applied

sciences
Article
Framework for Identification and Prediction of Corrosion
Degradation in a Steel Column through Machine Learning and
Bayesian Updating
Simone Castelli and Andrea Belleri *

Department of Engineering and Applied Sciences, University of Bergamo, Viale Marconi 5, 24044 Dalmine, Italy;
[email protected]
* Correspondence: [email protected]

Abstract: In recent years, structural health monitoring, starting from accelerometric data, is a method
which has become widely adopted. Among the available techniques, machine learning is one of the
most innovative and promising, supported by the continuously increasing computational capacity of
current computers. The present work investigates the potential benefits of a framework based on
supervised learning suitable for quantifying the corroded thickness of a structural system, herein
uniformly applied to a reference steel column. The envisaged framework follows a hybrid approach
where the training data are generated from a parametric and stochastic finite element model. The
learning activity is performed by a support vector machine with Bayesian optimization of the hyper-
parameters, in which a penalty matrix is introduced to minimize the probability of missed alarms.
Then, the estimated structural health conditions are used to update an exponential degradation
model with random coefficients suitable for providing a prediction of the remaining useful life of the
simulated corroded column. The results obtained show the potentiality of the proposed framework
and its possible future extension for different types of damage and structural types.

Keywords: structural health monitoring; machine learning; support vector machine; Bayesian
optimization; remaining useful life; synthetic data

Citation: Castelli, S.; Belleri, A.

Framework for Identification and
Prediction of Corrosion Degradation 1. Introduction
in a Steel Column through Machine In recent years, the topic of the permanent monitoring of structures using the latest
Learning and Bayesian Updating. technologies has become increasingly important. Traditional techniques and visual in-
Appl. Sci. 2023, 13, 4646. https:// spections are increasingly being joined or replaced by techniques that take advantage of
doi.org/10.3390/app13074646 technologies, such as signal processing, operational modal analysis, the Internet of Things
Academic Editor: Zengshun Chen (IoT), machine learning (ML), deep learning (DL), and big data, among others, which
allow for the production of continuous and sometimes more sensitive and reliable results
Received: 31 January 2023 compared to traditional techniques.
Revised: 24 March 2023
The discipline called structural health monitoring (SHM) aims to use these techniques
Accepted: 30 March 2023
for structural damage diagnosis, i.e., methods to identify the probability of existence,
Published: 6 April 2023
location, and severity of damage. The two most common strategies are data-driven and
model-based approaches. The model-based approach is generally more computationally
expensive and may even be prohibitively expensive for use in large-scale models. Instead,
Copyright: © 2023 by the authors.
data-driven methods allow faster processing, sometimes even in situ, based on statistical
Licensee MDPI, Basel, Switzerland. learning algorithms [1,2]. Early efforts considered and compared the standard deviation of
This article is an open access article the signal with the norm of outliers, while today, techniques based on machine learning
distributed under the terms and and neural networks are experimented [3–14]. Environmental vibration-based methods are
conditions of the Creative Commons among the most promising of the various methods for damage identification presented in
Attribution (CC BY) license (https:// the literature. A vibration-based damage identification method and a genetic algorithm
creativecommons.org/licenses/by/ that can identify the location and severity of structural damage under the influence of
4.0/). temperature and noise variations are presented in [15], while a method combining a support

Appl. Sci. 2023, 13, 4646. https://doi.org/10.3390/app13074646 https://www.mdpi.com/journal/applsci

Appl. Sci. 2023, 13, 4646 2 of 20

vector machine (SVM) with moth–flame optimization (MFO) for damage identification
under varying temperature and noise conditions and a practical example of its use on a
real bridge are proposed in [16]. Some authors have focused on damage identification by
the Bayesian method; however, this method still suffers from some limitations, such as
that the objective function, based on natural frequencies and modal shapes, is limiting
due to its limited number of sensors and that the sampling method still needs to be
improved. In [17], a Bayesian method based on a new objective function with autoregressive
coefficients (FAR) was developed, in which the sampling using the standard Metropolis–
Hasting–(MH) algorithm was improved by introducing particle swarm optimization (PSO),
obtaining a hybrid Markov chain–Monte Carlo (MH–PSO) sampling method and proposing
a method capable of sampling by greatly reducing the computational burden, as in [18],
through the use of a simple population MH algorithm (SP–MH). In order to quantify
the value of information extracted from a SHM system to implement it in a decision-
making tool, studies were conducted by [19], in which a heuristic model was used for
life-cycle optimization by the sequential updating of structural reliability based on the
identification of deterioration and the estimation of its evolution using a classical Bayesian
model updating method. In [20], a framework is proposed to transfer knowledge obtained
through synthetic data creation from earthquake simulations for various damage classes
to real data with exposure limited to a single health state via a domain adversarial neural
network (DANN) architecture. Again, to overcome the limitations imposed by knowledge
of the goodness–of–fit class data set alone, the authors of [21] developed a vibration-
based SHM framework for damage classification in structural systems to overcome this
limitation. The model is trained to acquire richness and knowledge in the learning task
from a source domain with extensive and exhaustive datasets and transfer that knowledge
to a target domain with much less information. Specifically, the proposed procedure is
based on creating a model that learns the lower-level features that characterize vibration
records from the rich audio dataset and then specializes its knowledge on the chosen
structural dataset. Other studies have focused on the detection and localization of structural
damage using decision tree ensembles (DTEs); in particular, the authors of [22] developed
a methodology, based on a set of decision trees, which belongs to the class of vibration-
based approaches as a method for the assessment of the health of a structure, obtained
by analyzing the dynamic properties of the structural system, in particular, mode shapes
and natural frequencies. The proposed damage detection method was validated for three
different test cases, comprising both numerical simulations and experimentally recorded
data. In [23], a new generalized auto-encoder (NGAE) is proposed and supplemented with
a statistical pattern–recognition approach using power cepstral coefficients of acceleration
responses. This method was validated using numerical simulations and experimental
data and shows better performance than the traditional auto-encoder (TAE) and principal
component analysis (PCA).
In this paper, we investigate the ability of a framework based on hybrid-based models
to identify damage under environmental conditions, herein applied for damage identifica-
tion due to the generalized corrosion of a steel structure [24]. The proposed framework,
implemented in the MATLAB environment [25], is able to identify the damage through
supervised learning algorithms, such as the SVM, which can classify damage states with
a good accuracy, even under stochastic conditions. The accuracy achieved in the damage
prediction is a function of the goodness of selected features and artificial intelligence (AI)
hyperparameters. For the AI training, stationary signals are generated through response
history dynamic analyses on a finite element (FE) model.
The second part of the paper explores the possibility of estimating the remaining
useful life (RUL) of a structural system using a Bayesian updating model. The potential of
condition monitoring to support high-level decision making, such as equipment replace-
ment, maintenance planning, and spare parts management, is indeed a topic of growing
interest. In order to make the most effective use of condition information, it is useful to
identify a degradation signal, i.e., a quantity computed from sensor data that is suitable
Appl. Sci. 2023, 13, x FOR PEER REVIEW 3 of 21

Appl. Sci. 2023, 13, 4646 replacement, maintenance planning, and spare parts management, is indeed a topic 3 ofof20
growing interest. In order to make the most effective use of condition information, it is
useful to identify a degradation signal, i.e., a quantity computed from sensor data that is
suitable for estimating the condition of the structure and thus providing information on
for estimating the condition of the structure and thus providing information on how this
how this condition is likely to evolve in the future. The approach studied was successfully
condition is likely to evolve in the future. The approach studied was successfully applied
applied to a simple structural system to test the concept.
to a simple structural system to test the concept.
2.2.Methodological
MethodologicalFramework
Framework
This
Thissection
sectiondescribes
describes the
the methodology
methodology used usedin inthis
thisarticle.
article.One
Oneofofthe
themain
main aspects
aspects for
for the application of this framework is the creation of a database of signals
the application of this framework is the creation of a database of signals that includes the that includes
the response
response of the
of the studied
studied structure
structure in different
in different damage damage configurations.
configurations. Since itSince
is notitpossible
is not
possible to obtain such a database through actual records, a parametric
to obtain such a database through actual records, a parametric FE model is analyzed. FE model is ana-
lyzed. Starting
Starting fromdatabase,
from this this database, it is possible
it is possible to continue
to continue the AI training
the AI training with super-
with supervised ML
vised ML techniques,
techniques, since the
since the signals aresignals
labeled.areAtlabeled.
this point,At the
thissecond
point, part
the second part focuses
of the work of the
work
on thefocuses on theofestimation
estimation RUL, i.e.,of byRUL, i.e., bythe
recording recording the real
real signals andsignals and by identify-
by identifying the class
ing the classand
of current of current and past
past damage, it isdamage,
possibleit tois estimate
possible theto estimate the RUL
RUL through through a
a degradation
degradation
mathematical mathematical
model. Sincemodel.
it is notSince it isto
possible not possible
have to have
real data real data
available, available,
the previous the
model
previous model was
was considered and considered
modified toand modified
simulate to simulate of
the progression thedamage
progression ofcase,
(in this damage (in
uniform
this case, uniform
corrosion corrosionsteel
on the reference on the reference
column) oversteel
time.column)
Figure 1over
showstime.
theFigure
general1 shows the
framework
just described.
general framework just described.

Generalframework
Figure1.1.General
Figure frameworkinvestigated.
investigated.

