Structural Safety 108 (2024) 102448

Contents lists available at ScienceDirect

Structural Safety
journal homepage: [Link]/locate/strusafe

System-reliability-based sizing and shape optimization of trusses

considering millions of failure sequences
Lucas A. Rodrigues da Silva a, *, André J. Torii b, André T. Beck a
a
Department of Structural Engineering, University of São Paulo, São Carlos, SP, Brazil
b
Latin American Institute of Technology, Infrastructure and Territory (ILATIT), Federal University of Latin American Integration (UNILA), Foz do Iguaçu, PR, Brazil

A R T I C L E I N F O A B S T R A C T

Keywords: System-Reliability-Based Design Optimization (S-RBDO) of structures considering progressive collapse is a

Optimal design complex problem, as the number of potential failure sequences increases geometrically with the degree of static
Truss sizing and shape optimization indeterminacy of the structure. Existing methods for identifying failure sequences in structural systems are
Progressive collapse
computationally expensive and prone to missing some critical failure sequences, especially within an optimi­
System reliability
Failure sequence
zation framework. In this context, identifying the most critical failure sequences to simplify the problem is
fundamental. Herein, we propose a novel system-reliability-based framework for sizing and shape optimization
of trusses. The procedure identifies minimal cut sets using the recently developed null space method, which has
been proven more efficient than traditional failure path-based methods. The most probable failure sequence is
selected from each identified minimal cut set. System reliability is estimated using the dominant failure se­
quences for the whole structure, selected based on their correlations using the Probabilistic Network Evaluation
Technique (PNET). Craziness-Based Particle Swarm Optimization (CRPSO) is employed as the optimization al­
gorithm. Numerical examples involving hundreds to millions of failure sequences demonstrate applicability and
efficiency of the proposed framework on truss optimization problems with different material post-failure be­
haviors. Results suggest that, in a system-reliability analysis considering progressive collapse, the most critical
failure sequences are those obtained from minimal cut sets. Furthermore, results show that the procedure pro­
posed herein can outperform other frameworks based on traditional failure path-based methods. Simple truss
sizing and shape optimization is considered herein, but the conclusions have immediate relevance to the optimal
design of realistic structures considering progressive collapse.

1. Introduction system-reliability target allows for competition between different failure

modes, as shown in [13–17]. System Reliability-Based Design Optimi­
In recent decades, the optimization of truss structures has been zation (S-RBDO) [18–21] is especially relevant for the study of optimal
extensively investigated [1–5]. Most of these studies focused on deter­ design considering progressive collapse, a topic that has been widely
ministic optimization. However, for more effective management of risk investigated recently [14–16,22–24]. However, only a few studies have
of structural failures, uncertainties in loads and material properties need addressed S-RBDO of trusses considering progressive collapse [25–28].
to be properly addressed in the optimization. The popularization of Among the technical challenges of system-reliability-based optimiza­
structural reliability methods [6,7] has made this possible, leading to the tion, one can cite [29]: (a) statistical dependence between components,
development of the so-called Reliability-Based Design Optimization failure modes and failure sequences; (b) consideration of internal force
(RBDO) methods. In this context, recent papers have incorporated redistribution after local failures; (c) computational cost, especially
component reliability constraints into truss sizing, shape [8–10] and considering the thousands to millions of potential failure sequences.
topology [11,12] optimization. In S-RBDO considering progressive collapse, system failure is
Introducing system-reliability constraints in the optimization leads described by sequences of component failures. For real structures with a
to a more consistent formulation, as most civil structures remain func­ high degree of static indeterminacy, the number of potential failure
tional after failure of one or two components. Moreover, imposing a sequences easily reaches the millions, making the assessment of system

* Corresponding author.
E-mail address: araujolucasrs@[Link] (L.A. Rodrigues da Silva).

[Link]
Received 25 September 2023; Received in revised form 5 February 2024; Accepted 8 February 2024
Available online 15 February 2024
0167-4730/© 2024 Elsevier Ltd. All rights reserved.
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

reliability a computational demanding task. However, since most of n: good systems, where k = n characterizes a series system, k = 1 char­
these sequences are extremely unlikely, system reliability can be eval­ acterizes a parallel system, and 1 < k < n characterizes actual structural
uated based only on the most critical failure sequences [30,31]. This is systems. Each combination of element failure events leading to system
the premise of well-stablished failure path-based methods for identi­ failure is called a failure sequence. Every failure sequence can be rep­
fying structural failure modes or failure sequences, such as the incre­ resented as a parallel sub-system. Consequently, failure of a statically
mental load method [32,33], the β-unzipping method [34,35], and the indeterminate structure is represented as a series system, with each
branch-and-bound method [36]. Recent studies have proposed new component being a failure sequence. Hence, the system failure event is
techniques for identifying critical failure sequences, such as the B3- given by:
method [37], the selective search technique [38], the differential ⋃⋂
evolution-based search strategy [39] and surrogate-model-based Fsys = Fj , k = 1, 2, ⋯, nFS (3)
k j∈Ck
methods [40,41]. Yet, these techniques have two main limitations: (a)
they are very time-consuming for large problems, requiring several where Fj is the j th component failure event, Ck is the index set of ele­
structural analyses, and (b) the identified sequences are not necessarily
ments in the kth failure sequence and nFS is the number of failure se­
the most probable ones. To overcome these limitations, Wei and Deng
quences.
[42] recently developed a method for identifying failure sequences
Based on Eqs. (1) and (3), we can define the system failure domain
based on the null space of the structural equilibrium matrix. Using this
Ωf,sys as:
technique, the authors could efficiently identify critical failure se­
{ ⋃⋂ }
quences and compute system reliability of trusses. However, the prob­
Ωf ,sys (d) = x| gj (d, x) ≤ 0 , k = 1, 2, ⋯, nFS (4)
lems addressed by Wei and Deng [42] were limited to the assessment of k j∈Ck
system reliability (no optimization).
In this paper, we propose a novel framework for system-reliability- Hence, the system failure probability pf,sys can be formulated as:
based sizing and shape optimization of trusses using the null space ∫
[ ]
method and a craziness-based particle swarm optimization algorithm, pf ,sys (d, X) = P Fsys = fX (x)dx (5)
objectively addressing progressive collapse. Numerical examples Ωf ,sys

