Beruflich Dokumente
Kultur Dokumente
Aram Soroushian∗, †
Structural Engineering Research Center, International Institute of Earthquake Engineering
and Seismology (IIEES), 19395 Tehran, Iran
SUMMARY
The integration step size is the main algorithmic parameter in time integration analysis. Nowadays, for
time integration with the complete records of digitized excitations, the integration steps cannot be set
larger than the excitation steps. Considering the practical importance of this restriction, and with the aim
of structural dynamic analysis with less computational cost, this paper intends to extend conventional
time integration analyses to analyses, if needed, carried out with steps larger than the excitations steps. In
view of few simplifying assumptions, and presenting a new theorem on responses convergence, a
technique is developed and a computational procedure is set. Being based on convergence-oriented
redefinition of digitized excitations, the implementation of the new procedure is simple and, though,
sacrificing some accuracy, can considerably reduce the total computational cost. Copyright q 2008 John
Wiley & Sons, Ltd.
KEY WORDS: time integration; digitized excitations; integration step size; computational cost;
convergence; accuracy
1. INTRODUCTION
The behaviors of many structural systems are being governed by the effects of inertial forces, e.g.
building structures’ behaviors in regions with high seismic risk. In order to analyze the structures’
dynamic behaviors, it is an accepted approach to define the structural models, set the mathematical
models, and, discretizing the mathematical models in space [1–3], try to analyze the resulting
∗ Correspondence to: Aram Soroushian, Structural Engineering Research Center, International Institute of Earthquake
Engineering and Seismology (IIEES), 19395 Tehran, Iran.
†
E-mail: a.soroushian@iiees.ac.ir, aram.soroushian@gmail.com
Additional constraints : Q
where, t and tend imply the time and the duration of the dynamic behavior; M is the mass matrix;
fint and f(t) stand for the vectors of internal force and excitation; u(t), u̇(t), and ü(t) denote the
vectors of displacement, velocity, and acceleration; u0 , u̇0 , and fint0 define the initial status of the
model (regarding the essentiality of considering fint0 in Equations (1), also see [4]); and finally, Q
represents some restricting conditions, e.g. additional constraints in problems involved in impact
or elastic–plastic behavior [5, 6], all in view of the degrees of freedom set for the model. The most
versatile approach to analyze Equations (1) is time integration by an approximate method [7, 8].
Since time integration cannot in general be based on exact formulations [8, 9], by implementing
different approximations, many integration methods are developed in the past decades; e.g. see
[10–15]. The Newmark family [16], central difference [17], and HHT [18] are the most popular
time integration methods, see [2, 8, 10, 19]. To briefly review the process of time integration, the
sizes of time steps or a criterion for adaptive time stepping, e.g. see [20], should be set before
any computation. Then, the analysis starts from the initial conditions (if needed, after applying
a starting procedure [21]), and with marching along the time axis (see Figure 1), approximate
responses are being determined at discrete time stations, sequentially, based on the formulation of
the integration method. Accordingly, time integration is by its nature a computationally expensive
analysis tool, leading to approximate responses. As an immediate consequence, in the evaluation
of time integration methods, special attention should be paid to (also, see [22, 23]):
1. Convergence of the computed approximate value of any arbitrary component of the response,
Ua , to the corresponding exact value, U [24–27], i.e.
lim Ua =U (2)
t→0
(t is a positive-definite parameter, linearly controlling the sizes of time steps throughout
the integration). In view of the Lax equivalence theorem [24, 25], for well-posed problems
[19, 26], the approximately computed responses converge only when they are both numerically
consistent and numerically stable [19, 24–26]. Equation (2) is also conceptually equivalent
to the convergence plot in Figure 2 [26, 27], where, from the definition of error [28],
E = |Ua −U | (3)
t
Time station : 0 1 2 ………………… i-1 i ……..
Figure 1. Typical arrangement of time steps and time stations at time integration analyses.
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
A TECHNIQUE FOR TIME INTEGRATION ANALYSIS 2089
log (E )
1
q
−∞
log (∆t)
Figure 2. Typical convergence plot for computed responses converging to exact responses.
(Equation (3) is meaningless in real computations unless being considered relatively), and,
from the definition of the rate of convergence [26], q (the slope of the linear section in the
convergence plot) is the positive integer, satisfying
E −Ct q
q1, lim =0 (4)
t→0 t q
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
2090 A. SOROUSHIAN
where the two terms in the parentheses, respectively, represent the requirements of convergence
and accuracy; h is the minimum step size essential for preserving convergence (numerical stability
and consistency) [9, 19, 26], and T standing for the smallest period of oscillations, when being
divided by 10 (or 100), implies the existing accuracy-related comment, for the selection of t
[37–39] (not having the responses before the analyses, when, also not having any information
about the least dominating periods of vibration, the smallest natural periods of vibration at t = 0
are, in general, being considered as the T in Equations (5)). Efforts for relaxing Equations (5)
and implementing larger t are reported in the literature, e.g. adaptive time stepping [20, 40] and
commenting the implementation of unconditionally stable time integration methods (h = ∞), see
[19, 41]. Nevertheless, in real problems, convergence and accuracy are not necessarily the only
restrictions affecting the selection of the t. An additional restriction is the step size by which some
excitations are recorded [8, 9, 37]; not all excitations are available analytically, e.g. the current
seismological instrumentations and computational methods can result in only digitized ground
strong motions, recorded at specific, temporally separate time instants (see [42]). This restriction
affects Equations (5) by f t, the step by which the digitized excitation is recorded, i.e.
