BachelorarbeitenimWintersemester2021/22 ...

Institut für Systemtheorie und Regelungstechnik - Universität StuttgartProf. Dr.-Ing. Frank Allgöwer

21. September 2021

Bachelorarbeiten im Wintersemester 2021/22am Institut für Systemtheorie und Regelungstechnik

Liebe Studierende,

Sie sind auf der Suche nach einer Bachelorarbeit (BA) und begeistert von Regelungstechnik? Dann lesenSie unbedingt weiter, denn hier werden Sie über die Möglichkeit informiert, eine BA am IST anzufertigen.

Wie läuft eine Bachelorarbeit am IST ab?

• Neben dem Bearbeiten des Themas gehört das Anfertigen einer schriftlichen Ausarbeitung sowiedas Halten einer 20 minütigen Präsentation zum festen Bestandteil jeder BA.

• In den meisten Studiengängen ist das Hören von neun Vorträgen verpflichtend. Neben den Vor-trägen Ihrer Mitstudierenden bietet das IST daher drei Seminarvorträge an, die Ihnen wichtigeSchlüsselqualifikationen für die BA vermitteln.

• Der Start einer BA ist flexibel. Die Seminarvorträge werden nur zu Beginn des Semesters angebotenund sollten ggbf. bereits vor dem Start der BA gehört werden.

• Alle BAs können sowohl auf Deutsch als auch auf Englisch angefertigt werden.

• Einige Themen können auch in einer Gruppe von 2-3 Personen bearbeitet werden.

Wie kann ich mich auf eine ausgeschriebene Arbeit bewerben?

• Schicken Sie Ihre Bewerbung zentral an [email protected]. Wenn Sie an meh-reren Arbeiten Interesse haben, machen Sie eine Prioritätenliste. Bedenken Sie, dass die QualitätIhrer Bewerbung ausschlaggebend dafür sein kann, ob Sie ihr gewünschtes Thema erhalten.

• Bitte bewerben Sie sich nur auf Arbeiten, die zu Ihren Interessen und Fähigkeiten passen.

• Der Bewerbungsschluss ist Montag, der 10. Oktober 23:59. Spätestens dann erhalten Sie eineAntwort auf Ihre Bewerbung.

• Natürlich haben Sie die Möglichkeit, die jeweiligen Betreuer*innen kennenzulernen, bevor Sieoffiziell zu einem Thema zusagen müssen.

Welche Themen sind gerade ausgeschrieben?

• Am IST gibt es ein breites Spektrum an BAs, die von praktischen Arbeiten im Labor überSimulations- und Programmierarbeiten bis hin zu theoretischen Forschungsarbeiten reichen.

• An dieses Dokument sind alle aktuell ausgeschriebenen BAs angehängt, eine Übersicht finden Sieauf der nächsten Seite.

• Bei Fragen zu einem Thema können Sie sich direkt beim Betreuer melden.

• Bitte beachten Sie, dass die nächsten BA Themen erst wieder zum SS22 ausgeschrieben werden.

Bei Fragen zu einem ausgeschriebenen Thema melden Sie sich bitte direkt bei den jeweiligen Betreu-er*innen, bei organisatorischen Fragen wenden Sie sich bitte an mich.

Mit besten GrüßenLukas Schwenkel

1

Institut für Systemtheorie und Regelungstechnik - Universität StuttgartProf. Dr.-Ing. Frank Allgöwer

21. September 2021

Bachelorarbeiten im Wintersemester 2021/22am Institut für Systemtheorie und Regelungstechnik

Inhaltsverzeichnis1 Evaluation of event-triggered control mechanics 3

2 Bayesian neural networks vs. Lipschitz neural networks 4

3 Modelling of a 3-DOF-helicopter 5

4 Cloud-based encrypted control 6

5 Inferring system properties from inter-communication times of event-triggered con-trol 7

6 Controller design for a LEGO-Mindstorms robot 8

7 Swing-up control of an inverted pendulum 9

8 Neural network approximation of balancing controller for autonomous e-scooter 10

9 Comparison of data-driven and identification-based control 11

2

Universitat StuttgartInstitut fur Systemtheorie und Regelungstechnik

Prof. Dr.–Ing. Frank Allgower

Open Thesis (BA)

