=Paper=
{{Paper
|id=Vol-2215/paper20
|storemode=property
|title=Towards Smart Factory: Multi-Agent Integration on Industrial Standards for Service-oriented Communication and Semantic Data Exchange
|pdfUrl=https://ceur-ws.org/Vol-2215/paper_20.pdf
|volume=Vol-2215
|authors=Ronald Rosendahl,Ambra Calá,Konstantin Kirchheim,Arndt Lüder ,Nikolai D'Agostino
|dblpUrl=https://dblp.org/rec/conf/woa/RosendahlCKLD18
}}
==Towards Smart Factory: Multi-Agent Integration on Industrial Standards for Service-oriented Communication and Semantic Data Exchange==
https://ceur-ws.org/Vol-2215/paper_20.pdf
Towards Smart Factory: Multi-Agent integration on
industrial standards for service-oriented
communication and semantic data exchange
Ronald Rosendahl∗ , Ambra Calà∗ , Konstantin Kirchheim∗ , Arndt Lueder∗ and Nikolai D’Agostino†
∗ Otto-v.-Guericke University, Magdeburg, Germany
Email: [ronald.rosendahl, ambra.cala. konstantin.kirchheim, arndt.lueder]@ovgu.de
† CENIT AG Digital Factory Solutions, Stuttgart, Germany
Email: n.dagostino@cenit.de
Abstract—The concept of smart factories in the industry 4.0 cesses form an ideal basement for the acquisition, structuring
(I40) paradigm is based on the concept of cyberphysical systems and management of system related data at the runtime of a
which is more specifically elaborated within the concept of the facility. In the area of discrete manufacturing the data model
smart manufacturing component. In that concept an arbitrary
production asset is assigned to an administration shell, that covers and metamodeling capabilities of the file format Automa-
partial models and services for arbitrary functions. The desired tionML evolved which supports the modular thinking in terms
attributes of smart factories to be self-organized, self-optimized, of mechatronical engineering, high level of model dynam-
self-learning, etc. is based on the idea of an intelligent co- ics and concepts for building common models and libraries
operation between the smart manufacturing components. How- for different engineering domains and disciplines. In several
ever, the actual performing of this co-operation is not appropri-
ately elaborated yet. In this paper the approach of a Multi-Agent- research activities and industrial applications AutomationML
System to be applied as a well elaborated mean for intelligent proved a suitable data model for the multi disciplinary modular
co-operation of smart manufacturing components shall overcome engineering of production systems as well as the transition of
this lack. Furthermore the application of existing standards data from the design and virtual commissioning phases of a
for service-oriented industrial communication infrastructure by system to its runtime. In [4] a concept of an AutomationML
OPC-UA and for semantic data exchange by AutomationML
provides existing means for realizing the required connectivity and OPC UA based implementation of I40 components is
and interoperability for intelligent co-operation that shall realize being discussed. It is shown how AutomationML projects are
self-X capabilities of smart factories. automatically translated into namespaces of OPC UA in a
Index Terms—Agents, Flexible manufacturing systems, Knowl- standardized way following the specification given in DIN
edge representation, Service Spec 16592 [5]. Combining this auto-generated component
description in a standardized communication architecture with
I. I NTRODUCTION agent technology seems a promising candidate for the cre-
Driven by increased demands for shorter life cycles and ation of dynamically adaptable system modules thus easily
a higher degree of individualization of products the manu- bringing smart applications to the production domain. An
facturing industry is looking to raise the dynamics in their actual research question is to evaluate how the model and
systems. With the evolving technological capabilities strong workflows out of the engineering processes at design time
social and organizational changes in the management of all associated with AutomationML fit the requirements of the
life cycles of those artificial systems are being expected. This runtime management of a system specifically in an automated
issue was addressed by a funding of the German government way performed by Multi-Agent-Systems (MAS). Therefor it
to enforce a 4th industrial revolution and introduced the has to be proven which methodological and technological
term ”Industrie 4.0” (I40) [1]. In the domain of production adjustments or extensions are needed to support or enable
systems engineering the requirements for I40 are in a first step the AutomationML ecosystem to be a valuable part of the
being addressed by an increase in modularization of system development and operation of an MAS.
