Infrastructures 09 00040 v2
Infrastructures 09 00040 v2
Article
Enhancing the Damage Detection and Classification of Unknown
Classes with a Hybrid Supervised–Unsupervised Approach
Lorenzo Stagi 1, *, Lorenzo Sclafani 2 , Eleonora M. Tronci 3 , Raimondo Betti 4 , Silvia Milana 1 , Antonio Culla 1 ,
Nicola Roveri 1 and Antonio Carcaterra 1
1 Department of Mechanical and Aerospace Engineering, Sapienza University of Rome, 00184 Rome, Italy;
[email protected] (S.M.); [email protected] (A.C.); [email protected] (N.R.);
[email protected] (A.C.)
2 Department of Physics and Astronomy “G. Galilei”, University of Padua, 35131 Padua, Italy;
[email protected]
3 Department of Civil and Environmental Engineering, Northeastern University, Boston, MA 02115, USA;
[email protected]
4 Department of Civil Engineering and Engineering Mechanics, Columbia University,
New York, NY 10027, USA; [email protected]
* Correspondence: [email protected]
Abstract: Most damage-assessment strategies for dynamic systems only distinguish between undam-
aged and damaged conditions without recognizing the level or type of damage or considering unseen
conditions. This paper proposes a novel framework for structural health monitoring (SHM) that
combines supervised and unsupervised learning techniques to assess damage using a system’s struc-
tural response (e.g., the acceleration response of big infrastructures). The objective is to enhance the
benefits of a supervised learning framework while addressing the challenges of working in an SHM
context. The proposed framework uses a Linear Discriminant Analysis (LDA)/Probabilistic Linear
Discriminant Analysis (PLDA) strategy that enables learning the distributions of known classes and
the performance of probabilistic estimations on new incoming data. The methodology is developed
Citation: Stagi, L.; Sclafani, L.; Tronci, and proposed in two versions. The first version is used in the context of controlled, conditioned
E.M.; Betti, R.; Milana, S.; Culla, A.;
monitoring or for post-damage assessment, while the second analyzes the single observational data.
Roveri, N.; Carcaterra, A. Enhancing
Both strategies are built in an automatic framework able to classify known conditions and recognize
the Damage Detection and
unseen damage classes, which are then used to update the classification algorithm. The proposed
Classification of Unknown Classes with
framework’s effectiveness is first tested considering the acceleration response of a numerically sim-
a Hybrid Supervised–Unsupervised
Approach. Infrastructures 2024, 9, 40. ulated 12-degree-of-freedom system. Then, the methodology’s practicality is validated further by
https://doi.org/10.3390/ adopting the experimental monitoring data of the benchmark study case of the Z24 bridge.
infrastructures9030040
Keywords: structural health monitoring; damage detection; cepstral coefficients; probabilistic linear
Academic Editors: Carlo Rainieri,
discriminant analysis; Z24 bridge
Ahmet Emin Aktan, Branko Glisic and
Ivan Bartoli
permanently affect its performance. This includes, for example, intentional or unintentional
changes to the boundary conditions and system connectivity, reductions in the stiffness of
single or multiple system elements, or the loss of strength in supporting components [1].
The concept of damage involves a comparison between the reference, healthy condition
of a system and an unknown, potentially damaged one. Such a comparison enables the
identification of changes or deviations in a system’s properties and performance. Therefore,
monitoring a system’s condition over time is crucial to detect any signs of this deviation
and take appropriate actions to prevent further deterioration or catastrophic consequences.
Among the various strategies in SHM [2], the vibration-based approaches are among
the most appealing [3,4]. They concentrate on gathering data directly from the dynamic
response of the investigated system responses (e.g., displacement or acceleration), often
without dealing with the construction of a model of the system or having prior knowl-
edge. By focusing on the global response of the system, these methods have proven to be
particularly effective for systems that have complex geometries or are made of materials
that are not homogeneous. Given the collection of the system’s response via the use of
sensors directly installed onto the structure, it is often possible to conduct short- or long-
term monitoring campaigns, enabling real-time monitoring and online evaluations of any
potential damages.
Such continuous monitoring of a system can be performed using either model-based
or signal-based approaches. The model-based approach involves creating a physical or
mathematical model that represents the dynamic behavior of the real system [5,6]. Real-time
data are then used to update the parameters of the model so that its response accurately
reproduces the measured one. The damage can be detected through a comparison between
the data simulated via modeling and the real-time experimental data of the original system.
However, while model-based strategies work well with simple systems, they become less
reliable when dealing with articulated ones characterized by complex dynamics which are
challenging to model [7].
On the other hand, recently developed machine learning tools have enhanced the
opportunity to bypass the creation of complex and sophisticated models and to extract
damage-sensitive parameters directly from the monitored dynamic response [8,9]. Such
parameters are generally referred to as damage-sensitive features (DSFs) [10]. In this
context, damage detection can be considered a problem of statistical pattern recognition,
where the goal is to identify and characterize different behavioral and dynamic patterns in
the DSFs [11], some of which are potentially associated with damage occurrence [12].
Selecting and estimating a reliable DSF is essential for the implementation of a suc-
cessful statistical pattern recognition framework. The selected feature ought to be strongly
correlated with the structural properties of the system, to detect structural changes due
to the effect of damage, while being insensitive to changes due to noise or variations in
operational and environmental conditions [13–16]. Modal parameters (e.g., natural frequen-
cies or mode shapes [17]) and Auto Regressive model coefficients [18,19] are examples of
popular and common DSFs often employed for structural health monitoring applications.
However, in the last decade, new features, commonly used in a different domain, have
been introduced and successfully implemented in damage detection problems: cepstral
coefficients (CCs). These coefficients were first presented in 1962 by Bogert et al. in an
attempt to coin a method able to identify the presence of an echo in a sound signal [20].