involving hundreds to millions of failure sequences demonstrate the

efficiency and applicability of the proposed framework in optimization The solution of Eq. (5) is not simple because there is statistical de­
problems with different material post-failure behaviors. pendency between the random variables, between elements of a failure
The remainder of this paper is organized as follows. Section 2 pre­ sequence and even between different failure sequences. Moreover, real
sents the formulation of the optimization problems addressed herein. A structures are composed of hundreds of elements and have a degree of
description of the proposed framework and methods is provided in static indeterminacy (gH ) typically larger than one. The number of po­
Section 3. In Section 4, three numerical examples are investigated: an tential failure sequences grows exponentially with the number of ele­
11-bar planar truss with 496 failure sequences, a 21-bar truss bridge ments and with gH , reaching the order of millions for gH as low as 4. This
with over half-million failure sequences and a 25-bar spatial truss tower makes evaluation of Eq. (5) a computationally demanding task. To
with over ten million failure sequences. Conclusions and main findings overcome these challenges, several methods have been proposed in the
are summarized in Section 5. last decades [32–39]. In the following, two of the most well-known
techniques for identifying failure sequences are briefly discussed.
2. Problem formulation A widely used technique for identifying critical failure sequences and
assessing system reliability is the β-unzipping method [34]. This method
2.1. Structural system reliability relies on an element-level reliability analysis, in which the minimum
reliability index is selected (βmin ). Critical elements within an interval
Let X ∈ RnRV and d ∈ RnDV be vectors of structural system parameters. [βmin , βmin + Δβ], where Δβ is a defined parameter, are identified. The
Vector X contains the nRV random variables of the problem (e.g., loads, system branches into new subsystems, assuming that each critical
material strengths, member dimensions, model uncertainties) and vec­ element has failed. For each subsystem, a new element-level reliability
tor d contains the nDV design variables, whose values will be determined analysis is performed, and the branching procedure repeats until the
to maximize system performance, minimize structural weight, cost, etc. structure collapses. The identified elements are then combined in a
Structural element dimensions, partial safety factors and parameters of parallel-series system for system reliability assessment. The β-unzipping
inspection and maintenance programs are examples of variables in method is easy to implement, but it strongly depends on parameter Δβ.
vector d. Adopting a small value may result in not detecting the most critical
Additionally, let gi (d, X) be the limit state function corresponding to sequences, whereas adopting a sufficiently large value can identify all
failure sequences at a significant increase in computational cost.
the ith element. The limit state equation gi (d, X) = 0 defines the
The branch-and-bound method [34,36] is an efficient technique for
boundary between the failure domain Ωf,i and the safety domain Ωs,i :
identifying critical failure sequences compared to the β-unzipping
Ωf ,i (d) = {x|gi (d, x) ≤ 0 }, method. It operates on the premise that the probability of a failure
sequence decreases monotonically as new elements are added to it. The
Ωs,i (d) = {x|gi (d, x) > 0 }, i = 1, 2, ⋯, nLS (1) search procedure also branches the original system into new subsystems,
similar to the β-unzipping method, but this branching is based on the
where x denotes a realization of X and nLS is the number of limit state probabilities of partial failure sequences, composed of elements that
functions. have been removed from the structure. Based on the probability of the
The failure probability corresponding to the ith element is given by most critical identified failure sequences, some partial failure sequences
[7]: can be discarded – an operation known as bounding. Thoft-Cristensen
∫ and Murotsu [34] suggest neglecting partial failure sequences with
pf ,i (d, X) = P[gi (d, X) ≤ 0 ] = fX (x)dx (2) probabilities smaller than 10− γ PfpM , where PfpM is the maximum of the
Ωf ,i
probabilities of the identified complete failure sequences and γ ≥ 0. The
where P[.] is the probability operator and fX (x) is the joint probability branch-and-bound method allows the identification of failure sequences
density function of the random variable vector. in decreasing order of probability. However, depending on the bounding
Structural systems and sub-systems may be characterized as k-out-of- criterion, important failure sequences can be prematurely discarded.

2
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

Furthermore, evaluating the probabilities of partial failure sequences degree of static indeterminacy (gH ), the number of mechanisms (m) and
whenever a new element is removed can increase the computational the rank of equilibrium matrix (rM ):
costs compared to the element-level reliability analysis performed in
m = nDOF − rM (8)
β-unzipping method.
In this paper, the null space method proposed by Wei and Deng [42]
gH = nel − rM (9)
is used to identify the failure sequences. Based on their occurrence
probabilities and on statistical correlations, only the most relevant The degree of static indeterminacy represents the number of states of
failure sequences are considered when evaluating Eq. (5). Details about self-stress of the structure, which can be considered as regions of statical
these techniques are presented in Section 3. For a comprehensive review indeterminacy. If a structure has at least one state of self-stress, it means
about other techniques, the reader is referred to Song et al. [29]. that it is pre-stressable. Similarly, a mechanism can be seen as a region of
instability.
2.2. System-reliability-based size and shape truss optimization The null space of M represents the set of element internal forces in
equilibrium with zero applied nodal loads, i.e., the states of self-stress of
In the Deterministic Design Optimization (DDO) formulation, un­ the structure. This subspace has dimension s. An orthonormal basis
certainties are implicitly addressed through partial safety factors. While matrix V for the null space can be obtained by performing singular value
this formulation yields mechanically optimal solutions, it may decomposition (SVD) on M. Throughout this paper, we refer to V as the
compromise structural safety, as the effects of uncertainties are not null space basis matrix, which is an nel × gH matrix. Each row of V
properly considered [13]. On the other hand, the Reliability-Based corresponds to a structural element. Also, from the definition of null
Design Optimization (RBDO) formulation arose as an alternative to space:
properly incorporate the effects of uncertainties in optimization prob­ MV = 0 (10)
lems [19]. In this formulation, the deterministic constraints in DDO are
replaced with reliability constraints. For the progressive collapse anal­ A straightforward method for identifying structural failure sequences is
ysis of truss systems, a more consistent formulation is achieved by to remove the columns of M corresponding to the failed elements and to
replacing the element reliability constraints with a single system reli­ compute the rank of M. Then, the number of mechanisms m can be
ability constraint, resulting in the System-Reliability-Based Sizing and evaluated using Eq. (8). If m = 0, there are no mechanisms and new
Shape Optimization formulation (S-RBSSO): elements can be removed. If m > 0, a mechanism is formed, making the
find: d* structure unstable, and the removed elements constitute a failure
∑ sequence. The procedure can be repeated for the identification of new
which minimizes : W(d) = ρi Ai Li (6) sequences. However, evaluating the rank of M whenever an element is
removed can be computationally expensive, especially for structures
subject to:pf,sys (d, X) ≤ pfT,sys with many elements or large gH .
Amin ≤ Ai ≤ Amax Exploring the properties of the equilibrium matrix and its null space,
i , i = 1, 2, ⋯, nel
Deng and Kwan [45] demonstrated that an element k is necessary to
i

ξmin ≤ ξj ≤ ξmax maintain geometric stability in a structure if:

j j , j = 1, 2, ⋯, nnd

Vk = 0 (11)
where: d = {A1 , ⋯, An , u1 , ⋯, um } is the design vector containing the
element cross-sectional areas (A ∈ Rnel ) and the nodal coordinates where Vk is the row vector of V corresponding to the element k.
considered as design variables (ξ ∈ Rnnd ); nel and nnd are the number of Based on [45], Wei and Deng [42] presented the necessary and
elements and the number of nodal coordinates; W(d) is the structural sufficient condition such that k > 1 elements to be removed from a truss
weight; ρi and Li are the specific mass and the length of the ith bar, structure form a cut set. This condition is given by:
respectively; Amin and Amax are the lower and upper bounds of the ith bar
i i r(Vk ) < k (12)
cross-sectional area; ξmin
j and ξmax
j are the lower and upper bounds of the
jth nodal coordinate; and R sys = 1 − pfT,sys is the target system reliability. where r(.) is the rank operator and Vk in this case is the submatrix of V
which contains the row vectors corresponding to the k removed ele­
3. New S-RBSSO framework ments.
It is interesting to note that matrix Vk (k × s) is significantly smaller
3.1. System reliability assessment than M (nDOF × nel ). Hence, the failure sequences of a truss structure can
be identified more quickly by evaluating the rank of Vk than the rank of
In a truss structure, the internal element forces (N) are related to the M. However, since the number of failure sequences can be very large,
applied nodal forces (P) through the equilibrium matrix (M), as follows: this procedure can still be computationally demanding. Based on the
condition expressed in Eq. (12) and exploring the linear dependence
MN = P (7) properties of the row vectors of V, Wei and Deng [42] proposed the null
space method, which can identify the minimal cut sets of truss structures
where M is a nDOF × nel matrix; N and P are, respectively, nel × 1 and in an efficient way. Cut sets are combinations of element failures
nDOF × 1 vectors; and nDOF is the number of degrees of freedom (DOFs) of (regardless of order) that lead to system collapse. A cut set is said to be
the structure. minimal when, if any element failure is removed from the set, the
The equilibrium matrix in Eq. (7) and its subspaces offer valuable remaining elements no longer form a cut set. Similarly, a failure
insights into the geometric stability of trusses, as demonstrated by Pel­ sequence is said to be minimal if it is obtained by permuting the ele­
legrino and Calladine [43] and Pellegrino [44]. In linear algebra, the ments of a minimal cut set. The null space method proposed by Wei and
rank of a matrix M, denoted as rM , represents the maximum number of Deng [42] is used in this paper to identify the minimal cut sets of the
linear independent columns in M, which is equal to its maximal number truss structures. The procedure comprises the following steps:
of linear independent rows. This value can be determined by trans­
forming M to row echelon form using Gaussian elimination, counting • Step 1: Assemble the equilibrium matrix M of the structure;
the zero rows, and subtracting it from the total number of rows. Pelle­
grino and Calladine [43] presented Eqs. (8) and (9), which relate the

