You are on page 1of 6

A Flexible Multi-Agent System Architecture

for Plant Automation


Quibin Feng, Aleksey Bratukhin, Albert Treytl, Thilo Sauter

Abstract— Flexibility has become a key factor for ERP


manufacturing to keep competitive. Software agent ERP

technology can be widely used to improve the flexibility of a


plant and the Manufacturing Execution System (MES) used to Communication via Web Services

control production. This paper introduces a multi-agent MES Product Data


Repository
Order Agent
Supervisor Information
Resource Agent
Supervisor
system design consisting of four types of agents covering the Collector

three layers of the plant automation pyramid from Enterprise


Resource Planning (ERP), Manufacturing Execution System Order
Agent
Order
Agent
Ability Broker

(MES), to the field control layer. The architecture and the Communication
communication protocols of each agent are presented and the via ACL
Order Order
increase in flexibility for mass customized and highly Agent Agent Resource
Agent
Resource
Agent
dynamically changing products such as car manufacturing or
manifold production are given. Field control
Resource Resource

I. INTRODUCTION Fig. 1. PABADIS’PROMISE automation pyramid

A S [1] and [2] mentioned, manufacturing process


development can be divided into three periods: pre-
computer numerical control, computer numerical control
(Architecture for Distributed Dynamic Manufacturing
Scheduling) [7], AIMS (Agile Infrastructure for
(CNC), and knowledge epochs. In the pre-CNC epochs, the Manufacturing System) [8], MetaMorph I [9], MetaMorph
market was characterized by local competition and there II [10], and PABADIS (Plant Automation Based on
were small demands for product variations. Manufacturing Distributed System) [11] can be named as representatives
put focus on increasing production rate. In the CNC epoch, for multi-agent systems (MAS) used to introduce flexibility
the emphasis was changed to cost reduction and product to manufacturing.
quality control. In the knowledge epoch, with intensified Based on the results of EU project PABADIS, EU
global competition and improvement of computer and funded project PABADIS based Product Oriented
information technology, manufacturing is required to be Manufacturing Systems for Reconfigurable Enterprises
able to response rapidly to a fluctuating market. Flexibility (PABADIS’PROMISE) advances the use of multi agent
is a key factor to maintain the competitiveness of a plant. systems (MAS) to allow for the new paradigm of “the order
As defined in [3], flexibility is the “ability to better meet is the application” [12]. The PABADIS’PROMISE MAS
customer needs by modifying existing products”. covers all three levels of plant automation pyramid: ERP,
Conventional centralized manufacturing models show big MES, and field control devices. This paper will concentrate
deficiencies to satisfy frequently changing requirements on the description of individual agents in the
introduced by markets. Enterprises are therefore adopting PABADIS’PROMISE MAS as well as the communication
advanced models and technologies characterized as protocols between them.
distributed, collaborative and flexible [4] such as software This paper is structured as the following: First the
agents. Software agents are very attractive due to their PABADIS’PROMISE system architecture is introduced.
native properties of autonomy, communication, Then the architecture of each agent and the
coordination, and reaction [5, 6]. Exemplarily ADDYMS PABADIS’PROMISE MAS are presented. Finally, the
communication protocols between the PABADIS’
PROMISE agents are described.
Quibin Feng is with the Heifei University of Technology Heifei, Heifei,
China and worked at the Austrian Academy of Sciences, Research Unit for II. PABADIS’PROMISE SYSTEM ARCHITECTURE
Integrated Sensor Systems when writing this paper.
fengqibin@hotmail.com The PABADIS’PROMISE architecture [5] covers all
Aleksey Bratukhin, Albert Treytl and Thilo Sauter are with the Austrian three layers of the plant automation pyramid as shown in
Academy of Sciences, Research Unit for Integrated Sensor Systems.
[Aleksey.Bratukhin, Albert.Treytl, Thilo.Sauter]@oeaw.ac.at Figure 1. The novelty of the approach in
(Corresponding author is Aleksey Brathukin, bratukhin@fiss-oeaw.at, PABADIS’PROMISE is to gain flexibility by shifting the
phone +43-2622-23420-31)

