ORIGINAL ARTICLE

Buckling Analysis of Cylindrical Shells using Stochastic

Finite Element Method with Random Geometric
Imperfections
Zheng Li1, Hartmut Pasternak2, Karsten Geißler1

Correspondence Abstract

Dr.-Ing. Zheng Li Thin-walled cylindrical shells often exhibit buckling failure and the experimental buckling load is usually
Technical University of Berlin lower than calculation results from classical theory and simulation without geometrical imperfection.
Department of Steel Structure Besides, test results with carefully conducted similar specimens still have substantial scatter due to
Gustav-Meyer-Allee 25
imperfection sensitivity. The nonlinear analysis with FEM can obtain a high-precision result comparing
13355 Berlin
Email: [email protected] the experiment. However, this is almost impossible in practical engineering. The initial geometric
imperfections of cylindrical shells are complex and random properties. Theoretically, these geometric
imperfections can be described using a random field or a Fourier series representation. After that a
random sample of the buckling bearing capacity of the cylindrical shell can be obtained using the
stochastic finite element method within the subsequent analysis. This paper presents the investigation
of buckling analysis of cylindrical shells under axial compression considering the randomness of initial
geometric imperfections, and the stochastic FEM is employed to calculate the statistical sample of the
shell-buckling load based on a series of random fields and Fourier series representation of geometric
imperfections.

Keywords

Cylindrical shell, Buckling, Stochastic Finite Element Method, Random field, Geometric imperfection,
Fourier series representation, Maximum Entropy fitting method

1 Introduction However, the dispersion of experimental results and non-negligible

errors between the theoretical analysis and test results still exist. The
Thin-walled structures like cylindrical shells are widely used in civil main reason for this is that it is difficult to obtain the exact geometric
engineering and aerospace shell structures. It is well known that this dimensions of the practical shell structure and accurate modelling.
type of structure tends to buckle under axial compression, whereby this Nowadays, the three-dimensional (3D) scanning technique has been
buckling occurs without prior notice. The buckling phenomenon is used to directly obtain actual geometric dimensions of shell specimens.
mostly accompanied by a sudden and severe drop in the bearing capacity
of the cylindrical shell. In recent decades, lot of researchers have
investigated shell stability from different viewpoints. However, the
phenomenon of incomplete agreement of buckling load of cylindrical
shells between theory and experiment under axial compression is a long-
standing problem in shell structures [1]. Additionally, the buckling load
results obtained from tests has a large dispersion. With the
establishment of the nonlinear Donnell-Mushtari-Vlasov theory for large
deflections of isotropic thin shells [2], pre-buckling consistency theory [3]
and the initial post-buckling theory [4], combined with the nonlinear
analysis based on finite element method (FEM), stability issues for
cylindrical shells have been addressed at the theoretical level.

1. Institute of Civil Engineering, Technical University of Berlin,

Berlin, Germany
2. Institute of Civil Engineering, Brandenburg University of Figure 1 Geometryy of shell specimen via 3D scanning, Source: [13]
Technology, Cottbus, Germany

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and
reproduction in any medium, provided the original work is properly cited.
Funding Statement: Open Access funding enabled and organized by Projekt DEAL.
WOA Institution: TECHNISCHE UNIVERSITAET BERLIN
Consortia Name: Projekt DEAL