3
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

• Step 2: Compute the matrix V (null space basis matrix) consisting of Table 1
the null space basis vectors of M, performing SVD on M; Minimal cut sets of the 3-bar truss for varying values of θ.
• Step 3: If there are zero column vectors in VT , remove these columns θ =0
◦
θ =1
◦
θ = 45
◦
θ = 89
◦
θ = 90
◦

from VT and add the corresponding elements to Θ as minimal cut

[2] [1, 2] [1, 2] [1, 2] [1]
sets, where Θ is the set containing the minimal cut sets; [1, 3] [1, 3] [1, 3] [1, 3] [2, 3]
• Step 4: Compute the row echelon form of VT , denoted as V ̃ T , and – [2, 3] [2, 3] [2, 3] –

identify b maximal linearly independent subsets by selecting

̃ T;
different columns of V
• Step 5: For each maximal linearly independent subset identified in of influence for the ith bar due to the residual strength of the lth bar,
Step 4, construct a matrix VTb and remove the corresponding columns applied to the nodes of bar l as forces; ηl is a factor that represents the
̃T; material model of bar l (for η = 0 the material is perfectly-brittle; for η =
of V
T 1 we have an elastic perfectly plastic material).
• Step 6: For each remaining column vector vp in V ̃ compute the
In this manuscript, interest is in developing insight into the optimal
( T )− 1
combination coefficients γ p , such that: γp = Vb vp . If there are design of structural systems considering progressive collapse involving
zero combination coefficients in γ p , obtain a minimal cut set by millions of failure sequences. As the solution of actual structural design
selecting elements corresponding to the nonzero components of γp problems is secondary, the choice of probability distributions for the
and that corresponding to vp ; random variables does not need to be the most realistic. On the other
hand, we want to be able to efficiently evaluate failure probability of
• Step 7: Repeat the procedure until all maximal linearly independent
individual bars, and correlation between element failures and between
subsets are found.
failure sequences. One reasonable choice to achieve this goal is to keep
limit states as linear functions of Gaussian random variables.1 Hence, for
The equilibrium matrix depends on both the connectivity of the el­
computational convenience, we represent Si N(μSi , σ Si ) and Pj N(μPj , σPj )
ements and the nodal coordinates. During the shape optimization pro­
cess, nodal coordinates can change, possibly leading to different as normally distributed random variables, enabling the efficient
minimal cut sets for the optimal candidate solutions. Consider the 3-bar computation of element failure probabilities using Cornell reliability
truss depicted in Fig. 1 as an example. Table 1 presents the identified indices [7,47]:
minimal cut sets for this structure as a function of the angle θ of bar 3. E[gi (d, X) ]
Note that, except at the extremes of the interval (θ = 0 and θ = 90 ),
◦ ◦ pf ,i (d, X) = Φ(− βi ), with βi = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (15)
Var[gi (d, X)]
the minimal cut sets are the same. Hence, by setting appropriate bounds
for the nodal coordinates (design variables), minimal cut sets will not In Eq. (15), βi is the reliability index, E[.] is the expected value operator
change during the optimization process, and can be identified based on and Var[.] is the variance operator, which are given by Eqs. (16) and
the initial configuration of the structure. This strategy is adopted herein (17), respectively:
to reduce the computational cost of S-RBDO. ( )
In this paper, the failure sequences used to evaluate system reliability ∑ ∑
nF i− 1
E[gi (d, X) ] = μSi Ai − sign(Ni ) bij μPj + ail ηl μSl Al (16)
in Eq. (5) are obtained by permuting the elements of the minimal cut sets j=1 l=1
in Θ. Let F k be the kth failure sequence. The limit state function of the ith
element in F k is [46]: ∑
nRV (
∂gi
)2 nRV ∑
∑ nRV
∂gi ∂gi
Var[gi (d, X) ] = σ2Xi + ρXi Xj σ Xi σ Xj (17)
gi (d, X) = Ri (d, X) − |Ni (d, X)|, i = Ck1 , Ck2 , ⋯, Ckm (13) i=1
∂Xi i=1 j=1,j∕
=i
∂Xi ∂Xj

where: Ck = {Ck1 , Ck2 , ⋯, Ckm } is the index set of the m elements in the where μXi and σXi are the mean value and the standard deviation of Xi ,
kth failure sequence; Ri (d, X) and Ni (d, X) are, respectively, the axial and ρXi Xj is the correlation coefficient between the random variables Xi
strength and the normal force on the ith bar. and Xj .
Herein, different material post-failure behaviors and internal force As mentioned in Section 2.1, a failure sequence can be represented as
redistribution after an element failure are addressed. Therefore, Eq. (13) a parallel system. Hence, the failure probability of F k is:
can be rewritten as: [⋂ ]
(
∑
nF ∑
i− 1
) P[F k ] = pCk = P gi (d, X) ≤ 0 = Φm (− βk , ρk ) (17)
gi (d, X) = Si Ai − sign(Ni ) bij Pj + ail ηl Sl Al (14) i∈Ck

j=1 l=1
In Eq. (17), Φm is the m-dimensional standard normal cumulative dis­
where: Si is the allowable stress of the ith bar; bij is the coefficient of tribution function; βk is the vector containing the reliability indexes of
influence for the ith bar due to the jth force Pj ; nF is the number of forces; l the elements in F k ; and ρk = [ρij ] is the correlation matrix of F k , with
is an index for the bars of F k that failed before bar i; ail is the coefficient ρij = αTi αj , where αi is the vector of direction cosines of the normal to the
hyper-plane defined by the limit state equation gi = 0 in standard
normal space. The reliability index for the kth failure sequence is given
by:

βCk = Φ− 1 (pCk ) (18)

where Φ− 1 is the inverse standard normal distribution function.

As shown in the following, the use of PNET method for assessing

1
As future research we will address handling of other probability distribu­
tions, using other strategies to compute failure probabilities, such as the First
Fig. 1. 3-bar truss. Order Reliability Method (FORM).

4
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

system reliability requires the correlation coefficients between different equations:

failure sequences. Determining these correlations is challenging because
xi (t) = xi (t − 1) + vi (t) (21)
the limit state functions of the failure sequences are not known. To
overcome this challenge, we employ the equivalent planes method ( )
vi (t) = vi (t − 1) + c1 r1 xpbest (t − 1) − xi (t − 1) (22)
proposed by Gollwitzer and Rackwitz [48]. This method numerically i

derives an equivalent vector of direction cosines αCi for a given failure ( )

+ c2 r2 xgbest (t − 1) − xi (t − 1) , i = 1, ⋯, Npop , t = 1, ⋯, Nit
sequence i. The correlation coefficient ρCi Cj between failure sequences i
and j can then be computed as: where: vi (t) represents the velocity of the ith particle at iteration t; c1 and
T
ρ Ci Cj = α α Cj (19) c2 are, respectively, the cognitive and social weighting coefficients; r1
and r2 are random numbers uniformly taken from the interval [0,1]; Npop
Ci

Since the failure sequences resulting from a minimal cut set share many is the population size; Nit is the maximum number of iterations.
random variables, they are expected to be strongly correlated. Hence, On the other hand, in the CRPSO method the velocity expression in
the failure sequence with the highest probability of occurrence is Eq. (22) is modified by introducing a new set of random numbers and an
selected from each minimal cut set as the representative failure operator called craziness velocity:
sequence. With the probabilities of occurrence and the correlation co­

( ) ( )
vi (t) = r2 sign(r3 )vi (t − 1) + (1 − r2 )c1 r1 xpbest
i (t − 1) − xi (t − 1) + (1 − r2 )c2 (1 − r1 ) xgbest (t − 1) − xi (t − 1) + P(r4 )sign(r4 )vcraziness
i (23)

efficients between the representative failure sequences identified for where: r1 , r2 , r3 and r4 are random numbers uniformly taken from the
each minimal cut set, system failure probability in Eq. (5) is computed interval [0, 1]; sign(r3 ) is a function defined as:
using the Probabilistic Network Evaluation Technique – PNET [47]. This {
− 1, r3 ≤ 0.05
technique is based on the premise that the failure sequences with cor­ sign(r3 ) = (24)
1, r3 > 0.05
relations coefficients larger than a threshold correlation (ρ0 ) are
considered to be perfectly correlated. Within each group of strongly
vcraziness is a vector of parameters uniformly taken from the interval
correlated failure sequences, the one with the highest probability of [ i min max ]
vi , vi ; P(r4 ) and sign(r4 ) are functions defined as:
occurrence is assumed as the critical failure sequence. On the other
hand, failure modes with correlations smaller than ρ0 are considered {
1, r4 ≤ PCR
statistically independent. Based on this, the system failure probability P(r4 ) = (25)
0, r4 > PCR
can then be estimated as:
{
∑
nCFS − 1, r4 ≥ 0.5
sign(r4 ) =
pf ,sys = P[F i ] (20) 1, r4 < 0.5
i=1

and PCR is a predefined probability of craziness.

