You are on page 1of 11

Acta Mech. Sin.

(2012) 28(3):782792
DOI 10.1007/s10409-012-0081-z

RESEARCH PAPER

Dierential quadrature time element method for structural dynamics


Yu-Feng Xing Jing Guo

Received: 27 July 2011 / Revised: 20 January 2012 / Accepted: 27 February 2012


The Chinese Society of Theoretical and Applied Mechanics and Springer-Verlag Berlin Heidelberg 2012

Abstract An accurate and ecient dierential quadrature


time element method (DQTEM) is proposed for solving ordinary dierential equations (ODEs), the numerical dissipation
and dispersion of DQTEM is much smaller than that of the
direct integration method of single/multi steps. Two methods
of imposing initial conditions are given, which avoids the
tediousness when derivative initial conditions are imposed,
and the numerical comparisons indicate that the first method,
in which the analog equations of initial displacements and
velocities are used to directly replace the dierential quadrature (DQ) analog equations of ODEs at the first and the last
sampling points, respectively, is much more accurate than
the second method, in which the DQ analog equations of
initial conditions are used to directly replace the DQ analog
equations of ODEs at the first two sampling points. On the
contrary to the conventional step-by-step direct integration
schemes, the solutions at all sampling points can be obtained
simultaneously by DQTEM, and generally, one dierential
quadrature time element may be enough for the whole time
domain. Extensive numerical comparisons validate the eciency and accuracy of the proposed method.
Keywords Dierential quadrature rule Direct integration
method Time element Phase error Artificial damping
1 Introduction
The discretization of the spatial domain of structural dynamic problems governed by partial dierential equations
The project was supported by the National Natural Science Foundation of China (11172028, 10772014).
Y.-F. Xing () J. Guo
The Solid Mechanics Research Center,
Beihang University, 100191 Beijing, China
e-mail: xingyf@buaa.edu.cn

results in a system of ordinary dierential equations (ODEs)


with time as the independent variable. Various methods are
available for numerical integration of ODEs. Most commonly used one, based on the finite dierence concept, is
the direct integration methods (DIM) [1] which is based on
the following two ideas. First, the dynamic equations are
sought at discrete time or sampling points within the interval
of solution. The second is that a variation of displacements,
velocities, and accelerations within each time interval t is
assumed. And the variation form within each time interval
determines the accuracy, stability, and eciency.
The commonly known drawbacks of DIM such as the
classical RungeKutta (RK) methods of dierent stages and
orders [2, 3], the Wilson method [4, 5], the Newmark
method [6, 7] and so on, are numerical dissipation and dispersion or artificial damping and phase errors, although
the symplectic integration methods [8] have the energyconserving property.
Alternatively, the ODEs can be discretized using time
finite element methods (TFEM), which are based on the
Hamiltons variational principles ([9, 10] for example) or
Hamiltons law of varying action ([11, 12] for example) or
Gurtins variational principle ([13] for example) and the
weighted residual methods ([14, 15] for example). The features of TFEM include: (1) In principle, desired accuracy of
arbitrary order can be achieved by taking high-order polynomials as primary functions; (2) The geometric property of
the time domain is simple, there is no need to handle complicated boundary shapes; (3) It is suitable for nonlinear problems and leads to more feasibilities; (4) The formulations of
TFEM are generally low order.
People generally constructed TFEM based on the variational principles and the weighted residual methods. In a different way, the ODEs can also be solved using the dierential quadrature (DQ) method [16, 17] wherein the ODEs are
discretized by DQ rule. The dierential quadrature method
(DQM) [18] is a simple and highly ecient numerical technique for initial-value problems, and could yield accurate results using a considerably small number of sampling points.

