[diss] Sähkötekniikan korkeakoulu / ELEC

Permanent URI for this collectionhttps://aaltodoc.aalto.fi/handle/123456789/53

Browse

Recent Submissions

Now showing 1 - 20 of 1103
  • Towards cloud-based virtual commissioning of distributed automation systems
    (2025) Lyu, Tuojian
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-31
    Virtual Commissioning (VC) of distributed automation systems faces critical challenges in manual configuration complexity, performance prediction accuracy, and validation scalability. To address these limitations, this thesis develops CloViC (Cloud-based VC), an integrated framework that unifies advanced performance analysis, cloud computing, and AI-enhanced validation into a single web platform. Using containerization and orchestration technologies, the approach dynamically deploys soft PLC instances, OPC UA servers, and supporting services in cloud environments to minimize manual configuration and enable analysis of system performance. A key innovation is the automated benchmarking process that extracts comprehensive performance knowledge bases from systematic measurements across heterogeneous platforms. By measuring control performance metrics such as execution time, CPU load, and jitter with varying computational complexities, this approach enables accurate performance analysis during VC rather than solely focusing on functional requirement validation. The thesis introduces LLM4VC framework, an AI-assisted testing analyzer as one of services in CloViC, that transforms raw runtime logs into standardized, machine-readable formats through customized Python scripts. Enhanced with context-aware insights from LLM, this analyzer automatically processes distributed system logs against Machine-Readable Expected Behavior generated from engineering specifications. This AI-enhanced approach enables automated anomaly detection, fault diagnosis, and performance assessment without manual intervention.CloViC provides scalable backend services including automatic I/O simulation, real-time performance monitoring, and intelligent system configuration management. The platform is deployed on Azure Kubernetes Service as microservices, enabling efficient and flexible VC experiences accessible through web interfaces. Experimental validation through representative case studies that are Hot Water Tank control and EnAS Test Control Application demonstrates significant improvements in VC efficiency and accuracy. Automated benchmarking reveals performance variations in six heterogeneous platforms, while AI-assisted analysis achieves reliable fault detection across multiple failure scenarios. CloViC accelerates development cycles by providing immediate feedback on performance metrics and supports data-driven distribution strategies through continuous monitoring, establishing a new foundation for agile, adaptive distributed automation system development in Industry 4.0 environments.
  • Sensitivity of high-fidelity neural interfaces to perturbations
    (2025) Taleshi, Mansour
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-31
    High-density surface electromyography (HD-sEMG) coupled with motor-unit decomposition (MUD) can decode spinal-level neural commands for intuitive human-machine interfaces, yet its fidelity deteriorates under sweat, movement, or sensor drift. This dissertation assesses the robustness of such interfaces and neural drive under physiological and technical perturbations and explores strategies to mitigate performance degradation. Physiological perturbations involved acute blood flow restriction (BFR), where its effects on MU discharge, neural drive, force tracking, and common synaptic were assessed. Technical perturbations included simulated additive white Gaussian noise (WGN), channel loss, and electrode shifts on the recorded clean. Decoding performance was evaluated using global EMG features and MU-level features with pattern recognition (linear discriminant analysis (LDA), deep neural network (DNN)) and regression methods (logistic regression (LR), artificial neural network (ANN), DNN). An algorithmic mitigation strategy, musclesynergy-guided channel clustering, was also explored. Key findings show that BFR alters MU firing, neural drive, and coherence (alpha band decreased, delta increased) while subjects largely maintained force tracking performance. For signal degradation, amplitude-based global EMG features (root mean square (RMS) and mean absolute value (MAV)) were most robust to WGN and channel loss, but electrode displacement caused the most significant performance drop for featurebased decoding. MUD algorithm was highly sensitive to sever global WGN (81% yield reduction) but resilient to localized perturbations. However, MU-driven decoding was generally more sensitive to perturbations than feature-based decoding. Synergy-guided clustering substantially increased extracted MU yield (69%) and improved kinematic decoding accuracy. In conclusion, the thesis assesses the sensitivity of HD-sEMG interfaces and highlights the importance of signal quality and robust features. It demonstrates that physiology-informed signal processing, such as synergy clustering, can enhance MU extraction and decoding reliability and enable the development of more robust neural interfaces for real-world use.
  • Computational representations for user interfaces
    (2025) Jiang, Yue
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-17
    Traditional graphical user interfaces (GUIs) often follow a “one-size-fits-all” design, failing to accommodate the diverse needs and contexts of all their users. What if interfaces could instead dynamically understand and adapt to individual users, enhancing their capabilities across a wide range of tasks and contexts? As the diversity of user needs expands, the challenge of designing GUIs that accommodate varying contexts becomes increasingly complex. Data-driven AI methods may offer a way to design GUIs that align with users’ goals. Current AI methods, however, often fall short of capturing the full complexity of human needs, particularly when considering domain-specific knowledge and user-specific requirements. This dissertation contributes to the development of human-centered neural representations for interactions that combine domain knowledge and data-driven learning for GUIs. This dissertation centers on the development of intelligent GUIs in two primary areas. First, we focus on creating computational representations that capture the essential properties of UI design. Specifically, these representations integrate domain-specific knowledge into AI models, allowing for design expert guidance while ensuring that users retain control over their interactions. Moreover, we develop AI models that simulate and predict human behaviors to facilitate automatic personalized adaptation. These models encompass various human behaviors, including eye movements and user interactions. Simulating these behaviors enables personalized optimization and enhances the AI to adaptively respond to user needs.