T
t = min h, , f t for linear systems
10
(6)
T
t = min h, , f t for nonlinear systems
100
In problems with digitized excitations, f t may be the governing term in Equations (6). In such
cases, although, implementing the small step size, equal to f t, as the integration step size is
computationally more expensive than implementing the step size obtained from Equations (5)
(by leading to a higher total number of integration steps), with the latter step size, considering
the complete excitation history, is not established yet. (Simply omitting the inter-integration-step
excitations, though will cancel the computational cost, caused by the dominating f t, is vague
from the points of view of accuracy and convergence.) A practical potential engineering case in
this regard is the seismic analysis of advanced structural systems of modern buildings, bridges,
tunnels, etc. Seismic excitations are in general being applied as ground accelerations, available
as digitized records [42, 43]. Since specific engineering experience regarding the recent modern
structural systems (now many all around the world) is rare, and besides financial aspects are of
important issues especially for large constructions, such structural analyses also need to take into
account the nonlinear behavior (see [8, 37, 43]), the possibilities of very near natural frequencies,
unproportional dampings [8, 37], etc. Therefore, as the only almost completely versatile structural
dynamic analysis tool [8, 9, 11, 37], time integration should be implemented in the analyses. In the
stage of structural analysis, in general, after some nondetailed modeling (with moderate number
of degrees of freedom) and analysis of the structural systems, based on the obtained results, the
systems are remodeled in further detail with much higher number of degrees of freedom, and
the analyses are being repeated for further specific detailed studies. In the latter analyses, the
dominating periods of oscillations (the T in Equations (5) and (6)) are already approximately known
from the first analyses, and if the dynamic effects are considered important, the analyses of the
structural systems subjected to ground accelerations should be, according to the current knowledge,
carried out with steps not larger than the excitation steps (the f t in Equations (6)). In view of
Equations (6), when f t is much smaller than min(h, T /(10 or 100)), considerable computational
cost should be spent mainly because of the smallness of f t (not convergence or accuracy).
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
A TECHNIQUE FOR TIME INTEGRATION ANALYSIS 2091
t
Integration stations : 0 1 2 ......
Integration steps : 1 2 ......
Considering the practical importance of convergence, computational cost, and accuracy [22, 23],
the objective of this paper is to indirectly take into account the inter-integration-step excitations
(if existing), in conventional time integration analyses, i.e. arrive at a time integration analysis
procedure for considering all the excitation histories, without directly integrating Equations (1)
with steps equal to f t. In other words, by reasonably sacrificing a part of the accuracy achievable
by time integration with t = f t, most of the computational cost difference between analyses,
respectively, with step sizes t = f t and t = min(h, T /(10 or 100))> f t is intended to be
saved, also without negative effects on convergence. With no specific relevant documentation in
the literature, first the following four simplifying, still practical, assumptions are imposed on the
study (see Figure 3, for the first three):
1. The excitation steps, f ti , i = 1, 2, . . . , are equally sized, i.e.
2. The integration steps, ti , i = 1, 2, . . . , also known as the time steps, are equally sized, i.e.
3. The excitation steps are embedded by the integration steps (the first time station, i.e. t0 , is a
station for both excitation and integration),
t
∃ n ∈ Z +, = n<∞ (9)
f t
4. The f(t) in Equations (1) is a digitized representation (recorded at separate time instants) of
an actual excitation, g(t), smooth [44], with respect to time, i.e.
(hence, the temporal derivative of f(t) (see [45]), though rarely known, exists).
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
2092 A. SOROUSHIAN
Then, the main theory is explained in Section 2; where, after presenting, attempting to prove,
and numerically studying a basic statement (theorem) on responses convergence, a technique for
indirectly considering inter-integration-step excitations in the integration process is introduced and
its parameters are set by a brief study on accuracy, computational cost, and computational ease.
Later, in Section 3, a computational procedure for practical implementation of the new technique
is presented. This paper is then proceeded, numerically demonstrating the capabilities of the
technique, and finally, the achievements are summarized in Section 5.
Responses convergence trend, with or without implementing the technique to be presented in this paper
Responses convergence trend, after implementing the technique to be presented in this paper
The range of integration steps, not acceptable for time integration, according to the current literature
Integration step size, when the excitation is digitized at separate instants (Eqs. (6))
Figure 4. For f t<min(h, T /(10 or 100)), the typical convergence trend planned for the
errors of time integration analyses.