Evaluation of event-triggered controlmechanisms

Description:

Networked control systems(NCS) are control systemsin which some or all linksin the feedback loop are replaced by a shared communicati-on network. The classical periodic implementation of feed-back controllers with periodic transmissions at predefinedtransmission times may however lead to a waste of networkrecources for NCS. An alternative, called event-triggeredcontrol (ETC), is to transmit state information at time in-stants that are determined by a state-dependent trigger ru-le. Thus, transmissions are only triggered for ETC when thisis required due to the state evolution of the system. Ma-ny results confirm the advantageousness of ETC, as long asthe effect of ETC on the network utilization is neglected.Often ignored is the fact that ETC can lead to unbalancednetwork utilization which may result in worse control per-formance. The goal of this thesis is to evaluate existing ETCmechanisms from the literature in a simulation study and tocompare the resulting control performance to periodic con-trol. In this context, the effect of the ETC on the networkload shall be explicitly considered.

Prerequisites:

• Basic knowledge of feedback control (e.g. lectures Sys-temdynamik and Einfuhrung in die Regelungstechnik)

• Interest in theoretical problems and numerical simula-tions

Supervisor:

MichaelHertneckRoom 3.240

Area:

Networked ControlSystems

Properties:

Type: BA

Beginning:

At any time

Weitere Informationen: www.ist.uni-stuttgart.de/lehre/bama

Aushang vom 6. September 20213

Universit�at Stuttgart

Institut f�ur Systemtheorie und Regelungstechnik


Open Thesis (BA)

Bayesian neural networks vs.Lipschitz neural networks

Description:

Due to their susceptibilityto adversarial perturbations,neural networks (NNs) arehardly used in safety-criticalapplications. There havebeen multiple approaches totackle this problem. A mea-sure of robustness to such in-put pertubations is the Lipschitz constant of the input-outputmap defined by an NN. Hence, one approach to increase theNN’s robustness is to enforce an upper bound on the Lips-chitz constant by adding a matrix inequality constraint to thetraining objective. An alternative approach uses Bayesian neu-ral networks, i.e., stochastic neural networks that model theuncertainty in the parameters by learning a distribution overthe weights. This thesis consists of understanding both ap-proaches, implementing, comparing and potentially combiningthem to get an improved overall method.

Prerequisites:• Basic course on machine learning• Experience with Matlab and/or Python• Interested in neural networks and coding

Supervisor:Patricia Pauli,HenningSchluterRoom 3.234

Area:

Robustness of neuralnetworks

Properties:

Type: BA

40% literature20% theory40% simulation

Beginning:anytime

Further information on www.ist.uni-stuttgart.de/lehre/bama

Aushang vom September 13, 2021

4

Universität StuttgartInstitut für Systemtheorie und RegelungstechnikProf. Dr.–Ing. Frank Allgöwer

Open Thesis (BA)Modelling of a 3-DOF-helicopter

Description:

The shown three degrees offreedom (3 DOF) helicopteris part of the PraktikumKonzepte der Regelungstech-nik in the MSc EngineeringCybernetics. The prevalentCovid-19 pandemic has unfortunately shown that access tothis real plant cannot be expected at all times. For thisreason, it would be preferable to have an accurate model ofthe helicopter at hand which imitates the true behaviour.The goal in this thesis is to create a detailled model of the3 DOF helicopter, e.g. based on Euler-Lagrange equations.Ideally, the model should consider acting forces as well asphysical limitations. Further, it could be possible to usemodel identification techniques to obtain unknown parame-ters. This thesis complements the earlier project ”Visualisa-tion of a 3-DOF-helicopter”, which can be used to validatethe derived equations.

