Highlights

Model management to support systems engineering workflows using ontology-based knowl-

edge graphs

• A model management strategy to store and provide access to knowledge about system engineering workflows
that uses ontology-based knowledge graphs.

• An ontology that defines basic concepts and their relationships to support the management of system engineering
tasks.
• A toolchain that provides capabilities for system engineering design, enact and reason on workflows and their
related artefacts and data.
Model management to support systems engineering workflows
using ontology-based knowledge graphs

ARTICLE INFO ABSTRACT

Keywords: System engineering has been shifting from document-centric to model-based approaches where
model management assets are becoming more and more digital. Although digitisation conveys several benefits,
ontology it also brings several concerns (e.g., storage and access) and opportunities. In the context of
process modelling Cyber-Physical Systems (CPS), we have experts from various domains, executing complex
knowledge graph workflows, manipulating models in a plethora of different formalisms, each with their own
methods, techniques and tools. Storing knowledge on these workflows can reduce considerable
effort during system development not only to allow their repeatability and replicability, but also to
access and reason on data generated by their execution. In this work, we propose a framework to
manage modelling artefacts generated from workflow executions. The basic workflow concepts,
related formalisms and artefacts are formally defined in an ontology specified in OML (Ontology
Modelling Language). This ontology enables the construction of a knowledge graph that contains
system engineering data where we can apply reasoning on it. We also developed several tools to
support the system engineering during the design of workflows, their enactment, artefact storage
considering versioning, querying and reasoning on the stored data. These tools also hide the
complexity of manipulating the knowledge graph directly. Finally, we have applied our proposed
framework in a real-world system development scenario of a drivetrain smart sensor system.
Results show that our proposal not only helped the system engineering in fundamental difficulties
like storage and versioning but also could reduce time to access relevant information, and new
knowledge that can be inferred from the knowledge graph.

1. Introduction
Systems engineering processes are becoming increasingly complex. Managing knowledge about processes is
crucial for companies. A relevant factor is the heterogeneity of current system development processes. A diverse
spectrum of formalisms, languages, and tools are used for different purposes to build better systems. INCOSE’s
Systems Engineering Vision for 2035 details that knowledge should be considered as a critical asset [1]. One of the
mentioned challenges in current systems engineering practices is the huge fragmentation of tools and domain-specific
approaches to enable collaboration and analysis. The shift from document-centric processes to model-based strategies
brings opportunities, but also comes with challenges on dealing with all digital information. Therefore, knowledge
management is crucial to reach their vision.
A simple example that several companies face is to keep knowledge of standard procedures or successful cases.
Facilitating the repeatability and replicability of workflows is critical to achieving companies’ goals. They may lose
valuable information due to circumstances like employees changing roles, turnover, and so on. Thus, having a well-
formed way to keep this knowledge may reduce costs and effort. Moreover, several models and artefacts are generated
during the system development process. Most of the time, several versions of the same documents need to be stored and
made available for future reference. These artefacts are generated or required by activities, which in turn, are part of
a workflow executed by engineers. Nevertheless, they must be consistent, traceable, stored, and available for different
purposes like analysis, testing, simulation, or just as a reference for future activities.
Experiments play an essential role in systems engineering because they may provide evidence that the system
works according to expectations. However, the inherent complexity of executing these procedures brings challenges.
For instance, discovering what activities and in which order they should be executed for a given procedure, what
artefacts are created during their execution, which choice led to which result (and vice versa), what was the latest
("most correct”) version of a model, and how to find a particular artefact required by a given activity, are some of
several possible obstacles that may hinder the development of systems. Therefore, explicitly modelling workflows and
their related artefacts, together with a knowledge management mechanism, may optimise the design process.
ORCID (s):

First Author et al.: Preprint submitted to Elsevier Page 1 of 31

 Model management for system engineering using ontology and knowledge graphs

Product Lifecycle Management (PLM) software records artefact versions and simple relations between them,
possibly linked to the explicit enactment of workflows. However, PLM software usually focuses on CAD artefacts, and
has limited support for other file formats. Further, the possible links between data are usually ad-hoc and hard-coded.
Relations are not captured in a “meta-model”, and exist only implicitly as “whatever the code does”. As a result, it is
difficult to reason about complex relations between objects that span multiple links. The only way to navigate them, is
to write more code, repeating the same patterns. This ultimately makes these systems hard to maintain.
In [2], the authors propose a multi-paradigm modelling framework for model-based systems engineering. Accord-
ing to it, formalisms and transformation between them should be specified in a Formalism Transformation Graph
(FTG). Workflows are specified in terms of Process Models, and the result of executing them is recorded in a Process
Trace, which details events related to workflow activities and artefacts. Although the elements are clearly described,
they are not formalised. There is no implementation of the semantics of these workflows, nor concrete mechanisation
of the approach. Possible query descriptions are listed, but no concrete implementation is proposed.
Recently, the industry has shown an increasing interest in Knowledge Graphs (KGs), which are structures to
represent data and their relationships for a particular domain. This data can be spread to different sources. Usually,
KGs data conform to concepts defined in Ontologies. They provide a means to define properties on elements of a
particular domain.
We take advantage of these concepts to leverage the ideas presented in [2] using an ontology to serve as a formal
definition of systems engineering elements. This ontology also allows the construction of a knowledge graph to relate
systems engineering process data, which in turn enables several capabilities like versioning, querying and reasoning on
this data. Further, the system engineer does not need to worry about the underlying notations and data structure to use
these services because we have implemented several tools to support the design and enactment of workflow models,
and also to access the information from models, enactments, and artefacts. Finally, we have applied our framework in
a real-world context during the development of a drivetrain smart sensor system, where the engineer used our tools to
execute experiments during the system development process. Besides the results while executing these experiments,
our tools and methodology also allow these experiments to be repeated and replicated in the future.
In summary our major contributions are the following:

• An extensible ontology for model management of arbitrary artefacts, their internal structure, links between them,
and versioning information.
• Based on this ontology, an ontology for workflow models, and for traces produced by the enactment thereof.
These traces link to the artefacts produced and consumed by the enactment.
• A model repository, based on our ontology, that allows querying, reasoning, and services to be executed on the
stored data.
• A modelling environment for the creation of workflows in our custom workflow language.

• An engine the enactment of these workflows, linked to our repository.

The remainder of the paper is structured as follows. Section 2 presents some background on model management,
ontologies and knowledge graphs. Section 3 describes our framework using a spring-mass-damper system as the
running example. Section 4 explains the architecture and implementation of our framework, and the several tools
developed to support model management for systems engineering workflows. Section 5 shows how we applied our
framework in a real-world case for a drivetrain digital twin system. Finally, Section 6 discusses related work and
Section 7 gives a summary and presents the final remarks of our work.

2. Background
In this section, we present topics that are relevant to our proposed framework. First, we discuss in Section 2.1 the
challenges of model management. Section 2.2 describes what knowledge graphs are, their relationships with ontologies,
and some standards and languages used to define them.

First Author et al.: Preprint submitted to Elsevier Page 2 of 31

 Model management for system engineering using ontology and knowledge graphs

2.1. Model Management

Given the shift from document-centric to model-centric development of systems, the governance over the models is
becoming a concern of extreme relevance. Therefore, the availability of tool support to tackle this complexity is crucial.
A Model Management System (MMS) comes to provide a means to create, store, manipulate and access models [3].
This term emerged in the early 1980s as an extension to decision support systems to facilitate the use of analytical tools
in providing solutions [4]. However, such systems can be used for diverse purposes like model elaboration, storage,
testing, validation, execution, maintenance, integrity, interoperability, abstraction, extraction, and provisions for model
security and model sharing.
Baldwin et al. [5] consider three modelling task dimensions: model formulation, model representation and model
processing. The first regards understanding the problem and eliciting possible models to help solve it. As our focus
is the replicability of experiments, we need an understandable way to represent the steps performed during these
experiments together with elements that make artefacts and activities explicit. Therefore, an intuitive language to
represent workflows covering both data (artefacts) and control flow is crucial to our goals.
Model representation refers to the structures, both data and architectural properties, used to represent knowledge
relevant to the models. For instance, model parameters, model structure (e.g., relationships between elements),
relationships to other models and data. Bharadwaj et al. [6] classify model representation into three categories:
database approach, graph-based approach, and knowledge-based approach. In the database approach, the models are
organised according to a data model hiding from the users the physical details of the model base. Some examples of
this approach are the Entity-Relationship (ER) framework [7] and the Relational model framework [8]. The graph-
based approach represents knowledge in terms of one or more graphs or digraphs. According to Bharadwaj et al. [6],
this approach has several advantages, including conceptual clarity, ease of programming and manipulation, and more
effective communication between analysts and decision-makers. Some instances of this method include the structured
modelling framework [9], graph grammars [10, 11], and logic graphs [12]. Finally, the knowledge-based approach
applies Artificial Intelligence (AI) techniques to model management like semantic networks, first-order predicate
calculus and production rules. Some advantages of this approach are automatic model formulation, answering user
queries with the support of inference methods, and the information about modelling being made explicit and accessible
instead of embedding it within an algorithm.
In a semantic network, objects and relationships between them are represented as nodes and arcs in a directed
graph. Inference over the knowledge base can be performed by traversing the edges of the graph. It is a convenient
mechanism to express concepts, facts and their relationships. One of the main representatives of this approach is the
semantic web [13], which defines a set of languages and standards to allow the processing of information from the
Web by computers (as opposed to most of the current Web, which is mostly targeted at human consumption) by using
machine-understandable formats [14]. This is to be achieved by the usage of metadata incorporated into the web data
to allow reasoning over it. We discuss some of these standards in Section 2.2.
The last modelling dimension is model processing, which discusses the computation of the designed models to
deliver an adequate solution. Clearly, this capability is highly dependent on the model formulation and representation
because the mechanisms can be triggered for processing according to how the model is created and the used notations.
The processing can be related to different purposes, for example, model consistency, property analysis, reasoning
capabilities, simulation, and optimisation methods. From the user perspective, it must be an interactive process where
the MMS explains its processing results in a user-friendly language. From the system perspective, the goal is to generate
and evaluate a solution to the given model utilising the most appropriate techniques.
Model management has become more significant nowadays, given the investment in Model-based strategies.
Nevertheless, the concern from the 1980s still remains but on a much larger scale. Large amounts of models are being
generated and must be managed by MMSs. New functions are emerging to minimise complexity during the system
design process. For instance, pattern identification can be applied to automatically propose models during the early
design, the usage of available inference tools to reason over model properties, and facilitate the traceability between
the different elements represented inside the models.

2.2. Ontologies and Knowledge Graphs

Recently, companies considerably increased their interest in semantic applications to use in their specific domains.
For Model Management, it has not been different because we need even more sophisticated mechanisms to tackle the
evolving complexity and scale of systems. The terms Ontology and Knowledge Graph have been surfacing in the last

First Author et al.: Preprint submitted to Elsevier Page 3 of 31

 Model management for system engineering using ontology and knowledge graphs

years as structures to support Model Representation and facilitate Model Processing (dimensions presented earlier in
Section 2.1).
When managing our models, we need sound mechanisms to identify and relate the model elements. Hence, an
instrument to classify the possible model elements and their relations is essential. On the other hand, we need a
concrete realisation of these elements and their relationships. Ontologies serve the former purpose as they define types
of existing things for distinguishing domains and properties to describe them, while Knowledge Graphs can be seen as
the materialisation of these elements and relationships. In the following sections, we detail these two topics.