Dierential quadrature time element method for structural dynamics

The literature on DQM up to 1996 may be found in a survey paper [19]. And there is an increasing interest in applying DQM to solve dynamic problems [16, 2023]. Since the
governing ODEs for structural dynamic problems are of the
second order with two initial conditions, the application of
DQM is not as straight-forward as the first order ODEs [22].
The method of imposing initial conditions is an important
issue.
The imposition of the initial/boundary conditions in
DQM can be very tricky [19]. Several methods have been
used to deal with initial conditions, such as expressing
the initial conditions as dierential quadrature (DQ) analog
equations at the sampling grid points [16, 24, 25], modifying DQ rule [16, 26], modifying the trial functions [27, 28],
or modifying the weighting coecient matrices [21]. But
the second and the last have been shown to be identical by
Refs. [16, 22]. Hitherto, there are problems in feasibility,
generality and simplicity in imposing the initial conditions.
Especially no paper investigated the accuracy of dierent
methods in imposing initial conditions.
The advantage of the Fungs method [16, 22] is the accuracy of state variables at the end of time interval (which is
not a sampling point) can reach the order of 2n using only
n + 1 sampling points, additionally, unconditional stable and
non-dissipative algorithm can be constructed. However, the
disadvantages of the Fungs method [16] are also evident,
such as the accuracy at the sampling points is of the order
of n; the powers of polynomials of displacement and velocity are the same, hence the dierential relation between
displacement and velocity is violated, the initial velocity is
not exactly satisfied in general; In addition, the variables at
the end of time interval must be interpolated by the known
variables at sampling points. In a similar way to Ref. [16],
the beam problem subjected to a moving load was studied in
Ref. [17], but the imposing methods of initial conditions and
the analysis of stability were not involved.
In this context, this paper proposes an ecient and accurate time element method based on the DQ rule for solving
ODEs of structural dynamics, called the dierential quadrature time element method (DQTEM). The novelties of the
present study are as follows:
(1) The numerical dissipation and dispersion are caused
by lower order solution schemes including DIM and TFEM,
and accumulated linearly step by step. And it is well known
that the high order element is more ecient and accurate in
standard FEM, this motivates us to develop high order time
element method to solve numerical dissipation and dispersion problems based on DQM.
(2) Give two methods of imposing initial conditions,
and investigate their accuracy.
(3) Study the spectrum radius and phase error of
DQTEM.
(4) Compare the cost of DQTEM with the well-known
RK method and the widely used Newmark method.
An outline of the present paper is as follows. In Sect. 2,

783

the DQ rule is described briefly, the ChebyshevGauss


Lobatto sampling points are given and the DQM solution for
ODEs is derived. Section 3 presents the methods of imposing initial conditions, and their advantages compared with
the existent methods are highlighted. In Sect. 4, the stability
and accuracy of DQTEM are studied. The present results are
compared with analytical solutions and those of RK method
and Newmark method in Sect. 5. Conclusions are drawn in
Sect. 6.
2 DQM solution of ordinary dierential equation
2.1 The DQ rule
For clarity and completeness, DQM is briefly reviewed below, whose details can be found in Ref. [19] and a previous
paper [29] of the author.
The DQM idea is that the partial derivatives of a field
variable at the k-th discrete point in the computational domain is approximated by a weighted linear sum of the field
variable along the line which passes through the k-th point
and is parallel to the coordinate direction of the derivatives.
For example, the m-th derivatives of the field variable x(t) at
point tk is approximated as
x(m)
k =

N


A(m)
k j x j,

k = 1, 2, , N,

(1)

j=1

where A(m)
k j are the weighting coecients associated with the
m-th order derivatives, and N is the number of sampling
points. Note that Eq. (1) can be understood as the high order
dierence of the derivatives using variables. A key problem to DQ rule is the computation of weighting coecients
which can be obtained directly from an explicit formula [30].
The coecients for the first order derivatives are given by
(1)
A(1)
k j = l j (tk )

N
 
N


(t

t
)
(t j ti ),
k
i

i=1,ik,
j
i=1,i
j

(t

ti )
i=1,ik k

k  j,
(2)
k = j,

where l j is the Lagrange polynomials. The weighting coefficients of higher order derivatives can be obtained using the
recurrence relationship as
A(m)
kj =

N


(m1)
A(1)
.
ki Ai j

(3)

i=1

A crucial factor for improving the accuracy of the


DQ solutions is how to space the sampling points. It has
been well shown [31] that the so-called ChebyshevGauss
Lobatto points, first used by Shu and Richards [32] and used

784

Y.-F. Xing, J. Guo

widely since then, are more accurate than the equally spaced,
Legendre, and Chebyshev points in a variety of problems.
The time sampling points are given by
tk = t0 + h(t t0 ).

(4)

For the ChebyshevGaussLobatto sampling points, h is


1
k1

h = 1 cos
.
(5)
2
N1
2.2 The DQM solution
In the present study we employ DQM to discretize the structural dynamics ODEs as
M x + C x + Kx = f (t),

(6)

where M = [Mi j ]nn , C = [Ci j ]nn and K = [Ki j ]nn are the
mass, damping and stiness matrices, respectively; n is the
dimension or the degree of freedom (DOF) of the system.
The superposed dot denotes dierentiation with respect to
time. x(t) and f (t) are the displacement and the prescribed
load vector, respectively, and have the forms
x T = [x1 (t) x2 (t) xn (t)],
f T = [ f1 (t) f2 (t)

fn (t)].

(7)

Equation (6) can be obtained through spatial discretization


method as finite element methods, finite dierence method
and so on. Since each DOF is a function only of time, therefore its derivative based on DQ rule can be given as
xi |t=tk = xik =

N


A(1)
k j xi j ,

(8)

j=1

where xi is an element of displacement vector x(t), and xi j


is the value of xi at time sampling point t j . For the sake of
convenience, let us define
x Ti = [xi1 xi2 xiN ],
f Ti = [ fi1 fi2

fiN ].

(9)
(10)

Then the velocity vector x i and acceleration vector x i can be


given, respectively, by using DQ rule, as
x i = Ax i ,

(11)

x i = Bx i ,

(12)

where the weighting coecients matrices of order N N are


A = A (1) ,

(13)

B = A (2) .

