Accepted Manuscript

The Influence of Manufacturing Variances on the Progressive Failure of Fila-

ment Wound Cylindrical Pressure Vessels

Brian Ellul, Duncan Camilleri

PII: S0263-8223(15)00605-4
DOI: http://dx.doi.org/10.1016/j.compstruct.2015.07.059
Reference: COST 6640

To appear in: Composite Structures

Please cite this article as: Ellul, B., Camilleri, D., The Influence of Manufacturing Variances on the Progressive
Failure of Filament Wound Cylindrical Pressure Vessels, Composite Structures (2015), doi: http://dx.doi.org/
10.1016/j.compstruct.2015.07.059

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
The Influence of Manufacturing Variances on the
Progressive Failure of Filament Wound Cylindrical
Pressure Vessels

Brian Ellul1 & Duncan Camilleri

Department of Mechanical Engineering, Faculty of Engineering, University of Malta, Msida

MSD2080, Malta

Abstract
Developing modelling methodologies to characterize the post first-ply failure behaviour is an
ongoing tough challenge. In this study the results of a progressive failure algorithm based on
a sudden mode-dependant degradation methodology are presented and applied to filament
wound cylindrical pressure vessels subject to an internal pressure undergoing first ply-failure,
post failure and ultimate failure. In this case, the material properties are degraded according
to the failure mode detected from the homogenised stresses at lamina level. The numerical
models were corroborated with experimental tests results. Influences due to manufacturing
divergences from near perfect samples such as material thickness and voids are also
investigated. The results show that the laminated pressure vessels investigated are able to
sustain considerable higher pressures beyond the predicted first-ply failure of approximately
three times as much. However the progression of failure and pressure vessel response was
found to be highly dependent on manufacturing divergences. This property of fibre reinforced
composites gives rise to the importance of successfully modelling the post first-ply failure of
these novel materials to enable the engineer to carry out analyses of structural integrity
beyond their design loads and possibly exploit their advantages and increase confidence in
their design and application.

Keywords: Numerical Modelling; Progressive Ply Failure Analysis; Manufacturing

Variances; Voids and Imperfections; Filament Wound Pressure Vessels

Nomenclature
Symbol Unit Description

V % Cross-sectional area expansion at mid-span of the pipe

 ij
Poisson’s ratio in the principal coordinate system 1,2,3 shown in figure 3
(i,j=1, 2,3)

1
Corresponding author: [email protected]

1
 ° Off-axis angle

i ,  j
Pa Stresses in the material coordinate system
(i, j = 1,2,6)

cv Coefficient of variation

Ei
Pa Young’s modulus in the principal coordinate system 1,2,3 shown in figure 3
(i=1, 2,3)
Fi , Fij
Pa Material strength tensors found experimentally
(i, j = 1,…,6)
Gij
Pa Shear moduli in the principal coordinate system 1,2,3 shown in figure 3
(i,j=1, 2,3)
Hi
Hashin’s failure indices
(i=1, 2, 6)

kn Random number which is multiplied with all the material constants of layer n

S Pa Shear strength

std( ) Standard deviation of the variable in the brackets

t mm Average pipe wall thickness

t new mm New randomly generated pipe wall thickness

tCSM , t DR mm Thickness of the layer denoted by the subscript

new new
tCSM , t DR mm New randomly generated thickness of the layer denoted by the subscript

Thickness of the layer n (n = CSM, DR) reduced by respective void content

t no _ void , n mm
Vv
new
t no _ void , n mm New randomly generated tno_void,n

Average void content obtained from the material characterisation test

Vv %
specimens
Unidirectional strengths. Subscripts c and t denotes in compression and
X, Y, Z Pa
tension respectively

Abbreviation Description
CSM Chopped strand mat
DOF Degrees of freedom
DR Direct roving
FE Finite element

2
FEM Finite element method
FPF First ply failure
FRP Fibre reinforced polymers
FSDT First-order shear deformation theory
GDIS Random number having a normal distribution
LVDT Linear variable differential transformer
PFA Progressive failure analysis
RAND Random number having a uniform distribution function
SPF Second ply failure

1 Introduction
Fibre reinforced composites became widely used in various industries, from relatively simple
civil applications such as waste management piping to the state-of-the-art aerospace
applications. This wide range of application is attributed to their high customisability which
leads to other advantages like high strength to weight ratio resulting in lower material costs.
On the other hand, considering that they are tailored to cater for specific needs, the design
phase is perhaps the most crucial factor for the successful implementation of these novel
materials.

Modelling of composites proves to be a tough challenge for engineers and scientists since
they are composed out of different materials and phase constituents that interact together to
give a global response. The numerous publications on this subject, which are summarised in
several review publications and the extensive research published after the world-wide
composite failure exercise carried out in the early 2000’s [1], are a proof of the tough task. As
an example, this challenge is reflected in standards related to the design of filament wound
pipes or pressure vessels where they stipulate relatively high partial safety factors [2]. In view
of this, adequate modelling of these complex systems is necessary to exploit their advantages
and increase confidence in their design and application.

Standards such as the EN 13923:2005 [3] defines a design method for filament wound fibre
reinforced polymer pipes. The design load is based on the lowest pressure which produces the
first ply failure according to a predefined failure criterion irrespective of the failure mode and
location. A safety factor of 4 or more is applied to the first ply failure load to ultimately
stipulate the corresponding design load leading to a very conservative design. It is widely
known that composites fail progressively thus their structural performance isn’t fully
compromised after the first predicted failure. The mechanical performance beyond the first
ply failure load can be analysed using a progressive failure analysis (PFA). A PFA is an
iterative process where the plies of a laminate fail progressively. During the failure
progression, the stresses are redistributed to compensate for the loss in structural strength.
This procedure can be solved analytically for simple structures while other more robust
methods, such as the finite element method (FEM), are used for more complex problems.
Numerous publications tackling the solution to the tough challenges encountered during the
modelling of progressive failure of laminated composites can be found in [4, 5].

