Anda di halaman 1dari 29

Presentation for the course in Multibody Dynamics

Flavia Buonanno February 2012

Techniques for the efficient analysis and simulation of large models, reducing computational requirements without damaging the dynamical properties of the model suitable with Finite Elements spatial discretization

Departing from the linear equation of motion

The idea is to find a lower dimension space Such that the state vector can be approximated as The projection versus the lower dimension yields

With the reduced system matrices

The reduced model mass and stiffness matrices can be derived energetically by substituting the coordinate change into kinectic and potential energy expressions of the original model

The transformation T will take different forms depending on the transformation technique utilized

The equation of motion for the undamped system can be partitioned Where: Main assumption: for the lowest frequency modes, inertia forces on the slave dof are much less important than the elastic forces transmitted by the master dofs slave dofs are assumed to move quasi-statically wrt the motion of the master dofs In order to obtain a relationship between xs and xm, inertia terms are neglected (static nature) and it is assumed that no force is applied on the slave dofs From the lower partition: Expressing the state vector of the full model in terms of the master dof, the transformation matrix is obtained as:

Is a good approximation (exact for static analyses) for the low eigenvalue spectrum of a FE structure, the quality of the eigenvalue approximation decreases as the mode number increases In general, when high frequency motion is considered, the influence of the inertia terms is significant the basic assumption of static reduction is contradicted and the method becomes inaccurate. The quality of the eigenvalue approximations is highly dependent on the location of the preserved dof (master) The mass of the reduced system is not effectively preserved and generally reduced model frequencies are higher than those of the original full space model

The structure of the transformation matrix is similar to the Guyan reduction, with the difference that, this time, the inertia information of the model is considered Applying the Laplace transformation to the equation of motion of the undamped system, we obtain the description in frequency domain:
for undamped system

The system can be rewritten and partitioned in master and slave as in the Guyan reduction, with

B(w) is called the dynamic stiffness matrix

Assuming again that no forces are applied on the slave dof, and following the same steps as for the static case, the transformation matrix is obtained

For w=0, it reduces to the Guyan reduction With respect to its static counterpart, it is a better candidate for approximating high frequency motion, though, T depends of the choice of an initial frequency As the transformation depends on w, the choice of this value should not be conducted randomly, but it should constitute a value out of the models frequency spectrum (the free vibration equation with the dynamic stiffness matrix involved is not satisfied) The accuracy of the method is limited to a frequency range around the initial value selected for initializing the method has high accuracy for systems under periodical excitations

As its name indicates, this method is an effort for improving the already presented reduction methods, by including the inertia terms of the original model in the definition of the transformation matrix From the dynamic equation of a free vibrated Guyan reduced model: From the definition of the Guyan transformation: Both relations are substituted into the equation for the lower partition For obtaining the following transformation matrix

The transformation obtained can be expanded, evidentiating its direct dependence on the static transformation matrix obtained from Guyan

Taking the Guyan condensation as basis, the static information is perturbed by introducing the inertia terms as pseudo-static forces The generated reduced model capture the dynamics of the original model better than the Guyan condensation, as expected from the introduction of a summand (that adds information to the static matrix) containing the inertia info from the original model A dynamic counterpart can be found by taking as basis the dynamic reduction, and, with analogous passages, we obtain the similar transformation matrix for the dynamic variant:

The IRS variants shown rely on the reduced models of Guyan or dynamic condensation. Once the transformation is computed, an improved estimate for the transformation matrix can be obtained by applying an iteration scheme given by:

Also known as substructuring

division of a structure into components

Reduced-order models of the components are obtained and then assembled into a reduced order model of the entire structure Primary uses of dynamic substructuring: - Couple reduced order models of moderately complex structures - Test verification of FE models of components - Implement computation of dynamics of very large FE models

The individual substructure models are transformed from physical to component modal coordinates using a set of chosen basis functions normal modes (eigenproblem), constraint modes, attachment modes One of the most diffused CMS methods in structural analysis is the Craig Bampton Method, based on normal modes and constraint modes

The physical dofs for a singular component can be partitioned into a set of interior dofs and a set of interface or boundary dofs :
Free interface Fixed interface

Fixed Interface Normal Modes:

Constraint Modes:

Mass normalised eigenvectors of the component with the interface dofs fixed, which form the columns of the fixed interface modal matrix ; a subset of k modes are kept

Static displacement of the component due to a unit displacement of one interface dof with all other interface dofs fixed, where is a matrix of displacements of the interior dofs and is the constraint mode matrix

CM space includes both fixed interface normal modes and static constraint modes
The interior physical coordinates are transformed into the fixed interface modal coordinates , and the physical interface are retained, but denoted coordinates as constraint coordinates

After transformation, the equation of motion for the generic component

remains:

The component modal mass and stiffness matrices


Where: constraint modal mass and stiffness matrices for component coupling matrix diagonal matrix of kept modal eigenvalues

For the synthesis of two components and , compatibility of displacements at the boundary is given by Imposing the coupling conditions through a transformation matrix, the assembled stiffness and mass matrixes are obtained

Most applications of CMS employ one from two approaches:


- Constraint mode: employ constraint and fixed interface normal modes (Hurty and Craig-Bamptom) - Attachment mode: attachment and free-interface normal modes (McNeal and Rubin)

Component models based on the used of fixed-interface plus constraint modes are essentially superelements all physical boundary coordinates are retained as independent generalized coordinates, greatly facilitating component coupling CMS generates reduced order models which capture very well the dynamics of the original model with relatively few component modes, no wonder is implemented as a standard and most reliable reduction algorithm in several FE packages The components couple in a quite straightforward way, and the sparsity patterns of the resulting system matrices allow for faster computation times The constrained modes embodying the quasi-static effect of high-level normal modes can compensate the error caused by normal modes truncation at certain extents and accelerate convergence. The reduced model depends on the sum of the normal and constrained modes. The more dof at the interface, the more accurate the description of structural motion.

The basis is the modal matrix of the discretized FE structure, where n is the dimension of the model, and q the number of computed eigenvectors By making use of the modal transformation (relation between state vector and modal coordinates) and partitioning, it holds From the first line of the equation, the modal coordinates can be expressed in terms of the master dof, by using the pseudo-inverse definition

Then, by substituting into the initial modal transformation, the SEREP transformation matrix is obtained

Due to the nature of the transformation, an analytical expression for the reduced mass and stiffness matrices can be derived The reduced system matrices can be computed
directly by using only the eigenvalue data

SEREP exactly preserves all the modes selected from the original model, and the reductions quality depends on the selection of the full eigenvectors It has two main drawbacks: - Its application is only feasible if the modal matrix of the original model is available: sparse algebra techniques or lumped mass approximation can be used to overcome this problem - The system is truly equivalent when m=q. Different choices can be made, but careful must be taken.

The reduction methods presented so far are based on the selection of master and slave dofs, where - master are kept after reduction and therefore constitute the dof of the reduced model - slave are the unwanted dof to be removed from the ODE of the original model The selection of master and slave dof sets is model-dependent and requires much experience

Other reduction techniques non dependent on the selection of master dofs have been originated within control theory, such as Krylov Subspace Method and Balanced Truncation approach. Though, those methods were initially developed for LTI first order systems

It is not advised the direct aplication of KSM or BT , since there is no guarantee for preservation of reduced model structure Modifications have been introduced to derive second order methods

FEW BASIC CONCEPTS Krylov Subspace: A is a constant nxn matrix, b a nx1 start vector. The Krylov subspace is the subspace spanned by the q column vectors as defined below

Given the set of equations of a MIMO system, and its transfer matrix obtained after Laplace transformation Input matrices
Output matrix Input , output and state vectors

Two important quantities can be defined: System Moments: coefficients of the Taylor expansion of Reproduces low frequency range H around zero System Markov Parameters: replacing the eigenvalue term s by 1/lambda, the Taylor expansion about lambda=0 yields a different expansion

Reproduces high frequency range

A reduced q dimension subspace is seeked (q<<n) in order to approximate the state space vector of the original system For then projecting the system in the subspace spanned by another coordinate transformation matrix The KSM algorithm provides certain Krylov subspaces for the assignment of the projection matrices, such that the moments of the original and reduced model match they are INVARIANT parameters For an undamped mechanical system, it can be demonstrated that

The method is implemented by using the Arnoldi algorithm for calculation of basis vectors and Gram-Schmidt orthonormalization for obtaining the orthonormal basis vectors

The associated reduced model preserves structure and stability properties of the original model

Eigenvalues are well approximated, generally better than the rest of the reduction approaches already presented captures very well the dynamics of the original model It is independent from the position of master and slave dof, thus the method minimizes user intervention and has better numerical properties since the band diagonal structure of the system matrices is not damaged However, the dependence of the Arnoldi algorithm on the definition of a starting vector leads to the derivation of different projection matrices when different starting vectors are applied a suitable algorithm for obtaining a start vector is essential

Large dimension FE discretized systems: due to high accuracy required, or geometry complexity leading to smoother meshes the efficiency of the reduction is affected by the systems large dimension

The efficiency can be measured through its goodness in capturing the dyamics of the original model, and the computation time likewise high for large systems For coping with this, the reduction can be accomplished stepwise in two stages: - In the first stage, a reduction is applied generating a still large reduced order model, computed relatively fast and able to capture well the dynamics of the original model, up to certain frequency - the dimension is still reduced by applying a further reduction method Choice of the algorithm for first ROM is determinant for the efficiency of the two step method it should be able to create an accurate reduced model with a fairly low computation time The master and slave allocation constitutes an essential aspect for this first stage: the method should be m non dependent (to ensure that an inadequate selection of dofs would not affect negatively the first step) but simultaneously the set of m dof must be included in the first total set of dof