(14)

It is noteworthy that the weighting coecients are dependent


only on the coordinates of time sampling points. Substitution
of Eqs. (11) and (12) into Eq. (6) leads to
MXB T + CX A T + KX = F,
where

(15)

X T = [x 1 x 2 x n ],
FT = [ f 1 f 2

(16a)

f n ],

BX T = [Bn 1 Bn 2 Bn n ],
AX T = [ An 1 An 2

(16b)

An n ].

When there is no damping, Eq. (15) can be transformed


to Sylvester equation whose solution can be found by the
method of Bartels and Stewart [33]. According to the theory
of matrix analysis, Eq. (15) can be transformed into another
easy to solve form as
GZ = Q,

(17)

where Z = cs(X ) and Q = cs(F) are the column expansions


of matrices X and F, respectively, that is
Z T = [x11 xn1 x12 xn2 x1N
Q T = [ f11

fn1 f12

fn2

f1N

xnN ],

fnN ].

Moreover, the coecients matrix of Eq. (17) is given by


G = B M + A C + E K ,

(18)

where E is a N N unit matrix, and represents the Kronecker product. Another form in terms of sub-matrix of
Eq. (18) is
G jm = B jm M + A jm C + E jm K ,

(19)

where G jm of order n n are the sub-matrices of G, j, m =


1, 2, , N. Correspondingly, Q j and Z j of order n 1 are
the sub-matrices of Q and Z , respectively, and
Q Tj = [ f1 j f2 j

fn j ],

(20)

Z Tj = [x1 j x2 j xn j ].

(21)

The above solution method is called the dierential quadrature time element method (DQTEM).
Remarks:
(1) Equation (17) includes the dynamic equilibrium equations DQ analog equations of at any time sampling
points.
(2) The requests for highly accurate and ecient solution
methods may be more necessary for the dynamics problems without damping, and in these cases, Eq. (15) can
be directly solved by the method of Bartels and Stewart [33].
(3) The state variables at all sampling points are solved simultaneously in DQTEM, which results in a great decrease, even to the extent of negligibility, in the accumulation of numerical dissipation and phase errors if the
sampling points are more enough, not like the lower order step-by-step DIM.
(4) For nonlinear dynamic system with less number of DOF,
it is convenient to solve Eq. (17) directly. Using DQTEM
in conjunction with iteration method.

Dierential quadrature time element method for structural dynamics

785

(5) When numbers of DOF and the sampling points are


large, the storage of matrix G may become a challenge
to the capacity of computer, unless there is a skillful solution strategy.

G Nm = A1m I,

3 The methods of imposing initial conditions

where I is a nn unit matrix. The DQTEM with this method


of imposing initial conditions is denoted by DQTEM1.

When the dynamic problems are solved, the imposition of


initial conditions has the same importance as the imposition of boundary conditions in the analysis of structural static
problems and free vibrations. When DIM and mode superposition method are used to solve ODEs, the imposition of
initial conditions is not complicated, but it becomes tricky
when DQM is used.
The methods of imposing initial conditions include expressing the initial conditions as DQ analog equations at the
sampling grid points [16, 24, 25], modifying the trial functions [27, 28], modifying DQ rule [16] and modifying the
weighting coecient matrices [21] as briefly reviewed in
Sect. 1. Our methods given below are based on but dierent from the method in Refs. [16, 24, 25], and convenient for
practical use compared with those available methods. In the
following, we first present two practical and simple methods
for imposing initial conditions in DQTEM, then give a few
remarks.
The initial conditions corresponding to the first time sampling point (usually t = 0) are as follows
x(t)|t=0 = x 0 ,
t=0 = x 0 ,
x(t)|

(22)

where x 0 and x 0 are the initial displacement and velocity


vectors, respectively. Using DQ rule, Eq. (22) can be rewritten as

xi1

xi2

0
0
0
0
1

..

A11 A12 A13 A1(N1) A1N


xi(N1)

xiN

xi0
, i = 1, 2, , n,

(23)
=
xi0
which are the DQ analog equations of the initial conditions.
3.1 The first method of imposing initial conditions
In this method, the first n equations of G, corresponding to
the first sampling point, are replaced by the DQ analog equations of initial displacement conditions in Eq. (23), the last
n equations of G, corresponding to the last or terminal sampling point, are replaced by the DQ analog equations of ini-

tial velocity conditions in Eq. (23), that is


G 11 = I,

G 1m = 0,

Q 1 = x 0,

Q N = x 0 ,

m = 2, 3, , N,

m = 1, 2, , N,

(24)
(25)

3.2 The second method of imposing initial conditions


Alternatively, the first n equations of G, corresponding to
the first sampling point, are also replaced by the DQ analog
equations of initial displacement conditions, but the second n
equations of G, corresponding to the second sampling point,
are replaced by the DQ analog equations of initial velocity
conditions, as
G 11 = I,

G 1m = 0,

G 2m = A1m I,

Q 1 = x 0,

Q 2 = x 0 ,

m = 2, 3, , N,

m = 1, 2, , N.

