Structures 34 (2021) 4435–4446

Contents lists available at ScienceDirect

Structures
journal homepage: www.elsevier.com/locate/structures

A methodological approach towards evaluating structural damage severity

using 1D CNNs
Mohammad Almutairi a, Nikolaos Nikitas a, *, Osama Abdeljaber b, Onur Avci c, Mateusz Bocian d
a
School of Civil Engineering, Leeds University, LS2 9DY Leeds, UK
b
Department of Building Technology, Linnaeus University, 351-95 Växjö, Sweden
c
Civil, Construction and Environmental Engineering, Iowa State University, IA 50011, United States
d
Wrocław University of Science and Technology, Faculty of Civil Engineering, 50-421 Wrocław, Poland

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

Keywords: Evaluating the severity of structural damage is a critical component of Structural Health Monitoring (SHM).
Damage evaluation CNNs Convolutional Neural Networks (CNNs) have been used before to detect structural damage and evaluate its
Deep learning model updating severity by utilising only raw vibration data. However, these vibration-based CNN applications were limited to
Dynamic monitoring
discrete user-defined levels of damage. To provide a more accurate representation of structural damage, this
paper aims to design and validate a framework for evaluating structural damage severity within a continuous
range of damage levels, using 1D CNNs and distributed raw acceleration data. To this purpose, a simple Finite
Element (FE) cantilever model with non-rigid rotational spring support was adopted. Damage was simulated at
the support as reduction of the rotational spring stiffness. The performance of the proposed framework was
assessed under different excitation scenarios and data pre-processing techniques. The results demonstrate the
ability of 1D CNNs to evaluate damage severity with high accuracy. By estimating the reduced value of the
rotational spring stiffness, the proposed framework can also be used towards FE model updating in parallel with
damage severity evaluation.

1. Introduction with Neural Networks (NNs) to detect structural damage. Similarly,

natural frequencies and mode shape curvature were used to detect
Integrity and safety of civil infrastructures are essential for the so­ damage using NNs [20]. The main challenge of utilising ML tools on SDD
ciety to well function. During their service life, civil infrastructures are is that DSFs have to be extracted manually in advance.
susceptible to several types of damage due to operational and environ­ The advent of Deep Learning (DL), which is an advanced subfield of
mental factors. Frequent structural damage detection (SDD) is essential ML, tackled this challenge by combining the feature extraction task with
to avoid structural failure and posing risk to public safety. the classification task into one automated task. Many researchers used
The most conventional SDD method, which is no other than regular DL algorithms (e.g. Convolutional Neural Networks -CNNs-) to detect
visual inspection, highly depends on an engineer’s expertise. Besides its structural damage by utilising only raw vibration data [21–24].
subjectivity, visual inspection is also expensive, time consuming, and Teng et al. [24] detected and localised structural damage of a labo­
limited to accessible structural members. To avoid these drawbacks, ratory steel frame using a CNN approach, which was trained by nu­
vibration-based SDD has been proposed and applied in recent years merical vibration data obtained by a Finite Element (FE) model of the
[1–16]. frame, and validated by experimental vibration data of the laboratory
The hypothesis of any vibration-based SDD is that the occurrence of frame. They simulated the numerical damage by reducing the elastic
damage alters structural vibration features. Common damage-sensitive modulus of the frame rods to half, to simulate 50% damage. Then, the
features (DSFs) are natural frequencies, mode shapes, and their de­ structure was excited by impact hammer and each excitation was
rivatives [17,18]. Machine learning (ML) algorithms have been utilised repeated five times to acquire sufficient data. This process (i.e. intro­
to detect structural damage by learning from the evolution of DSFs. For ducing damage and exciting the damaged structure) was repeated for
example, Mehrjoo et al. [19] used natural frequencies and mode shapes several locations on the frame. Then, a 1D CNN algorithm was trained to

* Corresponding author.
E-mail address: [email protected] (N. Nikitas).

https://doi.org/10.1016/j.istruc.2021.10.029
Received 28 February 2021; Received in revised form 12 September 2021; Accepted 8 October 2021
Available online 23 October 2021
2352-0124/© 2021 Institution of Structural Engineers. Published by Elsevier Ltd. All rights reserved.
M. Almutairi et al. Structures 34 (2021) 4435–4446