where nCFS is the number of identified critical failure sequences. To avoid unnecessary oscillations in the last iterations, Kar et al. [49]
suggest choosing PCR = 0.3 and vcraziness i as a small value. In this study,
3.2. Optimization algorithm we consider vcraziness
i = 0.001v max
i and cognitive and social weighting
coefficients are c1 = c2 = 2.05.
In this paper, we employ an improved Particle Swarm Optimization Finally, system reliability constraint in Eq. (6) is handled using the
(PSO) method, called Craziness based Particle Swarm Optimization penalty function technique with the following penalty scheme [51,52]:
(CRPSO) [4,49], to solve the S-RBSSO problems. The PSO, originally {
W(d), if pf ,sys ≤ pfT,sys
proposed by Kennedy and Eberhart [50], is a population-based heuristic ̃
W(d, X) = (26)
(1 + ε1 v)ε2 W(d), otherwise
algorithm inspired by the social behavior of a swarm of birds. This
method has three main advantages: (a) it has fast convergence; (b) it is where: W(d,
̃ X) is the objective function of the unconstrained prob­
easier to implement than other heuristic algorithms; (c) it uses a few lem; ε1 = 1.5 • (1 + it/itmax ) and ε2 = 1.5; and v is the violation of the
control parameters. Besides, PSO can handle non-differentiable objec­ system reliability constraint, given by:
tive functions, such as those addressed in this paper. However, PSO and
pf ,sys
most other population-based algorithms suffer from problems of pre­ v = max(0, − 1) (27)
pfT,sys
mature convergence and stagnation. To overcome these limitations, Kar
et al. [49] modified the conventional PSO by introducing a new velocity In this paper, we employ the CRPSO as the optimization algorithm.
expression that includes random numbers and a craziness velocity However, it is worth noting that other heuristic optimization algorithms
associated with a predefined probability of craziness. could be used within the proposed framework. Hence, for comparison
In conventional PSO, each particle is a candidate solution of the with future studies, in the results section we present the statistical results
optimization problem. The position xi (t) of the ith particle at an iteration of several runs for each example. Fig. 2 illustrates the flowchart of the
t is updated using two main information: the best previous position of CRPSO and system reliability assessment in the proposed framework.
the ith particle, xpbest
i , and the best global position among all the particles,
xgbest . Mathematically, the positions are updated using the following

5
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

Fig. 2. Flowchart of the proposed framework.

4. Numerical results and discussion the optimal solutions of (a) the maximum size of the failure sequences,
(b) different material post-failure behaviors and (c) different correla­
In this section, we demonstrate applicability and efficiency of the tions between random variables. Finally, in the third example, we
proposed framework in different S-RBSSO problems. Three numerical demonstrate the applicability of the proposed framework in a large
examples in increasing order of complexity are addressed: an 11-bar problem with millions of failure sequences. To the best of our knowl­
planar truss (Fig. 3(a)); a 21-bar planar truss bridge (Fig. 3(b)); and a edge, no similar study has addressed the optimal design of trusses with a
25-bar spatial truss tower (Fig. 9 (a)). In the first example, we compare degree of static indeterminacy greater than two.
the optimal solutions resulting from the proposed framework to those For all problems, Young’s modulus and material density were
obtained by using β-unzipping method [34] for identifying critical assumed to be 200GPa and 7, 850kg/m3 . Random variables were
failure sequences. In the second example, we investigate the effects on assumed to follow a Gaussian distribution. The mean values of the loads

6
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

Fig. 3. Numerical examples: (a) 11-bar truss; (b) 21-bar truss.

were set to 100kN for the planar problems and 50kN for the spatial
problem. In both cases, the coefficients of variation for the loads were Table 2
Optimal results for the 11-bar planar truss obtained by: proposed framework,
assumed to be 0.10. The mean values and the coefficients of variation of
β-unzipping with Δβ = 8, and β-unzipping with Δβ = 4.
the allowable stresses were set to 25MPa and 0.05, respectively, for all
problems. A threshold correlation of 0.85 was adopted in PNET. More­ Design Variables Proposed framework β-unzipping β-unzipping
(Δβ = 8) (Δβ = 4)
over, a target system reliability βT,sys = 3.0 (PfT,sys = 1.35 × 10− 3 ) was
A1 (cm2) 0.12 0.10 0.15
assumed. Buckling failure of the elements was disregarded. The popu­
A2 (cm2) 0.98 1.84 4.97
lation size and the maximum number of iterations of CRPSO were set to A3 (cm2) 8.35 11.85 2.93
50 and 100, respectively. Due to stochastic nature of CRPSO, 100 in­ A4 (cm2) 9.69 8.88 6.09
dependent optimization runs were performed for the first two problems A5 (cm2) 0.18 6.31 3.41
and 20 runs were performed for the third problem. Moreover, to check A6 (cm2) 8.53 9.83 11.45
A7 (cm2) 7.73 0.22 3.89
the feasibility of optimal solutions, system failure probabilities were A8 (cm2) 6.99 4.69 1.07
post-processed using Monte Carlo simulation (MCS) with 106 samples. A9 (cm2) 1.88 2.91 5.41
Finally, as discussed in section 3.1, the repetitive identification of A10 (cm2) 0.19 8.14 4.57
minimal cut sets would result in a prohibitive computational cost. A11 (cm2) 5.71 0.29 3.05
y4 (cm) 296.54 299.34 292.37
However, as demonstrated by the 3-bar example in Fig. 1, the minimal
y5 (cm) 195.37 178.82 171.98
cut sets remain unchanged when the nodal coordinate bounds are y6 (cm) 6.59 5.51 11.35
properly defined. Hence, in the following sections, the bounds adopted Best weight (kg) 82.55 82.64 79.49
for the design variables are presented for each example. The dependency Worst weight (kg) 92.53 102.56 95.11
Average weight (kg) 86.82 87.51 84.43
on the bounds of design variables, as well as the use of Gaussian random
Weight CV (%) 3.85 5.06 5.86
variables, are limitations of the proposed framework that require Average CPU time (s) 694.78 2218.55 403.42
attention in future studies. βsys,MCS 3.10 3.13 2.40

7
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

Fig. 4. Best optimal solutions for: (a) Proposed framework, (b) β-unzipping with Δβ = 8, (c) β-unzipping with Δβ = 4. Line thickness proportional to cross-
section area.

4.1. Example 1: 11-bar planar truss with gH = 3 and 496 failure

Table 3
sequences
Twenty most critical failure sequences and corresponding reliability indexes for
the optimal 11-bar solution obtained with the proposed framework.
The truss under study is depicted in Fig. 3 (a). This truss has 3 de­
grees of static indeterminacy and a total of 496 failure sequences.2 The β-unzipping method (Δβ = 8) Null space method
Failure sequence Reliability Failure sequence Reliability
optimization using the β-unzipping method was performed considering index index
Δβ = 4 and Δβ = 8. Additionally, the elements were considered
7→5 3.34 7→5 3.34
perfectly ductile (η = 1). The cross-sectional areas of all bars and the
7 → 10 3.34 7 → 10 3.34
vertical coordinates of nodes D, E and F were considered as continuous 9 → 7 → 10 3.35 7 → 11 3.36
design variables, with bounds set at [0.10 − 30]cm2 and [5 − 300]cm. 9→7→5 3.35 9→8 3.45
Hence, this problem involves a total of 14 design variables. 7 → 9 → 10 3.35 11 → 5 3.49
7→9→5 3.35 11 → 10 3.49
Results of the optimization using both frameworks are presented in
9 → 7 → 11 3.36 6→2→3 3.61
Table 2 and Fig. 4. The design variable values corresponding to the best 7 → 11 3.36 6→2→5 3.61
solutions (with the lowest objective function values) are presented in 7 → 9 → 11 3.36 6→2→7 3.61
Table 2 for each method. In addition, statistical results including the 9→8 3.45 6→9 3.61
worst weight, average weight, and coefficient of variation (CV) in 100 11 → 10 3.49 6 → 2 → 10 3.61
11 → 5 3.49 6 → 2 → 11 3.61
runs are presented. In Fig. 4, the best optimal trusses are shown. System 11 → 7 3.50 9→2→3 3.71
reliability indices obtained with MCS (βsys,MCS ) and the average CPU time 9 → 11 → 10 3.50 9→2→5 3.71
are also presented. 9 → 11 → 5 3.50 9→2→7 3.71
From the data in Table 2, it can be observed that the three best 9 → 11 → 7 3.50 9 → 2 → 10 3.71
11 → 9 → 10 3.50 9 → 2 → 11 3.71
weights are similar. As depicted in Fig. 4, the optimal shapes (especially
11 → 9 → 5 3.50 9→4 3.79
solutions (a) and (b)) are also quite similar. Since node F is unloaded, its 11 → 9 → 7 3.50 8→2→5 4.03
removal from the structure would not compromise equilibrium. As we 6→9 3.61 8→2→7 4.03
are not addressing topology optimization problems, the optimization Total number of 496 Total number of 120
algorithm deals with this node by overlapping nodes C and F, as if they failure sequences failure sequences

