Self-organizing Service Structures for Cyber-physical Control Models
with Applications in Dynamic Factory Automation
A Fog/Edge-based Solution Pattern Towards Service-Oriented Process Automation
Maximilian Engelsberger and Thomas Greiner
Institute of Smart Systems and Services (IOS3), Pforzheim University, Tiefenbronnerstr. 65, Pforzheim, Germany
Keywords:
Dynamic Service Management, Service Orchestration, Service Placement, Service Embedment, Bubble
Model, Edge/Fog Computing, Factory Automation, Cyber-physical Production Systems, IT/OT Convergence,
Industry 4.0, IoT / IIoT.
Abstract:
The convergence of information technology and operational technology is a strong force in fabric automation.
Service-oriented architectures and cloud computing technologies expand into next generation production sys-
tems. Those technologies enable a lot of new possibilities; such as high agility, global connectivity and high
computing capacities. However, they also bring huge challenges regarding flexibility and reliability through
increasing system dynamics, complexity and heterogenity. New solution patterns are needed to conquer those
challenges. This paper proposes a new fog-oriented approach, which shows how future production systems,
that are often called cyber-physical production systems, can deal with dynamically changing services and in-
frastructure elements. The goal is to provide an adequate degree of flexibility and reliability across the whole
production lifecycle. Therefore, an event property model (“bubble model”), a multi-criterial evaluation metric
and extensions to Kuhn-Munkres and Add algorithm are described. The overall concept is evaluated by an
application example from the field of process engineering. With the help of practical case studies and dynamic
system simulations, qualitative results are gained.
1 INTRODUCTION
Service-oriented architectures (SOA) and cloud com-
puting technologies expand into next generation pro-
duction systems (Broy, 2010). As part of a bigger
transformation process, information technology (IT)
and operational technology (OT) converge. This gen-
erates a wide range of new possibilites within fac-
tory automation, which is traditionally very static,
isolated and often based on proprietary technologies
(Kowalewski, 2015). With a wide use of IT in fab-
ric automation, a lot of standards and quasi-standards
can be used. Thereby, a central technology is the In-
ternet Protocol (IP) family (Mor et al., 2016). This
leads to completely new possibilites like better system
agility, higher vertical connectivity, the use of huge
computing capacities and storage space etc. (Rug-
gerie et al., 2016). This transformation, which is
frequently called Industry 4.0, will finally result into
a lot of new business cases. Nevertheless, such a
process of change brings up a lot of risks to an ap-
plication domain, which is naturally depending on
high reliability and robustness. Eminent factors that
endanger reliability are increasing system dynamics,
complexity and heterogenity. This ranges from small
sensor nodes to big datacenters and from tiny PID-
controllers (Proportional-Integral-Derivative) to huge
optimization algorithms. In cyber-physical produc-
tion systems (CPPS) all physical processes and struc-
tures are mapped to virtual control models, which rep-
resent and control state and behavior of those entities
(Alur, 2015). These models may be encapsulated into
services, which need to be managed by the system.
Especially, all questions of service orchestration and
infrastructure utilization must be automatically recon-
figured under dynamic changes. Those might be
• Changes on functional/non-functional elements
• Changes on infrastructure and network
• Changes on sensors, actuators and smart objects
If IT will be useful, it must provide answers to the
biggest cost factors in factory automation through-
out the complete lifecycle of a production plant: En-
gineering efforts, production stops, maintenance etc.
Therefore, it is meaningful to automate as many of
the technical low-level tasks assigned to this costs, as
238
Engelsberger, M. and Greiner, T.
Self-organizing Service Structures for Cyber-physical Control Models with Applications in Dynamic Factory Automation - A Fog/Edge-based Solution Pattern Towards Service-Oriented
Process Automation.
DOI: 10.5220/0006365502660274
In Proceedings of the 7th International Conference on Cloud Computing and Ser vices Science (CLOSER 2017), pages 238-246
ISBN: 978-989-758-243-1
Copyright © 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
Cloud
Fog/Edge
Field/Mist
Embedment / Placement
Physical IT-/OT-node
Service with C/P-model
Physical network link
Meshed logical connections
Requests with data
/
Figure 1: CPPS system topology showing physical and log-
ical structures of services and infrastructure.
possible. From a service-oriented point of view, many
of those event-triggered tasks can be summarized by
the following questions:
• Where is data acquired?
• Where can data be processed?