3
There are various factors affecting the eventual failure of a structure such as micro scale
defects in the material resulting in cracks or voids, inappropriate material selection,
unpredicted loadings or perhaps defects induced during the manufacturing or installation. All
these factors can introduce different types of defects which eventually will compromise the
integrity of the structure. From a modelling point of view, these factors need to be simplified
such that the problem can be dealt with more efficiently. This is one of the major purposes of
progressive failure analyses where the complex failure process of laminated composites is
simplified into an iterative process composed of two main parts;

1. The failure criterion which predicts when failure occurs. Depending on the type of the
criterion, it may also define the type of failure. Extensive literature reviews on failure
criteria were carried out by various authors such as [6-15].
2. The material degradation model which defines how the material properties are
degraded on the macro scale to simulate failure in the micro scale. A review on
various degradation models can be found in [16].

The uncertainty of various parameters such as material composition, winding angles and wall
thickness all contribute to the probability of failure of a composite structure [17]. The
propagation of uncertainty in design parameters was studied by [18, 19] where they presented
a multi-scale nondeterministic progressive failure model. The micro and macro domains
where connected via a set of coupling variables and the uncertainties related to these coupling
variables were obtained from the uncertainties of the material and geometry parameters.

1.1 Scope of this study

In this study a probabilistic nondeterministic progressive failure algorithm applied at the

meso-scale utilising the Tsai-Wu failure criterion [20] coupled with a sudden degradation
model based on the Hashin’s failure indices [21], is presented. The uncertainties of the
material properties obtained from the material characterisation tests are included in the model
via two different methods;

1. Variability in the material properties of the initial structure.

2. Geometric imperfections of the initial structure.

The aim of this paper is to analyse the effect of these two types of initial imperfections
applied to cylindrical filament wound laminated pipes with closed ends. An asymmetric lay-
up composed from three direct roving (DR) layers wound over a chopped strand mat (CSM)
layer [CSM/-82.7°/±54.3°] (angles are with respect to the axial direction) is analysed. The
progressive failure algorithm is implemented into the FE package ANSYS ® R15.0 [22] and
the results of the simulations are corroborated with experimental tests carried out by the same
authors [23].

This paper is sectioned in the following order. A description of the progressive failure
algorithm used including a detailed description of the main components. This is followed
with an explanation of the pipe specimens under investigation together with details of the FE
4
models used in subsequent simulations. Finally the results are presented and discussed
accordingly.

2 Progressive failure algorithm

The progressive failure algorithm employed in this study is based on an iterative method
specifically used to simulate the structural performance of composites based on the work
done on laminated plates by the authors [24]. A generic flowchart of the progressive failure
algorithm is shown in Fig. 1 (left). The algorithm begins with the setting up of the initial
problem, which usually entails the creation of the structure including a set of boundary
conditions and loadings. The initial problem is solved where the initial applied load is set
below the predicted initial failure. After solving, the structure is checked against the chosen
failure criterion for any possible failures. If failure is detected at any part of the structure, the
corresponding material properties in that particular region, lamina and direction are degraded
according to a degradation methodology. After degrading the material accordingly the
problem is solved again with the degraded material at the same loading. The structure is
checked for any failure once again and the material is degraded if more failures are detected.
Once no failures are detected, the load is incremented and the problem is solved again
iteratively. This process is repeated until the solution cannot converge which indicates the
ultimate failure of the structure.

START

Problem
setup
n = # of elements plies

Add initial
imperfections
START
Solve analysis
Problem
setup
Solution
Laminate
Solve analysis converged
No failure
?

Solution Laminate Yes

converged? No failure Calculate the elements’ END
stress/strain after
Yes Degrade establishing equilibrium
Increment
END material x = element/ply #1
load
Calculate the elements’ properties Any
stress/strain after x=n
x=x+1 failures
establishing equilibrium ? Yes Yes
?
No
No
Check element/ply x
Any failed with Failure Criterion Increment
ply/ load
No Yes
element?

element /
ply
No
failed?

Yes

Degrade material properties

Fig. 1. (left) Flowchart illustrating the main components of the PFA and their connections between them.
(Right) Flowchart of the PFA used throughout this paper.

5
A more detailed flowchart of the actual progressive failure algorithm as implemented in this
study can be seen in Fig. 1 (right). The notable difference from the generic PFA is the
inclusion of initial imperfections to the initial problem. More details on how these are
implemented can be found in the section 2.2

2.1 Problem setup

In the following section, a brief explanation of the tests carried out by Ellul et al. [23] is
given for completeness. Furthermore, details regarding the finite element model used
throughout the simulations presented in the paper are presented in section 2.1.2.

2.1.1 The experimental test specimens

The experimental tests referred to in this paper, carried out by B. Ellul et al. [23], consisted of
four filament wound pipes with end domes having an asymmetric layup [CSM/-82.7°/±54.3°]
and pressurised at a rate approximately 0.5 bar/s. Apart from other measurements such as
strain gauges, the expansion at the mid-span of the pipes during the pressurisation was
recorded via three LVDTs equally distributed around the circumference. The setup is shown
in figure 2 (right). The expansion was calculated by assuming that the deformation of the
pipes during expansion was circular therefore the expansion in the cross-sectional area could
be calculated from the three displacements recorded by the three LVDTs. These results are
used and compared to the numerical models.

