Browsing by Department "Informaatio- ja tietoliikennetekniikan laitos"

Filter results by typing the first few letters
Now showing 1 - 20 of 38
  • Advances and New Applications of Spectral Analysis
    (2023) Esfandiari, Majdoddin
     School of Electrical Engineering | Doctoral dissertation (article-based)
    Spectral analysis is a mathematical tool for modeling signals and extracting information from signals. Among the areas where it finds applications are radar, sonar, speech processing, and communication systems. One of applications of spectral analysis is to model random signals with the so-called rational models like autoregressive (AR) model. Spectral analysis is used for estimating direction-of-arrival (DOA) in array processing as well. In addition, spectral analysis is used for channel estimation in communication systems such as massive/mmWave MIMO systems. It is because channels in these systems can be modeled with multipaths' gains and angles. This thesis proposes methods for the problems of noisy AR parameter estimation, DOA estimation in unknown noise fields, and massive/mmWave MIMO channel estimation and data detection with one-bit ADCs. The AR model offers a flexible yet simple tool for modeling complex signals. In practical scenarios, the observation noise may contaminate the AR signal. In this thesis, five methods for estimating noisy AR parameter estimation are proposed. In developing the methods, the concepts such as eigendecomposition (ED) and constrained minimization are used. The most common assumption about the structure of the observation noise in DOA estimation problem is the uniform noise assumption. According to it, the noise covariance matrix is a scaled identity matrix. However, other noise covariance matrix structures such as nonuniform or block-diagonal are more accurate in some practical situations. A generalized least-squares (GLS)-based DOA estimator that takes into consideration the signal subspace perturbation and also enjoys a properly designed DOA selection strategy is proposed for the case of uniform sensor noise. For the case of nonuniform noise, a non-iterative subspace-based (NISB) method is developed which is computationally efficient compared to state-of-the-art competitors. Moreover, a unified approach to DOA estimation in uniform, nonuniform, and block-diagonal sensor noise is presented. The use of one-bit ADCs instead of high-resolution ADCs is considered as an elegant solution for reducing power consumption of large-scale systems such as massive/mmWave MIMO and radar systems. In this thesis, we use the analogy between binary classification problem and one-bit parameter estimation to develop algorithms for massive MIMO and mmWave systems. In this regard, a method called SE-TMR is developed for one-bit mmWave UL channel estimation. Another method called L1-RLR-TMR is also offered for one-bit mmWave UL channel estimation. At last, the concept of AdaBoost is combined with Gaussian discriminant analysis (GDA) for developing computationally very efficient channel estimators and data detectors. It is shown that the proposed methods which use approximate versions of GDA as weak classifiers in iterations of the AdaBoost-based algorithms are exceptionally efficient, specifically in large-scale systems.
  • An AI-based Framework to Optimize Mobile Services
    (2023) Naas, Si-Ahmed
     School of Electrical Engineering | Doctoral dissertation (article-based)
    In current digital networks, novel verticals require diversity and extreme efficiency from network infrastructures. Recent studies have shown the key role of Artificial intelligence (AI) to improve resource management and to reduce operational costs of mobile services and networks. Furthermore, the rapid development of connected Internet of Things devices has created a variety of services. Since the complexity of mobile services has increased, efficient designs for these services are required to satisfy increasing power and processing requirements. Mobile services, however, face numerous challenges, including a shortage of data, inefficient data processing, and a lack of multi-service collaboration mechanisms. These setbacks have delayed the development of commercial applications that rely on mobile networks. Therefore, this thesis presents a framework, based on the "global brain" concept, to orchestrate various mobile services. The framework enhances mobile services for two scenarios: 1) low data rate (e.g., ambient intelligence) and 2) high data rate (e.g., virtual reality (VR) streaming). The distinction between low and high data rates is based upon bandwidth requirement. For low data rate scenarios, this thesis investigates the usefulness of emotion-based data for static (e.g., customer satisfaction) and dynamic (e.g., location-based recommendations) environments. We begin by presenting a practical multimodal emotion recognition system based on audio and video information, which improves emotion recognition accuracy. The results indicate that the integration of emotions is vital for customer satisfaction assessment and recommendation systems. For high data rate scenarios, we address VR streaming due to its high computational cost and latency. Notably, improved viewport (VP) and gaze prediction schemes are critical for enabling smoother VR streaming, as these features are the foundation of the user's experience while wearing the headset. Thus, the global brain proposes novel VP and gaze prediction schemes which produce high prediction accuracy and reduced resource consumption. Finally, this thesis proposes a novel knowledge sharing mechanism that enables mobile services to improve their models by learning from others and incorporating global knowledge, which significantly increases inference accuracy. To prevent sensitive information leakage during knowledge sharing among service providers, we propose a lightweight multi-key homomorphic encryption scheme that allows mobile services to protect their knowledge (i.e., deep neural network weights).