A study on a rod was performed by Koutsovasilis and Beitelschmidt, comparing the different reduction techniques presented Characteristics of the Model: - Discretized with FE in ANSYS - Tetrahedral elements - A total of 23835 nodes, with UX, UY and UZ dofs each n = 71505 - Master nodes are m=10 (black dots) For validity check, comparison between full and reduced model is done it is useful to compare the eigenfrequencies and eigenvectors of both models by a modal analysis For comparison of eigenvectors, the transformation matrix should be available, so as to compare eigenvectors of the same dimension: expand the reduced eigenvector to the dimension of the full model or viceversa Different Modal Correlation Criteria can be applied in order to make the comparison

Modal Correlation Criterion: Eigenfrequency Difference and Normalized Relative Difference


The lower the value, the better the reduction method

Modal Correlation Criterion: Modified Modal Assurance Criterion

Gives information about the angle of both compared eigenvectors, both mass normalized and reduced to the same dimension 100% means absolute correlation, but a modMAC correlation equal or larger than 80% indicates a qualitative succesful reduction

Modal Correlation Criterion: Stiffness Normalized Vector Difference

Gives information about the relative vector difference of stiffness-normalized eigenvectors Similar to the normalized relative frequency difference, since both modal stiffnesses are compared, which are eigenfrequency dependent The smaller the value, the best is the comparison

Modal Correlation Criterion: Normalized Modal Difference


Gives the deviation of single coordinates of compared eigenvectors vector coordinates that result in the worst correlation isolates the

User is able to define at this or near this node position a master dof and reapply the reduction procedure, by minimizing the bad coordinate correlation Based on Modal Scale Factor

Results are shown for some eigenvectors

Modal Correlation Criterion: Normalized Modal Difference

Modal Correlation Criterion: Normalized Modal Difference

SEREP and Krylov were the techniques delivering better eigenvalues results Guyan is the least reliable method for approximation of high frequency motion, due to its static nature IRS has good correlation results for the lower and some higher frequency motions by iterative scheme, the IRS transformation matrix can be improved giving good approximation results For CMS, the success depends on the selected number of eigenmodes for the internal structure (for the rod, 5 eigenmodes were taken) a selective choice of CB modes belonging to the whole eigenmodes spectrum (from lower, medium and higher ranges) seem to improve the results In general, the results could have been improved for a different choice of master dofs Krylov is a promising reduction method since there is no such dependence, and the user intervention is minimized, by just deciding the max dimension of the reduced system

Koutsovasilis P., Beitelschmidt M. Comparison of model reduction techniques for large mechanical systems. Multibody System Dynamics 20 (2008), pp. 111--128. Koutsovasilis P., Beitelschmidt M. Model Reduction of Large Elastic Systems: A Comparison Study on the Elastic Piston Rod. Proc. of 12th World Congress in Mechanism and Machine Science, IFToMM 2007, Besancon, France. Avitabile P., Model Reduction and Model Expansion and Their Applications Part 1 Theory, Proceedings of the Twenty-Third International Modal Analysis Conference, Orlando, Florida, Feb 2005. R. D. Cook, D. S. Malkus, M. E. Plesha, Concept and Applications of Finite Element Analysis, John Wiley, NY, 4th. Edition, 2001. L. Hinke, F. Dohnal, B.R. Mace, T.P. Waters, N.S. Ferguson, Component mode synthesis as a framework for uncertainty analysis Journal of Sound and Vibration 324 (12) (2009), pp. 161178. M.I. Friswell, S.D. Garvey, J.E.T. Penny, Model reduction using dynamic and iterated IRS techniques, Journal of Sound and Vibration, 186 (1995), pp. 311323. Craig R. R., Coupling of substructures for dynamic analysis: an overview, Proceedings of 41st AIAA/ASME/ASCE/ASC Structures, Structural Dynamics, and Materials Conference, Atlanta, GA, 38 April 2000, pp. 171179. Bai Z. Krylov subspace techniques for reduced-order modeling of large-scale dynamical systems, Applied Numerical Mathematics 43 (2002) 944. M. Paz, Modified dynamic condensation method. Journal of Structural Engineering 115 (1989), pp. 234238. Zhang N. Dynamic condensation of mass and stiffness matrices. Journal of Sound and Vibration 188 (4) (1995), pp. 601615. T. Bechtold, E. B. Rudnyi, J. G. Korvink, Fast Simulation of Electro-Thermal MEMS: Efficient Dynamic Compact Models, Series: Microtechnology and MEMS, 2006, Springer, ISBN: 3540346120.

Anda mungkin juga menyukai