Eigen Value Problems
A typical finite element formulation would yield the governing
equations as
In the case of free vibration, there is no external loading {F} = 0
and damping has negligible role. Therefore, for undamped free
vibration problems,
Assuming harmonic vibration at a frequency ωi
{Xi} = {Ui} sin (ωit) , we have
This is the basic equation, in matrix form, governing the
undamped free vibration of the structure.
Given the stiffness and mass matrices [K] and [M], we have to
find the natural frequencies ωi , and mode shapes {Ui}, i =
1,2,3,……. n.
These represent certain characteristic states of free vibration
of the system, each such state characterised by its eigenvalue
ωi and eigenvector {Ui}.
The above equation is a set of n equations in (n + 1)
unknowns, viz., ωi and n elements of {Ui}.
So a unique solution cannot be obtained
We choose to determine ωi and the ratios of elements of {Ui}.
Thus the mode shapes are defined only within a multiple of
themselves, i.e. if we have an eigenvector { U }, then α { U }
( α 0 ) is also an eigenvector.
There exist n pairs of natural frequencies ω and mode shape
{U}.
If the structure is excited at the frequency ω, it would vibrate
such that the various points will have “relative” amplitudes as
given in the mode shape eigenvector {Ui}.
For example, the first mode and fifth mode of a simply supported
beam would be as shown in figures below
If the beam is modelled with four beam elements, then the finite
element nodal points will vibrate as shown in the following
equation:
The “mode shape” only indicates
the relative amplitudes of vibration
(viz., overall shape of vibration)
and may readily be “scaled” to any
amplitude.)
Another way of interpreting the natural frequency and mode
shape is that, if the structure were given an initial displacement
to all its d.o.f. according to the relative amplitudes given in
the mode shape, and left free to vibrate on its own, it will
vibrate at the natural frequency ω, always maintaining these
relative amplitudes.
Since the finite element mesh of the structure has “n” d.o.f.,
our model would yield n pairs of natural frequency ω, and
mode shape {Ui} (i = 1, 2, ..., n).
The general form of the governing equations for the
undamped free vibration is given by
Which can be rewritten as
Again the eqn
The form of representation
is known as the standard form of eigenvalue problem,
whereas the form
is known as the nonstandard form
Cholesky Factorization of a Matrix
A symmetric matrix [M] can be factorized by Cholesky
factorization method i.e. it can be expressed as
where [L] is a lower triangular matrix.
Example
Equating the corresponding elements, we get
Some Properties of Eigenpairs
Basis vectors
The eigenvectors constitute the least number of linearly
independent mode shapes of the system.
Any deflected shape of the structure can thus be represented
as a linear combination of these eigenvectors { Ui }, i.e.,
Orthogonality of eigenvectors (for symmetric matrices)
…….Eq (1)
…….Eq (2)
…….Eq (3)
…….Eq (4)
Eigenvalue shifting
Thus we see that by effecting a shift in the eigenvalues of the
system, we are not changing the eigenvectors.
This property is used in typical computer procedures, to avoid zero
eigenvalues (i.e. rigid body modes) as also to improve the
convergence of two close eigenvalues.
Sturm sequence property
Cholesky factorise such that
An important property of [D] is that the number of negative
elements in [D] equals the number of eigenvalues smaller
than μ.
This property is typically used to check our eigenvalue
solution.
If we have obtained, say m eigenvalues, we can use a trial
value (say, μ) slightly higher than the highest eigenvalue
obtained and perform a Sturm sequence check.
A Sturm check fail is indicated if the number of eigenvalues
below μ turns out to be more than m.
This means that our eigenvalue solver missed out some
eigenvalues, indicating that we need to redo the analysis.
Solution of Eigenvalue Problem
There are essentially three groups of methods of solution of
eigenvalue problems:
1. Determinant based methods
2. Transformation based methods
3. Vector iteration based methods
1. Determinant based methods
The determinant based methods are primarily based on the
following:
In principle, we can take a trial value of λ and compute the
determinant |[A] - λ[I]|.
With several trial values, we can generate a plot such as the one
shown in fig.
By suitably monitoring
the sign changes in the
value of the
determinant, we can
iterate towards the
eigenvalues λ .
Such a scheme is, however, never implemented in practice because
of the heavy computational cost.
Evaluation of each determinant of size (n x n) requires of the order
of n3 floating point operations, and several iterations may be
required to determine all the eigenvalues of interest.
2. Transformation based methods
Transform [A] into a diagonal matrix, using a series of matrix
transformations of the type
[A] = [T]T [A][T]
where [T], the transformation matrix, is usually an orthogonal
matrix, i.e. [T]T = [T]-1.
If we are able to transform [A] completely into a diagonal matrix,
then the elements on the diagonal themselves are the required
eigenvalues.
3. Vector iteration based methods
Vector iteration based methods, involve assuming a trial
eigenvector and performing repeated matrix manipulations to
converge to the desired eigenvector.
Inverse Iteration Scheme
Assume a trial vector {U} and premultiply it by [A] to get a value
of ω2.
Continue with iterations so that ω2 converges to the smallest
value.
It should be noted that in this scheme only the lowest
eigenvalue is calculated.
Algorithm :
(1) Determine [K] and [M] for the given problem
Inverse Iteration with Gram-Schmidt Deflation
Here we have to determine one eigenvalue say by real inverse
iteration scheme described above.
Then we proceed to determine the remaining eigenvalues
one by one as described below.
Simultaneous Iteration Method
In this method, we find all the eigenvalues simultaneously.
Algorithm :
(1) Determine global [K] and [M] matrices.
(2) Assume a trial vector set
Note : Order of this matrix should be same as the number of
eigenvalues to be found out.
(3) Compute
(4) Cholesky factorize
Repeat steps (3) to (8) till we get a satisfactory level of
convergence of [λ].
This is achieved when
(i) non-diagonal elements are of negligible value and
(ii) change in diagonal elements from one iteration to the
next one is minimal.
Find the natural frequency of axial vibrations of a bar of uniform
cross section of 20 mm2 and of length 1 m. Take E = 2 x 105
N/mm2 and ρ = 8000 kg/m3. Take two linear elements. Solve
using inverse iterative scheme
Divide the whole domain of the problem into two elements.
The element matrix equation for the problem is given by
Element matrix equation is same for both element, assembling
them together, we get the global matrix equation as
Impose global boundary conditions i.e. U1 = 0
Global matrix equation reduces to
Find the natural frequency of axial vibrations of a bar of uniform
cross section of 20 mm2 and of length 1 m. Take E = 2 x 105
N/mm2 and ρ = 8000 kg/m3. Take two linear elements. Solve
using simultaneous iterative scheme .
Divide the whole domain of the problem into two elements.
The element matrix equation for the problem is given by
Element matrix equation is same for both element, assembling
them together, we get the global matrix equation as
Impose global boundary conditions i.e. U1 = 0
Global matrix equation reduces to
Iterations to be continued till we get convergence of X. This means
non-diagonal elements of [λ] to be almost zero and diagonal
elements to be almost equal in two successive iterations.
Values of λ in successive iterations