Proceedings of 5th IEEE International Conference on Industrial Informatics


2007, 1047-1052, (C) IEEE 2007
OA/RA/RAS/IC/AB OA/RA/PDR/IC/AB/ERP

RA Interface OA Interface
architecture architecture

Communication Behavior
Communication Behavior Order status protocol knowledge
Ability protocol knowledge
description Scheduling and
Scheduling and Domain History Domain
allocation knowledge
allocation knowledge
Device status Process Order
Resource Decision Decision
segments management
management making policy making policy
Database Order
Database Control module Repository Repository
decomposition
Control module
Interface

Fig. 3. OA architecture
Field device

Fig. 2. RA architecture from the basic purpose of the MES – to provide a link
between the customer (at the ERP) and the resource (field).
decision-making process from the ERP system down to the Therefore, the MAS in PABADIS’PROMISE can be
MES layer. The technology to provide this flexibility in the divided into three main categories:
MES is an agent based system described in this article.
• ERP interface (OAS, RAS)
Furthermore, the field control layer is integrated into the
• MES kernel (OA)
architecture by a set of plug-and-participate mechanisms
• Field control interface (RA)
that allow for flexibility and adaptability of the shop floor.
Such components as PDR, IC and AB are the database
To fulfill this task the PABADIS’PROMISE architecture
support and are not discussed in the next chapter that
consists of eight components (Figure 1):
describes the architecture of the four main agents.
• Enterprise Resource Planning (ERP) system:
responsible for the order management at ERP level and
III. PABADIS’PROMISE MAS ARCHITECTURE AND
interface to the superordinate layers of a company and
AGENT ARCHITECTURE
the customer.
• Order Agent Supervisor (OAS): responsible for the According to the categories in [14] and the
creation and initialization of order agents and for the functionalities of each agent, the OA and RA are designed
supervisory control of their life cycles. as a hybrid collaborative and deliberate agent. OAS and
• Order Agent (OA): responsible for the execution RAS do not need to take complex decisions and mainly
control of a production order. monitor and forward control commands to OA and RA
agents. They are therefore designed as reactive agents.
• Resource Agent (RA): responsible for the management
of a resource. A. Resource Agent (RA)
• Resource Agent Supervisor (RAS): responsible for The RA is responsible for the overall management and
supervision and control of resources. It can be contacted operation of a resource. As a deliberate agent, the RA
by ERP or other applications such as SCADA to makes its decision based on the belief-desire-intention
monitor, configure, control maintenance and manage (BDI) model. Based on the general architecture of BDI
resources. agent [15], the architecture of the RA is shown in Fig. 2.
• Information Collector (IC): responsible for the The RA is the link between the MAS and the field device
collection and provision of the order and resource controlled by it. It therefore consists of two interfaces: a
related data. standardized FIPA interface to the agent system and a
• Ability Broker (AB): responsible to make all resource resource specific interface to the field level to retrieve and
abilities and resource availability known in the system. sent information to the controlled device.
It provides facilities to collect abilities announced by a At start-up the RA initializes the resource and registers
resource manager and a look-up service to query for its abilities in the AB. Also when there are changes of its
certain abilities. own abilities and availability, the RA is responsible to
• Product Data Repository (PDR): maintains a database inform the AB about changes. During operational state the
of bill of material and bill of operation for all products. RA is responsible for processing production requests from
It is the “cookbook” of the system. OAs. It therefore needs to schedule and allocate its
The core of the PABADIS’PROMISE architecture is the underlying resource and to supervise exception in the
MES layer that is represented by the MAS and has two execution of the allocated processes on the resource. In
primary goals: 1) to execute the ERP orders and 2) to case of exceptions the RA also must inform OAs booked
control the field layer devices. These general goals come on it to reschedule their work flow. The repositories of the
ERP System OAS architecture
1000 cars
Create and initialize
OAS 1 Message input superOA Message output
OA 1 Supervise OA
1000 cars
Lot decomposition Create and submit
the order report
OA 1.1 ... OA 1.98 ... OA 1.1000
1 car (assembly) 1 car (assembly) 1 car (assembly)
RAS architecture
Operations decomposition Message input Message output
Data collection
OA 1.98B OA 1.98E OA 1.98W1 ... OA 1.98W4
car body engine 1 wheel 1 wheel Resource supervision