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
A TECHNIQUE FOR TIME INTEGRATION ANALYSIS 2093
lead to approximate responses (with acceptable computational features) for the original problem,
see [46].
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
2094 A. SOROUSHIAN
Considering Equation (14) throughout the interval 0t < ti , along with the assumption in the
statement of the theorem, i.e.
e0 = Ō (16)
i
ei = Ai−k (L(Pk )−k ) (17)
k=1
ei = ei1 +ei2
i
ei1 = Ai−k L(Pk ) (18)
k=1
i
ei2 = − Ai−k k
k=1
In view of Equation (14) and the definition of local truncation error [50], the i , in Equation (14),
is the local truncation error, at step i, in absence of the excitations’ approximation at ti , i.e.
In other words, i , being dependent on the computed responses, e.g. see [51, 52], is also under the
influence of (Pi ). As a result, while ei1 reflects the explicit effect of the excitations’ approximations
on the response error at ti , ei2 denotes the corresponding implicit effect of (Pi ). This leads to
the point that, if ei1 does not converge with the rate of the order of accuracy, ei2 may also be
affected and does not converge with the rate of the order of accuracy. Hence, also from Equations
(11)–(14), we specifically need to study the essentialities of convergence with the rate q0 for the
ei1 in Equations (18).
In this regard, consider the formal definition of ei , e.g. see [19],
(n d −1)
T
vi −v(ti ) v̇i − v̇(ti ) v̈i − v̈(ti ) v −v (n d −1) (ti ) (n d )
ei = −1 −2
··· i vi −v (n d ) (ti ) (20)
t d
n t d
n t d
n t
where, n d denotes the number of the temporal derivatives of the displacement implemented in
the definition of yi in Equation (11), also the number of the temporal derivatives needed in the
(n )
integration process; the superscript T is the matrix transposition operator; and vi , v̇i , v̈i , . . . , vi d ,
respectively, stand for the vectors implying displacement, velocity, acceleration, and other temporal
derivatives of v at t = ti in correspondence with the oscillatory mode under consideration.
Equation (20) can also typically be considered valid for ei1 . From Equations (17) and (18) and
the general structure of the LP vector, e.g. see Table I, it is apparent that, when the requirements
of convergence in absence of excitations’ approximations are provided, i.e. the exact response
associated with the exact excitation is asymptotically achievable and Ak 1, ∀k ∈ Z + [10, 19, 26],
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
A TECHNIQUE FOR TIME INTEGRATION ANALYSIS 2095
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
2096 A. SOROUSHIAN
q = q0 , it is essential to set q
with attention to Equation (23). With a similar argumentation, for
values of q
not satisfying Equation (23), if the conventional requirements of numerical stability
and consistency [19, 26] are provided and n s = 1, responses would converge with the rate of q
.
Even regardless of the four assumptions stated in Section 1, single-step integration scheme,
linear behavior, and classical (or as referred to in some references proportional) damping [2, 8–10]
are the assumptions in the discussion above (see Equation (11)). The last two assumptions, i.e.
linear behavior and classical damping, can likely be relaxed by implementing more complicated
mathematics, considering the nonlinearity residuals in the formulation, see [27, 53], and the fact
that time integration converts the nonlinear mathematical models to piece-wisely linear models
[21, 27, 54, 55]. The first assumption, i.e. single-step integration, though probably affects Equation
(23), can be disregarded in view of facts such as: (1) considering adaptive time stepping and prob-
lems involved in nonlinearity, single-step integration is the conventional practice, recommended in
the literature by times, see [11, 41]; (2) many multi-step integration schemes can simply be refor-
mulated as single-step methods, e.g. see [21, 56] for the Houbolt method (also see [30]). Hence, the
proof attempt can be considered concluded.
(In view of the Lax equivalence theorem [25], it is reasonable to consider the new requirement
as a part of the consistency requirement. In other words, excitations’ approximations negatively
affect consistency unless Equation (23) is satisfied, i.e. for the convergence of the responses
computed for a well-posed problem, besides numerical stability, consistency should be maintained
in the sense that both the satisfaction of the equation of motion and the exact excitation should be
asymptotically achievable; the convergence rate would then equal the minimum of the rates of the
two approaches.)