2.2.1. Ontologies
The motivation for the usage of Ontologies came from the necessity of clarifying terms and relationships used
in different areas. It is not unusual to have different teams, sometimes in the same company, that have distinct
understandings of the same term or use different terms for the same entity. This can lead to lots of confusion
and communication overhead. When we think about digitisation for instance, in automating processes, a uniform
understanding is even more important to allow machine-readable structures that can be processed by a computer.
Ontologies support this necessity by determining an explicit knowledge base of concepts and their relationships.
For that, it uses several components like concepts, their relations, attributes, instances and axioms [15]. Concepts are
the types or classes of elements, e.g., "car" and "person". While "is driven by" could be considered a relation between
"car" and "person". A car can have an attribute "mass", which specifies the weight of the car. The instance mechanism
is a concrete entity of a particular concept, for instance, a car with a given model, car plate and all of its features
(attributes) specified. And axioms (also known as facts) provide properties that are assumed valid for concepts and
relations.
Besides the goal of making the information explicit and improving data interoperability, ontologies are also
employed to retrieve additional knowledge from the current base through the use of reasoners. This way, new
information can be deduced using logic inference. For instance, "A car X is driven by a person P" + "A car is driven
by at most one Driver" => "P is a Driver". Finally, it also allows checking the consistency of partial descriptions of a
product from different developers [16, 17].
During the 2000s, research and applications on ontologies considerably increased due to its pivotal role in the
definition of Semantic Web standards. The meaning of the term "Semantic Web" is quite controversial, and it lacks
consensus. Some see it as a field of research [14], while others regard it as a technological stack. According to W3C’s
vision, it is the Web of linked data, i.e., a set of data structures (graphs) connected through the Web according to
their identifiers and using a machine-readable format to allow its processing [13]. Although these graphs may refer to
different domains, they interoperate through the usage of standards. In this context, ontologies are a major enabler for
data integration, sharing, and discovery. Moreover, they have been envisioned to be reusable, i.e., the same ontology
supposedly can be applied in similar environments. Although the Semantic Web proposes this idea of collaboration
between the different networks of data, the technology stack can be used by institutions to internally define how their
data is organised and also how to process these data according to their needs. In addition, institutions can decide what
parts of their data will be publicly made available.

2.2.2. Knowledge Graphs

While ontologies provide a way to define concepts and relationships, a Knowledge Graph can be understood as
a materialisation of the network of elements that conforms to the ontology definitions. We consider real data about
individuals to build it. It is as a form of model representation that mixes graph-based and logic-based approaches,
as described in Section 2.1. The term emerged in 2012 when Google utilised it to describe the searches on notable
entities so that relevant information about them is displayed on the result search page. Figure 1 displays what is shown
by Google when we search by Tony Hoare. Users also could navigate the nodes of the graph that are connected, e.g.,
we could access Bill Roscoe result page because he is listed as one of Hoare’s notable students.
After Google, other major companies decided to invest in such a notion, such as Facebook, Microsoft, and IBM [18],
and the term has been established in the industry since. However, more than a novel approach, it took well-established
research concepts (some mentioned in Section 2.1) and articulated them on an industrial scale with different emphases.
Among the techniques employed in knowledge graphs, Federation and Virtualisation play relevant roles in boosting
this technology. The former provides capabilities for the connection between graphs from different architectures,
implementation and data sources. This way, it is possible to navigate these integrated graphs to return a search result.
The motivation for Virtual Knowledge Graphs is to maximise scalability of operations over the graph, given that all the

First Author et al.: Preprint submitted to Elsevier Page 4 of 31

 Model management for system engineering using ontology and knowledge graphs

Figure 1: Illustration of Google Knowledge Graph usage.

data does not need to be exposed to the end users. Just a conceptual representation of the domain of interest is presented
instead, resulting in a reduced graph. Nevertheless, it brings challenges in integrating the different data sources and
the KG, besides assuring the integrity and completeness of all the data [19].
The way Knowledge Graphs are structured may also vary. For instance, Google does not disclose the representation
of its KG. It provides only an API to access the contents1 . Another alternative is to use off-the-self solutions that
provide mechanisms to build your own knowledge graph [20, 21]. Although good support is provided by these tools,
we are somehow tied to their internal representations and encoded functionalities to manage the knowledge graph.
Finally, there is the Semantic Web approach standardised by W3C, where graphs are represented using RDF (Resource
Description Framework). It is an open and well-defined standard that has been used by a large community. The Linked
Open Data Cloud is an enormous open-access knowledge graph that can serve as a reference for structuring possible
data domains2 , and it follows the Semantic Web standards, including RDF. Also, it can be integrated with other
standardised languages (e.g., SPARQL for querying and OWL to define ontologies) to deliver features to the end
user.

2.2.3. RDF
The Resource Description Framework (RDF) is a W3C standard that specifies a syntax to create directed typed
graphs [22]. These graphs are structured according to what is called RDF triplestores, which are statements that relates
three elements, a subject, a predicate, and an object. Figure 2 illustrates the structure of an RDF triplestore for the
car example previously mentioned. The subject "Car X" is a node of the graph representing some resource, the object
"Person P" is also a node in the graph, and the predicate "is driven by" is an arc from the subject to the object to
express a relationship between them. All the elements in RDF must be uniquely identified by an IRI (Internationalized
Resource Identifier). The main difference between a subject and an object is that the latter can be either a reference to
a resource or a literal value.
Although simple, this structure is flexible and generic, with a high level of expressiveness, which is suitable
for building graph databases. It has several serialisation formats, e.g., based on XML or JSON. However, the
most commonly used one is the Turtle (Terse RDF Triple Language) format. The RDF/Turtle specification in
Listing 1 shows a simplified example of the relationship between "CarX" and "PersonP". Here we use prefixes to
improve readability. The subject "CarX" is of type <http://example. com/car/Car>, while "PersonX" is of type
<http://example.com/people/Person>. "CarX" has a model "ModelX", which is a string, and it is driven by
"PersonX".

Listing 1: Turtle Example.

1 https://cloud.google.com/enterprise-knowledge-graph/docs/search-api
2 https://lod-cloud.net/

First Author et al.: Preprint submitted to Elsevier Page 5 of 31

 Model management for system engineering using ontology and knowledge graphs

Figure 2: Representation of an RDF triplestore.

@prefix car: < http: // example . com / car # > .

@prefix people: < http: // example . com / people # > .
@prefix rdf: < http: // www . w3 . org /1999/02/22 - rdf - syntax - ns # > .

car:CarX rdf:type car:Car ;

car:hasModel " ModelX " ;
car:isDrivenBy people:PersonX .

people:PersonX rdf:type people:Person .

The main purpose of these specifications is to have a machine-readable notation to describe how different resources
are defined and interact with each other. In addition, it allows integration with different other standards (e.g., OWL and
XSD) or different ontologies referenced by the prefixes.

2.2.4. SPARQL
While RDF supports the specification of elements and their relationships in terms of a graph-like structure, using
text matching to search through this structure is not convenient. SPARQL is a SQL-like language that supports querying
RDF graphs [23]. Queries are defined according to triplestore patterns with the possibility of using a diverse set of
operators like filtering, unions, optional patterns, among others. SPARQL queries can be executed from an endpoint
that should be made available to execute queries over a graph. One of the main capabilities of SPARQL is that queries
can retrieve information from graphs on different data sources available in different endpoints.
Listing 2 displays a simple illustration of a SPARQL query related to our car example. It returns the car and the
person that drives the car. The elements prefixed with ’?’ indicate variables that should match possible values in the
search. The statement "?c a car:Car" matches a subject ("?c") that is typed by "car:Car" (the object). The term ’a’
between the subject and the object is just a syntactic sugar for the predicate "rdf:type".

Listing 2: SPARQL Example.

PREFIX car : < http :// example . com / car # >
SELECT ? c ? p WHERE {
? c a car : Car .
? c car : isDrivenBy ? p
}

2.2.5. OWL
The Web Ontology Language (OWL), or its latest revision OWL 2 [24], provides capabilities to define ontologies,
initially for Web resources, but now it can be applied in a broader context. It describes a domain in terms of classes
(types), properties (relations) and individuals (instances), and it supports the detailed description of features from these
objects. Its core is based on a description logic designed to support deductive reasoning [25].
The official concrete syntax for OWL is RDF/XML, but it is also common to use the RDF/Turtle. Compared to
RDF, OWL is much more expressive, given its larger vocabulary to express classes (types), properties (relations),
individuals (instances), and data values. For instance, we can state class hierarchies, equivalences, or disjoints. It is
also possible to specify properties for a relation, e.g., symmetry, transitivity, reflexivity, and restrict their cardinality.
Another considerable difference is the computability and reasoning capabilities of OWL. For instance, if we say that
Robert is of rdf:type Professor, and Professor is a subClassOf Person, then I can now infer that Robert is also of type

First Author et al.: Preprint submitted to Elsevier Page 6 of 31

 Model management for system engineering using ontology and knowledge graphs

(a) OML vocabulary example. (b) OML description example.

Figure 3: Example of OML specification.

Person. Moreover, we can check that constraints over properties are also satisfied. Thus, we can check the consistency
of our ontology. Finally, it is also possible to make your ontology modular by reusing definitions from other files using
the owl:import clause.
Although OWL expands RDF in some sense, it is still a notation created for machines. Some tools, like Protégé [26],
help in the usage of the language, but it still requires considerable effort to work directly with OWL using RDF/XML
or RDF/Turtle serialization formats.

2.2.6. OML
OML (Ontological Modelling Language) can be seen as a Domain-specific language (DSL) for OWL [27]. It
provides a more user-friendly syntax that is mapped to several patterns of an OWL2 subset. Therefore, it inherits OWL
expressivity, modularity, extensibility, and description logic semantics. It has been developed by the Jet Propulsion
Laboratory of NASA with the aim of helping engineers define systems engineering vocabularies and using them to
describe systems.
OML provides a textual language that allows the creation of vocabularies and descriptions. The former declares
concepts, properties and relations, that is, definitions at the type level. The latter contains individuals that conform to
the vocabularies, that is, elements at the instance level. Figure 3 shows an example of a vocabulary for a simple car
(3a) and a possible instance of this vocabulary (3b). A concept is the basic mechanism to describe an entity. An aspect
is a concept that cannot have a related instance. It is similar to the notion of an abstract class in object orientation. Both
aspects and concepts can have properties when we want to relate them to a literal. On the other hand, relations are
used to specify a relationship between two concepts. The same constraints that are supported in OWL can be used in
the relations. For instance, we declare the relation HasWheel from a Car to a Wheel to be asymmetric and irreflexive
in Figure 3a. In the description scenario illustrated in Figure 3b, we instantiate a car named car1 and its four wheels
with their respective masses and relations.
OML has an Eclipse-based environment to support the development of ontologies called Rosetta [28]. There, users
not only can design their ontologies but check their consistency using embedded OWL reasoners. Rosetta translates
OML to OWL in order to reason over the ontologies and provide some feedback on possible non-conforming elements.
Moreover, it is also possible to load the ontologies and related instances in an RDF-based knowledge graph through
its integration with Apache Jena and Apache Fuseki [29] to execute queries using SPARQL.
A particular difference between OWL and OML is that the former considers the open-world assumption, that
is, statements can be either true, false or unknown. The absence of information is considered unknown information,
so it is not possible to state that something missing is false [30]. This is particularly important in the context of the
Semantic Web because the data must exist but is not available in the ontology yet. However, for some applications (e.g.,

First Author et al.: Preprint submitted to Elsevier Page 7 of 31

 Model management for system engineering using ontology and knowledge graphs

integrity constraints, validation and data structures), closed-world semantics are more suitable. OML allows closed-
world semantics via the usage of its vocabulary bundle, which is an ontology that bundles a set of vocabularies [27].