(26)
(27)

The DQTEM with this method of imposing initial conditions


is denoted by DQTEM2.
In order to demonstrate the dierences between the
present method and the method used in Refs. [16, 24, 25],
consider a single DOF system of
x + 2 x + 2 x = f.

(28)

Using DQTEM to solve Eq. (28) with initial condition (22),


Eq. (28) transform to the forms of

0
1
x1 x0

G21
x2 f2
G

G
22
2N

.
. .
.
.
.
..
..
..
..
.. = .. ,

G(N1)1 G(N1)2 G(N1)N xN1 fN1

AN1
AN2

ANN
xN
x0
for DQTEM1 ,
and

AN1

G31

..
.

G N1

(29)

AN2

ANN

G32

G3N

..
.

..

..
.

G N2

for DQTEM2.

G NN

x1

x2

.
..


xN1

xN


x0

x0


= f3

.
..

fN

(30)

When using the method in Refs. [16, 24, 25] to solve Eq. (28)
with initial condition (22), one has to express x1 and xN by
other variables from Eq. (23) as

0
0
x1 1

x0 0

xN
x0
A11 A1N
A12 A1(N1)

786

Y.-F. Xing, J. Guo

x2
..
.
xN1

(31)

Then discarding the DQ analog equations of governing differential equation (28) at the first and the last sampling
points, one has

B21 + 2A21

B22 + 2A22

B2N + 2A2N

B31 + 2A31

B32 + 2A32

B3N + 2A3N

..
.

..
.

..

..
.

B(N1)1 + 2A(N1)1

B(N1)2 + 2A(N1)2

B(N1)N + 2A(N1)N

Substituting Eq. (31) into Eq. (32), the variables x2 , x3 , ,


xN1 can be solved, and xN is obtained from Eq. (31) finally.
Remarks:
(1) When the derivatives at sampling points are approximated using DQ rule, the accuracy near the two end sampling points is lower than the accuracy at interior points,
so DQTEM1 is more accurate than DQTEM2, which is
validated by numerical results in Sect. 5.
(2) It can be seen that the method in Refs. [16, 24, 25] is
tedious in the imposition of initial conditions compared
with the present method.
4 Stability and accuracy
Stability and accuracy should be taken into account when investigating the eectiveness of an iterative or step-by-step
solution method [1, 34]. Here the spectral approach is employed as usual to examine the stability and accuracy of
DQTEM1 and DQTEM2 for the second-order ODEs.
4.1 Stability analysis
In stability analysis, it is convenient to work with the following single DOF system
x + (h)2 x = 0,

(33)

x1
x2

x3
x2

.. + ..
.
.

xN
xN1

f2
f3
..
.
fN1

(32)

The stability analysis of DQTEMs is the same, hence only


DQTEM1 is dealt with, see Eq. (29), then Eq. (33) is discretized to be

1
0
0

G21
G22
G23

G2N

..
..
..
..
..

.
.
.
.
.

G(N1)1 G(N1)2 G(N1)3 G(N1)N

A11
A12
A13

A1N

x1 x0

x2 0

.
.
.. = .. ,
(37)

xN1 0

x0
xN
from which one can solve x as

x1

x2

..
.

xN1

xN


S 11

S 21



..
=
.

S
(N1)1

S N1

where denotes the angular frequency. A superposed dot


denotes dierentiation with respect to dimensionless time ,
where = t/h, h and t represent the time element (or time
step size) and the real time, respectively. Since Eq. (33) is a
single DOF system, then x is given by
x T = [x1 x2 xN ],

(34)

x0

.. ,
.

x0

S 12

S 1(N1)

S 1N

S 22

S 2(N1)

S 2N

..
.

..

..
.

..
.

S (N1)2

S (N1)(N1)

S (N1)N

S N2

S N(N1)

S NN

(38)

where xi (i = 1, 2, , N) denote the values of the displacements at time sampling points, and

where S = G 1 . Based on DQ rule in Eq. (11), the velocity


of the N-th sampling point can be expressed as

G = B + (h)2 E,

(35)

x N = x0 + x0 ,

Z = x,

(36)

where

Q = 0.

(39)

Dierential quadrature time element method for structural dynamics

= AN1 S 11 + AN2 S 21 + + ANN S N1 ,

(40a)

= AN1 S 1N + AN2 S 2N + + ANN S NN .

(40b)

From Eqs. (33) and (34), one can find the relations between
(xN , x N ) and (x0 , x0 ) as

xN
x0
(41)

= D
,
x N
x0
where the Jacobi matrix D has the form of

S N1 S NN
D =
.

