Examensarbete vid ISY

Det är både roligt och spännande att utföra exjobb eftersom du då omsätter de kunskaper du tillägnat dig under studietiden. Exjobbet ger dig en möjlighet till inblick i näringslivet och ditt personliga initiativ är viktigt då exjobbet är ett tillfälle att skapa kontakter med presumtiva arbetsgivare och samarbetspartner. Exjobbet mynnar även ut i en offentlig akademisk avhandling.

Anvisningar

För studenter

För handledare och examinatorer

Frågor kan ställas till exjobbshandläggaren exjobb@isy.liu.se. Frågor kring kvalitetssäkring, regelverk, och synpunkter på exjobbsbeskrivningar kan också ställas till exjobbskoordinator@isy.liu.se.

Förslag till examensarbeten

Förslag på examensarbeten finns i exjobbsdatabasen

När du vill vidare diskutera ett exjobbsförslag tar du kontakt med en av examinatorna på ämnesområdet. Om du är osäker på vilket område ditt exjobb hör hemma i eller vill diskutera beskrivningen kan du kontakta någon av studierektorerna eller exjobbskoordinator@isy.liu.se för vägledning.

Framläggningar

2026-05-28 kl 13:15 i Systemet

Foundation Models for Automated Ki67 Cell Detection

Författare: Viktor Axén, Filip Nygren
Opponenter: Mikael Lundgren, Edvard Wetind
Handledare: Anmar Karmush
Examinator: Maria Magnusson
Nivå: Avancerad (30hp)
2026-05-28 kl 13:15 i Zulu

Adaptive Control and Simulation of Hydraulic Flow in a Tiltrotator

Författare: Jonathan Byman, Axel Stockhaus
Examinator: Farnaz Adib Yaghmaie
Nivå: Avancerad (30hp)
2026-05-29 kl 10:15 i Big Conference Room in Visionen, LiU

Path Planning Optimization for Navigation of an Underwater Vehicle Utilizing Bottom Terrain Positioning

Författare: Oscar Jemsson
Opponent: Erik Sjöberg
Handledare: Ashwani Koul
Examinator: Gustaf Hendeby
Nivå: Avancerad (30hp)

Autonomous underwater vehicles cannot rely on global navigation satellite systems (GNSS) for positioning and must use onboard sensors to mitigate the drift accumulated by their inertial navigation systems. Terrain-aided navigation addresses this by matching seabed observations against a bathymetry chart, but the quality of the resulting position fix depends strongly on the local terrain. When an autonomous underwater vehicle (AUV) follows a path through featureless terrain, localization degrades regardless of sensor quality, making path selection a critical factor in navigation performance.

This thesis investigates whether information-theoretic path planning can improve terrain-aided navigation (TAN) localization by guiding the vehicle through terrain that is more informative for position estimation. A Fisher information matrix based cost is derived from the Doppler velocity log (DVL) sensor model and combined with travel distance in an A* path planner. The method is evaluated in a simulated environment using a particle filter with a strapdown inertial navigation system (INS) with DVL aiding for localization, and bathymetry data from Lake Vänern, with problem scenarios based on Saab's AUV62.
Results show that the particle filter-based TAN reduces position error relative to unaided INS. Information-theoretic path planning reduces this error further, often with negligible additional path length. Accounting for terrain informativeness during mission planning is therefore a practical and low-cost strategy for improving navigation performance in rudder-steered AUVs equipped with a DVL.
2026-05-29 kl 15:00 i ISY Systemet, B-huset

High-Speed Frequency-Domain CDC Filters with Reconfigurable Filter Length

Författare: Zhaoqi Li
Opponent: Erik Helmer
Handledare: Mikael Henriksson
Examinator: Oscar Gustafsson
Nivå: Avancerad (30hp)

To meet the requirements for real-time operation and flexibility in chromatic dispersion compensation (CDC) for high-speed coherent optical communication systems, traditional fixed-structure frequency-domain CDC filters suffer from limited adaptability as well as high power and hardware cost. This thesis proposes a high-speed reconfigurable length frequency-domain CDC filter architecture based on a cascaded structure consisting of a fixed 192-point P-FFT and a reconfigurable C-FFT with C = 2/4/8. Combined with
overlap-save streaming processing, the architecture enables real-time chromatic dispersion compensation at a sampling rate of 60 GSa/s with 128-way parallel input. The work
completes the RTL design and implementation of the reconfigurable data-routing unit, reconfigurable FFT, and compensation modules, and evaluates the hardware through synthesis and simulation in a 28-nm CMOS process at 0.70 V and 553 MHz. The results show that the proposed architecture operates correctly, supports dynamic adjustment of
the filter length without interrupting the data stream, and achieves significantly reduced reconfigurability overhead in high-parallelism modes. At C = 8, the additional power
overhead relative to the fixed architecture is only 33.52%, while the proposed design also shows a clear area advantage. The architecture therefore, provides a practical solution
that balances high-speed processing capability, low power consumption, and reconfigurability for ASIC design in high-speed coherent optical communication systems.
2026-06-01 kl 09:00 i Systemt, B-huset

