You are on page 1of 10

Engineering Structures 31 (2009) 1677–1686

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

SDOF response model parameters from dynamic blast loading experiments


K. Fischer, I. Häring ∗
Fraunhofer Ernst-Mach-Institut, Efringen-Kirchen, Germany

article info abstract


Article history: In this paper, two minimization problems are presented that determine parameters for best practice single
Received 12 September 2008 degree of freedom (SDOF) models, based on the blast response of structural members. The inputs for the
Received in revised form model are dynamic experimental data as obtained from pressure and laser deflection measurements. The
10 December 2008
first approach compares model and experimental deflection and minimizes the differences between the
Accepted 11 February 2009
Available online 28 March 2009
two. Within this approach, a method is introduced to determine the best parameters of typical structural
resistance functions for SDOF models. Examples for the determination of up to three parameters are
Keywords:
given. The second approach uses a set of basic functions in terms of deflection and velocity to describe
Blast loading the resistance function as a linear combination of functions. Using experimental data, a matrix equation
SDOF model is obtained from the single degree of freedom force equation. Taking advantage of the minimization
Resistance function conditions, a second matrix equation is derived, then numerically inverted and solved for the response
Parameter determination basis function expansion parameters. In a systematic way the response function may be improved, e.g.
Dynamic experimental data by adding more terms to polynomial basis functions. Both procedures determine SDOF and in particular
Minimization problem response function parameters from dynamic data and also indicate whether they are well selected.
Pressure-impulse diagram Combining parameter data of multiple events and failure criteria, the overall model and failure behavior
Iso-damage curve
are determined in a P–I diagram for the example of a standardized laminated glass pane.
© 2009 Elsevier Ltd. All rights reserved.

1. Introduction effective mass, load-mass factors or dynamic design factors, and


the failure model. If two-dimensional visualization is included
The behavior of structures due to blast loading is conveniently the pressure-impulse shape is considered to be a part of the
represented with single degree of freedom (SDOF) models. These model [13]. The loading consists of the pressure–time history and
simplifying models try to evaluate a structure using only one phys- an effective area expression. The resistance may depend on the
ical degree of freedom, the horizontal deflection [1]. Advantages deflection and velocity locally in time, or in a functional way on
that still indicate the use of these state-of-practice [2] or common the complete deflection- and velocity-time history.
practice [3] models include the quick evaluation for assessment of Besides the mass and loading data and their parameterization
structures (within larger scenarios) [4], intuitive transition from the SDOF model depends on the functional form and the
static to dynamic models, use of static experimental data and avail- parameters of the resistance function. Often the best parameters
able data collected for these models. are determined from static experiments and used together with
SDOF models are available for all kinds of structural members. several kinds of factors to determine a dynamic SDOF model.
Examples are glass [5], masonry [6], walls and slabs [3], and However, also dynamic data are available. It is often only used
concrete [7]. Models are available for whole buildings [8]. Also, to determine failure criteria and factors that modify the static
the consequences for personnel due to blast loading may be response behavior. So it is desirable to develop procedures that
modeled [9]. determine ideal dynamic parameters of dynamic SDOF models and
The SDOF model which can be interpreted as an idealized structural resistance functions.
assumption, can be classified in three parts: material assumptions, The main part of the paper introduces two approaches to
loading assumptions and geometry assumptions [10]. Within these determine the response function parameters for the dynamic
three groups a complete SDOF model may be considered to response from dynamic experimental data. Both approaches use
consist of the initial condition, the pressure force term or loading the experimental deflection data. In the first case the model and
function [11], the resistance force term or resistance function [12], experimental deflections are compared. In the second case it is
tested whether experimental deflection, velocity and acceleration
data fulfill the model equations. It is shown how such models can
∗ Corresponding author. Tel.: +49 7628 9050 638; fax: +49 7628 9050 1638. be used for the fast determination of response behavior for the
E-mail address: haering@emi.fhg.de (I. Häring). structural members.
0141-0296/$ – see front matter © 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engstruct.2009.02.040
1678 K. Fischer, I. Häring / Engineering Structures 31 (2009) 1677–1686

This paper is organized as follows: In Section 2 we introduce In the elastic case, we let the plastic response Rmax and
the models for which parameters will be determined from the maximum elastic elongation xe go to infinity, such that
dynamic experimental data; Section 3 defines and solves the Hooke’s parameter k remains constant. In many cases a smooth
numerical minimization approach. It is shown how model behavior elastic–plastic (see straight line in Fig. 1(a)) response is preferred
is visualized in P–I diagrams; Section 4 details the analytical in comparison to the ideal elastic–plastic response and can be
minimization of the basis function approach. realized with the arctangent-function [6]
!
2. Parameter dependence of SDOF models describing blast
  2Ap0 π k̃x
R k̃, p0 ; x = arctan . (4)
loading response of structural members π 2p0

We assume that the movement of a structural member may be Here the parameter p0 is the maximal stationary plastic response
resembled by an equivalent system showing the same deflection pressure and k̃ = p0 /xe determines the initial elastic resistance.
due to the dynamic blast loading as the real system. The general For small deflections, R(k̃, p0 ; x) behaves linear with Hooke’s
model for which parameters are determined reads parameter k̃A. A p0 is the plastic asymptote. Elastic–plastic
response functions are used to describe the response of pre-
v
   
d x
= , fabricated concrete slabs [18] and beams [19]. In [6] dynamic
dt v Ap(t ) − R(c1 , . . . , cNc ; x, v) /(mkLM )

response of masonry walls is described using the arctangent
(1)
x(t = 0)
   