Fig. 5. Order and Resource Agent Supervisors architecture


OA 1.98BF OA 1.98BD1 OA 1.98BD2 ...
car frame car door car door

Fig. 4. Order decomposition functionality is to start and supervise the production of the
lot and therefore only implements functionality for lot
RA support the above actions. Especially the behavior decomposition and limited scheduling. ProductOAs do the
knowledge that allows the RA to efficiently react on main workload in producing and assembling the final
exceptions and perform alternative actions and the domain products they are responsible for. Their functionality
knowledge required for scheduling decisions must be includes order management and scheduling as well as
mentioned here. Details on scheduling in further decomposition of products. Finally, processOAs are
PABADIS’PROMISE can be found in [18]. responsible for fulfilling the production of subcomponents1
and include functionality for resource allocation and fine-
B. Order Agent (OA) grained planning. The depth of operation decomposition is
The Order Agent (OA) is the key component of the not limited by the system, but naturally will be limited by
PABADIS’PROMISE concept that autonomously organization of production and the overhead introduced by
organizes and executes a production order given by the agent creation and agent-to-agent communication and
ERP system. Following the paradigm “The order is the hence functionalities assigned to a lotOA, productOA or
application” an OA is responsible for one directly assigned process OA might be implemented in a single agent
product. In opposite to centralized approaches where the instance.
complete production is planned, the OA only takes Hierarchical decomposition of an order and instantiation
decisions for its on product. The main goal of this approach of the respective agents bring the following advantages:
is not to reach a total optimum in production rather, but the • An individual agent does not become too complex and
aim is to react in short time to changes in the order flow, is easy to develop.
e.g., late order freeze, or on the shop floor such as • The code especially of a processOA is simple and small
maintenance or machine failures. In order to improve decreasing the requirements for resource limited
utilization of the plant, OA show a benevalent behavior devices such as RFID or embedded platforms.
especially during re-scheduling. Especially active RFID tags hosting OAs are of importance
The architecture of OA is shown in Fig. 3. It is quite to the PABADIS’PROMISE concept since they allow
similar to RA’s due to the fact that both RA and OA are physical association of product and responsible order agent.
deliberate agents. The repositories form the base for all Favors of this approach are easy product identification and
decision making and the database holds production status location and storage of product relevant data and a
and all relevant data to execute the process segments. The responsible agent together with the product that helps to
functionality an OA has to offer range from order reduce overheads in highly dynamical environment.
decomposition, i.e. splitting a lot into products and Detailed information about the combination of agent and
subcomponents, to the management and execution of the RFID and their advantages would go far beyond the scope
process segments needed to fulfill an order to scheduling of of this paper and can be found in [16].
process segments and allocation of required resources.
Depending on the complexity of a production order an C. Order and Resource Agent Supervisor (OAS, RAS)
OA can create further OAs to fulfill subtasks of its The Order Agent Supervisor is an interface between ERP
production order. I.e., simple orders will be completely and MES. It has the following functionalities:
done by a single OA whereas complex orders such as the • Instantiation of ERP orders: An OAS receives
production of a car usually decompose into multiple levels production orders from the ERP and initializes them by
of OAs (see Fig. 4). The following three flavors of an OA
can be identified: lotOA, productOA, and processOA: The 1
The distinction between product and subcomponent is only done to
lotOA is only required for bigger production lots. Its main clarify the level of detail and complexity. From the agent implementation
point of view they are identical and the term Work in Progress is used.
creating appropriate OAs. Depending on the order product and also for organizing the production scheduling
complexity and the decomposition strategy, either a for this product, changes can be introduced at any time. In
lotOA or directly a productOA is applied. this respect changes can be performed as long as a
• Supervision of production orders: The OAS gets the production step is not executed or previous completed steps
reports about the order execution from its OAs and prevent the change. Changes will not influence other
