Assessing Impacts of Vine-Copula Dependencies:
Case Study of a Digital Platform for Overhead Cranes
Janusz Szpytko
a
and Yorlandys Salgado Duarte
b
AGH University of Science and Technology, Krakow, Poland
Keywords: Vine Copula, Risk Assessment, Cranes.
Abstract: Usually, components in a system degrade simultaneously and, for processes such as maintenance, predictions
of common failures due to degradation are needed to achieve accurate assessments for decision making. Vine
copula approach used in this paper is one way of approaching dependency modelling, offering in addition,
thanks to its features, flexibility when lack of data is an issue. Knowing that a multivariate vine copula
approach does not have a regular structure, in this paper, we propose an algorithm to simulate correlated
random numbers of a multivariate vine copula combining bivariate copulas, and the subject of study is the
evaluation of the impact of the vine copula dependency structure in a risk-oriented Monte Carlo simulation
model implemented in an online digital platform to support the maintenance strategies of a set of overhead
cranes.
1 INTRODUCTION
The amount of software and IT applications in the
modern industry is growing exponentially. Usually,
these IT applications or digital tools replace well-
established and well-known complex decision-
making processes with optimal and handy
programmed codes based on mathematical models,
ensuring with this level of integration, quick and
optimal decisions.
The industry with continuous processes is one of
the sectors with more applications. The reason is
linked with the high levels of interactions,
interoperability, and complexity in processes such as
maintenance and operation. For instance, introducing
IT applications is a necessity today in this sector of
industry.
In this paper, we are presenting other IT solutions
in an industry with continuous processes and the
object of the application is the maintenance strategies
of cooperative overhead cranes in a steel plant.
The overhead crane system operates under hazard
conditions, and these machines are critical devices in
the production line. These overhead cranes ensure the
movement of heavy loads within sectors of the
production line.
a
https://orcid.org/0000-0001-7064-0183
b
https://orcid.org/0000-0002-5085-3170
Even in the presence of high redundancy, when
one of the cranes fails for unexpected reasons, it can
be a critical situation for the steel plant.
The digital platform adopted in this work, is a
unique engineering practical application created
based on individual requirements to support the
maintenance decision making for a set of overhead
cranes, with the idea of minimizing the risk of
interaction between scheduled maintenance and
unexpected crane failures.
The tool is fully implemented in MATLAB and is
ready to be run on a personal computer. The main
sources of information related to the digital platform
are described in references Szpytko, J. and Salgado
Duarte, Y. (2020a), Szpytko, J. and Salgado Duarte,
Y. (2020b) and Szpytko, J. and Salgado Duarte, Y.
(2021). While the first reference is dedicated to
introducing the platform, the other two store the result
of the parametrization and sensitivity to the major
model variables achieved for a specific dataset.
To contextualize the digital platform and its
relationship to previous work, and to avoid gaps in the
description, we present the Digital Twins framework
to detail where the contribution is focused.
As we know, according to Grieves postulates, the
Digital Twins framework is composed of five
Szpytko, J. and Salgado Duarte, Y.
Assessing Impacts of Vine-Copula Dependencies: Case Study of a Digital Platform for Overhead Cranes.
DOI: 10.5220/0010709900003062
In Proceedings of the 2nd International Conference on Innovative Intelligent Industrial Production and Logistics (IN4PL 2021), pages 187-196
ISBN: 978-989-758-535-7
Copyright
c
2021 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
187
dimensions: physical object, virtual counterpart,
connection, data, and services.
In our case, the physical object is the coordination
of the maintenance decision making process for a set
of overhead cranes. The virtual counterpart is the Risk
Model and the Optimization Routine implemented in
the digital platform. The connection is made up by the
layers related to data processing. The data are the
historical degradation data, lifecycle maintenance,
system structure information, etc., collected by the
SCADA (Supervisory Control And Data Acquisition)
and SAP (Systems, Applications & Products in Data