  • Deep learning-based metal and scatter artifact reduction in conebeam computed tomography
    (2025) Agrawal, Harshit
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-24
    Cone-Beam Computed Tomography (CBCT) provides high-quality three-dimensional X-ray imaging and offers advantages such as reduced radiation dose, lower cost, and a smaller physical footprint compared to Multi-Detector CT (MDCT). Due to these features, CBCT is well-suited for a range of clinical applications, including dentistry, orthopedics, interventional radiology, and image-guided therapies, as well as for use in mobile clinics and remote deployments, thereby contributing to broader accessibility in healthcare. However, CBCT image quality is often compromised by inherent artifacts, including scatter and metal artifacts, which pose significant challenges to diagnostic applications. The emergence of deep learning methods presents a promising avenue for addressing these imaging challenges, potentially offering substantial improvements. However, practical constraints, such as the need for large training datasets, and the seamless integration of deep learning models into existing artifact correction pipelines, must be addressed to ensure clinical feasibility. The main purpose of this thesis is to develop clinically applicable deep learning techniques to mitigate metal and scatter artifacts in CBCT imaging. For instance, to overcome the challenge of data scarcity, simulated datasets are leveraged for network training. Additionally, lightweight Convolutional Neural Network (CNN) models are introduced to facilitate efficient integration into established artifact correction workflows. To ensure real-world applicability, the proposed methods are evaluated using real CBCT datasets. The research is structured around four key contributions. Publication I introduces a learning-based inpainting method of metal traces to reduce metal artifacts. Publication II employs simulated data to train a neural network for metal trace segmentation to improve the effectiveness of existing inpainting-based metal artifact reduction methods. Publication III presents a neural network for scatter estimation under clinically relevant variations in the Field of Measurement (FOM). Finally, in Publication IV, an ultrafast scatter estimation approach is proposed for deployment in mobile CBCT systems and on-device applications. The findings demonstrate that the developed models substantially enhance the state-of-the-art in artifact correction, advancing the clinical viability of CBCT imaging through deep learning-driven lightweight solutions.
  • Enhanced photodiode performance via surface nanoengineering
    (2025) Setälä, Olli
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-24
    Modern silicon photodiodes suffer from non-ideal quantum efficiency, which means that they are not able to detect every individual photon that hits the detector surface. The inefficiency typically emerges from optical losses, but also electrical losses, i.e., charge carrier recombination, may sometimes have an impact. This dissertation studies the potential of employing recent surface nanoengineering developments to improve the quantum efficiency across a range of detection applications. More specifically, surface nanostructuring, or so-called black silicon (b-Si), can be utilized to eliminate the surface reflectance, whereas atomic layer deposited (ALD) thin films are known provide excellent surface passivation and enable induced-electric-field-based charge collection as an alternative to conventional externally doped junctions. In this thesis, we apply these methods to design, fabricate and characterize a collection of Si detectors that achieve near ideal quantum efficiencies in a variety of applications. First, we investigate the possibility to fabricate an efficient b-Si photodiode with an externally doped front junction. Such a device is realized utilizing optimized implanted junction and plasma-etched b-Si. Second, our aim is to make b-Si photodiodes with more scalable and cost-efficient b-Si fabrication method while retaining the efficient electrical performance. Metal-assisted chemical etched b-Si combined with an ALD Al2O3-induced electric-field-based charge collection is identified as a suitable surface design and integrated successfully into a Si photodiode. Third, the advantages of the surface nanoengineered surfaces are transferred to multi-pixel devices by employing a b-Si surface in an image sensor. Fourth, the induced-electric-field-based charge collection is recognized as an interesting option for eliminating the junction-originated dead layer in Si charged particle detectors. Such a device utilizing an ALD Al2O3-induced electric field is fabricated and characterized using alpha particle radiation. We find out that applying the novel surface designs to otherwise conventional Si detectors provides major improvements, especially in terms of quantum efficiency, in all fabricated detector types. The quantum efficiencies reach near 100% over a wide wavelength range (200–1000 nm), including even parts of the conventionally difficult UV region. Importantly, the proposed surface designs do not introduce major drawbacks in other significant detector parameters. The surface nanoengineering steps are also easy to integrate into conventional processing lines providing a simple route for their direct application to commercial devices.