0 response function.
= ,
v(t = 0) 0 An arbitrary piecewise linear response function is given by 2N
parameters
where m and A respectively are the constant mass and geometri-
cally presented area of the structural member, and c1 , . . . , cNc are Rk+1 − Rk
R (x1 , . . . , xN , R1 , . . . , RN ; x) = Rk + (x − xk ) , (5)
Nc parameters that determine the response function R. The deflec- xk+1 − xk
tion x, the velocity v , as well as the pressure–time history p are
if xk ≤ x ≤ xk+1 , for k = 0, . . . , N − 1. In Eq. (5) we set x0 = 0 and
time-dependent. To ensure an equivalent system the dynamic de-
R0 = 0. The resistance (5) becomes point symmetric with respect
sign factor kLM is included.
to the origin if we set R(· · · ; x) = −R(· · · ; |x|) for x < 0.
The equivalent ersatz-system (1) is realized with a lumped mass
To describe the hysteresis behavior of an idealized elas-
and an idealized load resulting in one-dimensional deflection,
tic–plastic model beyond the first positive maximum deflection of
whereas the actual structural member has several degrees of
xmax without the simplification of (3)–(5), we introduce a response
freedom. Therefore design factors kM , kL are introduced to gain an
functional that depends on the deflection history [13,20]
equivalent system with respect to deformation, kinetic energy and
done work of external forces [14], in such a way that R (k, Rmax , xmax ; x)
mE = kM m, FE = kL Ap(t ), RE = kR R(c1 , . . . , cNc ; x, v), (2) R

kx = max x,
 0 ≤ x ≤ xe and (i)
where mE , FE and RE are the equivalent mass, load and resistance = xe (6)
of the ersatz system. The resistance force resembles the internal Rmax ,
 xe < x ≤ xmax and (ii)
Rmax + k (x − xmax ) , x < xmax and (iii).
forces of the real structure that try to keep the structure in the
unloaded start position. It can be shown that approximately kR = A resistance according to (6) may only be computed knowing that
kL [9]. Then by defining kLM = kM /kL = kRM = kM /kR we one: (i) either is in the elastic interval; (ii) has increasing deflection
achieve (1). Further we assume here for simplicity a constant load- values and is in the plastic plateau interval; or (iii) has decreasing
mass factor for elastic, elastic–plastic and plastic response [3]. An deflection after having reached a first positive maximum deflection
elaborate derivation of the dynamic design factors can be found in the plastic regime. Thus, knowing only the local deflection x in
in [12,14]. According to [14] kLM ∈]0, 1]. Here we use kLM = 0.78 general is not sufficient to compute the resistance and we may
as appropriate for an area loaded and four-sided fixed glass pane. speak of a resistance functional. Knowing the deflection history
x(t 0 ), 0 ≤ t 0 ≤ t, the information necessary for (6) may
2.1. Response functions without damping be extracted. For further piecewise linear response functionals
see [21].
It is common and convenient to disregard the damping effect
for studying the structural response under a pulse load with short 2.2. Basic function expansion of the response function
duration [11]; except of the series expansions in Section 2.2,
the response models discussed in this paper are only deflection Here we present a series expansion approach to the response
dependent. In most cases it is sufficient to restrict the response function including damping effects in contrast to the models of
function domain up to the first deflection maximum in the Section 2.1. The local approximation of the response function
direction of the loading, because the studied structures are assumes that (generalized) Taylor series expansion about zero,
designed to resist one explosion in their life time [15], and in terms of deflection and velocity exist [22]. The number and
in most cases also fail before the first maximum. Exceptions kind of coefficients allowed in the series expansion determines
are resonant like failure models resulting from complex loading the approach. For a deflection dependent resistance function, a
with a strong under pressure phase and resistance functions polynomial series expansion of the resistance function with Nx
that are asymmetric with respect to deflection [16], especially coefficients ck , k = 1, . . . , Nx , reads
retrofitted structures [17]. All found models assume R(x0 , v0 ) = 0
Nx
corresponding to the assumption of a stable minimum for weak X
Rx (x) = ck xk . (7)
loading at zero.
k=1
For modeling ideal elastic and ideal elastic–plastic resistance
we use (see also Fig. 1(a)) [8] In a similar way we define a velocity dependent damping function
Rv (v) with Nv coefficients dk . Section 4 of this paper will focus on
Rmax
(
kx = x, 0 ≤ x ≤ xe the computation of the coefficients of
R(k, Rmax ; x) = xe (3)
R c1 , . . . , cNx , d1 , . . . dNv ; x, v = Rx (x) + Rv (v).

Rmax , x > xe . (8)
K. Fischer, I. Häring / Engineering Structures 31 (2009) 1677–1686 1679

Fig. 1. (a) Linear-/ elastic–plastic- and arctangent response model. (b) Piecewise linear responsense model. (c) Response functional.

Eq. (8) is a special case of a mixed linear basis set the generation of P–I diagrams and iso-failure curves such locally
defined loadings, as well as failure criteria may be used. In practice
  XNx X
Nv
often the distance l to a free field high explosive source and its
; x, v = aij xi v j .