© 2022 The Authors. Published by Ernst & Sohn GmbH. · ce/papers 5 (2022), No. 4
https://doi.org/10.1002/cepa.1803 wileyonlinelibrary.com/journal/cepa 653
 25097075, 2022, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cepa.1803 by CAPES, Wiley Online Library on [02/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
654 |
The lower bound curve is derived from a large number of scattered test kernel 𝐶𝐻𝐻 , which are the solution to the homogeneous Fredholm
data [5,6], which form the basis of design curves in current standards. At integral equation of second kind [88]. In general, the covariance matrix
present, the "worst" geometric imperfection shape is commonly utilised of random field can be approximated by the autocorrelation function
to calculate the shell buckling load using numerical analyses and 𝜌𝐻𝐻 (𝑿1 , 𝑿2 ), i.e.:
formulae. According to EN 1993-1-6 [7] and EN 1999-1-5 [8] standards,
structural engineers need to introduce information on geometric 𝐶𝐻𝐻 (𝑿1 , 𝑿2 ) = 𝜎(𝑿1 )𝜎(𝑿2 )𝜌𝐻𝐻 (𝑿1 , 𝑿2 ) (2)
imperfections into calculations to ensure the safety of shell structures.
However, these design standards continue to have drawbacks; for where 𝜎 means the standard deviation of random field. According to
example, the lack of determination of different types of imperfections, reference [12], isotropic exponential, squared exponential and sine
functions can be employed as autocorrelation coefficient function for
including misalignment, out-of-roundness, and radial deviation [9].
numerical analysis in most multidimensional homogenous and isotropic
Theoretically, to meet structural reliability requirements and improve
random fields. In this paper, the isotropic exponential function as shown
rationality and the accuracy of shell structures, the design of cylindrical
in Eq. (3) are used for the random field of geometric imperfections.
shells should be based on statistical distribution values as opposed to
lower bound values obtained by the "worst" imperfection [10]. |𝑿1 −𝑿2 |
𝜌𝐻𝐻,1 (𝑿1 , 𝑿2 ) = exp (− ) (3)
𝐿𝑐
Many attempts to resolve this conundrum through deterministic theory,
experimentation, and probabilistic analysis have remained and Since discrete random field is approximated by finite truncation terms,
numerous of them are not completely successful. Hence, this paper will this means that this approximate calculation ignores the high order
focus on the analysis from the perspective of initial geometric terms in the tail and the random field will inevitably lose some
imperfection, and use two methods to describe the randomness and uncertainty information. Therefore, the accuracy of the discrete random
spatial correlation of geometric imperfections. Then, the statistical fields need to be evaluated. In general, the difference between the
samples of nonlinear buckling load of cylinder will be obtained based on variance of the discrete random field and the target variance, as shown
obtained using the non-invasive stochastic finite element method in Eq. (4), can be used as a measure of the accuracy of the truncated
(SFEM). After that the optimal probability function of cylinder bearing series.
capacity will be calculated based on the maximum Entropy fitting
̂ (𝑿,𝜃)]
𝑉𝑎𝑟[𝐻(𝑿,𝜃)]−𝑉𝑎𝑟[𝐻
method. 𝜖𝜎 2 (𝑿) = (4)
𝑉𝑎𝑟[𝐻(𝑿,𝜃)]

2 Representation of random geometric imperfection Figure 2 displayed the relationship between truncated terms 𝑀 and the
standard deviation of discrete random field as well as the error measure
A stochastic approach considering geometric imperfect for shell for the discretization 𝜖𝜎 2 (𝑿). It is obvious that the discretization error
structures goes already back to the 1950s. One important step in the decreases monotonically with the truncation order and the rate of
analysis of the influence of the imperfection properties randomness on decrease is constantly decreasing. This means that the increased
bearing capacity of shell structure is the representation of the input of computational effort after reaching a certain truncated terms 𝑀 cannot
the imperfection uncertainty. Recently, there are two main ways to significantly improve the accuracy of discrete random fields. In this paper,
represent geometric imperfection which contain random properties, the truncated terms 𝑀 is predefined as 2000 considering the balance
which treats the Fourier coefficients of the measured imperfection as between accuracy and consumed computing resources.
random variables, namely random field and Fourier series
representation with correlated random coefficients.

2.1 Random field and discretization

A random field can be seen as an extension of a random process in the

spatial domain, which consists of an infinite number of random variables
at points in the continuous domain. In general, the random field is a
useful mathematical model to accurately describe the uncertain
variables. With the continuous development of computer technology
and finite element theory, the stability analysis of cylindrical shell is now
commonly performed using commercial finite element software. Hence,
the random field need to be discretized using a finite number of random
variables and the purpose is to approximate the random field by the
finite set of random variables. Thus, obtaining an approximate discrete
Figure 2 Relationship between discretization and truncated terms.
random field with accurate statistical properties is the premise of Source: Z. Li
studying the bearing capacity of shell structures with uncertain
geometric imperfections. The discretization of random field can be Besides, the accuracy of discrete random field is also affected by the type
simulated by using the Karhunen-Loève (K-L) series expansion [11] with of statistical distribution of the random field, especially random fields
truncated expansion after a finite M terms. with high skewness. Although an arbitrary unknown distribution is nearly
or approach to Gaussian distribution in most cases, because of the
̂ (𝑿, 𝜃) = 𝜇(𝑿) + ∑𝑀 ̂ central limit theorem. The small probability events in the random field
𝐻 𝑖=1 √𝜆𝑖 ξi (θ)𝜑
̂ 𝑖 (𝑿) . (1)
of geometric imperfection, which are sensitive to the tail regions of
probability distributions, can be sensitive to the buckling bearing
where 𝐻 ̂ (𝑿, 𝜃) means a truncated random field in probability space; 𝑿
capacity of shell structures. Theoretically, the non-Gaussian random
is the position vector defined over the space domain and 𝜃 is primitive
field with small skewness can be transformed using a translations
randomness; 𝜇(𝑿) is the mean function of the random field; 𝜉𝑖 (𝜃)
processes or some mapping method based on the preceding
means a statistically uncorrelated random variable with zero mean; 𝜆𝑖
discretization method of random field.
and 𝜑𝑖 (𝑿) are the eigenvalue and the eigenfunction of the covariance
 25097075, 2022, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cepa.1803 by CAPES, Wiley Online Library on [02/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
| 655