In this work, only the damage associated with a uniform corrosion pattern was
In this work, only the damage associated with a uniform corrosion pattern was con-
considered, without taking into account the influence of environmental variables on the
sidered, without taking into account the influence of environmental variables on the ex-
extracted features, such as temperature, humidity, wind load, and others. The only random
tracted features, such as temperature, humidity, wind load, and others. The only random
variable considered is the mass acting on the steel column. It is plausible that if the method
variable considered is the mass acting on the steel column. It is plausible that if the method
can identify the damage despite the noise due to mass variations, it will also be possible to
can identify the damage despite the noise due to mass variations, it will also be possible
identify the damage in the case of simultaneous environmental variables. In addition, the
to identify the damage in the case of simultaneous environmental variables. In addition,
specification of a uniform corrosion pattern could be considered a limitation. However, the
the specification of a uniform corrosion pattern could be considered a limitation. How-
objective of this article is not to study the corrosion phenomenon in depth, but to develop a
ever, the objective of this article is not to study the corrosion phenomenon in depth, but
framework for damage identification. Therefore, corrosion was used here as a degradation
to develop a framework for damage identification. Therefore, corrosion was used here as
phenomenon to test the damage progression identification framework.
a degradation phenomenon to test the damage progression identification framework.
3. Damage Detection Using Machine Learning Techniques
3. Damage Detection Using Machine Learning Techniques
This section highlights an approach to monitor and detect damage using ML tech-
ThisIn
niques. section highlights
the context an approach
in which we work, to monitor
SHM meansand detectandamage
training using ML
AI to associate tech-or
a state
niques.
class of damage with a vector of measurements of the structure of interest. The vectorsorof
In the context in which we work, SHM means training an AI to associate a state
class of damage should
measurements with a vector
consistofofmeasurements
quantities, i.e.,of the structure
features, which ofareinterest.
sensitive The
to vectors
the typeofof
measurements
damage being shouldstudied.consist of quantities,
The features could be,i.e.,
forfeatures,
example,which areidentified
the first sensitive fundamental
to the type
offrequencies
damage being studied.statistical
of vibration, The features could be,
parameters of for
the example,
recorded the firstwavelet
signal, identified funda-
coefficients,
mental frequencies
or artificial scalar of vibration,computed
quantities statistical from
parameters
seismicofsignals
the recorded
recorded signal, wavelet
at the co-
structure,
efficients,
among others.or artificial
Once thescalar quantities
appropriate computed
features from seismic
are determined, whichsignals
is notrecorded at the
straightforward,
structure, among others.
a “map” between Onceand
the features thetheappropriate
damage class features are determined,
of the structure whichtois
can be created not
extract
straightforward,
mathematical models a “map”thatbetween the features
can automate the SHM and the damage
process class of the
and eliminate the intervention
structure canof
human
be createdtechnicians
to extractas mathematical
much as possible. The goal
models that of SHM
can is to gather
automate the enough
SHM processinformation
and
to take appropriate corrective and preventive actions to restore and preserve the artifact, or
at least to ensure its safety.
Appl. Sci. 2023, 13, x FOR PEER REVIEW 4 of 21

Appl. Sci. 2023, 13, 4646 eliminate the intervention of human technicians as much as possible. The goal of SHM 4 of 20is
to gather enough information to take appropriate corrective and preventive actions to re-
store and preserve the artifact, or at least to ensure its safety.
The approach discussed here for solving the classification level is based on the idea
The approach discussed here for solving the classification level is based on the idea
of pattern
of pattern recognition
recognition (PR). (PR). In
In aa general
general sense,
sense, aa PR PR algorithm
algorithm isis simply
simply an an algorithm
algorithm that that
assigns a label, a class label, to a sample of measurement data. The
assigns a label, a class label, to a sample of measurement data. The class labels encode class labels encode the
location and rating of the damage. The training method, in which
the location and rating of the damage. The training method, in which the diagnosis is the diagnosis is trained
by linking
trained by labels
linking and a fulland
labels set aoffull
measurement data (including
set of measurement those of the
data (including different
those of thedam-dif-
age classes), is called supervised learning. A supervised learning
ferent damage classes), is called supervised learning. A supervised learning approach, approach, unlike unsu-
pervised
unlike learning, imposes
unsupervised learning,very stringent
imposes veryrequirements on the data needed
stringent requirements for learning.
on the data needed
First, as described earlier, measurement data are needed for every
for learning. First, as described earlier, measurement data are needed for every possible possible damage con-
figuration, and such data are difficult to find in the civilian domain.
damage configuration, and such data are difficult to find in the civilian domain. One One option would be
physical FE modeling. However, FE modeling could be
option would be physical FE modeling. However, FE modeling could be hampered by hampered by several problems,
such as:problems,
several the a priori choice
such of the
as: the damage
a priori choicetype,of and the model
the damage could
type, andbethecomplex in geom-
model could be
etry and material
complex in geometry definition and, therefore,
and material definition theand,
analysis couldthe
therefore, takeanalysis
a lot of time.
couldMoreover,
take a lot
it could
of time. be difficult it
Moreover, tocould
modelbe the damage
difficult toaccurately,
model the especially in the caseespecially
damage accurately, of nonlinearities
in the
or in the case of fatigue cracks with cyclic opening–closing phenomena.
case of nonlinearities or in the case of fatigue cracks with cyclic opening–closing phenom- Finally, there are
uncertainties about present and future loading and environmental
ena. Finally, there are uncertainties about present and future loading and environmental conditions. An alter-
native solution
conditions. could be tosolution
An alternative make small-scale
could be to ormake
full-scale copies and
small-scale then damage
or full-scale copiesthem.
and
the number
then damageofthem.copiesthe should
numbercorrespond
of copies to should
the combination
correspond of the different
to the types ofofdam-
combination the
age that can
different typesbe of
assumed
damagebased on their
that can varyingbased
be assumed severity. However,
on their varyingthisseverity.
solutionHowever,
is simply
not feasible
this solutionfor construction
is simply works.for
not feasible The proposed framework
construction works. Theis proposed
shown schematically
framework in is
Figure schematically
shown 2. Briefly, the in framework
Figure 2. uses a parametric
Briefly, the framework FE model
uses a(here using the
parametric FE software
FE model (here
using
OpenSeesthe FE software
[26]) for damageOpenSees [26]) for
definition anddamage definition
stochastic variablesandforstochastic variables
environmental for
condi-
environmental
tions. The FE model conditions.
is used The FE model
to train is usedusing
a ML model, to train a MLmethod.
the SVM model, using
DetailstheareSVMpro-
method.
vided in Details are provided
the following sections.in the following sections.

Figure 2.
Figure 2. Framework
Framework for
for machine
machine learning
learning training.
training.

4. Support Vector Machine (SVM)

Once the features have been extracted, it is possible to proceed with the AI training. In
this case, it was decided to use a supervised learning approach using SVM [27–33]. SVM
belongs to a class of ML algorithms called kernel methods or kernel machines (here, we
considered Gaussian, linear, and polynomial kernels). Before training the actual classifiers,
 4. Support Vector Machine (SVM)
Once the features have been extracted, it is possible to proceed with the AI training.
In this case, it was decided to use a supervised learning approach using SVM [27–33]. SVM
belongs to a class of ML algorithms called kernel methods or kernel machines (here, we
Appl. Sci. 2023, 13, 4646 5 of 20
considered Gaussian, linear, and polynomial kernels). Before training the actual classifi-
ers, we need to divide the data into a training set and a validation set. The validation set
is used to measure accuracy during model development. A SVM constructs an optimal
we need to divide
hyperplane the data
(by solving into a training
a quadratic set and aproblem)
optimization validationasset. The validation
a decision surfacesettoismax-
used
to measure
imize accuracy distance
the separation during model development.
between two classes. A SVM constructs
The support an optimal
vectors refer hyperplane
to a small
(by solving
subset of theatraining
quadratic optimization
observations problem)
that are usedasasa support
decision for
surface to maximize
the optimal the separa-
position of the
tion distance
decision surface.between two classes. The support vectors refer to a small subset of the training
observations that are used
Only the support as support
vectors selected for
from thethe
optimal position
training data areof needed
the decision surface.the
to construct
Only the support vectors selected from the training data are needed to
decision surfaces. Once the decision surfaces are identified, the remaining training data construct the
decision
are surfaces. Once the decision surfaces are identified, the remaining training data are
not considered.
not considered.
5. Case Study
5. Case Study
5.1.
5.1.FE
FEModel
Model
AAsimple
simplecantilever
cantilever column
column embedded
embedded at the basebase
at the andand
supporting
supportinga vertical load and
a vertical load
mass at the top was selected to test the concept. The column is 6 m high
and mass at the top was selected to test the concept. The column is 6 m high and has a and has a circular–
pipe cross-section
circular–pipe with an outer
cross-section withdiameter
an outerofdiameter
330 mm of and330a thickness
mm and aofthickness
20 mm. The of 20dam-
mm.
age was simulated simply as a uniform, general reduction in thickness
The damage was simulated simply as a uniform, general reduction in thickness due to due to corrosion,
without
corrosion,capturing
withoutthe electrochemical
capturing corrosion process.
the electrochemical corrosion The structure
process. Thetype and geometry
structure type and
could
geometrybe related
could beto related
columns of industrial
to columns buildings,
of industrial residential
buildings, buildings,
residential bridges,
buildings, and
bridges,
wind turbines.
and wind The column
turbines. was modeled
The column by 20 beam
was modeled by 20 elements
beam elements(Figure(Figure
3). All 3).
the All
infor-
the
mation, such as
information, input,
such mass, and
as input, mass, damage conditions,
and damage are generated
conditions, in the MATLAB
are generated in the MATLABenvi-
ronment
environmentand then
and input into OpenSees
then input [26] to[26]
into OpenSees perform the actual
to perform analysis.
the actual Once the
analysis. Once syn-
the
thetic signals are computed, they are brought back into MATLAB for
synthetic signals are computed, they are brought back into MATLAB for further processing. further processing.
The
Thecase
casestudy
studypresented
presentedherehereallows
allowsusustototest
testthe
theframework
frameworkfor fordamage
damagedetection
detection
and
andsubsequent
subsequentestimation
estimationofofthe theremaining
remaininguseful
usefullifelifeafter
aftera asimulated
simulatedcorrosion
corrosionprocess.
process.
However,
However,the thecase
casestudy
studyisisintended
intendedas asaabenchmark
benchmarkfor forany
anytypetypeofofdegradation/damage.
degradation/damage.

Figure
Figure3.3.Scheme
Schemeofofthe
theFE
FEmodel.
model.

Theexternal
The externalloading
loadingcondition
conditionisischaracterized
characterizedby
byananambient
ambientvibration
vibrationatatthe
thebase,
base,
representing a stochastic Gaussian white noise process. Sixty synthetic signals were
representing a stochastic Gaussian white noise process. Sixty synthetic signals were de-defined
(Figure
fined 4). 4).
(Figure

5.2. Feature Extraction

A series of response history dynamic analyses is performed to obtain the structural
vibrations in terms of the acceleration time history of the top of the column (since it is
basically a 1 degree of freedom, DOF, system). Then, these signals are analyzed and
processed to extract features (usually scalar quantities) that can quantify the health of the
structure. These features can be extracted either in the time domain or in the frequency
domain. The features analyzed are reported below. Each time series is stored with a health
code that is later used for artificial intelligence (AI) training.
Appl.
Appl.Sci.
Sci.2023,
2023,13,
13,x4646
FOR PEER REVIEW 6 6ofof21
20

Figure
Figure4.4.Gaussian
Gaussianstochastic
stochasticprocesses.
processes.