2.1.2 Finite element model

The main types of elements available to model thin laminated shells are solids, layered shells
or a hybrid element between a solid and a shell. While the solid and hybrid elements are more
accurate to compute stress fields, shell elements are far more computationally efficient due to
their formulation. However, shell elements impose certain restrictions and are specific to
certain cases where they can be used, in particular it is assumed that the strain developed
through the thickness is linear. Two formulations are commonly available for shell elements
which are the classical Love-Kirchhoff [25] and the first order-shear deformation theories
[26, 27]. In this case, the first-order shear deformation theory (FSDT) was adopted which is
particularly suited for thin shells without imposing a large amount of processing
requirements. The 4-noded SHELL181 structural multi-layered elements were used to model
the pipe. This element assumes a state of plane-stress and the kinematics supports membrane
stretching. A cylinder with a diameter of 250mm and 800mm long was modelled in ANSYS
[22] which correspond to the actual dimension of the pipes under investigation. A mesh
sensitivity analysis was conducted and showed that a mesh with 40 elements in the
circumferential direction and 20 elements in the axial direction are sufficient to capture the
mechanical performance of the pipes. The overall thickness of the pipe wall was set equal to
the ones reported in [23] whilst the thickness of the direct roving layers was kept constant at
t DR  0.783mm if not otherwise noted. The FE model was held in 3D space by a set of
boundary conditions that impede rigid body motion whilst leaving the model to deform freely

6
under the action of the applied loads. The model was loaded internally under a pressure load
corresponding to the applied pressure which is varied during the progressive failure
algorithm. Furthermore, a force on the edge of the pipe acting axially outward was applied to
simulate the stress developed due to the built-in end-domes trying to rip apart the pipe during
pressurisation. Figure 2 shows the FE model in ANSYS with the loads together with the
experimental setup.

Pressure
Air vent
transducer
v
v

Pressurised water
LVDTs placed around from pump
the mid-section
DR layers Built-in
CSM layer end-dome

Fig. 2. (left) FE model in ANSYS R15.0 showing the mesh using element SHELL181 illustrating the four
layer stacking. (right) The experimental setup showing on of the pipes ready to be pressurised.

SHELL181 elements having 6 degrees of freedom (DOF) at each node, three transitional and
three rotational, were employed. The rotational DOFs at the free edges were coupled together
to simulate the stiffness offered by the end-domes at the edges.

As post-processing, the Tsai-Wu failure index was calculated after each solution for each and
every element and layer. If failure is detected, then the Hashin’s indices are calculated and
the failure mode established. The material is degraded accordingly as listed in section 2.4.
Furthermore, the average expansion at the mid-span of the pipe was recorded after each
solution to be able to obtain a plot of the expansion with respect to the applied internal
pressure.

2.2 Initial imperfections

It is widely known that imperfections arising from the manufacturing process are one of the
main sources for damage initiation. This fact shouldn’t be underestimated when modelling
composites since these novel materials are intrinsically more prone to defects. This fact is
reflected in the certain standards related to fibre reinforced composite structures such as [3],
where they stipulate relatively high safety factors during the design process. The reason
behind this is the high uncertainties related to various aspects such as manufacturing
procedures, raw material properties and their interaction. All these aspects affect the final
performance of the composite structure. On the other hand, analytical or numerical models
usually model a perfect structure thus over predicting the performance of the structure under
investigation. In view of this, imperfections for failure initiation are applied on the initial
structure before solving the first iteration and the algorithm continues as described in section
2.

7
In this study, two types of imperfections are implemented. The first imperfection
methodology, referred to as “Type 1”, involves the random modification of the layer
thickness whilst the second type, “Type 2”, involves the modification of the material stiffness
of each layer. These are explained in the subsequent sections.

2.2.1 Ply thickness imperfections method (Type 1)

With this type of imperfection, the thickness of some or all the layers for each and every
element is modified randomly and independently. Three separate cases were analysed which
are detailed hereunder.

2.2.1.a Case A

As a preliminary implementation of the algorithm, the thickness of the chopped strand mat
new
layer ( tCSM ) was varied randomly between ±2% and ±20% of the nominal CSM layer
thickness ( tCSM ) which were chosen arbitrary at the preliminary stage. The thickness of the
other DR layers was kept fixed at t DR  0.783mm each. This variation in CSM layer thickness
was applied across all the elements defining the pipe.

2.2.1.b Case B

In this case, the thickness of the CSM layer ( tCSM ) was varied according to the thickness of
the individual pipe under test. The following table list the average thicknesses ( t ) of each
pipe and their corresponding standard deviation ( std (t ) ) and coefficient of variation ( cv ).

Average Coefficient of
Standard deviation
Thickness variation
t std (t ) cv
[mm] [mm] [%]
Pipe A 3.345 0.2149 6.425%
Pipe B 3.355 0.3189 9.505%
Pipe C 3.350 0.4292 12.812%
Pipe D 3.415 0.2996 8.773%
Table 1: Average thickness of the pipe walls together with the corresponding standard deviation and
coefficient of variation.

In this case, the CSM layer thickness was varied according to two types of distributions, the
uniform distribution, similar to the one mentioned in Case A, and the normal distribution.

For the uniform distribution, the simulations were similar to those in Case A but the limits for
new
the new thickness of the CSM layers ( tCSM ) were set equal to:

8
 tCSM  (1  cv )  tCSM
new
 tCSM  cv

For the normal distribution, the new thickness of the CSM layer was obtained by first
generating a random variable having a normal probability density function with mean and
standard deviation corresponding to the average pipe thickness ( t ) and standard deviation (
std (t ) ) listed in table 1 respectively. The generated random number was set equal to the new
thickness of the pipe ( t new ). Secondly, the thickness of the direct roving layers was kept
constant at t DR  0.783mm each and subtracted from the newly generated thickness t new such
new
that the final new thickness of the CSM layer ( tCSM ) is;

new
tCSM  t new  3t DR

It is important to note that since there are no theoretical bounds on the generated random
variable the new thickness for the CSM layer could be zero or even negative. In view of this
shortcoming, a lower limit was set such that when tCSM new
 0 the new thickness is set to
new
tCSM  0.1 tCSM .