components. One essential building block in this discussion is Within this paper it will be presented how the requirements
the Smart Manufacturing Component (former: I40 component) for the acquisition of system information by Multi-Agent-
[2] which is being developed on the idea of Cyber Physical Systems will be met by a service based online model of
Production Systems (CPPS) and constitutes the concept of an the system represented in AutomationML. After a brief in-
administration shell which combines the presentation of the troduction to the underlying technologies for the architectural
physical components of a production system along with its proposal given in this paper the requirements are collected that
digital representation [3]. This modularity renders the creation need to be matched to apply an AutomationML knowledgebase
and adjustments of systems a semi-automatable process. The for MASs in the domain of production systems. Special
necessary information and data models within such design pro- attention is spent to the modeling capabilities for semantical
124
concepts of the system. In the following section it is being operation like variable customization requirements, changes
explained how this requirements are met by the proposed in plans and schedules, changes in technology, and equipment
technologies and which auxiliary methodologies or technolo- failures [7]. Other common application scenarios appear in the
gies have to be applied to the architecture. The final section transportation and material-handling systems, production man-
presents a prototypic implementation which is exploited to agement of frequently disrupted operations, coordination of
prove the solution in a lab size production system. organization with conflicting goals and frequently reconfigured
environments [13]. A more comprehensive set of MAS appli-
II. I NDUSTRIAL S TANDARDS AND ARCHITECTURES FOR
cations in industry can be found in [14]. Different requirements
THE PROPOSED SMART MANUFACTURING COMPONENT
in the application domains can be derived also from other
A. Industrial Agents and Architectures emergent reference control architectures developed within
Agent technology is a suitable technology to decentralize the ongoing projects like IMPROVE [15], PERFoRM [16],
control automation architectures in production systems in BaSys4.0 [17] and INTER-IoT [18], [19] and [20]. Besides
order to dynamically react to changing requirements according the general requirements of flexibility, scalability, modularity
to the I40 paradigm. Especially, MAS [6] have been used in the and standardized interface as well as common information
past years to develop modular, flexible, robust, adaptive, recon- model, other requirements can be included depending on the
figurable and responsive complex production systems, based application use case, such as: data curation; inter-enterprise
on the decentralization of control functions over a community data exchange; privacy, integrity and security; service detec-
of distributed, autonomous and cooperative agents. Agents tion and orchestration; and real-time communication between
applied to the industrial domain are called Industrial Agents architecture participants.
[7] and share the same characteristics of software agents, such
B. AutomationML
as intelligence, autonomy and cooperation. Depending on the
application scenario and the degrees of importance, agents can In the year 2006 a leading automotive OEM initiated
fulfill different requirements in production control systems. several cooperations of developer teams of their engineering
Agents have been used historically in industry to connect departments, service contractors and suppliers. The goal was
the cyber to the physical part. In fact, several distributed to overcome identified waste of resources in the engineering
control automation architectures based on MAS technology chain of production systems. One major problem was that
have been developed in the past years within research projects, at some transition points between the engineering phases
highlighting their benefits in production systems in terms of there was no digital transmission of the engineering data
flexibility and reconfigurability [8]. These research projects at all. The manual submission by printed documents lead
have derived MAS-based reference architectures focusing on to avoidable and error-prone reinterpretations of information
control, planning and scheduling, and supervision solutions, to digital representations. The solution proposal was to de-
positioning agents at the bottom levels of the ISA-95 and velop a data format which addresses the special needs of a
predominantly at MES level. Some of the realized MAS gapless transmission of digital representations of engineering
architectures need to be mentioned, such as GRACE project information along the whole development cycle of discrete
[9], which developed a multi-agent architecture that operates manufacturing systems. To evolve and promote this format in
at all stages of a production line integrating process control the year 2009 the ”AutomationML” society was founded. At
with quality control. The MAS architecture was applied on a the time of development object orientation and model driven
washing machine production assembly line and tested against development where state of the art paradigms and prototypic
planned and unforeseen variation of variables. IDEAS project object models gained some popularity with the ascent of
[10] developed a fully distributed and pluggable mechatronic web technologies in languages like JavaScript. In consequence
environment, based on agent technology, capable to self- the main requirements to the data model of AutomationML
organize itself. The assembly system operated with a totally that where derived together with production system engineers
distributed multi-agent control system. PRIME project [11] were:
developed a multi-agent architecture using plug-and-produce • object orientation to support the modeling of systems
principles for semi-automatically configuring production sys- in terms of mechatronical engineering and to represent
tems through innovative human-machine interaction mecha- a system component with several distinct (mechani-
nisms. The multi-agent architecture here is used for module cal, electrical, information-technological) aspects as one
integration including legacy systems. Finally, ARUM project unique self-contained entity,
[12] also proposed an agent-based architecture aiming at • prototypic inheritance to support the reuse of model el-
minimizing the response time to unexpected events during the ements with the capability to be easily evolved, avoiding
ramp up phase of plant. As seen in the past projects, main the complexity of managing consistent class hierarchies
requirements of MAS comprise the specific hardware integra- at the same time,
tion, reliability, fault-tolerance, scalability, industrial standard • weak inheritance rules affecting instance data regarding
compliance, quality assurance, resilience, manageability and the fact that engineering information is assembled at
maintainability. For industries that aim at mass production of design time and its associated data will be incomplete
individual customized products, MAS can support everyday and inconsistent meanwhile,
125
• object relations for the modeling of dependencies be- software. In this section the requirements to a knowledgebase
yond the hierarchical structures as well as for such an integration are discussed. After introducing some
• object roles for semantical richness of the model. basic principles that are being expected as general require-
Regarding this requirements AutomationML was defined as ments for data models applicable with MAS, more specific
an open, XML-based format in a series of standards (IEC concepts will be pointed out. Those address the requirements
62714) for the model based description of production systems derived from the aforementioned use-cases and reference ar-
and components. Given the mentioned capabilities Automa- chitectures for MAS. These requirements are evaluated here
tionML has proven a valuable data model for the exchange of against the capabilities of the AutomationML concepts of
engineering information of different disciplines in the design the systems engineering phases. On the basis of these facts
stage of a system [21], [22], [23]. It integrates sets of very implementation gaps are identified and enhancements sug-
heterogeneous information of the system being modeled and gested to probe the limitations of the proposed solution of
forms a sound repository of machine readable descriptions of an AutomationML based distributed system model for the use
the system. In several applications it has shown that systems in MASs.
