The topics below are open for the undergraduate graduation design project (KON 4901E / KON 4902E) and for MSc and PhD work. The topic codes are the same as in the course documents. Please contact me for further detail and for the current availability of a topic.

Undergraduate graduation design project topics (KON 4901E / KON 4902E)

A. Railway signalling, interlocking and safety

• L1 — Web based railway traffic control centre and interlocking
  A web based traffic control centre interface (HTML/CSS/JS, React, PHP) will be developed for a railway line. The interface will connect to a HIMA safety PLC over Modbus to provide point, signal and route control together with status monitoring. The interlocking logic resides on the PLC, while visualisation and commands are handled on the web side.
  Environment: React/JavaScript, PHP, Modbus TCP, HIMA PLC
• L2 — Interlocking design with Petri nets and implementation on a PLC
  A Petri net based interlocking model will be built for a given station or line. The model will be examined for deadlock freeness and basic safety properties, and then transferred to a PLC within a fail-safe programming discipline.
  Environment: Petri net tools (CPN Tools, TINA or PIPE), PLC
• L3 — Interlocking design with finite automata and functional safety analysis
  The same interlocking problem will be addressed through a finite automaton approach. In addition, a basic hazard analysis (FMEA or STPA) will be carried out within the framework of EN 50126/50129, and the results will be related to the notion of safety integrity level (SIL).
  Environment: libFAUDES (or MDESops), PLC
• L4 — OpenETCS based train driver interface (DMI) design and driving simulation
  An ERTMS/ETCS driver machine interface (speed indicator, planning area, permitted speed curve) will be implemented for a given line on the basis of openETCS specifications. The interface will then be connected to a simple train dynamics model so that driving along the line can be simulated.
  Environment: Python/Qt or C#/WPF; openETCS documents
• L5 — SCADE – HIMA integration for an interlocking test tool
  A desktop (C#) and/or mobile (Android) application developed with SCADE will connect to a HIMA PLC over Modbus and provide point, signal and route control and monitoring. The aim is to obtain a test tool that exercises interlocking software systematically.
  Environment: SCADE Suite, C# or Android, Modbus
• L6 — Comparative simulation of fixed block, moving block and virtual coupling operation
  Three signalling principles will be modelled separately for the same line and the same vehicle data. Minimum train headway and line capacity will first be computed and compared under the absolute (brick-wall) braking assumption. A sensitivity study will then show how much the result changes under the relative braking assumption.
  Environment: MATLAB/Simulink or Python
• L7 — Control of a train convoy under virtual coupling
  Gap and speed tracking control will be designed for a convoy of trains running one behind the other. Convoy stability and the string stability property will be examined under different communication topologies, namely predecessor following, leader following and bidirectional.
  Environment: MATLAB/Simulink
• L8 — Fail-safe design and verification of a level crossing control system
  The closing and opening logic of a level crossing will be modelled. A fail-safe design will be produced taking train detection delay and component failures into account, and the required safety properties will be established by model checking.
  Environment: UPPAAL or NuSMV; PLC