3. Requirements for a Model Management Framework

In this section, we state the key requirements that a model management framework for system engineering must
provide in our view. We illustrate our requirements with a running example of the engineering of a mass-spring-damper
system, which we introduce first.

3.1. Running example

Figure 4: Simple representation of the system.

In order to illustrate the framework, we provide an example consisting of a mass-spring-damper system and all
related components3 as shown in Figure 4. A mass-spring-damper system is a commonly used model in physics and
engineering to describe the behaviour of systems that involve the interaction between a mass, a spring, and a damper.
It is often used to analyse and predict the motion and response of mechanical systems, such as vibrating structures
or suspension systems. Such systems are described in detail in chapters 3 and 4 of Dynamical Systems for Creative
Technology [31].
The system consists of three main components:

• Mass (𝑚): The mass represents the physical object or body in the system. It is typically a point mass, meaning
it is concentrated at a single location. The mass interacts with the other components and experiences forces that
affect its motion. It slides horizontally in our example and is depicted using a purple rectangle on green wheels.
• Spring (𝑘): The spring represents an elastic element in the system. It provides a restoring force proportional to
the displacement of the mass from its equilibrium position. This force is directed opposite to the displacement,
attempting to bring the mass back to its resting position. The stiffness of the spring is quantified by the spring
constant (k), which determines how strong the restoring force is for a given displacement.
• Damper (𝑏): The damper represents a dissipative or damping element in the system. It exerts a resistive force that
opposes the motion of the mass. The damping force is proportional to the velocity of the mass, with the constant
of proportionality being the damping coefficient (𝑏). The damping coefficient determines how quickly the system
dissipates energy and reduces oscillations. We don’t depict a damper explicitly. In our case, the damping force
is provided by friction.
When the system is at rest, the mass is in its equilibrium position, where the net force acting on it is zero. If the mass
is displaced from this position, the spring exerts a force proportional to the displacement, trying to restore the mass to
its equilibrium. Simultaneously, the damper opposes the motion by exerting a force proportional to the velocity.
3 /mass-spring-damper

First Author et al.: Preprint submitted to Elsevier Page 8 of 31

 Model management for system engineering using ontology and knowledge graphs

The behaviour of the system depends on the parameters mass, spring constant, and damping coefficient, as well
as the initial conditions. When subject to an external force or disturbance, the system exhibits various types of
motion, including oscillations, transient responses, or steady-state responses[32]. The dynamics of the system can
be described by a differential equation. For example, the equation of motion for the mass-spring-damper system is
typically represented by Newton’s second law:

𝑑2𝑥 𝑑𝑥
𝐹 (𝑡) = 𝑚 ∗ +𝑏∗ +𝑘∗𝑥
𝑑𝑡2 𝑑𝑡
Here, 𝑥 is the displacement of the mass from its equilibrium position, 𝑡 is time, and 𝐹 (𝑡) represents any applied
2
external forces. The terms involving the derivatives 𝑑𝑥 𝑑𝑡
and 𝑑𝑑𝑡2𝑥 account for the velocity and acceleration of the mass,
respectively.
By solving these equations, engineers and scientists can analyse the response of the system, predict its behaviour
under different conditions, and design appropriate control strategies to achieve desired outcomes, such as minimizing
vibrations or optimizing performance.
For our example, we want to find a spring constant 𝑘 such that an initially excited system is damped enough for the
oscillation to have an amplitude lower than 0.05 between 5 and 7 seconds after the experiments starts. We can model
this system as a Causal Block Diagram model, as seen in Figure 5. Each of the integrator blocks represents part of the
equation 𝑦(𝑛) = 𝑓 − 𝑎1 𝑦𝑛−1 − ... − 𝑎𝑛−1 𝑦′ − 𝑎𝑛 𝑦 [33]. By repeating the simulation of this model with different parameter
values for 𝑘, a solution can be found.

1 OUT1 IN1
⅟ OUT1
0 OUT1

IC

∫
IC OUT1
IN1
∫
OUT1

100 OUT1 INx

∑ OUT1
INx
∏ OUT1 IN1

2 OUT1 IN1
- OUT1
INx
∏ OUT1

100 OUT1 IN1

- OUT1
INx
∏ OUT1

Figure 5: Causal Block Diagram model of the mass-spring-damper system.

A plot of an execution trace is shown in Figure 6. We want to find a spring constant for which the oscillations are
sufficiently damped after 5 seconds. The blue rectangle in the figure illustrates this area of interest.

3.2. Workflow modelling and enactment

In the previous section, we have introduced a simple engineering problem. We now look at an example of an
explicitly modelled workflow intended to solve the “mass-spring-damper” engineering problem. The workflow model
is shown in Figure 7, on the left. It is a model in the FTG+PM language [2], a very simple workflow language that
follows Moody’s “Physics of Notation” [34]. While explaining the workflow model, we give a very rough overview of
the FTG+PM language. For a complete description, we point the reader to [2].
At the highest level, the model consists of two activities: CreateRequirements and CreateModelAndEstimatePa-
rameters. Activities connect to each other via control flow (thick blue arrows) and data flow (thin green arrows).
Control flow specifies the order in which activities can execute. Data flow specifies the artefacts (and their types) that
are expected to be produced and consumed by the activities.
Our engineering workflow starts with writing down the requirements of the system. This is modeled by the first
activity CreateRequirements, which has no data input ports (it consumes no data), and one data output port. When it
ends, it must produce via this port an artefact named ‘requirements’ of type ‘xopp’ (Xournal++4 , a sketching app — an
4 https://xournalpp.github.io/

First Author et al.: Preprint submitted to Elsevier Page 9 of 31

 Model management for system engineering using ontology and knowledge graphs

Figure 6: System trace of a single mass-spring-damper causal block diagram simulation.

example of a requirements document for the mass-spring-damper problem is shown in Figure 8), and passes control to
the next activity, CreateModelAndEstimateParameters, which consumes the requirements artefact. When this activity
ends, it produces a Causal Block Diagram (CBD) model with the right parameters (example for mass-spring-damper:
Figure 5), that describes the behaviour of a system that is valid with respect to the requirements. The workflow ends
here.
To make it more interesting, we can refine the last activity further with another (nested) workflow model, shown
on the right of Figure 7. This model consists of three activities:
1. The activity DefineCBDModelAndParams, consumes the requirements, all previous versions of the CBD model
(if there are any, will be explained later), and produces a new version of the CBD model.
2. The activity RunCBDSimulation runs the new version of the CBD model, produced by the previous activity, for
a duration of 7 seconds. As output, it produces a table (e.g., a CSV file) containing a trace of the values of the
simulation’s variables (for mass-spring-damper, these will be 𝑥, 𝑣 and 𝑎) at every timestep.
3. The activity ValidateCBDSimulationResults checks if the simulation trace that was produced by the previous
activity is consistent with the requirements. The two possible outcomes (valid or invalid) are modelled in our
workflow model as two respective control flow output ports. If the model’s behaviour was valid, the workflow
ends. Otherwise, we perform another iteration of creating a new version of the CBD model, simulating it, and
validating it against the requirements. The requirements are left unaltered.
Even this simple workflow comes with artefacts of heterogeneous file formats, and will result in multiple versions
of these artefacts (because of the iteration cycle). Obviously, more complex engineering problems will lead to much
more complex workflows, with multiple domain experts working in parallel, performing real-world experiments and
simulations, involving even more different file formats and artefact versions. Often, these workflows exist only in the
minds of engineers, and they are informally communicated. We instead propose to model them explicitly, like we have
done here.
Workflow models in the FTG+PM language are highly reusable and hierarchically composable. For instance, the
activity CreateModelAndEstimateParams is reusable in any context where a CBD model needs to be created and its
parameters estimated in an iterative fashion according to a set of requirements. The enactment of a workflow model
produces an event trace, consisting of events like “activity started” and “activity ended”. These events can be linked
to the artefact versions consumed/produced respectively.

First Author et al.: Preprint submitted to Elsevier Page 10 of 31

 Model management for system engineering using ontology and knowledge graphs
cin

Legend
cin
Start of
Workflow
cin
reqs reqs
End of
prev_cbd : DefineCBDModelAndParams Workflow
reqs : CreateRequirements
cbd Control Flow

requirements : cout modelAndParams : cout Data Flow:

xopp cin cbd CBD cin "Consumes"/
"Produces"

reqs cbd name Activity

: RunCBDSimulation : Type
: CreateModel
AndEstimateParameters sim_trace
name :
cbd Artefact
simulationTrace Type
cout
: TabularData cin
cout
Control
modelAndParameters Input/Output
: CBD Port
invalid
sim_trace
: ValidateCBDSimulationResults Data
Input/Output
reqs Port

valid
: CreateModel
AndEstimateParameters

cout

Figure 7: Running Example: Mass-spring-damper Workflow Model.

3.3. Model Repository

In order to persist event traces and artefacts produced during the enactment of workflow models like from our
previous example, we need a model repository. We now motivate a number of requirements for this repository to
satisfy: versioning, traceability, typing/introspection and finally querying.

3.3.1. Versioning
A lot of engineering work is done in iterative, incremental cycles. When a workflow model contains cycles,
multiple versions of artefacts are produced. For instance, in our running example (Figure 7), each time the activity
DefineCBDModelAndParams ends, a new version of a CBD model is produced. Not only do we want to link the
new CBD model to the “end activity” event, we also want to link it to the previous version of the CBD model. This
versioning information needs to be recorded, and be queryable.

3.3.2. (Fine-Grained) Traceability

Our overall goal is to increase traceability in data produced during systems engineering, and be able to query
this data in a powerful manner. Traceability is a broad concept, that takes many forms [2]. One form of traceability is
linking artefact versions to the activities that produced them during workflow enactment. Other kinds of traceability are
links between artefacts and their meta-models (which are also just artefacts). These links can also exist at a finer level,
between the elements of artefacts and the elements of meta-models. Another example of fine-grained traceability are
the correspondence links between the concrete and abstract syntax of a model (example will be given in Section ??).
For yet another example of fine-grained traceability, consider the CBD model in Figure 5. Every connection (line)
between the different blocks represents a signal that changes with every timestep during simulation. The activity
RunCBDSimulation in Figure 7 performs this simulation, and produces a table with the signal values at every timestep.
Every column in this table can be linked (“caused by”) to a signal in the CBD model. Further, we could link the
CBD’s constant blocks representing the parameters 𝑚, 𝑏 and 𝑘 to the text labels “R1: ...”, “R2: ...” and “R3: ...” of the
requirements in the Xournal++ document (Figure 8).

First Author et al.: Preprint submitted to Elsevier Page 11 of 31

 Model management for system engineering using ontology and knowledge graphs

Mass-spring-damper System Requirements

m
k

Requirements

R1: The mass "m" of the block must be 0.01 kg.

