat shows how the curve meets the end point; S, avector of interpolation points s and its power up to 3,

n þ 1 points P0, , Pn, to be interpolated with nCatmull-Rom spline segments, the tangent value Tican

inter-To solve this problem a projection is needed

Given a target plane axþ by þ cz þ d ¼ 0 and itsnormal vector V¼ (a, b, c), the distance between apoint (x0, y0, z0) and the target plane is

D¼ax0þ byffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi0þ cz0þ d

a2þ b2þ c2

The corresponding projection (xp, yp, zp) of a point (x0,

y0, z0) onto a target plane can be calculated by

Figure 7 The virtual wall model

Figure 6 Painting on the bunny with different sizes of sphere tools: (a) Painting with a larger sphere tool; (b) Painting with asmaller sphere tool

With this method, boundary curves can be exactly

constrained onto the mesh model surface (Figure 12)

Figure 13 shows the generation of a smooth 3D

boundary curve from a sample painted region on a

pumpkin mesh model

5 Cutting surface generation

To perform volume separation of the mesh model, a

cutting surface is needed A closed smooth boundary

curve is obtained from the previous steps; the remaining

work here is to find a smooth surface to interpolate a

boundary curve There have been lots of works done in

developing techniques for creating such a surface from

curves, many of which are employed in commercial

CAD software like lofted surface (Figure 14)

However, these techniques are not guaranteed to

produce good surface parameterisations In general,

parametric curve and surface fitting to point cloud

data can be formulated as a least-squares problem The

equation for a B-spline surface (Eck and Hoppe 1996)

can be written as:

Sðu; vÞ ¼Xn

j¼0

Xm k¼0

Xj;kBjðuÞBkðvÞ ð8Þ

Where u and v are the two parameter values of the

surface, Xj,k are the familiar grid of B-spline control

points and B, B are B-spline basis functions

Figure 10 A Hermite curve

Figure 11 Smooth boundary curve not lying on the meshmodel

Figure 8 Boundary tracing and smoothing: (a) A painted region with rough boundary; (b) Automatic boundary smoothing; (c)User-guided boundary smoothing

Figure 9 Automatic boundary smoothing: (a) Remove a

triangle at the boundary; (b) Add a triangle

Figure 12 Boundary curve constrained to the mesh afterpoint projection: (a) Point project; (b) Projected curve

A point constraint can be easily achieved by:

Pi¼ Sðui; viÞ ¼Xn

j¼0

Xm k¼0

Xj;kBjðuiÞBkðviÞ ð9Þ

To accomplish the fitting, we may uniformly take

sampling points on the boundary curves P0 PLand

apply constraints, respectively Writing in matrix form,

it gives:

Where Bn(vi)¼ [B0(vi)B1(vi) Bn(vi)] and Xn(i)¼

[Xi.0Xi.1 Xi.n]

Solving the above equations will result in an accurate

parameterisation of the input data set, but the control

points Xj,kare unknown A typical solution is to apply

iterative optimisation by guessing an initial set of

parameters Because the constraint is dimensionally

infinite, we can only find the approximation with

minimum error using least square method

There may be numerous methods to generate thecutting surface based on the boundary points Apartfrom the above description, methods based on surfacetrimming (Farouki 1987, Hamann and Tsai 1996) mayalso be used In some cases, the soft touch materialpart on the multi-material products is thin enough to

be represented by surface offset Therefore, instead ofcreating the cutting surface, the boundary curve is used

to obtain a trimmed surface, which is then offset inside

to form a thin volume

6 Implementation

An experimental system is implemented using Visual

Cþþ together with a commercial desktop hapticdevice PHANToM (SensAble 2010) The paintingmodule in the implemented system consists of twomain loops The haptic toolkit OpenHaptics is used inthe haptic loop for force rendering, while OpenGL isemployed in the painting loop handling the visualdisplay and user interface

The most important aspect in developing a 3Dpainting system is to maintain an intuitive, precise and

responsive interface To guarantee the required 1 kHzupdate frequency of the haptic device, force computa-tion is decoupled from the rest of the applications.Only computations related to force feedback likecollision detection and force rendering are performed

in the haptic loop Other operations are done in thepainting loop that runs at 30 Hz for visual display.Figure 15 shows a complete multi-material productdesign process Figure 15(a) shows a conceptual

Figure 13 Painting on the pumpkin with boundary tracing and smoothing: (a) Paint on a mesh model; (b) Trace boundarypoints; (c) Boundary curve fitting

Figure 14 A lofted surface

P0

P1

PL

2664

377

5¼

B0ðu0ÞBnðv0Þ B1ðu0ÞBnðv0Þ    Bmðu0ÞBnðv0Þ

B0ðu1ÞBnðv1Þ B1ðu1ÞBnðv1Þ    Bmðu1ÞBnðv1Þ

3775

Xnð0Þ

Xnð1Þ

XnðmÞ

2664

377

toothbrush design model A portion of the model is

marked up for assignment to a different material, or

colour as in Figure 15(b) Based on the smooth

boundary curve of the marked region, a cutting surface

is generated and cut the volume into two as shown inFigure 15(c) Each volume in Figure 15(c) mayrepresent a different material and colour In actualdesign, some mechanical interlocks between the twovolumes must be designed on the cutting surface.Normally, these interlocking features are in regularshape and can be easily added by designers as in Figure15(d), where a rib is added to the upper volume and acavity of the same size is added to the bottom volume.Figure 16(a) shows a conceptual mouse model.Colour and material assignment is done in Figure16(b) with boundary smoothed A surface cuts into thevolume and separates it into two lumps as in Figure16(c) Note that each individual lump can be furtherdecomposed as in the upper lump in Figure 16(c) InFigure 17(a), a conceptual penguin toy model isshown Its jaw is painted red as in Figure 17(b) Acutting surface generated from the boundary cuts the

Figure 15 Design of a multi-material toothbrush: (a) A

conceptual toothbrush design model; (b) Paint on the model

and smooth the boundary; (c) Cut the model into two volumes;

(d) Add interlocking features at the cutting surfaces

Figure 16 Design of a multi-material mouse: (a) Aconceptual mouse design model; (b) Paint on the mousemodel; (c) Volume decomposition along the paintingboundary