The establishment of a random field requires statistical distribution interpolation method [15] was employed to modify the imperfection
properties and spatial correlations, which can be described by the values in the connection position. Figure 3 presents the a) original and b)
probability density function of random field variables and correlation modified imperfection random field.
lengths. The correlation length determines the decay of the mutual
influence of two different points in random field and is a measure for a 2.2 Fourier series representation
random field which are required to describe the random field with a
given correlation property. Therefore, a correlation length equal to Numerous studies shown that the actual geometry of the cylinder along
infinity describes a completely correlated random field and a correlation the circumferential direction usually appears to be periodic, so the
length tending towards zero describes an uncorrelated random field, Fourier series method [16] can be used to analyze the geometric
which means that all random variables representing the random field are imperfection of shell structures. Usually, the geometric imperfections
fully in dependent. In general, the radial deviations comparing perfect consist of a series of axial and circumferential waves based on Fourier
cylinder can be defined as geometric imperfections in cylindrical shell series with a large number of sine and cosine functions. The radial
structures. According to the references [13], the measured geometric deviation 𝑤(ℎ, 𝜃) ) of imperfection can be defined as a Fourier
imperfections are composed of a non-uniform random field. This implies representation with the half wave cosine approach, as shown in Eq. (7).
that the correlation structure of the geometric imperfections can be
𝑘 𝑁 𝑙𝑁 𝑘𝜋ℎ 𝑙𝜃 𝑙𝜃
decomposed into two one-dimensional issues along the axial and 𝑤(ℎ, 𝜃) = 𝑡 ∗ ∑𝑘=0 ∑𝑙=0 𝑐𝑜𝑠 ( ) ∗ (𝐴𝑘𝑙 𝑐𝑜𝑠 ( ) + 𝐵𝑘𝑙 𝑠𝑖𝑛 ( )) (7)
𝐻 𝑅 𝑅
circumferential direction and the correlation structure can be described
with two separate correlation lengths. According to the fact that the where ℎ and 𝜃 are the coordinates in the axial directions and
defects in the circumferential direction of the cylindrical shell are circumferential phase shift angle, respectively. 𝐻 and 𝑅 mean the height
dominated by low frequencies, Schenk et. al. [14] believe that the and radius of the ideal circular cylindrical shell. 𝐴𝑘𝑙 and 𝐵𝑘𝑙 are the
correlation lengths 𝐿𝑐,𝑎𝑥𝑖𝑎𝑙 and 𝐿𝑐,𝑐𝑖𝑟𝑐𝑢𝑚 can be approximated using the coefficients for the Fourier series of the half-wave cosine approach.
following equations.
The coefficients 𝐴𝑘𝑙 and 𝐵𝑘𝑙 can be regarded as random parameters
𝐿𝑐,𝑎𝑥𝑖𝑎𝑙 = 0.6 ∙ 𝐻 (5) within the probabilistic analyses. However, these parameters are
strongly correlated. Therefore, it is necessary to construct a coefficient
𝐿𝑐,𝑐𝑖𝑟𝑐𝑢𝑚 = 0.06 ∙ 2𝜋𝑟 (6) vector and map the corresponding correlation. According to Eq. (8), the
random vector 𝑪 is composed of the Fourier coefficients 𝐴𝑘𝑙 and 𝐵𝑘𝑙 .

𝑪 = [𝐴𝑘𝑙 , 𝐵𝑘𝑙 ] (8)

In general, the relation between a correlated random vector 𝑪 and a

random vector 𝝃 with mean value 0 and variance 1 without dependence
can be described by the Mahalanobis transformation [17] given by Eq.
(9).