  • Integrated mmWave transceivers and circulators
    (2025) Naghavi, Saeed
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-16
    Solutions for delivering the growing wireless capacity demands of future fifth-generation (5G) and sixth-generation (6G) networks typically involve expanding available bandwidth and improving spectral efficiency. In this context, millimeter-wave (mmWave) communication has attracted significant attention as a promising technology to meet the high data rate requirements of these networks by offering multi-GHz channel bandwidths. The first prototype complementary metal-oxide semiconductor (CMOS) integrated circuit (IC) developed as part of this thesis explores mmWave transceiver IC design techniques through the design and implementation of a hybrid heterodyne/homodyne receiver (RX) and a two-step transmitter (TX) featuring a wide modulation bandwidth. At the same time, full-duplex (FD) wireless is another emerging paradigm that can theoretically double the spectral efficiency by supporting simultaneous transmission and reception of radio signals at the same frequency. Equipping mmWave systems with FD capability will further improve the network’s spectral efficiency, allowing them to address the growing capacity needs more reliably. However, the main challenge in FD mmWave systems is the self-interference (SI) signal that leaks from a radio’s transmitter into its receiver. As a result, effective multi-stage self-interference cancellation (SIC) is required across the entire signal path, including the antenna (ANT) interface, radio frequency (RF)/analog front end, and digital domain. This thesis primarily focuses on the antenna interface, highlighting the role of integrated circulators in FD mmWave transceivers. The second and third prototype CMOS ICs reflect the innovations and contributions of this thesis toward advancing the state of the art in the field of integrated circulators across the following key areas: 1) compact die area and mmWave operation, 2) high power handling, 3) new theoretical framework for integrated circulators with unequal port impedances, and 4) single-ended implementation of switched all-pass network (APN) circulators. Additionally, the research work in this thesis was conducted in close collaboration with the antenna group, leading to the introduction of the concept of a radiating circulator for the first time. In the proposed radiating circulator structure, the reciprocal phase shift cells employed in conventional structures are eliminated, and the required functionality is achieved by co-designing the antenna with the non-reciprocal branch on the IC side. As a result, this approach offers improved efficiency and a more compact die area compared to conventional, separately designed circulators. To demonstrate the feasibility of the proposed radiating circulator in a practical setting, it has been integrated into a more comprehensive system, incorporating a low-noise amplifier (LNA) and a power amplifier (PA) to form an FD transceiver.
  • Detecting changes in distributions in large-scale streaming data
    (2025) Halme, Topi
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-17
    As modern technological systems generate increasingly large volumes of streaming data, the ability to analyze information and take actions in real time is essential. A key challenge in sequential signal processing and data analysis is the detection of rapid changes or anomalies in the statistical properties of data. It has applications in a variety of branches of engineering, such as wireless communications, sensor networks, radar, power grid, environmental monitoring and the Internet of Things. Effective real-time monitoring for changes is crucial for situational awareness, adaptation, and even security. The problem of real-time change detection in streaming data is known as a quickest change detection (QCD) problem. The fundamental goal is to detect changes as quickly as possible, subject to a constraint on the rate of false alarms. While the change detection problem has been extensively studied especially with regard to univariate time-series, the large-scale and multistream nature of modern applications present new challenges that are addressed in this thesis. First, we consider change detection when the distribution of the data after the change depends on a high-dimensional unknown parameter vector. In this setting, typical methods experience performance degradation since the high-dimensional parameter is difficult to estimate accurately in real time. We propose new methods that utilize the James-Stein shrinkage estimator to obtain better estimates of the parameter vector and show that substantial performance improvements are obtained when detecting a mean-shift in high-dimensional Gaussian data. Performance gains are shown analytically in both asymptotic and finite sample regimes, and in simulations. The direction of the shrinkage can be selected by the user, and it may depend on the observed data. It is shown, that the magnitude of the performance gain depends on the distance between the shrinkage target and the true post-change probability model. Second, we study the intersection of quickest change detection and multiple hypothesis testing. A centralized decision maker receives local decision statistics from a large number of sensors and runs multiple change detection tasks simultaneously, corresponding to e.g. different spatial locations or frequency bands. In this setting, conventional QCD error criteria developed for single stream settings may not capture the overall system performance accurately. We propose methods that provide control of the false discovery rate, a scalable and interpretable error criterion relevant to applications. In a Bayesian formulation, a method that minimizes average total detection delay subject to a constraint on the false discovery rate is derived. In addition, the proposed methods reduce the amount of data transmission between the sensors and the decision maker by adaptively choosing only a fraction of the sensors for monitoring at each time step. The thesis also contributes to detection problems involving spatially propagating phenomena that are observed with potentially mobile sensor networks. The objective is formulated as a dynamic programming problem and the structure of the optimal stopping rule is derived. We further propose a simpler, practically implementable threshold-based algorithm, that corresponds to a limiting form of the optimal test and establish its asymptotic optimality. The performance of the threshold test is studied when the data by the sensors corresponds to energy detector statistics, and in a scenario involving attenuation of propagating signals. Overall, the contributions of this thesis advance the theory and practical applicability of signal processing and change detection methods in large-scale, multi-stream environments, enabling more efficient and responsive monitoring and decision-making systems.
