Search

Digital rock physics (DRP) er en moderne metode for å karakterisere de fysiske egenskapene til ulike typer stein. Ved å modellere korn-, flerfaset- og porevolum i forksjellige bergarter, kan en assistere institusjoner med å finne egnede lokasjoner for lagring av hydrokarboner. Utover dette kan nøyaktige modelleringer av strukturen i bergarter hjelpe olje- og gassindustrien med å velge potensielle lokasjoner for utbygging av oljefelt. DRP blir fanget opp ved å benytte micro-CT for å produsere 2D eller 3D bilder. I nyere tid har dype læringsmetoder blitt effektive på å oppnå gode resultater på oppgaver som tidligere kun kunne løses av mennesker. Semantisk bildesegmentering er en datasyns oppgave med mål om å assosiere hver pixel-verdi i et bilde med en klasse. I denne avhandlingen utforsker vi å anvende deep convolutional encoder-decoder-nettverk til å utføre automatisk semantisk bildesegmentering av gråtonebilder av bergarter. Dataen vi har jobbet med består av tre forskjellige bergarter, Bentheimer, Berea og Carbonate. Vi presenterer eksperimenter der vi kombinerer dataaugmenterings-teknikker for ytterligere å utvide tilgjengelig data. Resultatene fra de forskjellige arkitekturene blir sammenlignet med hverandre, og modellen som gir best resultat oppnår 88.6% intersection over union (IoU). Den beste arkitekturen utvides til å inkludere flere bilder som ligger ved siden av hverandre, for å gi nettverket mer informasjon. Vi undersøker også om et nettverk trent på en eller to bergarter kan generaliseres og oppnå gode resultater på en bergart som ikke var inkludert i treningsprosessen. Digital rock physics (DRP) is a modern approach for characterizing physical rock properties. By modeling grain, multi-phase and pore space in different rock types, institutions can be assisted in locating suitable locations for hydrocarbon storage. Furthermore, producing accurate models of rock structures can economically assist the oil and gas industry in selecting potential reservoir extraction sites. DRP is captured using non-destructive imaging technology, micro computed tomography (micro-CT), producing 2D or 3D images. Deep learning has in recent years become very successful at accurately and efficiently performing tasks that previously could only be accomplished by humans. Semantic image segmentation is a computer vision task with the goal of assigning class labels to each pixel in an image. This thesis explores the application of multiple deep convolutional encoder-decoder networks to perform semantic image segmentation of gray-scale rock images. The provided datasets consist of three different types of sandstone, Bentheimer, Berea and Carbonate. We present experiments of combining data augmentation techniques to further increase the richness of the dataset. Results of the different architectures are compared to each other, and the best performing model achieves an intersection over union (IoU) score of 88.6% (with a categorical accuracy of nearly 98%). The best performing network is further extended to include multiple adjacent images as channel inputs to provide additional information. We investigate if a network trained on one or two rock types can generalize and perform accurate predictions on one not included in training.

A Smart Transformer (ST) is a power electronics based transformer developed for the electrical distribution grid, where it can provide many advanced grid services. The advanced functionality comes with the concern of lower reliability affected by the complexity and the higher number of components. To overcome the reliability concerns, this work proposes to utilize a modular architecture for routing the power in the cells such that cells with shorter remaining lifetimes are less stressed and the time to the next failure in the system is delayed.

The Smart Transformer (ST) is a possible solution to obtain intelligent nodes in the electrical grid, which can be used for the grid management and increase the capacity for the integration of renewable energy sources. A problem for the application of the ST in the distribution grid is the expected lower reliability in comparison with the traditional transformer. To address this problem, the knowledge of power system, power electronics and reliability is combined in this work. Following the "Physics of failure" approach, the most frequently failing components are identified, their load profile in the electrical distribution grid is analyzed and finally solutions are developed to improve the reliability. The power semiconductors are found to be the most prone to fail components and most of their failure mechanisms are found to be affected by thermal cycling. For this reason, thermal stress analysis is performed for the three-stage ST. As an opportunity to increase the reliability, active thermal control is introduced, which is a software based solution for the reduction of the thermal stress during operation. The existing approaches from literature are reviewed and categorized into control of the power converter losses and the control of the device loading. For increasing the reliability by control of the power converter losses, one algorithm is introduced and validated for hard switching power converters and one algorithm is introduced for soft switching power semiconductors. Controlling the thermal stress of modular building blocks in a modular power converter, referring to power routing, is proposed. The capability of the algorithm is investigated analytically for series connected and parallel connected modular building blocks. For the validation, the influence of the power routing on the loading of the single cell is demonstrated experimentally for series connected, parallel connected and medium frequency transformer coupled cells in modular power converters.