• Which requirements need to be fulfilled?
• Where is the processing output needed?
Thus, new design patterns for production systems
are needed, answering those questions for all service-
based control models in a situation-aware manner.
Mechanisms of self-adaptation must orchestrate the
services utilizing the cloud, the fog and the local
operational field. The goal is to automatically re-
arrange the service structure on the system’s phys-
ical computing infrastructure, in order to reach the
needed flexibility and reliability over the whole sys-
tem lifecycle. The contribution of this paper is a new
fog/edge-oriented approach, showing how future pro-
duction systems can deal with a wide range of dy-
namic events. The method described here presents an
advanced variant of the approach described in (En-
gelsberger and Greiner, 2017), now revised from a
service-oriented field of view and extended by par-
tial corrections and improvements. In section 4.1,
an event property model is proposed, which allows
to characterize different events in a CPPS by a set of
different properties. Further, in section 4.2, a multi-
criterial matching metric is described, which allows
to evaluate functional and non-functional properties
of service-based control models. Further, in sections
4.3 and 4.4, two algorithmic extensions are described,
realizing dynamic service operations in order to op-
erate against disturbing run-time events. Finally, in
sections 5 and 6, the concept is evaluated by practi-
cal case studies from the field of process engineering
and evaluated through dynamic system simulations,
where quantitative results are gained.
2 RELATED WORK
The service-oriented management of IT-resources is
a well known concept in cloud computing. It brings
a lot of flexibility in terms of scaling the provid-
ing service resources to its current needs by utiliz-
ing the available (virtualized) hardware infrastructure
(Buyya, 2010). Fog/Edge computing techniques dis-
solve the borders between local and remote cloud
computing resources (De Meer et al., 2006). Existing
approaches are limited to classical IT-resources, like
storage space, computing capacity and bandwith etc.
(Dubois et al., 2016), (Wang et al., 2015). They do
not consider the special properties and requirements
of CPPS like:
• Field process dynamics
• Real-Time- and Safety-Criticality
• Heterogeneous infrastructure and services
• Sensor- and actuator properties
• Cyber-physical model properties
There are also many approaches using load balanc-
ing techniques for increasing reliability and stabil-
ity (Oueis et al., 2015). In load balancing, there are
working jobs (loads) which are distributed to cur-
rently available worker nodes (Oueis et al., 2015). In
contrast to load balancing, the proposed method does
not only handle dynamic loads, but creates and opti-
mizes the service-structures underneath those work-
ing packets, see Fig. 1. Further, the presented topic
has of course a strong relation to the field of ser-
vice orchestation (Jain et al., 2016). Most methods in
this field consider dynamic changes in the services or
IT-infrastructure, but they do not offer suitable solu-
tions to dynamic events in the OT-infrastructure, like
from sensors, actuators or smart objects. In contrast
to that, the described approach allows to model and
handle functional and non-functional, CPPS-relevant
service-resources, as described in the resource type
model (Engelsberger and Greiner, 2016). This model
expands the classical service-resource model from
cloud computing with the service-resources appearing
in a CPPS. An approach for self-organizing service
structures for the application in CPPS is described in
(Engelsberger and Greiner, 2016) as well. In this ap-
proach, service-replication and placement is covered,
but a multi-criterial evaluation metric and an approach
for dynamic embedment of service-resources is miss-
ing. Service embedment allows to evaluate the most
suitable service instance, which currently exists in the
system, in order to perform a certain task, with re-
spect to actual requirements and run-time properties.
Service placement allows to (re-)arrange the services
Self-organizing Service Structures for Cyber-physical Control Models with Applications in Dynamic Factory Automation - A
Fog/Edge-based Solution Pattern Towards Service-Oriented Process Automation
239
itself, if a task’s requirements are not longer fulfilled,
on a given node. The missing model of different
event properties and their combination are limitations
of earlier approaches in this field (Engelsberger and
Greiner, 2017). Some of the existing work focus on
techniques, which are described as service placement
in this work: (Mor et al., 2016), (Oueis et al., 2015),
(Wang et al., 2015). Other authors focus on tech-
niques, which are described as service embedment in
this work: (Taherkordi and Eliassen, 2014), (Bosman
et al., 2015), (Zhang and Cai, 2010). A comprehen-
sive method, combining embedment, placement and
a multi-criterial evaluation metric as a comprehensive
solution pattern for CPPS is still not visible.
