Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Renewable energy resources will form the cornerstone of power systems of the future and present a key solution to the grand challenge of mitigating climate change. However, the proliferation of renewables has brought up a myriad of challenges involved in operating the power system with notable shares of inverter-based resources (IBRs). This has sparked several concerns pertaining to stability, protection, and continuity of power supply at large. As a solution to a wide body of such concerns, a consensus is forming toward the adoption of grid-forming (GFM) inverter technology to interface renewables with the grid. In the GFM paradigm, IBRs do not follow the grid, rather, they form it and offer voltage and frequency regulation much alike conventional synchronous generators. Yet, their performance during --- in particular unbalanced --- faults and their impact on power system protection remains largely uncharted territory. The present work investigates the (unbalanced) fault behavior of grid-forming inverters, and their effect on traditional distance protection. First, a novel grid-forming inverter control architecture is proposed which exhibits better performance during faults in terms of grid support. Enhanced virtual-impedance and current-reference saturation limiting strategies are introduced, along with an elaborate and generic control-design procedure. Electromagnetic transient simulations in MATLABextsuperscript{extregistered}Simulink of an all-inverter IEEE 14-bus network illustrates the voltage support benefits achievable with the proposed GFM controls. Second, the effect of grid-forming inverter controls on the performance of traditional distance protection is investigated. Different incarnations of GFM controls are taken into consideration for this analysis. Full-order EMT simulations, analytical steady-state fault models, and a hybrid hardware-software setup illustrate the superiority of the proposed GFM controls in terms of protection. Finally, a novel method is proposed for detecting a fault on a line connected to IBRs. The method is independent of the inverter's control structure and does not require high-bandwidth communication. As such, the method is agnostic to the presence of inverter-based resources in the grid.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

The 2010s have seen the first large-scale successes of computer vision "in the wild", paving the way for industrial applications. Thanks to the formidable increase of processing power in consumer electronics, convolutional neural networks have led the way in this revolution. With enough supervision, these models have proven able to surpass human accuracy on many vision tasks. However, rather than focusing exclusively on accuracy, it is increasingly important to design algorithms that operate within the bounds of a computational budget - in terms of latency, memory, or energy consumption. The adoption of vision algorithms in time-critical decision systems (such as autonomous driving) and in edge computing (eg{} in smartphones) makes this quest for efficiency a central challenge in machine learning research.How can the optimization of existing models be improved, in order to reach higher accuracy without affecting the processing requirements? Alternatively, can we search for models that fit the processing requirements while improving the accuracy on the task? In this thesis, we consider both of these questions, which are two sides of the same coin. On one hand, we develop novel methods for learning model parameters in a supervised fashion, improving the accuracy on the target task without affecting the efficiency of these models at test-time. On the other, we study the problem of model search, where the model itself must be selected among a family of models in order to achieve satisfactory accuracy under the resource constraints.Chapter 3 introduces the probably submodular framework for learning the weights of pairwise random graphical models. Graphical models are expressive and popular models, used notably in semantic segmentation. However, their inference is NP-hard in general. In order to ensure efficient inference, it is necessary to constrain the weights learned during training. Popular tractability constraints are definitely submodular constraints; they ensure that the local potential functions of the model are submodular for any input at test-time. We show that these constraints are often too conservative. Rather than enforcing that the graphical model is submodular for any input graph, it is sufficient to ensure submodularity with high probability for the data distribution of the task. We show on several semantic segmentation and multi-label classification datasets the superiority of this approach, validating the corresponding gain in model expressivity and accuracy, without compromising the efficient inference at test-time.Chapter 4 presents improved optimization methods to reduce the test-time error of semantic segmentation models, by introducing novel task-specific losses. In recent years, convolutional neural networks have dominated the state of the art in semantic segmentation. These networks are usually trained with a cross-entropy loss, which is easy to use within first-order optimization schemes. However, segmentation benchmarks are usually evaluated under other metrics, such as the intersection-over-union measure, or Jaccard index. A direct optimization of this measure, while challenging, can yield a lower error rate. Such gains are relevant to applications, as the Jaccard index has been shown to be closer to human perception, and benefits from scale invariance properties. Using the Lovász extension of submodular set functions, we develop tractable surrogates for the optimization of the Jaccard index in the binary and multi-label settings, compatible with first-order optimizers. We demonstrate the gains of our method in terms of the target metric on binary and multi-label semantic segmentation problems, using state-of-the art convolutional networks on the Pascal VOC and CityScapes datasets.Chapter 5 considers the problem of neural architecture search, where one wants to select the best-performing model satisfying the computational requirements among a large search space. We aim to adjust the channel numbers of a given neural network architecture, ie{} the number of convolutional filters in each layer of the model. We first develop a method to predict the latency of the model given its channel numbers, using a method relying on least-square estimation of the predictor without the need to access low-level details about the computation on the inference engine. We then build a proxy for the model error that decomposes additively over individual channel choices, by using aggregated training statistics of a slimmable model on the same search space. The association of the pairwise latency model and the unary error estimates leads to an objective that can be optimized efficiently using the Viterbi algorithm, yielding the OWS method. A refinement of OWS, named AOWS, adaptively restricts the search space towards optimal channel configurations during the training of the slimmable network. We validate our approach over several inference modalities and show improved final performance of the selected models within given computational budgets.Overall, this thesis proposes novel methods for improving the accuracy/efficiency tradeoff of contemporary machine learning models, using methods derived from first principles and validated through experimentation on several contemporary computer vision problems. This research paves the way towards a smarter usage of the computational resources of machine learning methods, curbing the trend for "wider and deeper" models in order to face the challenges of time-critical and carbon-neutral AI.