2.2.1.c Case C

Varying the thickness based on the thickness of the actual pipe means that the pipe needs to
be manufacture beforehand to obtain the variation in the thicknesses. Furthermore the void
content wasn’t considered when calculating the material properties from the material
characterisation tests. In this case, the average void content and its standard deviation of the
test specimens used to obtain the material properties are taken into account. Table 2 lists the
average void fraction and its standard deviation with respect to the layer type.

Average void content Standard deviation

Vv std(Vv)
[%] [%]
Direct roving (DR) 4.95 1.46
Chopped strand mat (CSM) 4.89 0.40
Table 2: Average void content and its standard deviation with respect to the type of layer [23].

The average void content (Vv) was used to reduce the thickness of the respective layer (tn
where n = DR, CSM) such that;

tno_void,n = (1 – Vv) × tn ( n = DR, CSM )

where tno_void,n is the thickness of the layer without the voids having material type n. This
reduction in layer thickness was applied throughout all the elements.

9
The standard deviation of the void content std(Vv)n is used to calculate the upper and lower
new
bounds between which the new random thickness of each respective layer t no _ void ,n can vary

randomly.

tno _ void ,n  (1  std (Vv ) n  tno

new
_ void ,n  t no _ void ,n  std (Vv ) n where n  DR, CSM

In contrast with the previous two cases, the thicknesses of all the layers were varied in this
case.

2.2.2 Material stiffness modification method (Type 2)

The imperfections are included by modifying the material properties of each element. For
each element and layer type, namely the direct roving and chopped strand mat, a random
number kn (n  DR, CSM ) having a uniform distribution function is generated whose limits
are calculated from the standard deviation of the void content (std( Vv )) of the material
characterization specimens respectively listed in table 2;

1  std (Vv ) DR  k DR  1  std (Vv ) DR

1  std (Vv )CSM  kCSM  1  std (Vv )CSM

The generated random number k n is multiplied with all the material constants of the
respective layer type. A new set of random numbers ( k DR and kCSM ) is generated for each
element therefore the material properties varies from element to element.

2.3 Failure criterion

Numerous failure criterion theories have been put forward by various researchers [6-15].
These theories can be categorised according to their scale of application either macro or
micro length scale.

The failure criterion used in this paper is the Tsai-Wu [20] generalised quadratic failure
criterion proposed by S. W. Tsai and E. M. Wu and perhaps it is the most popular failure
criterion used to detect failure in laminates [13]. This criterion is applied at the macro
properties of the lamina. Similar to most of the macro-mechanics based failure theories, this
criterion treats the composite as a homogenous material where stiffness is predicted
according to the average properties across the lamina with respect to the material parameters
obtained from material characterisation tests.

10
Fig. 3. Schematic of a fibre-reinforced composite showing the principal directions 1, 2 & 3 with respect to
the fibre orientations and the off-axis angle θ between the principal directions and the material’s
coordinate system (x,y,z).

Considering the unidirectional strengths denoted by Xc, Xt, Yc, Yt, Zc and Zt where the capital
letter denotes the material principal direction corresponding to directions 1, 2 & 3
respectively shown in Fig. 3 while the subscript denotes whether in compression (c) or in
tension (t), the shear strength denoted by S, the tensor strength criterion is given by:

Fi i  Fij i j  1 i,j  1,...,6 (1)

where Fi and Fij are material strength tensors found experimentally while σi and σj are the
stresses in the material coordinate system (Fig. 1). Under planes stress conditions eq. (1) can
be expressed as a polynomial in terms of the material strengths by;

F11  F2 2  F6 6  F1112  F22 22  F66 62  2F121 2  1 (2)

by applying a uniaxial tensile stress to failure in the fibre direction, eq. (2) reduces to:

F1 X t  F11 X t2  1 (3)

Similarly, by applying a uniaxial compressive stress to failure in the fibre direction, eq. (2)
reduces to:

F1 X c  F11 X c2  1 (4)

By solving equations (3) and (4) simultaneously, F1 and F11 can be expressed in terms of the
material strength parameters. Similarly, the other coefficients in eq. (2) can be found by
means of uniaxial tests with the exception of F12 . Tsai-Wu [20] proposed an off-axis uniaxial
test to evaluate F12. Considering the difficulty and expense to perform the required off-axis
uniaxial tests, Narayanaswami and Adelman [28] proposed to set F12 = 0 and neglect the term
F12 since it has minimal influence on the predicted strength [28]. Finally, the coefficients for
the Tsai-Wu criterion, eq. (2), in terms of the lamina strengths are:

11
 1 1 1
F1   , F11  
Xt Xc Xt Xc
1 1 1
F2   , F22   (5)
Yt Yc YtYc
1
F6  0 , F66 
S2

Since the Tsai-Wu failure criterion is a generalised criterion, it lacks to link the predicted
failure with the mode of failure of the lamina. In view of this drawback, the Hashin’s indices
Hi (i=1, 2, 6), proposed by Hashin [21] are used to distinguish between the modes of failure.
The highest Hi index shown in Eq. (6) determines the mode of failure.

H1  F11  F1112 H 2  F2 2  F22 22 H 6  F66 62 (6)

where H1 and H2 represents fibre and matrix failure respectively while H6 represents fibre-
matrix interface failure.

2.4 Material degradation model

A sudden material degradation model was used throughout this study. This means that as
soon as failure is detected with the Tsai-Wu criterion (eq. 2), the mode is identified according
to the Hashin index (eq. 6) and consequently the respective material stiffness are reduced to
near zero. The material properties were not reduced to zero exactly for computational
purposes otherwise the solution doesn’t converge. Table 3 lists the modifications made to the
material properties with respect to the failure mode.

