You are on page 1of 11

Engineering Structures 304 (2024) 117621

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Machine learning-based analysis of historical towers


Hamed Dabiri a, *, 1, Jessica Clementi b, Roberta Marini c, 2, Gabriele Scarascia Mugnozza b, 3,
Francesca Bozzano b, 4, Paolo Mazzanti b, 5
a
Department of Civil and Environmental Engineering, Imperial College London, SW7 2AZ, UK
b
Department of Earth Sciences, Sapienza University of Rome & CERI Research Center, P.le Aldo Moro 5, 00185 Roma, Italy
c
Natural Hazards Control and Assessment (NHAZCA srl.), Via Vittorio Bachelet 12, 00185 Rome, Italy

A R T I C L E I N F O A B S T R A C T

Keywords: Determining frequency of buildings has a crucial impact on proper analysis and design of structures. Conven­
Natural frequency tional methods (i.e., in situ and numerical analysis) are generally costly and time-consuming. This study aims at
Towers developing Machine Learning (ML)-based methods for obtaining frequency of structures with the focus on ma­
Machine learning
sonry towers (i.e., bell towers). To this end, a database including the results of either in situ or numerical analysis
Artificial Intelligence
on over 90 masonry towers available in literature is collected. Additionally, frequencies of 18 towers located in
Structural analysis and design
Venice, Italy are measured by site survey, and added to the database. Parameters with the highest influence on
tower’s frequency, namely height, plan dimensions and modulus of elasticity are defined as input variables for
predicting natural frequency of a tower by ML-based techniques including Decision Tree (DT), Random Forest
(RF), XGBoost and K-Nearest Neighbors (KNN). The models’ performance is analyzed by comparing the corre­
lation between the predicted and real values. Moreover, the models’ accuracy is assessed through common
performance metrics and Taylor diagram, and the most accurate model is introduced accordingly. Results
highlighted the high capability of ML-based approaches for predicting towers’ frequency, and the XGBoost model
exhibited the highest accuracy. In the second part of the paper, the values predicted by the most accurate model
are compared to those calculated by the equations proposed by design Codes (i.e., NTC008, NCSE, ASCE) and
literature. Lastly, for deeper investigating the performance of masonry towers a sensitivity analysis is performed
by the proposed XGBoost model, and an equation is suggested for calculating their natural frequency based on
their height and plan width.

1. Introduction degradation phenomena affecting the mechanical properties of mate­


rials, presence of cracks and damages, effects of successive construction
Masonry towers are a Cultural Heritage building typology wide­ phases or modifications, effects of thermal variations), but is recognized
spread worldwide, especially on the Italian peninsula, where they are a also as one of the most vulnerable to dynamic actions (e.g. traffic-
distinctive feature of many of its historic centers and countryside. The induced micro-tremors, swinging of bells, wind and especially seismic
great variety of uses reserved for masonry towers (namely bell towers, events, as number of collapses occurred in the past decades have pointed
civic towers, watch towers and tower-houses) has resulted in a consid­ out [1–4]). This is due to both their structural configuration - as they
erable number of structural configurations. They are usually high and usually present high vertical compressive stresses induced by gravita­
slender structures located either nearby churches and other religious tional loads, often close to the compressive strength of the material - and
sites or free-standing. Moreover, such structural typology is not only the mechanical properties of the masonry.
characterized by the common issues of historic buildings (i.e. Structural safety assessment of historic masonry towers is one of the

* Corresponding author.
E-mail address: h.dabiri@imperial.ac.uk (H. Dabiri).
1
ORCID: 0000–0002-5654–2429
2
ORCID: 0000–0002-2899–0232
3
ORCID: 0000–0002-2917–0324
4
ORCID: 0000–0002-0297–842X
5
ORCID: 0000–0003-0042–3444

https://doi.org/10.1016/j.engstruct.2024.117621
Received 3 October 2023; Received in revised form 23 December 2023; Accepted 4 February 2024
Available online 13 February 2024
0141-0296/© 2024 Elsevier Ltd. All rights reserved.
H. Dabiri et al. Engineering Structures 304 (2024) 117621