  • Modeling human decision-making in naturalistic settings
    (2025) Putkonen, Aini
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-10
    The widespread availability of interactive devices implies that choices are often based on information presented in graphical user interfaces (GUIs), such as when booking a flight or a hotel online. These tasks involve selecting an option among competing alternatives within visual displays, necessitating evidence accumulation to support decision-making. Modeling how people make decisions in such settings is valuable for both theory and practice. A better understanding of naturalistic decision-making could, for instance, aid in developing practical applications like decision support systems. However, much of the previous work on modeling decision-making focuses on controlled settings that differ significantly from everyday decision contexts. Additionally, the process of information gathering in information-rich interfaces is under-studied, limiting our understanding of decision-making in such settings. This thesis aims to bridge this gap by extending models of decision-making from controlled environments to naturalistic settings to explain and predict choices, using the case of GUIs. The contributions of this thesis can be summarized in three main claims. Firstly, I argue that applying existing cognitive models in naturalistic settings—where the experimental design cannot be controlled—must be undertaken cautiously. This is due to the quality of model fitting being dependent on the often limited and uncontrolled set of tasks presented to the users. Such circumstances can lead to potential issues with parameter recovery, as demonstrated with the application of two classic decision-making models under risk to naturalistic game logs. Secondly, I contend that understanding the information gathering process preceding a choice is critical for effectively modeling decision-making in GUIs. I discuss how information gathering may be studied through representative eye-tracking studies where the stimuli preserve the structure of the naturalistic environment. In these settings, users' visual attention, the order in which they focus on elements, and their reaction times are influenced by the features of the GUI. I present evidence of these observations from two empirical studies focusing on visual search and browsing. My third claim states that information gathering should be integrated into models of naturalistic decision-making. This can be achieved by representing decision-making as a partially observable Markov decision process (POMDP) solved using reinforcement learning (RL). The proposed approach is based on the concept of computational rationality, which I argue is suitable for modelling naturalistic choice for three reasons: it appropriately captures mechanisms that explain information gathering, relaxes data requirements, and allows for the incorporation of machine learning-based elements for enhanced predictive accuracy. I demonstrate this approach through the example of multi-attribute choice, reproducing various context effects in property selection. This thesis concludes by discussing the implications of these findings for our theoretical understanding of decision-making in naturalistic settings, connecting the results to the Adaptive Interaction framework. Additionally, I consider workflows as a tool to further advance modeling of decision-making in naturalistic environments.
  • Holistic renovation strategies for cold climate buildings: simulation-based insights into decarbonization and cost-effectiveness
    (2025) Hu, Xinyi
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-10
    The building sector is a major contributor to global energy use and CO2 emissions, accounting for approximately 40% of total energy use in the European Union, and 21% in China. In cold climate regions, where space heating dominates end-use demand, renovating existing buildings serves as a critical strategy for achieving long-term decarbonization goals. Despite growing interest in this area, there remains a need for systematic and transferable frameworks to assess holistic renovation strategies under various building types and energy contexts. This thesis proposed a simulation-based multi-objective optimization framework to identify cost-effective and low-carbon renovation solutions. Two case studies demonstrated its applicability: a rural residential building in northern China and a public kindergarten in Finland, which were evaluated using life-cycle cost (LCC) and annual CO2 emissions as key performance indicators. Given the local heating practices in rural China, the first phase assessed the impact of heating operation modes on the performance of individual envelope renovation. Compared to the prerenovation baseline, cost-optimal solutions achieve a 30−40% emission reduction. Under intermittent heating, LCC increases by 7%, while continuous heating leads to an 8% decrease. To improve indoor air quality during heating season, the second phase evaluated the technoeconomic and environmental performance of four mechanical ventilation systems. Exhaust ventilation system is identified as the most cost-effective option in life-cycle terms, while balanced ventilation system incorporating heat recovery and earth-air heat exchanger achieves the lowest CO2 emissions. The inclusion of earth-air heat exchangers reduces reheater capacity and defrost operations, with enhanced performance under colder conditions. The optimization method was then extended to holistic renovation packages that integrate envelope upgrades, ventilation improvements, and decentralized energy systems. Biomass-based system and photovoltaic (PV)-combined heat pump system achieve annual 53−90% emission reduction and 6−18% LCC savings compared to pre-renovation. Their cost-optimal configurations offer comparable LCCs to envelope-only solutions, with additional emission reductions of 15−52%. The method was further applied to a Finnish hybrid energy system combining renewable-powered heat pumps, PV generation, and district heating (DH). PV-combined ground-source heat pump with zero-emission DH from hydrogen waste heat achieves the lowest LCC, reducing costs by up to 23% compared to DH-only cases. This lower DH pricing scheme also reduces heat pump capacity by 10% points. Both decentralized and hybrid renewable-powered electrification strategies offer resilient, cost-effective solutions for decarbonizing cold climate buildings, with increasing benefits as the grid continues to decarbonize. This thesis presents a comprehensive, adaptable framework for identifying renovation strategies and offers practical guidance for carbon-neutral, economically viable building systems in cold climates.