  • Application of Signal Processing Methods for Precision Impulse Voltage and Partial Discharge Measurements
    (2023) Havunen, Jussi
     School of Electrical Engineering | Doctoral dissertation (article-based)
    High-quality grid components are the basis for reliable power grid. Different high-voltage tests are applied to grid components according to international standards after manufacturing to ensure the claimed performance. Accuracy of the used measuring systems is crucial when more cost-efficient and better-quality products are being developed. Accuracy and traceability to SI system can be verified by calibration services provided by national metrology institutes, like VTT MIKES in Finland. This thesis presents impulse voltage and partial discharge measuring systems whose performance has been evaluated and improved using signal processing methods at VTT MIKES. Impulse voltage tests are performed to test how a test object withstands lightning strike or switching overvoltage in the grid. Test voltage is generated by an impulse generator and is measured across the test object using a measuring system based on a voltage divider. Approved measuring systems used for testing need to be periodically calibrated traceable to national or international measurement standards. Commercial systems are primarily designed to withstand the stresses of the industrial environment with sufficient measurement uncertainty for testing. Therefore, their performance is not sufficient for references used in calibration. This thesis presents three implemented methods to improve the performance of an impulse voltage measuring system. The first method is based on deconvolution, and it corrects the measured signal in frequency domain by utilizing the measured step response of the system. Time parameter errors of 1 % were improved to be circa 0.1 %. The second method is a time domain correction of the drooping response of a damped-capacitive voltage divider. Correction is based on the measured time constant of the low-voltage arm improving the used system from not-approved to low-uncertainty reference. The third method reduces the distortion caused by the signal cable of a resistive voltage divider by shortening the signal cable or by matching the signal cable only on the divider end. Insulations used in high-voltage systems may have imperfections. Under high-voltage stress, these can cause a local electric field enhancement so that the intrinsic field strength is exceeded causing a localized electrical discharge. Partial discharge can cause premature ageing of the insulation or lead to unrecoverable damage. To enable accurate detection of partial discharge, each test setup must be separately calibrated using a partial discharge calibrator. The lowest charges of calibrators have been difficult to calibrate with reasonable measurement uncertainty using the traditional calibration methods. This thesis presents a calibration method based on charge-sensitive preamplifiers which allows to measure small charges more accurately. New calibration services introduced, lowest calibration limit is extended from 1 pC down to 0.01 pC, with typical uncertainty of 1 %.
  • Applications of Interferometric Measurements and Photoacoustic Detection in Optical Metrology
    (2023) Sharma, Sucheta
     School of Electrical Engineering | Doctoral dissertation (article-based)
    Measurement loops and detectors are irreplaceable constituents of an experimental process in optical metrology. Interferometry and photoacoustic methods can be applied as key techniques for developing such measurement and detection systems. In this thesis, two interferometric measurement arrangements are presented. The phase modulation process in interferometry is important for the measurement speed. Mechanical phase modulation, in such cases, suffers from limitations with regard to driving voltage amplitude, increased modulation frequency and system noise. In this work, the potential of electro-optic phase modulation has been assessed for developing a multi-wavelength interferometric sensor to replace the mechanical phase modulation system. The results not only show the suitability of the electro-optic sensor to improve the measurement speed of the multi-wavelength interferometer, but the sensor is also able to operate at considerably low driving voltage. Another interferometric method driven by an optical system consisting of a mirror array has been presented in this thesis to measure the surface parallelism of step gauge faces. Contact methods are commonly used for this purpose. However, the research gap lies in the available options for non-contact methods to carry out such measurements. The alignment sensitivity is a major factor that controls the accuracy of the measurement with the presented array of mirrors which are arranged as a periscope and a triangular prism reflector-type configuration. Hence, in this work, methods to monitor the alignment sensitivity and estimate corresponding corrections have been presented. The theoretical and experimental studies on the performance of the custom-built optical system show that the system has the potential to serve as a suitable tool for non-contact surface parallelism measurement of step gauge faces. For the development of detection instruments, a silicon cantilever-based photoacoustic radiation detector is presented in this work. Photoacoustic detectors are widely employed for power meter applications, however, in most available cases, the pressure sensors have limitations on the highest detectable radiation power. The photoacoustic detector presented in this work, has shown radiation detection capability with a linear dynamic range of nearly six orders of magnitude and highest detectable power in the milliwatt level which can be further extended with suitable adjustments of the detection parameters. The spectral coverage of the system was initially tested from ultraviolet to infrared region, and in the latest work the radiation sensitivity has been successfully demonstrated in the terahertz range with proper absorber materials. A numerical model for designing the cantilever pressure sensor has also been developed to improve the detection sensitivity. It is concluded that the cantilever-based photoacoustic detector can be a good solution for power measurement applications which require a broad spectral coverage and large dynamic range with robust pressure sensing element offering high damage threshold.