were the same node.

The S-RBSSO solution using β-unzipping framework with Δβ = 4 by the coefficient of variation. Besides, the average CPU time was
outperformed the others in terms of the best weight, average weight, and significantly lower for the proposed framework. To interpret this dif­
the average CPU time. However, post-processing reveals that the best ference, in Table 3 we present the most critical failure sequences (i.e.,
optimal solution obtained with this approach is not feasible as system sequences with highest probabilities of occurrence) identified by each
reliability constraint is not respected. This finding may be related to the method for the optimal solution in Fig. 4 (a). The reliability indexes of
fact that β-unzipping method with Δβ = 4 could not identify all the those sequences and the total number of identified failure sequences are
critical failure sequences, yielding optimal solutions with unacceptable also presented. Moreover, Fig. 5 shows the correlation coefficients (Eq.
levels of safety. Furthermore, this result shows that β-unzipping is (19)) between the most critical failure sequences identified using the
strongly parameter-dependent since, depending on the value of Δβ, the β-unzipping method.
identified sequences may not be the most critical. Other methods such as As shown in Table 3, seven of the top twenty critical sequences are
the branch-and-bound [36] and the selective search technique [38] common to both methods. However, the other thirteen sequences
suffer from a similar limitation. identified by β-unzipping method are not minimal. As an example, the
On the other hand, the optimal solutions obtained with the proposed first sequence, 7 → 5 is minimal as it is derived from the minimal cut set
framework and with the β-unzipping framework using Δβ = 8 satisfied [5,7]. However, this is not the case for sequences 9 → 7 → 5 and 7 → 9 →
system reliability constraint. This result indicates that both methods 5, because if element 9 was removed, they would still be failure se­
could identify the critical failure sequences. The solutions obtained with quences. Besides, the correlation matrix in Fig. 5 indicates that these
the proposed framework presented the lowest variability, as indicated sequences are perfectly correlated with the corresponding minimal
failure sequence. Similar observations apply to other sequences, such as
9 → 7 → 10 and 7 → 9 → 10, which are strongly correlated with 7 → 10.
2
The total number of failure sequences was identified using the β-unzipping This result suggests that many sequences identified by the β-unzipping
method with Δβ = 8. The null space technique only allows one to find the method are redundant. Therefore, in a system reliability analysis
minimal failure sequences.

8
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

9 9 7 7 9 7 9 9 9 11 11 11

7 7 7 7 9 9 7 7 9 9 11 11 11 11 11 11 9 9 9 6

5 10 10 5 10 5 11 11 11 8 10 5 7 10 5 7 10 5 7 9
1
7 5 1 1 1 1 1 1 1 1 1 0.29 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.64

7 10 1 1 1 1 1 1 1 1 1 0.29 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.67 0.64

9 7 10 1 1 1 1 1 1 1 1 1 0.31 0.67 0.67 0.67 0.67 0.67 0.68 0.67 0.67 0.68 0.65
0.9
9 7 5 1 1 1 1 1 1 1 1 1 0.31 0.67 0.67 0.67 0.67 0.67 0.68 0.67 0.67 0.68 0.65

7 9 10 1 1 1 1 1 1 1 1 1 0.31 0.67 0.67 0.67 0.67 0.67 0.68 0.67 0.67 0.68 0.65

7 9 5 1 1 1 1 1 1 1 1 1 0.31 0.67 0.67 0.67 0.67 0.67 0.68 0.67 0.67 0.68 0.65 0.8
9 7 11 1 1 1 1 1 1 1 1 1 0.29 0.68 0.68 0.68 0.68 0.68 0.68 0.68 0.68 0.68 0.62

7 11 1 1 1 1 1 1 1 1 1 0.29 0.68 0.68 0.68 0.68 0.68 0.68 0.68 0.68 0.68 0.62

7 9 11 1 1 1 1 1 1 1 1 1 0.29 0.68 0.68 0.68 0.68 0.68 0.68 0.68 0.68 0.68 0.62 0.7

9 8 0.29 0.29 0.31 0.31 0.31 0.31 0.29 0.29 0.29 1 0.29 0.29 0.29 0.3 0.3 0.3 0.3 0.3 0.3 0.34

11 10 0.67 0.67 0.67 0.67 0.67 0.67 0.68 0.68 0.68 0.29 1 1 1 1 1 1 1 1 1 0.64

11 5 0.67 0.67 0.67 0.67 0.67 0.67 0.68 0.68 0.68 0.29 1 1 1 1 1 1 1 1 1 0.64 0.6

11 7 0.67 0.67 0.67 0.67 0.67 0.67 0.68 0.68 0.68 0.29 1 1 1 1 1 1 1 1 1 0.63

9 11 10 0.67 0.67 0.67 0.67 0.67 0.67 0.68 0.68 0.68 0.3 1 1 1 1 1 1 1 1 1 0.64
0.5
9 11 5 0.67 0.67 0.67 0.67 0.67 0.67 0.68 0.68 0.68 0.3 1 1 1 1 1 1 1 1 1 0.64

9 11 7 0.67 0.67 0.68 0.68 0.68 0.68 0.68 0.68 0.68 0.3 1 1 1 1 1 1 1 1 1 0.64

11 9 10 0.67 0.67 0.67 0.67 0.67 0.67 0.68 0.68 0.68 0.3 1 1 1 1 1 1 1 1 1 0.64
0.4
11 9 5 0.67 0.67 0.67 0.67 0.67 0.67 0.68 0.68 0.68 0.3 1 1 1 1 1 1 1 1 1 0.64

11 9 7 0.67 0.67 0.68 0.68 0.68 0.68 0.68 0.68 0.68 0.3 1 1 1 1 1 1 1 1 1 0.64

6 9 0.64 0.64 0.65 0.65 0.65 0.65 0.62 0.62 0.62 0.34 0.64 0.64 0.63 0.64 0.64 0.64 0.64 0.64 0.64 1
0.3

Fig. 5. Correlation matrix of the twenty most critical failure sequences identified with β-unzipping method (Δβ = 8) for the optimal 11-bar truss.