In thisExtraction
5.2. Feature paper, the following features are extracted and analyzed:
1. Apeak seriesamplitude
of response history dynamic analyses is performed to obtain the structural
2.
vibrations peak in frequency
terms of the acceleration time history of the top of the column (since it is
3. damping
basically a 1 degree coefficient
of freedom, DOF, system). Then, these signals are analyzed and pro-
4. peak band
cessed to extract features (usually scalar quantities) that can quantify the health of the
5.
structure. natural frequency
These features can be extracted either in the time domain or in the frequency
domain. For The features
dynamic analyzed are
identification, thereported
frequency below.
domain Each time series is (FDD)
decomposition storedalgorithm
with a health[34]
code
is used,thatwhere
is laterthe
used for artificial
search intelligence
for the resonance (AI) training.
peaks is automated using a local maximum
search In algorithm.
this paper, the following features are extracted and analyzed:
1. Onceamplitude
peak the features were extracted, they were plotted on graphs so that their quality
could
2. peak frequency An excellent feature allows easy discrimination of the healthy or
be quantified.
unhealthy
3. damping state of the structure. The algorithm, one-way analysis of variance (ANOVA),
coefficient
was
4. used
peak band to quantify the goodness-of-fit of the extracted features. This is to determine
if
5. the data from
natural frequency different groups have a common mean. ANOVA thus makes it possible
to find out if different groups of an independent variable have different effects on the
For dynamic identification, the frequency domain decomposition (FDD) algorithm
response variable. The one-way ANOVA is a simple case of a linear model. Such a model
[34] is used, where the search for the resonance peaks is automated using a local maximum
has the form:
search algorithm. yij = α j + eij (1)
Once the features were extracted, they were plotted on graphs so that their quality
where:be quantified. An excellent feature allows easy discrimination of the healthy or un-
could
yij is anstate
healthy observation: i represents
of the structure. Thethe observation
algorithm, number,
one-way and j of
analysis represents
variancea(ANOVA),
different group
was
of thetopredictor
used quantifyvariables y;
the goodness-of-fit of the extracted features. This is to determine if the
αi represents
data the population
from different groups have mean for the j–th
a common group;
mean. ANOVA thus makes it possible to find
e
out is the random error, independent
ij if different groups of an independent variable and having a normal distribution
have different with
effects onzero
the mean
responseand
2


constant The
variable. variance
one-way eij ∼ANOVA
N 0, σ is . a simple case of a linear model. Such a model has the
form:ANOVA assumes that all distributions are normal or with small deviations from this
assumption. In this study, this assumption is satisfied.
In this section, we examine how the 𝑦𝑖𝑗AI
= responds
𝛼𝑗 𝑒𝑖𝑗 to training, including damage with (1)
multiple classes. Specifically, the goal is to predict a corrosion thickness with a sensitivity
where:
of 1 mm and a maximum of 5 mm. Figure 5 shows the training scheme used; 600 response
𝑦 is an observation: i represents the observation number, and j represents a different
history analyses were performed by integrating as many stochastic white noise processes,
group of the predictor variables y;
where 1 mm of thickness is removed every 100 analyses to simulate the layer lost to
𝛼 represents the population mean for the j–th group;
corrosion. FaultCode 0 corresponds to the healthy state, while from FaultCode 1, the removal
𝑒of 1ismm the ofrandom
thicknesserror,
wasindependent
started; in this andway,
having a normal identifier
the FaultCode distribution alsowith zero mean
corresponds to
and constant
mm of corrosion loss.variance 𝑒 ~𝑁(0, 𝜎 ).
ANOVA assumesthe
To best simulate thattypical
all distributions
conditionsare of anormal or with
structure, the small deviations
mass lumped from
at the this
top of
assumption.
the column was In this study, this
introduced assumption(while
stochastically is satisfied.
it was kept constant in the single analysis
to keepIn thisthesection, we examineinvariant
mass assumptions how the during the i–th
AI responds to training,
record). A including damage
plot of the with
introduced
mass change
multiple is shown
classes. in Figure
Specifically, the 6. This
goal is condition
to predict represents
a corrosiona thickness
challengewithfor AI prediction,
a sensitivity
as1the
of mm AIandmust distinguishofand
a maximum recognize
5 mm. Figurethe change
5 shows theintraining
featuresscheme
due to aused;
change
600inresponse
stiffness
(in this analyses
history case duewere to corrosion)
performed from byaintegrating
change in mass.
as many stochastic white noise processes,
where 1 mm of thickness is removed every 100 analyses to simulate the layer lost to cor-
rosion. FaultCode 0 corresponds to the healthy state, while from FaultCode 1, the removal
 25

Thickness [mm]
20
Appl. Sci. 2023, 13, x FOR PEER REVIEW 7 of 21
15

10 F aultC ode F aultC ode F aultC ode F aultC ode F aultC ode F aultC ode
0 1 2 3 4 5
5
Appl. Sci. 2023, 13, 4646 7 of 20
0 of 1 mm of thickness was started; in this way, the FaultCode identifier also corresponds to
0 mm 100of corrosion200
loss. 300 400 500 600
Number analyses

25 Figure 5. Outline of analysis performed for AI training.

Thickness [mm]
20

15 To best simulate the typical conditions of a structure, the mass lumped at the top of
10 F aultC odethe column was introduced
F aultC ode F aultC ode stochastically
F aultC ode(whileF aul
it was
tC odekept constant
F aultC odein the single analysis
0 to keep the1 mass assumptions
2 invariant 3 during the4 i–th record). 5A plot of the introduced
5
mass change is shown in Figure 6. This condition represents a challenge for AI prediction,
0
0 as100
the AI must distinguish
200 and300
recognize the400
Number analyses
change in features
500 due to a 600
change in stiffness
(in this case due to corrosion) from a change in mass.
Figure 5. Outline of analysis performed for AI training.

60 To best simulate the typical conditions of a structure, the mass lumped at the top of
the column was introduced stochastically (while it was kept constant in the single analysis
Mass [kN/g]

40 to keep the mass assumptions invariant during the i–th record). A plot of the introduced
mass change is shown in Figure 6. This condition represents a challenge for AI prediction,
20 as the AI must distinguish and recognize the change in features due to a change in stiffness
(in this case due to corrosion) from a change in mass.
0
0 100 200 300 400 500 600
Number of analyses
60
100
Mass [kN/g]

40

50
20

0
00 100 200 300 400 500 600
46 48 Number50of analyses 52 54 56
Mass [kN/g]
100
Figure6.6. Stochastic
Figure Stochasticvariation
variationin
inmass.
mass.

50 Experience
Experience and and testing
testing have
have shown
shown thatthat even
eventraining
training AI AIwith
withthe thesame
sameparameters
parameters
and
and features results in different AI with different performance. Therefore, the
features results in different AI with different performance. Therefore, the question
question
arises:
arises: what
what are
are the
theparameters
parametersthat thatcan
canbest
bestbebeused
usedto totrain
trainAIAIwith
withthethesame
samefeatures?
features?
0 An
Anoptimization
optimizationproblem
problemisisthen then configured
configuredfor forthe
the AI
AI training
training parameters,
parameters, properly
properly
46
called 48 50 52 54 56
calledhyperparameters
hyperparameters hereafter,
hereafter, with
withthethegoal
goalofof
minimizing
minimizing thethe
error function
error in classifi-
function in clas-
Mass [kN/g]
cation. The problem takes the form of a huge and incredibly complex
sification. The problem takes the form of a huge and incredibly complex multidimensional multidimensional
optimization
optimization
Figure problem,
problem,
6. Stochastic and
andinititmass.
variation also
also has
has aastochastic
stochasticobjective
objective function,
function, since
since the
the outcome
outcome
varies
varies each time an AI is re–trained with the same parameters. This problem cannot
each time an AI is re–trained with the same parameters. This problem cannot bebe
addressed
addressed with
Experience classical
and testing
with classical deterministic
have shown
deterministic optimization algorithms.
that even training
optimization In In
AI with
algorithms. this
the
this work,
same
work, Bayesian
parameters
Bayesian op-
optimization
and features
timization is is used.in different AI with different performance. Therefore, the question
results
used.
arises: what are the parameters that can best be used to train AI with the same features?
5.3. Bayesian Optimization
An optimization problem is then configured for the AI training parameters, properly
called The adopted Bayesian
hyperparameters optimization
hereafter, with theinternally manages athe
goal of minimizing Gaussian process in
error function model
clas-
(GP) of the objective function and uses evaluations of the function itself
sification. The problem takes the form of a huge and incredibly complex multidimensional to train the model.
Bayesian
optimizationoptimization
problem,uses andan acquisition
it also function, which
has a stochastic objectivethe function,
algorithmsince uses to
thedetermine
outcome
the next point to evaluate. The acquisition function can
varies each time an AI is re–trained with the same parameters. This problem compensate for sampling
cannot be at
unexplored
addressed with points or explore
classical areas thatoptimization
deterministic have not yet algorithms.
been modeled [35].work,
In this A summary
Bayesian of op-
the
actual treatment
timization is used. of this technique can be found here. The Bayesian optimization algorithm
attempts to minimize an objective function f ( x ) in a finite domain. The function can be
deterministic or stochastic (as assumed in this research), which means that it can give
different results when evaluated at the same point.
The key elements of the process are:
1. a Gaussian process of f ( x ) [36]
2. a Bayesian updating procedure to change the GP with each new evaluation of f ( x )
Appl. Sci. 2023, 13, 4646 8 of 20

3. an acquisition function a( x ) (based on the GP model of f ) that is maximized to

determine the next point of x for the evaluation of the objective function
4. an evaluation of yi = f ( xi ) for randomly taken points of xi within the limits of the
variables
5. an update of the GP of f ( x ) to obtain a posterior distribution
6. a definition of the new point x that maximizes the acquisition function on a( x )
The interested reader is referred to [37–40]. The algorithm implements a stop criterion
when the maximum number of allowed iterations is reached.