represented in AutomationML exploiting the full modeling
A. Structure of information
capabilities form a directory of component descriptions in
a systemic way. Consequently the idea came up to use this Not just to the application of MAS one major principle of
data objects across further life cycle phases of the underlying information modeling is the Divide and Conquer paradigm.
system. Several research projects are currently concerned with It helps to manage complexity by breaking huge systems
the integration of this data at the runtime of a system [24]. down to simpler ones and have the responsibility for the
Since the majority of agent based approaches rely on an subsystems assigned to a squad of professionals (algorithms,
information model of the system they are being applied to, experts, 3rd party companies) who not just share the load but
AutomationML seems to be a reasonable candidate for the are specialized to a certain task. To integrate this partially
acquisition and management of system information enabling managed subsystems and make them work as a whole it is
a repository of multi purpose system descriptions [25]. necessary to conceptualize elements that represent the part and
aggregate its related information. These elements are required
C. OPC-UA Information Models to be coherent compositions of data based on a common
OPC UA (Open Platform Communications Unified Archi- data model and to have an identity. These elements are
tecture) is a platform independent architecture for commu- subordinate to other elements which represent the (sub-)system
nication of devices and systems in industrial communication they constitute and recursively form a part of.. hierarchy.
networks defined in the series of standards (IEC 62541). It AutomationML meets these requirements with the concept
focuses on providing a unified service oriented infrastructure of an InternalElement (IE) which may contain an unlimited
for the transport of data and the modeling of information in number of Attributes. The standard specifies a mandatory
order to facilitate vendor independent interoperability between identifier for each of this elements. The identifier is defined
devices. OPC UA models information as a full meshed net- to be unique and shall be implemented using Globally Unique
work, consisting of nodes that represent for example types or Identifier (GUID) that per definition persists along the life-
objects, and references that describe relation between these cycle of the distributed set of elements. Each InternalElement
nodes. Following this approach, object-oriented concepts like may contain an unlimited number of other InternalElements
inheritance and type hierarchies can be described in a way that constituting a tree structure which is contained within a so
enables clients to query and manipulate instance- and type- called InstanceHierarchy. This way an MAS may browse and
related information in a uniform manner, while also allowing access data of certain assets within the system in a structured
the extension of existing type hierarchies. OPC UA imposes way.
no further limitations on how to model information in order to
allow the representation of arbitrary existing rich information B. Types of information
models [26]. A basic principle in all MAS architectures is the specializa-
This integration of vendor-specific information models into tion of agents to purpose and asset. If agents with different
OPC UA is specified in so called companion specifications. purpose access the same asset they may require different
DIN Spec 16592 specifies the translation of AutomationML kinds of contained data with distinct semantical meaning not
projects into OPC UA information models, enabling the online satisfied by the definitions of the underlying data model and
exchange of AutomationML models along with all of OPC a static structure. These use cases require the attributes to be
UAs features concerning data management, multi user support, recognized by a name and to have a type to qualify the value
access methods, security etc. [27]. as well as a unit to quantify the value. In AutomationML there
is a mechanism for dynamic typing and only spatial, spatio-
III. R EQUIREMENTS ON L IFE -C YCLE -S PANNING temporal and kinematic information are statically defined.
I NFORMATION M ODELS FOR THE SHOP FLOOR Therefore as part of the basic concepts each InternalElement
Multi-Agent-Systems performing tasks on industrial appli- may declare a Frame attribute which specifies its location
ances require suitable descriptions of the underlying hard- and by transformation parameters in the three-dimensional space
126
relative to its parent. Spatio-temporal and kinematic data is as slopes and cycles up to full meshed networks of elements.
embedded by a specific linking mechanism to instances of the AutomationML has built-in mechanisms for the definition of
open standard data model Collada (ISO/PAS 17506) originally connection points of elements called ExternalInterface and
for the purpose of the virtual commissioning. This mechanism InternalLink objects that establish connections between them.
may be useful for autonomous navigation of movable system In [29] a method is described to exploit this linking mechanism
parts but is a very special use case for agents and therefore out to define directed graphs in AutomationML. By defining an
of scope of this document. To dynamically define the meaning InternalElement being an edge object and inserting it in
of data each Attribute contains an entry Unit to explicitly between two links to the related elements a connection with
quantify a measure and an entry RefSemantic to add a string rich semantical capabilities may be specified. That way agents
based hint for the semantical interpretation to qualify the value. may discern routes through the system, dependencies as well
Additionally AutomationML declares the extended concept of as logical or physical connections from the model enabling a
PropertySets. The definition of PropertySets allows to map variety of applications.
proprietary attributes of AutomationML objects with seman-
tically predefined attributes. These attributes are managed in E. Semantical depth
shared libraries of roles. The RoleClassLibrary concept will To express commonalities of elements for algorithmic ma-
be explained in detail in the paragraph about semantical depth. nipulations class based inheritance is a very popular concept.