Performance analysis of OFDM based ISAC systems under deceptive jamming attacks

Författare: Oscar Wilkens
Opponent: Niklas Lennarth Greger Eriksson
Handledare: Diana Pamela Moya Osorio
Examinator: Erik G. Larsson
Nivå: Avancerad (30hp)
2026-06-01 kl 13:15 i Visionen

Event-based Early Detection for YOLO Models - Comparison of Event-Handling Methods and Real-Time Use in UAV’s

Författare: Filippa Jernberg, Rebecca Lundgren Nilsson
Opponenter: Lovisa Andersson, Emmy Lindgren
Handledare: Bryan Adams
Examinator: Maria Magnusson
Nivå: Avancerad (30hp)

This thesis investigates whether the low-latency properties of event cameras can enable fast, real-time object detection. The study focuses on YOLO models, known for their rapid inference, and includes an examination of a recurrent YOLO variant. Since YOLO is image-based, various event-to-image methods are explored. Findings indicate that non-recurrent models are faster, making them more suitable for early detection. Effective accumulation methods achieve a mAP@50 of ~0.55. Potential applications include UAVs, so the models are also tested on aerial event data. The findings from these tests reveals the need for UAV-specific training to improve performance. Future work could explore shorter event accumulation intervals to increase fps, analyze trade-offs between model size, version, mAP, and FPS, and optimize the ReYOLOv8n training setup to enhance mAP
2026-06-01 kl 13:15 i ISY Systemet

Pilot-Power Control for Interference Suppression in Reciprocity-Based Systems

Författare: Tim Ljungberg
Handledare: Dexin Kong
Examinator: Erik G. Larsson
Nivå: Avancerad (30hp)
2026-06-01 kl 14:15 i Transformen

Sensor Management for Covert Anti-Drone Tracking

Författare: William Larsson
Opponent: William Olsson
Handledare: Rasmus Uhlin
Examinator: Gustaf Hendeby
Nivå: Avancerad (30hp)

Abstract: The presence of adversarial or rogue drones is becoming increasingly common and problematic in military and civilian scenarios. In the military domain, rogue drones are often used to scout covert positions and destroy expensive radar equipment. Restricted airspace around critical civilian infrastructure, such as airports, may also be violated, resulting in closures and delays. To combat this threat, this thesis, conducted at the Swedish Defence Research Agency (FOI), explores how sensor management can be applied to radars to minimize the risk of exposing a protected covert position and losing expensive equipment. Four different drone flight scenarios were realistically simulated. Simulated detections from two radars and two radio frequency (RF) sensors were fused using an extended Kalman filter (EKF) by global nearest neighbor (GNN) and joint probablistic data association (JPDA) trackers. A novel sensor management solution is suggested, involving splitting a reward function into separate interpretable components. Each component is designed to capture different aspects, such as radar use and track quality. The proposed sensor manager uses breadth-first search (bfs) to maximize the reward function, thereby balancing track quality with radar use. The results show that the proposed method reaches almost as good performance as the non-covert baseline of always using the radars, while on average decreasing active radar use by ∼79–85 %.
2026-06-03 kl 10:15 i Transformen

Fuzzing IPv4 modules on FPGAs: Design and comparison of on-chip and host-based fuzzers

Författare: Alexander Lindskog, Filip Ripstrand
Opponenter: Cornelia Calota, Daniel Söderström
Handledare: Petter Källström
Examinator: Kent Palmkvist
Nivå: Avancerad (30hp)
2026-06-03 kl 13:15 i Stora Visionen

Deep Learning-based Identification of Anatomical Landmarks

Författare: Valdemar Bång, Philip Gustafsson
Opponenter: Simon Hansson, Edvard Nilsson
Handledare: Ioannis Athanasiadis
Examinator: Maria Magnusson
Nivå: Avancerad (30hp)
2026-06-03 kl 13:15 i Transformen (B-hus, A-korridor, ing. 27, övre plan)