Numerical illustration: In view of both the plan to implement Equations (23) and (24) in arriving
at the objective of this paper and the discrepancy existing in the literature (regarding the effects
of excitations’ approximations on the rate of convergence) [47, 48], presenting a simple numerical
example sounds instructive. In this regard, an illustrative example is studied below, once with
q
= 1 and then again with q
= 2. Consider the simple problem in the following equation (studied
in the literature by times [39, 57, 58]):
Mü+Ku = f(t), 0t<6.72
u(t = 0) = u̇(t = 0) = Ō
(25)
2 0 6 −2 0
M= , K= , f(t) =
0 1 −2 4 100
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
A TECHNIQUE FOR TIME INTEGRATION ANALYSIS 2097
and considering q
= 1 (or 2), the new problem
Mü+Ku = f̃(t), 0t<6.72
u(t = 0) = u̇(t = 0) = Ō
(28)
2 0 6 −2 0
M= , K= , f̃(t) =
0 1 −2 4 100+(10w)q
w = t (29)
and decreasing the size of the time steps, t, the new excitation at Equations (28) would converge
to the excitation of Equations (25) with the rate q
. Therefore, when selecting integration methods
with q0 = 2 and n s = 1, for instance, the average acceleration, the single-step central difference, the
Wilson-, and the generalized- methods, for time integration analysis of Equations (28), in order
that the computed responses converge with the rate q = q0 = 2 to the exact responses (reported in
Equations (26)), it is essential to provide q
= 2 (see Equation (24)). Analyzing the new problem
(Equations (28)) using the four integration methods, each in the two cases q
= 1 and q
= 2, several
times with different step sizes, all satisfying Equations (5), i.e. t = 0.1, 0.05, 0.025, . . . , and then
analyzing the errors in view of different norms [59], result in the convergence plots depicted in
Figures 5 and 6, as an apparent evidence for the validity of Equations (23) and (24).
1.E+00 1.E+00
Error (%)
Error (%)
1.E-01 1.E-01
q = 1 ≠ q0 q = 1 ≠ q0
1.E-02 1.E-02
1.E-03 1.E-03
0.001 0.01 0.1 1 0.001 0.01 0.1 1
(a) Integration step size, ∆ t (sec) (b) Integration step size, ∆ t (sec)
1.E+01 1.E+01
1.E+00 1.E+00
Error (%)
Error (%)
1.E-01 1.E-01
q = 1 ≠ q0 q = 1 ≠ q0
1.E-02 1.E-02
1.E-03 1.E-03
0.001 0.01 0.1 1 0.001 0.01 0.1 1
(c) Integration step size, ∆ t (sec) (d) Integration step size, ∆ t (sec)
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
2098 A. SOROUSHIAN
(a) (b)
(c) (d)
Additional constraints : Q
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
A TECHNIQUE FOR TIME INTEGRATION ANALYSIS 2099
where,
t =0: f̃i = f(t = 0) = f0
0<t<tend : f̃i = af(ti )+(1−a)[b1 (f(ti+d1 )+f(ti−d1 ))+b2 (f(ti+d2 )
(33)
+f(ti−d2 ))+· · ·+bn
(f(ti+dn
)+f(ti−dn
))]
t = tend : f̃i = f(tend )
and
n
1
bk =
k=1 2
ci
di = (34)
t
t
ci = i , i = 1, 2, . . . , n
Equation (23), and specifically, Equation (24), would be satisfied; consequently, the responses
computed for Equations (32) by time integration with steps equal to t would converge with a
rate equal to the order of accuracy q0 to the exact responses of Equations (1) (see Figure 4),
when the requirements established for convergence in the absence of excitations’ approximations
[19, 25, 26] are provided. This is specifically true for q0 = 2 and n s = 1, which is the most common
and recommended case in the literature, see [41]. Now, if we relate Equations (33) and (34) to the
assumptions in Section 1 and the objective of this paper, by considering 1<n<2n
(n is defined in
Equation (9)), we can claim that, by implementing Equations (33) and (34), the excitations at all
stations digitized at steps t/n, are used in defining the new excitation, digitized at steps t; in view
of Equation (23) and the conventional values of q0 = 2 and n s = 1, time integration of Equations
(1), considering one of the two excitations (see Equations (1) and (32)), would result in responses
converging with the rate of order of accuracy q0 to the exact response of Equations (1). This
implies arriving at an inherently computationally inexpensive technique for time integration with
steps larger than the steps of excitation, with the cost of sacrificing some accuracy. Nonetheless,
the n
, addressed above, and the a and bk at the right-hand sides of Equations (33) and (34) are
still undetermined. The remainder of this section deals with these parameters, from the standpoints
of accuracy, and computational cost and ease.
Nevertheless, not to consider the inter-integration-step excitations more than once, and besides,
for further accuracy (in view of the notion of the Taylor expansion as the starting point for
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
2100 A. SOROUSHIAN
Figure 7. Replacement of the original excitation with the excitation in Equations (33)–(35).
the formulation suggested for f̃), the smallest acceptable value for n
, such that not to miss any
excitation station, seems to be the best selection. Hence, as displayed in Figure 7,
t = t : n
= n −1
⎧n
⎪
⎨2, n = 2 j, j ∈ Z +
Parameters a and bk : both directly affect the accuracy, and for both, in view of specifically
Equations (30) and (33), an appropriate theoretical selection will depend on the excitations and
the temporal derivatives of the excitations. Therefore, in this seemingly first study on the subject,
considering that the parameter a is indeed a controller of the contribution of the original excitation at
ti (in the new excitation at ti ) and that bk , and specifically, (1−a)bk , imply the contribution of
the original digitized excitations at t +dk and t −dk , not located at integration stations (in the
new excitation), it seems to be a simple, reasonable, and computationally inexpensive selection to
consider the constant unbiased contributions below,
a = 1−a
(36)
(1−a)bk = const.