central issues in the conservation of the national and global architectural


1
heritage. Their preservation is paramount because of their historic, frequency = (2)
0.0151 × h1.08
architectural, social and symbolic value for a given “Heritage commu­
nity” of people who value specific aspects of cultural heritage which 1
they wish, within the framework of public action, to sustain and transmit frequency = (3)
0.01137 × h1.138
to future generations (Faro Convention, art. 2). Indeed, the feeling of
Conventional techniques used for determining frequency (both
place and belonging is strongly linked to Cultural Heritage buildings:
experimental and numerical), however, are generally challenging. In the
their damage and partial/total collapse can have a huge impact on social
case of historical towers, for instance, installing measurement tools is
cohesion, sustainable development and psychological wellbeing [5].
not allowed because of their high importance to governments in terms of
In the viewpoint of structural engineering, the natural frequency of
history, economics, and art. Moreover, the instruments are often costly
buildings is one of the most significant parameters which reflect the
and require special skills for their application and acquiring data.
behavior of structures. For the purpose of designing structures, natural
Analytical approaches (e.g., change detection through image process­
frequency should be determined for applying dynamic loads (i.e.,
ing), on the other hand, are time-consuming since the data are obtained
earthquake). Furthermore, the frequency of buildings is obtained for
by extensive analysis on large databases.
better understanding of their behavior. As an example, comparing fre­
In recent decades, Machine Learning (ML)-based models have been
quency of a building before and after retrofitting could help engineers
introduced as inexpensive, accurate and quick substitutions for experi­
realize its performance enhancement under dynamic loads [6].
mental and analytical techniques. The implementation of ML-based
Due to specific structural configuration of masonry towers, particu­
models for evaluating performance of cultural heritages [16], masonry
larly their vulnerability to earthquake-induced damage, there has been a
buildings [17,18], and towers have been discussed in literature. In a
growing interest in evaluating response of slender masonry towers (e.g.,
recent study [19], power regression and Artificial Neural Network
bell towers) in the last decades. Performance of tall masonry buildings
(ANN) are applied for predicting fundamental frequency of the first
are generally analyzed through experimental techniques or numerical
bending mode of a tower considering its geometry, boundary condition
simulations; in the former, the tower response against ambient vibra­
and material. Additionally, empirical equations were proposed for
tions induced by different factors (e.g., wind, traffic, micro-seismic
calculating frequency based on different parameters. It was concluded
events, and other human activities) is monitored by seismic stations
that the below equations which are functions of tower height (h) returns
equipped by seismometers [7,8]. In the latter, towers behavior is
higher accuracy compared to those which are dependent on geometry,
assessed by analyzing parameters like frequency, modal shapes,
boundary condition and material:
maximum displacement, drift ratio, and stress distribution on the tower
model verified by results of the similar experimental tests [6,9,10]. frequency = 42.12h− 0.893
(all towers) (4)
As an example, D’Ambrisi et al. [9] clarified that the behavior factor,
q, values provided by the Italian codes overestimate the actual q values frequency = 58.57h− 1.026
(isolated towers) (5)
obtained from a FE analysis. Moreover, they reported that the outcomes
of nonlinear dynamic analysis are in line with those acquired by 0601
frequency = 11.87h−effective (bounded towers) (6)
nonlinear static analysis. The displacement capacity given by nonlinear
dynamic analysis, however, is less than those obtained by nonlinear frequency = 190.90h− 1.378
(isolated brick towers) (7)
static analysis which takes into account the considerable influence of
large frequencies on towers response parameters (e.g., base shear). In In another ML-based study, Ivorra et al. [20] estimated the variation
another study conducted by Yanık et al. [6], a noticeable difference of groundwater level by the application of ANN based on the data ac­
between the natural frequency of FE analysis compared to the experi­ quired by accelerometers installed in different heights of a tower. Their
mental test is observed. They enhanced the FE model’s accuracy by results showed that ANN could be considered as a reliable method for
modifying the springs’ stiffness on the base level and mechanical analyzing performance of slender buildings. Implementing both ML and
properties of the tower’s walls. Lacanna et al. [7] has also estimated FEA for assessing and monitoring performance of cultural heritages over
parameters, namely natural frequency, damping ratio and modal shapes, time has also been discussed in some studies [21]. Standoli et al. [22]
by Enhanced Domain Decomposition (EFDD). Additionally, they pro­ proposed a novel method combining FEA and generic algorithm for
posed an automated technique for determining modal parameters in predicting material properties of masonry buildings, and consequently
real-time using seismic stations. facilitating monitoring their dynamic performance.
The results of experimental and numerical studies could be used for Although estimating the fundamental period of structural frames by
generating equations which return an accurate estimation of towers ML-based techniques has been rarely researched [23], no similar study
frequency. Bartoli et al. [11] proposed an equation (Eq. (1)) for deter­ has been carried out on towers so far. This study, as a result, aims at
mining the first natural frequency of towers based on their effective introducing quick, simple in implementation and most importantly ac­
height (Heffective, m) and plan width (w, m). The comparison between the curate ML-based methods which could be a reliable substitution for
real and calculated frequencies confirmed an acceptable reliability of expensive and time-consuming conventional techniques for acquiring
Eq. (1): natural frequency of masonry towers. To this end, the paper is structured
in three main parts: (i) firstly, ML-based techniques, namely DT, RF,
frequency =
150 × w
(1) XGBoost and KNN are developed for predicting natural frequency of
h2effective masonry towers; then, the most accurate model is introduced through
deep comparison and analysis of models; (ii) the values predicted by the
In another study, Bartoli et al. [12] collected a large database on
most accurate model are compared to the frequencies calculated by the
masonry towers which were used to propose a more complicated
equations available in design codes and literature; (iii) and finally, a
equation for calculating the main frequency of historic masonry towers
sensitivity analysis is performed, and an equation is proposed for
by considering the façade opening ratio which affects mass and stiffness
calculating frequency of a tower by its height and plan width. The
of towers. Other equations, which are a function of only height, have
research flowchart is given in Fig. 1.
been also proposed by researchers [13,14], as given in Eqs. (2)–(3). It
has been clarified that Eq. (3) returns the most accurate results [15], and
2. Definitions and code provisions
therefore it is used in this study as well:
Frequency is simply the inverse of period (1/T) where T is the

2
H. Dabiri et al. Engineering Structures 304 (2024) 117621

Fig. 1. Research flowchart.