3 BASIC ALGORITHMIC
APPROACHES
Because of the flat and decentral organizational
structure of CPPS, meshed service-structures evolve.
These services and their interconnections can be mod-
eled and manipulated effectively by the use of graph
theory. For different management operations, there
exist algorithmic approaches.
The main task performed by the service embedment
is to match the requirements of a set of given pro-
cess control sequences to the actual properties of a
set of currently available service instances. Such
n:m-shaped matching problems can be solved by the
Kuhn-Munkres or Hungarian algorithm, which is well
known in graph theory (Thulasiraman et al., 2016).
It solves a combinatorial optimization problem, by
maximizing the matching - or sum of selected edge-
weights - between two bipartite sets of vertices. That
edge-weights allow to assign arbitrary metrics be-
tween each possible two-set of vertices. Because of
the bipartite structure of the graph, which is a must for
the algorithm, it is not possible to have more than one
edge between one possible pair from partition 1 and
partition 2, see Fig. 2. Thats why another approach is
needed, in order to apply the proposed multi-criterial
metric on the given single edge-weight graph struc-
ture.
Within service placement, the vertices of a meshed
service-structure has to be placed on a meshed node-
structure of the physical infrastructure. There are al-
ready existing methods to place groups of providing
services on groups of consuming services, in a way,
that the overall costs are minimized. This is equiva-
lent to the p-Median problem (Krishna, 2014). One
possible solution method is the Add algorithm, which
needs an input data structure in form of p providers
and q consumers (Domschke and Drexl, 1996). Be-
1
0
A
1
B
2
C
w1=[c1, c2, c3, ...]
w2
w3
w4
w5
Partition 1:
Servicetype candidates
from control sequences
Partition 2:
Global available
service instances
Figure 2: Combination of bipartite graph matching problem
with multi-criteria matching metric.
cause of the flat and decentral structure of a CPPS,
services are often providers and consumers at the
same time. Hence, there is no fixed hierarchical struc-
ture between the services, but a meshed one, which
changes its structure over time. This problem struc-
ture is not directly applicable to existing methods,
like the Add algorithm. Thats why a new approach
is needed, which analyzes the given service structure
and transforms it in a form, where already existing
solutions can be applied.
4 IMPROVED APPROACH FOR
SELF-ORGANIZING SERVICE
STRUCTURES IN DYNAMIC
FACTORY AUTOMATION
The following subsections explain the improved ap-
proach for self-organizing service structures in dy-
namic factory automation. In 4.1, a new event prop-
erty model is described. In 4.2, the multri-criterial
matching metric is introduced. The metric is used in
4.3 for the extension of the Kuhn-Munkres algorithm.
Finally, in section 4.4, the extension of the Add al-
gorithm for the use of meshed service structures is
described.
4.1 Event Property Model
In CPPS, dynamic events may occur during the
factory’s run-time. In the described event property
model (“bubble model”), run-time starts with the
first module going in operation and ends with the
last module, going out of operation. This includes
all events during production phase, as well as
functional/non-functional changes and maintenance.
During production-phases, the sensitive modules
and their downstream dependancies are locked
against dynamic management operations in order to
CLOSER 2017 - 7th International Conference on Cloud Computing and Services Science
240
Intention
Domain
Organization
Function-TypeDirectness
Source
Planned Unplanned
Direct Indirect
Intrinsic Extrinsic
Cyber Physical-IT Physical-OT
Composition Behavior
Non-
functional
Functional
Figure 3: Event property model (“bubble model”) with rel-
evant properties.
not endanger running charge-processing. Only on
emergency exceptions, it is possible to remove the
management locks: (A) To prevent process-faults
or (B) to minimize time to get back to a working
processing. The model itself allows to characterize
different dynamic events, from various sources, in
a formalized way, see Fig. 3. In the following,
certain properties and assigned valid values are given,
followed by a short explanation:
• Intention: An event may be planned or unplanned
• Source: An event may be triggered system-
intrinsic or -extrinsic
• Domain: An event may occur in the Cyber-,
Physical-IT- or Physical-OT-domain
• Organization: An event may be described as
compositional or behavioral
• Directness: An event may be triggered direct or
indirect by other events
• Function-Type: An event might be of functional
or non-functional type
With the given event property model, it is possible
to characterize and compare different run-time events
by their type, domain, intention etc., by merging sin-
gle bubbles to complex structures. This can be used
to construct a complete set of proto-cases, showing all
principal types of events, which the proposed method
can deal with. The model is applied to the case stud-
ies, discussed in section 5.
