Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: Nine different schemes of dynamic relaxation method are compared in the paper. Schemes with viscous
Available online 14 July 2015 damping and schemes with kinetic damping are used. Kinetic damping with a peak in the middle of the time
step and kinetic damping with the parabolic approximation of the peak are considered. They are also used
Keywords:
in three different ways of cable approximation. The cable is approximated as a tension bar, a catenary and
Cable structures
Nonlinear analysis
a parabolic cable element. The efficiency and stability of each method are compared to three selected 3D
Dynamic relaxation examples of cable structures and one existing structure. The effect of mass distribution along the structure is
Kinetic damping also of interest and is studied in the paper.
Viscous damping © 2015 Civil-Comp Ltd. and Elsevier Ltd. All rights reserved.
Iteration
1. Introduction of fictitious parameters – i.e. mass and damping. If the masses are
too small (in relation to the stiffness of the structure), then the in-
The dynamic relaxation method (DRM) is an iterative process that stability of the iteration may occur and the analysis will not converge
is used to find static equilibrium. DRM is not used for the dynamic to the equilibrium state. On the contrary too large fictitious masses
analysis of structures; a dynamic solution is used for a fictitious lead to the slow and time consuming calculations. The iterative DRM
damped structure to achieve a static solution. The method relies on a algorithm converges very fast when using the values of damping co-
discretized continuum in which the mass of the structure is assumed efficients close to the critical values. Too large values of damping also
to be concentrated at given points (nodes) of the structure. Resid- lead to the slow and time consuming calculations. For this reason the
ual forces (the difference between internal and external forces) are technique of kinetic damping is employed (with the viscous damping
also calculated at these nodes. Nodal displacements are calculated on coefficient taken as zero).
the basis of Newton’s second law of motion, in which residual forces The paper compares the effectiveness of different solution strate-
and fictitious variables are used. The diagonal mass matrix and the gies for the static analysis of cable structures based on the DRM. The
diagonal damping matrix attenuation are considered and, therefore, factors having an impact on the stability of the method and the speed
the nodal displacement equation may be written for each node sep- of the convergence, such as distribution of the fictitious mass along
arately. And this, in particular, is the main advantage of this method: the structure, using the fictitious damping factor and the choice of
DRM does not require the assembly and storage of the global stiffness the time step are studied in the paper. Three different ways of cable
matrix of the structure. DRM is quite suitable for solving large-scale approximation are used – the cable is approximated as a tension bar,
nonlinear problems such as cable structures. a catenary and a parabolic cable element. The efficiency and stabil-
The DRM theory was first described by Day [1]. This theory was ity of each solution DRM strategy are compared to three selected 3D
further developed by adding a rule for determining the mass to each examples of cable structures and one existing structure.
node [2]. Using kinetic damping is another effective method that was This study is a follow-up to study [5]. It has been extended to in-
described by Topping [3] and Lewis [4]. Here is also mentioned that clude the design of the roofing of the Barrandov tram stop in Prague.
the dynamic relaxation method is more stable and more efficient than
other stiffness matrix approaches for structures with large degrees of
freedom. 2. Approximation of cable
The stability of the DRM, the speed of convergence and the CPU
time of the solution can be radically affected by the suitable choice A cable can be approximated as a tension bar, a catenary (sev-
eral tension bars) and a perfectly flexible element (bending moments
along all of its length are equal to zero). Homogeneous material
∗
Corresponding author. Tel.: +420 24354500. with a constant cross-section throughout its length is assumed in all
E-mail address: maca@fsv.cvut.cz (J. Máca). cases.
http://dx.doi.org/10.1016/j.advengsoft.2015.06.009
0965-9978/© 2015 Civil-Comp Ltd. and Elsevier Ltd. All rights reserved.