B. Traffic, intersections and intelligent transportation

• L9 — Traffic signal control at a single intersection with deep reinforcement learning
  A single intersection will be modelled in SUMO. Queue length, waiting time and vehicle count will be taken as the state, and phase selection will be learned by a DQN style agent. Results will be compared against fixed time and actuated classical control under the same scenario.
  Environment: SUMO and TraCI, Python
• L10 — Coordinated traffic signal control with multi-agent reinforcement learning
  Each intersection in a road network with several junctions will be modelled as a separate agent. Agents will decide in a coordinated manner using traffic density, queue length and information arriving from neighbouring intersections. Depending on scope, the effect of rain, snow or low visibility conditions on performance may also be studied.
  Environment: SUMO, Python, multi-agent DQN
• L11 — Interpretable traffic signal control, comparing fuzzy logic and PI control with learning
  Fuzzy logic and PI based adaptive signalling will be designed for the same intersection and compared against reinforcement learning based control. The emphasis lies not only on performance but on whether a traffic engineer can inspect the resulting rule, that is on interpretability.
  Environment: SUMO; MATLAB Fuzzy Toolbox or Python
• L12 — Reservation based intersection management without traffic lights
  The traffic light will be removed entirely and the intersection will be managed by an agent. Vehicles approaching the intersection will reserve its zones for given time intervals. A conflict free scheduling algorithm will be developed and compared against classical signal control.
  Environment: SUMO or a purpose built simulation
• L13 — Development of an intersection crossing decision mechanism for autonomous vehicles
  An autonomous vehicle approaching an intersection will decide to proceed, slow down, stop or wait by considering traffic density, the positions and speeds of other vehicles, the signal state and possible conflicts. Either a learning based or a reservation based approach may be chosen.
  Environment: SUMO or CARLA
• L14 — Reinforcement learning based decision making system for an autonomous vehicle
  Lane keeping, car following, speed adjustment, safe distance keeping and reaction to sudden events will be developed by learning for an autonomous vehicle in the CARLA environment, and tested under different traffic densities and road conditions.
  Environment: CARLA, Python
• L15 — Smart charging station selection system for electric vehicles
  Vehicles will be directed to the most suitable charging station by considering battery level and position, traffic conditions, and the occupancy and waiting times of the stations. Deep reinforcement learning (DQN, PDQN or similar) will be used.
  Environment: Python; optionally SUMO
• L16 — Area wide traffic control on a real map with agent based route selection
  A real area taken from OpenStreetMap will be imported into SUMO. Both the control of the signals in that area and the selection of routes by vehicles following agent recommendations will be treated together. How route guidance affects the performance of signal control will be examined.
  Environment: SUMO, OpenStreetMap, Python
• L17 — Timetabling and real time regulation of a public transport fleet
  Departure times for an urban bus line will be optimised under passenger demand and travel time variability. In addition, a real time regulation strategy based on holding or speed adjustment will be developed to prevent bus bunching.
  Environment: Python or MATLAB

C. Controller design and laboratory applications

• L18 — PID-PIR type controller design and experimental application on a magnetic levitation rig
  PID and PIR (proportional integral resonant) controller structures will be designed for the magnetic levitation and/or ball and beam rig. Parameter regions guaranteeing stability under parametric uncertainty will be determined and the designs will be tested on the rig.
  Environment: Control Laboratory rig; MATLAB/Simulink
• L19 — Control of classical laboratory rigs by different methods (ball and beam, inverted pendulum, tension control)
  The selected rig will be controlled by different controller design methods such as PID control, pole placement, variable structure control and fuzzy control. The methods will be compared experimentally under the same performance criteria.
  Environment: Control Laboratory; MATLAB/Simulink
• L20 — Rebuilding the quadruple-tank rig and controlling it by multivariable pole placement
  The project is deliberately split across the two semesters. In the first semester (KON 4901E) the rig will be repaired, or rebuilt from scratch if necessary: pump and valve selection, level measurement, data acquisition hardware and computer connection, a safe enclosure, and calibration. In the second semester (KON 4902E) pole placement methods for multi-input multi-output systems will be studied, with particular attention to pole placement by dynamic feedback in two-input two-output systems. The required software will be developed and applied to the rig. The non-minimum phase operating point of the rig will also be addressed.
  Environment: Control Laboratory; MATLAB or Mathematica; data acquisition card, workshop facilities
• L21 — Determination of stability regions for fractional order PIλDμ controllers
  Fractional order systems and this class of controllers will be studied. Software computing the set of controller parameters that guarantees stability will be developed, controllers giving a critically damped response will be sought, and the results will be tested on the thermal process rig.
  Environment: MATLAB or Mathematica; thermal process rig
• L22 — n-dimensional gridding in controller design with parallel computing
  When more than two PID or PI-PD parameters have to be swept, gridding ceases to be practical. In this study the sweep will be moved to a GPU or to multi-core parallel computation and the resulting speed-up will be measured.
  Environment: MATLAB Parallel Toolbox; Python or CUDA
• L23 — FPGA based PID controller design
  A PID controller will be implemented in fixed point arithmetic on a Xilinx FPGA board. The effects of sampling period, word length and quantisation will be examined, and the design will be tested on a rig in the departmental laboratory.
  Environment: AMD Vivado (formerly Xilinx Vivado), VHDL or Verilog; laboratory rig
• L24 — Comparison of data-driven (model free) control methods on a laboratory rig
  Methods that obtain a controller directly from input-output data without building a system model (VRFT, model free adaptive control, DeePC) will be studied and applied to the same rig. They will be compared against classical model based design in terms of the amount of data required, sensitivity to noise and performance.
  Environment: MATLAB or Python; laboratory rig