Processing) systems.
Among all the dimensions mentioned, the
contribution of this paper impacts only the virtual
counterpart and connection. While the impact on
connection is addressed by the contribution Szpytko,
J. and Salgado Duarte, Y. (2022) and somehow it is
needed to refer to the impact in this paper, here we
will be focusing on the virtual counterpart impacts,
specially, the Risk Model.
The digital platform is composed of three layers,
the Data Processing, the Risk Model, and the
Optimization Routine to ensure, given the input
settings, the best scenario available for the system.
The Data Processing layer has the duty to collect,
filter and reshape the raw data on an online basis,
allowing to run the model smoothly and without
human intervention. Reference Szpytko, J. and
Salgado Duarte, Y. (2020a) point out how the process
works and at the same time alludes in some way to
how the data are connected to the variables in the Risk
Model.
In the filter and reshape steps, a formal flow data
processing diagram is applied to capture the
dependencies between overhead cranes through the
time-to-failure records of each crane analyzed, and
copula approach is the method selected to address the
measurement of dependencies. Reference Szpytko, J.
and Salgado Duarte, Y. (2022) describes in detail how
the dependency structure is built and validated for use
by the Risk Model.
The Risk Model uses the estimated dependency
structure to simulate potential failures in the overhead
cranes. The simulated stochastic vectors convolute
the maintenance scheduling and then, using an
Optimization Routine, the Risk Model is stressed by
reducing the interaction between the scheduled
maintenance and the failure predictions. As a result,
the achieved maintenance scheduling, one of the main
outputs of the digital platform, ensures that planned
maintenance routines are well-coordinated under the
minimum system failure criterion.
In this Risk Model, failure simulation is a weighty
variable and accurate predictions are needed to
achieve the expected results. Therefore, the
dependency structure estimation and consequently
the simulations resulting from the estimated structure
are crucial in this Risk Model.
Usually, to capture the dependencies between
components (cranes) within a system (set of cranes),
a common frame window is needed for the
measurement (time, in our case). This requirement is
indispensable and sometimes ends up as a limitation
in many applications in practice. Knowing the
dependency measurement limitations, and knowing
that, in our case, the time-to-failure marginals
between overhead cranes are shifted because these
machines have different life cycles, within the copula
approach family, vine copula is chosen to measure the
dependencies.
The selected approach guarantees a wide family
of options and flexibility when lack of data is an issue
because dependencies are measured in pairs, as
detailed in the reference Szpytko, J. and Salgado
Duarte, Y. (2022).
The vine copula approach does not have a
standard multivariate structure because is composed
by concatenations of pairwise bivariate copulas,
therefore, is a challenge generate random numbers
from a non-standard structure, and as we statement
above, accurate simulations are required for the Risk
Model.
In this paper, we present an algorithm for
simulating dependent random numbers given an
estimated vine copula structure. Most of the
contribution is aimed at discussing the algorithm
before it is used in practice. That said, artificial data
generated by a given vine copula structure will be
used to test the impact of the algorithm on the Risk
Model, then the link to previous contributions and the
results of the algorithm will be described.
The testing framework proposed and discussed in
the paper with an artificial vine copula structure is not
so far from the real case study. Usually, when real
data are used, the impacts are reflected in the
estimated parameters in each bivariate copula
(pairwise marginals of time-to-failure records) and in
the final concatenation between the pairwise bivariate
copulas. The range of potential copulas to be selected
during the estimation of the structure with real data in
each concatenation is the same family used in the
artificial structure. Therefore, whatever the final
structure, the algorithm will be able to simulate
dependent vectors of random values.
The remaining sections are organized as follows:
first, a broad description of the copula approach used
EI2N 2021 - IFAC/IFIP International Workshop on Enterprise Integration, Interoperability and Networking