Oppenheim [21] addressed the discrete-time formulation of the cepstrum and its complex
counterpart, the complex cepstrum. Despite such features being commonly used in acous-
tics [22,23], they have been successfully implemented in different forms and fields. They
have been particularly successful in mechanical systems as features for detecting damage
in bearing and gearboxes [24–26]. Several studies address them as features for monitoring
the health conditions of civil and infrastructural systems [16,27–30], demonstrating how
these features might represent valid alternatives to the most common approaches. In 2021,
Morgantini et al. [31] investigated their connection to the modal characteristics of the struc-
ture and how they can be successfully used in a damage assessment strategy. The use of
Infrastructures 2024, 9, 40 3 of 24
power cepstral coefficients as DSFs was further refined by Li et al. [32] in 2023. Li proposed
a New Generalized Auto-Encoder integrated with a statistical-pattern-recognition-based
approach that uses the cepstral coefficients extracted from structural acceleration responses
as DSFs. It has been proven that the main advantages of the cepstral features consist of
their natural dimensionality reduction and their simple and direct extraction process, from
which follows a lower required user expertise and a lower computational burden.
This study addresses a supervised damage detection and classification framework
using cepstral features as damage-sensitive features. Supervised learning is a machine
learning paradigm in which labeled data are used to train classification or regression al-
gorithms [33–35]. In the context of structural health monitoring and damage detection,
supervised learning methods face limitations because they require information on both
normal and damaged conditions during the learning process. For this reason, unsupervised
approaches are usually preferred for these kinds of applications [36–40]. However, except
for dynamic systems in a controlled environment, it is often impossible to have sufficient
observations from the system in all the possible damage scenarios that could occur. The
model would likely be trained on insufficient information and misclassify unknown dam-
aged conditions. In particular, these data are rarely available for civil engineering structures
for which it is neither reasonable nor cost-effective to intentionally damage them to provide
training data [41].
The objective of this paper is to enhance the benefits of a supervised learning frame-
work and address the challenges that arise while implementing these strategies in a struc-
tural health monitoring (SHM) context. The proposed approach is a new and innovative
method that combines supervised tools within an unsupervised process. This method
can automatically identify and classify unknown damage conditions and use them to
strengthen the existing knowledge about the health condition of the system. The proposed
framework is based on the use of a Linear Discriminant Analysis (LDA) [42]/Probabilistic
Linear Discriminant Analysis (PLDA) [43] strategy that allows one to learn the distributions
of known classes and to perform probabilistic estimations on new incoming acceleration
data. The combination of dimensionality reduction via LDA and the probabilistic learn-
ing framework of PLDA applied on cepstral features has shown promising outcomes for
speaker recognition applications and damage detection in civil systems [29,44,45]. The
methodology is further developed here in an unsupervised hybrid approach and proposed
in two versions. The first can be used in the context of controlled condition monitoring or
for post-damage assessment in the case of well-known classes. The second is built in an
automatic framework, where unknown damage classes, different from the healthy reference
state, are recognized, classified, and used to update the classification algorithm via PLDA.
The sections below describe the proposed framework for damage detection and clas-
sification and its application. Section 2 presents the structural features adopted for this
work and the theoretical background behind LDA and PLDA techniques, followed by the
detailed functioning of the proposed algorithm. In Section 3, a numeric case study is pre-
sented to affirm the algorithm’s effectiveness, while Section 4 shows real-world application
results using the Z24 bridge experimental dataset.
2. Methodology
The proposed methodology follows a statistical pattern recognition framework to
detect and classify damage. It does so by comparing the cepstral features extracted from
dynamic responses of the system in an unknown state against those extracted from the
system in a healthy condition. Firstly, the methodology focuses on the extraction of
the CCs from the acceleration responses of the monitored system, and then it creates a
classification model based on the statistical distribution of the CCs representative of the
healthy condition (training phase). Once the model is created, new incoming CCs, extracted
from the system in an unknown state, are compared to the training ones (testing phase)
and if the distributions do not match, the incoming data are labeled as damaged.
Infrastructures 2024, 9, 40 4 of 24
where F and F −1 are the discrete Fourier transform and the inverse discrete Fourier trans-
form, respectively. The simplicity of the extraction process grants optimal computational
efficiency, suitable for real-time monitoring or large-scale applications, as well as requiring
lower user expertise, as the decision-making is limited to the number of coefficients to
extract. These reasons make cepstral coefficients appealing alternatives to the more com-
monly used DSFs. However, a drawback is that when using such features, the intuitive
physical relationship between these coefficients and the system’s structural properties is
less apparent, making interpreting the coefficients less intuitive.
From the analytical formulation, for which the reader is referred to [31], it is possible to
better understand the connection to the structural properties as well as the input excitation
and localization of the sensor. In particular, thanks to the presence of the logarithm in
Equation (1), the coefficients are described as a sum of terms that separate the different
contributions.
Figure 1 outlines in five essential steps the extraction process of the CCs from an input
sampled signal x [n] (acceleration, for example). Each signal first undergoes a phase of
framing, where the signal is divided into equal-length frames, followed by a windowing
step, where a Hamming window is applied to each frame to reduce riddle effects at the
frame’s onset and offset. From such windows it is then possible to extract the cepstral
coefficients simply by calculating the log-power spectrum and, subsequently, applying
the inverse discrete Fourier transform. The output of this process is a set of correlated
coefficients characterizing the frequency content of the system in a concentrated time-
dependent form. Because the first coefficient has proven to be particularly sensitive to the
input force, it is a common rule to neglect it to highlight the contribution related to damage.
The extraction process highlights another important property of these coefficients. By
analyzing small signal frames sequentially, these coefficients are capable of capturing varia-
tions caused by factors like input changes or evolving system properties. This makes them
suitable for handling non-stationary signals with varying frequency components, mak-
ing them an attractive option for describing systems characterized by complex dynamics
or inputs.
Infrastructures 2024, 9, 40 5 of 24
x[n], n = 0,...,N-1
c[d], d = 0,...,D-1
SW = ∑ ∑ (⃗xi − m ⃗ k )T
⃗ k )(⃗xi − m SB = ∑ nk (⃗mk − m ⃗ )T
mk − m
⃗ )(⃗ (2)
k i ∈Ck k
where m⃗ k and m⃗ represent the mean vector of class k and the mean vector of the entire
dataset, respectively. LDA aims to find the linear transformation ⃗x −→ WT⃗x that maximizes
the between-class variance relative to the within-class variance. It can be shown that such a
hyperplane can be found by solving the following eigenproblem:
⃗ = ΛSW w
SB w ⃗, (3)
The desired coordinate system is defined by the eigenvectors associated with the
largest eigenvalues. The dimensionality of the latent space can be chosen by selecting
the desired number g of components to keep. Λ is the diagonal matrix that describes the
change of coordinates from the original system to the latent space.