R2: The friction coefficient "b" of the damper must be 0.02 Ns/m.

R3: The spring constant "k" must be such that almost zero oscillation
(equilibrium) occurs in the time frame of [5, 7] after impulse release.

Figure 8: Running Example: mass-spring-damper Requirements (as Xournal++ document).

In order to persist all these different kinds of traceability links, between arbitrary artefact elements, we want to treat
all data in our repository in a unified manner, as one large graph.

3.3.3. Typing/Introspection
By knowing the types of artefacts and their elements (called introspection in programming languages), interactive
exploration and query-building of our model repository becomes greatly enhanced. The system can answer questions
like “What is the structure of objects of type X?”, and “What operations can be performed on it?”, and “Give me all
objects of type X that implement requirement Y”.
For instance, given a CBD model, the system knows that it can be simulated, producing a table of signal values
(via the activity RunCBDSimulation). Hence, for a given CBD model, the system can automatically suggest a query
for the user to retrieve all simulation traces produced by that CBD model.
To support introspection, we explicitly store the types of all artefacts and their elements. They are just one kind of
traceability link.

3.3.4. Consistency checking

Many kinds of inconsistencies can occur in model-based systems engineering. In the general case, we want to be
able to express arbitrary consistency rules over arbitrary model elements. One special case of consistency checking
is conformance checking, where it is verified whether an artefact really conforms to the constraints (this includes
multiplicities) imposed by its type (defined in e.g., a meta-model).
As an example of conformance checking, the meta-model of CBD could have a rule saying that for every block,
all of its input ports must be connected to some output port. This makes sense, because otherwise no value could be
computed for this block. Suppose the user forgot to connect the IC port of one of the integrator blocks. When parsing
the CBD model, it should be detected that the model is invalid.

3.3.5. Query language

A final requirement for our model repository is to support a query language. With a query language, relatively
complex data retrieval operations can be expressed in a relatively simple manner. Query languages can be powerful:
for instance, existing queries may be composed into new queries. Query languages are usually easier to understand for
non-technical users than having to use an API and write code.

First Author et al.: Preprint submitted to Elsevier Page 12 of 31

 Model management for system engineering using ontology and knowledge graphs

Queries can be broadly categorised into 3 categories:

1. Queries aiding in traceability
• Queries for horizontal traceability
• Queries for vertical traceability
• Queries for fine-grained traceability
• Queries for versioning
2. Queries to execute services
3. Queries to explore the structure

Horizontal traceability Horizontal traceability refers to traversing the FTG+PM++ from left to right, or vice versa.
Each artefact belongs to a type, and all parts of the workflow are typed. For example, we know the requirements artefact
is typed by a Xournal++ file formalism xopp.

Vertical traceability Vertical traceability allows for following the workflow. We can clearly see how the process
progresses when looking at the spring-mass-damper process model in Figure 7. By following the blue control flow
facilities, we relate each of the activities and achieve full vertical traceability within the process model. Here, we know
that the RunCBDSimulation activity is always preceded by the DefineCBDModelAndParams activity and succeeded
by the ValidateCBDSimulationResults activity.

Fine-grained traceability Whenever a relation refers not directly to an artefact, but the inner structure of one, we
consider it a fine-grained relation. Any query using these relations is therefore a fine-grained traceability query. An
example here could be the relation between a column of the simulationTrace TabularData artefact instance and a
specific block in the corresponding modelAndParams CBD artefact instance. This would give us a better understanding
as to which blocks are responsible for generating which data.

Versioning Enactment of a workflow containing a circular dependency will create multiple versions of an artefact
for an activity. These artefacts are related to each other, so earlier versions can be returned in a query. If we enacted
multiple iterations of the workflow, we would end up with multiple simulationTrace artefact instances. Each of these
artefacts would have a direct relation to its precursors from previous workflow iterations.

Services By using special keywords, we can call external services which perform part of the query. These service
calls could include looking for outliers in the simulationTrace or even running, validating, or generating traces of the
CBD.

Exploration Some queries are performed to get the structure of data. This includes knowing which types are available
and what their properties are. We are able to check which formalisms and/or transformations are available to us. For
the spring-mass-damper running example, these would include, but are not limited to, exploring the structure of CBD,
xopp, TabularData and ProcessModel5 artefacts.

Composing categories Simple queries can be composed into (more) complex queries. For instance, we can construct
a query on the spring-mass-damper system to figure out the initial model (first version) whose later versions led to the
correct parameter estimation and its value.

4. Implementation
In this section, we explain our prototype implementation, and explain how it satisfies the requirements described
in Section 3.
5 process models are also just artefacts

First Author et al.: Preprint submitted to Elsevier Page 13 of 31

 Model management for system engineering using ontology and knowledge graphs

4.1. Architecture Overview

To implement our model repository, we decided to use an OML/OWL/RDF stack. By using RDF, we have a
unified graph-like data structure for all our models and the traceability links between them (solving 3.3.2), that can
be queried by the existing SPARQL language (solving 3.3.5). By using OWL on top of RDF, we can define simple
consistency rules (solving 3.3.4) and typing information for introspection (solving 3.3.3). We do not use OWL directly,
but instead generate it from OML documents, for reasons explained in Section 2.2.6. OML distinguishes between type-
level (“vocabularies”) and instance-level (“descriptions”) documents. The descriptions are generated from artefacts in
different file formats via the use of adaptors. Their respective vocabularies were written by hand.
OML has no built-in solution for versioning. It foresees using a text-based versioning system like Git for versioning
the OML documents themselves. The main problem with this approach is that the versioning information remains
external to OML, and therefore also to the generated OWL/RDF. It would be impossible to use this information in
(SPARQL) queries. Therefore, we use the “poor man’s approach” of simply adding for every new artefact version a
new OML description document, and OWL/RDF data generated from it, to our model repository (solving 3.3.1).

4.1.1. Components
Our implementation consists of the following components (Figure 9):

OML Vocabularies are a set of hand-written “meta-models” that all the instance-level data (artefacts and workflow
event traces) must conform to.
Adaptor Service transforms artefacts from different file formats to RDF (by first transforming to OML).
Fuseki is an RDF database and SPARQL query engine. It stores all the RDF data.
Workflow Enactment Engine is a web service that interprets workflow models in our FTG+PM language (Sec-
tion 4.2), and saves execution traces directly in Fuseki.
Draw.io Frontend is a plugin for the diagramming tool draw.io that allows the user to navigate traceability links (of
the model repository) from/to diagrams for which we have parsers.
Federation Services allow large datasets to be queryable via SPARQL without having to transform their data to RDF
in advance (this would not scale).
Graph Explorer is an application for exploring the knowledge graph with the use of SPARQL queries.

We will now explain these components in more detail.

4.2. OML Vocabularies

We have defined an ontology using OML (described in Section 2.2.6) that supports the design of diagrams in a
context based on two generic vocabularies for objects and links. Objects represent nodes used in diagrams, and links are
the edges connecting two objects. According to these two vocabularies, support for other languages can be provided.
Figure 9 how these two OML vocabularies are used to create language-specific vocabularies. The FTG+PM abstract
syntax can be described using the metamodel_object_diagrams.oml vocabulary, as well as their concrete syntax in
Draw.io. We use this same mechanism to describe the metamodels for formalisms used in our framework. For instance,
causal block diagrams, FTG, process models, process traces, textual documents, tabular data and Xournal++ (xopp)
are examples of formalisms currently supported. For those formalisms we do not have a metamodel yet, we have created
a generic formalism file, which can represent any type of file to be managed by our framework. However, fine-grained
traceability will not be supported for these formalisms as we do not have knowledge about their internal structure.
Especially formalism transformation graphs, process models and process traces are vital to support the description
of experiment workflows in our framework. We show in Figure 10 a graphical illustration of part of the OML
vocabularies created to define process models and traces that were presented in Section .
A pm:Model represents a workflow, and it is composed of pm:Activity, pm:Artefact and pm:ControlNode elements.
An pm:Activity describes a task to be executed, and it conforms to one ftg:Transformation, which is part of the
FTG vocabulary and represents a generic description of an action that can receive inputs and provide outputs. A
pm:Activity has ports, which can be control or data ports. Ports can also be input or output to an activity. For instance,
a pm:CtrlInputPort is an entry point to execute a pm:Activity, while a pm:CtrlOutputPort indicates the control flow

First Author et al.: Preprint submitted to Elsevier Page 14 of 31

 Model management for system engineering using ontology and knowledge graphs

OML "Vocabularies" Legend

metamodel_correspondence_links.oml data flow

conformance
metamodel_object_diagrams.oml dependency

inheritance
Federation Service
metamodel_
drawio.oml
some large dataset
original data
metamodel_ metamodel_ (user-created)
some_DSL.oml xopp.oml
generated data

Adaptor Service service

subquery results
model_abstract_syntax.oml transformation
DSL
parser
(e.g., FTG,
PM, CBD, ...) model_correspondence_links.oml

Fuseki SPARQL Server

model_concrete_syntax.oml OML files OWL/RDF
draw.io OML builder process trace
model dataset
parser dataset
model.drawio
Graph query results
update
Explorer
requests

sketch_concrete_syntax.oml diagram Draw.io

Xournal++ frontend
parser
any file
sketch.xopp Workflow Enactment
Engine

Figure 9: Architecture.

to be followed when a pm:Activity terminates. While control ports indicate possible flows of execution, data ports
indicate the production or consumption of artefacts by activities. A pm:Artefact indicates a resource utilised during the
workflow, and it is typed by a ftg:Formalism, which represents a class of elements that conform to the same metamodel
in the FTG vocabulary. A pm:ControlNode represent an element used to coordinate the control flow of the diagram.
For instance, pm:Initial indicates the point where the workflow starts while pm:Final determines its termination. A
pm:Fork can create concurrent flows while pm:Join synchronises several input concurrent flows in a single output
control flow.
Elements of process traces can be a pt:Event or a pt:Artefact. A pt:StartTrace event represents the beginning of
executing a trace described by a pm:Model, and pt:EndTrace represents the end of the trace. A pt:StartActivity indicates
the moment an activity starts by a control flow received in its pm:CtrlInputPort, while the pt:EndActivity means the
termination of a particular activity that will continue the flow by its pm:CtrlOutuputPort. Finally, a pt: Artefact is a
concrete version of a resource detailed by pm:Artefact, and that is generated/consumed during a pm:Model enactment.
Given that the workflow can have loops, the same artefact can be generated but with different versions (Versionable).
Although Figure 10 offers an intuition about the entities and relationships we have defined, it does not provide the
complete details and properties specified in our ontology. For example, to determine that a pt:StartTrace event has
not any other event before, we specify the property restricts relation isPrecededBy to max 0 in the context
of a pt:StartTrace. In this case, describing that the relation isPrecededBy does not happen. The opposite happens for
pt:EndTrace where we use the property restricts relation isFollowedBy to max 0. The complete OML ontology that
supports our framework is available in our Git repository6 . There, the vocabularies for the FTG, PM, and PT will be
available, but also the ones for the previously mentioned supported formalisms.
6 /DTDesign/src/main/examples/oml

First Author et al.: Preprint submitted to Elsevier Page 15 of 31

 Model management for system engineering using ontology and knowledge graphs

Figure 10: Part of the OML Vocabularies for the FTG+PM language.