forwards them to the ERP system. The OAS also reacts products since the scheduling is also “local” to one agent
on status requests or order changes from the ERP and and dynamically created based on the latest status of
accordingly supervises the OAs. To find a compromise resources. The second aspect is resource management. Due
between granularity and overhead within to the fact that there are no fixed relations between OAs
PABADIS’PROMISE, a milestone based reporting is and RAs, failed resources can be dynamically replaced,
introduced. E.g. a customer might not be interested if a therefore, easing maintenance, change of equipment etc.
certain screw is fixed, rather he wants to know if his car Also load-distribution among multiple resources of the
is painted red or if he can still change the color. same kind is done automatically since OAs request an
The implementation of the OAS is a reactive agent that ability and not a certain resource. The authors are aware
contains several simple, parallel-organized competence that optimization is a crucial aspect since a flexible yet
modules. Based on the general architecture of a reactive purely utilized shop floor is not acceptable. Ongoing
agent described in [17], the architecture of the OAS is investigations show promising approaches that allow
shown in Fig. 5. combining good optimization together with higher
The purpose of having the RAS is to provide additional flexibility due to the simpler system concept. Investigated
access points to the shop floor of the system apart from the issues are ERP design that is ability and not resource
ERP – OAS – OA – RA mechanism. These mechanisms centric and local scheduling domains to reduce
can be used for different purposes such as coordination of communication effort. Agent systems from other domains
the resources or integration of legacy systems for data indicate that the above goals can be reached for highly
collection or influencing the behavior of a resource such as customized products and highly dynamic shop floors. The
SCADA or operator control stations. Similar to the OAS, PABADIS’PROMISE project therefore installs a field test
the RAS is also a reactive agent. to verify the expected results.
The architecture of RA is shown in Fig. 5. The Resource
supervision module also provides a possibility for a human V. COMMUNICATIONS IN PABADIS’PROMISE MAS
machine interface. The Data collection module is Agents in the PABADIS’PROMISE MAS have to
responsible for collecting necessary data for human coordinate with each other to fulfill their production tasks.
supervision and resource management. The successful coordination is based on an efficient and
effective communication. In the following sections, the
IV. FLEXIBILITY BY COOPERATION proposed agent-centric approach is described in detail and
The approach taken in PABADIS’PROMISE is to the communication protocols are described.
maintain necessary flexibility required in future
A. RA Communications
manufacturing by introducing cooperation between agents.
A single agent has limited knowledge and resources. For a Fig. 7 is the sequence diagram of the communication
complex task, several agents need to coordinate with each links of the RA. Via the communications with the AB, the
other to find a solution. The multi-agent system (MAS) of RA can register and change its abilities and availability.
PABADIS’PROMISE provides a communication The RA can also look up needed abilities in the AB.
(cooperation) infrastructure to allow the four main agents Additionally the RAS can be informed to record and notify
to fulfill their tasks. Agents are organized in the superordinate systems about ability changes and
hierarchical way (Fig 1 and Fig 4) but collaboration exceptions.
between different hierarchy levels is allowed. The OAS The RA can also communicate with other RAs for
first receives the production order from the ERP. It creates scheduling as well as with OAs as shown in the Fig 6,
one or multiple OAs corresponding to the order. According depending on the different mechanisms for scheduling and
its production data, missing information is complemented resource allocation.
from the PDR. Then the OAs query for needed abilities at B. OA Communications
RAs using the AB and finally set up a schedule for their OAs are key components of the MES layer in a
“private” production. PABADIS’PROMISE system. As described before, there
The (advanced) flexibility introduced by the MAS is are three types of OAs in PABADIS’PROMISE system.
based on two independent aspects: first is order flow The lotOA receives the production order from OAS. Both
management. Since an OA is only responsible for its own lotOA and productOA need to contact PDR to collect the
lotOA productOA PDR processOA AB RA

askForProductData
productData
askForProcessData
processData
askForFunctionBlock
functionBlock