  • Optical modification of 2D materials
    (2025) Varjamo, Suvi-Tuuli
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-10
    Two-dimensional layered materials, characterized by their atomic thickness and diverse properties, are a promising material platform for advancing many technological fields, including physics, quantum technologies and medicine. In the past twenty years, the family of two-dimensional materials has expanded and nowadays encompasses thousands of materials that range from insulators to conductors, with physical properties for applications across the range. However, their industrial scale applications still face issues, mainly stemming from the difficulty of obtaining cost-effective and fast large-scale fabrication methods. Optical modification methods are processes that utilize the energy of light to drive fabrication and property modification processes, such as doping and patterning. They provide a great alternative to conventional fabrication methods as they are sustainable, fast, cheap, and can provide highly localized modification without the use of masks. This thesis focuses on advancing the field of optical modification by exploring optical defect engineering of both monolayer and multilayer heterostructure transition metal dichalcogenides. Femtosecond laser irradiation is found to enhance the nonlinear optical responses of transition metal dichalcogenide flakes by resonances with defect states while also creating three-dimensional structures in the flake. On the other hand, continuous wave irradiation of transition metal dichalcogenide heterostructures is found to simultaneously modify both constituent materials at different rates, leading to enhanced photoluminescence and drastic changes in electrical properties. These findings not only deepen our understanding of the interaction between light and two-dimensional materials but also highlight the untapped potential of optical modification methods. By applying the results in practical applications, this work further establishes optical modification methods as a true alternative for scalable and sustainable technologies that harness the unique properties of two-dimensional materials for next-generation innovations.
  • Deep reinforcement learning-driven optimization for UAV-enabled wireless networks
    (2025) Bai, Yu
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-10
    Unmanned Aerial Vehicles (UAVs) have become essential components of modern wireless networks, supporting critical applications such as emergency communications, Internet of Things (IoT) deployments, surveillance, and disaster recovery. UAV-enabled wireless networks leverage UAVs' inherent advantages, including high mobility, rapid deployment, and superior line-of-sight communication, significantly enhancing network coverage, robustness, and operational efficiency. Nevertheless, these advantages introduce intricate optimization challenges, particularly in dynamic UAV deployment scenarios, complex trajectory planning, and interdependent resource management decisions. Traditional static optimization methods face difficulties in effectively managing these dynamic, non-convex optimization problems characterized by highly coupled variables, necessitating more advanced and adaptive solutions. Deep Reinforcement Learning (DRL) addresses these challenges effectively by integrating deep learning’s robust feature extraction capabilities with reinforcement learning’s adaptive decision-making strength. This thesis examines DRL-driven optimization in three critical scenarios of UAV-enabled wireless communication applications. First, we explore dynamic multi-UAV deployment for adaptive wireless coverage, aiming to optimize UAV operational modes, transmission power, and movement strategies to balance power consumption and ground user coverage effectively. A multi-modal feature-based DRL model addresses these coupled optimization challenges efficiently. Second, we address UAV-assisted data collection in backscatter wireless sensor networks, where a UAV equipped with a directional movable antenna (MA) enhances communication efficiency. A tailored DRL approach is employed to jointly optimize UAV trajectory and MA orientation, minimizing the total data collection time and associated energy consumption. Lastly, we focus on UAV-enabled integrated sensing and communication (ISAC) systems designed for time-critical missions. The primary objective is to minimize the age of information (AoI) by jointly optimizing UAV trajectory planning and beamforming strategies. A DRL framework incorporating Kalman filtering for target tracking and regularized zero-forcing beamforming is developed, effectively balancing sensing accuracy with communication quality. Overall, this thesis establishes adaptive DRL frameworks specific to diverse UAV scenarios, enhancing the adaptability, efficiency, and performance of UAVenabled wireless networks.
  • Mapping, tree detection, localization, and autonomous flight of unmanned aerial vehicles in forest applications
    (2025) Ouattara, Issouf
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-10
    Two forest management tasks which are important for healthy forest growth are the cleaning of seedling stands and the prevention of moose damage. These management tasks are still largely done manually. They are time-consuming, labour-intensive, and likely to face a labour shortage in the near future. This thesis has two aims: (1) to develop approaches that use images collected by a UAV platform to support the semiautonomous cleaning of a seedling stand and (2) to develop methods that enable UAVs to autonomously and selectively spray a bio-based repellent to prevent moose damage. In the first two articles of the thesis, deep learning methods are used to detect seedling trees in the images collected by a low-cost UAV platform, thereby creating a map of the seedling trees. A graphbased registration approach is developed to localize the forest cleaning machine within the map. A human-machine interface is also developed to incorporate the registration algorithm and to show in real-time the seedling trees and the cleaning tool to the machine operator. The approaches developed in these two articles demonstrate the feasibility of the UAV-assisted semiautonomous cleaning of seedling stands. The last three articles of this thesis develop approaches for state estimation, loop closure detection, seedling tree detection from LiDAR data, and path planning. A pose-graph state estimation approach is developed, achieving accurate position, attitude, and consistent velocity estimates. A novel loop closure method using surface variation features is developed to correct the drift in pose estimation. A region-growing algorithm is developed to segment the individual seedling trees from the LiDAR point cloud. Finally, a path planning utilizing an octree data structure, and the informed RRT* approach is used to plan collision-free paths enabling the UAV to fly from the top of a tree to another. These approaches solely rely on IMU and LiDAR sensors; they are fully implemented on a single board computer onboard the UAV platform, and they achieve real-time performance. Several real-world experiments have been conducted to test the methods developed in this thesis, including the autonomous flight in a seedling forest stand.