Failure Mode E1 E2 E3 G12 G23 G13 ν12 ν23 ν13

Fibre 0 0 0 0 0 0 0 0 0
Matrix – 0 0 – – – 0 – 0
Fibre-Matrix – – – 0 – 0 0 – 0
Table 3. Reduction in material stiffness degradation with respect to the failure mode. Ei, Gij and νij (i,j
=1,2,3) are the Young’s moduli, shear moduli and the Poisson’s ratio respectively corresponding to the
directions illustrated in Fig. 3. (–) means that the property was unaltered.

Fibre failure is the most server failure out of the three failure modes where all the stiffness
constants are reduced to zero and occurs when H1 is the highest out of the three Hashin’s
indices listed in equation (6). When H2 is the highest, the transverse (E2) and out of plane (E3)
Young’s moduli are reduced to zero together with the Poisson’s ratios related to the fibre
direction. When failure occurs between the fibres and matrix due to shear loading recognized
through the H6 index, the shear moduli and the Poisson’s ratios related to the fibre direction
are reduced to zero while the remaining properties are unmodified. Note that, if a layer is
predicted to fail in either the matrix or interlaminar shear it can fail again until fibre failure is
detected and all stiffness properties are reduced to zero.

12
3 Results and comments
The results are categorised into four distinct sections.

1. A set of simulations where no imperfections were implemented which resembles a

classical PFA.
2. Imperfections are introduced on the thickness of the CSM layer where it is varied
randomly following a uniform distribution at two arbitrary levels i.e. ±2% and ±20%
of the nominal CSM layer thickness.
3. Imperfections are applied on the thickness of the CSM layer based on the variation of
the thickness of the pipes. These simulations were performed both using a uniform
random thickness and normal distributed random thickness.
4. Two sets of PFA simulations for each pipe were carried out, In the first instance the
thicknesses of all the layers were varied and in the other the material properties of all
the layers were varied. The variations were based on the fluctuations of the void
content of the specimens used during the material characterisation tests [23].

3.1 No imperfections (Classical PFA)

A PFA simulation without any imposed initial imperfections was carried out for each pipe
listed in table 1. This means that the thickness of the direct roving layers was fixed at
t DR  0.783mm whilst the thickness of the chopped strand mat layer was set constant at
tCSM  t  3tDR where t is the average thickness of the pipe listed in table 1. The experimental
and PFA results of the mid-span expansion ( V ) versus the applied pressure for all the pipes
are illustrated in figure 4.

13
 Mid-span expansion vs. applied pressure
1.4% 1.4%

1.2% 1.2%
Volumetric change [%]

Volumetric change [%]

1.0% 1.0%

0.8% 0.8%

0.6% 0.6%

0.4% 0.4%

0.2% Pipe A 0.2% Pipe B

0.0% 0.0%
0 5 10 15 20 25 30 35 40 45 0 5 10 15 20 25 30 35 40 45
Pressure [bar] Pressure [bar]
1.4% 1.4%

1.2% 1.2%
Volumetric change [%]

Volumetric change [%]

1.0% 1.0%

0.8% 0.8%

0.6% 0.6%

0.4% 0.4%

0.2% Pipe C 0.2% Pipe D

0.0% 0.0%
0 5 10 15 20 25 30 35 40 45 0 5 10 15 20 25 30 35 40 45
Pressure [bar] Pressure [bar]

Experimental PFA without imperfections

Fig. 4. Plots of the mid-span expansion ( V ) versus the applied pressure (bar) for the four pipes listed in
table 1. Each plot compares the experimental results [23] with the corresponding classical PFA i.e.
without any initial imperfections.

From figure 4, one can notice that the expansion of the actual experimental pipes with respect
to the applied internal pressure is linear until a particular pressure level corresponding to first
ply failure. Following this a number of step changes are observed and the relationship
between applied pressure and volume change is no longer linear. This reflects progressive
failure. Similarly in the case of the numerical models step changes are also observed however
apart from two to three significant step changes the relationship between the expansion and
the applied pressure is relatively linearly smooth suggesting that ply failure occurs throughout
the whole lamina at the same time.

First and second ply failure load results summary

Wall First Ply Failure (FPF)
Test [bar] ±0.05bar 2nd ply failure from PFA
Thickness
Specimen [bar] ±0.05bar
[mm] ±80µm
Experimental [23] Classical PFA
A 3.345 13.3 12.4 21.7
B 3.355 18.2 12.6 21.7
C 3.350 13.8 12.4 21.7
D 3.415 18.0 12.5 22.2
Table 4: List of first and second ply failure loads obtained from the PFA without imperfections (termed
“Classical PFA”) compared with the experimental results [23].

14
The first sudden drop in expansion occurred in the range of 12.4 bar and 12.7 bar for all the
pipes. This slight drop in expansion is the first ply failure load and corresponds to the matrix
failure of the three direct roving layers. Comparing these results with the experimental ones
reported in [23] (see table 4) one can notice that the FPF load from the PFA is slightly
underestimated.

Fig. 5. Plot showing the results of the mid-span expansion ( V ) versus the applied pressure (bar) for
pipe A obtained from the experimental [*ellul*] and PFA without imperfections.

The second and most pronounced expansion drop occurred between 21.7 bar and 22.2 bar
except for pipe B where the most pronounced expansion drop occurred at 32.2 bar. This
pronounced drop in expansion corresponds to the matrix failure of the chopped strand mat
layer. It is difficult to compare the second ply failure (SPF) load obtained from the PFA with
the experiment results since the actual pipes experience more gradual progressive failure.
This behaviour can be seen more clearly in figure 5 which shows the expansion with respect
to the applied pressure of pipe A.