6. Results
For the case study considered, the classification error function is shown in Figure 7,
which contains the output of Bayesian optimization:
each blue dot corresponds to an estimate of the minimum classification error computed by
the optimization process when all sets of hyperparameter values tested so far, including
the current iteration, are considered. The estimate is based on an upper confidence interval
of the current classification error objective model.
Each dark blue dot corresponds to the minimum observed classification error computed so
far by the optimization process.
The red square indicates the iteration corresponding to optimized hyperparameters. Such
hyperparameters do not always yield the minimum observed classification error. Bayesian
hyperparameter optimization chooses the set of hyperparameters that minimizes a higher
confidence interval of the objective model of the classification error, rather than the set that
minimizes the classification error.
The yellow dot indicates the iteration corresponding to the hyperparameters that yield the
lowest observed classification error.
Appl. Sci. 2023, 13, x FOR PEER REVIEW 9 of 21

Then, the best point is extracted as the optimized AI.

0.8 Estimated minimum classification error

Observed minimum classification error
Bestpoint hyperparameters
0.7 Minimum error hyperparameters
Minimum classification error

O p tim izatio n R esu lts

0.6 Multiclass method: O n e-vs-O n e
Box constraint level: 2.0657
Kernel scale: 1.911
0.5
Kernel function: G au ssian
Standardize data: tru e

0.4

0.3

0.2

20 40 60 80 100 120 140 160 180 200

Iteration

7. Classification error
Figure 7. error function.
function.

For
For sake
sakeof
ofbrevity,
brevity,the
theresults
resultsof of
this AIAI
this in the form
in the of scatter
form plots,
of scatter parallel
plots, plots,plots,
parallel and
ROC curves are omitted; instead, the corresponding confusion matrix after
and ROC curves are omitted; instead, the corresponding confusion matrix after validation validation is
shown
is shown in Figure 8. The trained AI achieves an accuracy of 73.3 percent and commits aa
in Figure 8. The trained AI achieves an accuracy of 73.3 percent and commits
maximum
maximum error
error in
in estimating
estimating thethe corroded
corroded thickness
thickness of
of 11 mm.
mm.

0 16 4

1 2 15 3
 Figure 7. Classification error function.

For sake of brevity, the results of this AI in the form of scatter plots, parallel plots,
and ROC curves are omitted; instead, the corresponding confusion matrix after validation
Appl. Sci. 2023, 13, 4646 9 of 20
is shown in Figure 8. The trained AI achieves an accuracy of 73.3 percent and commits a
maximum error in estimating the corroded thickness of 1 mm.

0 16 4

1 2 15 3

True Class 2 7 11 2

3 3 15 2

4 4 14 2

5 3 17

0 1 2 3 4 5
Predicted Class

Confusionmatrix.
Figure8.8.Confusion
Figure matrix.

ItIt isis worth

worth noting
noting that
that aa distinction
distinction should
should be
be made
made between
between the
the different
different types
types ofof
errors: the error in the upper triangular matrix corresponds to an overestimation of the
errors: the error in the upper triangular matrix corresponds to an overestimation of the
actual damage, which would lead to false alarms in real applications, while the errors in
actual damage, which would lead to false alarms in real applications, while the errors in
the lower triangular matrix are associated with an underestimation of the actual damage,
the lower triangular matrix are associated with an underestimation of the actual damage,
which may lead to missed alarms. While these two types of errors are the same in terms of
which may lead to missed alarms. While these two types of errors are the same in terms
damage quantification, they can have very different consequences: a false alarm is costly
of damage quantification, they can have very different consequences: a false alarm is
because of the inconvenience it causes, while a missed alarm can jeopardize the safety
costly because of the inconvenience it causes, while a missed alarm can jeopardize
Appl. Sci. 2023, 13, x FOR PEER REVIEW 10 ofthe
21
of the work and people. Based on these considerations, it is interesting to define how to
safety of the work and people. Based on these considerations, it is interesting to define
translate these technical observations so that AI can learn from them. The solution found
how to translate these technical observations so that AI can learn from them. The solution
is to insert penalties during the learning phase that limit the number of missed alarms in
found is
alarms to insert penalties during the9 shows
learning phase that limit used.
the number of AI
missed
favor ofinfalse favor of false
alarms. alarms.
Figure Figure
9 shows the penaltythematrix
penalty matrix
used. Then,
Then, the AI wasthe was
re–trained
re–trained through optimization.
through optimization.

0 0 1 2 3 4 5

1 2 0 1 2 3 4
True Class

2 3 2 0 1 2 3

3 4 3 2 0 1 2

4 5 4 3 2 0 1

5 6 5 4 3 2 0

0 1 2 3 4 5
Predicted Class

Penaltymatrix.
Figure9.9.Penalty
Figure matrix.

The resulting
The resultingAIAI achieves
achievesan
an accuracy
accuracyof
of 75.0%
75.0% and
and commits
commits aa maximum
maximum error
error in
in
estimating the corroded thickness of 1 mm. The introduction of the penalty matrix could
estimating the corroded thickness of 1 mm. The introduction of the penalty matrix could
also lead to a decrease in accuracy (although only to a very small extent) because its
also lead to a decrease in accuracy (although only to a very small extent) because its ob-
objective is not to increase accuracy, but to improve structural safety. Figures 10 and 11
jective is not to increase accuracy, but to improve structural safety. Figures 10 and 11 show
show the classification error function and the confusion matrix of the AI trained with
the classification error function and the confusion matrix of the AI trained with the pen-
the penalty matrix, respectively. Figure 12 shows how well Bayesian optimization fits.
alty matrix, respectively. Figure 12 shows how well Bayesian optimization fits. Neverthe-
less, Bayesian optimization succeeds in finding one of the best possible AIs, but not the
absolute best, because AIs with higher accuracy can be seen from the figure. Thus, one
could argue that Bayesian optimization allows one to find an excellent (not the best) AI
without having to perform many iterations to understand its variability and then take its
 Appl. Sci. 2023, 13, 4646 10 of 20

Nevertheless, Bayesian optimization succeeds in finding one of the best possible AIs, but
not the absolute best, because AIs with higher accuracy can be seen from the figure. Thus,
one could argue that Bayesian optimization allows one to find an excellent (not the best)
AI without having to perform many iterations to understand its variability and then take
its maximum; the solid blue line remains close to the upper bound. Figure 12 shows
the accuracies obtained by training tens of thousands of AIs trained by hyperparameter
randomization. The accuracy of a Bayesian optimized AI with an additional penalty matrix
is shown in the dashed blue line; the resulting accuracy is very similar to that without a
penalty matrix. As mentioned earlier, this is not a problem, since the goal is to reduce the
probability of missed alarms. We also wanted to use this plot to examine the variability in
accuracy as the number of FE analyses performed for training changes. Specifically, we
Appl. Sci. 2023, 13, x FOR PEER REVIEW started with 600 training analyses and increased the number to 4800. It can be seen that the
11 of 21
variability of accuracy decreases as the number of analyses performed increases, and there
Appl. Sci. 2023, 13, x FOR PEER REVIEW 11 of 21
is no overfitting.

0.9 Estimated minimum classification error

0.9 Observed
Estimated minimumminimum classification
classification error error
Bestpoint
Observed hyperparameters
minimum classification error
0.8
Bestpoint hyperparameters
Minimum error hyperparameters
0.8
Minimum error hyperparameters
error

O p tim izatio n R esu lts

classificationerror

0.7
0.7 O p tim iMulticlass
zatio n R esumethod:
lts O n e-vs-O n e
Multiclass method: O n e-vs-O n e
Box constraint level: 0.15423
Minimum classification

Box constraint level: 0.15423

Kernel Kernel scale:
483 2.483
0.6
0.6 scale: 2.
Kernel Kernel
function:function: G au ssian
G au ssian
Standardize
Standardize data: trudata:
e tru e
0.5
0.5
Minimum

0.4
0.4

0.3
0.3

0.2
0.2
20 40 60 80 100 120 140 160 180 200
20 40 60 80 Iteration100 120 140 160 180 200
Iteration
Figure 10. AI classification error function trained with the penalty matrix.
Figure 10.10.
Figure AIAIclassification
classificationerror
error function trainedwith
function trained with the
the penalty
penalty matrix.
matrix.

0 13 6 1

0 13 6 1
1 1 14 5

1 1 14 5
True Class

2 2 16 2
True Class

2 2 16 2
3 3 13 4

3 3 13 4
4 2 14 4

5 4 2 14 20 4

0 1 2 3 4 5
5 Predicted Class 20

Figure
Figure 11. Confusion
11. Confusion matrix of AI trained
matrix with3with the penalty matrix.
0 1 of AI trained
2 the penalty
4 matrix. 5
Predicted Class

Figure 11. Confusion matrix of AI trained with the penalty matrix.

Appl. Sci. 2023, 13, x FOR PEER REVIEW 12 of 21
Appl.Sci.
Appl. Sci.2023,
2023,13,
13,4646
x FOR PEER REVIEW 12
11 of 21
of 20

Figure 12. Variability of accuracy.

Figure 12.
Figure 12. Variability
Variabilityof
ofaccuracy.
accuracy.
7. Remaining Useful Life Estimate
7.
7. Remaining
Remaining Useful
Useful Life
Life Estimate
Estimate
Predictive maintenance is a maintenance approach that uses predictive mathemati-
Predictive maintenance
Predictivemodels,
maintenance isisa amaintenance
maintenance approach that uses predictive mathematical–
cal–statistical
statistical models, typically
typically based based
on data data approach
onrecordedrecorded
by
that
by
sensors,
uses
to
predictive
sensors, detect mathemati-
toanomalies,
detect anomalies,
identify
cal–statistical
identify the models,
quantity typically
and quality based
of on data recorded by
damage/degradation, and sensors,
estimatetothe
detect anomalies,
remaining use-
the quantity
identify and quality
the quantity and of damage/degradation,
quality and
of damage/degradation, estimate
and the remaining useful life
estimate the remaining use-
ful life
(RUL). (RUL).
Figure Figure
13 shows13 shows
a simple a simple
flowchartflowchart
for RUL for RUL estimation.
estimation.
ful life (RUL). Figure 13 shows a simple flowchart for RUL estimation.