4.2 Functional- and Non-functional
Multi-criterial Metric
In this section, a multi-criterial evaluation metric for
functional- and non-functional pairs of requirements
and properties is described with some improvements
and corrections to (Engelsberger and Greiner, 2017):
First, a set of matching criteria c ∈ C is defined as the
following 4-tupel:
c = (δ, µ, ψ, ϕ) (1)
In the following, the parameters from the 4-tupel c ∈
C are defined: Given deviation
δ ∈ R
δ >= 0 (2)
describing the deviation between reference require-
ment
r
re f
∈ R
r
re f
> 0 (3)
and actual property
p
act
∈ R
p
act
>= 0 (4)
of a given pair of sequence and service instance. With
the following meanings:
• δ = 0 : no fulfillment
• (δ > 0) ∧ (δ < 1) : partial fulfillment
• δ = 1 : full fulfillment
• δ > 1 : overfulfillment
Next, given a mandatory flag
µ ∈
{
0, 1
}
(5)
In positive logic, 0 means not mandatory and 1 means
mandatory. Given
ψ ∈ R
ψ >= 0 (6)
defining the priority of the given criteria. Next, given
a flag
ϕ ∈
{
0, 1
}
(7)
defining if the given criteria is functional (with value
1) or non-functional (with value 0). In the following
steps, the calculation of the service matching metric
SMM(C), is described. In earlier work, an alternative
metric is named resource matching metric RMM(C),
see (Engelsberger and Greiner, 2017). For instance,
the relative deviation of the actual metric value from
the reference value is determined:
δ =
p
act
r
re f
(8)
As next step, the priority and mandatory-flag, which
are assigned to one criteria, are taken into respect:
ζ =
(
δ · ψ : (µ = 1 ∧ δ >= 1) ∨ (µ = 0)
0 : else
(9)
Therefore, the interim value ζ is the product of the
deviation δ and the priority ψ, if the mandatory-flag
Self-organizing Service Structures for Cyber-physical Control Models with Applications in Dynamic Factory Automation - A
Fog/Edge-based Solution Pattern Towards Service-Oriented Process Automation
241
µ equals 1 and the deviation δ is equal to or greater
1. This means if the mandatory flag is set, consider
the deviation only, if the requirements are fulfilled
or over-fulfilled (>= 1). Otherwise set ζ to zero. If
mandatory flag isn’t set, consider the deviation, inde-
pendent from its actual value. Next, the interim value
is normalized to the maximum value of all given cri-
teria:
ζ
norm
=
ζ
ζ
max
(10)
As a final step, the service matching metric SMM(C)
over all c ∈ C is determined:
SMM(C) =
N=
|
C
|
∑
n=1
ζ
norm
(11)
The resulting metric SMM(C) is used in the subse-
quently described algorithmic approach.
4.3 Extension of the Kuhn-Munkres
Algorithm With a Multi-criterial
Metric
This section describes the extension of the Kuhn-
Munkres algorithm with the multi-criterial metric
from section 4.2. The goal is to have multiple criteria
for a single matching between control-sequences and
service instances, as described in 3. In the following
paragraphs, the relevant CPPS architecture elements
are explained: A possible factory concept consists of
modular electro-mechanical production units, which
are networked via Ethernet (IEEE 802.3). Each of
the modules is equipped with its own Process Con-
trol Model (PCM), including a formalized set of con-
trol sequences, see (Engelsberger and Greiner, 2017).
In order to execute a specific sub-production task, a
set of requirements is assigned to each control se-
quence. An example sub-production task might be
“fill media reactor I with 1500ml of liquid media type
112 from inlet A”. Further, service-instances, which
are also named cyber-resources in (Engelsberger and
Greiner, 2017), exist within the global scope of the
CPPS. A set of functional and non-functional proper-
ties are assigned to each of that service instances. The
task, which dynamically evaluates the best-matching
service instance to each of the existing control se-
quences, is called embedment operation. Therefore, a
bipartite graph structure G(V, E) has to be defined. “A
graph G is called bipartite, if the number of vertices
in V is the union of two disjunct sets S and T , in such
way, that each edge in the set of E has exactly one
ending vertex in S and one vertex in T ”(Teschl and
Teschl, 2006) and (Engelsberger and Greiner, 2017):
G = G (S ∪T, E) (12)
In the described context, the vertices of set S are
control sequences with assigned requirements and the
vertices of set T are service instances with assigned
properties. Edges between a certain set of two vertices
from T and S represent one possible matching. The
aggregated metric SMM(C), containing its individual
criteria values, is now assigned to each of the existing
edges. In order to preserve correct functionality of
the matching algorithm, it is important not to violate
the rules of the bipartite graph structure, like it was
defined above. The proposed approach fulfills this re-
quirement. After executing the extended method with
the bipartite graph input data structure, the maximum
matchings between the sequences and instances are
known. Now, the modules can connect to the service
instances and send working packets to it in order to
solve sub-processing tasks.