M. Hüttner et al. / Advances in Engineering Software 89 (2015) 28–35 29
The force T can be calculated iteratively from Eq. (2), as given in [7]:
√
l2T 2c rQ b
g(T, r, l, c, s0 , Q ) =ln − + +
2rQ l lT lT
√
2c rQ a
− ln − − +
l lT lT
c √ 1 √
+ ( a − b) + ( a + b) − s0
4rQ 8T
T l2 c2 Qr2
− + + =0 (2)
Fig. 1. A bar element. EA r r 12T 2
where:
2.1. Tension bar
• l is the horizontal distance between the two end joints,
• c is the vertical separation between the joint j and the joint i (it
The bar connects the endpoints and carries only a positive normal
can be negative),
force. The internal force T in one bar element may be calculated using
• r is the distance between two end joints of the cable element,
the known Eq. (1) (Fig. 1):
• s0 is the slack length of the cable element,
EA • Q is the resultant of the vertical uniform load q acting vertically
T= (r − s0 ) (1)
s0 along the entire length of a parabolic curved cable, while Q = qs0 .
where: For reasons of clarity, the calculation introduces two more
substitutions:
• E is Young’s modulus of elasticity,
• A is the cross-sectional area, a = Q 2 r2 + 4c2 T 2 + 4l 2 T 2 + 4crQT ;
• r is the distance between two end joints in the chord direction b = Q 2 r2 + 4c2 T 2 + 4l 2 T 2 − 4crQT.
(current length),
• s0 is the non-elongated length of the element (slack length). 3. Dynamic relaxation
The force T acts as a normal force on the bar. If the force T
is negative, then it is set to zero. The deadweight of the strut The theory of this method was first described by Day [1]. This the-
has been assumed to be concentrated equally at its two end ory was further developed and its detailed overview may be found in
joints. Barnes [2], Topping [3] or Lewis [4].
The basic assumption of this theory is that the behaviour of a The basic Eq. (3) of motion for the joint i, the direction j (j corre-
cable may be approximated by a few bars. These bars are intercon- sponds to x, y and z directions) and the time t is:
nected by joints and sustain only positive normal forces. The be- Rti j = Mi j · v̇ti j + Ci j · vti j (3)
haviour of individual bars is described in Section 2.1. As found in
where:
[6], five bars well enough correctly describe the characteristics of
the cable. This approximation enables to describe more precisely • Rti j is the residual force at the nodal point i, in the direction j and
the behaviour of individual members (e.g. vertical deflections of the at the time t,
cable). • Mi j is the fictitious mass at the nodal point i and in the direction j,
• Ci j is the fictitious damping factor for the nodal point i and in the
2.3. Cable element direction j,
• vti j is the velocity at the nodal point i in the direction j and at the
The basic assumption of the analysis of a flexible elastic cable is time t,
that the cable is regarded to be perfectly flexible and is devoid of • v̇ti j is the acceleration at the nodal point i in the direction j and at
any flexural rigidity. The load on the cable, which must include at
the time t.
least self-weight, is distributed uniformly along the curve of the ca-
ble, which is assumed to be a parabola. The detailed analysis can be The basic unknowns are nodal velocities, which are calculated
found in [7–9]. from nodal displacements. The discretisation from the timeline with
For the purpose of the study, it is necessary to use an internal force the time step t will be performed. During the step t, a linear
T, which is always positive and whose significance is shown in Fig. 2. change of velocity is assumed. Acceleration during the step t is
thus considered to be constant. By substituting the above assump- (d) Determine velocities for each node.
tions and adjusting Eq. (3), velocity can be expressed at a new time (e) Determine current kinetic energy.
point t + t/2: (f) If max|R| < Rlim and Uk < Uk,lim , then stop the algorithm.
(g) Update the coordinates.
Mi j /t − Ci j /2 Rti j
v(i t+ t/2)
= v(i t− t/2)
+ (4) (h) Continue the DRM iteration from step c).
j j Mi j /t + Ci j /2 Mi j /t + Ci j /2
The current coordinates of the nodal point i at the time instant 3.3. Stability and fictitious masses
t + t may then be expressed as follows:
The factors having an impact on the stability of the method and
x(i t+t ) = xti + t · v(ixt+t/2) (5) the speed of the convergence are:
Similarly, equations may be written for the y and z coordinate di- • Distribution of the fictitious mass of a structure at joints,
rections. From the imbalance (between external and internal forces) • using the fictitious damping factor,
in the node i, we may calculate the residual forces Rti j (j = x, y, or z) • choice of the time step t.
for the corresponding node at the time t.
As is written in [3]: “Usually, the time step is a fixed value and the
xti − xtj
Rtix = Pix − Tkt other two factors are varied until stability of the iteration is achieved.
rkt If the time step t is too large or the masses are too small, then the in-
k
stability of the iteration may occur and the analysis will not converge
yti − ytj to the equilibrium state. Generally, convergence may be achieved by
Rtiy = Piy − Tkt reducing the time interval or increasing the fictitious masses. It has
rkt
k been shown by Barnes [2] that for any t convergence may usu-
ally be ensured by using fictitious masses defined by the following
Qk zti − ztj equation:”
Rtiz = Piz + − Tkt (6)
2 rkt t 2
k k
Mi j = Si j (9)
where: 2
where Sij is the largest direct stiffness of the ith joint in the j direction.