The SPF of the actual pipe A occurred at 19.4bar whilst the SPF of the corresponding PFA
occurs at 21.7bar. Beyond the SPF, the PFA predicts a linear expansion with respect to the
applied pressure until the final failure at 43.3bar. Comparing this prediction with the
experimental result, the expansion after the SPF at 19.4bar of the actual pipe was not linear
but rather jittery which is a clear indication that failure is progressing constantly. This region
is referred to as the instable region as indicated in figure 5 where the performance of the pipe
is unpredictable due to the progression of failures. This phenomenon can be seen in the
remaining pipes which is a characteristic behaviour of failing composites [29].

15
3.2 Preliminary simulations - Varying the thickness of the CSM (Case A)

In the following numerical models, the thickness of the chopped strand mat layer was varied
randomly. This modification was opted for more accurate representation of the instable
region that otherwise wasn’t captured with the classical PFA. At this preliminary stage, the
thickness of the CSM layer of all the elements of pipe D was varied randomly with a uniform
distribution. As detailed in section 2.2.1, two simulations where carried out where the new
thickness of the CSM layer was varied between the intervals ±2% and ±20% of the nominal
thickness. The results of these simulations of pipe D are illustrated in figure 6.

Applied Pressure [bar] vs. Vol. Change [%]

1.2%

1.0%
Volumetric change [%]

0.8%

0.6%

0.4%

0.2%

0.0%
0 5 10 15 20 25 30 35 40 45
Applied pressure [bar]
Experimental No Imperfections 2% of CSM tks 20% of CSM tks

Fig. 6. Plot showing the results of the mid-span expansion ( V ) versus the applied pressure (bar) of the
actual pipe D with respect to the PFA without any imperfections and the PFA with two levels of
imperfections (±2% and ±20%) applied on the thickness of the CSM layer.

The FPF corresponds to the matrix failure of the three direct roving layers whilst the SPF
corresponds to the matrix failure of the CSM layer which is similar to the simulations without
imperfections. From figure 5, one can notice that the FPF load of the simulation where the
thickness of the CSM layer was varied between ±20% of tCSM was lowered to 11.1 bar where
the FPF load of the ±2% simulation remained approximately equal to the simulation without
imperfections. Details of the FPF and SPF loads are listed in table 5.

FPF load SPF load

Pipe D
[bar] [bar]
Experimental 18.0 20.5
No imperfections 12.5 22.2
±2% of tCSM 12.4 22.0
±20% of tCSM 11.1 20.4
Table 5: List of first and second ply failure loads for pipe D.

16
The SPF load of the ±20% simulation was lowered such that it coincides with the
experimental result. The behaviour of both the ±2% and ±20% simulations beyond the SPF
load is instable which is in agreement with the experimental observations. The region ends
until all the matrix of the CSM layer of all the elements fails progressively. This is clearly an
improvement over the simulation without imperfections since the region where the pipe is
failing progressive can be predicted. The region following the instable region corresponds to
phase where the matrix of the CSM layer of all the elements failed. The simulations continue
until the first fibre failure occurs thus rendering an un-converged solution indicating global
failure of the pipe.

3.3 Varying the thickness of the CSM based on the pipe thickness (Case B)

In this section, the thickness of the CSM layer of all the elements was varied randomly using
uniform and normal distributions, denoted as RAND and GDIS respectively in the following
plots, as detailed in section 2.2.1.b. Figure 7 shows the results of the mid-span expansion with
respect to the applied internal pressure for all the pipes.

Mid-span expansion vs. applied pressure

1.40% 1.40%

1.20% 1.20%
Mid-span expansion [%]
Mid-span expansion [%]

1.00% 1.00%

0.80% 0.80%

0.60% 0.60%

0.40% 0.40%

0.20% Pipe A 0.20% Pipe B

0.00% 0.00%
0 5 10 15 20 25 30 35 40 45 0 5 10 15 20 25 30 35 40 45
Applied pressure [bar] Applied pressure [bar]

1.40% 1.40%

1.20% 1.20%
Mid-span expansion [%]

Mid-span expansion [%]

1.00% 1.00%

0.80% 0.80%

0.60% 0.60%

0.40% 0.40%

0.20% 0.20%
Pipe C Pipe D
0.00% 0.00%
0 5 10 15 20 25 30 35 40 45 0 5 10 15 20 25 30 35 40 45
Applied pressure [bar] Applied pressure [bar]

Experimental PFA without imperfections GDIS RAND

Fig. 7. Plots of the mid-span expansion ( V ) versus the applied pressure (bar) for the four pipes listed in
table 1. Each plot compares the experimental results [23] with the corresponding PFA simulations with
varying CSM layer thickness according to the respective pipe thickness. GDIS and RAND denote a
normal and uniform distribution of the random number generated to calculate the random CSM layer
thickness.

17
It can be noted from figure 6 that the FPF load of the GDIS simulations is lower than the
simulations with the uniform distribution random number. This is attributed to the fact that
there are no bounds on the generated thickness which means that the will be layers with very
small thickness. This will induce premature failure and hence lower the FPF load. This can be
seen even in the SPF load.

Accumulation of failures vs. applied pressure of pipe D

(a) Normal distribution (b) Uniform distribution
100% 100%
Matrix Matrix failure of the CSM
90% failure of layer 90%
Matrix Matrix failure
80% the DR 80%
% of total elements

failure of of the CSM

layers
70% 70% the DR layer
60% 60% layers
1st
fibre failure in the

FPF load = 12.1 bar

50% CSM layer at 22.8 bar 50%
FPF load = 8.6 bar

40% 0.20% 40%

30% 30%

20% 0.00% 20%

22 24 26 28 30
10% 10%

0% 0%
8 10 12 14 16 18 20 22 24 26 28 30 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40
Applied pressure [bar] Applied pressure [bar]
Unfailed Matrix Failure Fibre failure

Fig. 8. Area plots showing the accumulation and distributions of failure modes for pipe D according to the
random number distribution used; (a) normal distribution and (b) uniform distribution.