Prerequisites:• Experience in deriving equations of motion• Interest in modelling mechanical systems

Supervisor:Robin Strässer,Matthias HircheRoom 3.236

Area:

ModellingSimulation

Properties:

Type: BA60% literature20% implementation5% simulation15% experiments

Beginning:anytime


Aushang vom August 24, 2021

5



Open Thesis (BA)

Cloud-based encrypted control

Description: Plantx+ = f(x,u)Actuator Sensor

Controlleru = p(x)

u x

⟦u⟧ ⟦x⟧In cloud-based control, evalu-ation of the control law is out-sourced to a powerful exter-nal cloud computing service.Although this approach is highly flexible and allows to forgoexpensive hardware which runs the controller, there might beprivacy concerns since the cloud might not be trustworthy. Ifthe measurements, the control actions or the control policyare confidential, encrypted control can be used to computecontrol actions in the cloud without revealing any informa-tion about the data to the cloud provider. For this purpose,secure control schemes relying on multi-party computationhave been developed at the IST. The goal of this thesis is toimplement this scheme and possibly other encrypted controlschemes on real (external) servers, and to compare their real-time capabilities. Further, a software-in-the-loop simulation(or possibly even a hardware-in-the-loop simulation using oneof the IST’s experimental setups) using a cloud controller canbe set up, and the control performance be compared to a localcontroller.

Prerequisites:• Course Einfuhrung in die Regelungstechnik• Experience in Python and Matlab, having fun with pro-

gramming

Supervisor:Sebastian Schlor,StefanWildhagen

Area:

Encrypted Control,Multi-partycomputation

Properties:

Type: BA

15% literature15% theory60% implementation

Beginning:October 2021



6



Open Thesis (BA)

Inferring system properties frominter-communication times of

event-triggered control

Description:In networked control systems, measurements and control ac-tions are sent via a communication network between sensor,controller and actuator. Thereby, the controller can be out-sourced to a powerful external cloud computing service. Be-sides all the advantages of this flexible approach, there mightbe privacy concerns. If the measurements, the control actionsor the control policy are confidential, encrypted control can beused to compute control actions in the cloud without revealingany information about the data to the cloud provider.To reduce the amount of communication in the network, event-triggered control schemes were developed. There, measure-ments are only sent to the controller if certain triggeringconditions are met. In combination, encrypted event-triggerednetworked control is communication-efficient and data privacyis maintained. However, the communication events can beobserved.The goal of this thesis is to investigate which system propertiescan be determined if the transmitted data is hidden but thecommunication pattern is known.

Prerequisites:• Course Einfuhrung in die Regelungstechnik• Interest in systems theory, having fun with theoretical

work

Supervisor:Sebastian Schlor,StefanWildhagen

Area:

Networked ControlEncrypted Control

Properties:

Type: BA

40% literature50% implementation10% simulation

Beginning:October 2021



7



Open Thesis (BA)

Controller design for a LEGO-Mindstormsrobot

Description:

The LEGO®-Mindstormsrobots provide an easyentry into control forstudents. Therefore,usually simple controlstructures are implementedin order to stabilize asegway or to follow a line.Now the question risewhether more complexcontrol stuctures, astaught in an initial control lecture as Einfuhrung in dieRegelungstechnik, can be successfully implemented on theLEGO-Mindstorm brick. To this end, Matlab/Simulinkand more reliable sensors than from LEGO® will beconsidered.Subject of this project will be modelling a LEGO-Mindstorms robot, validation of the model and controllerwith Matlab/Simulink, and implementation of the controlleron the LEGO-Mindstorms brick using Matlab/Simulink.