220 Table 4
Computational costs for identifying minimal cut sets and assessing system reli­
ability of optimal solution for the 11-bar truss problem.
200
Objective Function (kg)

Steps CPU time

180 Minimal cut sets search 0.39 s

System reliability analysis (A + B) 0.55 s
Component reliability analyses (A) 0.35 s
160 Computation of probabilities of failure sequences (B) 0.20 s

140
considering progressive collapse, they could be represented by the cor­
120 responding minimal failure sequences, reducing computational cost.
Hence, the lower CPU time for the proposed framework can be explained
by the fact that the null space method only identifies the minimal failure
100
sequences, and these are sufficient for an accurate estimate of system
reliability.
80
0 20 40 60 80 100
The convergence history for an optimization run using the proposed
framework is depicted in Fig. 6. The curve illustrates that the neigh­
Iteration
borhood of the best solution is quickly found, with most of the iterations
Fig. 6. Convergence history for one optimization run of the proposed method. being spent on refining the solution, which is a known feature of heu­
ristic algorithms. Additionally, Table 4 presents the computational costs

9
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

Fig. 7. Best optimal solutions for the 21-bar truss bridge considering different material post-failure behaviors and correlations between random variables: a) brittle
structure B; b) ductile structure D; c) structure BC (brittle with correlation); d) structure DC (ductile with correlation).

for identifying minimal cut sets and assessing system reliability for the allowable stresses of bars within the same group were assumed to be
11-bar truss problem using the proposed framework. completely correlated (ρ→1), while the correlation coefficients of the
allowable stresses for bars in distinct groups were set at 0.7. Addition­
4.2. Example 2: 21-bar truss bridge with gH = 4 and over half a million ally, correlation coefficients between the applied loads were also
failure sequences assumed to be 0.7. The best optimal trusses and the post-processed
system reliability indices are shown in Fig. 7. As can be seen from the
In this section, we present the optimization results for the 21-bar figure, the best optimal shapes for structures D, BC and DC (Fig. 7 (b), (c)
truss shown in Fig. 3 (b). This example has 4 degrees of static indeter­ and (d), respectively) were very similar. All optimal trusses have thicker
minacy (gH = 4) and 537,048 failure sequences, identified using bars in the upper and lower chords and in the extreme verticals. Addi­
β-unzipping method with Δβ = 8. The vertical coordinates of nodes F-J tionally, the reliability indices indicate that the proposed framework
and element cross-sectional areas are the continuous design variables, provided optimal solutions that satisfy system reliability constraints.
with bounds set at [100 − 500]cm and [0.10 − 30]cm2 . Based on the In Table 5, we present the design variables corresponding to the best
symmetry of the structure, cross-sectional areas are grouped into eleven solutions and the statistical results for each optimization. Results for
different groups; hence, the problem has 14 design variables. uncorrelated random variables (structures B and D) indicate that, on
The optimization was initially performed considering both brittle average, the optimal solution considering ductile behavior had a lower
and ductile post-failure behaviors and different correlations between weight than its brittle counterpart. The same is observed for the best
random variables. Four structures were analyzed: a brittle structure weight. This difference can be attributed to the fact that ductile elements
(denoted by B) and a ductile structure (D), both with uncorrelated have non-zero residual strengths. Hence, a structure composed of such
random variables; and similarly, brittle (BC) and ductile (DC) structures elements requires a lower volume of material compared to a structure
with correlated random variables. For the BC and DC structures, composed of brittle elements. On the other hand, comparing the results
considering correlations between random variables reveals an inter­
esting outcome. For the brittle structure, the optimization algorithm
Table 5
Optimal results of the 21-bar truss bridge considering brittle and ductile material addresses the effects of correlations by changing the shape of the
post-failure behaviors and correlations between random variables. structure, as depicted in Fig. 7 (a) and (c), while the best and average
weight remain practically unchanged. However, for the ductile structure
Design variables B D BC DC
the optimal shape remains nearly identical, with increases in best and
2
A1, A4 (cm ) 2.42 0.80 0.19 0.11 average weights.
A2, A3 (cm2) 5.82 7.14 7.20 8.39
The influence of the maximum size of the failure sequences in the
A5, A8 (cm2) 5.72 7.27 8.12 9.52
A6, A7 (cm2) 7.40 8.48 9.04 10.49 optimal solutions obtained with the proposed framework is shown in
A9, A13 (cm2) 5.52 7.42 8.47 8.59 Fig. 8. Since the 21-bar truss analyzed in this section has 4 degrees of
A10, A12 (cm2) 3.65 2.17 2.65 3.24 static indeterminacy, the maximum number of elements in a failure
A11 (cm2) 2.70 2.54 1.05 0.94
sequence is 5. Hence, we performed the optimization considering brittle
A14, A21 (cm2) 4.47 7.67 7.95 8.64
A15, A20 (cm2) 3.91 1.11 0.25 0.12
and ductile post-failure behaviors with uncorrelated random variables
A16, A19 (cm2) 2.40 2.76 3.69 4.11 and different maximum sizes for the failure sequences. As shown in
A17, A18 (cm2) 1.18 0.12 0.43 0.26 Fig. 8(a), the average optimal weight did not change significantly when
yF, yJ (cm) 116.27 178.28 165.86 113.73 failure sequences with 4 and 5 elements were disregarded. The same
yG, yI (cm) 391.87 336.32 356.87 309.53
held true for system failure probabilities. As indicated in Fig. 8(b), sys­
yH (cm) 460.88 380.34 372.63 322.63
Best weight (kg) 235.16 226.60 237.54 250.52 tem failure probabilities for optimal solutions, considering failure se­
Worst weight (kg) 300.17 267.80 289.63 287.54 quences containing at most 3 elements, were very close to the target
Average weight (kg) 259.23 252.00 264.11 264.49 reliability. These findings suggest that the progressive collapse of
Weight CV (%) 8.81 5.90 6.50 4.37
optimal structures is mainly controlled by the shorter failure sequences.
Average CPU time (h) 2.59 4.00 2.77 4.01
βsys,MCS 3.01 3.33 3.17 3.20
Additionally, as the null space method used in this framework first

10
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

270 100
Brittle Brittle
Ductile Ductile
260 pfT,sys

Average weight (kg)

10-1

Failure probability
250

240 10-2

230
10-3
220

210 10-4
2 3 4 5 2 3 4 5
Maximum size of the failure sequences Maximum size of the failure sequences
(a) (b)

4
Brittle
Average CPU time (h) 3.5 Ductile

2.5

1.5

0.5

0
2 3 4 5
Maximum size of the failure sequences
(c)

Fig. 8. Optimization results for the 21-bar truss bridge, considering different maximum sizes of failure sequences: (a) average optimal weight; (b) system failure
probability; (c) average CPU time.

identifies the minimal cut sets and only then compute the probability of
Table 6
failure sequences, significant computational time can be saved in an
Computational costs for identifying minimal cut sets and assessing system reli­
optimization by disregarding failure sequences containing many ele­
ability of optimal solutions for the 21-bar truss problem.
ments, as illustrated in Fig. 8 (c).
Steps B D BC DC Table 6 presents the computational costs for identifying minimal cut
Minimal cut sets search 23.81 24.34 25.47 25.94 sets and computing system failure probabilities for the best optimal
s s s s
System reliability analysis (A + B) 18.29 26.73 18.45 27.55
s s s s
Table 7
Component reliability analyses (A) 16.82 25.13 16.91 25.88
s s s s Loading of the 25-bar spatial truss.
Computation of probabilities of failure 1.47 s 1.60 s 1.53 s 1.67 s Node Px (kN) Py (kN) Pz (kN)
sequences (B)
A 50 − 50 − 50
B 0 − 50 − 50
C 50 0 0
F 50 0 0

Fig. 9. 25-bar spatial truss tower problem: (a) Ground structure; (b) Best optimal configuration.

11
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

Table 8
Nodal coordinates and member grouping of the 25-bar spatial truss.
Node x (cm) y (cm) z (cm) Group Element (end nodes)

A − 92.25 0 508 A1 1(A,B)