This way the data models of the platform the agent and the But for the application of MAS in dynamic system scenarios
production system have been defined in are being decoupled statically typed systems of data structures like inheritance of
a semantically correct interpretation of the data by the agents classes are not practicable. In contrast to applications with a
is ensured. fixed purpose it is not possible to anticipate all occurrences
of variety and each necessary class structure in advance.
C. Views on information Therefor the information model is required to be capable
As already mentioned different types of agents access of runtime changes. But modifying the model structure
different information derived from different subsets of data dynamically inhibits the power of type safety. To avoid the
of an asset. Depending on the purpose an agent was designed necessity of runtime analysis a concept of an alternate access
for there are different decomposition structures of a system. pattern to identified structures is necessary. This concept
Therefore it is required to overcome the limitation of a single requires the possibility to indicate the occurrences of the
hierarchy and represent the same system elements in optional structural pattern by some kind of annotation to the model.
alternative hierarchies. To avoid the expensive processing of Patterns used in the application are documented in applica-
search and filter operations there are several mechanisms in tion specific pattern catalogs. A possible implementation of
AutomationML that may be used to explicitly define aggre- this requirements with prototypical objects was discussed in
gates or subsets of data. To organize the elements in different [30]. AutomationML is based on the CAEX standard which
decomposition structures a methodology was introduced in implements its infomration model on prototypical objects. But
[28] to implement three distinct views on the same system in AutomationML the identification of instances was changed
elements related to the focus on Product, Process and Resource to an explicit addressing scheme by IDs without adjusting
(PPR) demands. To represent the same element in more then the inheritance concept of CAEX which uses path to allocate
one hierarchy the standardized MirrorObject was used which base classes. Additionally the inheritance rules were loosened
basically is a pointer to the original element. To aggregate sets to enable the exchange of draft design data which is by
of data the same principle of pointers is used with the extended definition incomplete and possibly inconsistent. But at the
concept of a so called GroupObject which containerizes a same time it also enables more flexible ways to evolve the
set of MirrorObjects. In comparison to alternative hierarchies, data model with changing demands. This freedom requires
groups may be associated with Facets. A Facet is also one carefully defined processes and semantical enhancements. The
of the extended concepts of AutomationML and supports the semantical expressiveness and extensibility of AutomationML
specification of a sub-view of a parent element. It specifies a has been discussed in several publications [31], [32], [33]. The
subset of Attributes and ExternalInterfaces of an InternalEle- main concepts defined in the standard are the class libraries
ment. Different agents may access the different hierarchies for RoleClasses, InterfaceClasses and SystemUnitClasses. The
and signatures of an element to avoid intense search and concept of prototypic inheritance holds for class hierarchies
preprocessing operations. and allows the definition of commonalities. Interface- and
RoleClasses allow the semantical enrichment of ExternalIn-
D. Relations and networks of information objects terfaces and InternalElements. SystemUnitClasses are used to
In several use cases like self-organization, plug-and produce, define prototypes for InternalElements. But the classes are not
scheduling and logistics the agents have to rely on more meant to be inherited but used as a master copy. For imple-
complex relations between elements. While chains of elements mentation of typesafe access to the model additional concepts
(processes, routes, dependencies) may easily be represented have to be applied to link content to a standardized feature
in ordered sets hierarchical structures are not useful for the system or taxonomy [34]. Such classification schemes may
distinction between uni- and bidirectional relations as well be represented in RoleClassHierarchies where each RoleClass
127
represents an entry and defines the required attributes and OPC UA AML server has to be utilized that ensures the
interfaces of an element having this role assigned. Annotating encapsulation and realizes the behavior of a component to
instance data with standardized assignment of RoleClasses change its internal structure and state.
realizes the alternate access pattern while the weak inheritance There are (at least) 4 different events that entail agents to
brings the freedom of agility to the information model. This update the production systems model in a modular configura-
way the requirements of the application of MAS in dynamic tion:
system setups are met. • Integration of a new resource
• Change of resource
F. Consistency
• Removal of a resource
To ensure or at least asses the correctness of information • Exchange of a resource
and the agents’ decision based upon it, it is very important to These modifications implicate different changes to an Au-
express constraints on the content in the information model. tomationML model. When integrating a new or changing
Beyond the formal description of the data model by related an existing resource, it may occur that InternalElements,
XML-Schema definitions which ensures syntactical and basic ExternalInterfaces or SupportedRoles have to be added. In
structural correctness it is required to define logical and each of these cases, dependencies (SystemUnitClasses, Inter-
functional dependencies within the system. In [31] it has been faceClasses, RoleClasses) have to be resolved by importing
shown that the Object Constraint Language (OCL) is capa- missing libraries into the model. In order to actually integrate
ble of describing logical dependencies within AutomationML the added objects into the model, relations to other objects
models to proove their semantical correctness. In a different in the system have to be defined. In AutomationML, this
context MathML was used within AutomationML to describe is accomplished by connecting the objects interfaces with
functional dependencies in mathematical formulas. Adding so called InternalLinks. On the other hand, when removing
MathML and OCL definitions directly to the AutomationML elements from the model by either removing or changing an
model facilitates the dynamic enhancement of the model. existing resource, these InternalLinks have to be removed prior
G. Interactivity to the execution of the remove operation in order to keep
the model in a consistent state. When removing elements, it
The ideas of the I40 require production systems to evolve is reasonable to also remove unused dependencies from the
continuously throughout every stage of their life-cycle. As a model in order to keep it as slim and performant as possible.