R aij 1≤i≤Nx (9) net explosive quantity (NEQ) are easier to use than P and I. Such
1≤j≤Nv
i=1 j=1
mass–distance-diagrams may be computed on equal footing as P–I
Finally we note that the parameter determination, illustrated in diagrams. Again the locally unique determination of p(t ) in terms
Section 4 for the resistance function (8), is applicable to any linear of l and m is critical.
basis function expression of the form
3. Chi-square minimization for the determination of model

X parameters
R λ1 , . . . , λNλ ; x, v = λi fi (x, v).

(10)
i=1 Sections 3 and 4 focus on the computation of parameters
In (10) we require that R(λ1 , . . . , λNλ ; 0, 0) = 0, see e.g. (7)–(9). of the resistance functions (3)–(5) and (8) with the help of
dynamic experimental deflection and pressure data. In Section 3
a minimization problem is iteratively solved. At the beginning of
2.3. Failure model and P–I diagrams
Section 3, using the example of a blast loaded standard glass pane,
we will shortly describe the measurement process.
The response behavior of a SDOF model can be characterized
The numerical minimization could also be employed to
using the first positive maximum deflection xmax in the direction
determine parameters other than response function parameters.
of the loading. This maximum can be compared with a critical
Examples are load-mass factors for different phases of the loading
deflection xcrit that defines failure or damage of the structure [11]:
and parameters that model changing mass or area. At least with
xmax > xcrit ↔ structure fails. (11) the more flexible tri-linear response according to (5) in Section 3.3
we implicitly take account of non-constant response-mass factors.
Analogously failure criteria for velocity, acceleration, kinetic In Section 4, a single matrix inversion determines the parame-
energy in the direction of loading, and critical energy in a critical ters of the SDOF model. In this case explicitly A/(mkLM ) and, im-
time [23] can be defined. One may consider a number of critical plicitly together with the series expansion parameters, 1/(mkRM )
values to measure different levels of damage. In summary, given a are determined. The first parameter combination is constant, the
loading, a SDOF model (1) and a failure criterion (e.g. (11)) it may second non-constant.
be decided with a numerical solution of (1), whether a structure
fails. 3.1. Experimental data
The dependence of the model and failure behavior on char-
acteristic blast sizes is often visualized in two-dimensional dia- As indicated in Fig. 2(a), we assume that the pressure time
grams R t[24,25]. If peak overpressure p0 and specific blast impulse history measured at the side of the structure at the rigid frame
I = 0 d p(t )dt, where td is the duration of the first positive over- (see filled circle of Fig. 2(a)) is identical to the pressure time
pressure phase, are used one speaks of P–I diagrams. history as expected in the centre of the structure (see open circle in
Since the model behavior is determined by the pressure loading, Fig. 2(a)). This is especially valid for plane shock tube loading, (see
one has to require that for all combinations of p0 and I, the Fig. 3) [26]; it is less valid for free field experiments. The deflection
deflection-time history is uniquely determined. A generalized form data is obtained directly at the centre of the structure (see the
of a blast loading function reads [11] triangle in Fig. 2(a) and 3(b)).
  In Fig. 2(b) details of the investigated laminated glass are given.
t −γ t
The experiments are performed following the European Standards
p(t ) = p0 1−λ e td
. (12)
td DIN EN 13123-1, Annex A [27] and DIN EN 13541 [28]. The dashed
rectangle shows the presented area that is loaded. The presented
When we choose in (12) for example λ = 1, γ = 0 and neglect the
mass m = 16.5 kg and area A = 0.8 m2 are used throughout the
negative overpressure phase we get a triangular shape
paper. The area between the dashed and the solid line disappears
behind the frame of the laminated glass pane. So the presented area
  
t
p0 1− , 0 ≤ t < td

pp0 ,td (t ) = (13) and mass for the model is smaller than the whole glass pane.
td
Using pressure gauges with measurement error σp and
0, td < t ,

laser deflection measurements with measurement error σx , the
where td = 2I /p0 and hence the loading is determined by the experimental overpressure time history and deflection history
characteristic blast sizes. data are obtained:
If the maximum deflection is shown in a P–I diagram the iso-
(ti , pexp
i ) i=0,...,N −1 , σp ; (ti , xexp
i ) i=0,...,N −1 , σx .
 
(14)
damage curve is marked by xmax = xcrit . In a similar way, other
characteristic model sizes may be visualized. A straightforward Typical values of the experimentally obtained errors are σp =
generalization is to use multiple global failure criteria and multiple 0.8 kPa and σx = 0.4 mm. For simplicity, we assume in (14) that
global pressure loadings. A worst-case approach then determines the time grids of both data types are identical and equidistant.
the iso-damage curve. If a failure criterion is not valid for all P–I Typical numbers for experimental data sets are N = 6 × 104 and
points we may call it local in contrast to global, as used above. For 1t = 10−3 ms.
1680 K. Fischer, I. Häring / Engineering Structures 31 (2009) 1677–1686

Fig. 2. (a) Schematic of the experimental set up. (b) Dimensions of the framed 7.5 mm laminated standard glass pane.

Fig. 3. (a) Shock tube for a standard laminated glass pane. (b) Glass after a shock tube experiment.