In fact we have the following expressions

xN1 xN1
xN2
xN
= D N
,
= D N1

x N
x N1
x N1
x N2

x0
x1
,
= D 1
.
x1
x0

787

the cases for N = 3 and N = 4 are unconditionally stable, see


Tables 1 and 2 whose elements are obtained using time step
size h = 0.01 and (1) < 106 , and from which one can see
that the usage of ChebyshevGaussLobatto points can increase the accuracy and eciency. The elements of columns
with head No. in Tables 1 and 2 represent the number of
stable intervals with respect to dierent sampling point N.

(42)

,
(43)

And the following relation can be readily verified


D = D N D N1 D 1 .

(44)

But x1 , x2 , , xN are simultaneously solved from Eq. (38)


using (x0 , x0 ), this is dierent from the classical single stepby-step method as Newmark method or RK method in which
the results of a time step are the initial conditions of the next
time step.
DQTEM would be stable if spectral radius ( D) 1.
The curves of ( D) with respect to the number N of the sample points in one time element are plotted in Figs. 1 and 2 for
DQTEM1 and DQTEM2, respectively. It follows that ( D)
of DQTEM1 is larger than or equal 1 for dierent value of N,
and ( D) of DQTEM2 can be smaller than, equal or larger
than 1. It is apparent from Fig. 1 that when N became larger,
the upper limit h of the first stable interval is larger, and
there are several intervals of h with ( D) = 1 except for
N = 3, and similarly from Fig. 2 that when N became larger,
the upper limit h of the first stable interval is also larger, and

Fig. 1 Spectral radii for dierent values of N for DQTEM1

Fig. 2 Spectral radii for dierent values of N for DQTEM2

Table 1 The stable intervals of h with ( D) = 1 for DQTEM1


N

Interval 1

Interval 2

Interval 3

Interval 4

No.

Stability

0.00 < h 2.82

Conditionally

0.00 < h 2.95

3.27 h 5.65

Conditionally

0.00 < h 3.09

3.14 h 5.65

6.93 h 8.84

Conditionally

10

0.00 < h 9.41

9.43 h 12.53

12.69 h 15.65

16.16 h 17.81

Conditionally

15

0.00 < h 15.70

15.72 h 18.85

18.87 h 21.92

21.97 h 24.46

Conditionally

50

0.00 < h 78.53

78.55 h 81.67

81.69 h 84.78

84.84 h 87.91

23

Conditionally

100

0.00 < h 179.05

179.08 h 182.19

182.24 h 185.36

185.40 h 188.34

41

Conditionally

788

Y.-F. Xing, J. Guo


Table 2 The stable intervals of h with ( D) 1 for DQTEM2

Interval 1

Interval 2

Interval 3

Interval 4

No.

Stability

Unconditionally

Unconditionally

0.00 < h 0.11

h 3.13

Conditionally

10

0.00 < h 2.02

5.91 h 6.23

6.29 h 9.43

12.58 h 16.32

Conditionally

15

0.00 < h 5.87

8.96 h 12.56

12.59 h 14.88

17.37 h 18.84

Conditionally

50

0.00 < h 51.91

55.30 h 61.57

67.70 h 69.01

69.12 h 72.25

17

Conditionally

100

0.00 < h 133.24

136.41 h 142.96

149.15 h 150.78

150.80 h 153.93

29

Conditionally

Although there are several stable intervals for


DQTEM1 and DQTEM2 when N > 4, only the first interval is of practice for multi DOF dynamic system.

more superior the DQTEM with ChebyshevGaussLobatto


sampling points.

4.2 Accuracy analysis

5 Numerical analysis

Accuracy refers to the dierence between the numerical solution and the exact solution when the numerical solution
process is stable. If the eigenvalues of D remain complex,
as

(45)
1,2 = a ib = a2 + b2 ei ,

This section aims at demonstrating the high accuracy and efficiency of DQTEM.

= arctan(b/a),
(46)

where i = 1 and b  0, is the phase of DQTEM for


one time step or element. The single-step phase error is a
measure of numerical dispersion, and usually defined as
= h = h arctan(b/a).

(47)

The time step h of DQTEM is usually much larger than that


of the commonly used DIM as RK method, especially when
N is large. When h is larger than , the single-step phase
error of DQTEM can not be directly calculated by Eq. (47),
and should be calculated by
= mod(h/) |arctan(b/a)| ,
0 < mod(h/) 1/2,
= mod(h/) ( |arctan(b/a)|),
mod(h/) > 1/2,

(48)

(49)

where mod(h/) represents the remainder of h/. Note


that Eqs. (48) and (49) are suitable for all intervals in Tables 1 and 2, and are virtually negligible when N is large,
see Table 3 which shows that for each N, all upper limits
h of the first interval of DQTEM1 are larger than those
of DQTEM2. If one uses the same h in both schemes,
the phase errors of DQTEM1 are much smaller than that of
DQTEM2, and its reason has been given in Sect. 3.
The above discussion shows that DQTEM is an eective and accurate algorithm, the larger the time element, the

5.1 Example 1
Consider the free vibration of a two DOF system for which
the governing equilibrium equations are

2 x1 0
2 0 x1 6
(50)
+
= .