D. Vehicle dynamics and driving

• L25 — PID based controller design for an active suspension system
  State feedback controllers are commonly used in active suspension. In this study different PID structures (inner-outer loop, PI-PD, acceleration feedback) will be applied to an active suspension system and compared both theoretically and experimentally.
  Environment: MATLAB/Simulink; quarter-car rig
• L26 — Sky-hook and derived control methods in semi-active suspension
  Sky-hook, ground-hook and hybrid strategies will be applied to a semi-active suspension model. The trade-off between ride comfort and road holding will be examined under different road inputs.
  Environment: MATLAB/Simulink
• L27 — Modelling human driving of railway vehicles
  A model imitating human driving behaviour will be built from speed and position data collected from real driving by train drivers. Work will also be done on making the speed and position information collected in the field consistent with each other and with the acceleration and comfort rate constraints. Artificial intelligence and learning techniques may optionally be used.
  Environment: Python or MATLAB; field data
• L28 — Driving in compliance with speed limits, design of a predictive deceleration structure
  Where the speed limit drops, deceleration has to begin before the restricted zone in order to avoid a violation. In this study a Smith predictor like structure will be built taking dead time and vehicle dynamics into account, and tested for both a railway and a road vehicle.
  Environment: MATLAB/Simulink

MSc topics

• Y1 — Constraining learning based controller tuning by the stability region
  Open problem: PID controllers are tuned by reinforcement learning, yet stability is not proved. This gap is the first open problem on the robust control map.
  The set of controller parameters that guarantees stability and the desired performance will first be determined by parameter space and D-stability methods. The search space of the learning agent will then be restricted to that set, so that the stability guarantee is preserved without sacrificing the flexibility of learning.
  Environment: Parameter space, D-stability, reinforcement learning; MATLAB and Python
• Y2 — Analytical design of fractional order PID controllers
  Open problem: The field has largely surrendered to heuristic optimisation; analytical and parameter space based approaches are rare.
  Work will focus on deriving stability boundaries analytically for fractional order PID, on dominant pole placement, and on computing the parameter region that ensures stability.
  Environment: Fractional order systems, parameter space method; MATLAB or Mathematica
• Y3 — Guarantees under noise in data-driven control, with a laboratory application
  Open problem: The first question of the field: guarantees derived from noisy data are overly conservative.
  The guarantees provided by DeePC and behavioural theory based methods under noise will be examined, arrangements reducing conservatism will be tried, and the results will be tested on a real experimental rig.
  Environment: Behavioural theory, DeePC, model free adaptive control; MATLAB or Python
• Y4 — Assessing the capacity gain of virtual coupling together with a safety argument
  Open problem: The literature demonstrates the capacity gain but does not construct the safety argument; the relative braking assumption is accepted without question in most studies.
  The capacity gain obtained under the relative braking assumption will be recomputed taking communication delay and loss, and the difference in braking performance between the leading and the following vehicle, into account. The results will be tied to a risk argument within the EN 50126 framework.
  Environment: Railway simulation, RAMS framework; MATLAB or Python
• Y5 — Scale in interlocking verification, compositional model checking
  Open problem: Claims of large scale verification are mostly supported by academic examples; end to end demonstration on a real station is rare.
  A real station interlocking will be decomposed into modules. The modules will be verified separately and the safety property of the whole system will be derived through composition rules. The scaling behaviour of the method will be measured.
  Environment: Model checking (NuSMV, UPPAAL), compositional verification
• Y6 — Building and verifying Petri net models with large language models
  Open problem: Obtaining a Petri net model is still a manual task. Large language models are addressing this gap, but no verified solution exists yet. The map marks this as the new theme of 2026.
  A Petri net model will be generated from a set of operating rules given in natural language. The correctness of the generated model will be tested by formal verification and a generate, check and repair loop will be established.
  Environment: Large language models, Petri net tools, model checking
• Y7 — Quantification of STPA and its application to a railway subsystem
  Open problem: There is no standard path from qualitative hazard analysis to a numerical risk measure.
  Unsafe control actions obtained through STPA will be carried into a probabilistic framework, for example a Bayesian network, and a numerical risk estimate will be produced for a railway subsystem.
  Environment: STPA, Bayesian networks, RAMS
• Y8 — Learning based model predictive control, preserving guarantees in energy efficient train driving
  Open problem: Preserving the MPC guarantee under a learned model is the first question of the MPC field. The branch is comparatively small and competition is limited.
  An MPC scheme will be set up for energy efficient train driving. While the model is updated through learning, stability and feasibility will be preserved through the design of the terminal set and the constraints.
  Environment: MPC, learning based modelling; MATLAB or Python
• Y9 — Intersection management in mixed traffic and modelling of human driver behaviour
  Open problem: The case where connected and human driven vehicles coexist is one of the liveliest headings of the traffic branch; the transfer from simulation to the field is also an open issue.
  Human behaviour will be modelled in scenarios where autonomous vehicles operate alongside non-autonomous ones, and this model will be incorporated into the intersection management algorithm. The effect of the autonomous vehicle penetration rate on performance will be examined.
  Environment: SUMO or CARLA, behaviour modelling, reinforcement learning
• Y10 — Surrogate models and error bounds in pantograph-catenary interaction
  Open problem: Surrogate models demonstrate speed-up but do not provide an error bound.
  A surrogate model will be built for pantograph-catenary interaction. The main contribution will be the derivation of an error bound for the result the model delivers, and the experimental testing of that bound.
  Environment: Surrogate modelling, finite elements, statistical learning
• Y11 — Integrated line wide mitigation of rail potential and stray current
  Open problem: Local remedies exist, but there is no integrated design along the whole line.
  Feeding point placement, the earthing arrangement and the operating scenario will be treated together in order to minimise rail potential and stray current across the whole line.
  Environment: Traction power simulation, optimisation
• Y12 — Interpretable traffic signal control, combining a fuzzy rule base with learning
  Open problem: The traffic engineer cannot inspect the learned policy; the barrier to acceptance is institutional rather than technical.
  Either by converting the learned policy into a fuzzy rule base or by tuning the rule base through learning, both performance and inspectability will be achieved.
  Environment: Fuzzy logic, reinforcement learning, SUMO