  • Attention-based End-to-End Models in Language Technology
    (2024) Rouhe, Aku
     School of Electrical Engineering | Doctoral dissertation (article-based)
    Speech recognition specifically, and language technology more generally, have started to find everyday use. Challenging language tasks have become feasible through a continued growth in data resources and compute capacity, and through neural networks methods which are able to take advantage of this growth. As applications continue to integrate more deeply into our lives, it is important to understand and follow the many directions that these fields may take. At the turn of the 2020-decade, end-to-end models have received a lot of attention. End-to-end models hold  promise of simpler solutions, which nonetheless may scale better with data and compute. On the other hand, end-to-end models defy decomposing tasks into easier subproblems. This decomposition allows modular designs, which permit a wider variety of data sources to be used. It remains unclear whether the end-to-end models are truly an improvement over previous technologies. It is not straight-forward to compare end-to-end and decomposed solutions fairly, because of their many differences. This thesis proposes a principled approach for comparisons of such heterogeneous solutions and applies it to speech recognition. In their default configuration, the end-to-end models forego many useful data sources, and rely solely on expensive end-to-end labeled data. This thesis explores methods for leveraging additional data sources in speech recognition, canonical morpheme segmentation, and spoken language translation. Additional data sources are especially useful in low data and under-resourced tasks. These difficult tasks often need the structure imposed by decomposed solutions. This thesis investigates end-to-end models in an under-resourced speech recognition and a low data canonical morpheme segmentation task. The tasks explored in this thesis are connected through a shared architecture: attention-based encoder-decoder models. Though these attention-based models are most often outperformed by hidden Markov model speech recognition systems, they showcase remarkable flexibility. They succeed in speech recognition using just tens of hours and upto thousands of hours of data. They learn to exploit auxiliary speaker and segmentation-marker inputs. They perform spoken language translation in one step. They even yield the author a first place in a public benchmark competition.
  • Audio Decomposition for Time Stretching
    (2024) Fierro, Leonardo
     School of Electrical Engineering | Doctoral dissertation (article-based)
    Time-scale modification is a common audio signal processing task that involves changing the duration of a sound without altering its frequency content. This thesis explores transients and noise sounds in the context of audio processing and investigates the use of sound decomposition to improve the quality of time scaling for normal and extreme stretching factors. Traditionally, time-stretching methods often introduce artifacts, such as phasiness and transient smearing, especially when the stretching factor is large. To address the issue, this thesis introduced an improved method to decompose sounds into their constituent sine, transient, and noise components, and a different processing technique can be separately applied to each individual class. This allows for better preservation of transient features, even at extreme stretching factors, and improves the perceived quality of time-stretched audio signals compared to traditional methods. This thesis also presents an alternative audio-visual evaluation method for audio decomposition using an interactive audio player application, where access to the individual sinusoidal, transient, and noise classes is granted through a graphical user interface. This application aims at covering the shortcomings of misused objective metrics and promotes experimenting with the sound decomposition process by observing the effect of variations for each spectral component on the original sound and by comparing different methods against each other, evaluating the separation quality both audibly and visually. This thesis also discusses the motivation behind the use of the sines-transient-noise decomposition for time stretching by analyzing the performance drop in a well-known time-scale modification method due to incorrect transient and noise handling. This work shows that, by adopting the proposed three-way decomposition within its framework, the quality of the timestretching performance of such a method is increased. The noise component is typically overlooked by conventional time-scale modification methods. This thesis introduces a novel hybrid design using a deep learning model to generate the stretched noise component with high quality even for extreme stretching factors, when the sound is slowed down by more than four times as it happens for slow motion sport videos or synthesis of ambient music. Finally, a simple and effective solution named noise morphing is described, producing state-of-the-art results across a wide range of audio inputs and stretching factors.
  • Auditory-model-based assessment of the effect of head-worn devices on sound localisation
    (2024) Lladó, Pedro
     School of Electrical Engineering | Doctoral dissertation (article-based)
    Head-worn devices (HWDs), e.g. headphones, head-mounted displays or helmets, inherently introduce acoustic distortions to the sound reaching the ear canals, potentially degrading the localisation abilities of the listener. These distortions pose potential risks to safety, hinder spatial awareness, and may affect immersion in augmented reality applications. Traditional methods for assessing the degradation in localisation due to HWDs rely on listening experiments, which are time-consuming and require specific facilities. Consequently, alternative approaches are sought, particularly in the prototyping and development phases of HWDs. This thesis investigates the feasibility of utilising acoustic measurements and auditory models to estimate the degradation in localisation caused by HWDs. We examine the efficacy of existing static localisation models in predicting experimental data when HWDs are worn. These models demonstrate robustness in their predictions, despite their initial validation under open-ear conditions only. Furthermore, we propose two auditory models tailored to estimating degradation in localisation with HWDs. The first model combines a peripheral processing front-end with a shallow neural network that estimates perceived localisation from frequency-dependent interaural cues. The second model extends an existing static localisation model based on Bayesian inference to accommodate voluntary head rotations in an auditory-aided visual search task. We validate these models with experimental data and provide publicly available implementations. Our contributions aim to enhance automatic assessment tools for HWD quality by leveraging advancements in sound localisation research. The performance of these models is robust for unseen listening conditions, highlighting the importance of integrating evidence from hearing research into assessment methodologies. This motivates the need for ongoing fundamental research in sound localisation and the development of auditory models that incorporate such findings, with the overarching goal of enhancing the quality of spatial audio applications.  
  • Autonomous Sensing using Satellites, Multicopters, Sensors and Actuators (MULTICO) - Final Report
    (2023) Virrankoski, Reino (ed.)