fundamental period of the building. Fundamental period is highly tower are included in the database. Parameters, namely modulus of
influenced by parameters, namely structural properties and deforma­ Elasticity (E, MPa), tower height (m), plan width (m) and plan length
tional characteristics of resisting elements [24]. This study is focused on (m), which have the highest effect on frequency are considered as input
the first natural frequency (which will be denoted by ‘frequency’ here­ variables for predicting frequency (Fr, Hz) using four ML-based ap­
after for the sake of simplicity) of masonry towers, and therefore only proaches including DT, RF, XGBoost, KNN. The mentioned parameters
equations given for this building classification are discussed. The natural are schematically displayed in Fig. 2. Statistical properties of the input
period of towers is mainly determined by their geometrical properties and output parameters are provided in Table 2.
including “height” or “height and plan dimensions”. Many design codes Based on the information given in Table 2, the collected database
like American Society of Civil Engineering (ASCE) [24], and Italian covers a wide range of each variable, and this means that our database is
Building Code (NTC08) [25], provide equations which calculate sufficiently appropriate for developing generalized ML-based prediction
approximate fundamental period by a tower’s height. Several re­ models which could be used for predicting frequency of any tower with
searchers have also proposed equations which are dependent on only its features being included in the inputs’ range. As an example, the
tower height [13,15]. In some other design codes (e.g., Spanish Standard collected database included towers with heights between 18.5–98 m
(NCSE 2002) [25]), on the other hand, the fundamental period of a which cover a high portion of existing towers. The minimum and
tower is a function of both height and plan dimension. The most often maximum plan dimensions of the towers in the database are respectively
used equations are given in Table 1. 3 m and 14.5 m, and plan dimensions of most of the towers are included
in this range. The histogram of each parameter is also demonstrated in
3. Data collection and analysis Fig. 3 to better understanding the variables’ distribution.
Considering parameters with the highest effect on the output plays a
In order to develop ML-based models for predicting natural fre­ crucial role on the models’ accuracy and performance. In this study,
quency of towers, a database including over 100 masonry towers is feature selection is performed by evaluating two common correlation
utilized. 83 data are collected from the results of either in situ or nu­ coefficients: Pearson (Eq. (12)) and Spearman (Eq. (13)):
merical analysis on towers available in peer-reviewed international ∑
cov(X, Y) (xi − x)(yi − y)
publications [11–13,15,27–29]. Additionally, 17 towers located in ρx,y = = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
̅√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (12)
σx σy ∑ ∑
Venice, Italy were evaluated by remote sensing techniques, and the (xi − x)2 (yi − y)2
acquired frequencies are added to the database. It should be noted that

the data collected from numerical studies are highly dependent on 6 (xi − yi )2
various parameters (e.g., material heterogeneity, tower boundary con­ ρ = 1− (13)
n(n2 − 1)
ditions, soil type, nonlinearity and anisotropy due to cracks, and ge­
ometry) [30] which can affect reliability of the data. To address this where x and y are variables, cov (X,Y) is the covariance of variables, σ is
issue, the analytical simulations which are validated by a real-world standard deviation, x and y are mean values of the variables, and n is the
number of each variable. It should be noted that Pearson coefficients
reflect the linear correlation between two parameters while Spearman
Table 1
coefficients indicate nonlinear relationship between two variables. Both
The most used equations proposed in design codes and literature.
Pearson and Spearman coefficients take values between − 1 and 1: the
Reference Equation Equation No. higher the coefficient is, the more dependent the variables are. It is
Italian Building Code (NTC08)[25]
f(H) =
1 (8) worth noting that positive and negative coefficient represent direct and
0.05H0.75 reverse relationship, respectively [31,32]. Correlation coefficients of the
ASCE[24] 1 (9)
f(H) = input and output parameters are given in Table 3.
0.0488H√0.75
̅̅̅
Spanish Standard (NCSE 2002)[26] L (10) Considering the values reported in Table 3, tower height (with
f(H, L) = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
0.06H
H Pearson and Spearman coefficients of respectively 0.6885 and 0.7729)
2L + H has the highest influence on its frequency. This is in line with the
Ranieri and Fabbrocino[14] 1 (11)
f(H) =
0.01137H1.138 equations provided in design codes (e.g., ASCE [24]) and literature (e.g.,
Ranieri and Fabbrocino [14]) for calculating approximate frequency
H: tower height (m), L: base dimension in the oscillation direction

3
H. Dabiri et al. Engineering Structures 304 (2024) 117621

Fig. 2. Input and output parameters used for developing prediction models.

Table 2
Statistical properties of input and output variables.
Min. Max. Mean Median Variance Standard deviation

Modulus of Elasticity (MPa) 700.00 9000.00 2292.62 1900.00 1962610.42 1400.93


Tower Height (m) 18.50 98.00 44.77 41.50 276.72 16.63
Plan Width (m) 3.00 14.50 7.18 7.10 4.98 2.23
Plan length (m) 3.30 14.50 7.36 7.30 5.00 2.24
Frequency (Hz) 0.35 4.02 1.38 1.25 0.42 0.65