Figure 13. Flowchart

Figure 13. Flowchart for
for the
the remaining
remaining useful
useful life
life estimation
estimation algorithm.
algorithm.
Figure 13. Flowchart for the remaining useful life estimation algorithm.
In Figure 13, the first step is to collect data from the monitored system. Then, the
In Figure 13, the first step is to collect data from the monitored system. Then, the raw
raw data are pre–processed to feed into a data
damage fromdetection algorithm (this stepthe canraw be
data InareFigure 13, the first
pre–processed to step
feed is intoto collect
a damage detectionthe algorithm
monitored system.
(this stepThen,
can be based
based
data on ML,
areamong among
pre–processed others).
to this In
feedstep, this step,
into adifferent different
damagetechniques techniques
detection algorithm can be implemented,
(this step can both
be based
on ML, others). In can be implemented, both tradi-
traditional
on ML, and AI
among techniques,
others). In this tostep,
detect anomalies,
different classify can
techniques the type
be of anomalies or
implemented, damage,
both tradi-
tional and AI techniques, to detect anomalies, classify the type of anomalies or damage,
quantify
tional andand AIlocalize the damage,
techniques, to detect and estimate the
anomalies, remaining
classify the lifetime
type of of the structure.
anomalies or damage,It is
quantify and localize the damage, and estimate the remaining lifetime of the structure. It
advantageous
quantify to collect
and localize as large
the damage, and heterogeneous
andand estimate a dataset
the remaining as possible.
lifetime This is achieved
is advantageous to collect as large heterogeneous a dataset asofpossible.
the structure.
This is It
by
is collecting
advantageous data under
to different
collect as operating
large and conditions, such
heterogeneous a as environmental
dataset as possible.conditions,
This is
achieved by collecting data under different operating conditions, such as environmental
which
achieved allows the development
bywhich
collecting data of a more
under differentrobust algorithm.
operating In civil engineering,
conditions, such as representa-
conditions,
tive damage data allows
are hardly theavailable.
development of
Therefore, a more
a robust
mathematical algorithm.
model In environmental
can civil engineer-
conditions,
ing, which allows
representative damage thedata
development
are hardly ofavailable.
a more robustTherefore,algorithm.
a In civildeveloped
mathematical
be engineer-
model
to predict
ing, system response
representative damage asdata
a function
are of parameters
hardly available. and environmental
Therefore, a conditions.
mathematical model
can be Suchdeveloped
response to predict
signals system
are called response
synthetic as a function
signals and of parameters
should be and
integrated environmen-
with the
canconditions.
tal be developed to predict system response as a function of parameters and environmen-
recorded data
tal conditions. of the real structure. When the desired amount of real data is not available,
Such response
integration betweensignals
real data are(undamaged
called synthetic signals
structure) andand should be
simulated integrated
data (damaged with the
struc-
Suchdata
recorded response
of the signals
real are called
structure. When synthetic
the signals
desired and
amount should
of real be integrated
data is not with
available,the
ture) is generally
recorded data of sought.
the real The real When
structure. data are the affected
desired by measurement
amount of real data errors
is notdue to the
available,
integration between
measurement real data while
infrastructure, (undamaged structure)
the simulated data and aresimulated
affected data by (damaged
model errorsstruc-
due
integration
ture) between
is generally in real
sought. data (undamaged
The real data structure)
are affected and simulated
byphysical
measurement data (damaged
errors due struc-
to the
to simplifications the mathematical description of the phenomenon. Once the
ture) is generally
measurement sought. The
infrastructure, real the
while data are affected by measurement errors due due
to the
data are recorded or simulated, the nextsimulated
step (Figure data 13)are
is affected
to preprocess by model themerrors
to facilitate to
measurement infrastructure,
simplifications in the while the
mathematical simulated of
description datathe are affectedphenomenon.
physical by model errors Onceduethe to
the identification of health indicators. This step could be based on one or more techniques,
simplifications
data are recorded in the mathematical
or simulated, description
the next step of the
(Figuretrend physical
13) iselimination, phenomenon.
to preprocess Once
them to facilitate the
such as noise removal, outlier elimination, dummy frequency domain
dataidentification
the are recordedof orhealth
simulated, the next
indicators. Thisstep
step(Figure
could 13)
be is to preprocess
based on one or themtechniques,
more to facilitate
transformation, and many others. The next step is to identify the structural health by calcu-
the identification
such as noise of health
removal, indicators.
outlier elimination,This step couldtrend be based on one orfrequency
more techniques,
lating and comparing features that respond dummy elimination,
to the type of degradation/damage domain
analyzed.
such as noise removal,
transformation, and outlier
many elimination,
others. Theanext dummy istrend
stepoperation elimination,
to identify frequency domain
These features are used to distinguish healthy fromthe structural
a damaged health
one. In the by
transformation, and many others. The next step is to identify the structural health by
Appl. Sci. 2023, 13, x FOR PEER REVIEW 13 of 21

Appl. Sci. 2023, 13, 4646 12 of 20

calculating and comparing features that respond to the type of degradation/damage ana-
lyzed. These features are used to distinguish a healthy operation from a damaged one. In
the present work, the last and the previous steps are performed by applying AI. Finally,
present work, the last and the previous steps are performed by applying AI. Finally, the
the last two steps in Figure 13 define a relationship between the health status estimated
last two steps in Figure 13 define a relationship between the health status estimated by the
by the AI and the aging path of the structure, by allowing one to estimate the remaining
AI and the aging path of the structure, by allowing one to estimate the remaining useful
useful life (RUL) for the structure by imposing a minimum safety threshold and, thus,
life (RUL) for the structure by imposing a minimum safety threshold and, thus, defining
defining maintenance schedules.
maintenance schedules.
The RUL is defined as the time between the current situation and the minimum per-
The RUL is defined as the time between the current situation and the minimum
formance threshold established for the structural system. Depending on the situation, the
performance threshold established for the structural system. Depending on the situation,
timetime
the maymay be expressed
be expressed in days, flights,
in days, cycles,cycles,
flights, among others.others.
among For civil engineering
For works,
civil engineering
the RUL
works, theis RUL
specified in years.
is specified Algorithms
in years. for RUL
Algorithms forcalculation
RUL calculation can becandivided into three
be divided into
categories: similarity algorithms, survival algorithms, and degradation
three categories: similarity algorithms, survival algorithms, and degradation algorithms. algorithms. Figure
14 shows
Figure the different
14 shows methods
the different for calculating
methods for calculatingRUL. RUL.The Thechoice
choiceofofthe
themost
most effective
effective
method depends on the system and the available data.
method depends on the system and the available data. The degradation modelThe degradation model method
method is
used when undamaged state data are available and a safe threshold
is used when undamaged state data are available and a safe threshold can be defined. can be defined. It is
possible
It to use
is possible tothis
use information
this informationto update a mathematical
to update a mathematical modelmodelof degradation to a per-
of degradation to
a performance indicator and then use the history recorded up to that point tohow
formance indicator and then use the history recorded up to that point to predict the
predict
indicator
how will change
the indicator willin change
the future. In this
in the way,Init this
future. is possible
way, ittois statistically
possible toestimate the
statistically
remaining time before the indicator reaches the imposed threshold
estimate the remaining time before the indicator reaches the imposed threshold and, thus, and, thus, the RUL. A
linear degradation model can be used for this purpose if the
the RUL. A linear degradation model can be used for this purpose if the system does system does not experience
cumulative
not experience degradation.
cumulative The mathematical
degradation. Themodel describesmodel
mathematical the physical
describes process of deg-
the physical
radationofasdegradation
process a stochastic as linear process. linear
a stochastic If, on the other If,
process. hand,on thethe other
system experiences
hand, the systemcu-
mulative degradation,
experiences cumulativean exponentialandegradation
degradation, exponentialmodel is used,
degradation in which
model the in
is used, physical
which
process
the of degradation
physical is modeled by
process of degradation an exponential
is modeled by an stochastic
exponential process.
stochastic process.

Figure 14.
Figure 14. Methods
Methods of
of estimating RUL.
estimating RUL.

For civil
civil engineering
engineeringapplications,
applications,thethebest
bestmodels
models forfor
RULRULestimation
estimationareare
degradation
degrada-
models (i.e., (i.e.,
tion models models in which
models the failure
in which threshold
the failure is known,
threshold is known, Figure 14).14).
Figure Other types
Other of
types
models (i.e., similarity models and survival models, Figure 14) were
of models (i.e., similarity models and survival models, Figure 14) were initially initially discarded
because structural and infrastructure
infrastructure works
works are
are typically
typically unique,
unique, thus
thus it is difficult to have
the required data for the operating conditions, including degradation data. These models
require as
as input
inputaasingle
singlehealth
healthindicator,
indicator,which,
which, herein,
herein, is derived
is derived from
from thethe
rawraw
datadata
ob-
obtained from
tained from thetheAIAImodel,
model,i.e.,
i.e.,the
theestimated
estimatedcorrosion
corrosionthickness,
thickness, as
as indicated
indicated in in the
the
previous section. The degradation models used herein is the exponential model, which
implements a stochastic model with an exponential degradation over time [41–48]:
Appl. Sci. 2023, 13, x FOR PEER REVIEW 14 of 21

Appl. Sci. 2023, 13, 4646 13 of 20

previous section. The degradation models used herein is the exponential model, which
implements a stochastic model with an exponential degradation over time [41–48]:

𝑆(𝑡) = 𝜙 𝜃(𝑡)𝑒 ( ) ( )σ2 (2)

S(t) = φ + θ (t)e β(t)t+e(t)− 2 (2)
where:
where:
𝜙 is the intercept of the mathematical model at the origin
φ is the intercept of the mathematical model at the origin
𝜃(𝑡) is a random variable modeled with Lognormal distribution
θ (t) is a random variable modeled with Lognormal distribution
𝛽(𝑡) is a random variable modeled with Gaussian distribution
β(t) is a random variable modeled with Gaussian distribution
𝜖(𝑡) is the model noise implemented as a random variable with normal distribution
e(t) is the model noise implemented as a random variable with normal distribution
𝜎 is the variance of the noise
σ2 is the variance of the noise
8. Case
8. Case Study
Study Application
Application
The cantilever
The cantileversteelsteelcolumn
columnconsidered
consideredbefore
beforeis isanalyzed
analyzed with
with some
some modifications
modifications to
to test
test thethe
RUL RUL calculation.
calculation. First,
First, the the damage
damage of column
of the the column (i.e.,(i.e., the corrosion
the corrosion thickness)
thickness) was
was correlated
correlated with assuming
with time, time, assuming a continuous
a continuous degradation
degradation rate equal rateto equal to 80 μm/year,
80 micrometers/year,
which is the reduction associated with an urban–industrial
which is the reduction associated with an urban–industrial environment [49]. environment [49].
Figure 15
Figure 15 shows
showsthe theframework
framework developed
developed and implemented.
and implemented. Herein, the gray-high-
Herein, the gray-
highlighted portion allows creating benchmark data to test the RUL algorithm. The whole
lighted portion allows creating benchmark data to test the RUL algorithm. whole
repeated whenever
framework is repeated whenever the damage
damage evaluation
evaluation is required,
required, thus
thus allowing
allowing the
the
update of
update of the
theparameters
parametersofofthe thedegradation
degradationmodel.
model.As Asshown
showninin Figure
Figure 15,15,
thethe approach
approach is
is very
very similar
similar to AI
to the thetraining
AI training described
described in theinprevious
the previous
section. section. The lumped
The lumped mass atmass at
the top
thethe
of topcolumn
of the column was introduced
was introduced in stochastic
in stochastic terms, the terms,
signal theatsignal
the baseat the
wasbase was reas-
reassigned in
signed in each iteration as a pseudo–random white–noise signal to
each iteration as a pseudo–random white–noise signal to simulate environmental vibrations, simulate environmen-
tal vibrations,
and the damage and the damage parameterization
parameterization was modified towas modified
account for a to account for a time–de-
time–dependent damage
function
pendent (in this case
damage a function
function governing
(in this the corrosion
case a function governing rate).the corrosion rate).