Create mediaOA

initializeMediaOA
askForProductData
productData
askForProcessData
processtData
askForFunctionBlock
functionBlock

Create processOA

initializeProcessOA
askForNeededAbilities
eligibleAbilities

askForScheduling

schedulingResponse
askForAllocation
allocationResponse
askForAcess

accessResponse
sendFunctionBlock

functionBlockResponse
initializeApplication

Execute
application
askForApplicationStuate

applicationStatue

askForApplicationTermination

applicationTerminationResponse
applicationReport
orderReport
orderReport

Fig. 6. OA communications
needed information to decompose an order and plan the RAS is involved in the resource discovery mechanism,
execution. A productOA creates one or more processOA(s). which is trigged either by the RA or by the ERP, where the
A processOA then negotiates with RA(s) for scheduling RAS connects two entities to synchronise respective local
and resource allocation based on the information from the databases.
AB. After the production finishes, the processOA sends a
report to its productOA, that creates a report to its lotOA VI. CONCLUSIONS AND FUTURE WORK
and the lotOA submits a collected report to OAS. In the
Fig. 6, the communications of all three types of OAs are Within the PABADIS’PROMISE project a multi agent
shown. The horizontal communication between system was developed to increase flexibility in production.
productOAs is not included. RA AB RA RAS

registerAbility
C. OAS and RAS Communications changeAbility
schedulingRequest
Communication patterns of the OAS and the RAS are lookForNeededAbilities

given in Fig. 8 and Fig. 9. After receiving a production eligibleResources


askForScheduling
order the OAS creates a lotOA. When the production order
schedulingResponse
is finished, the OAS receives a report from the lotOA and
creates its report to the ERP. askForAllocation

allocationResponse
The RAS provides interfaces to the legacy systems such initializeApplication

as HMI as well as providing an interface to the ERP to Perform application


applicationReport
control the shop floor. The RAS can influence a resource
by changing the current schedule, current state, features, availabilityChange
Supervise
and supervises utilization of a resource, which are all
included in supervise commands. The RAS
communications are shown in Fig. 9. Additionally, the Fig. 7. RA communications
ERP OAS LotOA RA RAS ERP
Legacy system
productionOrder
Request
changedProducitonOrder (supervise)
createLotOA Status Request
(Resource Status)
initializeLotOA Order
decomposition, Ability description
scheduling or changes Resource Status
orderReport and production
orderReport control Resource Ability

