On Perceptually Colorless Artificial Reverberation

dc.contributorAalto-yliopistofi
dc.contributorAalto Universityen
dc.contributor.advisorSchlecht, Sebastian
dc.contributor.authorHeldmann, Janis
dc.contributor.schoolSähkötekniikan korkeakoulufi
dc.contributor.supervisorSchlecht, Sebastian
dc.date.accessioned2021-08-29T17:12:35Z
dc.date.available2021-08-29T17:12:35Z
dc.date.issued2021-08-24
dc.description.abstractWind Turbine (WT) blades undergo high operational loads, experience critical environmental conditions, and are susceptible to faults. These factors severely affect the efficiency of WTs, and increase the importance of developing an effective methodology of maintenance of WTs in practice. To detect exterior WT blade faults, visual surface inspection through images is the most common inspection method in practice. With recent advances in deep learning, the available high- resolution blade images acquired in past inspection campaigns represent an enormous potential to use neural networks to automate the analysis of these inspection images. This project explores the applicability of deep learning in the detection of faults in wind turbine (WT) blade images taken from the ground using a telescope. A two-stage approach is proposed to detect WT blade faults, first classifying the images based on the location on the blade, and then detecting the faults. Through collaboration with Xabet Digital Solutions, multiple datasets from available historical wind park inspections are built for training and optimizing. ResNet50 architecture together with transfer learning and data augmentation techniques is used for the localization stage. Successful object detection algorithms such as SSD are tested in the detection stage. Further, an architecture to deploy the proposed approach in a real-world scenario is presented. The design of the models and the deployment is motivated by the available resources and the specific application demands in the practical This work is a contribution to the quest to achieve colorless artificial reverberation started by Schroeder and Logan in 1961. By analyzing various feedback delay networks (FDNs), novel criteria for artificial reverberation were investigated. Particular focus was laid on the modal properties of FDNs, such as modal decay, frequency, phase, and excitation and its analysis through modal decomposition. Modal excitation distributions were modeled by synthetic Rayleigh distributions and evaluated by a psychoacoustic listening test. It was found that narrow modal excitation distributions can improve colorless artificial reverberation. The topology of a recently proposed allpass FDN was determined to range from a single allpass filter to a series allpass reverberator, enabling a flexible design space for its modal excitations. A colorless allpass FDN was designed to exhibit a narrow modal excitation distribution and was subjectively evaluated as more colorless than conventional FDN designs.en
dc.format.extent59+1
dc.format.mimetypeapplication/pdfen
dc.identifier.urihttps://aaltodoc.aalto.fi/handle/123456789/109345
dc.identifier.urnURN:NBN:fi:aalto-202108298581
dc.language.isoenen
dc.locationP1fi
dc.programmeCCIS - Master’s Programme in Computer, Communication and Information Sciences (TS2013)fi
dc.programme.majorAcoustics and Audio Technologyfi
dc.programme.mcodeELEC3030fi
dc.subject.keywordartificial neverberationen
dc.subject.keywordpsychoacousticsen
dc.subject.keywordphysical modelingen
dc.subject.keyworddigital signal processingen
dc.titleOn Perceptually Colorless Artificial Reverberationen
dc.typeG2 Pro gradu, diplomityöfi
dc.type.ontasotMaster's thesisen
dc.type.ontasotDiplomityöfi
local.aalto.electroniconlyyes
local.aalto.openaccessyes

Files

Original bundle

Now showing 1 - 1 of 1
No Thumbnail Available
Name:
master_Heldmann_Janis_2021.pdf
Size:
8.62 MB
Format:
Adobe Portable Document Format