• k is the index of the element entering the nodal point i.
The stiffness Sk of each member k (entering the nodal point i) is
• j is the second endpoint on the element k.
represented by two components – namely, the geometric stiffness
• Pix is the external load at the nodal point i in the direction x,
SkG and the elastic stiffness SkE .
• Piy is the external load at the nodal point i in the direction y,
• Piz is the external load at the nodal point i in the direction z, Ek A k T
Sk = SkE + SkG = + k (10)
• Qk is the resultant of the vertical uniform load for each cable ele- s0k rk
ment k,
Hence, the following Eq. (11) applies to the stiffness of the node Sij :
• xti , yti , zti are the current coordinates of the nodal point i,
• rkt is the distance between two end joints for each cable Si j = Sk, j (11)
element k. k
The internal force Tkt for each cable element k may be calculated where Sk, j is the stiffness Sk – from Eq. (10) – distributed in the direc-
from Eq. (2), and for each bar element from Eq. (1). tion j.
In order to start calculating, the velocity at the time point t/2
must be calculated. Using the initial conditions for the time t = 0 3.4. Damping factors
where v0i j = 0, we obtain:
As shown in [4], the iterative algorithm converges fastest when
t
v(i jt/2) = R0i j (7) using the critically damped mode. In an undamped mode, the struc-
2Mi j ture will oscillate around its position of equilibrium, and the viscous
where R0i j
are residual forces at the time t = 0. damping coefficient, known as critical damping, may be found from
At each time instant, it is also possible to calculate the kinetic en- Eq. (12):
ergy Uk throughout the structure: 4π Mi j
C = 2 Si j · Mi j = , (12)
n
m 2 Nt
(t+t/2) (t+t/2)
Uk = Mi j vi j (8) where N denotes the number of iterations required to complete one
i j cycle of oscillation. It may now be seen that in order to obtain the
where n is the number of joints and m is the number of dimensions value of the viscous damping coefficient, an additional computer run
(3D or 2D). is necessary, with C set to zero.
3.2. Iteration scheme 3.5. Kinetic damping with a peak in the middle of the time step
The acceptable error of residual forces Rlim and the error of kinetic The description of the method is based on publications [3] and
energy Uk,lim must be defined before starting the calculation. The it- [4]. “When the technique of kinetic damping is employed, the viscous
eration process can be outlined as below: damping coefficient is taken as zero. The system is brought to rest by
following a process stopping the iterations, whenever a peak in the
(a) Set initial conditions. kinetic energy of the entire system is detected, and then restarting
(b) Set fictitious masses, the fictitious damping factor and the size the computation from the current configuration, but with zero initial
of a time step. velocity.” In other words, if it is true that:
(c) Determine residual forces from stresses and loads for each
element. Uk(t−t/2) > Uk(t+t/2) (13)
M. Hüttner et al. / Advances in Engineering Software 89 (2015) 28–35 31
the coordinates are set to x(t− t/2) when the peak is assumed to have 3.7.4. Scheme C
occurred. Then This scheme is based on the theory of kinetic damping with a peak
t in the middle of the time step (Section 3.5). The mass for the whole
x(i t−t/2) = xti − · v(ixt−t/2) (14) structure is calculated from Eq. (18).
2
when the analysis is restarted the velocities must be calculated at the
3.7.5. Scheme D
mid-point of the first time steps as follows:
This scheme is based on the theory of kinetic damping with a peak
t t in the middle of the time step (Section 3.5). The discretized mass Mi
v(ixt+t/2) = R (15)
2Mix ix for each joint i is calculated separately from Eq. (19).
where the residual forces Rt are recalculated from the x(t− t/2) dis-
3.7.6. Scheme E
placement position.
This scheme is based on the theory of kinetic damping with a peak
in the middle of the time step (Section 3.5). The discretized mass Mi
3.6. Kinetic damping with parabolic approximation
for each joint i is calculated separately from Eq. (19), but the masses
Mi are recalculated after each restart of kinetic energy.
The idea is taken from Lewis [4]. She assumes that the trace of ki-
netic energy near the peak can be approximated by a parabolic curve.