  • Model-based reinforcement learning for integrated radar and communications systems
    (2025) Pulkkinen, Petteri
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-03
    Convergence between radar and wireless communications systems and growing demand for radio spectrum motivates the integration of radar and communications functionalities into unified, integrated sensing and communications (ISAC) systems. As sensing and communications tasks may have conflicting objectives, this integration poses new challenges for radio resource allocation and waveform co-design. A promising avenue for this integration is provided by multicarrier waveforms and systems with multiantenna architectures. However, time-varying channels and interferences as well as many degrees of freedom in frequency and spatial domains pose a need for efficient algorithms capable of adapting waveforms and resource use in real-time. Waveform optimization and resource allocation problems have been traditionally solved using structured optimization approaches that are incapable of learning from experiences and are susceptible to modeling deficiencies. Therefore, this thesis adopts a data-driven reinforcement learning (RL) approach for optimizing resource allocations and waveforms in ISAC systems. In particular, the focus is on model-based reinforcement learning (MBRL) that, unlike model-free RL, can utilize the rich structural knowledge about sensing and wireless communications systems to improve data-efficiency and interpretability. This approach builds on a constrained partially observable Markov decision process (C-POMDP) model that allows target states and radio channels to be dynamic and captures uncertainty in observations about the targets and radio channel conditions. Furthermore, C-POMDPs enable balancing between communications and sensing tasks via constrained formulation. MBRL approach is well-known in the field of control theory. However, fundamental differences, such as the complexity of the decision spaces and dynamical models, make applying traditional MBRL and control algorithms to ISAC systems difficult. Therefore, this dissertation develops a practical approach based on online learning as well as myopic and pseudo-myopic control strategies tailored for ISAC systems. Analytical bounds are derived and verified numerically for the myopic strategy to give rigorous performance guarantees. The myopic and pseudo-myopic strategies are also demonstrated in practical multicarrier and multiantenna ISAC resource allocation and waveform optimization problems. These problems involve allocating resources or powers in the frequency domain (sub-carriers or resource blocks) and beamspace domain (codebook of beams) while operating in dynamic target and radio environments. This dissertation shows that computationally efficient algorithms based on MBRL can be developed to solve practical resource allocation and waveform design problems in multicarrier and multiantenna ISAC systems. In particular, the data efficiency and interpretability are improved compared to the model-free RL approach. Furthermore, this online learning capability enables ISAC systems to adapt to dynamic radio environments to improve robustness and performance compared to traditional structured optimization approaches.
  • Towards decentralised platform governance with feedback economics and open blockchains
    (2025) Elo, Tommi
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-03
    The central thesis of this dissertation is that efficient decentralised governance is preferable to command-and-control hierarchies because decentralisation better fulfils the enlightenment values of egalitarian governance and liberty. Governance is the vital interdisciplinary activity of overseeing and controlling the direction of an organisation, such as a platform or community, thereby enabling collective action. However, most current business activity, including digital business, occurs in command-and-control hierarchies, that is, in institutions and companies which are centrally governed. Moreover, the creation of more decentralised institutions has proven elusive as power typically becomes rapidly centralised. As a cryptographic and technical solution, open blockchains offer a promising approach to alleviating this less-than-ideal centralising tendency. This dissertation maps the ways and extent to which open blockchains and distributed ledger technologies (DLTs) can help decentralise governance. The primary methodology utilised in this dissertation is System Dynamics. It allows the model of a system to include both social and technological processes; i.e., it can express both constructivist and realist worlds. It also describes the longitudinal aspects of the system, i.e., behaviour over time, which enables simulations. The results demonstrate the suitability of System Dynamics in DLT- and blockchain research and design. As a theoretical contribution, the dissertation developed System Dynamics (SD) simulation models for how network effects can create strong social norms, how open DLTs can increase collaboration resilience under competitive pressure and create shareable antirival units of account. The central result is the model of how these new type of tokens can increase allocative efficiency of the work of the digital community. Moreover, in decentralised identity creation, DLTs were shown to be useful as enablers of privacy. However, brute force denial of service by attackers directly utilising money was found to be a problem in today's non-access-controlled open DLT-based systems. All these results are relevant for well-functioning data markets and digital community building. Both are key accounting areas when we move from scarce (material) resources as the main definers of the economy, more towards the ultimate commons, which are antirival non-wearing resources, such as, data, information, knowledge, and wisdom.
  • Construction of few-angular spherical codes and line systems in Euclidean spaces
    (2025) Ganzhinov, Mikhail
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-10-03
    Spherical codes are finite non-empty sets of unit vectors in d-dimensional Euclidean spaces. Projective codes, also known as line systems, are finite nonempty sets of points in corresponding projective spaces. A spherical code or a line system is called few-angular if the number of distinct angular distances between vectors or lines of the code is small. The fundamental problem is to find a code with minimum angular separation between the vectors or lines as large as possible. In this dissertation few-angular spherical codes and line systems are constructed via different algebraic and combinatorial methods. The most important algebraic method is automorphism prescription (in different forms) while combinatorial methods include exhaustive isomorph-free generation of Gram matrices of spherical codes and weighted clique search in graphs with vertices representing orbits of vectors. We classify the largest systems of real biangular lines in d≤6 and construct two infinite families of biangular line systems which achieve equality in the second Levenshtein bound from irreducible representations of finite groups SL(2,q). We also construct several low-dimensional spherical codes with prescribed automorphisms and large minimum angular distances between the vectors and obtain new lower bounds for kissing numbers in dimensions 10,11 and 14.