Especially for the basis function approach, it is of importance to


use filtered data. In most cases weighted moving average filters are
a good choice, e. g. based on the normal distribution density fµ,σ .
For instance, the filtered deflection data (14) reads

N −1
1ti −nt σt ≤tj ≤ti +nt σt fti ,σx (tj )x(tj )
P
j =0
x̃(ti ) = , i = 0, . . . , N − 1, (15)
N −1
1ti −nt σt ≤tj ≤ti +nt σt fti ,σx (tj )
P
j =0

where the index function 1 is used and n+ is a scalar that


determines the smoothing window. Because of the large data size
it is convenient to use less data points, e.g. taking only every n-th
data point.
In Fig. 4, we see three examples for pressure–time histories
after a dynamic experiment that is used for SDOF model parameter
determination. Fig. 4(a) shows an experiment with weak loading
where no damage is observed. Such experiments are often used to
determine the natural frequency of the structure. Fig. 4(b) shows
a free field experiment. We obtain a high peak overpressure and
short positive blast duration. In Fig. 4(c) we see a shock tube
experiment with a high peak overpressure and long blast duration.

3.2. Chi-square sum and numerical minimization for parameter


determination

Our aim is to determine the best parameters for a given Fig. 4. Pressure–time histories obtained from shock tube and free field
experiments.
resistance function R(c1 , . . . , cNc ; x). For each set of parameters
c1 , . . . , cNc we compute with standard numerical routines (see
If the chi-square expression is small we have good agreement
e.g. [29]) the model deflection xmod (c1 , . . . , cNc ; t ) using (1) and the between experimental and model deflections. More precisely the
experimental pressure history from which we compute pexp (t ) by sum has a global minimum for the best choice of parameters.
interpolation. The next step is to compare the model deflections Thus the task of finding the best parameters of a generalized SDOF
with the experimental deflections. For each Nc -tuple (c1 , . . . , cNc ) model is reduced to finding a global minimum of the chi-square
we compute the chi-square sum: expression (16). It is important to note that (16) presents only
N −1  exp 2 one possible way to measure the agreement of the experiment
X xi − xmod (ti ; c0 , . . . , cNc ) and model. In a similar way one may use any p-norm, e.g. the
χ 2 (c0 , . . . , cNc ) = . (16)
i=0
σx Manhattan norm to be less sensitive to outliers.
K. Fischer, I. Häring / Engineering Structures 31 (2009) 1677–1686 1681

Fig. 5. Comparison measure of model and experiment and minimization for one
parameter.

Fig. 7. (a) Comparison function with no unique minimum in both dimensions.


(b) Comparison function with no unique minimum in one dimension.

Fig. 6. Comparison of model and experiment and minimization for two free
parameters.

This nonlinear, multidimensional minimization problem can


be solved by the Downhill-Simplex-Method due to Nelder and
Mead [30]. The method dispenses with derivatives, and thus,
comparatively requires a high number of function evaluations. For
that reason, it is in general inferior to more complex algorithms
as for instance Powell’s method [31]. Since the computational
problem is moderate, the Downhill-Simplex-Method presents a
good and stable choice.
For a function f : RN → R the algorithm determines the two
points of a given initial N-dimensional simplex with the greatest
(worst) and lowest (best) chi-square value. If the distance of the
two points is less than a defined tolerance the algorithm stops, if
not the worst point is replaced via extrapolation and the routine is
iterated.
It is essential to choose the starting simplex around a well Fig. 8. (a) Comparison of ideal-elastic (red line) and arctangent-elastic–plastic
educated guess to avoid absorption by local minima and to keep (green line) model deflection with experimental data (jagged black line for weak
loading). (b) Linear elastic (red line) and arctangent resistance function (green line)
the complexity low. Helpful are graphical visualizations of the chi-
with best parameters. (For interpretation of the references to colour in this figure
square sum, e.g. parameter diagrams. Fig. 5 shows the comparison legend, the reader is referred to the web version of this article.)
and minimization process for one parameter. In this example
we determine the ideal parameter for linear-elastic structure and the yellow square in the middle marks the global minimum
behavior. In the case of Fig. 5 the experiment was set-up according and hence the best parameters of the resistance function.
to Fig. 4(a) and tN −1 = 65.53 ms. For a given pressure–time and Often we find that the standard models have no good agreement
deflection-time history we search for the Hooke’s constant that with experimental data, see Fig. 7(a) and (b). In Fig. 7(a) we
generates the best experiment-model agreement. In the following have an example where in both parameter dimensions, many
comparison examples, see Figs. 5–7, we compare the models with combinations result in the same function value of the comparison
the experimental deflection of Fig. 8. measure. In this case the weak loading of Fig. 4(a) with tN −1 =
In Fig. 6 we see the chi-square sum for the loading of Fig. 4(a) 65.53 ms was used. In Fig. 7(b) we use the experimental
and tN = 65.53 ms that determines two parameters, Hooke’s pressure–time data of Fig. 4(a). In the experiment no failure is
constant and linear damping factor. In Figs. 5 and 6 the black observed except a damped oscillation. For a resistance function we
squares indicate the iteration-points of the Downhill-Simplex choose arctangent-elastic–plastic structure behavior (4). However,
method. The big squares show the starting points of the algorithm, the plastic plateau is not reached. We observe that the comparison
1682 K. Fischer, I. Häring / Engineering Structures 31 (2009) 1677–1686

Fig. 10. Deflection laser measurement for strong loading.

Fig. 9. (a) Comparison of linear-elastic and arctangent-elastic–plastic model


deflections with experimental data for strong loading. (b) Resistance functions with
best parameters.

function has no unique minimum (in the dimension of the plastic


response pressure parameter). Thus the minimization procedure
correctly reveals that the elastic response suffices to describe the
experiment.