4.3. Adaptor Service

The Adaptor Service is a REST endpoint where other services (most importantly the Workflow Execution Service)
can upload arbitrary files. Uploaded files are parsed and their internal structure converted to OML, which in turn is
converted to an OWL ontology, and added to Fuseki’s dataset. This is the main mechanism of adding new artefact
versions to our model repository. The originally uploaded files are also stored (in the adaptor service).
For every filetype to be converted to OML, we need a dedicated parser and OML generator. For some file types (e.g.,
Xournal++, a sketching application), we only convert the file’s contents to OML. In this case, the OML representation
contains exactly the same amount of information as the original file, and in theory, we could restore the original file
from the OML data (not implemented). For other file types (e.g., draw.io), the contents of the file (e.g., a diagram) may
be additionally interpreted as an instance of a DSL. For instance, a draw.io diagram representing a workflow model
can be seen as just a diagram consisting of shapes and connections (concrete syntax), but also as a process model, with
activities, control and data flow (abstract syntax). In this case, we not only convert the original file contents to OML,
but we additionally invoke a second DSL-specific parser, that attempts to construct an instance of the DSL’s abstract
syntax, in OML. Finally, we also store traceability links between elements of the generated OML files for concrete
and abstract syntax. These traceability links are stored in yet another OML file. All generated OML files are so-called
descriptions (i.e., OML’s instance-level), that conform to (manually written) vocabularies (i.e., OML’s type-level).
All of our parsers were written in Python, and OML generation makes use of the Jinja2 template library for Python.
The OML to OWL conversion is taken care of by a Gradle build script included with the Rosetta IDE for OML. During
OML to OWL conversion, the generated OWL output is checked for inconsistencies. Inconsistencies usually indicate
non-conformance between a description and its vocabulary. When an inconsistency occurs, the procedure fails. The
source code of the OML Generator Service is available in our Git repository7 .
7 /drawio2oml/src/master/backend

First Author et al.: Preprint submitted to Elsevier Page 16 of 31

 Model management for system engineering using ontology and knowledge graphs

Figure 11: WEE Architecture.

4.4. WEE - Workflow Enactment Engine

Once we have an environment to design workflows, it should be possible to execute them and record the traces
and their respective data. Manually designing the traces while executing the workflows would require too much effort.
Moreover, the generated process traces should be available to other tools for visualisation and information recovery.
Therefore, WEE has been developed to support both matters. First, it provides an environment to guide users through
the possible flows of execution related to previously designed process models, recording not only the tasks being
performed but also artefacts generated and used during the workflow enactments. WEE’s second purpose is to provide
access to the recorded knowledge on process traces and process models as a service. The latter is currently used by the
Drawio plugin to render information about the traces. However, any other tool that could benefit from these data could
use this service.
WEE has been conceived as a Web application and can be accessed via browser or through the REST API services.
Figure 11 shows WEE architecture, which is detailed in three layers. The Web Layer is responsible for the interaction
between users and other systems with the tool. It has a Web front-end that allows users to interactively enact selected
process models. This layer communicates with the Service Layer to invoke the appropriate services. The Service Layer
owns the business logic used to manipulate process models and traces. It also can request services from the Repository
Layer, which encodes the statements to query and update the knowledge graph stored in the Fuseki Server with data
from all process models and traces.
Using the Web Layer, the user can interact with the tool to enact a previously designed process model. First, the
user must select a PM to enact from a list returned from the knowledge graph. Then, the enactment process starts and
the tool displays the activities available based on the possible flows of the PM. For instance, considering our running
example of the Mass-Spring-Damper PM shown in Figure 7, the first activity to be executed is define_damper_req.
After concluding this task, the user must indicate its termination and upload the artefact springDamperRequirements
of type xopp, in conformance to what is described in the process model. Figure 12 displays the screen for this task. It
shows the process being enacted, the current activity, and the artefact that is required. After uploading the requested
file, the user can click the End Activity button. After starting or ending an activity, we update the current process trace
in the knowledge graph with the related events and artefacts provided by the user. The current events performed so far
are outlined on the right-hand side of the screen, as can be seen in Figure 12. The files are also stored using a local
storage service we have developed. Some types of artefacts are described in our ontology in OML. It can be seen as
meta-model descriptions of these types. For instance, if a CBD model is uploaded, we convert it to OML, and we
perform consistency checks before uploading their RDF representation to the knowledge graph in the Fuseki server (as

First Author et al.: Preprint submitted to Elsevier Page 17 of 31

 Model management for system engineering using ontology and knowledge graphs

Figure 12: Example of activity conclusion in WEE.

Figure 13: Example of starting an activity in WEE.

detailed in Section 4.3). This allows executing queries that not only link different CBDs, but also fine-grained queries
that relate a CBD to information from different artefacts, as discussed in Section 3.3.5.
Artefacts that have been uploaded by activities are also made available for other activities that require them.
Figure 13 shows the screen to start the following activity createCBD. As it needs the springDamperRequirements
artefact previously uploaded, we made it available for the user to download here. Although in this case the uploaded
artefact was required by the next activity, it may not always be the case. Artefacts may be available in different phases
of the workflow. However, the user does not need to worry about gathering artefacts to start an activity because WEE
makes them available in the expected task, which is quite convenient.
The process model enactment follows until the user reaches a final node, and the trace terminates. Nevertheless,
it is possible to stop an enactment at any time and resume it later at the point the user paused it. For instance, the
user may close the WEE window during an enactment, and on the next day, after selecting the PM he was enacting,
WEE shows a list of non-concluded enactments. Then, the user can choose the one he wants to continue enacting until
its termination. Once the traces are stored in the knowledge graph, it is possible to query them to search for specific
information, for example, the latest version of an artefact, or discover the activity that generated a particular artefact, but
also reason on the data from the enactment or from artefacts. This can be executed using Graph Explorer (Section 4.8)
or executing queries directly on Fuseki. However, for the latter, it is required knowledge not only of SPARQL but also

First Author et al.: Preprint submitted to Elsevier Page 18 of 31

 Model management for system engineering using ontology and knowledge graphs

of how our ontology is structured. Nevertheless, WEE also provides a REST API to read data from process models
and traces. Via HTTP calls it is possible to obtain all the traces of a particular process model, only finished traces, only
unfinished traces, and also get the nodes of a process model. The returned result can be JSON or XML descriptions of
the elements. Such a mechanism is useful for tool integration. For instance, it is used by the Draw.io plugin to render
visualisations of the process traces generated by WEE. Other tools could also benefit from the same API. More details
about WEE can be seen in its git repository, including tutorials on how to use the tool8 .

4.5. Draw.io Frontend

Draw.io Frontend is a plugin, written for the open-source diagramming tool draw.io. We use draw.io for the
authoring of various visual DSLs, and the plugin queries Fuseki in the background to find out whether a diagram
is an instance of such a DSL. When this is the case, it uses traceability links from concrete syntax (draw.io shapes)
to abstract syntax (DSL model elements), to inform the user about the types of elements, and possibly additional
inter-model traceability links. Exploration of these links happens via context menus. For instance, it is possibly to
navigate from a “StartActivity”-event in a process trace, to the control flow input port that was used to start the activity
(see screenshot in Figure 14). Its source code can be found in its Git repository9 .

Figure 14: Draw.io plugin extended context menu.

4.6. Fuseki
Fuseki is an RDF graph database and SPARQL query engine. In our implementation, Fuseki contains two separate
datasets: (1) the model dataset, which contains the data of all model versions, and traceability links between them, and
(2) the process trace dataset, which contains (possibly ongoing) event traces of enacted workflows, and all traceability
links between this data. The model dataset is generated via (again generated) OML data, whereas the process trace
dataset is inserted directly into Fuseki as OWL/RDF data, by WEE.
Other components of our architecture only read Fuseki’s data, via its SPARQL query engine. We provide example
SPARQL queries to alleviate the traversal of the FTG+PM. Prefixes are left out for the sake of brevity. Figure 15 is a
workflow which can be followed to get a general feel of the knowledge graph in question, and all its queries fall into
the structure exploration queries from Section 3.3.5.
Getting all the types is a good to place to start when encountering an unfamiliar knowledge graph. The types related
to the FTG+PM are of interest here. The query in Listing 3 will return all class types as a is a shortcut for rdf:type.

Listing 3: SPARQL query which returns all possible types.

SELECT DISTINCT ? type
WHERE {
? entity a ? type .
}

We proceed to delve deeper into one specific type once we have an overview of all available types in the Knowledge
Graph. Here, in Listing 4, we chose to find all available properties of the Formalism type. For our ontology, this returns
8 values:
1. http://www.w3.org/1999/02/22-rdf-syntax-ns#type
2. http://ua.be/sdo2l/vocabulary/base/base#hasGUID
3. http://ua.be/sdo2l/vocabulary/formalisms/pm#occursAsArtifact
8 /wee/wiki
9 /drawio/src/master/src/main/webapp/myPlugins/dtdesign.js

First Author et al.: Preprint submitted to Elsevier Page 19 of 31

 Model management for system engineering using ontology and knowledge graphs

kg
find_all_classes
: ClassQuery
query_result

all_classes :
ModelSet

all_classes

kg
knowledge_base : get_limitation_properties
KnowledgeGraph : PropertyQuery
query_result

limitation_properties :
ModelSet

properties properties classes

kg kg
find_which_use_description get_limitation_description
: HoldPropertyQuery : IndividualQuery

classes individuals

classes : individuals :
ModelSet ModelSet

Figure 15: Workflow model for extracting information from the Knowledge Graph using SPARQL queries.

4. http://ua.be/sdo2l/vocabulary/formalisms/cs_as#renderedAs
5. http://ua.be/sdo2l/vocabulary/formalisms/traceability_model#traceLinkFrom
6. http://www.w3.org/2002/07/owl#sameAs
7. http://ua.be/sdo2l/vocabulary/formalisms/ftg#isInputOf
8. http://ua.be/sdo2l/vocabulary/formalisms/ftg#isOutputOf

Listing 4: SPARQL query returning all properties of Listing 5: SPARQL query returning instance of a type
a certain type. and their name property.
SELECT DISTINCT ? property SELECT DISTINCT ? individual ? name
WHERE { WHERE {
? limitation a ftg : Formalism . ? individual a ftg : Formalism .
? limitation ? property ? value . ? individual ftg : isInputOf ? name .
} }
As we now have a list of all available properties for the ftg:Formalism we can proceed to query the parameters
which values we want.
To come back to the composite query example from Section 3.3.5, we present the query which in our system will
deliver on that promise. Listing 6 returns the artefact location. We do this by first looking for the activity type which
creates requirements. We then follow its link to the formalism it generates and filter for requirement formalisms. From
here on, we move to the process trace by finding the artefact which corresponds to the formalism type we just found.
The process trace we are looking for will contain an instance of this artefact. So we match and filter for the requirement
artefact instances in the process trace. We then return its location property to the end user.

Listing 6: Composite SPARQL query showing multiple categories being used in conjuntion.
SELECT DISTINCT ? prelement ? tag ? location
WHERE {
? individual a ? type .
? type rdfs : subClassOf * ftg : Transformation .
? individual base : hasGUID " C r e a t e M o d e l A n d E s t i m a t e P a r a m e t e r s " .