Complete Coverage Path Planning for Multiple UAVs

Författare: Adam Mejri, Jacob Persson
Opponenter: David Forslund, Axel Johansson
Handledare: David Axelsson
Examinator: Erik Frisk
Nivå: Avancerad (30hp)
2026-06-04 kl 10:15 i SH63

System-Level Optimisation of Heat and Electricity Utilisation from Gas Turbine Testing Using Thermal Storage

Författare: Sami Chouman, William Elfström Mackintosh
Examinator: Daniel Axehill
Nivå: Avancerad (30hp)
2026-06-04 kl 10:15 i Nollstället

Positive-Unlabeled Graph Learning for Financial Crime Detection - Using Graph Neural Networks to Identify Financial Crime in Sparsely-Labeled, Real-World Data

Författare: Karl Duckert Karlsson, Lukas Olof Ingemarsson
Opponenter: Alice Stattin, Sofie Wiklund
Handledare: Adrian Edin
Examinator: Danyo Danev
Nivå: Avancerad (30hp)
2026-06-04 kl 10:15 i C3

Modelling and Performance Evaluation of Cell-to-Cell Variations in Series-Connected Battery Systems for Automotive Applications

Författare: Simon Gustafsson
Handledare: Arvind Balachandran
Examinator: Lars Eriksson
Nivå: Avancerad (30hp)
2026-06-04 kl 13:15 i Systemet

Reactive Power Control in a Pulp and Paper Industry

Författare: Oscar Ljungberg, Gabriel Lyberg
Handledare: Arezou Safdari-Vaighani
Examinator: Christofer Sundström
Nivå: Avancerad (30hp)
2026-06-04 kl 14:30 i Systemet

Performance Evaluation and Control Strategy Analysis of a Microgrid

Författare: Elisa Rylander, Hannah Schmid
Handledare: Oskar Lind Jonsson
Examinator: Christofer Sundström
Nivå: Avancerad (30hp)
2026-06-04 kl 15:15 i ISY-Visionen

Map based drone positioning

Författare: Mikael Lundgren, Edvard Wetind
Opponenter: Viktor Axén, Filip Nygren
Handledare: Justus Karlsson
Examinator: Amanda Berg
Nivå: Avancerad (30hp)
2026-06-05 kl 09:15 i Visionen Stora Konferensrummet

Directional Visual Perception for Object Sonification in Indoor Environments

Författare: Edvard Nilsson
Opponent: Valdemar Bång
Handledare: Arvind Balachandran, Lars Nielsen
Examinator: Per-Erik Forssén
Nivå: Avancerad (30hp)
2026-06-05 kl 10:15 i Transformen

Visualization of multiple duel simulations

Författare: Viktor Thellgren
Handledare: Souad Mohaoui
Examinator: Ingemar Ragnemalm
Nivå: Avancerad (30hp)
2026-06-05 kl 13:15 i Transformen

Computationally Efficient Angle-of-Arrival Estimation on FPGA

Författare: Daniel Söderström
Handledare: Theodor Lindberg
Examinator: Oscar Gustafsson
Nivå: Avancerad (30hp)

The ability to track mobile devices and vehicles using signal direction estimation is a critical component for modern communications and signal intelligence purposes. This thesis investigates the implementation and optimisation of angle-of-arrival (AoA) estimation algorithms on a Field-Programmable Gate Array (FPGA). The theory focuses on three algorithms: Bartlett's method, Capon's method and MUSIC, utilising a uniform circular array (UCA) antenna geometry. The objective is to find a balance between mathematical complexity, execution time and algorithm precision to meet real-time requirements.

To achieve this, an architecture was developed using systolic array-based processing elements. The main contribution to the implementation is an optimised grid-search approach for estimating the angle-of-arrival utilising adaptive angle spaces to reduce the amount of iterations required without sacrificing essential resolution. The implementation was verified using Matlab and Python simulations, as well as with synthesis using AMD Vivado. Results show that Bartlett's method provides the balance between execution time and AoA accuracy required. Furthermore, a Pareto front analysis was utilised to determine the optimal parameter configurations, showing that wordlength and grid search parallelisation impacts the trade-off between execution time and resource usage. This thesis thus provides a robust proof-of-concept for efficient angle-of-arrival estimation in different communications sectors.
2026-06-05 kl 15:15 i Transformen

Re-Identifying User Equipment in LTE-Networks Using Machine Learning

Författare: Erik Karlstedt
Opponent: Martin Castro Bildhjerd
Handledare: Martin Dahl
Examinator: Danyo Danev
Nivå: Avancerad (30hp)