3.7.7. Scheme F
If Eq. (13) is true, the coordinates are set to x(t− β t) when the peak is
This scheme is based on the theory of kinetic damping with a
assumed to have occurred. Then
parabolic approximation (Section 3.6). The mass for the whole struc-
(t−β t )
xi = xti − β t · v(ixt−t/2) , (16) ture is calculated from Eq. (18).
where
3.7.8. Scheme G
KE3 − KE2
β= , (17) This scheme is based on the theory of kinetic damping with a
KE3 − 2KE2 + KE1 parabolic approximation (Section 3.6). The discretized mass Mi for
(t+t/2) each joint i is calculated separately from Eq. (19).
and where KE3 = Uk is the kinetic energy at the time point
(t−t/2) (t−3t/2)
t + t/2, KE2 = Uk , KE1 = Uk . When the analysis is
3.7.9. Scheme H
restarted the velocities must be calculated at the mid-point of the
This scheme is based on the theory of kinetic damping with a
first time steps from Eq. (15), where the residual forces Rt are recal-
parabolic approximation (Section 3.6). The discretized mass Mi for
culated from the x(t− β t) displacement position.
each joint i is calculated separately from Eq. (19), but the masses Mi
are recalculated after each restart of kinetic energy.
3.7. DRM schemes
4. Illustrative examples
Nine different DRM schemes will be used in the paper.
Schemes A, A∗ and B are based on the theory of viscous damp-
The above chosen DRM schemes were applied to three different
ing (Section 3.4). Schemes C–E are based on the theory of ki-
constructions. The initial prestress in the link is included as the ini-
netic damping with a peak in the middle of the time step
tial slack length. Unless otherwise indicated, the self-weight consid-
(Section 3.5), and schemes F–H are based on the theory of kinetic
ered is q = 0.001 kN/m. The time step t = 1s in all examples. In
damping with parabolic approximation (Section 3.6).
all cases, the computations were stopped when the vector of residual
forces was reduced to 0.1% (or 1% respectively by cable elements) of
3.7.1. Scheme A
its original value.
The discretized mass M chosen in this scheme is the same for each
All DRM schemes were programmed MATLAB R2012b and calcu-
node and all directions. The mass is calculated based on the max-
lations were carried out on the computer ASUS with processor AMD
imum direct stiffness of the node i and in the j direction. We may
E-450 1.65 GHz, 4GB RAM.
write
The results of the calculations (the number of iterations and the
t 2
M= (max Si j ) (18) CPU time) are presented in Tables 1–6. The maximum number of iter-
2 ations for one calculation was 50,000 iterations. When the total num-
The viscous damping coefficient C for the whole structure is cal- ber of iterations was higher than this limit the result is called “error”
culated from Eq. (12). and marked as “-” in Tables 1–6.
Table 1
The number of iterations – cable net (2 × 2).
Bar 76 71 427 47 47 38 39 39 47
5 bars 454 445 26,108 344 356 899 346 356 968
Cable 77 70 384 43 43 38 38 38 44
Table 2
Time of solution (CPU in seconds) – cable net (2 × 2).
Scheme A A∗ B C D E F G H
Bar 0.03 0.02 0.13 0.02 0.02 0.02 0.02 0.02 0.02
5 bars 0.56 0.56 31.85 0.43 0.46 1.39 0.44 0.55 1.47
Cable 15.29 14.26 69.19 10.3 9.69 8.57 8.92 9.86 10.83
Table 3
The number of iterations – hypar net.
Bar 177 169 8889 214 197 376 215 209 252
5 bars 1042 1014 – 1176 – 3078 1161 – 2894
Cable 135 129 5920 149 139 237 145 165 234
Table 4
Time of solution (CPU in seconds) – hypar net.
Scheme A A∗ B C D E F G H
Bar 0.12 0.12 6.9 0.15 0.14 0.32 0.15 0.15 0.32
5 bars 3.40 3.29 – 4.1 – 12.10 3.73 – 11.39
Cable 52.08 48.43 3247 55.80 52.83 88.38 55.09 66.86 86.92
Table 5
The number of iterations – spatial net.
A A∗ B C D E F G H
Bar 119 112 3557 119 135 257 123 122 425
5 bars 661 660 – 769 602 1098 727 – 1236
Cable – – 2381 199 226 201 263 371 182
Table 6
Time of solution (CPU in seconds) – spatial net.