     School of Electrical Engineering | D4 Julkaistu kehittämis- tai tutkimusraportti tai -selvitys
    This is a consortium level public final report of the project Autonomous Sensing using Satellites, Multicopters, Sensors and Actuators (MULTICO). The project duration was 1.4.2020-31.10.2022. It was funded by Business Finland New Space Economy program and by the participating companies and research organizations. Project consortium consisted of seven companies, one subcontractor, two research organizations and two advisory organizations. Participating companies were Nokia Oyj having Insta ILS Oy as a subcontractor, Saab Finland Oy, DA-Group, Huld Oy, Iceye Oy, Nsion Oy and Small Data Garden Oy. Participating research organizations were Aalto University and VTT Technical Research Centre of Finland. Finnish National Defence University and Finnish Air Force Academy participated in advisory role. Aalto University was coordinating the whole project entity. MULTICO project included research, algorithm development and software implementation of the developed algorithms for satellite free (GNSS free / GNSS denied) navigation, pattern recognition, radio tomographic sensing and speaker identification. Temporal networks using multicopter-mounted flying base station were also developed. Everything was integrated so that it can operate as one situational awareness system. Alternatively, one or several subsystems can also operate independently. The trunk part of the integrated system was tested at winter experiments in December 2021 and finally the whole system at the final experiments in September 2022.
  • Balancing privacy and utility of smart devices utilizing explicit and implicit context
    (2024) Zuo, Si
     School of Electrical Engineering | Doctoral dissertation (article-based)
    The swift evolution of communication technologies, coupled with advancements in sensors and machine learning, has significantly accelerated the pervasive integration of smart Internet of Things (IoT) devices into various aspects of our daily lives. Examples range from automating homes to optimizing industrial processes and improving healthcare. While these applications enhance quality of life and operational efficiency, they also raise concerns about user privacy due to the collection and processing of personal data. Ensuring the seamless and secure integration of these technologies is crucial. Balancing the benefits of smart applications with protecting user privacy is the key challenge. To address this issue, we present a general method as well as customized approaches for specific scenarios. The general method involves data synthesis, which safeguards privacy by substituting real data with synthetic data. We propose an unsupervised statistical feature-guided diffusion model (SF-DM) for sensor data synthesis. SF-DM generates diverse and representative synthetic sensor data without the need for labeled data. Specifically, statistical features such as mean, standard deviation, Z-score, and skewness are introduced to guide the sensor data generation. Regarding customized approaches for specific scenarios, we address both active (explicit context) and passive (implicit context) situations. Explicit context typically includes information willingly shared while implicit context may encompass data collected passively, with users potentially unaware of the full extent of information usage. Segregating explicit and implicit context aims for a balance between personalization and privacy, empowering users with enhanced control over their information and ensuring adherence to privacy regulations. In active scenarios, we focus on privacy protection in pervasive surveillance. We propose Point-Former, the example-guided modification of motion in point cloud to translate from default motion and gesture interaction alphabets to personal ones, to safeguard privacy during gesture interactions in pervasive space. In the passive scenario involving implicit context, we consider on-body devices and environmental devices. For on-body devices, we introduce \textbf{CardioID}, an interaction-free device pairing method that generates body-implicit secure keys by exploiting the randomness in the heart's operation (electrocardiogram (ECG) or ballistocardiogram (BCG) signals). For environmental smart devices, we propose GIHNET, a low complexity and secure GAN-based information hiding method for IoT communication via an insecure channel. It hides the original information using meaningless representations, by obscuring it beyond recognition. Building on GIHNET, we extend the use of data encryption and propose SIGN, which converts signatures into a Hanko pattern and uses it as an encryption method to generate digital signatures in pervasive spaces.
  • Capacity Planning for Vehicular Fog Computing
    (2023) Mao, Wencan
     School of Electrical Engineering | Doctoral dissertation (article-based)
    The strict latency constraints of emerging vehicular applications make it unfeasible to forward sensing data from vehicles to the cloud for processing. Fog computing shortens the network latency by moving computation close to the location where the data is generated. Vehicular fog computing (VFC) proposes to complement stationary fog nodes co-located with cellular base stations (i.e., CFNs) with mobile ones carried by vehicles (i.e., VFNs) in a cost-efficient way. Previous works on VFC have mainly focused on optimizing the assignments of computing tasks among available fog nodes. However, capacity planning, which decides where and how much computing resources to deploy, remains an open and challenging issue. The complexity of this problem results from the spatio-temporal dynamics of vehicular traffic, the uncertainty in the computational demand, and the trade-off between the quality of service (QoS) and cost expenditure. This dissertation focuses on capacity planning for VFC. The objective of capacity planning is to maximize the techno-economic performance of VFC in terms of profit and QoS. To address the spatial-temporal dynamics of vehicular traffic, this dissertation presents a capacity planning solution for VFC that jointly decide the location and number of CFNs together with the route and schedule of VFNs carried by buses. Such a long-term planning solution is supposed to be updated seasonally according to the traffic pattern and bus timetables. To address the uncertainty in the computational resource demand, this dissertation presents two capacity planning solutions for VFC that dynamically schedule the routes of VFNs carried by taxis in an on-demand manner. Such a short-term planning solution is supposed to be updated within minutes or even seconds. To evaluate the techno-economic performance of our capacity planning solutions, an open-source simulator was developed that takes real-world data as inputs and simulates the VFC scenarios in urban environments. The results of this dissertation can contribute to the development of edge and fog computing, the Internet of Vehicles (IoV), and intelligent transportation systems (ITS).
