Heat transfer assessment of an ice induced engine inlet using Machine Learning

Heat transfer assessment of an ice induced engine inlet using Machine Learning

Koolen, Sven (TU Delft Mechanical, Maritime and Materials Engineering)

Klein, S.A. (mentor)

Delft University of Technology

Mechanical Engineering | Energy and Process Technology

2018-12-14

Icing conditions are some of the critical operating phenomena that an aircraft engine can encounter within its flight envelope.22,000 feet is recognized in the industry as the upper limit for the existence of super-cooled liquid water. Above this level, the particles are no longer in liquid water form, but rather ice crystals, snowflakes, graupel or hail. In this thesis is tried to contribute to the research into the critical operating phenomenon by answering the following research question: How can heat transfer analysis of aircraft engine inlets in ice crystals conditions be improved by application of \acrlong{ml} algorithms?\\ This thesis takes as starting point the test results of a \acrfull{ge} component test with a sector inlet in icing conditions. The test was run through a \acrfull{doe} matrix with inputs variations of flow inlet parameters and heating of the component. The temperature measurements from the test are signals which are showing the behavior of the wall temperature given a combination of inlet conditions.\\Ice crystals which are induced in the inlet flow impinge on the wall and melt because of the heat flux through the wall. This process repeats itself until the energy delivered through the wall is not enough to melt all the ice particles which are impinging. At this moment ice starts to accrete. Two hypothesis where formulated. Both using different characteristics in the temperature signal to calculate an experimental factor $\zeta$. This experimental factor $\zeta$ is used to close the "5 heat equation" which describes the heat fluxes to the impinging ice crystals. This $\zeta$ is used as a target value in a \acrshort{ml} algorithm.\\The \acrshort{ml} algorithm uses data to extract a mapping function which maps the input variables to the target variable, referred to as training. The mapping function of hypothesis 1 was nonphysical. It showed unexpected relations between the input variables and $\zeta$. Therefore is hypothesis 1 rejected. The mapping function of hypothesis 2 was physical and this hypothesis is accepted. \\ Using this trained \acrshort{ml} algorithm, the tool was build and prediction could be made. Some zones within the inlet variables showed very accurate predictions other zones showed very poor predictions. An unsupervised learning algorithm, \acrfull{pca} was used, to look for relations between prediction accuracy and inlet variables. The foundings from the \acrshort{pca} let to the clear proof that the "5 heat equation" used by \acrshort{ge} explains the physics insufficiently. Another approach to the "5 heat equation" and further development of the tool are proposed.

Machine Learning
Heat transfer
Ice

http://resolver.tudelft.nl/uuid:9de9e391-6ddd-41f4-9009-839fb0757a59

2022-12-07

Student theses

master thesis

© 2018 Sven Koolen