1
𝑪 = 𝜮2 ∙ 𝝃 + 𝜧 (9)

where 𝜮 is covariance matrix and 𝜧 means mean value vector 𝑪. Since

the covariance matrix is a symmetric matrix, the roots of matrix 𝜮 can be
determined by the singular value decomposition, as shown in Eq. (10).

1 1
𝜮2 = 𝑼𝜦2 𝑽∗ (10)

where 𝑼 and 𝑽 are the unitary matrix with eigenvectors of 𝜮 and 𝜦 is a

diagonal matrix with the eigenvalues of 𝜮. The idea of transforming with
1
matrix 𝜮2 is similar to performing an inverse principal component
analysis [18]. From the perspective of matrix analysis, the uncorrelated
random vector is first rotated according to matrix 𝑽∗ . Since all variables
in the original vector 𝝃 are uncorrelated, this rotation will not cause
changes of random vector. In some reference, this operation is directly
omitted [19]. In fact, the Eq. (9) becomes the K-L series expansion. The
1
main transformation is to stretch uncorrelated random vectors by 𝜦2 in
all orthogonal directions, whereby the scale of stretch equal to the
standard deviation of each random variables. Then the random vector
will be rotated according to unitary matrix 𝑼 based on the inverse
Figure 3 Samples of random field for geometric Imperfections a) principal components.
original and b) modified imperfection. Source: Z. Li
Therefore, under the premise of known covariance matrix and mean
Furthermore, the discrete random fields are generally bounded， while
vector of coefficient vector for the Fourier series, coefficient vector
a random field of cylindrical shell is unbounded in the circumferential
samples can be randomly generated using this transformation based on
direction. Therefore, the connection between the head and tail of the
the primitive random vector 𝝃 and mean value vector of Fourier
discrete random field in the circumferential direction needs to be
coefficient 𝜧. The covariance matrix can be constructed using a similar
corrected to make it transition smoothly and reduce singularities in
method to those applied in random field theory based on the correlation
numerical calculations. In this paper, the inverse distance weighted
matrix 𝜌(𝑐𝑖 , 𝑐𝑗 ).
 25097075, 2022, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cepa.1803 by CAPES, Wiley Online Library on [02/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
656 |
𝐶𝑜𝑣(𝑐1 , 𝑐1 ) ⋯ 𝐶𝑜𝑣(𝑐𝑛 , 𝑐1 ) need to be analysed. Therefore, the geometric imperfections with spatial
𝜮=( ⋮ ⋱ ⋮ ) (11) correlation can be described by constructing a covariance matrix 𝜮
𝐶𝑜𝑣(𝑐1 , 𝑐𝑛 ) ⋯ 𝐶𝑜𝑣(𝑐𝑛 , 𝑐𝑛 ) based on coefficient of variation 𝑐𝑣 and the correlation length of Fourier
coefficient 𝐿 as well as mean value vector 𝜧 of Fourier coefficients.
𝐶𝑜𝑣(𝑐𝑖 , 𝑐𝑗 ) = 𝜎𝑖 𝜎𝑗 ∙ 𝜌(𝑐𝑖 , 𝑐𝑗 ) =𝑐𝑣 2 ∙ 𝜇𝑐𝑖 𝜇𝑐𝑗 ∙ 𝜌(𝑐𝑖 , 𝑐𝑗 ) (12)
Figure 4 displays the geometric imperfection samples generated using
2
Fourier series representation with different input parameters.
(𝑋𝑐𝑖 −𝑋𝑐𝑗 )
𝜌(𝑐𝑖 , 𝑐𝑗 ) = 𝑒𝑥𝑝 (− ) (13)
𝐿 3 Probabilistic Buckling analysis of cylindrical shells and
corresponding statistical distribution
where 𝑐𝑣 means 𝐿 the coefficient of variation and dimensionless
correlation length of the Fourier series coefficients. 𝑋𝑐𝑖 is the position 3.1 Numerical calculations of buckling load
vector for coefficient 𝑐𝑖 and 𝐿 can be seen as the correlation length of
In this paper, numerical method with arc-length approach based on
Fourier coefficient 𝐴𝑘𝑙 and 𝐵𝑘𝑙 .
Geometrically and Materially Nonlinear Analysis with Imperfections
According to experimental outcomes, the geometric imperfections of included (GMNIA) is employed used to obtain the ultimate bearing
cylindrical shells are basically dominated by a low frequency. Hence, the capacity of cylindrical shells with random geometric imperfection based
distribution of the Fourier coefficients is regular and can be fitted by a on random field and Fourier series representation. The non-intrusive
two-dimensional surface according to Eq. (15). Monte Carlo random finite element has high accuracy and robustness
for the strong geometric nonlinear analysis, so it is very suitable for
𝜇𝑐1 obtaining statistical samples of buckling bearing capacity of cylindrical
𝜇𝐴
𝜧 = [ ⋮ ] = [𝜇 𝑘𝑙 ] (14) shells. In order to make the numerical analysis and the random samples
𝐵𝑘𝑙
𝜇𝑐𝑛 of geometric imperfection to interact with each other, the meshing rule
of discrete random field or Fourier series representation and FE model
(𝑘−𝐾)𝑛 (𝑙−𝐿)𝑛
𝜇𝐴𝑘𝑙 = 𝐴 ∙ exp ( ) exp ( ) (15) are completely separated. The coordinates of the nodes in the finite
𝑁 𝑁
element model will be modified through a somewhat mapping
where 𝐴 means the largest average of all coefficients. 𝐾 and 𝐿 mean the relationship, such as inverse distance weighted interpolation method
positions of the largest coefficients in the axial and circumferential [33], to reflect the randomly generated geometric imperfection. In this
directions, respectively. 𝑁 and 𝑛 are the shape coefficients of the 2D paper, the generated geometric imperfection files with random
surface function, which can be determined by experimental properties and finite element model node information are firstly read,
investigation. and then the mapping relationship between each numerical point and
the associated imperfection value points is established according to Eq.
(1) in reference [15].