The gradual variation in thickness for the normal distribution gave rise to a gradual failure
progression when compared to the uniform distribution. This effect is clearly shown in figure
8 where the matrix failure of the DR layers started at 8.6 bar which marks the FPF load and
ended at 13.6 bar (indicated by the first light blue shading in figure 8a). In the case of the
uniform distribution, this failure occurred suddenly at 12.1 bar (indicated by the vertical
dotted line in figure 8b). Furthermore, in the case of the normal distribution fibre failure was
predicted to initiate at 22.8 bar. This is indicated in Figure 8a in the small plot with enlarged
scale since the fraction of fibre failures was too low to show in the full plot (0.16% of the
total elements experienced fibre failure at the maximum load reached i.e. 30.9 bar). These
fibre failures are the reason why the ultimate load is lower with respect to the simulation
using the uniform distribution. Although the gradual failing is more realistic with respect to
the experimental results, the simulations require more time to converge because the failures
are distributed over a wider range of iterations.

3.4 Varying the thickness or material stiffness constants of all the layers based on Vv of
the material characterisation specimens (Case C, Type 1 & 2)

Following the results obtained in the previous section, the uniform distribution was opted for
the following last set of simulations. Two simulations where carried out for each pipe, one in
which the material property of each and every layer was varied and in the other, the thickness
of every layer was varied. In each simulation, the thickness of the layer was reduced by the
void fraction Vv listed in table 2 respectively. Furthermore, the thickness or material property
18
of each layer was varied randomly where the limits were set equal to the std(Vv) of the
respective layer type listed in table 2. The results of the mid-span expansion with respect to
the applied pressure for pipe A are shown in figure 9.

Mid-span expansion vs. applied pressure (Pipe A)

1.4%

Experimental FPF 13.3 bar

1.2% Mid-span expansion vs. applied pressure (Pipe A)
1.4%
1.0%
Mid-span expansion [%]

1.2%
0.8%

0.6% 1.0%
Mid-span expansion [%]

Instable region
PFA
FPF
0.4% 0.8% SPF

0.2%
0.6%
0.0%
0 0.4%
5 10 15 20 25 30 35 40 Vary45Material Prop
Applied pressure [bar] Vary thicknesses
Experimental
0.2%
Experimental Vary material properties Vary layer thicknesses

0.0% ( V ) versus the applied pressure (bar) for pipe A with varying
Fig. 9. Plot of the mid-span expansion
the thickness of the layers and material
0 properties
5 independently.
10 15 20 25 30 35 40 45
Applied pressure [bar]
The FPF load for both simulations occurred at approximately 11.6 bar which is lower than the
reported FPF load of the experiment 13.3 bar. This is expected due to the inclusion of
imperfections. On the other hand, the SPF predicted by both the simulations (20.4 bar) is in
good approximation with the experimental one (19.4 bar).

FPF load SPF load

Pipe A FPF failure mode SPF failure mode
[bar] [bar]
Experimental 13.3 N/A 19.4 N/A
Matrix failure of the Matrix failure of the
No imperfections 12.4 21.7
DR layers CSM layer
Matrix failure of the Matrix failure of the
Vary material properties 11.6 20.4
DR layers CSM layer
Matrix failure of the Matrix failure of the
Vary layer thicknesses 11.7 20.4
DR layers CSM layer
Table 6: List of first and second ply failure loads together with the corresponding failure modes of pipe A.

Table 6 lists the FPF, SPF and the corresponding failure modes of both simulations with
respect to the PFA without imperfections and the experimental observations [23] of pipe A.
Although only the FPF and SPF for pipe A are listed, the numerical results for the other three
pipes were similar. The failure modes corresponding to both the FPF and the SPF are similar
to the ones obtained from the PFA without imperfections. Analysing the expansion of both

19
the simulations shown in figure 9, varying the material properties or the thicknesses of the
layers gave relatively better results with respect to the results obtained from the simulations
which were based on the variation of the pipes’ wall thicknesses using the normal distribution
shown in figure 7. Furthermore, it was found that for the above mentioned cases, varying the
material properties (Type 2) gave better predictions during the instable region with respect to
Type 1 imperfections. The irregular variations in the mid-section expansion during the
instable regions are attributed to the random variations included in the thickness or material
properties. This irregular failure progression is difficult to predict accurately and this can be
noted by comparing the experimental results. Although the pipes were manufactured with the
same materials and the same machine, every pipe exhibited a slightly different failure
progression. This randomness is featured in the results obtained with the proposed
modifications to the PFA. Nonetheless, applying imperfections of either type predicted
similar FPF loads, SPF loads and instability regions which are the salient objectives of a
PFA.

4 Conclusions
In this paper a probabilistic nondeterministic PFA applied to asymmetric filament wound
pressure vessels was presented. A set of simulations using the classical PFA were run and it
was shown that although the FPF load was predicted successfully, failure progression
thereafter was not. This is attributed to the fact the geometry in the modelling software used
is perfect and free from any defect. In reality, this isn’t the case and in fact this is shown in
the numerous experiments on fibre reinforced polymers where the load-deflection
relationships are not linearly smooth. In this paper this region is referred to as the instable
region where the behaviour of the structure is irrational thus almost unpredictable.

With the proposed modification, i.e. the inclusion of imperfections based on the void
fractions of the specimens used in the material characterisation tests, the PFA is improved
considerably since the instable region can be predicted.