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Colour Appearance Models (CAM) attempt to predict the colour appearance of a stimulus by taking the physical properties of the stimulus and its surroundings into account. The fundamental goal is to look for correlates between the measured optical spectral data of a stimulus and its surrounding and the corresponding perceptual attributes. There are three absolute colour attributes (Brightness, Colourfulness and Hue) and three relative colour attributes (Lightness, Chroma and Saturation).Most of the existing CAMs are developed to describe the perception of related surface colours which implicitly assume the presence of a light source illuminating the target and one or more other surfaces. One of them, CIECAM02 is recommended by the "Commission Internationale d'Eclairage" (CIE). Recently, a new CAM (CAM15u) was developed for unrelated light sources (i.e. self-luminous stimuli seen in a completely dark environment). However, at this moment, there is no colour appearance model for light sources seen in a specific luminous context (e.g. a traffic signal at daytime).A first crucial step is to gather relevant optical data (spectral radiances) and the corresponding perceptual data of the colour attributes of observers. With this new data a new and comprehensive colour appearance model will be developed for these self-luminous stimuli seen in relation to a self-luminous background using the spectral optical data of the stimulus and the background as basic input, and taking into account as much as possible input from the physiologic and neurologic processes of the visual system. This new model can immediately be used when investigating the brightness perception of LED-signalisation.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

The goal of this research was to develop and test integrated CMOS circuits for radiation tolerant time-based circuits with picosecond accuracy for nuclear applications and high-energy physics. The main applications for which these circuits were developed, are time-based readout interfaces in high-energy physics particle detectors, clock generation and data-transmission for these detectors. During this research, a radiation tolerant Time-to-Digital Converter (TDC) and a low noise clock synthesizer were designed and optimized for the particle detectors at CERN.A short overview on the radiation effects and mitigation techniques to ionizing radiation is given, together with a discussion on the practical aspects which are required in modern TDCs and frequency synthesizers.A high-resolution TDC is presented with a discussion on the design aspects and the practical implementations of the circuit which are required in nuclear environments. The TDC is based on a Delay-Locked Loop (DLL) that has two phase detection circuits to boost the recovery time after an energetic particle disturbs the circuit. The functionality of the DLL ensures that the timing resolution of the TDC remains the same after irradiation. Furthermore, this DLL has a new phase detector architecture which reduces static-phase offsets in the phase detectors through a correlated sampling mechanism which has been implemented for the first time in the time domain. The circuit was prototyped in a 40 nm CMOS technology and a 4.8 ps resolution was measured with a 4.2 mW power consumption.DLL based TDCs or serial communication links do require a low-noise, high-frequency reference clock. For a 64 channel TDC, a 2.56 GHz frequency synthesizer was designed to upconvert the 40 MHz reference clock of the Large Hadron Collider (LHC) at CERN to a 2.56 GHz high speed clock with a targeted rms jitter below 1 ps. A radiation hardened Phase Locked Loop was designed in which both an LC-tank oscillator and ring oscillator were present. The chip was prototyped in a 65 nm COS technology. These circuits were, in a next step, irradiated to make a comparison between ring and LC-tank oscillators in terms of noise, radiation damage and single-event effects. The devices were irradiated with X-rays up to 600 Mrad to study the Total Ionizing Dose effects on the circuits and were also irradiated with heavy-ions to study the single-event effects on the oscillators. The clock generator has a power consumption of 11.7 mW and had an integrated rms jitter of only 345 fs. Triple Modular Redundancy was used in the digital circuits to protect them from soft errors. A new phase detector architecture is presented which minimizes the error rate due to high-energy particles in frequency synthesizers. The devices were also tested for temperature variations from -25 °C up to 125 °C.From the results gathered in the radiation experiments, an improved LC-tank oscillator was designed which has a reduction of the sensitivity to single-event upsets of more than 600 times compared to a traditional implementation which is mainly due to the cross section of the tuning varactor of the oscillator. This technique was also experimentally verified.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Remote phosphor plates are increasingly used in LED modules to obtain auniform white light distribution by converting part of the blue LED light into light with a longer wavelength. Remote phosphor plates also havethe potential to improve the system performance of display backlights if luminescent materials with a narrow emission band (e.g. quantum dots) are considered.The goal of this research is to find new concepts and improved configurations for efficient illumination devices that incorporate remote phosphor convertors. This is done by using accurate simulations of the phosphor plate based on detailed characterization of the relevant material properties. In a second stage, the converter is integrated into a complete device and the total system performance is investigated. The resulting simulation models are validated by comparing them with real demonstrators. In this way new configurations with desiredproperties can easily be explored.