  • Measurements of temporally modulated LED light sources
    (2025) Mantela, Ville
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-09-19
    LED-based light sources have become the de facto light source type in recent years due to having superior energy efficiency when compared to traditional incandescent light sources, along with regulations made worldwide banning the sales of incandescent, fluorescent, compact fluorescent and mercury based light sources to consumers. This shift in lighting technology has created challenges in optical metrology, especially as the LED sources have different temporal modulation characteristics to incandescent light sources. Temporal Light Modulation (TLM) is the variation in illumination over time, which can cause different ill effects on human health under the illumination of these sources, ranging from annoyance to even epileptic seizures. The LED differs from incandescent sources in temporal behaviour as the diodes only work on direct current (DC), while incandescent light sources use the alternating current (AC) coming from the electric source, such as electrical outlet at homes. The electronics needed to the AC-DC conversion affects the temporal behaviour of the LED lights, resulting in various temporal light outputs. In this thesis, advancements in the measurements of temporally modulated lights are demonstrated. First, a novel set of Temporal Light Artefact (TLA) software was created to calculate the metrics of temporally modulated light more accurately than previously available methods. Secondly, a hyperspectral camera-based measurement method was tested to enable the measurement of more complex LED arrangements, including multiple sources, and measurement of TLA from different parts of spectrum. Finally, the Ecodesign regulation in the European Union, which sets limits for TLA parameters in consumer-grade LED light sources, was addressed by measuring LED lamps that were manufactured both before and after the regulation was implemented. It was found that the regulation was considered by the manufacturers, and lamp types with lower TLA values were adopted, thereby minimizing the risk of adverse effects to humans.
  • Sea-ice field analysis in polar regions for smart ships
    (2025) Sandru, Andrei
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-09-05
    Due to climate change, Arctic and Antarctic regions have experienced a diminishing sea-ice extent during the last decades. This situation poses both a risk factor for environmental disasters (e.g., loss of fauna, increasing sea levels), yet at the same time it presents new opportunities for ship operations (e.g., new and more efficient routes for cargo transit, research and tourism). However, the navigation or operation in ice-infested waters continues to pose a challenge, and detailed information about the sea-ice conditions is essential for navigation, to avoid in-water hazards (e.g., ice ridges, compression zones), but also to optimize parameters such as transit time and fuel consumption. In this research, traditional machine vision methods are used to develop an accurate sea-ice field analysis process by means of vision-based systems, capable of producing data describing local ice conditions such as concentration, sea-ice floe sizes and distribution. To support the orthorectification (i.e., obtaining a virtual bird’s eye view) and photogrammetry processes, a highly accurate method for attitude estimation from the horizon line is introduced, by means of a single monocular camera or through a visual-inertial sensor fusion. Then, methods for mapping highly dynamic environments using a laser scanner (LiDAR) were successfully implemented, to produce 3D point-cloud maps of the ice field in Antarctic conditions. Lastly, sensor fusion is used to produce highly detailed 2D maps of the sea-ice fields. These maps have been used so far in developing ship-ice interaction simulation models and represent the first attempt at digitizing and mapping sea-ice fields from imagery and other sensor data collected onboard a ship with decimetre-level accuracy. Additionally, during the study two experimental setups have been integrated, programmed, and instrumented onboard S.A. Agulhas II, to collect the required full-scale research data during various voyages to the Antarctic waters. The study aims at expanding the maritime industry’s knowledge and capabilities in ice-covered waters, by developing and improving algorithms for sea-ice field analysis and mapping using machine vision cameras and LiDARs. Such environments present real challenges, since there are no reliable ground-points available, and almost no static targets. Furthermore, the algorithms are implemented and evaluated in small- and full-scale real systems, not relying on simulators. These algorithms and methods support the automation of sea-ice monitoring, safe and efficient (semi- ) autonomous navigation in ice-covered waters, as well as the development of simulation models for ship-ice interaction.
  • Ultra-Low Power Circuits for Batteryless Energy Harvesting Systems and Thermal Compensation in Resistive In-Memory Computing
    (2025) Monga, Dipesh C.