Therefore, LDA is trained on a set of known classes and finds the lowest dimensional
space that enhances the separation between the classes. Once new incoming data arrive,
they are projected into the subspace obtained from the training data. At this point, classifi-
cation of the data can be performed. LDA could be also used for classification by computing
the distance in the latent space between the data point and the centroids of the clusters and
associating the new data to the nearest class. However, such a deterministic approach fails
when the incoming data are part of a previously unseen class. In order to recognize unseen
classes, a new, more flexible model is needed; one way to approach this problem is to use a
probabilistic approach by adopting Probabilistic LDA.
To better understand how PLDA works, let us consider an example ⃗x in the data
space and the corresponding class center ⃗y. The PLDA model can be represented by the
following expression:
⃗x = m⃗ + A⃗u, (4)
where ⃗u (⃗u ∼ N (·|⃗v, I )) represents the example of the class in the projected space, and
⃗v (⃗v ∼ N (·|0, Ψ)) represents the class itself in the same latent space (Ψ is the covariance
matrix in the LDA-transformed latent space). Therefore, the example ⃗x in the original data
space is related to its latent representation ⃗u via an invertible transformation A.
Once the data have been projected into the latent space, PLDA computes the probabil-
ity that the data come from one of the known classes by computing the probability that
the data point belongs to the normal distribution that describes the class. Lastly, PLDA
provides the log-likelihood ratio, also known as the PLDA score. To evaluate whether a
point belongs to one of the known classes, the likelihood score is defined as the logarithm
of R, the ratio between the likelihood of the point to be part of the same class and the
likelihood to be of a different class.
likelihood(same class)
R= (5)
likelihood(different classes)
Based on the log( R) score, it is possible to compare each new example with all the
other examples; a PLDA score matrix is defined for each tested observation which is
assigned to the class with the higher score.
2.2. Training
The first step of the proposed strategy involves the creation of a statistical model
characterizing the system in its undamaged condition. The definition of such an initial
model is obtained during the training phase, which consists of the following steps (Figure 2):
(1) the extraction of the cepstral coefficients for all the observations in all the known classes
characterizing the system; (2) a dimensionality reduction phase via the application of LDA;
(3) the training of the PLDA classification model using the LDA projected components and
the definition of the distribution of the PLDA scores for each class; (4) the identification of
the thresholds characterizing the PLDA scores for the known classes.
Therefore, the first parameter to define is the number and type of classes used in
the training phase to characterize the system’s health conditions. Since LDA finds the
coordinate system in which classes are separated as much as possible (i.e., minimizing the
within-class scatter while maximizing the variance between classes), it requires at least two
classes to operate.
The strength of the proposed methodology lies in its capacity to autonomously rec-
ognize classes that may not have been encountered during the training phase. However,
it does require a minimum of two known classes during training as a starting reference
condition. This assumption necessitates a more detailed characterization of the system’s
health condition beyond a single undamaged reference state. In the context of monitoring
civil structural or mechanical systems, this is a reasonable hypothesis as the structure in
its healthy condition can be further characterized according to different operational and
environmental conditions that change the system’s dynamic behavior. Alternatively, it is
possible to consider, in addition to the healthy state, low-intensity damaged conditions
that could effectively characterize structures after a few years of operational use. In the
present study, lower damage levels were considered additional classes for training the
model, aiming to simulate real-life scenarios where structures encounter progressively
severe damage, starting from minor damage. The identification and separation of this
second class from the reference undamaged scenario can be achieved, when the information
is not known beforehand, through a novelty detection unsupervised strategy that captures
deviations in the chosen damage-sensitive features from this initial reference state. Once
the second class is identified, it can be used in combination with the reference undamaged
Infrastructures 2024, 9, 40 7 of 24
condition to initialize the proposed LDA/PLDA framework. At the same time, the number
of components of the latent space is another crucial parameter for the simulation as it
enhances class separation. The proposed algorithm autonomously evaluates the relevant
components of LDA that effectively separate classes, following the procedure outlined
in [43].
Once the training classes are chosen, 70% of each class’s data are picked randomly
to be used to train the algorithm, while the remaining 30% are used later for testing. The
cepstral coefficients are then extracted from the training data and employed to define the
initial latent space using the LDA procedure. At this point, it is possible to calculate the
initial thresholds of the training classes, which is the starting point for the testing phase.
The computation of the thresholds is based on the analysis of the scores associated with the
likelihood of training data belonging to their respective classes. By modeling these scores
with the distribution that best fits the data, it is possible to define a limit value that enables
the assessment of whether a new point belongs to a particular class.
To make things clear, let us consider two classes for the training phase. The algorithm
computes the likelihood scores, hence, the probabilities, of the first class data to belong
to the first class. In regards to this study, it was verified that the distribution that best
modeled the scores is the generalized extreme value (GEV) distribution. Further details
on this distribution can be found in [48]. The case studies treated in the following sections
illustrate data distribution and the quality of fitting for each case. In this paper, the threshold
of each class was set to the 95th percentile of the GEV distribution associated with the
class. It is important to understand that the latent space and thresholds obtained from
the training serve only as an initial model and will be updated as new data are classified,
accommodating potential new classes.
After training the algorithm, the testing phase is conducted in two different ways.
The first method involves analyzing data extracted from each new health condition as
a set of available observations to be tested all together, to determine whether the group
belongs to a known class. The second procedure simulates real-time monitoring, analyzing
a single data point at a time to ascertain its class or identify if it belongs to an unseen
class. Despite the apparent similarities between these two strategies, the algorithm exhibits
notable differences depending on the method, and its performance can vary significantly.
The subsequent sections delve into the description of these two testing types, highlighting
their differences.
Subsequently, the mean of the likelihood scores for the entire data group is calculated.
This mean value provides an estimate of the probability of the entire group belonging to
each of the known classes.
The algorithm then identifies the class the group is most likely to be a part of, com-
paring the mean likelihood score to the threshold of that class. If the likelihood score is
below the threshold of one of the known classes, the observations are considered to belong
to the identified class or, in the case of multiple classes meeting this requirement, to the
most probable one. On the other hand, if the score exceeds all classes’ threshold values, the
group of data is assumed to be representative of a previously unseen class. Consequently,
a new class is created using these data, and the algorithm learns about this new class by
updating the latent space and the thresholds acquired in the training phase, following
the step described in Section 2.2. In this way, the subsequent groups of data are classified
considering the newly created class. In order to better visualize the procedure, a graphical
representation of the testing strategy is given in Figure 3.
