University:
Boston University
Course:
CAS PY 105 | Physics 1
Academic year:

2024
Views:

9

Pages:

12
Author:

gydraddolmxla

Fan broadband noise background Largest broadband contributor in a fan stage is from rotor wake interaction with FEGV (This also produces the tonal noise) fan FEGV • Low-order methods aimed at assisting in the design phase use : – Information about the rotor wake turbulence intensity and lengthscale (as well as mean flow) – Then the response of the FEGV is computed and subsequently the noise produced by the interaction • Rotor wake info currently taken from simulation or experiment … – Method not fully “low-order” Our project Main part: Create a surrogate model for the rotor wake flow (eliminating need for computation or experiment in order to define input for the low-order FEGV calculation) Use machine learning End goal: provide fan geometry, RPM, mass flow and perhaps some other information together with duct geometry and have the surrogate model provide the mean flow and turbulence intensity and length scale just upstream of the FEGV Secondary part: Improve the low-order FEGV response method: full scale validation, relaxation of some assumptions fan FEGV Machine learning We have considered 2 basic methods so far Method 1: single output Define rotor geometry on %radial strips: • chord, • stagger, • position of t.e. • t.e. bdy layer thickness Give RPM, mass flow Predict flow values on ordered grid points • Streamwise velocity mag • TKE (k) • Turbulent dissipation Provide flow values on portion of NxMxP grid in the gap region • Streamwise velocity magnitude • TKE (k) • Turbulent dissipation (e or w) Used splines, MARS, XGBOOST with decision tree, DNN Tested necessary values of N, M, P (3D grid in gap) Tested method for selecting which data to use for training/testing Random selection out of N, M, P (what %: 80% train, 20% test, etc.) Leave out entire N (axial locations), leave out entire mass flow cases, etc. Database for the ML thus far • Use RANS (k-w), rotor alone simulations in rotor frame • Geometries : based on NASA SDT – CAD (cold), 7808 RPM (hot), 11607 RPM (hot), 12657 RPM (hot) • 7 RPM : 50%, 60%, 70%, 77.5%, 87.5%, 95%, 100% • 7-10 mass flow rates at each RPM • Total of 268 cases Example: SDT cold. EXP data (digitized from NASA TM, fan alone performance) 100% 95% 87.5% 77.5% 70% 60% 50% ML method I 2 step process for the machine learning 1st : learn functions related to rectification of the wake 2nd: learn the flow values in the wake j=1 j=2 . . . j=57 . . . j=89 j=1 . j=21 . . . . . . j=88 j=89 2𝜋 22 j=7 j=8 . . . j=57 . . . j=95 j=65 j=66 . . . j=110 . . . j=153 Periodicity (a) Original data (at fixed radial %) j=7 j=8 . . . j=57 . . . j=95 Percentagewise (b) Centered data j=65 j=66 . . . j=110 . . . j=153 2𝜋 22 (c) Centered and percent data 100% ML method I Example: streamwise velocity, XGBOOST • When randomly use 80% of NxMxP gap points Very good reconstructions of points not used. HUB TIP Streamwise velocity Model: v_total ~ omega + r_te + r_new + th_te + th_new + x_te + th_percent + massFlow + lowBLnor + upBLnor + chord + bladeAngle • Test MSE: 37.7 • R-squared: 0.996 • k: 3 - 28 Data very near hub and tip are ignored • i: 2 - 30 ML method I Example: streamwise velocity, XGBOOST, some chosen at random from all wake location points When leave out some entire mass flow cases Streamwise velocity shape is fine, magnitude is not so good w Streamwise velocity Turbulence dissipation parameter is fine (mass flow not an important parameter!) ML method I Example: leave out entire rotor speeds and then try to predict Decision tree (XGBOOST) leads to “nearest neighbor” type predictions – shape ok, magnitude off Streamwise velocity (XGBOOST) DNN approach leads to good mean values on a passage but wake shapes are off Note : Impact on low-order calculations …. Wake shapes (passage profile) needed for tonal noise Average value on passage needed for broadband noise DNN works fine Streamwise velocity (DNN) Machine learning method II Method 2: multi output Use all rotor information as one input (no radial slicing) 28 Learn axial cut “image” as one output ML: Decoder part of Convolution Neural Network 100 256 175 1D tensor: 1X199 7X25X1 14x50x32 28x100x1 Rotor geometry • chord, • stagger, • position of t.e. • t.e. bdy layer thickness For each strip, combined into one vector + RPM, mass flow ML method II Example: CNN, TKE, random selection of axial slices to use for training and testing • Input: 199 28 * (r_te, th_te, x_te, lowBLnor, upBLNor, chord, bladeAngle) + omega + massFlow + x • k: 2 - 29 • i: 1 - 30 Used 80% of axial slices for training HUB TIP Must determine reason for jitter Will explore further options for numbers of layers in CNN etc. Final ML comment Historical empirical wake evolution fittings utilized rotor force In our inputs, we have utilized boundary layer trailing edge thickness and will also use rotor force in the future Only way to know these parameters currently is to run a simulation, which defeats our purpose We have efforts ongoing to : • learn the wake without these parameters as input • learn these inputs as a first step so they can be used in the wake ML Summary and future work Wake parameter surrogate model Have developed viable single output method : At this moment would suggest combination of DNN and XGBOOST to fully describe wake Have demonstrated potential for CNN for multi output Allows clearer use of rotor case as 1 input and forces “physical” connection of output values Must fully develop and test Rotor force data, inflow and outflow angles etc. have not been used as input yet Both methods will be tested using larger set of rotor input parameters Test ability to learn rotor force and rotor trailing edge boundary layer thickness Must increase database, include larger differences in geometry and test ML predictions Noise prediction (focus on low-order FEGV response) Full scale test data (CLEEN I) being curated for use in validation study Year 2 effort includes expansion of low-order capability: nonisotropic turbulence, circumferential asymmetry

Fan Broadband Noise Background

Related Documents

Recommended Documents

Description

Related Documents

Free up your schedule!

Take 5 seconds to unlock