The numerical calculation of buckling bearing capacity is carried out

using the Finite Element code ABAQUS®. The shell element S4R with 4
nodes and reduced integration points is used and all element size is
defined as 10 mm. The all degrees of freedom for the node on the upper
and lower edges of the numerical model are coupled with two reference
points at the center of the upper and lower edges. This both reference
points are defined as hinged. The elastic modulus and yield stress of the
steel are defined as 210 GPa and 355 MPa, respectively.

Figure 5 Sample size of random field versus statistical moment of bearing

capacity. Source: Z. Li

Figure 5 and 6 present the relationship of sample size of random field

Figure 4 Samples of Fourier series representation with various and Fourier series representation as well as the first four moments of
parameters. Source: Z. Li the cylinder bearing capacity, namely mean value, variance, skewness
and kurtosis, respectively. It is obviously that the first four moments
Generally, for the considered set of shells, the Fourier series describes show large fluctuations when the sample size is relative small and the
the imperfection patterns sufficiently accurate for 𝑁𝑘 =10 and 𝑁𝑙 =20 and fluctuation decreases as the sample size increases. When the sample size
hence, 2 ∙ (𝑁𝑘 + 1) ∙ (𝑁𝑙 + 1) = 462 ∙ coefficients for 𝐴𝑘𝑙 and 𝐵𝑘𝑙 are reaches 200, a relatively stable first four moments can be obtained,
 25097075, 2022, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cepa.1803 by CAPES, Wiley Online Library on [02/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
| 657

which is very important for the subsequent statistical distribution

function of bearing capacity and reliability analysis of shell structures.

Figure 6 Sample size of Fourier series representation versus statistical

moment of bearing capacity. Source: Z. Li

3.2 Maximum Entropy fitting method

It is well known that the reliability analysis of structures is based on the

statistical distribution functions of the action and resistance. In general,
the Probability Density Function (PDF) of structural resistance can be
approximated based on traditional fitting techniques. Usually, it is
Figure 7 Optimal fitted PDF of nonlinear buckling load a) random field
assumed that the sample belongs to some known random distribution
model b) Fourier series representation model. Source: Z. Li
function and then the null hypothesis theory is used to judge whether
the data conform to the assumed distribution. If the distribution is
checked for conformity, then the corresponding parameters of its
distribution can be estimated. However, the accuracy of the results
depends on the selected probability model and the sufficiency of the
sample. Overall, the correct distribution of random variables and
accurate corresponding statistical parameter are a prerequisite for
reliability analysis. In some cases, where the sample size is small, the
sample may accept some different distributions at the same time, and
the statistical parameters of the random variable may be different from
the true value.