x2
x2
2 4
0
0 1
The initial conditions are x T0 = [0 1], x T0 = [1 0], the

angular frequencies are 1 = 2 and 2 = 5. The relative


error used below is defined as

RelErr = abs(exact present) exact ,
(51)
5.1.1 Accuracy comparison of equally and unequally spaced
sampling points
There are two typical methods to space sampling points
when DQM is used as shown above, one is the equally
spaced points, and the other is the ChebyshevGauss
Lobatto points. Let N = 41 (the first stable interval is
0 < h 65.96 for DQTEM1), time element size h is 20 or
the time interval is 020. It can be seen from Figs. 3 and 4
that the DQTEM1 with ChebyshevGaussLobatto points is
much more accurate than that with the equally spaced points.
5.1.2 Accuracy comparison of two methods of imposing initial conditions
Assume N = 100, consider time domain [0, 50]. It follows
from Figs. 5 and 6 that DQTEM1 is much more accurate
than DQTEM2, their RelErrs are about 1014 and 1010 , respectively, for this case.

Dierential quadrature time element method for structural dynamics

789

Table 3 The comparison of h and single-step phase error of DQTEM1 and DQTEM2
N
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100

h1

h1 /T

DQTEM1
Phase error for h1

Phase error for h2

h2

DQTEM2
h2 /T
Phase error

3.09
9.41
15.70
25.12
34.55
43.97
53.40
62.82
72.25
78.53
91.10
97.38
109.95
116.23
128.80
135.08
153.91
153.93
169.64
179.05

0.49
1.50
2.50
4.00
5.50
7.00
8.50
10.00
11.50
12.50
14.50
15.50
17.50
18.50
20.50
21.50
24.50
24.50
27.00
28.50

3.165 9102
4.402 1103
1.275 3103
2.471 1103
2.321 2103
3.316 6103
2.335 7103
3.284 9103
2.077 5103
1.234 5103
1.798 7103
1.302 6103
1.557 9103
1.292 7103
1.357 8103
1.236 0103
1.274 3102
1.153 6103
6.228 8103
1.023 5102

2.096 0109
7.053 9109
1.677 4109
1.007 9105
1.728 2107
4.371 8109
6.810 71010
3.534 4108
1.541 8109
3.400 11010
4.186 7109
4.269 71010
1.070 41010
5.698 51010
1.721 71010
6.876 31010
1.616 21010
5.566 11010
1.440 81010
8.570 61011

0.11
2.02
5.87
18.05
21.23
24.73
30.01
42.93
46.45
51.91
64.81
68.72
75.97
86.87
91.77
105.47
109.45
124.20
127.74
133.24

0.02
0.32
0.93
2.87
3.38
3.94
4.78
6.83
7.39
8.26
10.31
10.94
12.09
13.83
14.61
16.79
17.42
19.77
20.33
21.21

Fig. 3 The displacement x1 and its RelErr for equally spaced points

Fig. 4 The displacement x1 and its RelErr for ChebyshevGaussLobbatto points

4.775 6108
5.245 0107
5.644 6108
7.040 7103
1.790 3104
5.423 0106
3.059 7107
1.446 3104
5.632 6106
2.365 3107
4.304 0105
2.427 5106
1.223 9106
9.603 0106
6.834 8107
3.035 5105
2.183 9106
7.600 8105
5.208 9106
2.135 0107

790

Y.-F. Xing, J. Guo

5.1.3 Eciency comparison of RK method and DQTEM


Computational cost is a crucial issue for an algorithm. Let
the time domain be [0, 100]. In order to arrive at the same
RelErr of 1010 , DQTEM1 and DQTEM2 need at least 148
and 162 sampling points, respectively. But h = 0.002 5 s
must be used in RK method, that is, the number of time steps
equals to 40 000. Note that the RK method used here is the
classical one of stage four and order four with equal time
step. Table 4 gives the CPU comparison used in calculation.
5.2 Example 2

Fig. 5 The RelErr of x1 for DQTEM1 with the ChebyshevGauss


Lobbatto points

Fig. 6 The RelErr of x1 for DQTEM2 with the ChebyshevGauss


Lobbatto points

The free vibration of a fixed-free uniform rod is analyzed


below. The Yangs module E = 125 GPa, density =
8 980 kg/m3 , diameter d = 0.1 m and length l = 1 m. The rod
is meshed into 20 uniform constant-strain spatial elements.
The initial velocity of the free-end is 1 m/s, all others are
zero, so the problem is an analogous counterpart of impact
problem.
Consider time domain [0, 3103 ], which is separated
into 10 equal DQTEM time elements with 50 Chebyshev
GaussLobbatto sample points in each element, so the size
of DQTEM element is 3104 . Note that the maximum frequency of the discretized rod is 20 = 2.578 9105 , based on
which one can see that 3 104 20 is within the first stable
interval (0.00 < h 78.53), see Table 1.
Since DQTEM1 is more accurate than DQTEM2, thus
only the solution of DQTEM1 for this dynamic problem is
compared with the exact solution obtained by the mode superposition method. DQTEM1 is also compared with the average acceleration scheme (a member of Newmark algorithm
family) since it is one of the most commonly used direct integration method, this comparison can validate the value of
the present method for engineering problem.