188
will be presented, followed by the description of the
algorithm developed to simulate dependent random
numbers. Then an artificial copula structure proposed
to test the developed algorithm is described, showing
its linkage with the Risk Model. Finally, the paper
ends with the conclusions section.
2 VINE COPULA APPROACH
In 2002, Bedford, T., Cooke, R. M. (2002) introduced
the vine copula approach as a generalization of the
Markov trees used to model high-dimension
distributions. The cited work is supported by solid
previous research in uncertainty analysis for
constructing high dimensions distributions by
Markov trees, and the main contribution of the paper
is the introduction of a vine copula as a graphical
representation of conditional dependence.
Most recently, Aas, K., Czado, C., Frigessi, A.,
Bakken, H. (2009) based on the work of Bedford,
Cooke, and Joe, clearly describes, and applies how a
multivariate distribution can be modelled by pairwise
copulas concatenations. In addition, and more aligned
with the discussion of the research presented, the
cited paper formalized a definition of how to simulate
a multivariate distribution from concatenated
bivariate copulas but leaves open the discussion on
the implementation in practice.
The work of Aas, K., Czado, C., Frigessi, A.,
Bakken, H. (2009) and the results shared are in the
field of economics, but it is not until the previous year
that Sun, F., Fu, F., Liao, H., Xu, D. (2020)
successfully applies the vine copula approach to
degradation data, same field of application as us.
The above references archive the theoretical
foundations used in the research presented, and the
goal of the paper is to contribute further on the same
by presenting an application of the vine copula
approach in a risk-oriented model.
As we stated before, in the vine copula case, it can
be difficult to generate random numbers with
dependence when they have distributions structures
that are not from a standard multivariate distribution.
In this paper we propose an algorithm to simulate
a vine copula once all the components (pairwise
bivariate copulas) and connections of the structure
have been estimated. The algorithm is fully described
in the Appendix: Generating random numbers with
vine copula, and the definition at the most granular
level of the bivariate copulas used is taken from
MATLAB. (2019) help, software used in the
implemented tool.
Knowing that five bivariate copulas can be fitted
(see Appendix: Bivariate copula densities), in the
next section, a discussion is presented to describe the
features of each copula, as well as the commonalities
between them.
3 COPULA FEATURES
Bivariate copula functions try to capture the
dependence between marginals through the copula
parameter. The reason why we have several densities
is related to the operating space of each copula
distribution.
In the case of the Archimedean copulas, the
parameter manages the dispersion of the random
numbers. For instance, higher parameter values result
in less dispersion of the random values. When the
parameter is close to one, the random values are
somehow independent.
On the other hand, Gaussian and t-copula are
elliptical copulas. In these cases, the correlation
parameter ρ controls the dispersion of the random
values. Values of ρ close to one, more correlated
marginals.
Within the elliptical copulas, t-copula has two
parameters, therefore t-copula offers another feature
more, the tail dependency. Figure 1 shows the scatter
plot of random values generated with a Gaussian
copula, setting the correlation parameter ρ = 0.8, and
with the same correlation parameter and the degrees
of freedom parameter υ = 3, Figure 2 shows the scatter
plot of random values generated with a t-copula.
Visible between these two figures is how the t-
copula can simulate values at the corners of the
distribution space. This flexibility is given by the
parameter degrees of freedom υ. For instance, fixing
the correlation parameter and changing the degrees of
freedom to higher values, in a t-copula density, results
in a Gaussian copula. That said, t-copula is more
flexible than Gaussian and can better fit the data, but
this flexibility results in a costly computation.
Bivariate Gaussian and t-copula are symmetric
copula distributions, but in the Archimedean cases,
only Frank, as shown Figure 3, remain with same
property, Gumbel, and Clayton, shown in Figure 4
and Figure 5, respectively, are not symmetric copulas.
The combination of these five copulas to be used
in the model ensures a wide range of possibilities and