while control != abort do
for all sequence controls SC do
G.addVertex(sc);
S.addVertex(sc);
for all service instances SI do
G.addVertex(si);
T.addVertex(si);
for all sequence controls SC do
for all service instances SI do
SMM = 1- SMM(si, sc);
e1 = G.addEdge(si, sc);
G.setEdgeWeight(e1, SMM);
matching = KuhnMunkresMinimalMatching(G,S,T);
4.4 Extension of the Add Algorithm to
Meshed Service Structures
This section describes the extension of the Add algo-
rithm in order to make it able to handle meshed ser-
vice networks as input data structure. This is done by
a series of additional isolation- and weighting-steps
of a given service-structure. The idea of the so called
placement operation is to optimize the requirements-
fulfillment of a CPPS by a dynamical placement or re-
arrangement of unconstrained service instances on the
physical infrastructure. Service-instances are uncon-
strained, if they havn’t any physical dependancies to a
certain node, like specific I/O-devices, real-time con-
straints etc. Typical unconstrained service instances
might be - but not limited to - tasks like analyzation,
optimization, archiving, monitoring etc.
The problem is equal to the p-Median problem, where
p ∈ N \ {0} providing services are placed in such a
position to q ∈ N \ {0} consuming services, that the
resulting distance costs are minimal (Laporte et al.,
2015). As definition for the distance costs the multi-
criterial matching metric from section 4.2 can be
CLOSER 2017 - 7th International Conference on Cloud Computing and Services Science
242
logical service structure
with hierarchical-topology
physical node structure
with mesh-topology
Placed logical service
structure on physical
node infrastructure
Add-Algorithm
Subtopology
Isolation
and Weighting
Subtopology
Consumer
Provider
Figure 4: Add algorithm with subtopology isolation and
weighting extension.
used with appropriate criteria. The metric allows
to define QoS-specific (non-functional) parameters
like actual traffic, latency, jitter, timing-constraints
etc., as well as functional parameters like sensor-
accuracy, actuator-degeneration-criteria and so on. As
described in (Engelsberger and Greiner, 2017), exist-
ing solution methods are able to solve the single-stage
p-Median problem, when a set of possible locations
for the providing services and a set of consuming ser-
vices with fixed locations are given. As explained, the
difference to this concept is, that in a realistic SOA-
based production system, certain services are provider
and consumer at the same time. An example might
be an agitation-optimization service, which consumes
data from a media-temperature service and provides
data to the agitator service. If this principle is ex-
panded to a complete production-system, this results
in a complex meshed service-structure, like in Fig. 4,
section 4.4. The Add algorithm needs an input data
structure in the form of p providers vs. q consumers,
so the meshed service structure needs to be analyzed
and transformed in an appropriate way. The following
algorithmic extension is proposed:
In step 1, the isolation and weighting process is done
by a graph-based breadth-first seach, which identifi-
cates the service instances without dependancies. The
found services are sorted by their placement priority.
In step 2, their dependancies are determined. With
this information, a new subtopology is built. The
subtopologies are used to generate a reference cost
matrix, by aggregating the connection requirements.