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-08-29
    Embedded systems, from wearable health monitors and implantable diagnostics to environmental sensor nodes, are becoming increasingly prevalent in modern life. These platforms are often expected to operate continuously under severe energy constraints, where frequent battery replacement is impractical. To address this, there is a growing need for ultra-low-power (ULP) circuits capable of harvesting ambient energy from sources such as radio frequency (RF) fields, biochemical reactions, and photovoltaic cells. Ensuring stable operation under these limited and varying energy conditions requires circuits with ULP consumption and robust performance under variations. This thesis presents an approach to the design, implementation, and experimental validation of ULP integrated circuits across multiple circuit blocks, tailored for energy-autonomous and flexible systems. The contributions span several key building blocks of energy-harvesting systems, including variation-insensitive voltage and current reference generators, RF-DC converters, low-dropout (LDO) regulators, and switched-capacitor (SC) DC-DC converters with finegrained, arithmetic progression-based voltage scaling. Further, the work introduces thermal compensation techniques to maintain computational accuracy for analog in-memory computing units under varying thermal conditions. The circuits presented in this thesis are designed and fabricated using a conventional CMOS in 65 nm and Pragmatic 600 nm flexible indium gallium zinc oxide (IGZO) based technology using unipolar TFT-based transistors. Variation-insensitive reference generators form the foundation for reliable biasing across the circuits presented in this work. To address this, the thesis implements amplifier-free, MOS-based voltage and current references that ensure stable operation under varying conditions. A dual-mode, all-NMOS circuit is developed to function both as a voltage reference and a temperature sensor, enabling efficient circuit reuse in energy and area constrained systems. This circuit is further extended to flexible electronics, with a voltage reference designed using IGZO thin-film transistors. For energy regulation, an LDO based on IGZO unipolar transistors is presented, offering efficient voltage regulation under a low quiescent current of 150 nA. An RF-to-DC converter targeting operation in 13.56 MHz is also developed to harvest energy from wireless sources. To support energy sources with variable and degrading outputs, such as biofuel or zinc-air cells, a reconfigurable switched-capacitor DC-DC converter is introduced, with arithmetic progression control of voltage conversion steps of 0.125. The proposed circuits are implemented for voided fluid volume sensing in smart diapers powered by urine-based energy harvesters. Additionally, voltage regulation using DC-DC converters operating from degrading and decaying energy sources has been designed. The circuits implemented are validated through system-level integration in practical applications. A smart diaper platform powered entirely by harvested urine energy demonstrates the feasibility of fully autonomous operation. Additionally, the thesis addresses thermal variability in analog in-memory computing arrays through two compensation techniques: one using programmable calibration, and another using on-chip thermal sensing for automatic adjustment. The circuits developed in this work enable energy-autonomous operation in batteryless systems and provide robust thermal stability for analog in-memory computing.
  • Challenges and solutions of metasurface based reconfigurable intelligent surfaces
    (2025) Shabanpour, Javad
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-08-29
    Reconfigurable Intelligent Surfaces (RISs) have emerged as a transformative technology in wireless communications, promising enhanced signal propagation, interference management, and energy efficiency. This thesis develops novel methodologies for designing RISs with improved electromagnetic parameters. In the first part of the thesis, a key focus is done on relationships between reflection locality, physical optics, and angular stability of a uniform metasurface. We show that angular stability of a uniform metasurface is crucial for a RIS implemented as its periodically non-uniform analogue (with tunable period). Through analytical modelling, numerical simulations, and experimental validation, our study results in such binary metasurfaces (BMSs) that possess following advantages compared to previously known ones: higher angular stability of the reflection phase, true polarization insensitivity of the scattering pattern, broader operation band, and what is not less important, possibility to accurately predict the operation characteristics with the use of simple analytical models. In the second part of the thesis, we aim to achieve higher scattering efficiency. With this purpose our research advances the discrete sheet impedance approach for periodically non-uniform metasurfaces of general type. We optimize their beamforming efficiency, overcoming the previously adopted limitations of conventional (diagonal matrix) techniques. A novel framework for RIS with multiple reradiation modes is developed, leveraging both unequal coefficient superposition and discrete impedance optimization. The methodologies proposed and validated in this thesis contribute to the realization of highperformance, cost-effective RISs for future wireless networks – 6G and beyond.
  • Unsupervised audio enhancement with diffusion-based generative models
    (2025) Moliner Juanpere, Eloi
    School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2025-08-22
    Audio recordings are often compromised by noise, reverberation, and other distortions, leading to loss of quality. Examples of this include historical music recordings affected by the degradation of analog media or speech recordings where reverberation reduces intelligibility. Audio enhancement and restoration techniques are used to recover and improve the acoustic quality of these recordings. At the time of this thesis, the state-of-the-art audio restoration methods are predominantly datadriven, with deep generative models demonstrating exceptional expressivity. However, most of these approaches rely on supervised learning, which, while successful, comes with inherent limitations. These include a restricted generalization to unseen degradations, as well as the need to train task-specific models for each different restoration scenario. This thesis explores an alternative unsupervised approach that employs unconditional generative models, specifically diffusion models. In this context, a single generative model, trained without prior knowledge of specific degradation processes, can be adapted to an endless variety of restoration tasks during inference, thus overcoming the limitations of task-specific supervised models. The first and second publications included in this thesis demonstrate the effectiveness of this approach in several known restoration problems, including music bandwidth extension, inpainting, and declipping, supported by objective and subjective evaluations. This thesis also addresses blind restoration problems, where the characteristics of the degradation are unknown. The third publication presents a blind approach to audio bandwidth extension for historical music restoration, where the lowpass filter degradation is automatically estimated and iteratively refined during the generation process. The fourth publication extends this work to generative equalization, enabling both the correction of spectral coloration and the regeneration of missing content. This method has shown significant improvements in the restoration of historical gramophone recordings, particularly for piano and singing voice performances. The final two publications focus on single-channel blind speech dereverberation. Here, speech signals affected by room reverberation are enhanced using a diffusion model trained on anechoic speech, combined with a parametric subband filtering model of room impulse responses. This approach allows for simultaneous estimation of both anechoic speech and the room impulse response. The method is evaluated on multiple datasets through objective and subjective experiments, demonstrating performance that matches or surpasses supervised baselines, particularly in conditions that differ from the training data.