  • Characterization of Predictable Quantum Efficient Detector
    (2023) Korpusenko, Mikhail
     School of Electrical Engineering | Doctoral dissertation (article-based)
    The aim of this thesis is to develop new characterization methods for Predictable Quantum Efficient Detector (PQED) at visible wavelengths and to extend the predictable responsivity range of PQED into short visible and ultraviolet (UV) spectral ranges. The thesis presents optical studies of PQEDs made of p- and n-type photo diodes, including determination of responsivity, reflectance, spatial uniformity, and bias-voltage dependent characteristic in expanded spectral range compared to previous studies. Investigations of internal quantum efficiency of silicon detectors were done at UV spectral range. A theoretical model for quantum gain in silicon photodiodes was developed for short visible and UV spectral ranges. The n-type PQED with Al2O3 layer coating was optically characterized at short and long visible wavelengths. Its responsivity was obtained through comparison measurements against reference p-type PQED. Reflectance losses were measured and compared with simulated values and found to agree. With these measurements n-type PQED responsivity is predictable in visible range. Novel p-type PQEDs with SiNx surface layer were characterized against reference p-type PQED. It was found out that SiNx PQEDs have excellent spatial uniformity and at least as high responsivity as the used reference PQED. Responsivity of PQED and Hamamatsu trap detector was studied in UV range by comparison measurements against room temperature pyroelectric radiometer. With predicted reflectance and recombination losses, absolute value of quantum gain was retrieved from measured responsivity of PQED. Based on measured data of quantum yield, a theoretical model for quantum yield was developed for short visible and UV wavelengths. Calculated quantum gain has very good agreement with measured values at short visible wavelengths. Separation of contributions of quantum gain and reflectance in responsivity allowed to estimate recombination losses of Hamamatsu photodiodes that was not reported before. A new way of analysing PQED photocurrent dependence on bias voltage was also proposed. Such data can be used for fitting of a 3D charge-carrier transport model which was utilized in a separate work to predict fundamental parameters of PQED photodiodes and to determine spectral responsivity with an unprecedented accuracy.
  • Connectivity for smart grids: Novel communications solutions in evolving electrical grids
    (2024) Borenius, Seppo
     School of Electrical Engineering | Doctoral dissertation (article-based)
    Increasing electrification and power grid evolution to allow integrating large amounts of renewable generation will be key enablers for creating a sustainable, carbon-neutral energy system. The expected increased demand for electricity will result from efforts to reduce the use of fossil fuels in the industry, heating, and transport sectors. The larger share of intermittent generation from renewable sources, such as wind and solar, as well as the decreasing number of traditional controllable inertia-providing generators will lead to higher system volatility. This volatility challenge is worsened by the lack of seasonal system-level electric energy storage capacity. In distribution grids, the volatility challenge and increased power system dynamics will necessitate expanded automation, which in turn will require enhanced connectivity solutions. This thesis contributes first by taking a forward-looking, system-level view in exploring possible power grid futures and then by identifying approaches for integrating electric power grids with Information and Communications Technologies (ICT) in these futures. The overall research problem of the thesis is defined as follows: How can communications solutions support the creation of sustainable resilient power grids by the 2030s? The research extensively utilises expert panels, formal scenario planning and value networks based on the Finnish power grid context as a case example. The thesis proceeds in three stages. The thesis first establishes multiple scenarios, i.e. possible power grid futures. These describe the potential evolution from the perspective of grid management and the services offered to customers. Thereafter, in the second stage, the thesis explores the role of both the latest as well as anticipated new communication technologies, the feasibility of applying these in future distribution grids, and the potential impact of softwarisation on power grid architectures. The more extensive use of ICT gives rise to new attack points for malicious actors and consequently increases the vulnerability of the electric energy system. The thesis continues by identifying the most significant cybersecurity risks and trends, followed by an examination of how well these risks and trends are currently analysed and understood in academia and industry. In the third stage, the thesis shifts the focus to the business level. The opportunities for various actors are explored by identifying the potential industry (business) architectures for the communications solutions required to manage future distribution grids. The results of this thesis should help stakeholders, such as actors within the energy and ICT sectors as well as regulators and politicians, to consider alternative futures in order to make correct decisions on which businesses to be in, how to invest until the 2030s, as well as how to ensure the reliability and cost efficiency of the electric power system.
  • Data-driven room-acoustic modelling
    (2024) Götz, Georg
     School of Electrical Engineering | Doctoral dissertation (article-based)
    The study of room acoustics has traditionally been of interest in architectural planning and design. With the spread of virtual- and augmented-reality technology, room-acoustic modelling has also become increasingly relevant for audio engines. The dynamic and fast-paced nature of such applications requires rendering systems to operate in real-time. However, accurate state-of-theart room-acoustic-simulation technology is often computationally expensive, limiting its use for audio engines. Data-driven methods offer the potential to bypass expensive simulations, while ensuring convincing perceptual experiences. This dissertation works towards data-driven audio engines by exploring the interaction between room-acoustic modelling and data-driven methods. It comprises five peer-reviewed publications that investigate automatic data acquisition, robust room-acoustic analysis in complex environments, and data-driven room-acoustics rendering. As sound propagates through a room, it interacts with various surfaces, leading to a gradual energy decay over time. The properties of this energy decay significantly influence the acoustic impression evoked by a room, making it a widely studied topic in room-acoustic research. The first part of this thesis provides an overview of sound-energy decay, its analysis, and challenges associated with complex geometries featuring multiple rooms and non-uniform absorption-material distributions. To this end, it introduces a neural network for multi-exponential sound-energy-decay analysis. Moreover, spatial and directional variations of sound-energy decay are investigated, and a compact representation to model them is proposed. The second part of this thesis is centred around data-driven methods and explores how they can be applied to room-acoustics research. After elaborating on the properties of room-acoustic data, techniques for its large-scale acquisition are investigated. Two of the contained publications describe autonomous robot systems for conducting room-acoustic measurements. While the first one describes the general idea and the design constraints of a practical system, the second one extends the measurement strategy to complex geometries featuring multiple connected rooms. An overview of commonly used machine-learning concepts is provided, focusing on the ones relevant for the included publications. Finally, several applications of data-driven methods in roomacoustics research are described, including a summary of a late-reverberation rendering system proposed in one of the appended publications.