3.3. Ideal elastic and elastic plastic arctangent response, tri linear
response
Fig. 11. (a) Comparison of complex experimental structure behaviour and a SDOF
model with a three parameter resistance function. (b) Tri-linear resistance function.
For further illustration we consider a test series of laminated
glass as shown in Figs. 2 and 3 In Fig. 8, the jagged line shows the
Fig. 11 shows the comparison up to the first positive maximum
deflection of the experiment and the green or red lines show the
in the direction of the blast load between the experiment of
results of the models.
Fig. 10 and a SDOF model with a piecewise linear resistance
In Fig. 8 we computed the best model parameters for an
function, describing the complex structure behavior of the two
experiment with weak loading where no damage is observed
materials. The two different slopes of the resistance function and
(see Fig. 4(a)) over the first positive deflection phase. There is
the breakage point of glass were determined. We see that the
no significant difference between the one and the two parameter
calculated value xb = 27.5 mm of glass breakage is in order of
SDOF model that both show very poor goodness-of-fit. Since the the experimental value 17 mm. Fig. 11(b) implies two ideal elastic
plastic structure behavior is not reached, the second parameter of regimes of the resistance function.
the arctangent function is not used. In summary both often used For the ideal elastic, arctangent elastic–plastic and the simple
parameterizations fail to describe the experimental data. tri-linear response functions, the loading scales with the constant
Fig. 9 shows the same SDOF models as in Fig. 8 with the presented area. The infinitesimal change of velocity is in all cases
best parameters for stronger loading (see Fig. 4(b)) where glass inversely proportional to the constant mass of the pane, see the
breakage is observed. We see that the two-parameter model right hand side of the second component of the first expression of
describes the behavior better than the linear-elastic model over the (1). Assuming that the response itself does not change geometries
whole deflection range. in the vicinity of the experimentally investigated geometry can be
Fig. 10 shows the deflection-time history of the strong loading covered. Only in the case of the arctangent response an overall
of Fig. 4(c). This experiment is an example for complex structure inverse mass area density A/m can be extracted. In practice it
behavior. With the minimization problem we try to fit a function is used to take account of geometry changes of typical window
that describes the behavior of two materials. In this case, glass and sizes including changes of the thickness. A similar factor could be
laminate. In the first milliseconds we have elastic behavior of the extracted in the two other cases. It is obvious that from the very
glass till the glass breaks, which we can see as a discontinuity of the beginning parameterizations of the SDOF model can be chosen
deflection. After the breakage point the behavior of the structure is that allow the determination of less sample specific parameters,
determined by the response behavior of the laminate. Note that we e.g. using a typical length scale, time and force of the experiment.
observed a huge maximum deflection larger than 100 mm. Of course the resulting (dimensionless) parameters can also be
In the two foregoing examples Figs. 8 and 9 we compared determined with the numerical minimization approach.
model and experiment up to the end of the first positive deflection
phase to obtain the best parameters for an elastic and elastic 3.4. Overall model elastic–plastic response
plastic model, respectively. A comparison only up to the first
maximum yields similar poor values for the goodness-of-fit up As described in Section 2.3 we compute the surface of maximum
to the maximum deflection and the following behavior is not deflection values and the iso-damage curve in Fig. 12 using
covered at all. Thus, describing complex structure behavior with (13) and xcrit = xfirst max . Table 1 lists the best parameters
a one or two parameter SDOF models leads to poor goodness-of-fit obtained from six experiments for the arctangent elastic–plastic
with dynamic data. Next we present a three parameter response function (4). The average of the parameters disregards the weak
function up to the first maximum deflection. loading experiment, as motivated in the discussion to Fig. 7(b).
K. Fischer, I. Häring / Engineering Structures 31 (2009) 1677–1686 1683

and backwards difference expressions, respectively, and the inner


matrix lines from central difference expressions [34].
For instance, the resistance function (8) leads in (17) to a linear
expression in terms of the parameters of the two polynomials,
Nx Nv
* A ck dk h * k
h i X X i
D(2) x = pi − xki − D(1) x ,
i m k=1
m k=1
m i

i = 0, . . . , N − 1. (19)

Now we determine the parameters of (8). As input we use


experimental data (14), then apply the filter (15) and restrict
ourselves to a reduced deflection data set. The notation for the
experimental deflection data is not changed after this preparatory
two-step process. For the examples in Section 4.2, Nx = 3 or 4 and
Fig. 12. P–I diagram, maximum deflection surface and iso-damage curve for the
best arctangent elastic -plastic model. Test data for laminated glass. Nd = 1. Thus the deflection-dependent response is not symmetric.
Both resistance expressions do not have functional behavior.
These averages of the parameters are used as the best choice of We employ a kind of chi-square sum that measures the
parameters for a SDOF model of the laminated glass pane. difference between the left hand side and the right hand side
The colors of the points in Fig. 12 show the experimental de- of (19):
flection of Table 1. The characteristic blast sizes were determined
χ 2 a0 , . . . , aNx +Nv

from pressure–time histories. Failed glass panes are indicated with
squares. If glass panes do not fail we use circles. The three top left
!2
N −1 X
X N −1 Nx
X Nx
X
(2)
points belong to free field experiments. Close to the experimental = Dij xj − a0 pi + ak xki + aNx +k ik v , (20)
data and for similar loading Fig. 12 offers a fast assessment of the i=0 j =0 k=1 k=1
standardized glass pane.
* T
where the parameters are summarized by a = a0 , . . . , aNx +Nv