The strength capability of a pipe is determined from the FPF load according to the EN
13923:2005 [3] therefore knowing where the structure is instable is essential from the
optimisation point of view. For maximum performance efficiency, the FPF load should be
approximately equal to the start of the instable region. In this way using the proposed
modified PFA methodology will enable the engineer to design more efficient FRP structures
utilising their full strength potential.

20
5 References
[1] Hinton MJ, Soden PD, Kaddour AS. Failure Criteria in Fibre-Reinforced-Polymer
Composites: The World-Wide Failure Exercise. Amsterdam, The Netherlands:
Elsevier Science Ltd, 2004
[2] Muc A, Fugiel T. Influence of Physical and Geometrical Nonlinearities on Failure
Analysis of Composite Pressure Vessels. In: Proc. ICCM13, Beijing, 25-29 June
2001.
[3] Filament-wound FRP pressure vessels. Materials, design, manufacturing and testing,
EN 13923:2005.
[4] Liu PF, Zheng JY. Recent developments on damage modeling and finite element
analysis for composite laminates: A review. Mater. Design 2010;31(8):3825–3834.
[5] Lapczyk I, Hurtado JA. Progressive damage modeling in fiber-reinforced materials.
Compos. Part A Appl. Sci. Manuf. 2007;38(11):2333-2341.
[6] Rowlands RE. Strength (Failure) Theories and Their Experimental Correlations. In:
Sih GC, Skudra AM, editors. Handbook of Composites 3rd ed. New York: Elsevier,
1985. p. 71–125.
[7] Nahas MN. Survey of Failure and Post-Failure Theories of Laminated Fibre
Reinforced Composites. J. Compos. Technol. Res. 1986;8(4):138– 153.
[8] Echaabi JF, Trochu F. Review of Failure Criteria of Fibrous Composite Materials.
Polym. Compos. 1996;17(6):786–798.
[9] Hinton MJ, Soden PD. Predicting Failure in Composite Laminates: The Background
to the Exercise. Compos. Sci. Technol. 1998;58(7):1001– 1010.
[10] Soden PD, Hinton JM, Kaddour AS. A Comparison of the Predictive Capabilities of
Current Failure Theories for Composite Laminates. Compos. Sci. Technol.
1998:58(7):1225–1254.
[11] Hinton MJ, Kaddour AS, Soden PD. A Comparison of the Predictive Capabilities of
Current Failure Theories for Composite Laminates, Judged Against Experimental
Evidence. Compos. Sci. Technol. 2002;62(12-13):1725–1797.
[12] Paris F. A Study of Failure Criteria of Fibrous Composite Materials. NASA/CR-2001-
210661, 2001.
[13] Icardi U, Locatto S, Longo A. Assessment of recent theories for predicting failure of
composite laminates. Appl. Mech. Rev. 2007;60(1-6):76 – 86.
[14] Icardi U, Ferrero L. A comparison among several recent criteria for the failure
analysis of composites. J Adv. Mater. 2008;40(4):73-111.
[15] Labossiere P, Neale K. Macroscopic failure criteria for fibre-reinforced composite
materials. Solid Mechanics Archives. 1987;12(2): 65 – 95.
[16] Garnich MR, Akula VMK. Review of Degradation Models for Progressive Failure
Analysis of Fiber Reinforced Polymer Composites. ASME Appl. Mech. Rev.
2008;62(1): 010801-010833.
[17] Soliman HE, Kapania RK. Probability of Failure of Composite Cylinders Subjected to
Axisymmetric Loading. AIAA J. 2005;43(6):1342-1348.
[18] Bahei-El-Din YA, Mullur A, Hajela P, Peters J, Dvorak G. Nondeterministic
Modeling of Progressive Failure in Laminated Composites. in Proc. 47th

21
 AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials
Conference, Newport:RI, 1-4 May 2006.
[19] Khire R, Hajela P, Bahei-El-Din YA. Handling Uncertainty Propagation in Laminated
Composites Through Multiscale Modeling of Progressive Failure. in Proc. 48th
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials
Conference, Honolulu:HI, 23-26 April 2007.
[20] Tsai SW, Wu EM. A General Theory of Strength for Anisotropic Materials. J. Comp.
Mater. 1971;5(1):58-80.
[21] Hashin Z. Failure Criteria for Unidirectional Fiber Composites. ASME J. Appl. Mech.
1980;47(2):329–334.
[22] ANSYS® Academic Research, Release 15.0.
[23] Ellul B, Camilleri D, Grech J, Muscat M. Filament Wound Composite Pressure
Vessels and Pipes Subject to an Internal Pressure – An Experimental and Material
Characterization Study. submitted for publication.
[24] Ellul B, Camilleri D, Betts JC. A progressive failure analysis applied to fibre
reinforced composite plates subject to out-of-plane bending. Mech. Compos. Mater.
2014;49(6):605-620.
[25] Love AEH. On the small free vibrations and deformations of elastic shells.
Philosophical trans. of the Royal Society (London), 1888;série A(17): 491–549.
[26] Reissner E. The effect of transverse shear deformation on the bending of elastic
plates. ASME J. Appl. Mech. 1945;12(2):69-77.
[27] Mindlin RD. Influence of rotatory inertia and shear on flexural motions of isotropic,
elastic plates. ASME J. Appl. Mech. 1951;18(1):31–38.
[28] Narayanaswami R, Adelman HM. Evaluation of the tensor polynomial and Hoffman
strength theories for composite materials,” J. Compos. Mater.1977;11(4):366-377.
[29] Pietropaoli E. Progressive Failure Analysis of Composite Structures Using a
Constitutive Material Model (USERMAT) Developed and Implemented in ANSYS
©. Appl. Compos. Mater. 2012;19(3-4):657–668.

22