Table 4 The CPU time of three methods


RelErr/1010
DQTEM1

RelErr/104

CPU/s

CPU/s

148

0.030 2

127

0.018 7

DQTEM2

162

0.037 5

141

0.027 9

RK

0.002 5

100/h = 4 104

7.597 5

0.05

100/h = 2 000

0.287 3

Figure 7 shows an excellent agreement of the free-end


displacement of DQTEM1 with the exact solution, but here
only 10 time elements are used, if DQTEM1 is taken as a
step-by-step direct integration method, a time element is a
time step. In order to reach the same accuracy, Newmark
method needs 180 000 time steps and much more time than
DQTEM1, see Table 5. Figures 8 and 9 present the RelErr of
DQTEM1 and the RelErr of the Newmark method, it follows
that the RelErr of Newmark method increases with the in-

crease of number of time steps, while the RelErr of DQTEM


not.
It is obvious that DQTEM requires less computational
cost and are more eective than RK and Newmark methods,
and some remarks are given as follows:
(1) When a small number of sampling points are used,
that is to say, the time step size is larger, the solutions of
such conventional schemes as RK method would encounter
a sharp drop of accuracy for the long-term responses due to

Dierential quadrature time element method for structural dynamics

791

the artificial damping and the phase error accumulation while


DQTEM can hold high accuracy and eciency. In other
words, the artificial damping and the phase error accumulation of DQTEM are significantly smaller compared with
the conventional step-by-step methods.
(2) For the whole time domain interested, one DQ time
element is usually enough if N is large enough, and the precision can be significantly improved by increasing a few number of the sample points.
(3) In the conventional step-by-step schemes, the results of the present time step are the initial conditions of the
next time step, so all the initial conditions are approximate
except for the first time step; and the DQTEM solutions for
all time sampling points are calculated simultaneously using
the same exact initial conditions when only one time element
is employed. Nevertheless, DQTEM can also serve as an accurate step-by-step method, as shown in Sect. 5.
(4) One can find from the above numerical results that
DQTEM can achieve more accurate solutions than other
step-by-step schemes by using a considerably smaller number of sample points. In some sense, DQTEM solutions can
be regarded as benchmarks to validate other lower order integration methods.
(5) The same problem can be solved using the ode45
(a RK solver) in MATLAB, by which one can obtain the results with only a relative dierence error of 103 at most.

Fig. 7 The displacement of free end of the rod

6 Conclusions
Fig. 8 The RelErr of xfreeend for DQTEM1 with the Chebyshev
GaussLobbatto points

Fig. 9 The RelErr of xfreeend for Newmark method


Table 5 The CPU time of two methods (RelErr 105 )
h

Time steps

DQTEM1 Newmark 1.666 710

CPU/s

10 (50 points of an element) 0.226 0


8

180 000

158.983 7

In this paper, the ideas, advantages and disadvantages of the


direct integration methods and the time finite element methods were reviewed. It was concluded that for structural dynamic ODEs, the numerical simulation would not be satisfactory for somewhat long time history by using the available
direct integration method of single/multi steps and TFEM
because of their accumulated artificial damping and phase
error.
A time element method called DQTEM was presented
for solving structural dynamic ODEs on the basis of DQ rule,
extensive numerical results shown that DQTEM possesses
advantages over TFEM and the direct integration methods of
single/multi steps which have been widely used nowadays.
Particularly, DQTEM with the ChebyshevGaussLobbatto
sampling points and using the first method of imposing initial conditions has much smaller artificial damping and phase
error, implying that this method is practical and exhibits a
better capability of long-term numerical simulation.

References
1 Bathe, K.J., Wilson, E.L.: Numerical Methods in Finite Element Analysis. Prentice-Hall, Englewood Clis, New Jersey

792
(1976)