4. Approach with basis function set
T
= m1 A, c1 , . . . , cNx , d1 , . . . , dNv , the velocities vi = Nj=−01 D(1)
ij xj ,
 P
i = 0, . . . , N − 1, are pre-computed and a0 = A/m is included as
This approach uses almost no physical information for an edu- an optional parameter.
cated selection of parameters. The examples given in Section 4.2 We search for the best tuple (a0 , . . . , aNx +Nv ) minimizing the
determine all parameters of the model from dynamic data. chi-square sum. This yields vanishing partial derivatives,

4.1. Determination of parameters by matrix equation ∂ 2


χ 2 = min ⇒ χ = 0, k = 0, . . . , Nx + Nv . (21)
∂ ak
First we present a discrete version of Eq. (1), absorbing kLM
into the loading and the response, based on finite difference Separating the ak , we find a system of linear equations,
expressions [33] employing the grid (14):
* *
A a = b, (22)
(2) * A 1 (1) *
h i  h i
D x = pi − R xi , D x , i = 0, . . . , N − 1. (17)
i m m i where
The finite difference matrices for the first and second derivative,
pi vi v 
N
pi vi1
 −p2 N
i pi x1i ··· pi xi x ···
−3 4 −1 
 −pi x1i x1i +1 ··· 1+N
xi x vi1 x1i ··· viNv x1i 
−1 0 1 0  . .. .. .. .. 
 .
 .N . . . . 
  
 .. .. ..  N −1 
1 . . .
X
Nv Nx 
    
N +1 N +N 1 Nx
(1)
D =
  (1)
 = Dij 0≤i≤N −1 , A=  −pi x
i
x
xi x
··· xi x x vi xi ··· vi xi 
2 1t  .. .. ..

vi1+Nv 

i=0  −p v 1 1 Nx
. . .
0≤j≤N −1 vi xi
1 1
··· v
i xi vi1+1 ···

   i i 
   . .. .. .. .. 
0 −1 0 1  .
. . . . .

1 −4 3
2 −5 4 −1  (18) −pi viNv viNv x1i ··· viNv xNi x viNv +1 ··· viNv +Nv

1 −2 1 0 
 = Aij 0≤i≤Nx +Nv (23)
 .. .. ..  0≤j≤Nx +Nv
1 
 . . .   
(2)
(2)
D =  = Dij 0≤i≤N −1 ,

1t 2 .. .. .. and

. . .
0≤j≤N −1

 
 
0 1 −2 1 *
N −1 X
N −1
Nv T
X
−1 4 −5 2 −D(2) ... N
xj vi1 ... xj v i .
 
b = ij xj p i xj x1i xj xi x
are later also used in the minimization problem for the basis i=0 j=0

function set approach, as well as the vectors of experimental


* * In (23) every matrix element is obtained by summing exper-
deflection x = [x0 , . . . , xN −1 ]T and pressure data p =
imental data. With (22) the parameter determination of the re-
[p0 , . . . , pN −1 ]T . The superscripts indicating experimental data are
sponse function (8) is reduced to the inversion of a system of linear
omitted throughout Section 4, in particular within the deflection
equations. Throughout several applications we found A always to
and velocity dependent response function R(x, v). Both finite
difference matrices of (18) have an error of order 1t 2 ; the first and be invertible for well filtered data and therefore a unique solution
*
last line of the matrix expressions are obtained from the forward a using Gauss–Jordan elimination [29].
1684 K. Fischer, I. Häring / Engineering Structures 31 (2009) 1677–1686

Table 1
Ideal parameter after minimization with respect to the several experiments.

Exp. no. Kind of exp. Result tN −1 (ms) Exp. max. deflection (mm) p0 (kPa) k̃ (kPa/mm) Figures. Citation

G748 Shock tube No break 65.53 10.5 1056.08 2.85 Figs. 4(a), 5–8(a), 13 [26]
G688 Shock tube Glass fracture 65.53 14.4 59.13 5.22 – [32]
FX14 Free field Glass fracture 65.53 18.9 137.72 68.24 Figs. 4(b), 9(a) [32]
FX11 Free field Glass fracture 65.53 15 257.59 12.34 – [32]
G684 Shock tube Glass fracture, 43.50 113.4 64.24 3.45 Figs. 4(c), 10 and 14 [26]
G749 Shock tube PVB tear 65.53 102.1 42.07 5.01 – [26]
Average: 112.15 18.852

Fig. 13. (a) Comparison of the model deflection (red line) using a cubic deflection dependent resistance polynomial and linear damping to experimental data (black line)
for experiment No. G748. (b) Cubic deflection dependent mass-specific resistance polynomial. (c) Linear velocity dependent mass-specific damping. (For interpretation of
the references to colour in this figure legend, the reader is referred to the web version of this article.)

