Welcome to the internal ISY pages

B-huset

The Department of Electrical Engineering (Institutionen för systemteknik, ISY) is central in our engineering programmes both for base knowledge and applied courses. The research is primarily focused on industrial needs and reaches from foundational questions to more application-centered issues.

Undergraduate education

The department offers more than 100 different courses within four undergraduate education areas: Images, Electronics, Automatic control, and Telecommunication. Within the university’s programs there are a number of specialisations coordinated by us.

Research

Research and research education is performed within the subject areas: Computer vision and learning systems, Electronics and computer engineering, Vehicular systems, Information coding, Communication systems och Automatic control.

Master’s thesis

Here you can read about how to find and finish a Master’s thesis with us.

Thesis defenses

2026-09-18 kl 10:00 i Systemet

Energy-Efficient Semi-Decentralized Federated Learning via Cooperative D2D Model Aggregation

Författare: Mehran Mamivand
Opponent: Lei Meng
Handledare: Yuhang Deng, Angelo Rodio
Examinator: Zheng Chen
Nivå: Avancerad (30hp)

Federated learning allows multiple devices to train a shared machine-learning model without sharing their raw data. However, frequent communication with a central server can be energy-intensive, especially in wireless edge networks. This thesis studies a semi-decentralized federated learning approach which combines intra-cluster aggregation via device-to-device (D2D) communication and global inter-cluster aggregation assisted by server coordination. The work investigates how local cooperation affects model accuracy, stability, and communication energy under different data-distribution scenarios. The results show that adaptive local coordination can reduce costly global communication while maintaining good learning performance, highlighting the potential of semi-decentralized federated learning for energy-efficient wireless edge intelligent systems.
2026-09-23 kl 15:15 i Systemet

Extension of Two Image Reconstruction Methods from Dual-Energy CT to Spectral CT and Simulation-Based Evaluation of Their Material Decomposition Accuracy

Författare: Yana Dashchechka
Handledare: Alexandr Malusek
Examinator: Maria Magnusson
Nivå: Avancerad (30hp)
2026-09-24 kl 10:00 i Systemet

Temporal Convolutional Network acceleration with FINN

Författare: Xiaoju Wang
Opponent: Edvin Jakobsson
Handledare: Ran Huo
Examinator: Jose Nunez-Yanez
Nivå: Avancerad (30hp)

Temporal convolutional networks (TCNs) are commonly used for time-series classification and typically employ one-dimensional causal convolutions. Deploying a quantized
causal TCN as a FINN-based FPGA dataflow accelerator is nevertheless challenging. The
main difficulties arise from one-dimensional operators, asymmetric causal padding, reshape operations, and the loss of datatype information during quantized graph transformations.
This thesis presents a QONNX export and FINN build flow for mapping quantized
causal TCNs to resource-constrained FPGAs. Temporal convolutions are reformulated
as two-dimensional convolutions over four-dimensional feature maps. The conventional
flatten–linear classification head is replaced by a convolutional classifier that preserves
the feature-map layout throughout the network. In addition, customized FINN transformations are introduced to convert causal padding into hardware-compatible FMPadding
layers, restore integer datatype annotations, and absorb compatible scalar operations into
thresholding layers.
The proposed flow is evaluated on the ECG5000 classification task using a PYNQ-Z2
FPGA. Three quantization configurations, W8A8, W4A4, and W2A4, are examined together with four W4A4 folding configurations and several batch sizes. The W4A4 implementation achieved a classification accuracy of 0.9116. The lowest measured batch-1
latency was 3.881 ms, while the Folding 4 configuration achieved 3.928 ms and provided
higher effective throughput when multiple samples were processed in a batch. These results show that quantized causal TCNs can be deployed within the existing FINN dataflow
framework without introducing a dedicated one-dimensional convolution engine.
2026-09-28 kl 14:15 i Transformen

Decentralized Target Tracking Using Direction-of-arrival Sensors - Analysis of Different State Representations

Författare: William Olsson
Handledare: Louise Lennartsson
Examinator: Gustaf Hendeby
Nivå: Avancerad (30hp)

Direction-of-arrival tracking is critical in many applications, for example, tracking with sensor arrays consisting of radio antennas, microphones, or hydrophones. Tracking a target's position generally requires multiple sensors, and decentralized networks are of particular interest due to their robustness and scalability. However, decentralized trackers rely on single-sensor direction-of-arrival estimates, which pose significant challenges due to nonlinearities and limited observability. To handle this, previous research has proposed alternative state representations to Gaussians in Cartesian coordinates, such as Gaussians in modified polar coordinates or particle filters. This thesis evaluates several such alternatives in a decentralized setting, in particular: trackers based on an extended Kalman filter using Cartesian, polar, and modified polar coordinates, and trackers based on particle filters.

Constructing a decentralized tracker using polar or modified polar coordinates requires modifications to classic fusion methods to handle the non-Euclidean geometry of a polar representation, namely, that bearings are periodic and ranges are positive. Three algorithms to handle this are proposed: Cartesian coordinate-fusion (CC-fusion), difference-fusion, and translation-fusion. CC-fusion relies on transforming the state and performing fusion in Cartesian coordinates, while difference and translation-fusion rely on wrapping the bearings. All methods successfully handle nonlinearities, and it is shown that difference and translation-fusion are equivalent.

To evaluate trackers, simulations were run using three different sensor networks and trajectories randomly sampled from a constant velocity motion model. The trackers were evaluated in scenarios with and without a known prior, and in scenarios with communication failures. The scenarios were designed to evaluate different aspects of tracker behaviour, rather than to replicate real-world conditions.

In these simulations, the evaluated representations offered small, if any, advantages over the standard Cartesian coordinate tracker,
though the similarity may be caused by the evaluated scenarios not being sufficiently challenging. Nonetheless, modified polar coordinates could still be an appropriate option, as they performed better in one of the scenarios with decreased communication while performing nearly identically to Cartesian coordinates in other scenarios. Polar coordinates, however, showed no indication of improved performance, consistent with previous literature where polar coordinates are rarely proposed. Of the proposed fusion methods, CC-fusion appears to be preferable for modified polar-coordinate trackers.

Particle filter-based trackers could match the Cartesian coordinate tracker's performance under suitable parameter choices, but that performance was highly dependent on those choices, and several configurations were prone to divergence. This suggests that any implementation should be preceded by an evaluation of parameters for the specific application. Given this sensitivity and the limited performance benefits over other trackers, together with the added bandwidth and computational requirements, particle filter-based tracking is unlikely to be justified for applications resembling those evaluated here.