Figure 15. Framework for estimating RUL.

Figure 15. Framework for estimating RUL.
The following parameters have been used:
1. corrosion rate: 80 micrometers/year
2. simulated operating period: 60 years
3. number of damage estimations for each simulated year (i.e., number of runs of the
trained AI to estimate the corroded thickness): 10
 The following parameters have been used:
1. corrosion rate: 80 μm/year
Appl. Sci. 2023, 13, 4646 2. simulated operating period: 60 years 14 of 20
3. number of damage estimations for each simulated year (i.e., number of runs of the
trained AI to estimate the corroded thickness): 10
4. input signal: pseudo–random white noise signal, with a length of 10 min and a sam-
4. input signal: pseudo–random white noise signal, with a length of 10 min and a
pling frequency of 200 Hz, for each damage identification run
sampling frequency of 200 Hz, for each damage identification run
5.5. safety threshold: 55 mm
safety threshold: mm of
of corroded
corroded thickness
thickness
Figure
Figure 16
16 shows
shows thethe comparison
comparison between
between thethe thickness
thickness lostlost in
in the
the corrosion
corrosion process
process
associated with the assumed corrosion rate, i.e., the thickness
associated with the assumed corrosion rate, i.e., the thickness reduction reduction entered in the
entered FE
in the
model, as well as the AI prediction. The discontinuity of the latter is
FE model, as well as the AI prediction. The discontinuity of the latter is due to the AIdue to the AI training
phase,
trainingwhere
phase, the recognition
where of a corrosion
the recognition thicknessthickness
of a corrosion with a sensitivity of 1mm was
with a sensitivity of 1mmset.
One could
was set. Onehave
couldopted
haveto opted
createtoa training scheme with
create a training scheme higher
withsensitivity, say 0.5 mm,
higher sensitivity, say 0.5in
the training phase. However, this would certainly have lowered
mm, in the training phase. However, this would certainly have lowered the accuracy, asthe accuracy, as the AI
would have struggled
the AI would more to
have struggled recognize
more a smaller
to recognize variation
a smaller in the in
variation extracted features
the extracted (e.g.,
features
natural frequency). The number of analyses to be performed is another
(e.g., natural frequency). The number of analyses to be performed is another important important aspect
affecting the choice
aspect affecting theof the training
choice scheme.scheme.
of the training As mentioned before, seeking
As mentioned before,aseeking
more sensitive
a more
and accurate
sensitive andAI results AI
accurate in an increase
results in aninincrease
the number of signals
in the numbertoofbesignalscreated, to which may
be created,
be far from
which may benegligible. Herein, theHerein,
far from negligible. proposed the AI represents
proposed a good compromise
AI represents between
a good compromise
training
betweentime, sensitivity,
training and accuracy,
time, sensitivity, even if it even
and accuracy, sets up
if ita sets
discontinuous trend. Thetrend.
up a discontinuous best
compromise
The best compromise between computational time and sensitivity will be the objectde-
between computational time and sensitivity will be the object of future of
velopments.
future developments.

6
Corrosion imposed
Health Indicator (AI)

4
Corrosion [mm]

0
0 10 20 30 40 50 60
Elapsed time [year]

Figure 16.
Figure Damage estimation
16. Damage estimation by
by AI
AI as
as time
time passes.
passes.

Figure 17 shows the output of the RUL algorithm (Figure 15). The algorithm is run for
Figure 17 shows the output of the RUL algorithm (Figure 15). The algorithm is run
each new identification, and at each time point, the RUL is recalculated by updating the
for each new identification, and at each time point, the RUL is recalculated by updating
parameters. The diagram above shows the comparison between the real RUL (RULreal ), i.e.,
the
the parameters. The diagram
remaining useful above shows
life calculated the comparison
by knowing between
the confidence the(5real
limit mmRUL (RULreal),
of corrosion)
i.e., the remaining useful life calculated by knowing the confidence limit (5 mm
and the corrosion rate function, as well as the RUL obtained by updating the degradation of corro-
sion)
model (RULest ). The confidence interval of the latter is also given. The bottom graph degra-
and the corrosion rate function, as well as the RUL obtained by updating the shows
dation model (RUL est). The confidence interval of the latter is also given. The bottom graph
the ratio RULest /RULreal in percent.
shows theimportant
One ratio RULest /RULreal
aspect in percent. stands out: the estimated RUL almost overlaps
immediately
with the actual RUL in the period between about 15 and 40 years. This proves the excellent
estimation of the forecast. In contrast, the estimation before 10 to 15 years and after 40 years
is unreliable, but not problematic, if we consider the nature of the phenomenon: in the early
years of the structure’s life, the algorithm will learn and update itself to detect the physio-
logical degradation trend of the structure (of course, in the case of damage/pathological
degradation, there is always an AI that can provide us with a snapshot of the current
damage state), and, in this time frame, it is less relevant to estimate the RUL; analogously,
Appl. Sci. 2023, 13, 4646 15 of 20

in the last moments of the structure life, it is obviously useless to estimate the RUL, since the
damage identification AI will report damage close to the imposed threshold. Instead,
Appl. Sci. 2023, 13, x FOR PEER REVIEW 16in
ofthe
21
intermediate life span, it is important to make an accurate RUL estimate in order to detect
and locate future damage and have enough time to effectively plan maintenance actions.

Remaining Useful Life

50
RUL real
RUL est
40
90% confidence interval RUL est
RUL [years]

30

20

10

0
RUL esti / RUL real [%]

0 5 10 15 20 25 30 35 40 45 50
200 [Years]

100

0
0 5 10 15 20 25 30 35 40 45 50
Elapsed time [years]

Figure 17.
Figure Estimated RUL.
17. Estimated RUL.

However,
One it is interesting
important to note that
aspect immediately the estimated
stands RUL is practically
out: the estimated RUL almost always lower
overlaps
than the actual RUL, i.e., it underestimates the actual RUL. From
with the actual RUL in the period between about 15 and 40 years. This proves the excellent a structural safety perspec-
tive, it is certainly
estimation better to be
of the forecast. In in this state
contrast, thethan in the opposite
estimation before state,
10 to i.e., overestimating
15 years and after the 40
RUL. is
years From a mathematical
unreliable, point of view,ifthis
but not problematic, we is due to two
consider the factors:
nature of the degradation
the phenomenon: model in
used is exponential, and a penalty matrix was used in the AI
the early years of the structure’s life, the algorithm will learn and update itself to detecttraining. The exponential
degradation
the physiological model, even if updated,
degradation trend willof the always provide
structure (ofacourse,
shorter inremaining
the caselife of than
dam-a
linear degradation model due to its curvature. From the
age/pathological degradation, there is always an AI that can provide us with a snapshotauthors’ point of view, it is not
recommended to use a linear model: first, because underestimating
of the current damage state), and, in this time frame, it is less relevant to estimate the RUL; the RUL is beneficial for
structural safety, and second, because it is not possible to know
analogously, in the last moments of the structure life, it is obviously useless to estimatethe shape of the degradation
function
the RUL,asince priori.theThus,
damageif we identification
use a linear degradation
AI will report model and theclose
damage degradation function
to the imposed
turns out to be exponential, we would overestimate the RUL to the detriment of safety. At
threshold. Instead, in the intermediate life span, it is important to make an accurate RUL
the same time, the introduction of the penalty matrix in the AI training phase leads to a
estimate in order to detect and locate future damage and have enough time to effectively
higher probability of overestimating damage, which in turn leads to an underestimation of
plan maintenance actions.
the RUL.
However, it is interesting to note that the estimated RUL is practically always lower
One aspect that could have a strong impact on RUL is the discontinuity of the damage
than the actual RUL, i.e., it underestimates the actual RUL. From a structural safety per-
estimated by the AI, as shown and discussed in Figure 16, due to the solution of a classifi-
spective, it is certainly better to be in this state than in the opposite state, i.e., overestimat-
cation problem. Future developments in this regard could include methods to reduce the
ing the RUL. From a mathematical point of view, this is due to two factors: the degradation
discontinuity of damage estimation so that a more faithful reconstruction of the degradation
model used is exponential, and a penalty matrix was used in the AI training. The expo-
history is available. Another solution could be to implement techniques for interpolating
nential degradation model, even if updated, will always provide a shorter remaining life
or averaging the damage classes resulting from AI to facilitate the RUL update algorithm.
than aAlinear degradation model due to its curvature. From the authors’ point of view, it
hypothesis on the possible use of RUL calculation throughout the life cycle of the
is not recommended
structure is presented to in
useFigure
a linear18.model:
All thefirst, because developed
techniques underestimating the RUL isinben-
and presented this
eficial for structural safety, and second, because it is not
paper are designed for estimating the remaining useful life. Once the construction possible to know the shape of of the
the
degradation function aand
structure is completed priori.
the Thus, if we use
monitoring a linear degradation
infrastructure is properly model
installed,andit the degra-
is possible
dation
to updatefunction
the FEturns
model outbased
to beonexponential,
the recorded wedata
would (FEoverestimate
model, which the RULend
could to the
up detri-
being
ment of safety. At the same time, the introduction of the
the same model used for the design of the structure). In this way, we obtain the FE penalty matrix in the AI training
model
phase
needed leads to aAI
for the higher probability
training. This FEofmodeloverestimating damage, be
will undoubtedly whichmore inaccurate
turn leads to an
than an
underestimation of the RUL.
outdated model as a result of the optimization problem. In this way, we will minimize
One aspect
the model error that
in thecould have
AI that a strong
will impact
be trained on the
with RULsignals
is the discontinuity
derived from of theFE
this damage
model.
estimated by the AI, as shown and discussed in Figure 16, due to the solution of a classi-
fication problem. Future developments in this regard could include methods to reduce
the discontinuity of damage estimation so that a more faithful reconstruction of the deg-
radation history is available. Another solution could be to implement techniques for
 structure is presented in Figure 18. All the techniques developed and presented in this
paper are designed for estimating the remaining useful life. Once the construction of the
structure is completed and the monitoring infrastructure is properly installed, it is possi-
ble to update the FE model based on the recorded data (FE model, which could end up
being the same model used for the design of the structure). In this way, we obtain the FE
Appl. Sci. 2023, 13, 4646 16 of 20
model needed for the AI training. This FE model will undoubtedly be more accurate than
an outdated model as a result of the optimization problem. In this way, we will minimize
the model error in the AI that will be trained with the signals derived from this FE model.
Once
Once the
thetraining
trainingphase
phaseis is
complete, wewe
complete, willwill
have an AI
have anthat can estimate
AI that the damage
can estimate (with
the damage
some
(with accuracy and therefore
some accuracy uncertainty
and therefore in the estimate)
uncertainty and weand
in the estimate) canwe
usecan
this information
use this infor-
as inputas
mation for estimating
input the remaining
for estimating useful life.
the remaining Atlife.
useful thisAt
point,
this every
point, time
everywe perform
time we per- a
registration, we willwe
form a registration, estimate the damage
will estimate over the
the damage previously
over trainedtrained
the previously AI andAI iteratively
and iter-
update
atively the RULthe
update calculation.
RUL calculation.

