Ph. D. Project
Title:
Data-driven modeling for building and updating digital twins of cyber-physical systems: application to the ANR
JUNEAU project
Dates:
2024/04/08 - 2027/04/07
Student:
Supervisor(s):
Other supervisor(s):
Pr. Virginie GOEPP (virginie.goepp@insa-strasbourg.fr)
Description:
The redesign/reconfiguration of industrial systems relies, among other things, on behavioral models, notably
discrete-event simulations (SED) [1, 2].The construction of these models is an important issue, as the
performance of the redesign/reconfiguration depends on them. Several aspects come into play in the
construction of these models, such as the time and expertise required to build them, as well as their validity,
credibility and degrees of precision and granularity, depending on the modeling objective and targeted
performance indicators [3].
In recent years, a number of studies [4-7] have looked at data-driven modeling/simulation as a means of
automating the construction and consequent acceleration of more accurate and complete models of complex
systems. The underlying principle of data-driven modeling/simulation is to build models using real data from
the system (sensor measurements, computer system logs, etc.). Obtaining the data needed for data-driven
modeling/simulation is a (major) bottleneck in the process [8]. Its main origins lie in the lack of knowledge
and/or absence of the data needed to build the model for the redesign/reconfiguration problem being
addressed. Eliminating this bottleneck requires the rapid availability of generic (reusable) models and suitable
measurement systems.
In this context, in contrast to existing work that proposes ad hoc data-driven simulation models, a first
objective is to propose generic models that can be rapidly adapted to specific needs. This requires first of all
characterizing the data needed for simulation models according to the redesign/reconfiguration problem
under consideration, and in parallel identifying the models that can be obtained from existing data, with regard
to the performance indicators envisaged.
The data produced by the real system will therefore be used to instantiate the proposed generic models, and
will also be used to connect the resulting simulation model with the real system, to the point of creating a
digital shadow. This model will necessarily have to be updated in line with minor/major changes in the real
system. A second objective, contiguous to the first, will therefore be to study the possibilities of modifying,
dynamically over time, the structure of the initial model with the real system, while respecting the model's
topology. In addition to the question of "how", there is also the question of "when" to carry out this update.
This will require the study of drift detection approaches between real/virtual systems, such as [9,10], or data
assimilation.
Particular attention wil need to be paid to the completeness of the data required to design/update simulation
models. If additional data needs to be obtained, it will then be necessary to specify the measurement system
requirements capable of meeting the information needs, with a view to assisting the human expert in the
implementation of data collection through ephemeral or non-ephemeral instrumentation, as a parallel layer to
the information systems already in place on the system of interest, rapidly implemented to acquire the missing
data at lower cost.
This doctoral project will be carried out on a research/action basis. Research/action is a research methodology
in which research activities and experimentation/observation in the factory mutually enrich each other. This
research method is based on loops comprising 3 phases: Action planning taking into account existing scientific
knowledge, Action/Observation to implement the proposals and Reflection to analyze the results obtained in
the field. The end of a loop can give rise to a new one, whenever necessary. Two fields of application are
envisaged for this work: on the one hand, the hospital world, with an application to the Saint-Etienne Pediatric
Emergency Department, which will be the main target of the thesis project. Secondly, and more
conventionally, in order to demonstrate the generalizability of the approach, we plan to apply this work to the
school factories of the Haguenau IUT (University of Strasbourg).
The techniques used will be those of discrete-event simulation (Witness, Arena, FlexSim software, etc.) and
business engineering for model characterization.
discrete-event simulations (SED) [1, 2].The construction of these models is an important issue, as the
performance of the redesign/reconfiguration depends on them. Several aspects come into play in the
construction of these models, such as the time and expertise required to build them, as well as their validity,
credibility and degrees of precision and granularity, depending on the modeling objective and targeted
performance indicators [3].
In recent years, a number of studies [4-7] have looked at data-driven modeling/simulation as a means of
automating the construction and consequent acceleration of more accurate and complete models of complex
systems. The underlying principle of data-driven modeling/simulation is to build models using real data from
the system (sensor measurements, computer system logs, etc.). Obtaining the data needed for data-driven
modeling/simulation is a (major) bottleneck in the process [8]. Its main origins lie in the lack of knowledge
and/or absence of the data needed to build the model for the redesign/reconfiguration problem being
addressed. Eliminating this bottleneck requires the rapid availability of generic (reusable) models and suitable
measurement systems.
In this context, in contrast to existing work that proposes ad hoc data-driven simulation models, a first
objective is to propose generic models that can be rapidly adapted to specific needs. This requires first of all
characterizing the data needed for simulation models according to the redesign/reconfiguration problem
under consideration, and in parallel identifying the models that can be obtained from existing data, with regard
to the performance indicators envisaged.
The data produced by the real system will therefore be used to instantiate the proposed generic models, and
will also be used to connect the resulting simulation model with the real system, to the point of creating a
digital shadow. This model will necessarily have to be updated in line with minor/major changes in the real
system. A second objective, contiguous to the first, will therefore be to study the possibilities of modifying,
dynamically over time, the structure of the initial model with the real system, while respecting the model's
topology. In addition to the question of "how", there is also the question of "when" to carry out this update.
This will require the study of drift detection approaches between real/virtual systems, such as [9,10], or data
assimilation.
Particular attention wil need to be paid to the completeness of the data required to design/update simulation
models. If additional data needs to be obtained, it will then be necessary to specify the measurement system
requirements capable of meeting the information needs, with a view to assisting the human expert in the
implementation of data collection through ephemeral or non-ephemeral instrumentation, as a parallel layer to
the information systems already in place on the system of interest, rapidly implemented to acquire the missing
data at lower cost.
This doctoral project will be carried out on a research/action basis. Research/action is a research methodology
in which research activities and experimentation/observation in the factory mutually enrich each other. This
research method is based on loops comprising 3 phases: Action planning taking into account existing scientific
knowledge, Action/Observation to implement the proposals and Reflection to analyze the results obtained in
the field. The end of a loop can give rise to a new one, whenever necessary. Two fields of application are
envisaged for this work: on the one hand, the hospital world, with an application to the Saint-Etienne Pediatric
Emergency Department, which will be the main target of the thesis project. Secondly, and more
conventionally, in order to demonstrate the generalizability of the approach, we plan to apply this work to the
school factories of the Haguenau IUT (University of Strasbourg).
The techniques used will be those of discrete-event simulation (Witness, Arena, FlexSim software, etc.) and
business engineering for model characterization.
Keywords:
data-driven simulation, digital twin
Department(s):
| Modeling and Control of Industrial Systems |