a= 1
2
1 (37)
bk =
2n
(Regarding the constant values selected for the three parameters above, it is also worth noting that
a profound study on n
, a, and bk would result in formulations depending on the excitation and/or
the response, and probably on their temporal derivatives. This may need the computation of the
three parameters at each integration step, which, requiring considerable computational cost, may
destroy the main advantage and justification of time integration with larger steps, i.e. less compu-
tational cost.) Consequently, in view of Equations (33)–(37) and the four assumptions addressed in
Section 1, the main formulation proposed in this paper for replacing excitations (digitized at steps
smaller than the integrations steps) with excitations digitized at integration stations and enabling
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
A TECHNIQUE FOR TIME INTEGRATION ANALYSIS 2101
integration with steps larger than the excitations digitizations steps is summarized in the following
equations:
ti = 0 : f̃i = f(ti )
1 1 n
where, n
can be obtained from Equations (35). Provided we have appropriately determined the
three parameters n
, a, and bk , we can now claim that the objective of this paper is attained
(as also implied with the ‘a’ in the start of the title of this paper, the proposed technique is only
one technique to materialize time integration with steps larger than the steps of excitation). To
verify the above claim and specifically to study the performance of the proposed technique, a
computational procedure is presented next.
3. PRACTICAL IMPLEMENTATION
3.1. Procedure
There is no claim that the selections made for the parameters in Section 2 are the best; only a
new technique for integration with steps larger than the steps of the excitations’ digitizations is
defined; this technique is acceptable from the points of view of convergence and computational
cost, and sounds reasonable for accuracy. A computational procedure for applying the proposed
technique is presented below:
(a) Select a conventional time integration method; if needed, apply a starting procedure, e.g.
see [21]; in view of Equations (1), determine the largest t and n, satisfying
T
t = n f tmin h, for linear systems
10
T
t = n f tmin h, for nonlinear systems (39)
100
t tend
where, n is a positive integer, h and T are defined in Equations (5), and the last relation is
a trivial restraint, essential in computerizing the procedure.
(b) If n = 1, considering
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
2102 A. SOROUSHIAN
procedure would be more effectual for methods of first or second order of accuracy; for methods
with higher orders of accuracy, in view of Equations (23) and (24), we can re-formulate the
proposed technique by determining bi from the first equation of Equations (34) and the set of
algebraic equations below:
2j
j = 0, 1, 2, . . . J : bi ci = 0 (41)
i=1
where, J is the sufficient number of the terms on the right hand sides of Equations (30) that when
omitted, the q
will change, such that Equation (23) can be satisfied (since, in general, J n
,
the bi with the above two restrictions exist). It should meanwhile be noted that the proposed
procedure is applicable to systems with linear or nonlinear behaviors. To explain better, after
the stage ‘b’ in the above procedure, the procedure is identical to conventional time integration
analyses; hence, under the four assumptions in Section 1, we can always implement the proposed
technique.
3.4. Accuracy
By implementing the new technique, responses accuracies may decrease, because of replacing the
exact excitations with approximate excitations. Hence, as implied in Sections 1 and 2.1, imple-
menting (or not implementing) the new technique is inherently a trade-off between computational
cost and accuracy, which, by the selections made in Section 2.4, is attempted to be reasonable. The
numerical study in the next section is dedicated to demonstrating that the theoretical discussion,
leading to the computational procedure in this section, materializes this objective.
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
A TECHNIQUE FOR TIME INTEGRATION ANALYSIS 2103
4. NUMERICAL STUDY
0.6
f ∆t = 0.01 sec
0.4
0.2
f (t)
0
-0.2
-0.4
-0.6
0 10 20 30 40
Time (sec)
Figure 8. The digitized excitation for the SDOF model in Equations (42).
0.2 0.2
Displacement (m)
Velocity (m/s)
0.1 0.1
0.0 0.0
-0.1 -0.1
-0.2 -0.2
0 10 20 30 40 0 10 20 30 40
Time (sec) Time (sec)
Figure 9. The exact response for the SDOF model in Equations (42).
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
2104 A. SOROUSHIAN
0.2 0.2
Displacement (m)
Velocity (m/s)
0.1 0.1
0.0 0.0
-0.1 -0.1
-0.2 -0.2
0 10 20 30 40 0 10 20 30 40
(a) Time (sec) Time (sec)
0.2 0.2
Displacement (m)
Velocity (m/s)
0.1 0.1
0.0 0.0
-0.1 -0.1
-0.2 -0.2
0 10 20 30 40 0 10 20 30 40
(b) Time (sec) Time (sec)
0.2 0.2
Displacement (m)
Velocity (m/s)
0.1 0.1
0.0 0.0
-0.1 -0.1
-0.2 -0.2
0 10 20 30 40 0 10 20 30 40
(c) Time (sec) Time (sec)
0.2 0.2
Displacement (m)
Velocity (m/s)
0.1 0.1
0.0 0.0
-0.1 -0.1
-0.2 -0.2
0 10 20 30 40 0 10 20 30 40
(d) Time (sec) Time (sec)
Figure 10. Responses for Equations (42) computed by the single-step central difference
method, for the new technique: (a) n = 2 (t = 0.02 s); (b) n = 4 (t = 0.04 s); (c) n = 15
(t = 0.15 s); and (d) n = 60 (t = 0.60 s).