2 Runge, C.: Uber


die numerische Auflosung von Dierentialgleichungen. Mathematische Annalen 46, 167178 (1895)
3 Kutta, W.: Beitrag zur naherungsweisen Integration totaler Differentialgleichungen. Z. Math. Phys 46, 435453 (1901)
4 Wilson, E.L., Farhoomand, I., Bathe, K.J.: Nonlinear dynamic
analysis of complex structures. Earthquake Engineering and
Structural Dynamics 1, 242252 (1973)
5 Bathe, K.J., Wilson, E.L.: Stability and accuracy analysis of
direct integration methods. Earthquake Engineering and Structural Dynamics 1, 283291 (1973)
6 Newmark, N.M.: A method of computation for structural dynamics. ASCE Journal of the Engineering Mechanics Divisions
85, 6794 (1959)
7 Wood, W.L., Bossak, M., Zienkiewicz, O.C.: An alpha modification of Newmarks method. International Journal for Numerical Methods in Engineering 15, 15621566 (1980)
8 Feng, K.: Dierence scheme for Hamiltonian formalism and
symplectic geometry. Journal of Computational Mathematics
4, 279289 (1986)
9 Fried, I.: Finite-element analysis of time-dependent phenomena. AIAA Journal 7, 11701173 (1969)
10 Argyris, J.H., Scharpf, D.W.: Finite elements in time and space.
Nuclear Engineering and Design 10, 456464 (1969)
11 Ri, R., Baruch, M.: Time finite element discretization of
Hamiltons law of varying action. AIAA Journal 22, 1310
1318 (1984)
12 Simkins, T.E.: Finite elements for initial value problems in Dynamics. AIAA Journal 10, 13571362 (1981)
13 Wilson, E.L., Nickell, R.E.: Application of the finite element
method to heat conduction analysis. Nuclear Engineering and
Design 4, 276286 (1966)
14 Zienkiewicz, O.C.: A new look at the Newmark, Houbolt and
other time stepping formulas. A weighted residual approach.
Earthquake Engineering and Structural Dynamics 5, 413418
(1977)
15 Zienkiewicz, O.C., Wood, W.L., Taylor, R.L.: An alternative single-step algorithm for dynamic problems. International
Journal for Numerical Methods in Engineering 8, 3140 (1980)
16 Fung, T.C.: Solving initial value problems by dierential
quadrature methodPart 2: second-and higher-order equations. International Journal for Numerical Methods in Engineering 50, 14291454 (2001)
17 Eftekhari, S.A., Khani, M.: A coupled finite element-dierential quadrature element method and its accuracy for moving
load problem. Applied Mathematical Modeling 34, 228237
(2010)
18 Bellman, R., Casti, J.: Dierential quadrature and long term integration. Journal of Mathematical Analysis and Applications
34, 235238 (1971)
19 Bert, C.W., Malik, M.: Dierential quadrature method in computational mechanics: a review. Applied Mechanics Reviews

Y.-F. Xing, J. Guo


49, 128 (1996)
20 Zong, Z.: A variable order approach to improve dierential
quadrature accuracy in dynamic analysis. Journal of Sound and
Vibration 266, 307323 (2003)
21 Tanaka, M., Chen, W.: Dual reciprocity BEM applied to
transient elastodynamic problems with dierential quadrature
method in time. Computer Methods in Applied Mechanics and
Engineering 190, 23312347 (2001)
22 Fung, T.C.: Stability and accuracy of dierential quadrature
method in solving dynamic problems. Computer Methods in
Applied Mechanics and Engineering 191, 13111331 (2002)
23 Shu, C., Kha, W.: Numerical simulation of natural convection
in a square cavity by SIMPLE-generalized dierential quadrature method. Computers and Fluids 31, 209226 (2002)
24 Bert, C.W., Wang, X., Striz, A.G.: Dierential quadrature for
static and free vibration analyses of anisotropic plates. International Journal of Solids and Structures 30, 17371744 (1993)
25 Shu, C., Du, H.: A generalized approach for implementing general boundary conditions in the GDQ free vibration analysis of
plates. International Journal of Solids and Structures 34, 837
846 (1997)
26 Fung, T.C.: Solving initial value problems by dierential
quadrature method-Part 1: first-order equations. International
Journal for Numerical Methods in Engineering 50, 14111427
(2001)
27 Wu, T.Y., Liu, G.R.: A dierential quadrature as a numerical
method to solve dierential equations. Computational Mechanics 24, 197205 (1999)
28 Wu, T.Y., Liu, G.R.: The generalized dierential quadrature
rule for initial-value dierential equations. Journal of Sound
and Vibration 233, 195213 (2000)
29 Xing, Y.F., Liu, B.: High-accuracy dierential quadrature finite element method and its application to free vibrations of
thin plate with curvilinear domain. International Journal for
Numerical Methods in Engineering 80, 17181742 (2009)
30 Quan, J.R., Chang, C.T.: New insights in solving distributed
system equations by the quadrature methods I. Analysis.
Computers in Chemical Engineering 13, 779788 (1989)
31 Bert, C.W., Malik, M.: The dierential quadrature method for
irregular domains and application to plate vibration. International Journal of Mechanical Sciences 38, 589606 (1996)
32 Shu, C., Richards, B.E.: Application of generalized dierential
quadrature to solve two-dimensional incompressible Navier
Stokes equations. International Journal for Numerical Methods
in Fluids 15, 791798 (1992)
33 Bartels, R.H., Stewart, G.W.: Solution of the matrix equation
AX + XB = C. Communications of the ACM 15, 820826
(1972)
34 Hughes, T.J.R.: The Finite Element Method: Linear Static and
Dynamic Finite Element Analysis. Prentice-Hall, Englewood
Clis, NJ (1987)

You might also like