maps the entire distribution space that the data may
Assessing Impacts of Vine-Copula Dependencies: Case Study of a Digital Platform for Overhead Cranes
189
have, as we can see in the scatter plot in the Appendix:
Figures.
Once an overview of copula features has been
described, in the next section we set a copula structure
for assessing the impacts on the Risk Model. Leaving
ready after the discussion, the implementation in
practice with real data.
4 IMPACTS IN THE MODEL
The parameterized scenario as well as the Risk Model
used as the base case for comparison in this
contribution is described in references Szpytko, J. and
Salgado Duarte, Y. (2020b) and Szpytko, J. and
Salgado Duarte, Y. (2021).
In papers cited above, the generation of random
numbers to simulate potential failures in the overhead
cranes were considered independent. Now the
random numbers will be generated using a given
dependency structure.
Figure 5 shows the Risk Model overview and the
convolution product definition to obtain the Loss
Capacity indicator (see references cited for more
details), and in the same diagram, we also point out
the impacted variable by the dependency structure
and its connection with the Risk Model.
The proposed algorithm for generating random
dependent numbers is an independent layer that
transfers the dependencies information to the
simulated vector of potential overhead crane failures.
As a result, the simulation has built-in dependency
information and considers common failure states
between overhead cranes.
In the system analyzed, 33 overhead cranes make
up the system, but only 26 cranes report historical
failures.
Therefore, the vine copula dependency structure
is composed of 25 bivariate copulas concatenated.
For testing purposes, and considering the whole
range of copulas available in our case, we propose to
repeat five times the following five copulas to build
the entire vine copula structure, also following the
order listed below:
- Gaussian copula with parameter ρ = 0.8.
- t-copula with parameter ρ = 0.8 and degrees
of freedom υ = 3.
- Frank, Gumbel, and Clayton copulas, all of
them with parameter θ = 10.
Taking the parametrization of the above described
vine copula and merging the dependency structure
information into the Risk Model adopted, we obtain
the results in Table 1 and Table 2.
Table 1: Risk value of each scenario evaluated.
Scenario η (%)
Independent Vine copula
Capacity Loss
(
tons/
y
ear
)
Capacity Loss
(
tons/
y
ear
)
1 95 19830.88 28191.50
2 94 15809.63 22643.14
3 93 12926.06 18656.09
4 92 10161.75 14868.16
5 91 7730.12 11490.93
6 90 5499.55 8433.81
7 89 3827.58 6146.27
8 88 2717.24 4601.36
9 87 1648.54 3165.55
10 86 1161.59 2297.41
11 85 782.20 1615.97
12 80 111.91 290.17
Base 75 14.61 47.78
Table 1 shows the exponential increasing impact
on the risk indicator assessment when the Risk Model
variables are changed sensitively. The risk indicator
Capacity Loss is the conditional expected value of the
convolution between the available capacity of the
overhead crane system and the number of overhead
cranes required to support the production line.
Table 2: Variance of each scenario evaluated.
Scenario η (%)
Independent Vine copula
Sample
Size
Sample
Size
1 95 202 436
2 94 205 459
3 93 211 470
4 92 224 482
5 91 232 478
6 90 294 569
7 89 384 747
8 88 645 912
9 87 1111 1359
10 86 1214 1424
11 85 1462 1732
12 80 5907 3946
Base 75 28108 14339
EI2N 2021 - IFAC/IFIP International Workshop on Enterprise Integration, Interoperability and Networking
190
Figure 7 shows the visual view of the impact.
When more cranes are required to support the
production line, the risk of Loss Capacity due to
possible unexpected failures increases.
As expected also, when we now consider the
common failures among the overhead cranes, the risk
values per scenario is higher. The reason relates to the
incorporation of the common failure probability
states in the assessment.
In addition, Figure 7 also shows that the impact is
even more severe, in states with a larger convolution
area (see Figure 7 in Appendix, blue line: simulated
independent failures and red line: simulated failures
considering the proposed dependency structure). The
assessment shows a clear impact when considering
common events.
On the other hand, Table 2 shows how the
variance of the estimator behaves. For example, since
the copula approach with the parameters set has a
smaller individual scatter of the random numbers than
the previously generated independent values, as a
result, the system estimator Loss Capacity has less
variance as well.
The results presented in this paper show in some