  • Enhancing Latency Reduction and Reliability for Internet Services with QUIC and WebRTC
    (2024) Li, Xuebing
     School of Electrical Engineering | Doctoral dissertation (article-based)
    This dissertation introduces innovative systems and algorithms aimed at improving latency reduction and enhancing reliability for Internet services. To structure the research effectively, the dissertation categorizes the lifecycle of a service connection into three stages: service discovery, delivery, and migration. In the context of service discovery and migration, edge computing employs server replication to achieve low latency via user proximity and high reliability through load balancing. The primary challenges lie in developing a mechanism for rapid connection establishment and consistent optimal end-user mapping. A focal point of this study is the examination of QUIC and its potential to enhance service discovery and migration. By integrating QUIC's handshake data flow with conventional service discovery protocols, two systems are proposed to reduce the latency in connection setup and enhance the effectiveness of end-user mapping, with a particular emphasis on leveraging anycast and Domain Name System (DNS) technologies. For anycast, a novel approach in cloud computing is introduced to anycast routing, incorporating enhanced capabilities for name resolution and load awareness. In DNS, a middleware solution in the 5G core network improves performance, notably in query delay, cache hit rates, and consistency, thereby refining DNS-based discovery in edge cloud computing. Furthermore, along with essential server-side modifications, the solution extends QUIC's zero-round trip time (0-RTT) handshake feature to facilitate 0-RTT service migration, significantly boosting migration efficiency. Regarding service delivery, the data generation and transmission behavior is governed by the system and network capabilities. The challenge resides in designing a control algorithm to ensure consistent low-latency and reliable packet delivery, facilitating the application requirements. This dissertation concentrates on optimizing delivery control mechanisms within real-time video streaming, using WebRTC as the testbed. It provides an exhaustive analysis of how control parameters affect streaming performance and application metrics, leading to the development of an algorithm for optimizing the parameter setting during the slow start phase of congestion control. Furthermore, a machine learning based streaming control solution is proposed to jointly control multiple parameters, serving as a more general solution. This work also introduces an open-source framework designed to facilitate future research on applying machine learning to WebRTC control. Aiming to improve latency and reliability, this dissertation investigates the integration of emerging technologies such as QUIC, 5G, and machine learning within the established framework of edge computing. This research emphasizes the importance of cross-layer design in the optimization of Internet services, identifying machine learning as a promising approach.
  • Generalized Accelerated Optimization Framework for Big Data Processing
    (2024) Dosti, Endrit
     School of Electrical Engineering | Doctoral dissertation (article-based)
    Large-scale optimization problems arise in different fields of engineering and science. Due to the large number of parameters and different structures that these problems can have black-box first-order methods are widely used in solving them. Among the existing firstorder methods, the ones that are most widely used are different variants of Fast Gradient Methods (FGM). Such methods are devised in the context of the estimating sequences framework and exhibit desirable properties such as fast convergence rate and low per iteration complexity. In this Thesis, we devise new estimating sequences and show that they can be used to construct accelerated first-order methods. We start by considering the simplest case, i.e., minimizing smooth and convex objective functions. For this class of problems, we present a class of generalized estimating sequences, constructed by exploiting the history of the estimating functions that are obtained during the minimization process. Using these generalized estimating sequences, we devise a new accelerated gradient method and prove that it converges to a tolerated neighborhood of the optimal solution faster than FGM and other first-order methods. We then consider a more general class of optimization problems, namely composite objectives. For this class of problems, we introduce the class of composite estimating sequences, which are obtained by making use of the gradient mapping framework and a tight lower bound on the function that should be minimized. Using these composite estimating sequences, we devise a composite objective accelerated multi-step estimating sequence technique and prove its accelerated convergence rate. Last, embedding the memory term coming from the previous iterates into the composite estimating sequences, we obtain the generalized composite estimating sequences. Using these estimating sequences, we construct another accelerated gradient method and prove its accelerated convergence rate. The methods devised for solving composite objective functions that we introduce in this thesis are also equipped with efficient backtracking line-search strategies, which enable more accurate estimates of the step-size. Our results are validated by a large number of computational experiments on different types of loss functions, wherein both simulated and publicly available real-world datasets are considered. Our numerical experiments also highlight the robustness of our newly introduced methods to the usage of inexact values for of the Lipschitz constant and the strong convexity parameter.