which is a function of a tower’s height. Although plan dimensions are relationship between inputs and output.
rarely considered in approximate equations (e.g., NCSE 2002 [26]),
their effect on frequency is not negligible, and is highly recommended to 4. ML-based models
be taken into account for acquiring a tower frequency. The other notable
point is that all the coefficients between inputs and the output are In order to develop ML-based models for predicting frequency of
negative. This simply means that the taller the tower is, or the larger the towers, the collected database was split into two sub databases: training
plan area is, the lower the frequency will be. The low correlation be­ (85% of the total data) which was used for training the relationship
tween E and frequency implies that no certain trend could be obtained between inputs and outputs to the models, and testing (15%) which was
based on the values in the collected database. A similar relationship used for assessing the accuracy of the models. Four different models,
between E and frequency could also be observed in the database given in namely DT, RF, XGBoost and KNN are developed, compared, and dis­
literature [11,12] where higher towers do not necessarily exhibit higher cussed. A brief definition of each model is provided in this section.
frequencies. To this end, the modulus of elasticity of a tower is not
generally considered in the equations proposed for calculating fre­
quency. It should also be highlighted that the Spearman coefficients are 4.1. Decision Tree (DT)
higher than the Pearson coefficient. This means that the nonlinear de­
pendency of inputs on frequency is higher than linear dependency. In Decision Tree (DT) is a supervised ML-based algorithm which could
this study, therefore, ML-based methods including DT, RF, and XGBoost be used for both classification and regression problems. It’s called a
are used which have the capability to solve problems with nonlinear "decision tree" because it is a tree-shape model that splits a dataset into
smaller sub-databases considering different features of the data. The

4
H. Dabiri et al. Engineering Structures 304 (2024) 117621

Fig. 3. Histograms of input and output variables.

methods.
Table 3
Pearson and Spearman (stared values) coefficients of input and output variables.
4.2. Random Forest (RF)
E (MPa) Height Plan width Plan length Fr (Hz)
(m) (m) (m)
Random Forest (RF) is a powerful ensemble ML-based technique
E (MPa) 1 0.1061* 0.2359* 0.1893* -0.0098*
which could be used for both classification and regression problems.
Height (m) 0.0095 1 0.5924* 0.4867* -0.7729*
Plan width 0.1511 0.5128 1 0.9148* -0.3730* An RF is formed by a series of decision trees, where each tree is
(m) trained on a different randomly selected subset of the training data and
Plan length 0.1025 0.4528 0.9486 1 -0.2935* with a subset of the features. This diversity among the trees helps the
(m) model generalize better to new, unseen data. This technique is known as
Fr (Hz) -0.0932 -0.6885 -0.3299 -0.2775 1
Bagging (Bootstrap Aggregating) in which individual trees are trained
on bootstrapped samples of the original training data. Bootstrapping
architecture of a DT model contains different parts, namely (i) root node: involves randomly selecting data points with replacement from the
the top node of the tree which represents the entire dataset, (ii) decision training set. The ultimate output of an RF model would be either a
nodes: which represent a decision or a test on a particular feature, (iii) number (average value of the values predicted by each DT for regression
branches: which represent the possible values of a feature or attribute, problems) or a class (the most voted class for a classification problem)
and (iv) leaf nodes: which are the ultimate nodes at the end of each [35,36]. It is worth mentioning that RF is less likely to overfit the
branch, and are the final class or value for classification or regression training data compared to a single decision tree by averaging predictions
problems, respectively [33,34]. It should be pointed out that the best from multiple trees and introducing randomness.
feature to split the dataset at each internal node affects the model’s Two parameters, namely number of estimators (Decision Trees) and
accuracy significantly. Other factors, namely minimum number of splits tree depth have a remarkable influence on the RF accuracy which are
and tree depth (distance from the root node to leaf node) has a notable respectively set to 183 and 10 which resulted in the highest accuracy.
influence on the models’ outputs. A reliable model returns the result in
the most homogenous subsets in terms of the target variable (for clas­ 4.3. XGBoost
sification) or the least squared error (for regression). This process is
usually performed based on parameters like Gini impurity, entropy, or XGBoost (eXtreme Gradient Boosting) is an advanced and highly
mean squared error. In this study, the maximum number of splits and effective machine learning algorithm that belongs to the family of
maximum tree depth are limited to 5 and 10, respectively. Although DT gradient boosting methods. It is used for both classification and
is popular for being simple to implement and easy to follow, it can be regression tasks and has gained significant popularity due to its excep­
prone to overfitting issues which could be addressed by ensemble tional performance and versatility. XGBoost is a gradient boosting