Jaynes first proposed the principle of maximum entropy in [73]. Based

on additional constraints under known information, he pointed out that
the distribution with the smallest deviation has the largest entropy.
Therefore, the maximum entropy principle provides a way to construct
an optimal probability distribution under known information. In this
paper, the maximum entropy approach is employed to fit the optimal
probability density distribution for the buckling bearing capacity of shell
structures based on calculated 200 samples. The ideal of the maximum
entropy fitting approach is to find a probability density function 𝑓(𝑥)
with the maximum entropy 𝐻 under the constraints [21]. This is
formulated as a constrained optimization problem:

Maximize H = − ∫ f(x)lnf(x)dx
{ (16)
subject to E{ϕn (x)} = ∫ ϕn (x)f(x)dx

where 𝜙0 (𝑥) = 1 and 𝜙𝑛 (𝑥) = 𝑥 𝑛 . 𝐸{𝜙𝑛 (𝑥)} is the 𝑛th moment of the
random variable. The conventional method to solve the optimization Figure 8 Influence of different parameters, a) standard deviation of
problem is to introduce the Lagrange multiplier 𝜆𝑛 . The 𝝀 = [𝜆0 ⋯ 𝜆𝑛 ] is random field ( 𝐿𝑐,𝑎𝑥𝑖𝑎𝑙 = 0.6 ∙ 𝐻, 𝐿𝑐,𝑐𝑖𝑟𝑐𝑢𝑚 = 0.06 ∙ 2𝜋𝑟 ) and b)
obtained by solving the set of 𝑛 + 1 nonlinear equations as follows: correlations length of Fourier series representation model (𝑐𝑣 = 0.8) on
PDF of cylinder bearing capacity. Source: Z. Li
∫ ϕn (x) exp[− ∑N n
n=0 λn ϕn (x)]dx = E{x } (17)
the distribution. All moments mean that the value of 𝑁 tends to infinity,
Therefore, the fitted PDF of buckling bearing capacity can be defined as which is not achievable in numerical operations. Usually, the use of the
a form containing Lagrange multipliers, as following, first four moments in reliability analysis, such as fourth-order moment
reliability method, has already a fairly high accuracy. In this paper, the
𝑓(𝑥) = 𝑒𝑥𝑝[− ∑𝑁
𝑛=0 𝜆𝑛 𝜙𝑛 (𝑥)] (18) optimal function of PDF is determined based on the first four moments.

Theoretically, the collection of all the moments can uniquely determine Figure 7 show the best fitted PDFs for random field and Fourier series
 25097075, 2022, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/cepa.1803 by CAPES, Wiley Online Library on [02/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
658 |
representation model using maximum Entropy fitting method based on Delft. Amsterdam: HJ Paris; 1945. English translation. Ohio: Air
the first four statistical moments with different sample size. Obviously, Force Flight Dynamics Laboratory; 1970 [in Dutch with English
more samples at the expense of large computational resources can get summary].
more accurate PDFs. The outcomes show that when the sample size is
greater than 200, the fitted PDF basically remains unchanged. [5] ECCS, (2013) Buckling of Steel Shells European Design
Recommendations 5th Edition, Brussels.
In general, all parameters in the random field and Fourier series
representation model will affect the fitted PDF. For example, Figure 8 [6] Rotter, J.M. (2004) Buckling of cylindrical shells under axial
shows the optimal PDF obtained with different standard deviations for compression, in: J.G. Teng, J.M. Rotter (Eds.), Buckling Thin Met.
random field and various correlation lengths for Fourier series Shells, Spon, London, pp. 42–87.
representation model. It is clearly that the mean value of ultimate
bearing capacity of cylindrical shells decreases with the increase of [7] Eurocode 3 (2017) Design of steel structures – Part 1-6: Strength
random field standard deviation. Besides, there is a nonlinear and Stability of Shell Structures; German version EN 1993-1-6:2007
relationship between the dimensionless correlation distance of the + AC:2009 + A1:2017, Beuth Verlag GmbH, Berlin.
Fourier series coefficients and the average value of the cylinder
nonlinear buckling load. The largest mean value of cylinder bearing [8] Eurocode 9 (2017) Design of aluminium structures – Part 1-5: Shell
capacity is obtained when the correlation coefficient takes about 0.75. structures; German version EN 1999-1-5:2007 + AC:2009, Beuth
Moreover, the obtained PDFs of buckling load may not conform to a Verlag GmbH, Berlin.
normal distribution, although the primitive random variables
[9] Rotter, J.M. (2017) Schalenbeulen im Wandel: Mechanik,
distribution is Gaussian.
Konstruktionsbemessungen und deren Wechselbeziehung,
4 Conclusion and Outlook Stahlbau. 86 pp.315–324.