  • HENGITYSTIEAEROSOLIPÄÄSTÖJEN MITTAAMINEN JA HUOMIOINTI TARTUNTATAUTIEN EHKÄISYSSÄ
    (2023) Saari, Sampo; Tuhkuri Matvejeff, Anna; Silvonen, Ville; Heikkilä, Paavo; Hakala, Jani; Sanmark, Enni; Oksanen, Lotta-Maria; Rönkkö, Topi; Taipale, Aimo; Laukkanen, Anne-Maria; Alku, Paavo; Geneid, Ahmed
     A4 Artikkeli konferenssijulkaisussa
    Patogeenien aerosolileviäminen on monimutkainen tapahtuma, jossa on paljon tekijöitä, jotka vaikuttavat infektioriskiin. Yksi tärkeä muuttuja on hengitysteistä vapautuvien aerosolihiukkasten määrä ja ominaisuudet. Tässä esityksessä keskitytään hengitysteistä syntyvien aerosolihiukkaspäästön mekanismeihin, dynamiikkaan, mittauksiin, alustaviin tuloksiin sekä niiden huomiointiin tartuntatautien ehkäisyssä. Kokeellista tutkimusta varten kehitettiin uusi siirrettävä mittausjärjestelmä, jolla pystyttiin mittaamaan koehenkilöiden hengitystieaerosoleja reaaliaikaisesti laajalla hiukkaskokoalueella. Laitteiston avulla pystyttiin mittaamaan absoluuttisia hiukkaspäästökertoimia koehenkilöille erilaisissa harjoitustilanteissa (mm. puhe, laulu, kuiskaus, yskiminen).
  • Human-in-the-Loop Design Optimization
    (2023) Liao, Yi-Chi
     School of Electrical Engineering | Doctoral dissertation (article-based)
    This dissertation presents novel computational methods and investigations to enable human-in-the-loop optimization (HILO) for a wider range of realistic applications, allowing designers to efficiently explore the design space of practical problems. Designing effective interaction techniques requires careful consideration of various parameters, that significantly impact user experience and performance. However, optimizing these parameters can be challenging due to the large, multi-dimensional design space, the unclear relationship between parameter settings and user performance, and the complexity of balancing multiple design objectives. Traditionally, designers perform manual optimization via iterative design processes, which can be time-consuming, and effortful, and does not guarantee the best outcome. HILO emerged as a more principled solution for design optimization, using a computational optimizer to intelligently select the next design instance for user testing. Despite some examples of HILO in the human-computer interaction (HCI) field, its application scope is limited to a single objective and for a single user. How to extend it for handling multi-objective problems, optimizing for a population, and supporting physical interfaces has remained unclear. Furthermore, conducting HILO does not eliminate the costs arising from human involvement, and practitioners have been reluctant to embrace a technique whose positive and negative qualities are poorly understood. This dissertation presents a set of computational methods and investigations that speak to these challenges. Pareto-frontier learning is utilized to handle multi-objective design tasks, and I introduce novel extensions for practical solutions of group-level Bayesian optimization. To reduce the effort and time in prototyping, I propose using physical emulation to render physical design instances, enabling HILO to be applied to the design of physical interactions. The dissertation presents user experiments and a design workshop conducted to enrich the understanding of Bayesian optimization-supported design processes' strengths and limitations. Finally, in light of the resource-intensive nature of user studies, a simulation-based optimization framework is proposed whereby artificial users evaluate design instances. With the ultimate goal of expanding HILO's utility in realistic and general design tasks, this dissertation opens new directions for future HILO research. One important path for exploration involves more advanced optimization techniques, such as methods that enable greater efficiency and support a high-dimensional design space. The project also spotlights the value of investigating better human-machine collaboration mechanisms in design optimization such that the designers can steer the optimization as required or fine-tune the suggestions proposed by the optimizer. Lastly, simulation-based optimization methods require further validation, and developing human-like models will be a crucial next step.
  • Improving Live Video Streaming Performance for Smart City Services
    (2024) El Marai, Oussama
     School of Electrical Engineering | Doctoral dissertation (article-based)
    Our world is rapidly moving in all its aspects toward a more digitized and connected life, including transportation, education, farming, and healthcare. A major enabler for such transformation is ICT-related tremendous innovations in networking, computation, and storage, both in software and hardware at affordable prices. Owing to these phenomenal advances, many revolutionary paradigms, such as multi-access edge computing, self-driving vehicles, and Smart Cities, have emerged, promising rosy prospects and a flourishing future. An eminent feature of these futuristic technologies is automation, where objects can communicate (i.e., sending and receiving data), understand their environment, and adapt to changing conditions by taking the right decisions. Also, stringent requirements (e.g., low latency communication) might be needed by many services for their proper functioning. To successfully accomplish these tasks, many paradigms (e.g., software-defined networking and machine learning techniques) should be involved at different levels (e.g., network and decision-making levels). Most of today's applications and systems (e.g., over-the-top and surveillance platforms) require video streaming as a key technology. Video streaming applications rank as the most bandwidth-intensive services, especially when delivered at higher resolutions, such as FHD and 4K. Fortunately, 5G technology is already available and promises higher bandwidth that can reach up to 20GB. In addition, it requires huge data storage spaces when historical data is needed, which no longer becomes an issue with the dawn of edge and cloud computing. The target consumer (i.e., humans or machines) might demand heavy computation resources, often requiring GPU processing, which is also nowadays readily available and affordable. This dissertation is all about harnessing video streaming technology for enabling Smart City services and paradigms, such as self-driving vehicles. Towards this end, we start by addressing the problem of improving video streaming performance in terms of delivered video quality, stall-free sessions, and low latency streaming, for various services, including video streaming services and some use cases of self-driving vehicles. As data is the fuel that empowers most Smart City systems and services, we propose a cost-efficient and sustainable solution to create the digital twin of city roads, which mainly relies on video streaming data. The proposed solution represents an essential step towards realizing the Smart City paradigm and would create a valuable data asset that feeds and benefits various systems and domains such as intelligent transportation systems and tourism. Owing to the extreme importance of situational awareness in Smart Cities, notably in dense urban areas, we leverage the proposed digital twinning solution and machine learning techniques to raise the awareness of connected vehicles about their surroundings, as well as overall street awareness per defined regions while accounting for the amount of transmitted data over the network to avoid video streaming performance degradation.