B 92.25 0 508 A2 2(A,D), 3(B,C), 4(A,E), 5(B,F)
C − 92.25 92.25 254 A3 6(B,E), 7(B,D), 8(A,C), 9(A,F)
D 92.25 92.25 254 A4 10(C,F), 11(D,E)
E − 92.25 − 92.25 254 A5 12(C,D), 13(E,F)
F − 92.25 − 92.25 254 A6 14(C,J), 15(F,G), 16(D,I), 17(E,H)
G − 254 254 0 A7 18(C,H), 19(D,G), 20(F,I), 21(E,J)
H 254 254 0 A8 22(C,G), 23(D,H), 24(E,I), 25(F,J)
I 254 − 254 0
J − 254 − 254 0

solutions outlined in Table 5 and Fig. 7, using the proposed framework. distinct groups were set at 0.7.
Interestingly, the time required to identify all minimal cut sets using the In Table 9, we present the design variables, the structural weight and
null space method is approximately 25 s, which is comparable to the the post-processed reliability index corresponding to the best optimal
CPU time reported by Kim et al. (2013) [38] (23.77 s) for identifying 10 solution. Statistical results for 20 optimization runs are also presented in
failure sequences in a very similar 25-bar truss bridge with the selective Table 9. The best optimal configuration is depicted in Fig. 9(b). As
search technique. Furthermore, the CPU time for assessing system reli­ shown in Table 9, the proposed framework could handle an S-RBSSO
ability using the proposed method was 27.55 s, significantly smaller problem involving over 10 million failure sequences. Besides, the post-
than the time reported by [38] (563.38 s) for the same 25-bar problem processed reliability index indicates that the best optimal solution sat­
but using the matrix-based system reliability method. isfies system reliability constraint. Computational costs for identifying
minimal cut sets and assessing system reliability for the optimal solution
are presented in Table 10.
4.3. Example 3: 25-bar truss tower with gH = 7 and millions of failure
sequences
5. Concluding remarks
In this section, we present the optimization results for the 25-bar
This paper presented a new framework for System-Reliability-Based
truss depicted in Fig. 9 (a). Details of the loading, member grouping
Size and Shape Optimization (S-RBSSO) of trusses, based on the null
and nodal coordinates of the 25-bar truss are presented in Tables 7 and
space method for identifying minimal failure sequences. The proposed
8. This structure has 7 degrees of static indeterminacy (gH = 7) and
framework aims at an efficient identification of the most critical failure
10,014,151 failure sequences (identified using β-unzipping method with
sequences, reducing the computational burden of evaluating system
Δβ = 8). Based on the symmetry of the structure, the element cross-
reliability in problems with millions of potential failure sequences. The
sectional areas were combined into 8 groups, as shown in Table 8.
applicability of the framework was demonstrated through three nu­
Nodal coordinates were grouped into 5 groups: the x, y and z coordinates
merical examples: an 11-bar truss, a 21-bar truss bridge and a 25-bar
of nodes C-F, the x and y coordinates of nodes G-J and the x coordinates
spatial truss tower involving 496, over half a million and over ten
of nodes A and B. Hence, this problem has 13 design variables. The
million failure sequences, respectively. The results presented for the 11-
design variables bounds are: [0.5 − 50]cm2 for the cross-sectional areas;
bar truss show that, in a system reliability analysis considering pro­
[50.8 − 152.4]cm for the x-coordinates of nodes C-F; [101.6 − 203.2]cm
gressive collapse, many failure sequences are redundant and can be
for the x coordinates of nodes G-J; [101.6 − 203.2]cm for the y-co­
represented by the corresponding minimal failure sequence. Conse­
ordinates of nodes C-F; [254 − 355.6]cm for the y coordinates of nodes G-
quently, S-RBSSO using the proposed framework reduced the compu­
J; [228.6 − 330.2]cm for the z-coordinates of nodes C-F. Besides, the el­
tational cost of the optimization process, compared to a similar
ements were assumed to be perfectly brittle. Allowable stresses of bars
framework based on the β-unzipping method. The lower CPU time for
within the same group were assumed to be completely correlated (ρ→1),
the proposed framework can be explained by the fact that the null space
while the correlation coefficients of the allowable stresses for bars in
method only identifies the minimal failure sequences, and these are
sufficient for an accurate estimate of system reliability. Importantly,
Table 9 some failure path methods for identifying failure sequences (e.g.,
Optimal results of the 25-bar spatial truss tower.
β-unzipping method, branch and bound method) are strongly
Design Variables Best Solution parameter-dependent and the identified sequences may not be the most
A1 (cm2) 0.90 critical. On the other hand, the method employed in the proposed
A2 (cm2) 2.69 framework does not involve any parameter setting.
A3 (cm2) 3.65 From the second example, we conclude that the proposed framework
A4 (cm2) 0.73
can address problems with both brittle and ductile post-failure behav­
A5 (cm2) 0.68
A6 (cm2) 0.59 iors. The results presented for the 21-bar truss further demonstrate that
A7 (cm2) 3.17 the computational cost of optimization using the proposed framework
A8 (cm2) 5.86 can be reduced even further by disregarding the failure sequences with
xC (cm) 102.86
xG (cm) 171.97
yC (cm) 155.28 Table 10
yG (cm) 265.23 Computational costs for identifying minimal cut sets and assessing system reli­
zC (cm) 273.01 ability of optimal solution for the 25-bar truss problem.
Best weight (kg) 173.69
Steps CPU time
Worst weight (kg) 219.79
Average weight (kg) 185.82 Minimal cut sets search 46.20 s
Weight CV (%) 8.35 System reliability analysis (A + B) 101.40 s
Average CPU time (h) 8.20 Component reliability analyses (A) 60.84 s
βsys,MCS 3.06 Computation of probabilities of failure sequences (B) 40.56 s