5
H. Dabiri et al. Engineering Structures 304 (2024) 117621

algorithm that systematically combines the predictions of multiple weak K= 2 is chosen for the proposed KNN model.
learners (typically decision trees) to create a strong predictive model. It
incorporates advanced techniques such as gradient boosting, regulari­ 5. Results and accuracy assessment
zation, and handling missing values, resulting in improved accuracy and
generalization [37,38]. 5.1. Visual comparison
High accuracy of XGBoost is due to implementing Gradient Boosting
Framework which means that XGBoost builds an ensemble of weak Frequency of the towers included in the collected database is pre­
learners (often decision trees) sequentially. It focuses on correcting the dicted by the developed ML-based models. Correlation between real
errors of previous models by emphasizing misclassified data points. values with those predicted by the models for both training (circle
In order to prevent overfitting in complex models, XGBoost includes markers) and testing (red star markers) databases is demonstrated in
Lasso and Ridge regularization terms in its objective function. It also Fig. 5. It should be explained that the green line represents the ideal line
allows the use of custom-defined objective functions, which is particu­ where predicted values are equal to real value (y = y). The gray space
larly useful for addressing specific problems or incorporating domain between the dashed red lines indicated the prediction error by 20% (i.e.,
knowledge. Other notable assets of this method are: (i) handling missing y = 0.8y and y = 1.20y). Based on the visual comparison made in Fig. 5,
values: XGBoost can automatically handle missing values during it could be realized that tree-based models (i.e., DT, RF and XGBoost)
training and prediction, reducing the need for explicit data preprocess­ return more accurate predictions compared to KNN. Otherwise stated,
ing; (ii) feature importance: It provides a measure of feature importance, the error (|y − y|) of almost all the frequencies predicted by DT, RF, and
aiding in feature selection and understanding the model’s decision- XGBoost is less than 20%. It could also be understood that the fre­
making process; (iii) parallel processing: XGBoost efficiently utilizes quencies predicted by XGBoost are more than those predicted by Dt or
parallel processing capabilities to accelerate training, making it suitable RF, which reflects higher reliability of XGBoost compared to other Tree-
for large datasets; (iv) scalability: XGBoost is designed to handle large based techniques.
datasets and can be efficiently parallelized for distributed computing. For better comparison of the models’ performance, the residual of
It should be noted that like DT and RF, maximum number of esti­ predicted values by each model (difference between each predicted and
mators and tree depth affect the XGBoost accuracy significantly which real value divided by its corresponding real value: |y − y|/y) is illustrated
are limited to 40 and 10, respectively. in Fig. 6. The training and testing databases are distinguished by green
and pink colors, respectively. It should be explained that the labels of
4.4. K-Nearest Neighbors data in a model are not the same as the labels of data in the other models
because training and testing sub-databases are selected randomly in
KNN (k-Nearest Neighbors) is a simple and intuitive machine order to eliminate any human effects on the models accuracy. Consid­
learning algorithm used for both classification and regression tasks. KNN ering the diagrams shown in Fig. 6, it could be claimed that all the
is straightforward to understand and implement, making it a good proposed models exhibit a high ability to learn the relationship between
choice for simple tasks and as a baseline for more complex algorithms. inputs and output which results in acceptable prediction reliability.
However, it can be computationally expensive for large datasets, and the
algorithm’s predictions can be influenced by noisy data or outliers in the
5.2. Accuracy assessment
training dataset [39,40].
It operates based on the principle of similarity and uses the notion
In order to evaluate the models more accurately, they are assessed
that data points with similar features tend to have similar outcomes.
through performance metrics and Taylor diagram.
Otherwise stated, KNN is a non-parametric algorithm that makes pre­
Most common performance metrics generally used for assessing the
dictions by finding the "k" training examples in its dataset that are closest
accuracy of a prediction model are R2-score, R-score, Root Mean Square
to a given test data point. The algorithm then assigns a class label (in
Error (RMSE), and Mean Absolute Error (MAE). These metrics are
classification) or calculates an average (in regression) based on the
mainly based on the difference between a predicted value (y) and the
outcomes of the k-nearest neighbors, providing a prediction for the test
real value (y), as could be realized from the Eqs. (14)–(17). Table 4
data point. As could be understood from the definition, the parameter "k"
reports the performance metrics of the proposed models.
is the most critical factor that influences the algorithm’s performance
significantly. Fig. 4 shows the number of “k” against the accuracy ob­ ∑
(̂y − yi )2
tained for both training and testing data. As also illustrated in Fig. 4, by R2 = 1 − ∑i i 2
(14)
i (yi − yi )
increasing the number of neighbors, the accuracy of prediction de­
creases. However, for high “k” values, the model might be overfitted. √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅

(̂y − yi )2
Considering the curve displayed for both training and testing databases, R = 1 − ∑i i 2
(15)
i (yi − yi )

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1∑ n
RMSE = (̂y − yi )2 (16)
n i=1 i

1∑ n
MAE = y − yi |
|̂ (17)
n i=1 i

The R2-score values of training databases given in Table 4 confirms


the high capability of the models for learning the relationship between
input variables and the output. The R2-scores obtained for the testing
databases of each model clarifies that the models are not overfitted and
are sufficiently reliable to be used for predicting frequency of any tower.
As discussed before, Tree-based model with higher R2-score is more
accurate than the KNN model. The XGBoost model with the highest R2-
score (0.97 and 0.85 for respectively training and testing databases)
Fig. 4. Number of neighbors (K) versus R2-score. exhibits the highest accuracy among all models. Considering the metrics

6
H. Dabiri et al. Engineering Structures 304 (2024) 117621

Fig. 5. Correlation between the real frequencies and those predicted by ML-based models.

of DT and RF, it could be realized that the RF results are slightly better 6. XGBoost vs commonly used equations
than those of the DT model. The KNN model, on the other hand, showed
the least accuracy (with R2 =0.70 for both training and testing In this section, the frequencies of towers in the collected database
databases). and predicted by the most accurate model (i.e., XGBoost) are compared
The developed models are deeper compared through Taylor diagram to those calculated by the equations available in design codes ([24], [25]
displayed in Fig. 7. Taylor diagram is an excellent approach for and [26]) and literature ([14]). Fig. 8 provides the correlation between
comparing performance of various models since it considers three pa­ the real values, and the frequencies given by NTC08, NCSE, ASCE and
rameters, namely strander deviation (vertical and horizontal axis), R2- Ranieri and Fabbrocino as well as the values predicted by XGBoost.
score (arc axe), and RMSE (circular curves centered at the real values) R2-score of each equation is also given in the corresponding diagram.
[41]. More precisely, models are positioned in a 2D space based on their Based on the visual comparison made in Fig. 8, it could be clearly
standard deviation, R2-score and RMSE, and their accuracy is assessed observed that the XGBoost model (with R2 =0.95) returns the most ac­
based on their distance to the real value (shown by a green star here). curate values compared to the results of equations. Consequently, this
The closer to the real value the model is, the more accurate it will be method could be introduced as an adequately reliable substitution for
[42]. Taking Fig. 7 into consideration, the XGBoost (brown point) is the design codes which provides equations for calculating approximate
closest model to the green star, and therefore is the most accurate model. frequency. It is noteworthy that the huge difference between the ML-
The KNN model (gray point) is the most far point to the real values, and based model and the given equation is because of the simplified equa­
is known as the least accurate model compared to other methods. tions which consider only height (or height and plan width) of a tower
Overall, it could be articulated that all the proposed models exhibi­ while other parameters like elasticity modulus and plan dimensions are
ted an acceptable level of accuracy (over 0.70) for predicting frequency also taken into account in ML models.
of masonry towers. The XGBoost model, however, possesses the highest It should be also stated that the equation proposed by Ranieri and
reliability, and is introduced as the most accurate ML-based model. Fabbrocino [14] with R2-score= 0.56 is more reliable than the equations
provided in the considered design codes. Furthermore, among the