In this paper, two models, i.e., random field and Fourier series [10] Knoedel, P.; Ummenhofer, T.; Rotter, J.M. (2017) Rethinking
representation model are discussed and developed to describe the initial imperfections in tanks and silos, in: Ce/Papers EUROSTEEL 2017,
geometric imperfections with randomness and spatial correlation for Ernst & Sohn, Copenhagen, Denmark, pp. 960–969.
cylindrical shell structures. Non-Invasive stochastic finite element
[11] Spanos, P.D.; Ghanem, R. (1989) Stochastic Finite Element
method is employed to simulated the nonlinear buckling load of
Expansion for Random Media, J. Eng. Mech. 115, pp. 1035–1053.
cylindrical shells based on geometrically and materially nonlinear
analysis with imperfections included approach. The first four statistical
[12] Li, C.C.; Kiureghian, A.D (1993) Optimal Discretization of Random
moments of the numerical bearing capacity samples are calculated and
Fields, J. Eng. Mech. 119, pp. 1136–1154.
the optimal probability density function of buckling load is obtained
using maximum Entropy fitting method based on the statistical [13] Li, Z.; Pasternak, H.; Jäger-Cañás, A. (2022) Beulverhalten von
moments. The results indicate that both models are able to quantify the ringversteiften Kreiszylinderschalen unter Axialdruck - Teil 1:
statistical properties of the bearing capacity due to the randomness of Experimentelle Untersuchungen und geometrische Imperfektions-
geometric imperfections. As such, based on the proposed approach, analyse, 97, pp. 38-46.
stochastic properties can be directly considered in numerical
calculations to develop a more precise and economical methodology for [14] Schenk, C.A.; Schüeller, G.I. (2003) Buckling analysis of cylindrical
the design of steel structures. shells with random geometric imperfections, Int. J. Non. Linear.
Mech. 38 pp. 1119–1132.
However, the parameters in these models require further study to reveal
their relationship to the statistical distribution function of the bearing [15] Li, Z.; Jäger-Cañás, A.; Pasternak, H. (2021) Buckling of ring-stiffened
capacity. Especially the model parameters need to be analyzed and cylindrical shell under axial compression: Experiment and numerical
obtained based on the measured geometric imperfection with a large simulation, Thin-walled structures, 1077888.
number of experiments. Besides, the sensitivity analysis of the
parameters of the both models to the bearing capacity of cylinder needs [16] Kriegesmann, B.; Rolfes, R.; Hühne, H.; Teßmer, J.; Arbocz, J. (2010)
to be performed in the future. Probabilistic design of axially compressed composite cylinders with
geometric and loading imperfections, Int. J. Struct. Stab. Dyn. 10,
Acknowledgements pp. 623–644.
The research leading to these results has received funding from the
German Research Foundation DFG (454652824). The authors thank for [17] Härdle W, Simar L. (2007) Applied multivariate statistical analysis.
the financial support. Springer Verlag.

[18] Jolliffe I.T. (1986) Principal component analysis. New York: Springer.
References
[1] Brush, D.O.; Almroth, B.O. (1975) Buckling of bars, plates, and [19] Wagner, H.N.R.; Hühne, C.; Elishakoff, I. (2020) Probabilistic and
shells, McGraw-Hill Book Company, New York. deterministic lower-bound design benchmarks for cylindrical shells
under axial compression, Thin-Walled Structures, 146, 106451.
[2] Donnell, L.H. (1934) Stability of thin-walled tubes under torsion,
Report National Advisory Committee for Aeronautic. [20] Li, Z.; (2018) Statistical size effect in steel structure and
corresponding influence on structural reliability, Brandenburg
[3] Stein, M. (1964) The Influence of prebuckling deformations and
University of Technology Cottbus-Senftenberg, Dissertation,
stresses on the buckling of perfect cylinders, NASA TR-190.
Schriftenreihe Stahlbau, Heft 13.
[4] Koiter, W. (1945) On the stability of elastic equilibrium. Ph.D. thesis,