Therefore, once the likelihood scores of one observation are computed and compared
with the thresholds, there are three possible scenarios: the data point is recognized as part
of (1) just one of the known classes, (2) of multiple classes, or (3) of none.
When a data point does not seem to belong to any of the seen classes, it is placed
on a waiting line, and a counter (NC-counter), tracking the non-classifiable data points, is
incremented. If the NC-counter reaches the previously defined value of NCpnts, the data
points waiting in the queue are used to create a new class. The algorithm learns this new
class by applying PLDA and re-computing the thresholds. Instead, if the counter is not
equal to NCpnts, the algorithm proceeds to the next data point.
If the data point appears to belong to more than one of the known classes, it is
labeled as transition data and put aside. Being part of multiple classes, the data could be
representative of a transition phase between two different states of the system as well as an
outlier of a particular damage scenario.
The third and final possibility happens when the data point belongs to one, and one
only, of the seen classes. In this case, a second counter is needed to keep track of the number
of data points recognized as part of the known classes (CD-counter). The primary purpose
of this second counter is to update the thresholds not every time a data point is added to a
specific class, but only when a group of data points is classified.
In this scenario, the algorithm places the data point on hold and increments the CD-
counter, then moves to the next data point. If this counter reaches its defined limit value, the
data points on hold are associated with the respective classes, and the thresholds, as well
as the latent space, are updated. Moreover, both counters are reset to zero. This resetting of
the NC-counter is crucial, as it ensures the previously explained misclassification margin.
Because of this, the limit of the NC-counter is indeed equal to Mis-NCpnts.
Additionally, the CD-counter is used to define an interval for re-scanning unclassified
data points. Every time the second counter reaches Mis-NCpnts, the previously unclassified
data points are rechecked, as updates to the thresholds may impact their classification.
The limitation of the algorithm lies in the definition of the two parameters that limit
the counters. The first, NCpnts, where higher values positively influence the accuracy of
the threshold of a potential new class but make its creation less likely. On the other hand,
the counter Mis-NCpnts is intuitively related to the dispersion of the data; it facilitates the
generation of a new class but reduces the accuracy of thresholds for already defined classes.
The choice of the values for the two parameters is closely related to the dataset and needs
to be determined by the user.
The algorithm’s performance was evaluated using two quality indices: binary and
multi-class. The binary performance index PIBinary measures the ability of the strategy to
discriminate between damaged and undamaged data, hence its damage detection capability.
It is defined as the number of correct predictions, made as a binary choice between damaged
and undamaged, over the total number of predictions (Equation (6)).
On the other hand, the multi-class performance index PI Multi−class is linked to the
classification of damage and the algorithm’s ability to differentiate between different
damage classes. In addition to the ratio between the correct predictions and the overall
predictions, it is important to assess whether the algorithm creates more or fewer classes
than expected. For this reason, the value of the multi-class performance index is divided
by a term dependent on the number of extra classes created or missed by the algorithm
(Equation (7)). Therefore, in cases where many classes are created or where classes are not
recognized, the results are strongly penalized as the index drops significantly.
Undamaged Structure (U) No stiffness in the first floor braces (D1) No stiffness in the first and
third floor braces (D2)
Figure 5. Representation of the 12-DOF numerical model and of the six damaged conditions.
emphasizing the overlap between data from sensors on the same floor. This pattern persists
when considering more coefficients, as depicted in Figure 7, where the first five CCs from
the sensors along the x-axis are plotted. Given the symmetry of the simulated system, the
same trend is also present for the sensors distributed along the y-axis direction, but for the
sake of brevity the results are presented only for the sensors in the x-direction.
Undamaged Condition
-0.4
-0.6
CC2
-0.8
-1
S-1 S-3 S-5 S-7 S-9 S-11 S-13 S-15
-1.2
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
CC1
-0.82 -1
S-13 S-15 S-9 S-11 -0.94 S-5 S-7 S-1 S-3
-0.46
-0.84
-0.96
CC2
CC2
-0.48
CC2
CC2
-0.86 -1.05
-0.98
-0.5 -0.88 -1
-0.52 -0.9 -1.02 -1.1
-0.65 -0.6 -0.55 0.1 0.15 0.2 0.4 0.45 0.45 0.5 0.55
CC1 CC1 CC1 CC1
0
-0.5
-0.5
CC2
-1
0.2
0
CC3
-0.2
-0.4
0.2
CC4
0.1
0
-0.2
CC5
-0.4
-0.6
-0.5 0 0.5 -1 -0.5 -0.4 0 0 0.1 0.2 -0.6 -0.4 -0.2 0
CC1 CC2 CC3 CC4 CC5
This strong dependency on the sensor location necessitates different training ap-
proaches depending on whether one is implementing testing strategy type 1 or 2. In the
first case, the effect of the sensors’ location is mitigated by analyzing all data simultane-
ously and considering the mean of the scores altogether. In this case, the algorithm tends
to distinguish between damage conditions even if not all sensors contribute equally to
differentiate classes.
On the contrary, the implementation of testing strategy type 2 requires more consider-
ation because, when examining one data point at a time, the significant separation between
data from different sensors makes the analysis challenging. Data from various sensors
Infrastructures 2024, 9, 40 13 of 24
Sensor 9 Sensor 13
-0.84 -0.4
-0.86
-0.45
-0.88
-0.5
-0.9
CC2
CC2
-0.92
-0.55
-0.94
-0.6
-0.96
U0 D1 D2 D3 D4 D5 D6 U0 D1 D2 D3 D4 D5 D6
-0.98 -0.65
0.1 0.15 0.2 0.25 0.3 -0.65 -0.6 -0.55 -0.5 -0.45 -0.4
CC1 CC1
Figure 8. First two CCs for sensor 9 and sensor 13 plotted by scenario.
Infrastructures 2024, 9, 40 14 of 24
-59 80
U0 D6 U0 D1 D2 D3 D4 D5 D6
60
-60
40
-61 20
LD2
LD2
-62
-20
60
-63 -40
-60
-64
-80
30
-150 -100
-65 -100
-150 -140 -130 -120 -110 -100 -90 -80 -150 -100 -50 0 50 100 150 200
LD1 LD1
Figure 9. First two components of the latent space. Data from sensor 13.