First Author et al.: Preprint submitted to Elsevier Page 20 of 31

 Model management for system engineering using ontology and knowledge graphs

? individual ftg : hasOutput ? formalism .

? formalism base : hasGUID " CBD " .
? formalism pm : occursAsArtifact ? iri .
? element pt : relatesTo ? iri .
? element a ? atype .
? atype rdfs : subClassOf pt : Artifact .
? element base : nextVersionOf * ? prelement .
? prelement pt : tag ? tag .
? prelement pt : hasLocation ? location .
}

4.7. Federation services

When working with existing large-scale systems, we don’t want to convert all data to RDF, as this would lead to a
blowup in data size. Therefore, we implemented middleware10 and services to solve the problem of having to store all
the data in the knowledge graph.

Contribution

Spendpoint Service Fuseki

User

Sparql query
1

Custom request
2

Custom response
3
opt [Additional data request]

Sparql sub-query
4

Sparql sub-query
5

Sparql answer
6

Spendpoint Service Fuseki

User

Figure 16: Order of operations for a custom service call.

The middleware intercepts SPARQL queries and performs intermediate operations, like calling a service. This
process is shown in Figure 16. We start with a user who wants to perform a query or service call. The query itself is
structured as in Listing 7. Note the SERVICE and BIND keywords which are used to perform the sub-query. Once the
query arrives at the middleware, it is processed and the relevant parts are extracted. The parameters required for the
subsequent service call are all present within the BIND instruction. Spendpoint then performs a call to the specified
service, passing on all the parameters. Once the service response is received, we proceed to parse it and convert it into
a graph representation, so it can be sent to the user in the form of a SPARQL response. Note that steps 1 and 4 can
also be reversed: the main query can be sent to the Fuseki server, which will forward the relevant part towards the
Spendpoint middleware.
10 /spendpoint

First Author et al.: Preprint submitted to Elsevier Page 21 of 31

 Model management for system engineering using ontology and knowledge graphs

This way, we can create models of the data, which is then populated on-demand by a service call. The middleware
can operate on any type of endpoint and converts the acquired data into SPARQL-compliant resources. We can add
service endpoints to expand the possible list of operations. For the proof of concept we provide 3 services:
1. An outlier service
2. A CSV to PARQUET converter
3. An example service which returns predefined data

Listing 7: SPARQL query for outlier detection.

SELECT ? outlier ? outlier_relation ? outlier_value WHERE {
SERVICE <{{ service_endpoint }} > {
SELECT ? outlier ? outlier_relation ? outlier_value WHERE {
BIND (
dtf : outlier (
" {{ ␣ outlier_file ␣ }} " ,
" {{ ␣ outlier_column ␣ }} " ,
" {{ ␣ outlier_artefact ␣ }} "
) AS ? outlier
) .
FILTER (? outlier_relation IN ( tabular : holdsContent , tabular : hasRowPosition ))
}
}
}

4.8. Graph Explorer

The Graph Exploring Tool (or GET)11 is a desktop application that allows exploring the knowledge graph using
SPARQL queries. The tool includes templates for quick traversal of compliant knowledge graphs. The tool separates
actions into multiple panels (Figure 17):
• The left Query Palette panel (red, leftmost) allows loading a predefined template. Hovering over the name of the
template will provide a tooltip with the description of the action carried out by that template.
• The centre and main Query Editor panel (green, upper) is where the template is shown once it has been chosen.
The template itself can be edited or the provided fields (blue) can be used to insert references. The reference
fields support dragging and dropping. The query can then be performed with the Query-button.
• The small one line panel (yellow, below green) shows the status. This includes messages whether a query was
successful. Any error will also be displayed right here.
• The most important panel is the Results panel (orange, bottom). Any information returned by the query will be
displayed here. You can drag and drop any of the fields in the result onto the parameters fields in the template.
Listing 7 is an example of a template which is used by the tool. The templated values are filled in using the GUI
and can be dragged and dropped from the results panel. Advanced users are also able to define their own templates or
perform custom queries. All templates need the query, its parameters, and an explanation of what the template does.

5. Case Study
In this section the obtained model management strategy is applied to a study looking for alternative measurement
strategies in a drivetrain setup, in the field of mechanical engineering. Section 5.1 explains the motivation to explore
alternative measurement strategies on existing drivetrains. Section 5.2 indicates the issues and difficulties that were
encountered in the study without the help of an external model management strategy. Section 5.3 explains how the
obtained model management strategy and environment can improve the workflow for exploring different models and
strategies for alternative measurement solutions. Section 5.4 describes how the outlier detection service can be applied
to the drivetrain case and indicates which type of aspects can be given to a data retrieval query. Finally section 5.5
summarizes the conclusions of using this tool in this type of applications.
11 /graph-exploring-tool

First Author et al.: Preprint submitted to Elsevier Page 22 of 31

 Model management for system engineering using ontology and knowledge graphs

Figure 17: Annotated Graph Exploring Tool user interface.

Figure 18: Drivetrain setup.

5.1. Motivation for alternative measurement strategies

The case study focuses on a study performed on a drivetrain setup consisting of two induction motors positioned
in a back-to-back configuration, with an interchangeable driveshaft [35]. In this setup, one motor serves as the drive
motor responsible for controlling the speed, while the other motor acts as a load motor, mimicking a specific torque
profile. Both motors are equipped with rotational position and acceleration sensors, and a torque sensor is mounted
between the driveshaft and the load motor. The setup is shown in Fig. 18.
The objective of the study was to explore an alternative approach for measuring the rotational states of motion,
and if possible, the torque as well. The reason to look for alternative solutions is the intrusiveness and high cost of the

First Author et al.: Preprint submitted to Elsevier Page 23 of 31

 Model management for system engineering using ontology and knowledge graphs

Volume serial number is 7ECF-EA90

C:.
├──2D corrected x and y sensors
│ └──codegen
│ └──mex
│ └──kinematic_estimator_2D_original
│ └──interface
├──2D lateral only cross coupling
│ └──codegen
│ └──mex
│ ├──kinematic_estimator_2D_original
│ │ └──interface
│ └──kinematic_estimator_2D_residual
│ └──interface
├──2D only lateral acceleration
│ └──codegen
│ └──mex
│ └──kinematic_estimator_2D_residual
│ └──interface
├──2D lateral only acceleration updated
│ └──codegen
│ └──mex
│ ├──kinematic_estimator_2D_original
│ │ └──interface
│ └──kinematic_estimator_2D_residual
│ └──interface
└──codegen
└──mex
├──kinematic_estimator_2D
│ └──interface
└──kinematic_estimator_2D_residual
└──interface

Figure 19: Original folder structure.

current state-of-the-art sensors. One potential method involves using accelerometers placed on the periphery of the
driveshaft, and combining the readings from these sensors with a physical model that describes the accelerometers’
rotation within the gravitational field. This method enables flexible sensor placement on existing drivelines, as the
sensors can be added on the driveshaft without having to dismount it.
During the model exploration phase, various aspects were considered, including sensitivity analyses of different
model parameters, ranging from basic to more complex models. Optimization methods were employed to determine
unknown parameters of both the model and sensors. Furthermore, the study evaluated whether the increased model
complexity and increased challenges in identifying the required parameters are justified by the improved accuracy.
Both simulation and experimental validation were conducted, which presented important challenges in managing both
the models and the obtained experimental data.

5.2. Difficulties in model management for research

Throughout the model exploration phase, extensive investigations were conducted involving various model
combinations and data sources, each yielding distinct results and conclusions. This process resulted in the creation
of an extensive folder structure with numerous subfolders, each containing a different approach. However, due to the
absence of a comprehensive structured framework to organize the required information, consistency was lost between
the different folders. This made it increasingly difficult to navigate through them and locate the necessary models.
Furthermore, since the direction in which the models would evolve was uncertain, a consistent naming convention was
not implemented. Consequently, the file names became more inconsistent with the content as the model exploration
process progressed. The following sections outline the issues encountered and explain how the lack of proper tools to
manage the relevant information resulted in significant time loss.

5.2.1. Folder structure and file naming

To show an example, Fig. 19 shows the folder structure in an early phase of the study, where different aspects were
investigated throughout the different folders. In the folder named ’2D lateral only cross coupling’, only the influence
of cross coupling in accelerometers in the lateral direction was taken into account. In the folder named ’2D only lateral
acceleration’, both the cross coupling and mounting angle were taken into account, also only in the lateral direction.
At the time, this way of naming appeared convenient and clear, but as more combinations of terms in the models

First Author et al.: Preprint submitted to Elsevier Page 24 of 31

 Model management for system engineering using ontology and knowledge graphs

C:.
Info.xlsx
kinematic_estimator_base_2D.m
kinematic_estimator_base_2D_CC.m
kinematic_estimator_base_2D_CC_dual_1_sensor.m
kinematic_estimator_deflection_2D_fixed_CC.m
kinematic_estimator_deflection_2D_variable_CC.m
kinematic_estimator_deflection_2D_variable_CC_reduced.m
Torque_estimator_dual_axis.m
Torque_estimator_single_axis_lateral.m

Figure 20: Original folder structure.

Figure 21: Kinematic models of different complexity.

were investigated, new ways of naming were needed. Those were not consistent with the original way of naming,
making it more difficult to keep track of the contents. Within both folders the goal was to investigate whether the
unknown parameters could be obtained through optimization. An important mistake here was that the model file named
’kinematic_estimator_2D_residual’ had different contents in the different main folders, being different parameters,
requiring different inputs, etc. Due to this, a large obstacle encountered during the exploration phase was that a file
with the correct model name was used, but it originated from a wrong folder that explored a certain strategy different
from the one that was envisioned. The reason this was not noticed was the different models having the same file
name in different main folders, making it unclear from which folder this model eventually originated. When selecting
this specific model for further deployment, unexpected results were observed, leading to a time consuming search
for the error in the model, as it was not clear what caused these unexpected results. Had all of the model files had a
file:///D/digitaltwin/paper/figs/core_models.txt[30/06/2023 14:46:28]
clear naming convention, or a consistent versioning strategy tracing each version to its corresponding conclusions and
validation strategy, this could have been avoided.
To deal with this issue at the time, after the error was found, all models were put together in one folder location, and
the path to these folders is added to the scripts that call for these models when evaluating them, preventing accidental
copy-paste errors. An example of this folder is shown in Fig. 20.

5.2.2. Model versions

An important issue with above mentioned strategy however is that one specific model can also have different
versions. Fig. 21 shows 3 different kinematic models each showing a different level of complexity in representing
rotating accelerometers. For each of these models, a different version can for example use a different discretization
method for the same physical model, or an adaptive covariance update instead of a constant one, etc. This also calls
for an external strategy to keep track of all versions with their respective conclusions. As a first solution to this issue, a
simple Excel file containing the version updates was used together with their conclusions and important remarks. This
however might not be sufficient in the long run, as not all relevant information to be able to navigate the model versions
later on might be available. Using this method to go back to a certain version, the model implementation itself needs
to be updated each time, where it would be more efficient to have copies of every version that can be accessed at any

First Author et al.: Preprint submitted to Elsevier Page 25 of 31

 Model management for system engineering using ontology and knowledge graphs

time. However, doing this in an ad-hoc manner, while having no consistent naming system or information storage tool
available, it can quickly become complicated and error-prone.