production system changes, agents have to be able to ma- These considerations introduce the need for some kind of
nipulate its model in order to provide up-to-date information dependency management. In order to encapsulate the model
to other entities. In a full fledged Self-X application on a from the agents, this maintenance should not be performed
dynamically changing production system the MAS has to by the agents themselves, but rather by the server that holds
rely on a powerful knowlegdebase to aquire and evolve the the model. This design adds up all the key concepts of ob-
system descriptions. This knowlegdebase is meant to contain ject orientation (identity, state, behavior, encapsulation). Each
all information that may be collected along the life-cycles object may be represented as an OPC UA node and may be
of the system. It ranges from static, centralized and well located on any resource hosting an OPC UA - AML Service.
structured models out of the systems engineering, along still Given this services agents may manipulate structure, views,
structured but frequently changing runtime-data to dynami- annotations and relations of elements forming knowledgebase
cally evolving and partially unstructured or even inconsistent systems.
support data. This tangle of data raises high demands on
the agility of the applied information models as well as the IV. I NTEGRATION OF P RODUCTION SYSTEM MODELS IN
information infrastructure for the aggregation and exploitation SHOP FLOOR FOR THE APPLICATION OF MAS S
of knowledge [35]. As we have shown before AutomationML As introduced in the previous section there are distinct sets
adresses the requirements on agility of the information models of requirements to a possible knowledgebase for MASs in a
with several concepts. The introduced changes in structural production system scenario. In the following three subsections
allocation, granularity, existence and availability of elements a testbed will be described that was used to evaluate how
in the system have additionally to be supported by the services the proposed solution meets those requirements. In the first
representing the model in an online application and form an subsection the general architecture for an integrated system of
agile information infrastructure. To enable a consistent and AutomationML, OPC UA and MASs will be described. In the
vivid runtime solution the characteristics of object orientation second subsection an example implementation of the proposed
have to be fully implemented serving the need to distribute architecture will be presented. In subsection three some use-
the information objects across the infrastructure. Identity and cases are described that were run on the presented lab size
state have already been applied but the encapsulation and production system.
behavior are partially described in the standard documents.
Also the server applications that were developed together with A. Architecture Proposal
the companion specification for AutomationML and OPC- The parts of the proposed architecture may be distinguished
UA didn’t address this characteristics. Therefor a customized into three major conceptual structures that are the execution
128
and control of the behavior of system parts, the management knowledgebase which is the cardinal gain of the proposed
of the integral system and the management of knowledge architecture. By the instantiation of the AutomationML model
about the system. The execution structure is located within within the information model of OPC UA all the knowledge
or close to the control system on the field or edge level of collected in the design process of the system may get published
the production system. Each appliance for the manufacturing to the servers of the proposed architecture and forms the
is self-contained with its control in industrial quality. Therefor distributed knowledgebase which is continuously evolved. For
it is connected to the hardware by field bus technologies and this purpose an OPC UA server software is used within the
meets constraints on real-time behavior and safety of the gov- architecture which serves AutomationML files according to
erned system part. For higher level applications each control the companion specification for AutomationML extended with
interfaces with a service platform e.g. a service API within the consistency preservation explained in section III G. A key
the control hardware. As a migration strategy a realization feature here is that within this model pointers to remote OPC
on state-of-the-art PLC runtime with an integrated OPC UA UA servers and nodes are represented. This way it realizes an
server per appliance is suggested. Due to the decoupling of arbitrary distributable model of the system with the expres-
components by vendor- and platform independent OPC UA siveness of AutomationML. Within this ecosystem the agents
architecture future technologies may transparently replace this are able to read AutomationML, browse the description of the
legacy systems. Each of these clusters is associated with field system, learn and reason upon the information and in a mature
level agents along with a model based description of their implementation enhance or improve the model. Thereby the
logical cooperation interfaces in AutomationML, rendering model dynamically evolves and satisfies the requirement of the
them intelligent system modules. The closer the agents and agility of information model and information infrastructure.
the model are integrated with the control hardware the more
modularity and adaptability is gained. B. Application Example
The second structure in the architecture is composed by The application presented here implements an examplary
hard- and software which enables the system to be run as version of MES agents in the sense of [36] and [14] integrated
an integrated whole. Its functions are on a higher level of with the mentioned technologies. It consists of a modular lab
abstraction of the production process and therefor called higher size production system of eight modules. The modules may be
level applications even though they are not necessarily superor- arranged in the laboratory arbitrarily. Each module represents
dinate to the field level. They contain the planning, operation, a manufacturing resource and is self-contained with a state-of-
supervision, service and maintenance of the production as well the-art control system and an integrated OPC UA Server which
as the production system. In this sense it heavily overlaps serves the control parameters of the PLC to the network (see
with traditional functions of manufacturing execution system figure 1).