The reference cost matrix is defined as a single-row
matrix R
N
res,x
with:
r
i
∈ [0, 100] (13)
In step 3, the costs of all physical connections in
the network are determined. These connection costs
are stored in the actual connection cost matrix A
N×M
pnet
with:
a
i j
∈ [0, 100] (14)
In step 4, the difference costs between R
N
res,x
and
A
N×M
pnet
are calculated. Therefore, a fitting function is
used, so all resulting values are within
d
i
∈ [0, 100] (15)
The fitting function is defined as
d
i j
= −
1
2
· (r
i
− a
i j
) + 50 (16)
The result is the difference cost matrix D
N×M
costs
with:
d
i j
∈ [0, 100] (17)
The difference cost matrix is a valid input data struc-
ture for the Add algorithm, as shown in the following
algorithm:
while allvisited != true do
V2V_queue += V_start;
BreadthFirstSearchIteration(G_srv,
V2V_queue, V_actual);
if !(IOconstrained(V_actual)) then
Unconstr_queue += V_actual;
for all V in Unconstr_queue do
CalculateQoSPriority(V);
SortByQoSPriorityDesc(Unconstr_queue);
for all V in Unconstr_queue do
if !(PartOfPlacedSubStruct(V)) then
G_subset = getWeightedSubset(G_srv,
G_phys, V);
placement = Add(G_subset);
The extended method is now able to place services
from meshed structures in such way on the physical
network nodes, that the connection costs are mini-
mized with respect to the service priority.
5 APPLICATION CASE STUDY
This section picks up the CPPS system description
from 4.3 and describes the derived case studies. The
CPPS testing facility itself is described in detail in
(Engelsberger and Greiner, 2016). It is a modular
and decentral system architecture with very heteroge-
neous system nodes, from very small real-time con-
trollers in the field to very huge IT-resources in the
Self-organizing Service Structures for Cyber-physical Control Models with Applications in Dynamic Factory Automation - A
Fog/Edge-based Solution Pattern Towards Service-Oriented Process Automation
243
Planned
Direct Cyber
Control model
join/update
New/changed
requirements
Select
operations
Measure
Analyze
Plan
Place, Embed
Execute
(direct) event(s):
Optimization model
join/update
Load new models,
switchover
Event properties:
Figure 5: Example 1: Agitation model updates.
Unplanned
Indirect Physical-OT
Pump #1 motor coil
anomaly detection
Pump #1 will
immediately fail
Use fault control model
(with special coil-control
pattern)
Load new model,
switchover
Measure
Analyze
Plan
Execute
Control model
join/update
New/changed
requirements
Select
operations
Measure
Analyze
Plan
Place, Embed
Execute
1st (direct) event:
2nd (indirect) event:
Event properties:
Figure 6: Example 2: Motor coil anomaly detection.
local fog and in the cloud. The described application
scenario comes from process engineering and repre-
sents a production process of liquid media. Fig. 7
shows a cut-out of this process with two processing
stages: The first is the boiling reactor, which is encap-
sulated in a production module. It is able to heat-up
liquid media to a given reference temperature. The
second is the agitation module, which is able to ag-
itate the media to a given reference viscosity. These
modules depend on field devices like networked sen-
sors, actuators and smart objects. Corresponding to
the event property model in 4.1, this group of devices
is summarized as physical OT domain. As described
in section 1, such systems are getting more and more
complex and fragile with size and number of dynamic
changes. In the subsequent sections, two examples
are given, what may change during run-time and how
the method can handle those dynamic events.
Boiling reactor
Agitation module
Physical OT-Domain
M
T1
R1
A1
P1 P2
H1
R1
R2
Figure 7: Relevant cut-out of the production process includ-
ing boiling reactor and sediment removal whirlpool.
5.1 Example 1: Agitation Model Update
The first example focuses on the agitation module.
Two parallel occuring events of the same type are
taken into account: (A) A control model join/update
and (B) an optimization model join/update, see Fig.
5. Both are part of the engineering or maintenance
process, happening during lifetime of the facility, but
not necessarily during a running operation. The con-
trol model of a CPPS-module represents the control
sequences, needed to control a certain aspect of the
system: The agitation unit, in this case. The control
model is encapsulated as a software service and in-
teracts with other services in the system-wide meshed
service-structure. The optimization model is a ser-
vice as well. In this case it is supposed to compute
and optimize the operational parameters for the agi-
tation control model, taking into account actual con-
ditions. This could be the actual media temperature,
which is served by another service from the boiling
reactor. Characterized by the event property model in
4.1, both events are of intention “planned”, by their
engineering character. Further, they are of direction
“direct”, because their update is the immediate cause
for the event. Finally, the event source is the “cyber-
domain”, because they are triggered by model code
changes and not by any physical resources.