7
H. Dabiri et al. Engineering Structures 304 (2024) 117621

Fig. 6. Comparison of the real frequency with its corresponding value predicted by different models.

Table 4
Performance metrics of the proposed models.
DT RF XGBoost KNN

Training Testing Training Testing Training Testing Training Testing

R2 0.93 0.88 0.95 0.70 0.97 0.85 0.70 0.70


R 0.96 0.94 0.97 0.84 0.98 0.92 0.84 0.84
RMSE 0.17 0.22 0.17 0.37 0.12 0.23 0.36 0.32
MAE 0.09 0.17 0.10 0.23 0.06 0.18 0.25 0.24

Fig. 7. Comparing the prediction models by Taylor diagram for training (left) and testing (right) databases.

considered Codes, the frequencies calculated by NCSE [26] (which 7. Sensitivity analysis
considers both height and plan width of a tower) are slightly more ac­
curate than those calculated by either NTC08 [25] or ASCE [24]. This The influence of height and plan dimensions of a tower’s frequency is
point clarifies that plan dimensions affect a tower’s frequency and investigated deeper through a sensitivity analysis in this section. To this
therefore is recommended to be considered for obtaining the frequency. end, San Marco tower located in Venice, Italy (see Fig. 1) is considered
as a case study (first row of Table 5). A database including 13 data is
generated by changing the height and plan width of the tower, as given
in Table 5. In the first 7 data (rows 2–8 of Table 5), the tower’s height is

8
H. Dabiri et al. Engineering Structures 304 (2024) 117621

Fig. 8. Comparing the accuracy of equations given in design codes and literature to that of XGBoost.

XGBoost model, which is the most accurate model, is used for obtaining
Table 5
the frequency of the tower. The results are given in Table 5 and shown in
The database generated for sensitivity analysis.
Fig. 9.
E (MPa) Height (m) Plan width (m) In terms of tower’s height, by increasing the height from 20 m to
1 1500 98 8.16 40 m, frequency decreases significantly (see Fig. 9a). The frequency of
2 1500 20 8.16 tower reduces slightly for height changes from 50 to 70, and eventually
3 1500 30 8.16 for very tall towers (H≥70 m) the frequency remains almost unchanged.
4 1500 40 8.16
5 1500 50 8.16
This reverse relationship between height and frequency is also reported
6 1500 60 8.16 in Table 3 in which Pearson and Spearman correlation coefficients are
7 1500 70 8.16 given.
8 1500 80 8.16 For more proper analysis of tower frequency when its plan area
9 1500 98 5
varies, the ratio of width (the smaller length of tower cross section) to
10 1500 98 6
11 1500 98 7 height (wtower/htower) is taken into consideration. It should be explained
12 1500 98 8 that the XGBoost model predicts the frequency based on the tower
13 1500 98 9 width, and then for discussing the results the ratio is considered. As
14 1500 98 10 could be observed in the diagram shown in Fig. 9(b), by increasing
wtower/htower the frequency increases. Similar to the effect of height on
changed from 20 m to 80 m by the step of 10 while its plan width frequency, by increasing tower width frequency increases slightly (up to
remained unchanged (8.16 m). For the other 6 data (rows 9–14), the wtower/htower≈0.21), and then jumps remarkably. As mentioned in Sec­
plan width is altered in the range of 5–10 m with the steps of 1 but their tion 7, the equation considering both height and plan width returns
height is fixed to 98 m. It is worth noting that the case study considered more accurate results. Consequently, an equation is proposed based on
for sensitivity analysis is not tapered along its height. The proposed the frequency change when the wtower/htower ratio changes for rectan­
gular cross section towers:

9
H. Dabiri et al. Engineering Structures 304 (2024) 117621

Fig. 9. Variation of frequency by changing (a) tower’s height, and (b) width to height ratio, (c); comparing the real frequencies to the frequencies calculated by Eq.
11 for 29 randomly selected towers (30% of the database).