4.2. Non-ideal elastic response and linear damping for weak and In practice, using a linearizing basis function approach is com-
strong loading putationally faster when compared to the numerical minimization
leading to model parameters as described in Section 3. This is due
to the experimental data which is used in sum expressions for set-
In Fig. 13(a) we compare the model deflection with experimen-
ting up the small matrix (23) that is inverted only once. The second
tal deflection data as obtained from solving (22) in the case of weak
approach is restricted to basis sets, the first approach can be used
shock-tube loading (see Fig. 4(a)) with Nx = 3, Nv = 1 and tN −1 =
for any kind of parameterizations of resistance functions and SDOF
65.5 ms. The presented area mass ratio A/m = 4.8 × 10−2 m2 kg−1
models, e. g. also including the dynamic design factor and changes
is in the order of the determined value a0 = 8.2 × 10−2 m2 kg−1 .
thereof. For the second approach, simultaneous inspection of the
Fig. 13(b) shows that a weak non-linearity of the resistance func-
basis set parameters and the deflection prediction can be used to
tion is required. Butterworth filtering was found to be superior to
obtain the best number of parameters, as well as best values. For
the normal filter (15).
suitable basis sets the parameters are expected to converge. This
Fig. 14 shows the case of strong loading (see Fig. 4(c)),
was the case for Fig. 13. As in the case of the minimization ap-
model deflection and resistance data, where Nx = 4, Nv= 1,
proach, average parameters can be determined for an overall SDOF
tN −1 = 63.7 ms, a0 = 5.8 × 10−2 kg m−2 , and normally
model and P–I diagram analysis.
filtered and reduced data were used. Similarly to Fig. 11
The basis function approach example does use almost no
we observe reasonable goodness-of-fit, especially for increasing
physical input for parameterization. We assumed a constant factor
deflection values. Also in the polynomial approach, without
A/m. No mass or geometry informations are extracted, for instance
implicit assumptions about the general shape of the resistance
as prefactor for the series expansions in (19). As in the case of
function, a strong reduction of the resistance (that can be
the numerical minimization approach we could have used a more
attributed to glass breakage) is observed after the first elastic
physics-based expansion parameter set that also takes account
increase. We have reasonable agreement between the mass-
of small geometry changes that allows for the transfer of the
specific deflection dependent resistance function Figs. 13(b) and
data to similar geometries. Within the basis function expansion
14(b), as well as between the velocity dependent resistance
approach the definition of parameters is more limited because of
Figs. 13(c) and 14(c). For strong loading, when reducing the
the linearization requirement of the formalism.
order of the deflection resistance polynomial, the model area over
mass ratio approaches the geometrical data. From a0 the dynamic
exp
loading factor can be derived, kLM = 0.83, which is compared well 5. Conclusions
with the theoretical load-mass factor of Section 2.
As Fig. 13(a) indicates models are feasible where the goodness- For the determination of SDOF models the following steps are
of-fit between SDOF model and experimental dynamic data is such proposed:
that one could start working with quantitative expressions for the (i) SDOF model selection; (ii) identify parameters that can be de-
goodness-of-fit, e.g. using the incomplete Gamma function [29]. termined from static data and geometry information; (iii) iden-
In the same direction the three-parameter deflection resistance tify general parameters that should be determined from dynamic
function of Fig. 11 and the six-parameter double polynomial data; (iv) generate experimental pressure and deflection data for
function of Fig. 14 could be compared quantitatively up to the different loading regimes; (v) apply minimization or basis set ap-
maximum deflection. proach to determine model parameters; (vi) select and combine
K. Fischer, I. Häring / Engineering Structures 31 (2009) 1677–1686 1685

Fig. 14. (a) Model deflection and experimental deflection for experiment No. G684. (b) Quadric deflection dependent resistance (c) Linear damping.