If the models change, so their requirements change
and it might become necessary to perform some
service-structure changes. In the given example a re-
arrangement of the place, where the service-models
are located, might increase the requirement fulfill-
ment. The control model needs to be placed on a
node, where all real-time-requirements are fulfilled,
in order to control the agitation unit. The optimiza-
tion model needs to be placed on a node, with a cer-
tain computing capacity, in order to process the op-
timization algorithm in an appropriate time. Finally,
the dynamic operations, calculated by the algorithm,
are deployed on the system and the switchover to the
new service configuration can be done.
CLOSER 2017 - 7th International Conference on Cloud Computing and Services Science
244
Figure 8: Results from the simulation-run with service em-
bedment algorithm based on a multi-criterial metric.
5.2 Example 2: Pump Motor Coil
Anomaly Detection
The second example focuses on the agitation module,
as well. One direct source event occurs: An anomaly
of media pump #1’s motor coil is detected, see Fig. 6.
The sensor says pump #1 will immediately fail. As an
reaction it is assumed to be possible to switchover to a
fault control model. The selection of the fault control
model triggers a second (now indirect) event, which
is to update the control model. This update brings
up new requirements, which the system needs to ful-
fill. The event is of “intention”unplanned, because no
one decided to produce this motor coil anomaly. It is
of directness “indirect”, because a first source event
triggers a second model join event. In this case, the
newly joined model needs to be placed on the system,
depending on its requirements. Further, the newly
joined model needs to be embedded into the other ser-
vices, in order to use it. Finally, the dynamic opera-
tions, calculated by the algorithm, are deployed on the
system and the switchover to the new service config-
uration can be done.
6 SIMULATION AND RESULTS
In this section, the previously described metric, al-
gorithms and use-case-scenarios are transformed into
software simulations, in order to evaluate the con-
cepts against non-adaptive service-structures. For this
purpose a simulation environment is written in Java,
where it is possible to describe infrastructure and ser-
vices from IT- and OT-domain and to realize mod-
els of production-processes via appropriate interfaces.
Furthermore, the simulation environment offers inter-
Figure 9: Results from the simulation-run with service
placement algorithm based on a multi-criterial metric.
faces to test and evaluate different metrics and algo-
rithms of adaptive behavior, as well as interfaces to
generate dynamic events and/or to collect them from
physical infrastructures.
In order to evaluate the embedment operation, 0 to
10 joins and leaves of control sequences and service
instances, as well as relevant properties and require-
ments are pseudo-randomly generated, within 20 sim-
ulation cycles. The results are shown in Fig. 8.
In order to evaluate the placement operation, joins and
leaves of control sequences and service instances, as
well as changes of the transport costs of the physi-
cal network are simulated. Therefore, 0 to 10 joins
and leaves of control sequences and service instances,
as well as relevant properties and requirements are
pseudo-randomly generated, within 20 simulation cy-
cles. The results are shown in Fig. 9.
Both graphs show the Requirements Fulfillment Index
(RFI), an evaluation metric which aggregates the re-
quirement fulfillment for all control sequences, across
the overall CPPS. The evaluation metric is calculated
by the following equation: RFI = AVG (
∑
SMM (C))
over all managed service instances.
7 CONCLUSIONS
Future cyber-physical production systems (CPPS) are
very heterogenous and dynamic across their whole
lifecycle. From the technology point of view, IT and
operational technology (OT) converge. That makes
new IP- and SOA-based technologies accessible in
automation applications. As a side-effect, CPPS get
more and more complex and fragile. In this paper, a
new solution-pattern is proposed, showing how future
production systems can deal with disruptive run-time
Self-organizing Service Structures for Cyber-physical Control Models with Applications in Dynamic Factory Automation - A
Fog/Edge-based Solution Pattern Towards Service-Oriented Process Automation
245
events by dynamic adaptations on service level, uti-
lizing the possibilities of the local operational field,
the local fog/edge and the global cloud. A new fog-
oriented approach is described, showing how an ad-
equate degree of reliability and flexibility is possible
under dynamic changes. For this purpose, an event
property model (“bubble model”), a multi-criterial
evaluation metric and two extensions to already ex-
isting algorithmic approaches are shown, which re-
alize operations of service structure adaptation. The
method is tested under realistic conditions by an ap-
plication example from the field of process engineer-
ing. Within two case studies and software simula-
tions, the concept is evaluated and quantitative results
are gained. The results show, that both of the de-
scribed algorithmic extensions are able to extend the
RFI against non-adaptive approaches. Further inves-
tigations are needed to learn more about mutual re-
actions between both algorithms and different CPPS-
specific cases.