  • IoT and DLT Integration—A Choice of Tradeoffs?
    (2023) Paavolainen, Santeri
     School of Electrical Engineering | Doctoral dissertation (article-based)
    The integration between Internet of Things (IoT) systems and Distributed Ledger Technologies (DLTs) seems to offer a possibility to address some of the shortcomings often encountered in the widespread deployment of IoT systems, as well as open a potential for novel business models. Existing research on IoT-DLT integration has, however, focused primarily on addressing functional problems on a high level and leaving many of the operational problems occurring in a real-world scenario out of scope. For example, a Raspberry Pi single board computer is often used as an analogue for an IoT device—even when it costs ten or hundred times more than an embedded processor inside a low-cost IoT device. Crucially, the cost of an IoT device is a major factor in the economic feasibility at industrial scale. This cost pressure implies that most IoT devices will be relatively cheap and as a consequence have only a meager amount of computing power, memory capacity, and network bandwidth, commonly referred to as constrained devices. Thus, it is important to consider not only macroscopic use cases, but to also address challenges low-cost and constrained devices face if we really want to enable DLT connectivity on a typical IoT device.  This dissertation describes different integration approaches used for IoT-DLT systems, and qualifies their applicability for constrained devices. Of particular importance is an integration method based on light protocols, which provide an enticing tradeoff of providing relatively high security while requiring substantially less resources as operating as a normal, fully functional peer on the DLT. Yet, some of these security tradeoffs can be shown to be worse than commonly assumed, leading to IoT devices being vulnerable to state injection attacks. This work proposes two new novel solutions to address such attacks: decentralised beacons and subset nodes. Decentralised beacons leverages on the existence of a trusted third party in IoT systems—the device owner—to provide scalable attestations of the DLT ground state to a low-power device, with a tradeoff of increased latency to DLT state changes. Subset nodes, in turn, addresses the latency issue by recognizing that most IoT applications will observe only a small subset of the whole DLT state, and by restricting its view to only this subset state, a higher level of security assurances can be reached with modest computing and storage requirement increases. These two methods are complementary and can be deployed separately or in combination.
  • Machine Learning based GNSS Spoofing Detection and Mitigation for Cellular-Connected UAVs
    (2023) Dang, Yongchao
     School of Electrical Engineering | Doctoral dissertation (article-based)
    Cellular-connected Unmanned Aerial Vehicle (UAV) systems are a promising paradigm in the upcoming 5G-and-beyond mobile cellular networks by delivering numerous applications, such as the transportation of medicine, building inspection, and emergency communication. With the help of a cellular communication system and Global Navigation Satellite System (GNSS), UAVs can be deployed independently or collectively in remote and densely populated areas on demand. However, the civil GNSS service, especially GPS, is unencrypted and vulnerable to spoofing attacks, which threatens the security of remotely- and autonomously-controlled UAVs. Indeed, a GPS spoofer can use commercial Software-Defined Radio (SDR) tools to generate fake GPS signals and fool the UAV GPS receiver to calculate the wrong locations. Fortunately, the 3rd Generation Partnership Project (3GPP) has initiated a set of techniques and supports that enable mobile cellular networks to track and identify UAVs in order to enhance low-altitude airspace security. The research works in this thesis leverage the potential of machine learning methods and 3GPP technique support to detect and mitigate GPS spoofing attacks for cellular-connected UAVs. The contributions of this thesis contain four parts. First, we propose a new adaptive trustable residence area algorithm to improve the conventional Mobile Positioning System (MPS) in terms of GPS spoofing detection accuracy under three base stations. Then, we deploy deep neural networks in the Multi-access Edge Computing (MEC) based 5G-assisted Unmanned Aerial System (UAS) for detecting GPS spoofing attacks, where the proposed deep learning methods can detect GPS spoofing attacks with only a single base station. Next, we analyze the max-min transmission rate for cellular-connected UAVs system theoretically and design an optimal Graphic Neural Network (GNN) to detect GPS spoofing attacks for cellular-connected UAV swarms. Finally, we employ a 3D radio map and particle filter to recover the UAV position and mitigate GPS spoofing attacks.