(MES), supervisory control and data acquisition (SCADA)
as well as enterprise resource planning (ERP) technologies. C
Chute
Here functions like “smartness” and “self-X” of the produc-
IE PLC
tion and production system are realized by MAS. While the
interfaces to the control system are realized using OPC UA OPC TCP
the implementation of cooperative behavior has been based
Conveyor 3
on the middleware of the MAS. This design enables the Modbus TCP P
separation of concerns between different levels of operation Machine
and especially the loose coupling between data describing
the structural conditions of the system and the operational
data of the application executed on the system. This avoids
complex integration dependencies between system modules A B
Conveyor 1 Conveyor 2
and their application and supports platform independence and
maintainability of system parts.
The third and most novel structure of this proposal is Fig. 1. architecture of a single component in lab size production system
the information infrastructure. Exploiting the power of the
applied middleware solutions it is well tailored to the in- The modules are connected to an Industrial Ethernet infras-
tended purpose. While the information out of the semantical tructure. In the scope of the whole system the components
context of cooperation stays within the implementation of are integrated with an aggregating OPC UA server that holds
the MAS the communication with the information repository an AutomationML model (see listing 1). It contains an object
and the executive parts of the system is performed using tree with an overall description of the system and points to
the open industrial OPC UA technology. Special attention the modules’ servers (see figure 2). Changes in the physical
should be spent to the fact that OPC UA is not only used setup of the system have to be observed and reported to this
as transmission protocol of the data. The strong mechanism server. The MAS is built on JADE technology. The agents
of self-descriptiveness using the OPC UA information models are equipped with the feature to surf OPC UA networks and
enables the establishment of the distributed machine readable interpret discovered AutomationML descriptions.
129
Listing 1. AutomationML segments of production systems’ resource topology static sequence on a module. It forms the base of future work
of the authors about skill based discovery and commission of
agents on the job shop.
[..]
Project
OPC UA Services
Logistic Resource
Resource Resource
2) OPC UA Servers with AutomationML backend: To gain
tionML models the interactivity had to be realized for the
tionML engine joined with the OPC UA Stack implementation
of the MILO project an OPC UA server implementation
was created that allows the object-oriented interaction with
AutomationML object hierarchies. Thereby the agents may not
just acquire information from the model but rather reason upon
[..] the structures based on inference rules to gain new knowledge
and communicate this knowledge through the models.
[..]
1) Field Level Components: Each of the 8 modules contains opc.tcp://pplab_master:48010/
a single-board computer which operates an industrial PLC
[..]
runtime with a built-in OPC UA Server. The PLC program is
61131 [37] and stored in the related PLCOpenXML exchange
format. For each module there is an AutomationML object tree
variables as well as a description of the module behaviors [..]
controlled by the code linking all together (see listing 2).
the PLC and thereby exposed to the outside as a service
of the module. Deploying this AutomationML project to the
PLC programming environment and an aggregating OPC UA ns=4;s=PROGRAM1.productType
Server capable of AutomationML hierarchies constitutes a
service based interactive module with a description of its
skills. For the prototype presented here the control code 5b196fe6-83a5-4b6e-b011-
f21e3436ce1b
implemented only indecomposable behaviors of the module
while sequences of behaviors were realized outside of the field
control components. At the moment this is realized by an agent
not further introduced in this paper that knows to execute a
130
3) Applied agent types: Within this infrastructure of mod- transport (to the cheapest available resource), configure
ular self describing modules an MAS was set up that was the machine(s) and execute the appropriate process.
taught to interpret and execute a production process description • Product Management Agent (PMA): The PMA man-
given in AutomationML. The agents automatically acquired ages production orders. When an order is placed, the
the information of the process order, the capabilities of the PMA spawns new Pas.
modules and the path the product may travel through the • Directory Facilitator: The DF provides a central dis-
system. Therefor the following agent types were designed. covery service for services offered by other agents in the
MAS architecture.
Resource
Resource C. Application Scenarios
Management
Agent
Agent The presented testbed primarily serves experiments for
(optimization) methods of the industrial engineering. Different
setups and production program plans are executed on the
Logistic system by the agents. According to the physical arrangement
Agent the setups are published and updated to the system in form
of AutomationML models. Additionally an agent monitors
process variables and calculates several KPIs to monitor the
performance of a solution. This way the production system
Logistic may be easily reconfigured to simulate different layouts and
Product
Management
Agent programs for a production process. A second scenario that
Agent
was successfully implemented for a doctoral thesis is a fault
detection and classification system [38]. The applied MAS
tracked KPIs and component state indicators. This data was
Product
evaluated against constraints to identify errors. Based on the
Management
Agent fault register recovery strategies where performed by other
agents. A third scenario was implemented which realized
Fig. 3. Applied agent types automatic reconfiguration of the system. The modules and
products were equipped with fiducials and a computer vision
• Resource Agent (RA): RAs manage resources. They system based on openCV was implemented that tracks the
are responsible for the initialization of resources, the presence and positional relation of products and resources. To
execution of processes and the prevention of multiple simulate a fault the fiducial of a resource was covered and got
resource allocation. Resource agents serve as gateways invisible for the agents this way. A specific agent recognizes
between the field level and the MAS. the change and updates the AutomationML knowledgebase.