Figure 18. Development

Figure Developmentprospective
prospectiveforfor
thethe
calculation of RUL
calculation overover
of RUL the entire life cycle
the entire of theof
life cycle struc-
the
ture.
structure.

Finally,
Finally,the
thepossibility
possibilityofof extending
extending thethe
proposed
proposed framework
framework to real
to structures involves
real structures in-
additional considerations. First, the effects of environmental conditions
volves additional considerations. First, the effects of environmental conditions may chal- may challenge
feature extraction
lenge feature compared
extraction to synthetic
compared signals.signals.
to synthetic Second,Second,
the measured data willdata
the measured be affected
will be
by measurement
affected errors due
by measurement to thedue
errors accuracy
to the of the sensor.
accuracy This
of the couldThis
sensor. affect the goodness
could affect the
of the proposed
goodness features and,
of the proposed if thisand,
features goodness is compromised,
if this goodness new features
is compromised, new must be
features
defined, or the existing
must be defined, ones must
or the existing onesbe improved.
must In fact,
be improved. this this
In fact, is not a major
is not a major problem
problem if
high-performance sensors, measurement systems and data transmission
if high-performance sensors, measurement systems and data transmission infrastructures infrastructures
that
that are
are used.
used. Assuming
Assuming that that this
this framework
framework is is widely
widely used,
used, one
one would
would needneed to to address
address
the computational power required for AI training, especially with respect
the computational power required for AI training, especially with respect to the creation to the creation of
the
of thevarious
varioussynthetic signals
synthetic signalsthrough
through dynamic
dynamic FE FE
analysis,
analysis,which
which is the realreal
is the bottleneck
bottleneck of
the whole
of the wholeprocess. Assuming
process. Assuming that that
the random and parametric
the random and parametrictraining in Figure
training 12, which
in Figure 12,
was
which performed for research
was performed purposes
for research and to
purposes andunderstand
to understand the the
effectiveness
effectiveness of Bayesian
of Bayes-
optimization, should be avoided and, instead, once the signals
ian optimization, should be avoided and, instead, once the signals are generated, are generated, we shouldwe
proceed with thewith
should proceed training
the through
training hyperparameter
through hyperparameteroptimization, the question
optimization, the arises
questionas
to how much computational power or how much time is required
arises as to how much computational power or how much time is required to generate to generate signals of
asignals
structure with multiple types of damage and different severity levels.
of a structure with multiple types of damage and different severity levels. As a As a simplified
estimate,
simplifiedwe can easily
estimate, determine
we can the number
easily determine the of analyses
number and thusand
of analyses thethus
computational
the compu-
time required.
tational time required.
Indeed, it is possible to calculate the total computational time with the following
Indeed, it is possible to calculate the total computational time with the following for-
formula:
mula:
TT = SAT · N A ID · D 0n,k (3)
𝑇𝑇 = 𝑆𝐴𝑇 ⋅ 𝑁𝐴 ⋅ 𝐷′ , (3)
where:
TT represents the total time to create the synthetic signals
where:
SAT represents the time to perform a single analysis with the FE model
NAID represents the number of analyses to be performed for the same damage configuration
to account for the variability of stochastic parameters
D 0n,k is the number of the arrangements with repetition where: n is the number of damage
severities and k is the number of the different types of damage to be investigated
When calculating analysis time for real-world use cases, it is possible to see how much
time is required for complex civil structures that easily reaches the order of hundreds, if not
thousands, of years. However, this is where methods to increase computational power, such
as parallel computing and cloud computing, come to the rescue. With these techniques, it is
possible to divide a large problem into smaller problems that can be developed and solved
in parallel. In fact, these techniques are easy to implement in the proposed framework, since
Appl. Sci. 2023, 13, 4646 17 of 20

each analysis is independent of the others. Even if, at first sight, the results just reported
give the impression that the application of this methodology is difficult, it is necessary to
consider the various advantages that can be obtained.

9. Conclusions
This paper proposed and investigated a framework for estimating structural damage
using machine learning techniques (ML) and the resulting remaining and useful life (RUL).
The applied ML technique involves supervised training, so knowledge of the data for the
undamaged and especially for the damaged structural system is required. Considering
the uniqueness of civil structural systems, training data for the different damage classes
are not available. Therefore, a hybrid-based approach was adopted here: a FE model is
used as a digital twin to generate synthetic damage data. The proposed, implemented,
and numerically tested framework seems to be able to estimate, quantify, and localize the
damage (in this case the corrosion thickness of a reference steel column) in a robust, reliable
way and with very good accuracy.
A reliable estimate of actual damage is a necessary prerequisite for RUL assessment.
In this context, a framework for updating the RUL of a structural system has been proposed
and investigated. To achieve this, a Bayesian approach is adopted for updating the parame-
ters of the chosen exponential degradation model with random coefficients. After updating
the degradation model, it is possible to determine the updated a posteriori distribution of
RUL of the structure. Finally, the framework was numerically applied to the reference case
study, with the trained AI serving as the basis for the probabilistic estimation of the current
state of health.
The main scientific contributions in this research are the development of an innovative
framework that allows a damage estimation AI to interoperate with the probabilistic RUL
estimation model. The application of the framework to a numerical, albeit simple, structural
system allowed validating the soundness of the method. Therefore, the added value of
performing a Bayesian update of the parameters to properly incorporate the information
into the RUL estimation model is significant.
In solving a supervised learning problem, such as the one presented in this paper,
a fundamental and crucial aspect becomes clear: do we have data about the damaged
structure (for all types and severities of damage)? If the answer is no, which is usually the
case, one needs to find a way to estimate it. In this paper, we attempted to solve a supervised
learning problem by developing a general framework that uses numerical FE modeling to
produce synthetic data of the monitored structure. Of course, these data will be affected by
model errors due to the various assumptions made in the development of the model itself.
For this reason, we refer to this technique as hybrid-based: an AI model is created from the
data all or part of which can be artificially generated by numerical simulations. To limit, but
not to eliminate model errors, finite element model update (FEMU) techniques can be used
to update the FE model before generating the synthetic data. In this way, model errors are
minimized for a newly built structure, but, more importantly, the developed ML method
could be used to monitor existing structures, thus already showing some degradation
(whether physiological or pathological), in which case it would be strongly discouraged to
proceed with AI training with a model representing the undamaged structure.
Another important point to consider when monitoring a structural system is its sen-
sitivity to environmental conditions (e.g., temperature and humidity sensitivity, among
others). Such conditions can cause features to change in a way that exceeds the change
caused by potential damage. For civil engineering structures, such as bridges, mass can
also have a significant effect on the modal properties of the infrastructure. Indeed, for
these structures, variable loads and thus variable masses could be an important part of
the loading on the whole structure. In the proposed framework, these random conditions
are integrated into the mass parameter, which is randomly chosen and not included as
a known parameter in the learning process. In this way, the AI is forced to distinguish a
Appl. Sci. 2023, 13, 4646 18 of 20

change in features due to uniform corrosion damage (i.e., a reduction in stiffness), as much
as possible, from the change in features associated with random conditions.
Another aspect in favor of using techniques, such as the one presented here, is the
possibility of in situ processing. Among other things, in situ processing has advantages
in terms of speed of processing and reduction of the amount of data sent. Namely, while
an enormous amount of computational time is required for training and especially for the
creation of synthetic signals, the resulting AI after the training process, on the other hand,
is very light and easy to process, even for very inexpensive CPUs. In summary, the use
of this framework could lead to significant benefits in terms of accuracy and reliability of
structure health estimations. Assuming that the development is extended to other types
of materials and structures and tested in practice, the only limitation in its use could be
the enormous computational cost. While on one hand, growth in available computational
power is expected, it is undeniable how easy it is to configure an AI with a disproportionate
need for signals. In this context, a simple equation has been reported that can be used
to estimate the time required to generate the necessary synthetic data at a given power
level. Alternatively, the required power (and, hence, the most appropriate engine) can
be estimated as a function of the desired AI accuracy, which depends on the number of
damage types and the number of damage severity levels selected.
Future research topics arising from these considerations could include improve-
ment/refinement of the framework, numerical application to more complex structures,
application to real structures that are also subject to changes in environmental condi-
tions, and integration with decision models to determine the optimal timing for mainte-
nance interventions.