Scheme A A∗ B C D E F G H
Bar 0.10 0.10 2.84 0.10 0.12 0.26 0.11 0.11 0.26
5 bars 2.70 2.64 – 2.94 2.28 5.10 2.75 – 5.85
Cable – – 1797 89.33 100.4 89.80 114.3 173.4 82.10
Table 7
The number of iterations – tram stop.
A A∗ B C D E F G H
Bar 488 364 1353 136 115 309 125 130 332
5 bars 11,461 10,181 – 2767 – 5320 2880 – 4446
Cable 448 331 1290 120 121 301 118 135 313
Table 8
Time of solution (CPU in seconds) – tram stop.
Scheme A A∗ B C D E F G H
Bar 0.36 0.28 1.11 0.12 0.10 0.45 0.10 0.12 0.33
5 bars 38.55 27.60 – 9.72 – 23.56 10.02 – 19.81
Cable 175.4 136.1 475.1 59.60 60.39 130.2 58.22 64.17 132.3
As it is a structure with a central symmetry, only one fourth of the The following cable lengths (selected cables) were considered:
structure was subjected to modelling. The canvas was represented
by forces exerted by the structure’s own weight and prestress in the
model. The background for the creation of the model was the geodetic Cable 1 – 11.221 m, Cable 2 – 13.618 m, Cable 3 – 14.340 m,
Cable 4 – 13.020 m, Cable 5 – 14.530 m, Cable 6 – 13.820 m.
survey of the existing structure made in November 2013.
The anchoring of the cables into anchor blocks was modelled as
a fixed support. The point where the cables were anchored into an- The load is applied at joints 18 and 19 where the forces exerted by
chor blocks was modelled as a sliding joint allowing the joint’s mo- the canvas act. These forces amount to: P18z = 12 kN and P19z = 16 kN.
tion only in the direction of the x and y axes. The structural scheme The computational accuracy considered with regard to the accu-
of the structure considered is evident from Fig. 7. Thus, the structure racy of the survey of the structure was Rlim = 0.1 kN.
has the total of 18 degrees of freedom. The number of iterations and the solution time for individual DRM
The parameters of cables are as follows: schemes are presented in Tables 7 and 8.
The forces acting in the real structure were measured and de-
Cables 1–7: E = 140 GPa, ∅23.6 mm, A = 230 mm2 , q = 0.021 kN/m
scribed in detail in [14]. To verify the computational model against
Cables 8–30: E = 140 GPa, ∅21.2 mm, A = 180 mm2 , q = 0.017 kN/m
reality, Scheme C (the most useful scheme) with the cable elements
M. Hüttner et al. / Advances in Engineering Software 89 (2015) 28–35 35
Table 9 In this context, it may be noticed that the methods based on ki-
A list of measured and computed forces in cables.
netic damping appear more stable and faster, which confirms the
Force T1 T2 T3 T4 T5 T6 conclusions presented by Rezaiee-Pajand [15].
The results confirm that it is beneficial to divide the same amount
Measurement 24.4 11.1 11.8 15.1 10.0 5.6
Scheme C – cable element 20.0 9.9 11.9 15.4 9.7 5.7
of mass into all nodes of the structure proportionally to the stiffest
node of the analyzed structure (Schemes C and F). This way of mass
distribution seems to be very universal in practical calculations and
Table 10
also faster in comparison with the schemes where the discretized
Summary of results.
mass is calculated separately for each joint (Schemes D, E, G and H).
Scheme Damping Total number Total number Total CPU The schemes with the recalculation of the masses after each restart of
of errors of iterations time (s)
kinetic energy (Schemes E and H) are very stable especially for struc-
C Kinetic – t/2 0 6083 232.6 tures with large number of degrees of freedom. On the other hand the
F Kinetic – β t 0 6180 253.9 total CTU time is lager in comparison with the Schemes C and F.
H Kinetic – β t 0 11,373 353.7 The results based on Schemes C and F, D and G and E and H are
E Kinetic – t/2 0 12,152 351.6
practically the same, so it is possible to conclude that the approxima-
A∗ Viscous 1 13,546 233.4
A Viscous 1 15,138 288.6 tion method of the kinetic energy peak has a negligible effect on the
D Kinetic – t/2 2 1981 226.4 total CTU time.
G Kinetic – β t 3 1565 315.2
B Viscous 3 50,309 5629.5 Acknowledgement