4
U0 Density Scores D6 Density Scores
GEV fitting 3 GEV fitting
3 2.5
Density
Density
2
2
1.5
1
1
0.5
0 0
1 1.5 2 2.5 3 3.5 4 4.5 1 1.5 2 2.5 3
Data Data
Figure 10. Fitting of GEV distribution on the density distribution of the scores for training classes U0
and D6.
Predicted Real
Scenario Training New or Existing Class Correct?
Class Class
U0 Seen Existing Class U0 U0 Yes
D1 Unseen New Class D1 D1 Yes
D2 Unseen New Class D2 D2 Yes
D3 Unseen New Class D3 D3 Yes
D4 Unseen New Class D4 D4 Yes
D5 Unseen Existing Class D4 D5 No
D6 Seen Existing Class D6 D6 Yes
when PImulty−class > 0.5, every damage scenario is recognized, as a consequence of the
index definition. Meanwhile, Mis-NCpnts does not appear to notably influence the output.
Another aspect to consider is the behavior of the standard deviations of binary and
multi-class indices, shown in Figure 12. Regarding the binary index, results are almost
unaffected by the random decision of training data, showing small standard deviation
values for every parameter combination. On the other hand, this parameter significantly
impacts the multi-class index.
1 1
0.8 0.8
Multi-class PI
0.95 0.8
Binary PI
0.6 0.6
0.94 0.6
0.93
0.92 0.4 0.4 0.4
0.91
8 0.2 8 0.2
20 6 20 6
30 4 30 4
2 0 2 0
NCpnts 40 Mis-NCpnts NCpnts 40 Mis-NCpnts
Figure 11. Mean of the binary (left) and multi-class (right) performance indices using data from
sensor 9.
1 1
0.8 0.8
SD Multi-Class PI
SD Binary PI
0.06
0.6 0.3 0.6
0.04 0.2
0.4 0.1 0.4
0.02
0
8 0.2 8 0.2
20 20 6
6
30 4 30 4
2 0 2 0
NCpnts 40 Mis-NCpnts NCpnts 40 Mis-NCpnts
Figure 12. Standard deviation of binary (left) and multi-class (right) performance indices using data
from sensor 9.
Similar results were obtained from data collected by sensor 13. Figures 13 and 14
depict the mean values of the performance indices obtained over 20 repetitions and their
standard deviations. The binary classification shows high performance and reliability,
being almost independent from both parameter combinations and the training set. In
contrast, the multi-class index achieves slightly lower values compared to sensor 9, but, at
the same time, it exhibits less dependence on Mis-NCpnts, showing a plateau for lower
values of NCpnts.
1 1
0.8 0.8
0.95
Binary PI
Multi-class PI
Figure 13. Mean of the binary (left) and multi-class (right) performance indices using data from
sensor 13.
Infrastructures 2024, 9, 40 17 of 24
1 1
0.8 0.8
SD Multi-Class PI
SD Binary PI
0.06
0.3
0.6 0.6
0.04 0.2
0.4 0.1 0.4
0.02
0
0.2 0.2
20 8
8 20 6
30 6 30 4
4 0 0
40 2 40 2
NCpnts Mis-NCpnts NCpnts Mis-NCpnts
Figure 14. Standard deviation of binary (left) and multi-class (right) performance indices using data
from the sensor 13.
setup number 5, which had no sensors on the piles. In addition, for all setups and in both
sub-databases, three reference sensors (R1, R2, R3) were placed as shown in Figure 15.
V Uni-axial
V+T Bi-axial
Zurich V+T+L Tri-axial
Setup 5
Utzenstrof
Koppigen
R3 R2 R1
Data were collected at a sampling frequency of f s = 100 Hz, with each acceleration
time history consisting of eight averages of 8192 samples. Consequently, the signal duration
for each dataset amounted to 10 min and 55 s.
were located at the midspan of the bridge, making them the most sensitive to the different
damage scenarios. Figure 15 displays the sensor information for setup 5.
Scenarios U01, D04, D06, D08, D13, and D14, highlighted in bold in Table 2, were
considered for the testing analysis. These scenarios are representative of various damages
in terms of types and severity, making them suitable for evaluating the classification perfor-
mances of the algorithm. When considering additional scenarios, the performance tends to
decline due to slight differences between damage conditions and resulting similarities in
the cepstral features.
Figure 16 depicts, on the left, the first two CCs from the selected scenarios extracted
considering the vertical sensors in setup 5. It can be noticed there is a considerable sep-
aration between undamaged condition U0 and the remaining scenarios. As expected,
condition D6, the new reference condition, defined after the reversible damage scenarios, is
separated from the original undamaged scenario, implying a sub-optimal restoration of the
system to the original reference state. This separation between classes is particularly rele-
vant in the second testing type, as the possibility of single data points being misclassified
is reduced, thereby facilitating the recognition of unseen classes. On the other hand, the
damage scenarios, even if sufficiently separated from the undamaged reference state, did
not show a clear separation between each other. Figure 16 displays, on the right, the first
two LDA components obtained when considering all the scenarios previously mentioned
in the training phase (U01, D04, D06, D08, D13, and D14). This representation, as for
the numerical case, is used to highlight the separation capabilities that the classification
PLDA model will achieve in its learning updating strategy when able to recognize new
unknown classes.
CCs from vertical sensor setup 5 LDA componets from vertical sensor setup 5
0 10
0
-0.5
-5
LD2
CC2
-10
-1
U0 -15 U01
D4 D4
D6 D6
D8 -20 D8
D13 D13
D14 D14
-1.5 -25
0.5 1 1.5 2 0 10 20 30
CC1 LD1
Figure 16. Comparison between CCs (left) and LDA (right) extracted from the vertical sensors of
setup 5.
Predicted Real
Scenario Training New or Existing Class Correct?