Prerequisites:

• Einfuhrung in die Regelungstechnik lecture or equiva-lent courses

• Experience with Matlab/Simulink is desired

Supervisor:

Tim MartinRoom 2.236

Area:

Implementation of con-trollers for real plants

Properties:

Type: BA

20% literature20% simulation20% modelling40% implementation

Beginning:

any time



8



Open Thesis (BA)

Swing-up control of an inverted pendulum

Description:

w

M2R

ϕ

One of the most popularexamples in controlengineering is an invertedpendulum on a cart.Therein, a rod is attachedto a cart and the taskis to balance the rodin its upright position by moving the cart. In this thesiswe want to design a tracking controller that can changeautomatically between the upright unstable and the hangingstable equilibrium. Hence, the controller should be able toswing up the rod into its upright position. One challenge inthe design is that the control inputs cannot be larger thanthe motor of the cart allows and that the cart itself hasonly a limited range to move. The goal of this thesis is tocompare several approaches, as for example a 2-DOF controlstructure, a model predictive controller, or an iterativelearning controller, first in simulation and then implementedon the real plant. This thesis is also suitable for group work.

Prerequisites:

• Courses Einführung in die Regelungstechnik• Some experience in Matlab/Simulink• Interested in designing and implementing controllers

Supervisor:

LukasSchwenkel,Matthias Hirche

Area:

Implementation

Properties:

Type: BA

30% literature30% simulation40% experiments

Beginning:

now



9



Open Thesis (BA)

Neural network approximation ofbalancing controller for autonomous

e-scooter

Description:

The interest in neural networks(NNs) has recently increased andNNs are applied to a wide rangeof problems. As universal func-tion approximators, they may inparticular be used to approximatecontrollers which are for examplecomputationally expensive. Forinstance, a model predictive con-troller requires to solve an online optimization problem inevery time step, whereas the evaluation of a NN is cheap.At the IST, a balancing controller has been designed thatstabilizes an autonomous e-scooter in the upright position. Inthis thesis, the student is to approximate the existing controllerwith an NN using imitation learning. Following, the student isto analyze the resulting feedback interconnection’s closed-loopstability or even enhance the training with stability constraints.Subsequently, the closed loop is to be simulated in Gazebo andfinally the NN controller to be implemented on the e-scooter.

Prerequisites:• Basic course in machine learning recommended• Knowledge in Python and/or Matlab• Experience with ROS and Gazebo helpful, not required

Supervisor:Patricia Pauli,David MeisterRoom 3.234

Area:

Neural networksBalancing controlAutonomous vehicles

Properties:

Type: BA

25% literature25% theory25% simulation25% experiments

Beginning:anytime



10



Open Thesis (BA)

Comparison of data-driven andidenti�cation-based control

Description:Data

SystemIdentification

ControllerDesign

Data-Driven Control

Most classical approaches to control theory are model-based.If the system is not known a priori, then various methodsexist to identify a model from measured data. Alternatively,controlling systems based directly on measured data withoutan intermediate system identification step is a research topicof increasing interest. In general, connections between di-rect (data-driven) and indirect (identification-based) controlapproaches are largely unexplored.For instance, both direct and indirect methods have beendeveloped for stabilizing unknown systems using data af-fected by bounded noise. In order to guarantee that the un-known plant is indeed stable in closed loop, these approa-ches robustly stabilize all systems which are consistent withthe available data. The particular choice of the noise boundand how it is included in the design condition can have largeinfluence on the success of the design. The goal of this the-sis project is to evaluate and compare different direct andindirect approaches for data-driven controller design in thepresence of bounded noise.

Prerequisites:

• Course Einfuhrung in die Regelungstechnik

• Experience in Matlab

Supervisor:

Julian BerberichRoom 2.236

Area:

Data-Driven ControlIdentification

Properties:

Type: BA

30% literature40% simulation30% experiments

Beginning:

now


Aushang vom 24. Marz 2021

11

BachelorarbeitenimWintersemester2021/22 ...

Documents

Transcript of BachelorarbeitenimWintersemester2021/22 ...