5.2.3. The need for experiment modelling

Another important problem source was the experimental data used to evaluate the models. To ensure accurate
parameter identification, specific criteria had to be met by the data associated with each model. These criteria
encompassed various aspects such as a minimum sampling frequency, the frequency spectrum content of the
acceleration, a threshold for rotational acceleration in constant speed profiles, and more. Moreover, not only the
requirements for the data but also the data source itself could vary. This variation included the use of different
accelerometers or even a distinct driveshaft, which had the potential to impact both the model and sensor parameters.
Consequently, it became essential to carefully document relevant details of the experimental data, such as its content,
the objectives of the conducted experiments, setup configurations, the actual equipment used, sampling frequencies,
and so on. When moving on to a model of different complexity, the data requirements may or may not change. Possibly
the required data might already exist from previous experiments. However, locating it can be a tedious task if it hasn’t
been appropriately linked to the necessary information, especially when dealing with a multitude of experiments. To
avoid having to inspect the data manually or redoing experiments of which the data already exists, it is essential that
all the required information is well documented for each experiment. All of the above mentioned aspects also apply
to simulations that emulate the sensor data. For simulations, scripts with parameter settings were used to emulate
the setup configuration. By keeping track of the exact content of the scripts, all these simulations are repeatable and
can be reused when needed. Section 5.3 describes how the developed tools can improve the model design workflow
by providing an environment that allows for uploading the essential files together with their documentation for each
activity in performing an experiment on the drivetrain.

5.3. Design flow improvement by integration of the obtained environment

The issues described in section 5.2 can be tackled by applying the model management framework as discussed in the
previous sections. For each step in the workflow, the reasoning behind it, the conclusions, and the exact data used can all
be documented in a structured way, where it is later accessible by executing queries on the resulting knowledge graph.
Fig. 22 shows part of the process model for conducting experiments on the drivetrain which was mentioned in section
5.1. The artefact ’drivetrain_limitations’ can be a document summarizing the setup and its limitations, a datasheet, or
other documentation. Given the limitations, a torque profile is created that matches the test requirements. Both files and
the generated torque profile can then be uploaded through the WEE environment as described in section 4.4 and linked
to the activity ’set_up_drivetrain_profile’. In this way, the data generated during this experiment will always be linked to
the setup, the test requirements and the actual torque profile. In the next activity ’Generate_matlab_control_parameters’,
the coefficients for a 10th order linear controller are generated from the torque profile and a script that generates
these coefficients. By uploading all artefacts in WEE, including the controller coefficients which are saved in tabular
form, all settings and correct files will again be linked to the data that will result from this experiment. The rest of
the process model is not shown, but follows a similar approach that links all relevant artefacts to the corresponding
activity. A similar process model can also be set up for simulations, where all relevant settings and simulated models
can then be included as artefacts. Such a process model can also be built for developing and evaluating a new physical
model of certain complexity. The WEE environment then allows for uploading each model version, with its associated
reasoning, conclusions and requirements for validation data, be it experimental or simulation based. This ensures that
all conclusions match with the correct model version and used data. This approach effectively deals with all issues
described in section 5.2. The Graph Exploring Tool (GET) can afterwards be used to efficiently query the resulting
Knowledge Graph (KG) for the required model or data, as well as execute services on the content of the KG. In this way,
a lot of time can be saved in retrieving both a required model, or data that meets specific requirements. The next section
briefly describes the outlier detection service in context of the drivetrain study. It also indicates some requirements that
can be given to a query to retrieve certain data.

5.4. Sub-application: outlier detection and querying for specific data

The developed framework not only allows for keeping track of all essential information, it also provides services
that can be executed on the existing data. For example, outlier detection as described in section 4.7 can be used to
check if a certain data set respects the limitations for outliers. Should a new physical model require a limitation on the
outliers of the rotational speed data for parameter identification, and this has not been included as information during

First Author et al.: Preprint submitted to Elsevier Page 26 of 31

 Model management for system engineering using ontology and knowledge graphs

Figure 22: Part of the process model representing the workflow for conducting experiments on the drivetrain.

the experiment, this service allows for querying for a data set that complies with this limitation. This eliminates the
need to manually inspect the available data or to redo an experiment for which the data already exists.
The GET also allows the user to retrieve experimental data that meets specific requirements by querying the
resulting KG. Such a query can for example demand the data to comply with all of the following requirements:
1. Time series data
2. Acceleration data (𝑚∕𝑠2 )
3. Measured with an accelerometer of type X
4. Measured on test setup Y
5. Motor speed of setup Y constant at 1000rpm
6. Sampling frequency of 2kHz
7. Minimum data length of 10s
This allows for efficient retrieval of this data, without having to inspect the data manually.

5.5. Summary case study

To summarize, the provided framework allows for significant improvements on following levels, compared to ad-
hoc methods:
1. Traceable link between the experimental data and the developed models: Establishing a traceable link between
experimental data and the developed models is essential for understanding how the models were constructed and
validated. By documenting the data sources, experimental setup, and measurement techniques used during the
data collection process, researchers can ensure that the models are based on reliable and representative data. This
traceability enables transparency and reproducibility, allowing others to verify and validate the existing models
or even newly developed models by referring back to the original experimental data.

First Author et al.: Preprint submitted to Elsevier Page 27 of 31

 Model management for system engineering using ontology and knowledge graphs

2. Traceable link between the experimental data and the equipment used to obtain it: In model development for this
kind of applications, the equipment used during experiments, such as the sensors, test rigs, or even simulation
tools, can significantly impact the quality and accuracy of the collected data. Therefore, it is crucial to keep track
of the link between the experimental data and the specific equipment used. This traceability provides insights
into the reliability, calibration, and limitations of the equipment, ensuring that the data is interpreted within the
appropriate context. It also helps in troubleshooting and identifying potential sources of error or bias in the used
experimental setup.
3. Model versioning with a link to a detailed description of the reasoning behind it: As the development of
the physical models evolves over time, it becomes necessary to manage different versions or iterations of the
models. Model versioning with the developed tools allows to track and document changes made to the models,
including modifications, improvements, or updates. By attaching traces to the documentation of the underlying
reasoning, assumptions, or methodologies to each version, the development process becomes more transparent
and understandable. This documentation facilitates collaboration, knowledge sharing, and informed decision-
making by providing a historical context for each of the model versions. It can also help future researchers to
build further upon the reasoning behind the models that were developed by other researchers.
All of above aspects are then part of an overarching KG. By having this KG available, together with a tool that
allows for domain specific querying, a lot of time can be saved during the model development process. As for the
drivetrain study described earlier for example, the mistake of using a file with the correct name from a wrong folder
would not have happened had all models been correctly linked to their reasoning and validation conclusions. Instead of
looking for the error in the model itself, or validating the model with new data, the KG could have been searched based
on the conclusions or reasoning behind this model to find the correct corresponding model. This mistake ultimately led
to a time loss of approximately two weeks, which could have been avoided had the proposed framework been available.

The main limiting factor of this framework is that for each new workflow the FTG and PM will need to be set up
and linked to the correct environment. However, this well outweighs the potential time loss by not having above
mentioned traceability available.

6. Related Work
Given the shift from document-centric to model-based system engineering, the interest in model management
strategies and frameworks has increased in recent years. However, the variety of disciplines performed during system
engineering activities makes use of a plethora of different notations, languages and formalisms. Therefore, it is essential
that model management frameworks provide support for multi-paradigm modelling. In addition to that, the description
of workflows for the large number of processes executed is also essential to context the usage of artefacts from different
formalisms in an organised flow of activities.
Regarding workflow management and enactment, Product Lifecycle Management (PLM) software proposes to
manage a product and its associated data through its development [36]. Predominantly it is used by design and
engineering teams working with CAD data and related artefacts. Its primary purpose is to integrate processes
for each stage of a product’s lifecycle to support supply chain management, but it is also used to improve team
collaboration and provide insights on product performance, customer feedback, and market trends. Usually, these
tools are commercial-off-the-shelf (COTS) developed by companies that also create CAD software, like Siemens,
Dassault and PTC [37, 38, 39]. Very little information is available on the inner workings of these tools, because they are
proprietary. Also, as their focus is on supply management, relations between data from artefacts may exist, but usually,
they are developed in an ad-hoc manner, i.e., based on customer requests, the tool providers can implement specific
data relations to provide the relevant functionality for the end-user. PLM tools usually support simple file versioning
and workflow modelling and enactment, but again this is typically developed in an ad-hoc manner: workflow models
are ‘special’, and their meta-models (and those of workflow traces) are implicit, as opposed to our solution which
has explicit meta-models for all models. As we use an ontology to define the supported formalisms, it is easier to
provide fine-grained traceability based on these relations. Also, having data structured in a knowledge graph gives
more flexibility to provide insights for the customer through querying instead of being software-dependent.
OpenMBEE (Open source Model-Based Engineering Environment) proposes an architecture to support an
integrated environment for engineering driven by connected models [40]. The initial motivation was to shift from
document-centric engineering processes to model-based strategies where documents could be automatically rendered

First Author et al.: Preprint submitted to Elsevier Page 28 of 31

 Model management for system engineering using ontology and knowledge graphs

from a unified model repository controlled by a Model Management System (MMS). Cross-References synthesise
relevant engineering information where model elements are transcluded [41]. Besides document generation, the central
model repository can support other techniques like model checking and simulation. Model Development Kits (MDKs)
allow the integration of different environments (e.g., Jupyter Notebook, MagicDraw), languages (e.g., SysML, MatLab)
and tools with the MMS. Currently, OpenMBEE supports model versioning and typing/introspection. Traceability is
limited, given that engineering workflows are not fully supported. Only SysML-based structure model elements are
linked using natural language processing, for instance, linking references in text requirements to SysML blocks or block
properties. Querying is supported by translating the SysML model elements from MMS to RDF-based knowledge
graphs and exposing a SPARQL endpoint. Workflow enactment is not currently supported.
DesignSpace [42] is another model repository aiming to bring models from different tools together in a unified,
graph-based (EMF-like) data structure, supporting fine-grained traceability, typing and incremental global consistency
checking. DesignSpace has an operation-based versioning system, meaning that it records edit operations on models
as they happen. To make this work with heterogeneous tools, adaptor plugins for these tools have to be written. A side-
effect of DesignSpace’s operation-based versioning is that every adaptor enables real-time/synchronous collaboration
in the tool it is written for, via DesignSpace’s central repository. Services can be added to DesignSpace that respond
to arbitrary changes. One such service is the consistency checker, which informs users of inconsistencies in real-time.
DesignSpace does not have workflow enactment, although it should be possible to write one as a DesignSpace service.
While DesignSpace has been used successfully in several industrial experiments, it is not clear if DesignSpace scales
to large amounts of dense data (e.g., very large tables containing execution traces), for instance by supporting data
federation as described in Section 4.7.

7. Conclusions
Companies face specific challenges, such as retaining knowledge of standard procedures and successful cases, and
managing the generation and storage of multiple versions of documents and artefacts. The role of experiments in system
engineering is significant in providing evidence of system performance, but we also acknowledge the complexity
involved in executing them.
We leveraged ontologies to define system engineering elements formally. This approach enables the construction of
a knowledge graph to manage system engineering data, and allows us to link artefact versions to executed workflows.
Our ontology-based framework consists of seven components, together solving our requirements for explicitly modeled
workflows and their automated enactment, and a model repository with versioning, fine-grained traceability, typing,
consistency checking and a query language.
We showed how our framework is applied to a spring-mass-damper running example and to a larger engineering
use-case involving the creation of smart sensors for a drivetrain.

