Gas turbine engine efficiency needs to be raised in order to decrease fuel consumption, greenhouse gas emissions, and expenses. Efficiency may be improved in two ways: by reducing engine weight and raising intake temperatures. At intake temperature, conventional nickel-based alloys are already on the verge of failure, meaning there is a need and demand of materials which can withstand higher temperatures. Silicon Carbide Ceramic Matrix Composites (SiC CMCs) are being investigated as a potential replacement for superalloys due to their superior physical properties, such as their low weight and high melting point (approximately one-third of superalloys' weight). However, using SiC CMCs has a serious disadvantage. The mass recession of the SiC is caused by the volatilization of silicon hydroxide, which is caused by oxidation and reactivity with water vapor under the working conditions of gas turbine engines. Therefore, a shielding layer is used to prevent oxidation of the SiC CMCs. This protective coating (EBC) goes by the name of Environmental Barrier Coating. Thermal spray techniques such as atmospheric plasma spray and suspension plasma spray, which employ powder as the feedstock, are used to deposit EBC on SiC CMCs. For EBC to perform well, the coating must be crystalline, reasonably thick to sustain harsh environment, and devoid of cracks. EBC was deposited in order to look at how the spray parameters affected the microstructure. SEM pictures were used to quantify the coating's porosity and the severity of the cracks. To investigate the production of thermally grown oxide (TGO) in the coating and substrate and check how EBCs perform under thermal cyclic fatigue loading, a thermal cyclic fatigue test was conducted. The XRD analysis is performed to ascertain the proportion of crystalline and amorphous phases in the coating, which unfortunately is still in the process to be completed. In the as-sprayed coating samples we can see that when there are more amount and larger pores, we see less number of cracks and vice versa. The effect of spray parameters can be seen on the coatings. Comparing to SPS trial 1, the SPS trial 2 coatings are denser with less number of cracks and has good adhesion. Still the SPS trial 2 coating did not achieve better microstructure in terms of density, and cracks compared to the APS coatings but further looking into the parameters, more desirable coatings can be achieved. After TCF testing, a layer of TGO was seen at the bond coat/topcoat interface, and there was no failure of the coating seen.

When ice is accumulated on a wind turbine's rotor blade its aerodynamics are altered, leading to reduced efficiency and sometimes altered pressure oscillations around the blade. These pressure oscillations can be detected as sound. With sound level measurements over a long time, combined with known ice conditions in the same period, the measured sound data can be used to classify the ice conditions. This master's thesis aims to investigate the possibilities of using sound level measurements at 36 frequency bands in the range 6.3–20 000 Hz along with machine learning and wind speed to detect icing on wind turbine blades. Four k-NN models have been trained and evaluated using two different data configurations that each treat two different means of normalization: one uses the raw sound level data in dBA which has been standardized using z-score. The other uses the wind power density Iwind = 0.5ρU3 instead of the reference sound intensity I0 = 10-12 W/m2 in the decibel formula L = 10log10(I/I0) to reduce the influence of wind speed on the data. The sound/wind speed hybrid data was also z-score standardized. Available data was from February 21st to March 3rd in 2023 and March 1st to April 3rd in 2024. In the summer of 2023, the leading edges of the rotor blades on the investigated wind turbine were renovated which might have altered the sound. Therefore, what is denoted as Data configuration A used 2024 data as training data while 2023 data was used solely for testing. Data configuration B on the other hand used data from March 1st to March 17th 2024 for training and data from April 1st to April 3rd 2024 for testing as the rotor blades were identical between those data sets. Wind conditions were also more similar between training and testing data for Data configuration B. The models were optimized using grid search, varying k, distance metrics and feature combinations of the 36 frequency bands, while maximizing the balanced accuracy, BA, of the model using 5-fold cross-validation. For Data configuration A, this resulted in a balanced accuracy in the testing stage at BAtesting = 0.535 using the dBA sound level data, and BAtesting = 0.601 using the data normalized with wind power density. For Data configuration B, balanced accuracy was BAtesting = 0.845 using the dBA sound level data, and BAtesting = 0.773 using the data normalized with wind power density. The main conclusion is that icing can be detected using sound level measurements, wind speed and machine learning although the models in this project generalize poorly partly due to limited data and partly due to how the models were constructed. The models perform better with wind speeds similar to the training data. / När is ackumuleras på vindturbinblad ändras aerodynamiken vilket leder till lägre verkningsgrad och ibland förändrade tryckoscillationer kring bladet. Dessa tryckoscillationer kan detekteras i form av ljud. Med hjälp av ljudmätningar över en längre tid, kombinerat med kända isförhållanden under tidsperioden, kan ljuddatan användas för att klassificera isförhållandena. Målet med detta examensarbete är att undersöka möjligheterna att använda ljudnivåmätningar vid 36 frekvensspann mellan 6,3–20 000 Hz tillsammans med maskininlärning och vindhastighet för att detektera isbildning på vindkraftverk. Fyra modeller baserade på algoritmen k-NN har tränats och utvärderats med två olika datakonfigurationer som vardera behandlar två olika metoder för normalisering: en använder obehandlad ljudnivådata i enheten dBA som har standardiserats med z-poäng. Den andra använder vindenergidensiteten Iwind = 0.5ρU3 istället för referensintensiteten I0 = 10-12 W/m2 i formeln för decibel L = 10log10(I/I0) för att begränsa vindhastighetens inverkan på datan. Ljud-/vindhybriddatan standardiserades också med z-poäng. Den tillgängliga datan var mellan 21 februari och 3 mars 2023 samt 1 mars till 3 april 2024. Sommaren 2023 renoverades bladen på det undersökta vindkraftverket vilket kan ha påverkat ljudet. Därför användes data från 2024 som träningsdata och data från 2023 som testdata i vad som benämns som Data configuration A. Data configuration B använde istället data från 1-17 mars 2024 för träning och data från 1-3 april 2024 för testning eftersom rotorbladen var identiska mellan de datamängderna. Vindförhållandena var också mer lika inom Data configuration B. Modellerna optimerades med grid search genom att variera k, avståndsmått, och vilken kombination av de 36 frekvensspannen som ingår i modellen. Balanserad träffsäkerhet, BA, är resultatet som maximerades genom 5-delad korsvalidering. För Data configuration A resulterade detta under teststadiet i BAtesting = 0,535 med omodifierad ljuddata och BAtesting = 0,601 då vindenergidensiteten användes som ljudets referensnivå. För Data configuration B var den balanserade träffsäkerheten BAtesting = 0,845 med omodifierad ljuddata och BAtesting = 0,773 då vindenergidensiteten användes som ljudets referensnivå. Den främsta slutsatsen är att isbildning kan detekteras med ljudnivåmätningar, vindhastighet och maskininlärning men modellerna som har tagits fram i detta projekt presterar relativt dåligt, delvis på grund av en begränsad datamängd och delvis på grund av hur modellerna har konstruerats. Modellerna presterade bättre för testdata med liknande vindförhållanden.