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
A TECHNIQUE FOR TIME INTEGRATION ANALYSIS 2105
Exact Excitation
0.6 0.6
0.4 0.4
0.2 0.2
f (t )
f (t )
0.0 0.0
~
~
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
0 10 20 30 40 0 10 20 30 40
(a) Time (sec) (b) Time (sec)
0.6 0.6
0.4 0.4
0.2 0.2
f (t )
f (t )
0.0 0.0
~
~
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
0 10 20 30 40 0 10 20 30 40
(c) Time (sec) (d) Time (sec)
Figure 11. The excitations implemented in the analysis of the SDOF model in Equations
(42), for the new technique: (a) n = 2 (t = 0.02 s); (b) n = 4 (t = 0.04 s); (c) n = 15
(t = 0.15 s); and (d) n = 60 (t = 0.60 s).
Table II. The study of computational costs associated with the responses in Figure 10.
Integration step size, t (s) Computational cost (s) Cost saved (%)
implementing the technique proposed here, the integration steps cannot be set larger than 0.01 s
(t = f t = 0.01 s). In view of these considerations, central difference time integration, without
implementing the new technique, with integration steps equal to 0.01 s, and then, time integration,
again, with the central difference method and implementing the new technique, several times, with
steps equal to t = 2 f t, t = 4 f t, t = 15 f t, t = 60 f t, set in stage ‘a’ of the proposed
procedure, would result in the responses depicted in Figure 10 (for each case, the excitation
obtained from stage ‘b’ of the proposed procedure is reported in Figure 11). Since, for linear
systems, u and u̇ completely represent the systems statuses, Figure 10 reveals the adequacy of the
analyses from the point of view of accuracy. This in addition to the computational costs saved
by the new technique reported in Table II demonstrate that while the losses of accuracies by
implementing the new technique are tolerable, the saved computational costs can be considerable.
Specifically, in order to display the independency of the proposed technique from the integration
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
2106 A. SOROUSHIAN
1.E+04 1.E+04
1.E+01 1.E+01
Error (%)
Error (%)
1.E-02 1.E-02
1.E-05 1.E-05
1.E-06 1.E-05 1.E-04 1.E-03 1.E-02 1.E-01 1.E+00 1.E-06 1.E-05 1.E-04 1.E-03 1.E-02 1.E-01 1.E+00
(a) Integration step size, ∆t (sec) (b) Integration step size, ∆t (sec)
Figure 12. Convergence of the responses computed by implementing the new technique in
time integration of Equations (42), with the: (a) central difference method and (b) HHT
( = −0.1, = 0.3025, = 0.6) method.
method, the study is repeated with the HHT ( = −0.1, = 0.3025, = 0.6) method [18, 41] and
led to the results conceptually similar to those reported in Figure 10 and Table II. The convergence
trends in analyses with both integration methods are reported in Figure 12. In view of Figure 2 and
the fact that, for both the central difference and the HHT ( = −0.1, = 0.3025, = 0.6) methods,
q0 = 2, the effect of the new technique on convergence (implied in Figure 12) is acceptable.
Figure 12, together with Figure 10 and Table II, clearly evidence the validity of the claim made
in this paper, i.e. the capability to time integrate Equations (1) (with acceptable computational
features), with steps larger than the excitations’ digitization steps, using all the original excitation
record, is achieved. (The differences between Figures 12(a) and (b) also imply that, as noted in
Section 2.1, the excitations defined via Equations (38) are a part of the numerical model.)
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
A TECHNIQUE FOR TIME INTEGRATION ANALYSIS 2107
u1 u2
m1 = 1 Kg
d m 2 = 10 m1
s
k 1 = k 2 = 0 .2 N m
Barrier Damping : Negligible
k1 k2
s =1 m
m1 m2
d = 0 .1 m
All impacts are elastic
(a) u g (ground acceleration)
0.4
0.3 f ∆ t = 0.02 sec
0.2
0.1 ⎡1 0 ⎤ ⎧1⎫
u g 0.0 f (t ) = −9.81⎢ ⎥⎨ ⎬ u g
-0.1 ⎣0 10⎦ ⎩1⎭
-0.2
-0.3
-0.4
0 5 10 15 20 25 30
(b) Time (sec)
Figure 13. The structural system under consideration as the second example:
(a) structural system and (b) excitation.
For the first degree of freedom For the second degree of freedom
0.3 0.4
Displacement (m)
0.2
Velocity (m/s)
0.2
0.1
0.0 0.0
-0.1
-0.2
-0.2
-0.3 -0.4
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec) Time (sec)
Figure 14. The exact response of the system introduced in Figure 13 when the displacement
constraints are removed (linear behavior).