Class Class
U1 Seen Existing Class 0 0 Yes
D4 Unseen New Class 1 1 Yes
D6 Seen Existing Class 2 2 Yes
D8 Unseen New Class 3 3 Yes
D13 Unseen New Class 4 4 Yes
D14 Unseen Existing Class 4 5 Yes
Table 4 illustrates the prediction results of the model when considering all the damage
scenarios for testing. The algorithm demonstrates the ability to differentiate between
various scenarios, with only a few instances of misclassification occurring between similar
scenarios in terms of severity and type. The classes from D11 onward are all associated
with the same damage condition. This outcome is expected, as these refer to the most
severe damage scenarios, and differentiation between them is challenging due to the overall
structural damage characterizing the bridge. Therefore, this does not significantly impact
the algorithm’s positive performance.
Predicted Real
Scenario Training New or Existing Class Correct?
Class Class
U1 Seen Existing Class 1 1 Yes
U2 Unseen New Class 2 2 Yes
D1 Unseen New Class 3 3 Yes
D2 Unseen New Class 4 4 Yes
D3 Unseen New Class 5 5 Yes
D4 Unseen New Class 6 6 Yes
D5 Unseen New Class 7 7 Yes
D6 Seen Existing Class 8 8 Yes
D7 Unseen New Class 9 9 Yes
D8 Unseen Existing Class 9 10 No
D9 Unseen New Class 11 11 Yes
D10 Unseen New Class 11 12 No
D11 Unseen New Class 13 13 Yes
D12 Unseen Existing Class 13 14 No
D13 Unseen Existing Class 13 15 No
D14 Unseen Existing Class 13 16 No
D15 Unseen Existing Class 13 17 No
1 1
0.8 0.8
0.6
Multi-class PI
0.96
Binary PI
0.6 0.6
0.94 0.4
0.92
0.4 0.2 0.4
0.9
0.88
0.2 0.2
40 20
20 20
0 40 0
10 20 10
Mis-NCpnts NCpnts NCpnts Mis-NCpnts
Figure 17. Mean of the binary (left) and multi-class (right) performance indices using data from
setup 5.
1 1
0.8 0.8
SD Multi-class PI
SD Binary PI
0.06 0.3
0.6 0.6
0.2
0.04
0.1
0.4 0.4
0.02
25
0.2 0.2
20 20
40 40
15
0 0
Mis-NCpnts 10 20 NCpnts Mis-NCpnts 10 20 NCpnts
Figure 18. Standard deviation of binary (left) and multi-class (right) performance indices using data
from sensor 9.
5. Conclusions
The present study addresses a novel damage assessment strategy for dynamic sys-
tems that adopts a hybrid supervised–unsupervised methodology to identify and classify
known and unknown healthy and damaged conditions. The strategy relies on training a
Infrastructures 2024, 9, 40 22 of 24
Probabilistic Linear Discriminant Analysis model starting from cepstral features extracted
from the measurement responses of the system, projected into a latent space where the
class separation is maximized. The classification model is then used to evaluate the health
condition of structural systems.
The classification performance of the model was evaluated based on its capacity
to identify the presence of damage as a binary decision and to classify it, even in the
case of new damage conditions. The framework being proposed takes into account two
testing strategies.
The first method analyzes groups of data altogether, assuming that all data in the
same block represent the same condition. This approach is useful for condition monitoring
scenarios where the damage conditions are already known or for post-processing damage
assessment. The second methodology, on the other hand, considers the analysis of a single
incoming observation data point, offering a damage assessment strategy compatible with
real-time monitoring strategies.
The proposed method was evaluated using two testing strategies on a simulated
shear-type model with plate number 12-DOF. Following the first testing approach, the
model was trained on the undamaged scenario and the lighter damaged condition, and
it demonstrated excellent performance in distinguishing between the known scenarios
and other damaged conditions. It proved to be able to recognize the classes seen in the
training phase correctly and, at the same time, identify unknown classes, update the prior
knowledge of the model, and recognize them afterward. Even when analyzing one data
point at a time, the model offered a robust classification accuracy for the binary and the
multi-class case with performance indexes above 90% and 80%, respectively, for the optimal
algorithm’s parameters.
A comparable set of results could be observed for the experimental data collected from
the Z24 bridge. The model proved to correctly identify the different damage categories and
differentiate them from the undamaged reference condition, processing the data by group
and for a single point at a time.
In the future, it is planned to upgrade the framework to explore the possibility of
conducting the training phase using only the undamaged condition and gradually intro-
ducing transitional intermediate damage scenarios, investigating a progressively increasing
degradation event for the system.
Author Contributions: Conceptualization, L.S. (Lorenzo Stagi), L.S. (Lorenzo Sclafani), E.M.T., R.B.,
S.M., A.C. (Antonio Culla), N.R. and A.C. (Antonio Carcaterra); methodology, L.S. (Lorenzo Stagi),
L.S. (Lorenzo Sclafani), E.M.T. and R.B.; software, L.S. (Lorenzo Stagi), L.S. (Lorenzo Sclafani), E.M.T.
and R.B.; validation, L.S. (Lorenzo Stagi), L.S. (Lorenzo Sclafani), E.M.T. and R.B.; formal analysis, L.S.
(Lorenzo Stagi) and L.S. (Lorenzo Sclafani); investigation, L.S. (Lorenzo Stagi), L.S. (Lorenzo Sclafani),
E.M.T. and R.B.; data curation, L.S. (Lorenzo Stagi), L.S. (Lorenzo Sclafani) and E.M.T.; writing—
original draft preparation, L.S. (Lorenzo Stagi), L.S. (Lorenzo Sclafani) and E.M.T.; writing—review
and editing, L.S. (Lorenzo Stagi), L.S. (Lorenzo Sclafani), E.M.T., R.B., S.M., A.C. (Antonio Culla), N.R.
and A.C. (Antonio Carcaterra); supervision, E.M.T., R.B., S.M., A.C. (Antonio Culla), N.R. and A.C.
(Antonio Carcaterra); project administration, E.M.T., R.B., S.M., A.C. (Antonio Culla), N.R. and A.C.
(Antonio Carcaterra). All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Data Availability Statement: The numerical data used in the present work can be reproduced
following the detailed information provided by the authors in [29]. The experimental data from the
Z24 bridge were provided by the Katholieke Universiteit Leuven.
Acknowledgments: The authors sincerely acknowledge the Katholieke Universiteit Leuven, Struc-
tural Mechanics Section, for providing the data collected within the SIMCES project.
Conflicts of Interest: The authors declare no conflicts of interest.