12
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

many elements. This also suggests that, in an optimal design considering [9] Burton SA, Hajela P. Reliability-based shape optimization of truss structures. 19th
AIAA Applied Aerodynamics Conference, American Institute of Aeronautics and
progressive collapse, special attention should be given to the shorter
Astronautics Inc.; 2001. [Link]
failure sequences. The applicability of the proposed framework was also [10] Torii AJ, Lopez RH, Biondini F. An approach to reliability-based shape and
demonstrated in a 25-bar spatial truss problem involving over ten topology optimization of truss structures. Eng Optim 2012;44:37–53. [Link]
million failure sequences. org/10.1080/0305215X.2011.558578.
[11] Chun J, Paulino GH, Song J. Reliability-based topology optimization by ground
Since the focus of the proposed framework is computation of corre­ structure method employing a discrete filtering technique. Struct Multidiscip
lations between element failure events and between failure sequences, Optim 2019;60:1035–58. [Link]
only linear limit state functions of Gaussian random variables are [12] Jalalpour M, Guest JK, Igusa T. Reliability-based topology optimization of trusses
with stochastic stiffness. Struct Saf 2013;43:41–9. [Link]
considered in this study. As a result, element failure probabilities were strusafe.2013.02.003.
computed using Cornell reliability index and system reliability was [13] Beck AT, Gomes WJDS. A comparison of deterministic, reliability-based and risk-
computed using PNET method. These limitations should be addressed in based structural optimization under uncertainty. Probab Eng Mech 2012;28:18–29.
[Link]
future work by employing more advanced techniques for assessing [14] Beck AT. Optimal design of redundant structural systems: fundamentals. Eng Struct
element failure probabilities, such as the First-Order Reliability Method 2020;219:110542. [Link]
(FORM). Future studies should also explore the advantages of the null [15] Beck AT, Ribeiro L da R, Valdebenito M. Risk-based cost-benefit analysis of frame
structures considering progressive collapse under column removal scenarios. Eng
space method and combine it with other techniques for assessing system Struct 2020;225. doi: 10.1016/[Link].2020.111295.
reliability. The data provided by the null space method can also be used [16] Beck AT, da Rosa RL, Valdebenito M, Jensen H. Risk-Based Design of Regular Plane
to obtain surrogate models, which can be employed to reduce the Frames Subject to Damage by Abnormal Events: A Conceptual Study. J Struct Eng
2022;148. [Link]
computational cost of S-RBSSO.
[17] Ribeiro L da R, Kroetz HM, Parisi F, Beck AT. Optimal risk-based design of
reinforced concrete beams against progressive collapse. Eng Struct 2024;300:
CRediT authorship contribution statement 117158. doi: 10.1016/[Link].2023.117158.
[18] Enevoldsen I, Sorensen JD. Reliability-based optimization of series systems of
parallel systems. J Struct Eng 1993;119:1069–84.
Lucas A. Rodrigues da Silva: Conceptualization, Formal analysis, [19] Schuëller GI, Jensen HA. Computational methods in optimization considering
Writing – review & editing. André J. Torii: Formal analysis, Writing - uncertainties - An overview. Comput Methods Appl Mech Eng 2008;198:2–13.
[Link]
review. André T. Beck: Conceptualization, Formal analysis, Supervi­ [20] Song J, Kang WH. System reliability and sensitivity under statistical dependence by
sion, Writing – review & editing. matrix-based system reliability method. Struct Saf 2009;31:148–56. [Link]
org/10.1016/[Link].2008.06.012.
[21] Nguyen TH, Song J, Paulino GH. Single-loop system reliability-based design
optimization using matrix-based system reliability method: Theory and
Declaration of competing interest applications. Journal of Mechanical Design, Transactions of the ASME 2010;132:
0110051–01100511. [Link]
The authors declare that they have no known competing financial [22] Adam JM, Parisi F, Sagaseta J, Lu X. Research and practice on progressive collapse
and robustness of building structures in the 21st century. Eng Struct 2018;173:
interests or personal relationships that could have appeared to influence
122–49. [Link]
the work reported in this paper. [23] Kiakojouri F, De Biagi V, Chiaia B, Sheidaii MR. Strengthening and retrofitting
techniques to mitigate progressive collapse: A critical review and future research
Data availability agenda. Eng Struct 2022;262. [Link]
[24] Beck AT, Stewart MG. Risk-based cost-benefit analysis of structural strengthening
to mitigate disproportionate collapse of buildings under abnormal blast loading.
Data will be made available on request. Structures 2023;57:105103. [Link]
[25] Prasad Varma Thampan CK, Krishnamoorthy CS. System reliability-based
configuration optimization of trusses. J Struct Eng 2001;127:947–56.
Acknowledgements [26] Cid Bengoa C, Baldomir A, Hernández S, Romera L. Multi-model reliability-based
design optimization of structures considering the intact configuration and several
partial collapses. Struct Multidiscip Optim 2018;57:977–94. [Link]
Funding of this research project by Brazilian agencies CNPq (Na­
10.1007/s00158-017-1789-y.
tional Council for Scientific and Technological Development, grant n. [27] Beck AT, Tessari RK, Kroetz HM. System reliability-based design optimization and
309107/2020-2), FAPESP (São Paulo State Foundation for Research, risk-based optimization: a benchmark example considering progressive collapse.
Eng Optim 2019;51:1000–12. [Link]
grant n. 2022/08009-6) and joint FAPESP-ANID (São Paulo State
[28] Rodrigues da Silva LA, Torii AJ, Beck AT. Hyperstatic and redundancy thresholds
Foundation for Research – Chilean National Agency for Research and in truss topology optimization considering progressive collapse due to aleatory and
Development, grant n. 2019/13080-9. Valuable comments by the epistemic uncertainties. Probabilistic Engineering Mechanics 2023;71:103384. doi:
anonymous reviewers are also cheerfully acknowledged. 10.1016/[Link].2022.103384.
[29] Song J, Kang W-H, Lee Y-J, Chun J. Structural System Reliability: Overview of
Theories and Applications to Optimization. ASCE ASME J Risk Uncertain Eng Syst
References A Civ Eng 2021;7. [Link]
[30] Felipe TRC, Haach VG, Beck AT. Systematic Reliability-Based Approach to
Progressive Collapse. ASCE ASME J Risk Uncertain Eng Syst A. Civ Eng 2018;4.
[1] Ohsaki M. Optimization of finite dimensional structures. Boca Raton: CRC Press;
[Link]
2010. doi: 10.1201/ebk1439820032.
[31] Felipe TRC, Beck AT. Dynamic analysis of failure paths of truss structures:
[2] Zegard T, Paulino GH. GRAND3 — Ground structure based topology optimization
Benchmark examples including material degradation. Mech Syst Signal Process
for arbitrary 3D domains using MATLAB. Struct Multidiscip Optim 2015;52:
2021:158. [Link]
1161–84. [Link]
[32] Moses F. Structural system reliability and optimization. Comput Struct 1977;7:
[3] Mortazavi A, Toğan V. Simultaneous size, shape, and topology optimization of truss
283–90.
structures using integrated particle swarm optimizer. Struct Multidiscip Optim
[33] Moses F. System reliability developments in structural engineering. Struct Saf
2016. [Link]
1982;1:3–13.
[4] Carvalho JPG, Lemonge ACC, Carvalho ÉCR, Hallak PH, Bernardino HS. Truss
[34] Thoft-Cristensen P, Murotsu Y. Application of structural systems reliability theory.
optimization with multiple frequency constraints and automatic member grouping.
Springer Science & Business Media; 1986.
Struct Multidiscip Optim 2018;57:547–77. [Link]
[35] Lu N, Liu Y, Beer M. System reliability evaluation of in-service cable-stayed bridges
1761-x.
subjected to cable degradation. Struct Infrastruct Eng 2018;14:1486–98. https://
[5] Kozłowski G, Sokół T. Enhanced growth method for topology and geometry
[Link]/10.1080/15732479.2018.1459743.
optimization of truss structures. Struct Multidiscip Optim 2022;65. [Link]
[36] Murotsu Y, Okada H, Taguchi K, Grimmelt M, Yonezawa M. Automatic generation
org/10.1007/s00158-022-03317-7.
of stochastically dominant failure modes of frame structures. Struct Saf 1984;2:
[6] Ditlevsen O, Madsen HO. Structural Reliability Methods. 3rd ed. Chichester: John
17–25.
Wiley & Sons Ltd; 2007.
[37] Lee Y-J, Song J. Risk Analysis of Fatigue-Induced Sequential Failures by Branch-
[7] Melchers RE, Beck AT. Structural Reliability Analysis and Prediction. 3rd ed.
and-Bound Method Employing System Reliability Bounds. J Eng Mech 2011;137:
Wiley; 2018. doi: 10.1002/9781119266105.
807–21. [Link]
[8] Bland JA. Structural design optimization with reliability constraints using Tabu
Search. Eng Optim 1998;30:55–74. [Link]
03052159808941238.

13
L.A. Rodrigues da Silva et al. Structural Safety 108 (2024) 102448

[38] Kim DS, Ok SY, Song J, Koh HM. System reliability analysis using dominant failure [45] Deng H, Kwan ASK. Unified classification of stability of pin-jointed bar assemblies.
modes identified by selective searching technique. Reliab Eng Syst Saf 2013;119: Int J Solids Struct 2005;42:4393–413. [Link]
316–31. [Link] ijsolstr.2005.01.009.
[39] Xing H, Jiang T, Hao P. An efficient dominant failure modes search strategy and an [46] Thoft-Christensen P, Murotsu Y. Application of Structural Systems Reliability
extended sequential compounding method of system reliability analysis and Theory. Springer Berlin Heidelberg; 1986. doi: 10.1007/978-3-642-82764-8.
optimization. Comput Methods Appl Mech Eng 2021:375. [Link] [47] Ang A, Tang W. Probability Concepts in Engineering Planning and Design, vol. 2.
10.1016/[Link].2020.113637. John Wiley & Sons; 1984.
[40] Yun W, Lu Z, Zhou Y, Jiang X. AK-SYSi: an improved adaptive Kriging model for [48] Gollwitzer S, Rackwitz R. Equivalent Components in First-Order System Reliability.
system reliability analysis with multiple failure modes by a refined U learning vol. 5. 1983.
function. Struct Multidiscip Optim 2019;59:263–78. [Link] [49] Kar R, Mandal D, Mondal S, Ghoshal SP. Craziness based Particle Swarm
s00158-018-2067-3. Optimization algorithm for FIR band stop filter design. Swarm Evol Comput 2012;
[41] Guan X, Xiang Z, Bao Y, Li H. Structural dominant failure modes searching method 7:58–64. [Link]
based on deep reinforcement learning. Reliab Eng Syst Saf 2022:219. [Link] [50] Kennedy J, Eberhart R. Particle Swarm Optimisation. Proceedings of ICNN’95 -
org/10.1016/[Link].2021.108258. International Conference on Neural Networks, 1995. [Link]
[42] Wei X, Deng H. Determining failure modes devoid of redundant elements for pin- ICNN.1995.488968.
bar structures using null space of equilibrium matrix. Struct Saf 2022:95. https:// [51] Simon D. Evolutionary Optimization Algorithms. John Wiley & Sons; 2013.
[Link]/10.1016/[Link].2021.102175. [52] Dang KD, Nguyen-Van S, Thai S, Lee S, Luong VH, Lieu QX. A single step
[43] Pellegrino S, Calladine CR. Matrix analysis of statically and kinematically optimization method for topology, size and shape of trusses using hybrid
indeterminate frameworks. vol. 22. 1986. differential evolution and symbiotic organisms search. Comput Struct 2022:270.
[44] Pellegrino S. Structural computations with the singular value decomposition of the [Link]
equilibrium matrix. vol. 30. 1993.

14