For the first degree of freedom For the second degree of freedom
6 20
Displacement (m)
Velocity (m/s)
3 10
0 0
-3 -10
-6 -20
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec) Time (sec)
Figure 15. The almost exact response of the system introduced in Figure 13.
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
2108 A. SOROUSHIAN
6 6
4 4
2 2
u1 0 u2 0
-2 -2
-4 -4
-6 -6
0 5 10 15 20 25 30 0 5 10 15 20 25 30
(b) Time (sec) Time (sec)
6 6
4 4
2 2
0 0
u1 u2
-2 -2
-4 -4
-6 -6
0 5 10 15 20 25 30 0 5 10 15 20 25 30
(c) Time (sec) Time (sec)
Figure 16. Responses for the system introduced in Figure 13 computed by the average acceleration method,
for the new technique: (a) n = 2 (t = 0.04 s); (b) n = 3 (t = 0.06 s); and (c) n = 4 (t = 0.08 s).
Table III. The study of the computational costs associated with the responses in Figure 16.
Integration step size, t (s) Computational cost (s) Cost saved (%)
Integration steps equal to f t (conventional) 6.296 Comparison basis
Integration steps equal to 2 f t 3.194 49.27
Integration steps equal to 3 f t 2.149 65.87
Integration steps equal to 4 f t 1.646 73.85
5. CONCLUSION
Considering structural dynamic systems subjected to digitized excitations, a technique and corre-
spondingly a computational procedure are proposed for time integration with steps larger than the
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
A TECHNIQUE FOR TIME INTEGRATION ANALYSIS 2109
excitation steps. The simplifying assumptions are the uniform distribution of the steps of inte-
grations and excitations, the latter embedded by the former, and the smoothness of the digitized
excitations. The objective is materialized by replacing the original excitations with new excitations,
digitized at larger steps, such that in time integration analyses, with losing some accuracy, consid-
erable computational costs are saved. After theoretically arriving at a technique for materializing
the integration, without destroying the convergence trend or disregarding any part of the excitation
history, the parameters of the technique are set, reasonably, leading to a computational procedure,
later implemented in the numerical study. The main achievements are as summarized below:
1. A technique is proposed that provides the capability of time integration, with steps larger
than the excitation steps, without discarding any part of the excitation record; based on the
new technique, the conventional procedure of time integration analysis is, for the first time,
extended to a procedure for time integration analysis, very less sensitive to the sizes of
digitized excitations steps (see Equations (6) and (39)).
2. As the key stage of the proposed computational procedure, prior to conventional time integra-
tion analysis, the original digitized excitations are being replaced with excitations digitized
at steps larger than (by an integer multiplier greater than one) the excitations’ digitization
steps. This stage is a part of the numerical model (analysis), applicable in time integration of
different types of equations of motion, with different types of conventional time integration
methods.
3. Considering convergence, computational cost and ease, and accuracy, the new technique does
not destroy convergence, considerably reduces the computational cost, is simply applicable,
and suffers from tolerable loss of accuracy.
4. As an intermediate, yet important, achievement, in the presence of approximations in the
right hand sides of semi-discretized equations of motion, for providing convergence of the
responses computed by time integration, with the rate of the order of accuracy, the excitations’
approximations should disappear with a rate not less than, and preferably equal to, q0 −1+n s
(see Equation (23)). This would rather be interpreted as a part of the consistency requirement.
Finally, the mere claim of the author is the introduction of a simple inexpensive tool for time
integration with steps larger than the excitations’ digitization steps, that, as a pioneering attempt,
can be considered successful.
ACKNOWLEDGEMENTS
Professor M. G. Ashtiany is gratefully appreciated for his question in a scientific session causing the need
to the research reported in this paper come into the view of the author. Sincere gratitude of the author is
also given to Dr O. Bahar and Dr M. Hosseini for the talks the author had with them while revising the
paper. Finally, the author is very thankful to the reviewer(s), who have caused many improvements in the
paper.
REFERENCES
1. Henrych J. Finite Models and Methods of Dynamics in Structures. Elsevier: The Netherlands, 1990.
2. Argyris J, Mlejnek JP. Dynamics of Structures. Elsevier: The Netherlands, 1991.
3. Belytschko T, Liu WK, Moran B. Non-linear Finite Elements for Continua and Structures. Wiley-Intersciences:
New York, U.S.A., 2000.
4. Soroushian A. New methods to maintain responses’ convergence and control responses’ errors in the analysis of
nonlinear dynamic models of structural systems. Ph.D. Thesis, University of Tehran, 2003 (in persian).
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
2110 A. SOROUSHIAN
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm
A TECHNIQUE FOR TIME INTEGRATION ANALYSIS 2111
35. Hörnell K, Lötstedt P. Time step selection for shock problems. Communications in Numerical Methods in
Engineering 2001; 17:477–484.