Author Contributions: Conceptualization, S.C. and A.B.; methodology, S.C.; software, S.C.; valida-
tion, S.C. and A.B.; formal analysis, S.C.; writing—original draft preparation, S.C.; writing—review
and editing, S.C. and A.B.; visualization, S.C. and A.B.; supervision, A.B. All authors have read and
agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Data are available upon a reasonable request.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Farrar, C.; Worden, K. Structural Health Monitoring: A Machine Learning Perspective; John Wiley & Sons: Hoboken, NJ, USA, 2012.
2. Worden, K.; Manson, G.; Fieller, N. Damage detection using outlier analysis. J. Sound Vib. 2000, 229, 647–667. [CrossRef]
3. Figueiredo, E.; Park, G.; Farrar, C.; Worden, K.; Figueiras, J. Machine Learning algoritms for damage detection under operational
and environmental variability. Strcutural Health Monit. 2011, 10, 559–572. [CrossRef]
4. Zang, Y.; Burton, H.; Sun, H.; Shokrabadi, M. A machine learning framework for assessing post–earthquake structural safety.
Struct. Saf. 2018, 72, 1–16. [CrossRef]
5. Abdeljaber, O.; Avei, O.; Kiranyaz, S.; Gabbouj, M.; Inman, D.J. Real–time vibration based structural damage detection using
one–dimensional convolution neurla network. J. Sound Vib. 2017, 388, 154–170. [CrossRef]
6. HoThu, H.; Mita, A. Damage detection method using support vector machine and first three natural frequencies for shear
structures. Open J. Civ. Eng. 2013, 3, 32514. [CrossRef]
7. Mita, A.; Hagiware, H. Quantitative damage diagnosis of shear structures using support vector machine. J. Civ. Eng. 2003, 7,
683–689. [CrossRef]
8. Gui, G.; Pan, H.; Lin, Y.; Yuan, Z. Data–driven support vector machine with optimization techniques for structural health
monitoring and damage detection. J. Civ. Eng. 2017, 21, 523–534. [CrossRef]
9. Chong, J.W.; Kim, Y.; Chon, K.H. Nonlinear multiclass support vector machine–based health monitoring system for buildings
employing magnetorheological dampers. J. Intell. Mater. Syst. Struct. 2014, 25, 1456–1468. [CrossRef]
10. Bornn, L.; Farrar, C.R.; Park, G.; Farinholt, K. Structural health monitoring with autoregressive support vector machine. J. Vib.
Acoust. 2009, 131, 021004. [CrossRef]
Appl. Sci. 2023, 13, 4646 19 of 20

11. Kim, Y.; Chong, J.W.; Chin, K.H.; Kim, J. Wavelet–based AR–SVM for health monitoring of smart structures. Smart Mater. Struct.
2012, 22, 015003. [CrossRef]
12. Sajedi, S.O.; Liang, X. A data framework for near reall–time and robust damage diagnosis of building structures. Struct. Control
Health Monit. 2019, 27, e2488.
13. Momeni, M.; Bedon, C.; Hadianfard, M.A.; Baghalani, A. An Efficient Reliability–Based Approch for Evaluation Safe Scaled
Distance of Steel Columns under Dynamic Blast Loads. Buildings 2021, 11, 606. [CrossRef]
14. Momeni, M.; Hadianfard, M.A.; Bedon, C.; Baghlani, A. Damage evaluation of H–section steel columns under impulsive blast
loads via gene expression programming. Eng. Struct. 2020, 219, 110909. [CrossRef]
15. Huang, M.; Gul, M.; Zhu, H. Vibration–based Structural Damage Identification under Varying Temperature Effects. J. Aerosp. Eng.
2018, 31, 04018014. [CrossRef]
16. Huang, M.; Lei, Y.; Li, X.; Gu, J. Damage identificaiton of Bridge Structures Considering Temperature Variations–Based SVM and
MFO. J. Aerosp. Eng. 2021, 34, 04020113. [CrossRef]
17. Luo, j.; Huang, M.; Xiang, C.; Lei, Y. Bayesian damage identification based on autoregressive model and MH–PSO hybrid MCMC
sampling method. J. Civ. Struct. Health Monit. 2022, 12, 361–390. [CrossRef]
18. Luo, J.; Huang, M.; Xiang, c.; Lei, Y. A Novel Method for Damage Identification Based on Tuning–Free Strategy and Simple
Population Metropolis–Hasting Algorithm. Int. J. Struct. Stab. Dyn. 2022, 23, 2350043. [CrossRef]
19. Kamariotis, A.; Chatzi, E.; Straub, D. Value of information from vibration–based structural health monitoring extracted via
Bayesian model updating. Mech. Syst. Signal Process. 2022, 166, 108465. [CrossRef]
20. Martakis, P.; Reuland, Y.; Stavridis, A.; Chatzi, E. Fusing damage–sensitive features and domain adaptation towards robust
damage classification in real buildings. Soil Dyn. Earthq. Eng. 2023, 166, 107739. [CrossRef]
21. Tronci, E.M.; Beigi, H.; Feng, M.Q.; Betti, R. A transfer learning SHM strategy for bridges enriched by the use of speaker
recognition x–vectors. J. Civ. Struct. Health Monit. 2022, 12, 1285–1298. [CrossRef]
22. Mariniello, G.; Pastore, T.; Menna, C.; Festa, P.; Asprone, D. Structural damage detection and localization using decision tree
ensemble and vibration data. Comput. Aided Civ. Inf. 2021, 36, 1129–1149. [CrossRef]
23. Li, L.; Morgantini, M.; Betti, R. Structural damage assessment through a new generalized autoencoder with features in the
quefrency domain. Mech. Syst. Signal Process. 2023, 184, 109713. [CrossRef]
24. Castelli, S.; Belleri, A.; Riva, P. Machine Learning technique for the diagnosis of environmental degradation in a steel structure.
In Proceedings of the SHMII 2021: 10th International Conference on Structural Health Monitoring of Intelligent Infrastructure,
Porto, Portugal, 30 June–2 July 2021; pp. 741–747.
25. Mathworks. MATLAB R2020a. Mathworks: Natick, MA, USA, 2020.
26. Hastie, T.; Tibshirani, R.; Friedman, J. The Element of Statistical Learning: Data Minimg, Inference, and Prediction; Springer:
Berlin/Heidelberg, Germany, 2009.
27. Crisitanini, N.; Shawe–Taylor, J. An Introduction to Support Vector Machines and Other Kernel–Based Learning Methods; Cambridge
University Press: Cambridge, UK, 2000.
28. Fan, R.E.; Chen, P.H.; Lin, C.J. Working set selection using second order information for training support vector machines. J.
Mach. Learn. Res. 2005, 6, 1889–1918.
29. Coleman, T.F.; Li, Y. A reflective newton method for minimizing a quadratic function subject to bounds on some of the variables.
SIAM J. Optim. 1996, 6, 1040–1058. [CrossRef]
30. Gill, P.E.; Murray, W.; Wright, M.H. Pratical Optimisation; SIAM: Philadelphia, PA, USA, 2019.
31. Kecman, V.; Huang, T.M.; Vogt, M. Iterative single data algorithm for training kernel machines from huge data sets: Theory and
performance. In Support Vector Machines: Theory and Applications; Springer: Berlin/Heidelberg, Germany, 2005; pp. 255–274.
32. Gould, N.; Toint, P.L. Preprocessing for quadratic programming. Math. Program. 2004, 100, 95–132. [CrossRef]
33. McKenna, F. OpenSees: A Framework for Earthquake Engineering Simulation. Comput. Sci. Eng. 2011, 13, 58–66. [CrossRef]
34. Brincker, R.; Zhang, L.; Andersen, P. Modal identification of output–only systems using frequency domain decomposition. Smart
Mater. Struct. 2001, 10, 441. [CrossRef]
35. Gelbart, M.A.; Snoek, J.; Adams, P. Bayesian optimization with unknown constraints. arXiv 2014, arXiv:1403.5607.
36. Rasmussen, C.E. Gaussian processes in machine learning. Summer Sch. Mach. Learn. 2003, 3176, 63–71.
37. Nocedal, J.; Wright, S. Numerical optimization.; Springer Science & Business Media: Berlin/Heidelberg, Germany, 2006.
38. Lagarias, J.C.; Reeds, J.A.; Wright, M.H.; Wright, P.E. Convergence properties of the nelder–mead simplex method in low
dimensions. SIAM J. Optim. 1998, 9, 112–147. [CrossRef]
39. Snock, J.; Larochelle, H.; Adams, R.P. Practical bayesian optimization of machine learning algorithms. In Advances in Neural
Information Processing Systems; MIT Press: Cambridge, MA, USA, 2012; Volume 25.
40. Bull, A.D. Convergence rates of efficient global optimization algorithms. J. Mach. Learn. Res. 2011, 12, 2879–2904.
41. Chakraborty, S.; Gebraeel, N.; Lawley, M.; Wan, H. Residual–life estimation for components with non–symmetric priors. IIE Trans.
2009, 41, 372–387. [CrossRef]
42. Gebraeel, N.Z.; Lawley, M.A.; Li, R.; Ryan, J.K. Residual–life distributions from component degradation signals: A bayesian
approach. IIE Trans. 2005, 37, 543–557. [CrossRef]
43. Lu, C.J.; Meeker, W.Q. Using degradation measures to estimate a time–to–failure distribution. Technometrics 1993, 35, 161–174.
[CrossRef]
Appl. Sci. 2023, 13, 4646 20 of 20

44. Wang, W. A model to determine the optimal critical level and the monitoring intervals in condition–based maintenance. Int. J.
Prod. Res. 2000, 38, 1425–1436. [CrossRef]
45. Nelson, W. Accelerated Testing Statistical Models, Test Plans and Data Anlysis; Wiley: Hoboken, NJ, USA, 1990.
46. Shao, Y.; Nezu, K. Prognosis of remaining bearing life using neural networks. Proc. Inst. Mech. Eng. Part I J. Syst. Control Eng.
2000, 214, 217–230. [CrossRef]
47. Ahmad, M.; Sheikn, A. Bernstein reliability model: Derivation and estimation of parameters. Reliab. Eng. 1984, 8, 131–148.
[CrossRef]
48. Gebraeel, N. Sensory–updated residual life distributions for components with exponential degradation patterns. IEEE Trans.
Autom. Sci. Eng. 2006, 3, 382–393. [CrossRef]
49. Pedeferri, P. Corrosione e Protezione Dei Materiali Metallici; Polipress: Michałowice, Poland, 2010. (In Italian)

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.