Infrastructures 2024, 9, 40 23 of 24
References
1. Farrar, C.R.; Worden, K. An introduction to structural health monitoring. Philos. Trans. R. Soc. A Math. Phys. Eng. Sci. 2007,
365, 303–315. [CrossRef]
2. Gharehbaghi, V.R.; Noroozinejad Farsangi, E.; Noori, M.; Yang, T.; Li, S.; Nguyen, A.; Málaga-Chuquitaype, C.; Gardoni, P.;
Mirjalili, S. A critical review on structural health monitoring: Definitions, methods, and perspectives. Arch. Comput. Methods Eng.
2021, 29, 2209–2235. [CrossRef]
3. Fan, W.; Qiao, P. Vibration-based damage identification methods: A review and comparative study. Struct. Health Monit. 2011,
10, 83–111. [CrossRef]
4. Hou, R.; Xia, Y. Review on the new development of vibration-based damage identification for civil engineering structures:
2010–2019. J. Sound Vib. 2021, 491, 115741. [CrossRef]
5. Brownjohn, J.M.; Xia, P.Q.; Hao, H.; Xia, Y. Civil structure condition assessment by FE model updating: Methodology and case
studies. Finite Elem. Anal. Des. 2001, 37, 761–775. [CrossRef]
6. Friswell, M.; Mottershead, J.E. Finite Element Model Updating in Structural Dynamics; Springer Science & Business Media: Berlin,
Germany, 1995; Volume 38.
7. Cabboi, A.; Gentile, C.; Saisi, A. From continuous vibration monitoring to FEM-based damage assessment: Application on a
stone-masonry tower. Constr. Build. Mater. 2017, 156, 252–265. [CrossRef]
8. Ying, Y.; Garrett, J.H., Jr.; Oppenheim, I.J.; Soibelman, L.; Harley, J.B.; Shi, J.; Jin, Y. Toward data-driven structural health
monitoring: Application of machine learning and signal processing to damage detection. J. Comput. Civ. Eng. 2013, 27, 667–680.
[CrossRef]
9. Tibaduiza Burgos, D.A.; Gomez Vargas, R.C.; Pedraza, C.; Agis, D.; Pozo, F. Damage identification in structural health monitoring:
A brief review from its implementation to the use of data-driven applications. Sensors 2020, 20, 733. [CrossRef]
10. Balafas, K.; Kiremidjian, A.S. Extraction of a series of novel damage sensitive features derived from the continuous wavelet
transform of input and output acceleration measurements. Proc. Sens. Smart Struct. Technol. Civil Mech. Aerosp. Syst. 2014, 9061,
386–394.
11. Worden, K.; Manson, G.; Fieller, N.R. Damage detection using outlier analysis. J. Sound Vib. 2000, 229, 647–667. [CrossRef]
12. Farrar, C.R.; Worden, K. Structural Health Monitoring: A Machine Learning Perspective; John Wiley & Sons: Hoboken, NJ, USA, 2012.
13. Cross, E.; Manson, G.; Worden, K.; Pierce, S. Features for damage detection with insensitivity to environmental and operational
variations. Proc. R. Soc. A Math. Phys. Eng. Sci. 2012, 468, 4098–4122. [CrossRef]
14. Gibbs, D.; Jankowski, K.; Rees, B.; Farrar, C.; Flynn, G. Identifying Environmental-and Operational-Insensitive Damage Features.
In Data Science in Engineering, Volume 9: Proceedings of the 39th IMAC, A Conference and Exposition on Structural Dynamics 2021;
Springer: Cham, Switzerland, 2022; pp. 105–121.
15. Wang, Z.; Yang, D.H.; Yi, T.H.; Zhang, G.H.; Han, J.G. Eliminating environmental and operational effects on structural modal
frequency: A comprehensive review. Struct. Control Health Monit. 2022, 29, e3073. [CrossRef]
16. Tronci, E.M.; Betti, R.; De Angelis, M. Damage detection in a RC-masonry tower equipped with a non-conventional TMD using
temperature-independent damage sensitive features. Dev. Built Environ. 2023, 15, 100170. [CrossRef]
17. Kim, J.T.; Ryu, Y.S.; Cho, H.M.; Stubbs, N. Damage identification in beam-type structures: Frequency-based method vs
mode-shape-based method. Eng. Struct. 2003, 25, 57–67. [CrossRef]
18. Yao, R.; Pakzad, S.N. Autoregressive statistical pattern recognition algorithms for damage detection in civil structures. Mech.
Syst. Signal Process. 2012, 31, 355–368. [CrossRef]
19. Bernagozzi, G.; Achilli, A.; Betti, R.; Diotallevi, P.P.; Landi, L.; Quqa, S.; Tronci, E.M. On the use of multivariate autoregressive
models for vibration-based damage detection and localization. Smart Struct. Syst. Int. J. 2021, 27, 335–350.
20. Bogert, B.P. The quefrency alanysis of time series for echoes: Cepstrum, pseudoautocovariance, cross-cepstrum and saphe
cracking. Proc. Symp. Time Ser. Anal. 1963, 209–243.
21. Oppenheim, A.V.; Schafer, R.W. Discrete-Time Signal Processing; Prentice-Hall: Hoboken, NJ, USA, 1999; Volume 2, pp. 878–882.
22. Davis, S.; Mermelstein, P. Comparison of parametric representations for monosyllabic word recognition in continuously spoken
sentences. IEEE Trans. Acoust. Speech Signal Process. 1980, 28, 357–366. [CrossRef]
23. Hasan, M.R.; Jamil, M.; Rahman, M. Speaker identification using mel frequency cepstral coefficients. Variations 2004, 1, 565–568.
24. Nelwamondo, F.V.; Marwala, T.; Mahola, U. Early classifications of bearing faults using hidden Markov models, Gaussian
mixture models, mel-frequency cepstral coefficients and fractals. Int. J. Innov. Comput. Inf. Control 2006, 2, 1281–1299.
25. Benkedjouh, T.; Chettibi, T.; Saadouni, Y.; Afroun, M. Gearbox fault diagnosis based on mel-frequency cepstral coefficients and
support vector machine. In Proceedings of the IFIP International Conference on Computational Intelligence and Its Applications,
Oran, Algeria, 8–10 May 2018; pp. 220–231.