This thesis investigates the re-identification of user equipment (UE) in LTE networks following
temporary identity changes, utilizing machine learning to analyse unencrypted
RNTI and TMSI metadata. Data was collected by connecting a target UE operated by the
author to a real base station. Unencrypted metadata from the base station was collected
using a software defined radio and data from the special UE was used to label samples as
positive if they belonged to the target UE and negative if they did not. Importantly, real
users were connected to the base station during collection but no user traffic was collected.
Further, unencrypted metadata related to other UE were considered negative samples and
no UE other than the target UE was re-identified. This ensured the privacy and integrity
of real users connected to the base station. Two machine learning models – a Long Short
Term Memory (LSTM) model and a Siamese neural network model – were employed and
compared for the task. The results showed that the LSTM model performed well on RNTI
data, managing to detect over 90% of all samples belonging to the target UE with a precision
of around 20%. The models based on Siamese neural networks performed much
worse. On RNTI data, around 70% of all positive samples were detected, but the precision
was only 0.001%. On TMSI data the Siamese models performed better, detecting around
90% of all positive samples with a precision of 0.0045%. Not enough TMSI data was collected
for training LSTM models. The thesis concludes that it may be possible to re-identify
UE in LTE networks after a change of temporary identity by analysing unencrypted metadata
in RNTI traffic with the use of machine learning, but that more research is needed to
definitively answer.
2026-06-05 kl 15:15 i Nollstället

Enhancing Subsea Inertial Navigation: Underwater current estimation

Författare: Jonathan Norrestam
Opponent: Oscar Johansson
Handledare: Joel Wendin
Examinator: Gustaf Hendeby
Nivå: Avancerad (30hp)

The absence of GPS/GNSS in the underwater domain makes accurate long-term navigation challenging. Inertial navigation systems (INS) are commonly used, but unaided, they suffer from inherent drift caused by sensor noise and bias, which is further exacerbated by unknown water currents. The Doppler velocity log (DVL) sensor can provide accurate velocity aiding if the seabed is within range, effectively reducing the amount of drift. However, its range is limited, and rough bathymetry can lead to measurement dropouts. In these scenarios, the DVL can measure the vehicle's velocity relative to the surrounding water using its water track (WT) mode. This thesis investigates the feasibility of utilizing these WT measurements for INS aiding.

The study addresses the problem by implementing current estimation capabilities driven by the DVL WT measurements. Different filtering techniques based on the error state Kalman filter (ESKF) architecture are used to model the unknown dynamics of the current. An extended version of the baseline ESKF with a fixed current dynamics model is first established. To address the challenge of correctly tuning this fixed model to unknown dynamics, two different interacting multiple model (IMM) approaches are then investigated: one using 2 models, and another using 4 models that decouple the current dynamics in the global navigation axes.

The implemented filters are evaluated in simulation using an autonomous underwater vehicle and compared to a baseline unaided INS under varying current conditions, sensor qualities, and movement patterns. The results demonstrate that current estimation has the potential to significantly reduce the long-term position drift of an unaided INS. While the fixed ESKF can perform well if tuned correctly, the 2-model IMM demonstrates superior adaptability to changing conditions and proves to be the most robust solution across all evaluated scenarios. This work concludes that DVL WT-aided current estimation, particularly using IMM techniques, can contribute to a more robust navigation solution for underwater missions.
2026-06-08 kl 13:15 i Stora Visionen

Mer än en nödlösning: En teknisk och ekonomisk studie av gasturbinbaserad reservkraft

Författare: Ida Edin, Karl Johansson
Handledare: Carl Steen
Examinator: Christofer Sundström
Nivå: Avancerad (30hp)
2026-06-08 kl 15:15 i ISY Visionen Stora konferensrummet

Semantic Segmentation of LiDAR Point Clouds Using Image Annotations

Författare: Johanna Nilsson
Opponent: Alexander Berntsson
Handledare: Anmar Karmush
Examinator: Yonghao Xu
Nivå: Avancerad (30hp)
2026-06-09 kl 10:00 i Stora Visionen

Semi-Autonomous Aircraft Tug Docking Function

Författare: Oscar Johansson
Opponent: Jonathan Norrestam
Handledare: David Axelsson
Examinator: Erik Frisk
Nivå: Avancerad (30hp)
2026-06-11 kl 13:15 i Zulu

Sim-to-Real Evaluation of Pallet Slot Corner Detection and Tracking Using Deep Neural Networks

Författare: Ferdinand Kouhia, Adil Shamji
Opponenter: Tobias Berglind, Oscar Sandblom
Handledare: Bryan Adams
Examinator: Mårten Wadenbäck
Nivå: Avancerad (30hp)