Fig. 8. OAS communications


Fig. 9. RAS communications

[3] Dixon J. R.: “Measuring Manufacturing Flexibility: An Empirical


Investigation”, European Journal of Operational Research, 60, 131-
Distribution of functionality between Order Agents (OA), 143. 1992.
Resource Agents (RA) and Resource and Order Agent [4] Shen W. M. and Barches J. P.: “Computer Supported Cooperative
Environment for Engineering Design: A Survey”, Tech. Report, 96-
supervisors (RAS, OAS) create small flexible entities that 122, 1996.
are used to map the functionalities of a Manufacturing [5] Jennings N. and Wooldridge M.: “Agent Technology: Foundations,
Execution System. Instead of centralized control by a Applications, and Markets”, Springer, 1997.
[6] Franklin S. and Graesser A: “Is it an Agent, or just a Program? A
single planning and control entity, the decision-making is Taxonomy for Autonomous Agents”, Proceedings of the Third
distributed among the agents. International Workshop on Agent Theories, Architectures, and
Negotiations and individual scheduling allow to fulfill Languages, 21-35, 1996.
[7] Butler J. and Ohtsubo H.: “ADDYMS: Architecture for Distributed
complex tasks and react on changes of orders as well on the Dynamic Manufacturing Scheduling: Artificial Intelligence
shop floor. Although communication and coordination are Applications in Manufacturing”, The AAAI Press, 199-214, 1992.
not centrally planned agent systems prove to work very [8] Park H., Tenenbaum J., and Dove R.: “Agile Infrastructure for
Manufacturing Systems: A Vision for Transforming the US
autonomous and efficient [10]. The next building steps are,
Manufacturing Base”, Defense Manufacturing Conference,1993.
based on the concept and specification given here, to
[9] Maturana F. and Norrie D.: “Multi-Agent Mediator Architecture for
integrate the system into conventional ERP systems, which Distributed manufacturing”, Journal of Intelligent Manufacturing, 7,
need to adopt the concept of abilities, i.e., planning 257-270, 1996.
operations independent of a specific resource, and criteria [10] Shen W.M., Maturana F., and Norrie D. H.: “MetaMorph II: An
Agent-Based Architecture for Distributed Intelligent Design and
to efficiently split orders. Such criteria should be (1) to Manufacturing”, Journal of intelligent manufacturing, 11(3), 237-
avoid too many layers of productOAs which may cause the 251, 2000.
decrease of flexibility; (2) to avoid too little layers of [11] Bratukhin A., Penya Y.K., and Sauter T.: “Intelligent Software
Agents in Plant Automation”, Proceeding of 3rd International
productOAs which will increase the complexity of each
NAISO Symposium on Engineering of Intelligent Systems, Spain,
agent. 77-83, 2002.
Multi agent systems provide the advantage in flexibility [12] Peschke J., Lueder A., and Kuehnle H.: “The PABADIS'PROMISE
and autonomy by introducing a certain degree of architecture – A New Approach for Flexible Manufacturing
Systems”, Proceeding of 10th IEEE conference on Emerging
indeterminism. To validate the usability and the degree of Technologies and Factory Automation, 491-496, 2005.
utilization the PABADIS’PROMISE project will install [13] Wooldridge M., Jennings N. R., and Kinny D: “The Gaia
field tests in areas of mass customizable goods and highly Methodology for Agent-Oriented Analysis and Design”, Journal of
Autonomous Agents and Multi-Agent Systems, 3(3), 285-312, 2000.
dynamic shop floors. In particular, car production and [14] Nwana H.S. and Ndumu D.T.: “An Introduction to agent
manifold production have been chosen for evaluating the technology”, BT Technology Journal, 14(4), 55-67, 1996.
properties order flexibility and shop floor flexibility [15] Chen Peishuai: “CORBA-based multi-agent system”, Dissertation of
ZheJiang University, 2002.
respectively in the architecture of PABADIS’PROMISE. [16] Bratukhin A., Treytl A.: “Agent technology enhancement by
embedded RFID for distributed production control”, .
REFERENCES [17] Rodney A. B.: “A robot layered control system for a mobile robot”,
IEEE Journal of Robotics and Automation, 2(1), 14-23, 1986.
[1] Mehrabi M.G., Ulsoy A.G., Koren Y.: “Reconfigurable
[18] Bratukhin A., Treytl A., Khan B. A: “Resource-oriented scheduling
Manufacturing Systems and Their Enabling Technologies”,
in the distributed production”, Proceedings of 5th IEEE International
International Journal of Manufacturing Technology and
Conference on Industrial Informatics,2007
Management, 11(1), 113-130, 2000.
[2] Mehrabi M. G., Ulsoy A. G., and Koren Y.: “Reconfigurable
manufacturing systems: Key to future manufacturing”, Journal of
Intelligent Manufacturing, 11, 403-419, 2000.

You might also like