way the potential impact of considering dependencies
in the adopted model and illustrate the range of
copulas that will be used in future states of research
and applied in practice.
It is important to clarify that in this work, the
figures related to the scatter plots were created using
the same parameterization described at the beginning
of this section.
5 CONCLUSIONS
The algorithm used to evaluate the impact of the
dependency structure on the adopted Risk Model
achieved the expected results. When common failures
between overhead cranes are considered and exist, the
scenario is more severe, and Table 1 and Figure 7 are
the evidence of the conclusion.
The research shows how the vine copula approach
can be applied to historical degradation data, and how
it can be used by the adopted digital platform under
study. This paper leaves the field ready to merge the
dependency measurement with real failure data (vine
copula structure), the risk assessment routines (Model
Risk and the algorithm proposed to simulate
dependent random numbers), and the maintenance
scheduling implemented in the digital platform.
Moreover, it is a clear application of the vine copula
approach.
This methodology can be extrapolated to another
dataset without much effort, following the same idea,
trying to measure certain dependencies between data
vectors corresponding to different components within
the same system.
In future steps of the research, as a continuation of
the presented work, we will share the application of
both stages (measurement and simulation) with real
failure data.
ACKNOWLEDGEMENTS
The work has been financially supported by the
Polish Ministry of Education and Science.
REFERENCES
Aas, K., Czado, C., Frigessi, A., Bakken, H. (2009) Pair-
copula constructions of multiple dependence.
Insurance: Mathematics and Economics. 44 (2009)
182–198.
Bedford, T., Cooke, R. M. (2002). Vines—A New
graphical model for dependent random variables. The
Annals of Statistics. 2002, Vol. 30, No. 4, 1031–1068.
MATLAB. (2019). version 9.7.0.1586710 (R2019b).
Natick, Massachusetts: The MathWorks Inc.
Sun, F., Fu, F., Liao, H., Xu, D. (2020). Analysis of
multivariate dependent accelerated degradation data
using a random-effect general Wiener process and D-
vine Copula. Reliability Engineering and System
Safety. 204 (2020) 107168.
Szpytko, J and Salgado Duarte, Y. (2020). “Integrated
maintenance platform for critical cranes under
operation: Database for maintenance purposes”.
Preprints of the 4th IFAC Workshop on Advanced
Maintenance Engineering, Service and Technology.
September 10-11, 2020. Cambridge, UK.
Szpytko, J. and Salgado Duarte, Y. (2020). “Exploitation
Efficiency System of Crane based on Risk
Management”. Proceedings of International
Conference on Innovative Intelligent Industrial
Production and Logistics, IN4PL 2020. 2-4 November
2020.
Szpytko, J. and Salgado Duarte, Y. (2021). Technical
Devices Degradation Self-Analysis for Self-
Maintenance Strategy: Crane Case Study. Proceedings
of INCOM 2021, June 2021, 17th IFAC Symposium on
Information Control Problems in Manufacturing.
Szpytko, J. and Salgado Duarte, Y. (2022). Digital Platform
for Overhead Cranes Maintenance Strategies:
Measuring Dependencies on Degradation Data with Vine
Copulas. Manuscript submitted to 14th IFAC
Workshop on Intelligent Manufacturing Systems.
Received July 20, 2021.
Assessing Impacts of Vine-Copula Dependencies: Case Study of a Digital Platform for Overhead Cranes
191
APPENDIX
Figures
Figure 1: Scatter plot of a Gaussian copula.
Figure 2: Scatter plot of a t copula.
EI2N 2021 - IFAC/IFIP International Workshop on Enterprise Integration, Interoperability and Networking
192
Figure 3: Scatter plot of a Frank copula.
Figure 4: Scatter plot of a Gumbel copula.
Assessing Impacts of Vine-Copula Dependencies: Case Study of a Digital Platform for Overhead Cranes
193
Figure 5: Scatter plot of a Clayton copula.
Figure 6: Risk Model architecture.
Figure 7: Dependency structure impact on the model.
EI2N 2021 - IFAC/IFIP International Workshop on Enterprise Integration, Interoperability and Networking
194
Generating Random Numbers with Vine Copula:
1.- For sampling ~u dependent uniform random
numbers [0, 1] with a vine copula, defining ~u
d,n
as a
d × n matrix, where d is the dimension and n is the
length of the random sample, first we need an w
independent uniform random sample [0, 1] as starting
point, then with the vine copula concatenation
estimated, we apply the following steps by u
d
component considered:
()
()
()
1
1
221
1
3312
1
11
,
,,
ddd
uw
uFuu
uFuuu
uFuu u
=
=
=
=
=