parameter data for overall SDOF model including failure criterion. [3] Morison CM. Dynamic response of walls and slabs by single-degree-of-
For the application of SDOF models two options applicable to the freedom analysis – A critical review and revision. Int J Impact Eng 2005;32:
1215–47.
investigated loading regime were discussed: (i) generate an iso- [4] Riedel W, Mayrhofer C. Customized calculation methods for explosion effects
damage curve which can be used as a fast look-up table; (ii) solve on structural building components. In: International symposium on structures
SDOF model for loading of interest and apply failure criterion. Es- under earthquake, impact, and blast loading 2008. Fraunhofer institute for
high-speed dynamics. Ernst-Mach-Institut; 2008.
pecially the latter option is flexible for use in quantitative hazard
[5] Wei J, Dharani LR. Fracture mechanics of laminated glass subjected to blast
and risk analysis tools, e. g. analysis of bombing explosions in urban loading. Theor Appl Fract Mech 2005;44:157–67.
environments. [6] Mayrhofer C. Reinforced masonry walls under blast loading. Int J Mech Sci
2001;44:1068–80.
The comparison of model deflection data obtained from
[7] Li QM, Ye ZQ, Ma GW, Reid SR. Influence of overall structural response on
SDOF models with simple common practice ideal-elastic and perforation of concrete targets. Int J Impact Eng 2006;926–41.
elastic–plastic resistance functions using best parameters fitted [8] Kinney GF, Graham KJ. Explosive shocks in air. Berlin [u.a.]: Springer-Verlag;
to experimental deflection revealed that the goodness-of-fit is 1985.
[9] Smith PD, Hetherington JG. Blast and ballistic loading of structures. Oxford:
rather poor for a standard laminated glass window. For strong Butterworth Heinemann; 1994.
loading a tri-linear resistance function was determined with much [10] Paz M, Leigh W. Structural dynamics. US: Springer; 2006.
better agreement using a numerical minimization procedure. [11] Li QM, Meng H. Pressure–impulse-diagram of blast loads based on dimensional
This indicates that more parameters or better parameterized analysis and single-degree-of-freedom model. J Eng Mech 2002;87–92.
[12] Biggs JM. Introduction to structural dynamics. New York: McGraw Hill; 1964.
resistance functions or SDOF models should be used to reproduce [13] Li QM, Meng H. Pulse loading shape effects on pressure–impulse diagram of
the experimental structure behavior. It was shown that the an elastic–plastic, single-degree-of-freedom structural model. Int J Mech Sci
numerical minimization procedure is sensitive to the kind of 2002;44:1986–98.
[14] D.o.t.A.t.N.a.t.A. Force. TM 5-1300 Structures of resists, the effects of accidental
parameterization and also indicates whether the parameterization
explosions. Departments of the Army, the Navy and the Air Force; 1990.
covers the physics of the blast loading and response. [15] Fallah AS, Louca LA. Pressure–impulse diagrams for elastic–plastic-hardening
The linear basis function approach allows a systematic increase and softening single-degree-of-freedom models subjected to blast loading. Int
of the number of parameters. In the case of weak loading a non- J Impact Eng 2006;34:824–42.
[16] Low HY, Hao H. Reliability analysis of RC slabs under explosive loading. Struct
linear polynomial cubic resistance function and linear damping, Safety 2001;23:157–78.
reproduced the damped oscillation. For strong loading a resistance [17] Razaqpur AG, Tolba A, Contestabile E. Blast loading response of reinforced
breakdown was found, reminding of the sawtooth-like resistance concrete panels reinforced with externally bonded GFRP. Composites: Part B
2006;38:535–46.
function as obtained from numerical minimization. The model-
[18] Alaoui S, Oswald C. Blast-resistant design considerations for precast,
experiment deflection agreement was reasonable over the first prestressed concrete structures. PCI J 2007;52:53–65.
positive deflection phase in both cases. The parameter definition [19] Schleyer GK, Hsu SS. A modelling scheme for predicting ths response of
was in both cases very sample specific. In particular the numerical elastic–plastic structures to pulse pressure loading. Int J Impact Eng 2000;24:
760–77.
minimization approach also operates with much more physically [20] Gantes CJ, Pnevmatikos NG. Elastic–plastic response spectra for exponential
motivated parameter definitions that support the application of blast loading. Int J Impact Eng 2003;323–43.
the SDOF model to similar geometries and materials. [21] Krauthammer T, Stevens DJ, et al. Reinforced concrete structures: Effects of
impulsive loads. Concrete Int 1990.
Future work could use better parameterizations of SDOF [22] Häring I, Kranzer C. Local and global approach to the blast response of
models, possibly including the dynamic design factors, the structures. In: 19th international symposium on military aspects of blast and
presented surface and mass of the glass pane. A better motivation shock. 2006.
[23] Gürke G, Kranzer C, Romani M. Application of the critical impulse in an critical
for the comparison time window and the compared experimental
time criterion for testing of blast resistant glazing simplified to the velocity
and model sizes would be desirable. Furthermore multiple failure criterion. In: 18th international symposium on military aspects of blast and
criteria could be determined and used simultaneously in SDOF shock. 2004.
model analysis within P–I diagrams thus covering all loading [24] Campidelli M, Viola E. An analytical–numerical method to analyze single
degree of freedom models under airblast loading. J Sound Vib 2007;302:
regimes. 261–86.
[25] Shi Y, Hao H, Li Z-X. Numerical derivation of pressure–impulse diagrams for
prediction of RC column damage to blast loads. Int J Impact Eng 2007;2–15.
References [26] Kranzer C. Experimentelle Untersuchung der Bewegung einer 7.5 mm
Verbundsicherheits-Glasscheibe bei Blastbelastung mit einem Laser-
Wegmesssystem. Fraunhofer-Institut für Kurzzeitdynamik. Ernst-Mach-
[1] Craig RR, Kurdila AJ. Fundamentals of structural dynamics. New Jersey: Wiley, Institut. Efringen-Kirchen; 2005.
J; 2006. [27] EN 13123-1. European standard for windows, doors and shutters – Explosion
[2] Gannon JC, Marchand KA. Approximation of blast loading and single degree- resistance – Requirements and classification – Part 1: Shock tube; 2001.
of-freedom modelling parameters for long span griders. In: International [28] EN 13541. European standard door glass in building – Security glazing –
conference on structures under shock and impact. 2006. p. 3–12. Testing and classification of resistance against explosion pressure; 2001.
1686 K. Fischer, I. Häring / Engineering Structures 31 (2009) 1677–1686

[29] Press WH, Teukolsky SA, Vetterling WT, Flannery BP. Numerical recipes, The [32] Kranzer C, Gürke G, Mayrhofer C. Testing of bomb resistant glazing systems
art of scientific computing. 3rd ed. Cambridge [u.a.]: Cambridge University experimental investigation of the time dependent deflection of blast loaded
Press; 2007. 7.5 mm laminated glass. Glass Processing Days. Finnland; 2005.
[30] Nelder JA, Mead R. A simplex method for function minimization. Comput J [33] Hirsch C. Numerical computation of internal and external flows. Fundamentals
1965;7:308–13. of numerical discretization, vol. 1. John Wiley & Sons; 1989.
[31] Brent RP. Algorithms for minimization without derivatives. New York: Dover, [34] Lynch DR. Numerical partial differential equations for environmental scien-
Prentice Hall; 1973. tists and engineers. New York: Springer; 2005.

You might also like