36. Koteras JR, Lehoucq RB. Estimating the critical time-step in explicit dynamics using the Lanczos method.
International Journal for Numerical Methods in Engineering 2007; 69:2780–2788.
37. Clough RW, Penzien J. Dynamics of Structures. McGraw-Hill: Singapore, 1993.
38. McNammara JF. Solution schemes for problems of nonlinear structural dynamics. Journal of Pressure Vessels
1974; 96:147–155.
39. Bathe KJ. Finite Element Procedures. Prentice-Hall: Englewood Cliffs, NJ, U.S.A., 1996.
40. Haitian Y. A new algorithm of time stepping in the non-linear dynamic analysis. Communications in Numerical
Methods in Engineering 2001; 17:597–611.
41. Hughes TJR. The Finite Element Method: Linear Static and Dynamic Finite Element Analysis. Prentice-Hall:
Englewood Cliffs, NJ, U.S.A., 1987.
42. Havskov J, Alguacil G. Instrumentation in Earthquake Seismology (Modern Approaches in Geophysics). Springer:
The Netherlands, 2005.
43. Zhongzhi S, Naeim F. The Seismic Design Handbook. Kluwer Academic Publishers: Hingham, MA, U.S.A.,
2001.
44. Brown JW, Churchill R. Fourier Series and Boundary Value Problems. McGraw-Hill: Singapore, 1993.
45. Apostol TM. Calculus, vol. I. Wiley: New York, U.S.A., 1967.
46. Eykhoff P. System Identification: Parameters and State Estimation. Wiley: The Netherlands, 1974.
47. Penry SN, Wood WL. Comparison of some single-step methods for the numerical solution of the structural
dynamic equation. International Journal of Numerical Methods in Engineering 1985; 21:1941–1955.
48. Zhou X, Tamma KK. Design, analysis, and synthesis of generalized single step single solve and optimal algorithms
for structural dynamics. International Journal for Numerical Methods in Engineering 2004; 59:597–668.
49. Weaver Jr W, Timoshenko SP, Young DH. Vibration Problems in Engineering. Wiley: New York, U.S.A., 1989.
50. Warburton GB. Formulae for errors for initial displacement and velocity problems using the Newmark method.
Earthquake Engineering and Structural Dynamics 1989; 18:565–573.
51. Li XD, Zeng LF, Wiberg N-E. A simple local error estimator and an adaptive time-stepping procedure for direct
integration method in dynamic analysis. Communications in Numerical Methods in Engineering 1993; 9:273–292.
52. Zienkiewicz OC, Boroomand B, Zhu JZ. Recovery procedures in error estimation and adaptivity in linear
problems. Earthquake Engineering and Structural Dynamics 1991; 9:871–887.
53. Soroushian A, Wriggers P, Farjoodi J. Time integration of nonlinear equations of motion—numerical instability
or numerical inconsistency? Fifth EUROMECH Nonlinear Dynamics Conference, Eindhoven, 7–12 August 2005.
54. Chen CC, Robinson AR. Improved time stepping analysis for structural dynamics calculations I. Treatment
of rapidly varying excitation and material nonlinearity. Journal of Engineering Mechanics (ASCE) 1993; 119:
2496–2513.
55. Bernal D. Locating events in step-by-step integration of equations of motion. Journal of Structural Engineering
(ASCE) 1991; 117:530–545.
56. Katona MG, Zienkiewicz OC. A unified set of single step algorithms. Part 3: the beta-m method, a generalization
of the Newmark scheme. International Journal for Numerical Methods in Engineering 1985; 21:1345–1359.
57. Soroushian A, Farjoodi J. More reliable responses for time integration analyses. Structural Engineering and
Mechanics—An International Journal 2003; 16:219–240.
58. Bismarck-Nasr MN, De Oliveira AM. On enhancement of accuracy in direct time integration dynamic response
problems. Earthquake Engineering and Structural Dynamics 1991; 20:699–703.
59. Noble B, Daniel JW. Applied Linear Algebra. Prentice-Hall: Englewood Cliffs, NJ, U.S.A., 1977.
60. Hairer E. Solving Ordinary Differential Equations. II. Stiff and Differential-Algebraic Problems. Springer:
New York, U.S.A., 1996.
61. Babitsky VI, Krupenin VL. Vibration of Strongly Nonlinear Discontinuous Systems. Springer: Germany, 2001.
62. Soroushian A, Farjoodi J, Mehrazin H. A new measure for the nonlinear behavior of piece-wisely linear structural
dynamic models. Proceedings of the 13th International Congress on Sound and Vibration, Vienna, 2006.
63. Mahin SA, Lin J. Construction of inelastic response spectra for single degree-of-freedom systems. Report No.
UCB/EERC-83/17, Earthquake Engineering Research Center, University of California, Berkeley, U.S.A., 1983.
64. Nau JM. Computation of inelastic spectra. Journal of Engineering Mechanics (ASCE) 1983; 109:279–288.
Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2008; 24:2087–2111
DOI: 10.1002/cnm