ACKNOWLEDGEMENTS
This work is supported by the Baden-Wuerttemberg
Ministry of Science, Research and the Arts (MWK)
within the scope of Cooperative Research Training
Group.
REFERENCES
Alur, R. (2015). Principles of Cyber-Physical Systems. MIT
Press.
Bosman, J., van den Berg, H., and van der Mei, R. (2015).
Real-Time QoS Control for Service Orchestration.
IEEE 27th International Teletraffic Congress.
Broy, M. (2010). Cyber-Physical Systems (acatec DISKU-
TIERT). Springer.
Buyya, R. (2010). Cloud Computing: Principles and
Paradigms. John Wiley & Sons.
De Meer, H., Sterbenz, J., and EuroNGI. (2006). Self-
Organizing Systems: First International Workshop,
IWSOS 2006 and Third International Workshop on
New Trends in Network Architectures and Services,
EuroNGI 2006, Passau, Germany, September 18-20,
2006, Proceedings. Computer Communication Net-
works and Telecommunications. Springer.
Domschke, W. and Drexl, A. (1996). Logistik: Standorte.
Oldenbourgs Lehr- und Handb
¨
ucher der Wirtschafts-
u. Sozialwissenschaften. De Gruyter.
Dubois, D. J., Valetto, G., Lucia, D., and Nitto,
E. D. (2016). Mycocloud: Elasticity through
Self-Organized Service Placement in Decentralized
Clouds. IEEE 8th International Conference on Cloud
Computing (CLOUD).
Engelsberger, M. and Greiner, T. (2016). Application-
independent Approach for the Dynamic Management
of IT-Resources in Cyber-Physical Systems.
Engelsberger, M. and Greiner, T. (2017). Handling Strat-
egy of Dynamic Resource Events in Cyber-Physical
Production Systems by a Multi-Criterial and Multi-
Operational Approach. Industrial Technology (ICIT),
2017 IEEE International Conference on.
Jain, P., Datt, A., Goel, A., and Gupta, S. C. (2016). Cloud
service orchestration based architecture of OpenStack
Nova and Swift. Advances in Computing, Communi-
cations and Informatics (ICACCI), 2016 International
Conference on.
Kowalewski, S. (2015). Cyber-Physical Systems - A UMIC
Perspective. Technical report, RWTH AACHEN UNI-
VERSITY.
Krishna, P. (2014). Challenges, Opportunities, and Dimen-
sions of Cyber-Physical Systems. Advances in Sys-
tems Analysis, Software Engineering, and High Per-
formance Computing.
Laporte, G., Nickel, S., and da Gama, F. (2015). Location
Science. Springer International Publishing.
Mor, N., Zhang, B., Kolb, J., Chan, D. S., Goyal, N., Sun,
N., and Lutz, K. (2016). Toward a Global Data Infras-
tructure. IEEE Internet Computing, 20(3).
Oueis, J., Strinati, E. C., and Barbarossa, S. (2015). The Fog
Balancing – Load Distribution for Small cell Cloud
Computing. Vehicular Technology Conference (VTC
Spring), 81:1–6.
Ruggerie, M., Malaguti, G., Dariz, L., and Selvatici, M.
(2016). In-Tractor Cloud: A Vision of Service-
Oriented System Design - Enabled by High-Speed In-
Vehicle Networks for a Safer Task- and Machine Man-
agement. SAE Technical Papers scopus(scholar).
Taherkordi, A. and Eliassen, F. (2014). Models@run.time
for Creating in-Cloud Dynamic Cyber-Physical
Ecosystems. volume 6. IEEE.
Teschl, G. and Teschl, S. (2006). Mathematik f
¨
ur Infor-
matiker. Springer.
Thulasiraman, K., Arumugam, S., Brandst
¨
adt, A., and
Nishizeki, T. (2016). Handbook of Graph Theory,
Combinatorial Optimization, and Algorithms. CRC
Press.
Wang, S., Urgaonkar, R., Zafer, M., He, T., Chan, K., and
Leung, K. K. (2015). Dynamic Service Migration in
Mobile Edge-Clouds.
Zhang, Y. and Cai, W.-d. (2010). Criticality-Driven QoS
Adaptive Dynamic Resource Management for Dis-
tributed and Embedded Safety and Mission Critical
Systems. International Conference on New Trends in
Information Science and Service Science (NISS), 4.
CLOSER 2017 - 7th International Conference on Cloud Computing and Services Science
246