Inferring wall pressure spectral model using neural networks
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Tato diplomová práce se zabývá modelováním spektra tlakových fluktuací u stěny v turbulentní mezní vrstvě založeným na datech. Současné semi-empirické modely predikují spektrum tlakových fluktuací u stěny z čistě lokálních veličin mezní vrstvy (přístup One-to-One). Historie proudění je v těchto modelech reprezentována pouze některými parametry. Nový přístup navrhuje použít k predikci spektra tlakových fluktuací u stěny v jednom bodě více bodů proti proudu tekutiny (přístup Many-to-One). Neuronové sítě jako modely založené na datech jsou navrženy pro model jak s přístupem One-to-One, tak i Many-to-One. Použitá databáze se skládá z jednoho experimentálního a tří numerických (proudění okolo controlled-diffusion profilu) souborů dat. Relevantní parametry mezní vrstvy jsou extrahovány z databáze a zpracovány do bezrozměrného vztahu. Nejprve se natrénují dva One-to-One modely s dopřednou neuronovou sítí. Oba dobře predikují spektra z databáze. Jeden z modelů je natrénován pouze na souboru dat s controlled-diffusion profily, aby byla k dispozici srovnávací architektura. Za druhé jsou natrénovány dvě Many-to-One dopředné neuronové sítě se dvěma a deseti pozicemi v mezní vrstvě. Model se dvěma pozicemi vylepšuje model One-to-One. Desetipoziční model takové zlepšení nepřináší. Oba tyto modely však predikují spektra poměrně dobře. Zatřetí, Many-to-One model konvoluční neuronové sítě je natrénován se vstupy s proměnným počtem pozic. Model však na proměnný počet pozic nereaguje dobře. Predikce modelu tedy nebyla dobrá. Kromě aplikace neuronových sítí je teoreticky zkoumáno použití modelů Many-to-One. Závěrem lze k novému přístupu říci, že model Many-to-One se dvěma pozicemi sice přináší určité zlepšení, ale není tak výrazné. Potvrzuje se, že model Many-to-One využívá hodnoty, které jsou proti proudu tekutiny a že modely Many-to-One mohou predikovat spektra tlakových fluktuací u stěny.
This master thesis deals with a data-driven approach to the turbulent boundary layer wall pressure spectral modelling. Current wall pressure spectra semi-empirical models predict the spectra from purely local boundary layer quantities (a One-to-One approach). The flow history is only represented by some parameters in these models. A novel approach suggests using multiple upstream points to predict the wall pressure spectra at one point (a Many-to-One approach). Neural networks as data-driven models are proposed to build a model with a One-to-One and Many-to-One approach. A database of one experimental and three numerical (controlled-diffusion airfoil) datasets is used. Relevant boundary layer parameters are extracted from the database and processed into a dimensionless relation. Firstly, two One-to-One models with a feedforward neural network are trained. Both of them predict the spectra from the dataset reasonably well. One of the models is trained only on the airfoil datasets to have a comparison architecture. Secondly, two Many-to-One feedforward neural networks are trained with two and ten positions in the inputs. The two positional model improves the One-to-One model. The ten positional model does not bring such improvement. Both of these models predict the spectra reasonably well. Thirdly, a Many-to-One convolution neural network model is trained with variable-length inputs. However, the model did not react well on the variable-length inputs. The prediction of the model was not great. In addition to these models, the use of the Many-to-One models is investigated in theory. To conclude the novel approach, the two positional Many-to-One model does bring some improvement, but it is not that significant. However, the investigation confirms that the upstream values are used in the Many-to-One models and that the Many-to-One models can predict the wall pressure spectra.
This master thesis deals with a data-driven approach to the turbulent boundary layer wall pressure spectral modelling. Current wall pressure spectra semi-empirical models predict the spectra from purely local boundary layer quantities (a One-to-One approach). The flow history is only represented by some parameters in these models. A novel approach suggests using multiple upstream points to predict the wall pressure spectra at one point (a Many-to-One approach). Neural networks as data-driven models are proposed to build a model with a One-to-One and Many-to-One approach. A database of one experimental and three numerical (controlled-diffusion airfoil) datasets is used. Relevant boundary layer parameters are extracted from the database and processed into a dimensionless relation. Firstly, two One-to-One models with a feedforward neural network are trained. Both of them predict the spectra from the dataset reasonably well. One of the models is trained only on the airfoil datasets to have a comparison architecture. Secondly, two Many-to-One feedforward neural networks are trained with two and ten positions in the inputs. The two positional model improves the One-to-One model. The ten positional model does not bring such improvement. Both of these models predict the spectra reasonably well. Thirdly, a Many-to-One convolution neural network model is trained with variable-length inputs. However, the model did not react well on the variable-length inputs. The prediction of the model was not great. In addition to these models, the use of the Many-to-One models is investigated in theory. To conclude the novel approach, the two positional Many-to-One model does bring some improvement, but it is not that significant. However, the investigation confirms that the upstream values are used in the Many-to-One models and that the Many-to-One models can predict the wall pressure spectra.
Description
Subject(s)
model pro spektra tlakových fluktuací u stěny, turbulentní mezní vrstva, strojové učení, neuronová síť, wall pressure spectral model, turbulent boundary layer, machine learning, neural network