• Resource Management Agent (RMA): The RMA man- Supervisory agents then successfully adjust the production
ages the Resource Agents. Whenever a resource is added and transportation routes in the system based on the dynamic
to or removed from the (model of the) factory,the RMA model. The mentioned scenarios have proven applicable and
will adjust the number RAs accordingly. give the opportunity to be scaled to industrial implementation.
• Logistic Agent (LA): A LA offers some simple transport
service to the LMA. (simple in the sense of its abstract V. C ONCLUSION
representation within the model) In this paper it has been shown how the open industrial
• Logistic Management Agent (LMA): The LMA offers standards OPC UA for service-oriented communication as
a complex transport service to the PA. If a good has to well as AutomationML for semantic data exchange meet the
be moved from one place to another, the LMA will (use requirements of the establishment of a Smart Manufacturing
sophisticated optimization methods in order to) find the Component (I40 component). Providing a single service inter-
shortest path from target to destination. Afterwards it will face to the semantical description as well as to the behavior
search for LAs that offer the simple transport services within a self-contained component the infrastructure of Self-X
necessary in order to execute the complex transport and is realized upon already available standardized technologies.
call for proposals. The cheapest transport offer will be The presented solution scales from legacy to modern hard-
accepted and scheduled. ware as well as from small facilities to huge geographically
• Product Agent (PA): The PA is responsible for the spread value networks. The applied use cases have shown
production of a single product. Upon creation, it will the applicability of the approach and the ease of integration
determine the production steps that have to be performed with innovative applications. This proposal enables a lot of
in order to create the product (with respect to certain opportunities for the implementation of a knowledge base
constraints, e.g deadlines, costs or quality). It will then system for production and will be the basement for further
sequentially find RAs that offer the execution of the work of the authors on third party application integration and
next production step, call for proposals, commission the dynamically evolving skillsets of agents.
131
R EFERENCES ate data exchange format?,” in 2015 IEEE 13th International Conference
on Industrial Informatics (INDIN), pp. 492–498, July 2015.
[1] H. Kagermann, J. Helbig, A. Hellinger, and W. Wahlster, Recommenda- [23] L. Berardinelli, S. Biffl, A. Lüder, E. Mätzler, T. Mayerhofer, M. Wim-
tions for Implementing the Strategic Initiative INDUSTRIE 4.0: Securing mer, and S. Wolny, “Cross-disciplinary engineering with automationml
the Future of German Manufacturing Industry ; Final Report of the and sysml,” vol. 64, 01 2016.
Industrie 4.0 Working Group. Forschungsunion, 2013. [24] AutomationML e.V., “Automationml - current projects.”
[2] P. I. 4.0, “Structure of the administration shell - continuation https://www.automationml.org/o.red.c/projects.html, 2018. Accessed:
of the development of the reference model for the industrie 4.0 2018-05-10.
component.” ebook (PDF) available at https://www.plattform- [25] DIN, DKE, VDE, “German standardization roadmap on
i40.de/I40/Redaktion/EN/Downloads/Publikation/structure-of-the- industrie 4.0.” Available at https://www.din.de/de/forschung-
administration-shell.html, 01.04.2016. und-innovation/themen/industrie4-0/roadmap-industrie40-62178,
[3] DIN Spec, “91345: Reference architecture model industry 4.0 (rami 18.05.2018.
4.0),” DIN Spec, 2016. [26] W. Mahnke, S.-H. Leitner, and M. Damm, OPC unified architecture.
[4] A. Lüder, M. Schleipen, N. Schmidt, J. Pfrommer, and R. Henßen, “One Springer Science & Business Media, 2009.
step towards an industry 4.0 component,” in 2017 13th IEEE Conference [27] M. Schleipen, R. Henßen, O. Sauer, N. D’Agostino, M. Damm, A. Do-
on Automation Science and Engineering (CASE), pp. 1268–1273, Aug gan, J. Ladiges, C. Ewertz, A. Gössling, T. Holm, S. Hoppe, A. Lüder,
2017. N. Schmidt, and R. Wilmes, “Harmonisierung im Kontext Industrie 4.0
[5] DIN Spec, “16592: Combining opc unified architecture and automation AutomationML und OPC UA,” 06 2017.
markup language,” DIN Spec, 2016. [28] M. Schleipen and R. Drath, “Three-view-concept for modeling process
or manufacturing plants with AutomationML,” in 2009 IEEE Conference
[6] M. Wooldridge, An Introduction to MultiAgent Systems. Wiley Publish-
on Emerging Technologies Factory Automation, pp. 1–4, Sept 2009.
ing, 2nd ed., 2009.
[29] A. Lüder, N. Schmidt, and S. Helgermann, “Lossless exchange of graph
[7] R. Unland, “Chapter 2 - Industrial Agents,” in Industrial Agents
based structure information of production systems by AutomationML,”
(P. Leitão and S. Karnouskos, eds.), pp. 23–44, Boston: Morgan Kauf-
in 2013 IEEE 18th Conference on Emerging Technologies Factory
mann, 2015.
Automation (ETFA), pp. 1–4, Sept 2013.