26. Hwang, Y.R.; Jen, K.K.; Shen, Y.T. Application of cepstrum and neural network to bearing fault detection. J. Mech. Sci. Technol.
2009, 23, 2730–2737. [CrossRef]
27. Balsamo, L.; Betti, R.; Beigi, H. A structural health monitoring strategy using cepstral features. J. Sound Vib. 2014, 333, 4526–4542.
[CrossRef]
28. de Souza, E.F.; Bittencourt, T.N.; Ribeiro, D.; Carvalho, H. Feasibility of Applying Mel-Frequency Cepstral Coefficients in a
Drive-by Damage Detection Methodology for High-Speed Railway Bridges. Sustainability 2022, 14, 13290. [CrossRef]
Infrastructures 2024, 9, 40 24 of 24
29. Tronci, E.M.; Beigi, H.; Betti, R.; Feng, M.Q. A damage assessment methodology for structural systems using transfer learning
from the audio domain. Mech. Syst. Signal Process. 2023, 195, 110286. [CrossRef]
30. Li, Z.; Lin, W.; Zhang, Y. Drive-by bridge damage detection using Mel-frequency cepstral coefficients and support vector machine.
Struct. Health Monit. 2023, 22, 14759217221150932. [CrossRef]
31. Morgantini, M.; Betti, R.; Balsamo, L. Structural damage assessment through features in quefrency domain. Mech. Syst. Signal
Process. 2021, 147, 107017. [CrossRef]
32. 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]
33. Cunningham, P.; Cord, M.; Delany, S.J. Supervised learning. In Machine Learning Techniques for Multimedia: Case Studies on
Organization and Retrieval; Springer: Berlin/Heidelberg, Germany, 2008; pp. 21–49.
34. Perfetto, D.; Rezazadeh, N.; Aversano, A.; De Luca, A.; Lamanna, G. Composite Panel Damage Classification Based on Guided
Waves and Machine Learning: An Experimental Approach. Appl. Sci. 2023, 13, 10017. [CrossRef]
35. Moradi, M.; Broer, A.; Chiachío, J.; Benedictus, R.; Loutas, T.H.; Zarouchas, D. Intelligent health indicator construction for
prognostics of composite structures utilizing a semi-supervised deep neural network and SHM data. Eng. Appl. Artif. Intell. 2023,
117, 105502. [CrossRef]
36. Fugate, M.L.; Sohn, H.; Farrar, C.R. Unsupervised learning methods for vibration-based damage detection. In Proceedings of the
18th International Modal Analysis Conference–IMAC, San Antonio, TX, USA, 7–10 February 2000; Volume 18.
37. Wang, Z.; Cha, Y.J. Unsupervised machine and deep learning methods for structural damage detection: A comparative study.
Eng. Rep. 2022, e12551. [CrossRef]
38. Roveri, N.; Carcaterra, A. Unsupervised identification of damage and load characteristics in time-varying systems. Contin. Mech.
Thermodyn. 2015, 27, 531–550. [CrossRef]
39. Rezazadeh, N.; de Oliveira, M.; Perfetto, D.; De Luca, A.; Caputo, F. Classification of Unbalanced and Bowed Rotors under
Uncertainty Using Wavelet Time Scattering, LSTM, and SVM. Appl. Sci. 2023, 13, 6861. [CrossRef]
40. Fogaça, M.J.; Cardoso, E.L.; de Medeiros, R. A systematic approach to find the hyperparameters of artificial neural networks
applied to damage detection in composite materials. J. Braz. Soc. Mech. Sci. Eng. 2023, 45, 496. [CrossRef]
41. Sarmadi, H.; Entezami, A. Application of supervised learning to validation of damage detection. Arch. Appl. Mech. 2021,
91, 393–410. [CrossRef]
42. Xanthopoulos, P.; Pardalos, P.M.; Trafalis, T.B.; Xanthopoulos, P.; Pardalos, P.M.; Trafalis, T.B. Linear discriminant analysis. In
Robust Data Mining; Springer: New York, NY, USA, 2013; pp. 27–33.
43. Ioffe, S. Probabilistic linear discriminant analysis. In Proceedings of the Computer Vision–ECCV 2006: 9th European Conference
on Computer Vision, Graz, Austria, 7–13 May 2006; Proceedings, Part IV 9, pp. 531–542.
44. Snyder, D.; Garcia-Romero, D.; Sell, G.; Povey, D.; Khudanpur, S. X-vectors: Robust dnn embeddings for speaker recognition.
In Proceedings of the 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Calgary, AB,
Canada, 15–20 April 2018; pp. 5329–5333.
45. Ananthram, A.; Saravanakumar, K.K.; Huynh, J.; Beigi, H. Multi-modal emotion detection with transfer learning. arXiv 2020,
arXiv:2011.07065.
46. Fisher, R.A. The use of multiple measurements in taxonomic problems. Ann. Eugen. 1936, 7, 179–188. [CrossRef]
47. Rao, C.R. The utilization of multiple measurements in problems of biological classification. J. R. Stat. Soc. Ser. B 1948, 10, 159–203.
[CrossRef]
48. Kotz, S.; Nadarajah, S. Extreme Value Distributions: Theory and Applications; World Scientific: Singapore, 2000.
49. Johnson, E.A.; Lam, H.F.; Katafygiotis, L.S.; Beck, J.L. Phase I IASC-ASCE structural health monitoring benchmark problem using
simulated data. J. Eng. Mech. 2004, 130, 3–15. [CrossRef]
50. Peeters, B.; De Roeck, G. One-year monitoring of the Z24-Bridge: Environmental effects versus damage events. Earthq. Eng.
Struct. Dyn. 2001, 30, 149–171. [CrossRef]
51. Roeck, G.D. The state-of-the-art of damage detection by vibration monitoring: The SIMCES experience. J. Struct. Control 2003,
10, 127–134. [CrossRef]
52. Brincker, R.; Andersen, P.; Cantieni, R. Identification and level I damage detection of the Z24 highway bridge. Exp. Tech. 2001,
25, 51–57. [CrossRef]
53. Krämer, C.; De Smet, C.; De Roeck, G. Z24 bridge damage detection tests. In IMAC 17, the International Modal Analysis Conference;
Society of Photo-optical Instrumentation Engineers: Bellingham, WA, USA, 1999; Volume 3727, pp. 1023–1029.
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.