7.1. Limitations and future work

There are a few limitations to our solution. Many are inherent to our technology choices of OML and OWL/RDF.

OML lacks reasoner and query language Another problem of using OML is that it needs to be transformed to
OWL/RDF, before it can be used by a reasoner (to perform a consistency check) or query engine. As a result, any
detected inconsistency results in an (often cryptic) error at the level of OWL/RDF, which the user must manually trace
back to the original OML document(s). Likewise, SPARQL queries work on generated RDF, and return RDF data, not
OML. To make OML more usable, we think a reasoner and query language should be created that works on the level
of OML (possibly built around an OWL/RDF reasoner and query engine).

Types are not instances OML makes a strict distinction between type-level and instance-level elements. A type
cannot be an instance at the same time. If our types evolve, we would want to store versioning links between the types.
It is impossible however to create links at the type level. At the type level, we can only express the possibility of having
links between the instances of types. Another consequence is that a type cannot have a type. However, OML is still a
young language, and it is our hope that this restriction will be relaxed in the future.

Poor man’s versioning does not scale As already explained in Section 4, OML has no builtin solution for
versioning. OML users are advised to use Git for versioning the OML documents (types and instances) themselves.
This makes it impossible to use a query language such as SPARQL to navigate an artefact’s history. A query such as “in

First Author et al.: Preprint submitted to Elsevier Page 29 of 31

 Model management for system engineering using ontology and knowledge graphs

which version was this element added?” cannot be performed. To make versioning information available from within
OML/OWL/RDF, we create a new copy of the artefact for every new version. This solution is sufficient to demonstrate
a proof-of-concept, but ultimately doesn’t scale. Since RDF is graph data, not text, we think a versioning system for
graphs would be most appropriate. To make it possible to access older versions of our graph data via queries, this
versioning system would have to be integrated with the query engine and reasoner. We are not aware of an existing
solution, and this could be an interesting ally for further research.

References
[1] INCOSE, Systems engineering vision 2035, https://www.incose.org/about-systems-engineering/se-vision-2035, accessed:
2023-06-12 (2022).
[2] R. Paredis, J. Exelmans, H. Vangheluwe, Multi-paradigm modelling for model based systems engineering: Extending the ftg + pm, in: 2022
Annual Modeling and Simulation Conference (ANNSIM), 2022, pp. 461–474. doi:10.23919/ANNSIM55834.2022.9859391.
[3] L. M. Applegate, B. R. Konsynski, J. F. Nunamaker, Model management systems: Design for decision support, Decision Support Systems
2 (1) (1986) 81–91. doi:https://doi.org/10.1016/0167-9236(86)90124-7.
URL https://www.sciencedirect.com/science/article/pii/0167923686901247
[4] J. J. Elam, J. C. Henderson, L. W. Miller, Model management systems: an approach to decision support in complex organizations, in:
International Conference on Interaction Sciences, 1980.
[5] A. Baldwin, D. Baldwin, T. Sen, The evolution and problems of model management research, Omega 19 (6) (1991) 511–528. doi:https:
//doi.org/10.1016/0305-0483(91)90002-B.
URL https://www.sciencedirect.com/science/article/pii/030504839190002B
[6] A. Bharadwaj, J. Choobineh, A. Lo, B. Shetty, Model management systems: A survey, Ann. Oper. Res. 38 (1–4) (1992) 17–67. doi:
10.1007/BF02283650.
URL https://doi.org/10.1007/BF02283650
[7] P. P.-S. Chen, The entity-relationship model—toward a unified view of data, ACM Trans. Database Syst. 1 (1) (1976) 9–36. doi:
10.1145/320434.320440.
URL https://doi.org/10.1145/320434.320440
[8] E. F. Codd, A relational model of data for large shared data banks, Commun. ACM 13 (6) (1970) 377–387. doi:10.1145/362384.362685.
URL https://doi.org/10.1145/362384.362685
[9] A. M. Geoffrion, An introduction to structured modeling, Manage. Sci. 33 (5) (1987) 547–588.
[10] C. V. Jones, An Introduction to Graph-Based Modeling Systems, Part I: Overview, ORSA Journal on Computing 2 (2) (1990) 136–151.
doi:10.1287/ijoc.2.2.136.
URL https://ideas.repec.org/a/inm/orijoc/v2y1990i2p136-151.html
[11] C. V. Jones, An introduction to graph-based modeling systems, part ii: Graph-grammars and the implementation, ORSA Journal on Computing
3 (1991) 180–206.
[12] S. O. Kimbrough, A graph representation for management of logic models, Decision Support Systems 2 (1) (1986) 27–37. doi:https:
//doi.org/10.1016/0167-9236(86)90118-1.
URL https://www.sciencedirect.com/science/article/pii/0167923686901181
[13] W. W. W. Consortium, Semantic web, https://www.w3.org/standards/semanticweb/, accessed: 2023-06-12.
[14] P. Hitzler, A review of the semantic web field, Commun. ACM 64 (2) (2021) 76–83. doi:10.1145/3397512.
URL https://doi.org/10.1145/3397512
[15] T. R. Gruber, A translation approach to portable ontology specifications, Knowledge Acquisition 5 (2) (1993) 199–220. doi:https:
//doi.org/10.1006/knac.1993.1008.
URL https://www.sciencedirect.com/science/article/pii/S1042814383710083
[16] S. Mostefai, A. Bouras, M. Batouche, Effective collaboration in product development via a common sharable ontology, International Journal
of Computational Intelligence 2 (01 2006).
[17] C. Bock, X. Zha, H. won Suh, J.-H. Lee, Ontological product modeling for collaborative design, Advanced Engineering Informatics 24 (4)
(2010) 510–524, construction Informatics. doi:https://doi.org/10.1016/j.aei.2010.06.011.
URL https://www.sciencedirect.com/science/article/pii/S1474034610000558
[18] N. Noy, Y. Gao, A. Jain, A. Narayanan, A. Patterson, J. Taylor, Industry-scale knowledge graphs: Lessons and challenges, Commun. ACM
62 (8) (2019) 36–43. doi:10.1145/3331166.
URL https://doi.org/10.1145/3331166
[19] G. Xiao, L. Ding, B. Cogrel, D. Calvanese, Virtual Knowledge Graphs: An Overview of Systems and Use Cases, Data Intelligence 1 (3)
(2019) 201–223. arXiv:https://direct.mit.edu/dint/article-pdf/1/3/201/683759/dint\_a\_00011.pdf, doi:10.1162/
dint_a_00011.
URL https://doi.org/10.1162/dint_a_00011
[20] Neo4j, From graph to knowledge graph: How a graph becomes a knowledge graph, https://neo4j.com/blog/
from-graph-to-knowledge-graph-how-a-graph-becomes-a-knowledge-graph/, accessed: 2023-06-12.
[21] Ontop, Ontop - a virtual knowledge graph system, https://ontop-vkg.org/, accessed: 2023-06-12.
[22] G. Schreiber, Y. Raimond, Rdf 1.1 primer w3c working group note, Online (2014).
URL https://www.w3.org/TR/rdf11-primer/

First Author et al.: Preprint submitted to Elsevier Page 30 of 31

 Model management for system engineering using ontology and knowledge graphs

[23] W3C, SPARQL 1.1 Query Language, Tech. rep., W3C (2013).
URL http://www.w3.org/TR/sparql11-query
[24] P. Hitzler, M. Krötzsch, B. Parsia, P. F. Patel-Schneider, S. Rudolph, Owl 2 web ontology language: Primer (second edition) (2012) W3C
Recommendation.
URL http://www.w3.org/TR/owl2-primer
[25] P. Hitzler, M. Krötzsch, S. Rudolph, Foundations of Semantic Web Technologies, Chapman and Hall/CRC Press, 2010.
[26] M. A. Musen, The protégé project: a look back and a look forward, AI Matters 1 (4) (2015) 4–12. doi:10.1145/2757001.2757003.
URL https://doi.org/10.1145/2757001.2757003
[27] M. Elaasar, N. Rouquette, Ontological modeling language, http://www.opencaesar.io/oml/, accessed: 2023-06-12.
[28] M. Elaasar, An eclipse ide that supports oml natively - rosetta, https://github.com/opencaesar/oml-rosetta, accessed: 2023-06-12.
[29] A. S. Foundation, Apache jena, https://jena.apache.org/, accessed: 2023-06-12 (2021).
[30] D. Allemang, J. Hendler, F. Gandon, Semantic Web for the Working Ontologist: Effective Modeling for Linked Data, RDFS, and OWL, 3rd
Edition, Vol. 33, Association for Computing Machinery, New York, NY, USA, 2020.
[31] J. Amerongen, Dynamical systems for creative technology (01 2010).
[32] R. Monroe, S. Shaw, On the transient response of forced nonlinear oscillators, Nonlinear Dynamics 67 (03 2011). doi:10.1007/
s11071-011-0174-4.
[33] J. Seck Tuoh Mora, M. Gonzalez-Hernandez, N. Hernández Romero, A. Trejo, S. Chapa, Modeling linear dynamical systems by continuous-
valued cellular automata, International Journal of Modern Physics C - IJMPC 18 (2007) 833–848. doi:10.1142/S0129183107010589.
[34] D. Moody, The “physics” of notations: Toward a scientific basis for constructing visual notations in software engineering, IEEE Transactions
on Software Engineering 35 (6) (2009) 756–779. doi:10.1109/TSE.2009.67.
[35] B. Forrier, F. Naets, W. Desmet, Broadband load torque estimation in mechatronic powertrains using nonlinear kalman filtering, IEEE
Transactions on Industrial Electronics 65 (3) (2018) 2378–2387. doi:10.1109/TIE.2017.2739709.
[36] A. Saaksvuori, A. Immonen, Product lifecycle management, Springer Science & Business Media, 2008.
[37] Siemens, Teamcenter plm siemens software, https://plm.sw.siemens.com/en-US/teamcenter/, accessed: 2023-06-12.
[38] Oracle, Product lifecycle management software oracle, https://www.oracle.com/scm/product-lifecycle-management/, accessed:
2023-06-12.
[39] PTC, Arena plm and qml solutions ptc, https://www.ptc.com/en/products/arena, accessed: 2023-06-12.
[40] OpenMBEE, Openmbee - open model-based engineering environment, https://www.openmbee.org/, accessed: 2023-06-12.
[41] T. B. Chammard, B. Regalia, R. Karban, I. Gomes, Assisted authoring of model-based systems engineering documents, in: Proceedings of the
23rd ACM/IEEE International Conference on Model Driven Engineering Languages and Systems: Companion Proceedings, MODELS ’20,
Association for Computing Machinery, New York, NY, USA, 2020. doi:10.1145/3417990.3421406.
URL https://doi.org/10.1145/3417990.3421406
[42] M. A. Tröls, L. Marchezan, A. Mashkoor, A. Egyed, Instant and global consistency checking during collaborative engineering, Softw. Syst.
Model. 21 (6) (2022) 2489–2515. doi:10.1007/s10270-022-00984-4.
URL https://doi.org/10.1007/s10270-022-00984-4

First Author et al.: Preprint submitted to Elsevier Page 31 of 31