[8] P. Leitão, “Agent-based distributed manufacturing control: A state- [30] H. Lieberman, “Using Prototypical Objects to Implement Shared Be-
of-the-art survey,” Engineering Applications of Artificial Intelligence, havior in Object Oriented Systems,” pp. 214–223, 1986.
vol. 22, no. 7, pp. 979 – 991, 2009. Distributed Control of Production [31] M. Schleipen, “A concept for conformance testing of automationml mod-
Systems. els by means of formal proof using ocl,” in 2010 IEEE 15th Conference
[9] P. Castellini, C. Cristalli, M. Foehr, P. Leitao, N. Paone, I. Schjolberg, on Emerging Technologies Factory Automation (ETFA 2010), pp. 1–5,
J. Tjønnås, C. Turrin, and T. Wagner, “Towards the integration of process Sept 2010.
and quality control using multi-agent technology,” in IECON 2011 - 37th [32] M. Sabou, F. Ekaputra, O. Kovalenko, and S. Biffl, “Supporting the
Annual Conference of the IEEE Industrial Electronics Society, pp. 421– engineering of cyber-physical production systems with the automationml
426, Nov 2011. analyzer,” in 2016 1st International Workshop on Cyber-Physical Pro-
[10] M. Onori, A. Maffei, and F. Durand, “The ideas plug and produce duction Systems (CPPS), pp. 1–8, April 2016.
system,” p. 339, 01 2013. [33] I. Grangel-González, D. Collarana, L. Halilaj, S. Lohmann, C. Lange,
[11] N. Antzoulatos, E. Castro, D. Scrimieri, and S. Ratchev, “A multi-agent M.-E. Vidal, and S. Auer, “Alligator: A Deductive Approach for the
architecture for plug and produce on an industrial assembly platform,” Integration of Industry 4.0 Standards,” in Knowledge Engineering and
Production Engineering, vol. 8, pp. 773–781, Dec 2014. Knowledge Management (E. Blomqvist, P. Ciancarini, F. Poggi, and
[12] P. Leitão, J. Barbosa, P. Vrba, P. Skobelev, A. Tsarev, and D. Kazanskaia, F. Vitali, eds.), (Cham), pp. 272–287, Springer International Publishing,
“Multi-agent system approach for the strategic planning in ramp-up 2016.
production of small lots,” in 2013 IEEE International Conference on [34] M. Obst, T. Holm, L. Urbas, A. Fay, S. Kreft, U. Hempen, and
Systems, Man, and Cybernetics, pp. 4743–4748, Oct 2013. T. Albers, “Semantic description of process modules,” in 2015 IEEE
[13] F. L. Almeida, B. M. Terra, P. A. Dias, and G. M. Goncalves, “Adoption 20th Conference on Emerging Technologies Factory Automation (ETFA),
issues of multi-agent systems in manufacturing industry,” in 2010 Fifth pp. 1–8, Sept 2015.
International Multi-conference on Computing in the Global Information [35] J. Scott, “A practical guide to microservices and containers.”
Technology, pp. 238–244, Sept 2010. eBook (pdf) available at https://mapr.com/ebooks/microservices-and-
[14] A. Lüder, A. Calá, J. Zawisza, and R. Rosendahl, “Design pattern for containers/, 2017.
agent based production system control – A survey,” in 2017 13th IEEE [36] V. G. F. . Spec, “2653 part3: Multi-agent systems in industrial automa-
Conference on Automation Science and Engineering (CASE), pp. 717– tion application,” VDI/VDE Spec, 2012.
722, Aug 2017. [37] I. E. Commission, “Programmable controllers - part 3: Programming
[15] “Innovative modelling approaches for production systems to raise vali- languages,” IEC Spec, 2013.
datable efficiency (improve)..” http://www.improve-vfof.eu/. Accessed: [38] H. Bayanifar, Agent-based mechanism for smart distributed dependabil-
2017-12-19. ity and security supervision and control of Cyber-Physical Production
[16] “Perform – production harmonized reconfiguration of flexible robots and Systems. 2017. xvi, 159 Seiten, Illustrationen.
machinery..” http://www.horizon2020-perform.eu. Accessed: 2017-12-
19.
[17] “Basissystem industrie 4.0..” https://www.basys40.de/. Accessed: 2017-
12-19.
[18] Y. Luo, Y. Duan, W. Li, P. Pace, and G. Fortino, “Workshop networks
integration using mobile intelligence in smart factories,” IEEE Commu-
nications Magazine, vol. 56, pp. 68–75, Feb 2018.
[19] G. Fortino, C. Savaglio, and M. Zhou, “Toward opportunistic services
for the industrial internet of things,” in 2017 13th IEEE Conference on
Automation Science and Engineering (CASE), pp. 825–830, Aug 2017.
[20] G. Fortino, W. Russo, C. Savaglio, W. Shen, and M. Zhou, “Agent-
oriented cooperative smart objects: From iot system design to implemen-
tation,” IEEE Transactions on Systems, Man, and Cybernetics: Systems,
pp. 1–18, 2017.
[21] A. Lüder, N. Schmidt, and R. Drath, Standardized Information Exchange
Within Production System Engineering, pp. 235–257. Cham: Springer
International Publishing, 2017.
[22] A. Lüder, N. Schmidt, and R. Rosendahl, “Data exchange toward plc
programming and virtual commissioning: Is automationml an appropri-
132