w
F(w, h) = 6.0342( )0.867 (18) recommended to be included in equations used for calculating fre­
h quency in further revision of design codes.
• Methods which are simple to be implemented and highly accurate
where w and h are respectively the width (the smaller length of tower
were utilized in this study. Further studies are necessary to evaluate
cross section) and height of a tower.
(i) the accuracy of other ML-based techniques (e.g., Artificial Neural
In order to evaluate the capability of Eq. (18) for calculating reliable
Network, Regression-based approaches, Support Vector machine,
frequency of towers, 29 databases are selected randomly. The frequency
etc.) for predicting frequency of towers, (ii) the influence of other
of each tower is calculated by Eq. (18) and compared by its corre­
parameters, namely opening ratio, material type (e.g., stone, brick,
sponding real frequency given in the database. According to the com­
etc.), tower geometry (e.g., geometrical tapering along the height),
parison made in Fig. 9(c) and considering the R2-score of the frequencies
tower constraints (e.g., existing other buildings beside the tower),
calculated by Eq. (18) (=0.75), it could be stated that the proposed
soil type, wall thickness, the widespread presence of cracks intro­
equation returns sufficiently accurate results. It is worth recalling that
ducing non-linearity and anisotropy effects, and presence of internal
accuracy of the equations available in current design codes is less than
floors and walls on the accuracy of prediction models, (iii) the
that of Eq. (18) (see Fig. 8).
capability of ML-based models for predicting other towers properties
(e.g., flexural Eigen-frequencies and torsional Eigen-frequency).
8. Summary and conclusions
CRediT authorship contribution statement
In this paper, frequency of masonry towers was analyzed through
ML-based techniques. Firstly, a database including details of over 100
Mugnozza Gabriele Scarascia: Conceptualization, Project admin­
towers was used to develop ML-based models, namely DT, RF, XGBoost,
istration, Supervision. Bozzano Francesca: Conceptualization, Project
and KNN. The models were compared, and the most accurate model was
administration, Supervision. Mazzanti Paolo: Conceptualization, Proj­
introduced. In the second part of the study, the values predicted by the
ect administration, Supervision. Dabiri Hamed: Conceptualization,
highest accurate model were compared to those calculated by the
Data curation, Formal analysis, Investigation, Methodology, Software,
equations available in either design codes or literature. Lastly, a sensi­
Validation, Writing – original draft, Writing – review & editing. Clem­
tivity analysis was performed for evaluating the impact of height and
enti Jessica: Conceptualization, Investigation, Writing – original draft,
cross section width of towers on first natural frequency. Consequently,
Writing – review & editing. Marini Roberta: Conceptualization, Data
an equation was proposed according to the results of sensitivity analysis.
curation, Investigation, Resources.
The main conclusions are listed below:

• All the ML-based methods exhibit high capability (with R2-score over Declaration of Competing Interest
70%) of learning the relationship between input and output pa­
rameters for predicting masonry towers natural frequency. The authors declare that they have no known competing financial
• Tree-based models (i.e., DT, RF and XGBoost) exhibited higher ac­ interests or personal relationships that could have appeared to influence
curacy than KNN method. Furthermore, the XGBoost model the work reported in this paper.
possessed the highest R2-score among all the models and could be Conflict of Interest and Authorship Conformation Form.
introduced as the most accurate method for predicting frequency of
masonry towers. • All authors have participated in (a) conception and design, or anal­
• ML-based models returned more accurate results compared to those ysis and interpretation of the data; (b) drafting the article or revising
obtained by available equations in design codes (e.g., ASCE, NTC, it critically for important intellectual content; and (c) approval of the
and NCSE) and literature. The reason is that available equations are final version.
based on height and plan width while ML-based models consider • This manuscript has not been submitted to, nor is under review at,
other effective parameters like plan dimensions in both directions another journal or other publishing venue.
and modulus of Elasticity. • The authors have no affiliation with any organization with a direct or
• Based on the reported correlation coefficients, and performed indirect financial interest in the subject matter discussed in the
sensitivity analysis in this study, the effect of cross section width on a manuscript
frequency tower is not negligible, and therefore is highly • The following authors have affiliations with organizations with
direct or indirect financial interest in the subject matter discussed in
the manuscript:

10
H. Dabiri et al. Engineering Structures 304 (2024) 117621

Data availability [20] Ivorra S, et al. A preliminary approach of dynamic identification of slender
buildings by neuronal networks. Int J Non-Linear Mech 2016;80:183–9.
[21] Younsi S, et al. Reconstructing missing InSAR data by the application of machine
Data will be made available on request. leaning-based prediction models: a case study of Rieti. J Civ Struct Health Monit
2023:1–19.
References [22] Standoli G, et al. Modal-based FE model updating via genetic algorithms:
Exploiting artificial intelligence to build realistic numerical models of historical
structures. Constr Build Mater 2021;303:124393.
[1] Stewart, J., et al., Engineering Reconnaissance of the 24 August 2016 Central Italy [23] Asteris PG, et al. Prediction of the fundamental period of infilled RC frame
Earthquake. Version 2. 2016. structures using artificial neural networks. Comput Intell Neurosci 2016;2016:20.
[2] D’Ayala, D., et al. Earthquake Engineering Field Investigation Team (UK)–Data and 20.
Discussion on the 2016 Central Italy Mission. 2019. 16th World Conference on [24] A.S.A.S.E.I., “Minimum designloads and associated criteria for buildings and other
Earthquake Engineering (16WCEE). structures.” ASCE 7-16, 2013.
[3] Poiani M, et al. Iconic crumbling of the clock tower in Amatrice after 2016 central [25] NTC2008, Norme tecniche per le costruzioni, D.M. 14/01/2008 , Gazzetta Ufficiale
Italy seismic sequence: advanced numerical insight. Procedia Struct Integr 2018; n. 29 del 04.02.2008, Suppl. Ord. n.30; 2008. 2008.
11:314–21. [26] NCRS-02, Norma de Construccion Sismorresistente: Parte General y Edification. Real
[4] Putrino, V., et al., The Mw 6.2 Amatrice, Italy Earthquake of 24 August 2016-A field Decreto 997/2002; de 27 Septiembre 2002. 2002.
report by EEFIT . 2019. [27] Pallarés FJ, et al. Structural health monitoring (SHM) and Nondestructive testing
[5] Sevieri G, et al. A multi-hazard risk prioritisation framework for cultural heritage (NDT) of slender masonry structures: A practical review. Constr Build Mater 2021;
assets. Nat Hazards Earth Syst Sci 2020;20(5):1391–414. 297:123768.
[6] Yanık Y, et al. Vibration testing and performance evaluation of Hagia Sophia bell [28] Standoli G, et al. Model updating of historical belfries based on oma identification
tower after recent restoration. Constr Build Mater 2022;347:128617. techniques. Int J Archit Herit 2021;15(1):132–56.
[7] Lacanna G, et al. Dynamic response of the Baptistery of San Giovanni in Florence, [29] Torelli G, et al. Analytical and numerical seismic assessment of heritage masonry
Italy, based on ambient vibration test. J Cult Herit 2016;20:632–40. towers. Bull Earthq Eng 2020;18(3):969–1008.
[8] Baraccani S, et al. Long-term seismometric monitoring of the two towers of [30] Porcu MC, Montis E, Saba M. Role of model identification and analysis method in
Bologna (Italy): modal frequencies identification and effects due to traffic induced the seismic assessment of historical masonry towers. J Build Eng 2021;43:103114.
vibrations. Front Built Environ 2020;6:85. [31] Dabiri H, et al. A machine learning-based analysis for predicting fragility curve
[9] D’Ambrisi A, Mariani V, Mezzi M. Seismic assessment of a historical masonry tower parameters of buildings. J Build Eng 2022;62:105367.
with nonlinear static and dynamic analyses tuned on ambient vibration tests. Eng [32] Kioumarsi M, et al. Compressive strength of concrete containing furnace blast slag;
Struct 2012;36:210–9. optimized machine learning-based models. Clean Eng Technol 2023;13:100604.
[10] Ferraioli M, Miccoli L, Abruzzese D. Dynamic characterisation of a historic bell- [33] Nisbet, R., J. Elder, and G.D. Miner, Handbook of statistical analysis and data mining
tower using a sensitivity-based technique for model tuning. J Civ Struct Health applications. 2009: Academic press.
Monit 2018;8:253–69. [34] Myles AJ, et al. An introduction to decision tree modeling. J Chemom: A J Chemom
[11] Bartoli G, et al. Semiempirical formulations for estimating the main frequency of Soc 2004;18(6):275–85.
slender masonry towers. J Perform Constr Facil 2017;31(4):04017025. [35] Breiman L. Random forests. Mach Learn 2001;45:5–32.
[12] Bartoli G, et al. On the role played by the openings on the first frequency of historic [36] Amit Y, Geman D. Shape quantization and recognition with randomized trees.
masonry towers. Bull Earthq Eng 2020;18(2):427–51. Neural Comput 1997;9(7):1545–88.
[13] Shakya M, et al. Empirical formulation for estimating the fundamental frequency of [37] Chen T, et al. Xgboost: extreme gradient boosting. R Package Version 0 4-2 2015;1
slender masonry structures. Int J Archit Herit 2016;10(1):55–66. (4):1–4.
[14] Rainieri, C. and G. Fabbrocino. Il periodo elastico delle torri in muratura: correlazioni [38] Chen, T. and C. Guestrin. Xgboost: A scalable tree boosting system. in Proceedings of
empiriche per la previsione. in XIV Congresso Nazionale L’Ingegneria Sismica in Italia, the 22nd acm sigkdd international conference on knowledge discovery and data mining.
Bari . 2011. 2016.
[15] Diaferio M, Foti D, Potenza F. Prediction of the fundamental frequencies and modal [39] Zhang M-L, Zhou Z-H. ML-KNN: A lazy learning approach to multi-label learning.
shapes of historic masonry towers by empirical equations based on experimental Pattern Recognit 2007;40(7):2038–48.
data. Eng Struct 2018;156:433–42. [40] Deng Z, et al. Efficient kNN classification algorithm for big data. Neurocomputing
[16] Mishra M. Machine learning techniques for structural health monitoring of heritage 2016;195:143–8.
buildings: A state-of-the-art review and case studies. J Cult Herit 2021;47:227–45. [41] Taylor KE. Summarizing multiple aspects of model performance in a single
[17] Abed MM, El-Shafie A, Osman SAB. Creep predicting model in masonry structure diagram. J Geophys Res: atmospheres 2001;106(D7):7183–92.
utilizing dynamic neural network. J Comput Sci 2010;6(5):597. [42] Dabiri H, Rahimzadeh K, Kheyroddin A. A comparison of machine learning-and
[18] Calledda C, Montisci A, Porcu MC. Optimal design of earthquake-resistant regression-based models for predicting ductility ratio of RC beam-column joints.
buildings based on neural network inversion. Appl Sci 2021;11(10):4654. Structures 2022 (Elsevier).
[19] Testa F, Barontini A, Lourenço PB. Development and Validation of Empirical
Formulations for Predicting the Frequency of Historic Masonry Towers. Int J Archit
Herit 2023:1–21.

11

You might also like