PhD topics

• D1 — A general theory of dominant pole placement for systems with delay
  Open problem: The scalar case with two delays has been solved; the general case remains open.
  The aim is to derive necessary and sufficient conditions for assigning a dominant pole pair in time delay systems in the general case, to secure the dominance of the assigned poles, and to turn the result into a computable design procedure.
• D2 — A conservatism measure and benchmark set for robust control
  Open problem: The claim of a less conservative condition appears in almost every paper, yet there is no standard of comparison and every author picks their own example. The same deficiency recurs together in robust control, in robust MPC and in the Petri net and learning branch.
  Using the fact that the parameter space method makes conservatism measurable, a measure giving the relative position of conditions will be proposed and published together with a benchmark problem set.
• D3 — A bridge from data uncertainty to parametric uncertainty
  Open problem: A non-conservative guarantee under noise is the first question of data-driven control.
  The Kharitonov and D-stability machinery will be carried over to data-driven control: the model set obtained from noisy data will be converted into an uncertainty structure in parameter space and handled with existing robustness tools.
• D4 — Allocation of a safety integrity level to a learning component
  Open problem: The first open problem of the functional safety field: there is no accepted route for assigning a SIL to a learning component.
  An allocation framework will be constructed leading from the performance distribution of a learning component to a safety integrity target. The framework will be instantiated on a railway function and supported by an inspectable safety argument.
• D5 — Placing the relative braking assumption of virtual coupling on a normative footing
  Open problem: Safety oriented studies appeared in 2026, but none ties the relative braking assumption to a normative requirement in the sense of EN 50126/50129.
  The technical conditions required for the relative braking assumption to hold, namely equality of braking performance, communication integrity and position uncertainty, will be derived through hazard analysis and converted into measurable requirements.
• D6 — From a verified model to certifiable code
  Open problem: EN 50716 extended formal proof up to SIL 1-2, but the tool chain is not ready; turning a synthesised supervisor into certifiable software is still done by hand.
  A tool chain and a proof structure will be developed for converting a supervisor produced from discrete event synthesis or from a Petri net model into certifiable code while preserving traceability.
• D7 — Discrete event control that is timed, probabilistic and under attack
  Open problem: Work addressing these three at once is almost non-existent. A unified framework for safety and security is likewise missing on the railway side.
  Timing constraints, probabilistic behaviour and the presence of an attacker will be treated in a common model, and supervisor synthesis and achievability conditions will be examined under this model.
• D8 — Preserving the model predictive control guarantee under a learned model
  Open problem: The first question of the MPC field. The branch is comparatively small and competition is limited.
  Joint design of the terminal set, constraint tightening and feasibility preservation while the model is updated through learning, together with measurement of the conservatism of the resulting conditions.
• D9 — Loss of validity over time in stochastic Petri net and digital twin models
  Open problem: The problem that digital twin models lose their validity over time was posed in 2026 and is not solved.
  A procedure will be developed for measuring the loss of validity as the model drifts away from field data, for detecting that drift early, and for recalibrating the model safely.