AIADAGA--144-Vansant---presentation-技术-课件

1、Latest developments in solving large simulation models in view of pass-by noiseAIA-DAGA March 20th 2013, MeranoLatest developments in solving large simulation models in view of pass by noise - OverviewReview of applicable simulation techniquesPerformance comparison for computing full vehicle PBN tar

2、get FRFs up to 4 kHzTPA based in-room pass-by noise model123Conclusion4Pass-by noise engineering In-RoomTime domain source contribution analysis=yk=NTFkiujQiHjiLW12krdx = 0 x = -= targets (measured & predicted)= indicators (measured)= sources (identified)Energetic TPA for PBN Engineeringu1(t)yk(t)+y

3、kn(t)Qn(t)Q1(t)yk1(t)FIR HjiLoads(synthesized)Indicators (measured)Contributions(synthesized)Targets(synthesized)123FIR characteristics calculated from Frequency domain TPAun(t)FIR FIR NTFkiFIR FIR characteristics derived frommeasured NTFsTotal PBN and contributionsEfficient Acoustic Source Quantifi

4、cationLimited amount of sensor requiredBandwidth: From 150 Hz to 8 kHzLatest developments in solving large simulation models in view of pass by noise - OverviewReview of applicable simulation techniquesPerformance comparison for computing full vehicle PBN target FRFs up to 4 kHzTPA based in-room pas

5、s-by noise model123Conclusion4Conventional (I)BEM: limited to medium sized models ( 25 knodes). Because BEM matrices are fully populatedMemory requirements O(n2)Direct Solving Factorization O(n3)Fast Multipole BEM: Fast Multipole ExpansionOctree Structure: BEM nodes are divided into cells at differe

6、nt levels. This happens at each frequency.For cells far away from each other (see x nodes and y nodes of 2 cells in picture), Greens kernel function is replaced with: Suppose we have 10 x nodes and 10 y nodes. 21 operations to get the influences conventional BEM would need 100 operations O(n (log(n)

7、2)Fast Multipole BEMRay AcousticsSound travels as rays: this assumption becomes more valid with increasing frequencyHowever difficult to capture accurately:Multiple diffraction: a ray hitting an edge should split into a new bundle of rays, each of which can diffract again on other edgesMultiple refl

8、ections: in highly reverberant spaces, smaller (energy) errors might build up after many reflections a good candidate for tyre and exhaust sources but not ideal for engine sources .8 copyright LMS International 2010Field Response for Warning Sound in front of bumper (Front Right) Comparison between

9、FMBEM and Ray Acoustics Multiple Cars650 HzRay Acoustics TessellatedRay Acoustics BEM MeshFast Multipole BEM (reference) 2500 HzField Response for Warning Sound in front of bumperComparison between FMBEM and Ray AcousticsFEM PML / AMLFinite layer of FEM elements forms a Perfectly Matched Layer which

10、 absorbs waves traveling outwards. The absorption is obtained by a transformationThis ensures no waves travel back into the FEM domain after reflection (to satisfy Sommerfeld) accurate solution obtained on solution anywhere else is obtained using a Kirchhoff surface integral, similar to the post-pro

11、cessing step in BEMA locally conform implementation allows using any convex shape EFFECT ON PERFORMANCE!LMS Virtual.Lab offers AML (automatically Matched Layer), in which the PML layer is automatically built on solver levelDirect MUMPS solver and iterative Krylov solver supportedFEM Adaptive Order (

12、FEM AO)Available in LMS Virtual.Lab 12 (Summer 2013) FEM AO, the next generation FEM Acoustic solver:Higher order shape functions are used to represent the pressure inside each elementAt order 10, an element can span more than 2 acoustic wavelengths.The solver automatically increases element order a

13、nd therefore nr of DOFs in the model with frequency fixed nr of DOFs for conventional FEM solvers. Important savings on time and memory in lower frequenciesThis allows for smaller model definition on pre-processor (LMS Virtual.Lab) which can be handled easierDiscretization only needs refinement in o

14、rder to capture accurately the geometry and (vibration) boundary conditionsLatest developments in solving large simulation models in view of pass by noise - OverviewReview of applicable simulation techniquesPerformance comparison for computing full vehicle PBN target FRFs up to 4 kHzTPA based in-roo

15、m pass-by noise model123Conclusion4Simulation model descriptionSeveral FEM models and a BEM model were built for a full vehicle (Chrysler Neon) including the surface of the powertrain to compute the target FRFs for a TPA PBN modelA reciproque approach is used with 19 monopole sources both left and r

16、ight from the vehicle located 7.5 m from the center line and from -6 m till 12 m from the back of the vehicle, all at 1,2 m above the ground. There were 2 receivers per tyre, 6 for the engine, and 2 for the exhaust (muffler and tailpipe) Comparison for Neon PassBy Noise Model up to 4 kHz3 frequency

17、ranges were defined to allow for 3 model sizes for conventional FEM and to adjust solver parallelization settings per range to optimally use the hardware. FEM AO updates the model size automaticallyFEM AO was about 2 times faster but needed more memory compared to the FEM models for 2 kHz. Many elem

18、ents in FEM AO had lower order for 2 kHz. AML requires additional DOFs for the FEM AO at low frequencies.The higher you go in frequency, the better the results get for both speed and memory for FEM AOFor the mid frequency range, FEM AML iterative solver needed 2h/freq and FMBEM 1h/freq. This is both

19、 remarkably slower than the FEM AO models with direct MUMPS solverFEM AO: evolution of model constitution, size and performance for the highest frequency rangeEvolution of model constitution. With higher frequencies more element of higher order are usedEvolution of total # DOF, memory and solving ti

20、me which are all frequency dependent for FEM AOAccuracy Comparison for Neon Pass By Noise Model up to 4 kHzPBN contribution results New ISO 362 Acceleration gear 3A set of simulation FRFs was used to propagate previously identified sources of a real vehicleTypical trends can be observed: tailpipe co

21、ntribution grows as the vehicle passes by, left tyres contribute more to left microphones compared to right tyres and vice versa, front tyres contributions peak earlier compared to rear tyresLatest developments in solving large simulation models in view of pass by noise - OverviewReview of applicable simulation techniquesPerformance comparison for computing full vehicle PBN target FRFs up to 4 kHzTPA based in-room pass-by noise model123Conclusion4ConclusionThis paper aimed at providing an overview of simulation methods to predict ext