where w is a settled random number sample with
length n, and
1
()F
is a cumulative bivariate copula
density.
In this research, we consider five copula densities
(see Appendix: Bivariate copula densities), therefore,
1
()F
depends on the k-th bivariate copula density
used. Below we describe the steps performed
depending on the copula used, where v
1
and v
2
variables represent the pair correlated vector in each
step and w independent uniform random sample or a
sample vector of the previous concatenation step:
a.- If k-th copula density is Clayton: v
1
= w,
then
1
1
21 1
1vvp v
θ
θ
θ
θ


+


=−+



where θ is the copula
parameter defined 0 < θ < and p is an independent
uniform random sample [0, 1].
b.- If k-th copula density is Frank: v
1
= w, then
1
1
2
1
1
ln
1
1
v
v
p
ee
p
v
p
e
p
θ
θ
θ
θ


+




=−

+


where θ is the copula
parameter defined -∞ < θ < and p is an independent
uniform random sample [0, 1].
c.- If k-th copula density is Gumbel: v
1
= w
5π, then by successive transformation we obtain:
v
2
= v
1
+ π /2,
,
ln
dn
ep=−
where p is an independent
uniform random sample, d = 1 and n = sample size,
2
1
cos
v
v
t
e
θ



=
where θ is the copula
parameter defined 1 θ < ∞ and,
()
1
2
1
sin
cos
v
t
g
vt
θ
θ






=



,
,
1
ln
dn
s
pg
θ
=−
where p is an independent
uniform random sample, d = 2 and n = sample size,
s
ve
=
,
v
2
= v
d = 2,n
d.- If k-th copula density is t: first, given ρ
parameter, a positive correlation matrix, apply the T
Cholesky-like decomposition for covariance matrix,
such as ρ = T
T
T, and set v
1
= z (w), where z ( ) is a
normalization function, which centers the data to
have mean equal to 0 and scales it to have standard
deviation equal to 1. Then by successive
transformations:
r = v
1, n
, p
1, n
where p is an independent random
sample,
r = r × T,
,
2
dn
x
η
η

Γ


=
where η is the degrees of
freedom, Γ() is the gamma distribution, d = 2 and n
sample size.
r = r / x,
v = t (r, η) where t ( ) is the cumulative t
distribution, then v
2
= v
d = 1, n
.
e.- If k-th copula density is Gaussian: first,
given ρ parameter, a positive correlation matrix,
apply the T Cholesky-like decomposition for
covariance matrix, such as ρ = T
T
T, and set v
1
= z (w),
where z ( ) is a normalization function, which centers
the data to have mean equal to 0 and scales it to have
standard deviation equal to 1. Then by successive
transformations:
r = v
1, n
, p
1, n
where p is an independent random
sample,
r = r × T,
v = Normal (r) where Normal ( ) is the
cumulative Gaussian distribution, then v
2
= v
d = 1, n
.
2.- End of sampling ~u dependent uniform random
numbers [0, 1] with a vine copula. As a result, a
matrix of uniform dependent random numbers is
obtained.
Bivariate Copula Densities:
In this document we present five possible selections
in the bivariate copula fitting process performed for
continuous variables. Below we describe the list of
Assessing Impacts of Vine-Copula Dependencies: Case Study of a Digital Platform for Overhead Cranes
195
probability copula density functions used in the
selection.
1.- Clayton:
()
()
1
12 1 2
,; 1cu u u u
θ
θθ
θ
−−
=+
where θ
is the copula parameter defined 0 < θ < ∞.
2.- Frank:
()
()()
12
12
1
ln 1 1
,;
1
uu
ee
cu u
e
θθ
θ
θ
θ
−−
−−
=
where θ is the copula parameter defined -∞ < θ < ∞.
3.- Gumbel:
()
()( )
1
12
ln ln
12
,;
uu
cu u e
θ
θθ
θ

−− +


=
where θ
is the copula parameter defined 1 ≤ θ < ∞.
4.- Gaussian:
() ()()
11
12 1 2
,; ,cu u u u
ρ
ρ
−−

Φ Φ

where ρ is a pairwise correlation value defined -1 < ρ
< 1.
5.- t-copula:
()()()
11
12 , 1 2
,;, ,cuu t tutu
υρ υ υ
υρ
−−

=

where ρ is a pairwise correlation value defined -1 < ρ
< 1 and υ is the degrees of freedom parameter defined
υ > 1.
EI2N 2021 - IFAC/IFIP International Workshop on Enterprise Integration, Interoperability and Networking
196