classify the damaged/undamaged rod cases. The algorithm showed good of evaluating structural damage severity using 1D CNN is explained
accuracy, 90%, on detecting and localising damage on the laboratory along with related works which include SDD and FE model updating
steel frame. The results indicate that 1D CNNs can learn from numerical applications using CNNs and NNs, respectively, to draw similarities for
responses of a simulated structure to detect damage in a laboratory the proposed approach. A numerical example was created and validated
equivalent when tested. in Section 4 to be adopted in sensitivity analysis, in which the perfor­
Similarly, Zhang et al. [21] used 1D CNNs to classify damaged/un­ mance of the proposed approach was investigated under different sce­
damaged members for three different structures; two laboratory struc­ narios including different number of sensors, excitation types, signal
tures and an in-service steel girder bridge. The damage was simulated in durations, frame sizes, and input scaling methods.
the first laboratory structure, an I-section steel beam, by adding magnets
on the web. In the second laboratory structure, a short steel girder 2. Overview of 1D convolutional neural networks (1D CNNs)
bridge, the damage was simulated in two ways: fixing of a steel plate on
the lower flange by two clamps and attaching mass blocks by magnets on CNNs are a class of feedforward artificial Neural Networks (NNs) that
the upper flange. Both laboratory structures were excited by impact are biologically inspired by the human visual cortex [27]. They are
hammer, at different locations, and each excitation was repeated 200 typically composed of two stages: feature extraction and classification/
times. The response under each excitation was used as input data for the regression [28]. The 1D CNN architecture is shown Fig. 1. This figure
1D CNNs. On the other hand, the damage on the in-service steel girder was inspired by previous works [29–32] and refers to CNN details used
bridge was simulated by attaching magnets on the lower flange. The hereafter.
bridge was excited by traffic and white noise excitation, on the lower The feature extraction stage consists of stacking and alternating
flange, for 40 min. The acceleration response was divided into frames, convolutional layers and pooling layers. The input data is convolved by
where each frame has 20,000 sampled data points representing 2-sec­ kernel filters. Each filter shares the same matrix of learnable weights
ond-long data, to be used as input. The algorithm shows a high accu­ over the entire input data to extract certain feature. The kernel filters
racy (97.2%− 99.5%) in damage detection and localisation. convolve along one dimension (i.e. vertical direction) and through all
Abdeljaber et al. [22] detected structural damage at joints of a lab­ input channels, as shown in Fig. 1. For this reason, the algorithm is
oratory steel grandstand-like frame using 1D CNNs. The damage was called 1D CNN. The convolution is calculated as follows [33,34]:
simulated by loosening the bolts at different joints, separately. Then, the
Feature mapcn = f (Input data*Wcn + bcn ) (1)
structure was excited by white noise excitation for 256 s. The acceler­
ation response, of each joint, was divided into frames to form input data. where * represents the convolution operation, Wcn is the weight matrix
The 1D CNN algorithm was trained by data of single damage cases (i.e. of the cnth kernel filter, bcn is the corresponding bias, and f(.) represents
one damaged joint at each case), and then was tested to detect double the activation function. The output of the convolutional process is the
damage cases (i.e. two damaged joints at each case). This study shows cnth feature map. It is worth noting that there are different types of
the ability of 1D CNNs to learn from a set of damage cases and apply the activation function such as the Rectified Linear Unit (ReLU), the sig­
learning on a different set of damage cases. moid, the hyperbolic tangent, the Scaled Exponential Linear Unit (SELU)
The aforementioned works were limited to determine whether a and others [35].
structure’s member is damaged or not. In other words, they do not As the data dimension increases by generating several numbers of
evaluate the damage level. To tackle this issue, Khodabandehlou et al. feature maps, pooling layer are used to reduce data dimension. The
[25] defined four levels of damage and trained 2D CNNs to detect each pooling layer is also adding robustness to small variance in the feature
level. A laboratory model of a reinforced concrete bridge, mounted on a maps. The most common pooling operators are max pooling and average
shaking table, was adopted in their study of different types of damage (e. pooling [33]. In a max pooling layer, for example, the max pooling
g. crack and buckling). White noise excitation was applied to the model. operator extracts the maximum value of a certain region on the feature
The synchronous acceleration measurements of 14 accelerometers were map. A sliding window defines this region over the feature map.
used as input data. The results show a high accuracy (87.5%–100%), on Following the feature extraction stage, the Fully Connected (FC)
identifying and classifying the damage into one of the four pre-defined layers process the extracted feature maps to conduct the classification/
levels. regression task. The FC layers are the classical NNs with learnable
Although CNN algorithms show good performance on classifying weights and biases, and an activation function [36]. In this stage, the
levels of damage, they are limited to the user-defined discrete levels estimation error is calculated by a loss function such as the Mean
because they are inherently trained to categorise the damage level to one
Squared Error (MSE). An optimisation function, such as Adam by
of the discrete levels only. The classification mechanism compromises Kingma and Ba [37], reduces this error by updating the learnable
the accuracy of evaluating the actual severity of damage. Almutairi et al.
weights and biases. The initial values of the learnable weights and biases
[26] proposed a framework for estimating a critical FE model parameter are randomly assigned to improve the learning process of the 1D CNNs
that represents structural damage within a continuous range of light
[38]. The optimisation process is conducted in supervised manner over a
damage levels using 1D CNNs and synthetic raw acceleration data. The dataset that contains input–output samples [36].
proposed framework was evaluated under a single scenario of excitation
and data pre-processing technique. To this end, this study is aiming to 3. 3. Training and testing the 1D CNNs
expand and evaluate the proposed framework thoroughly under
different scenarios of excitation and data pre-processing techniques. A The proposed CNN-based framework of evaluating structural dam­
simple FE model of a cantilever beam, with a non-rigid rotational spring
age consists of four stages: generating training data, pre-processing the
stiffness support, was adopted in this study. The structural damage levels training data, designing and training the 1D CNNs, and post-processing
at the support were simulated by reducing the stiffness of the rotational
the output data as illustrated in Fig. 2.
spring. A dynamic load was applied on the free end of the beam to Generating sufficient training data plays a critical role in a 1D CNN’s
generate multi-point synthetic acceleration responses.
performance. The training data include input–output samples. The
Since the proposed framework evaluates damage severity in the form output data are damage levels, represented by a variation on the value of
of reduction of an FE model parameter within a continuous range, this a FE model parameter (Pn ), while the input data are the corresponding
framework has the potential to be used towards FE model parameter acceleration responses for each damage level.
estimation for FE model updating purposes. The variation of FE model parameters should be selected carefully to
This paper is organised as follows. Background on 1D CNNs is pro­ reflect a wide range of conditions. For example, in the study of Levin and
vided at the beginning in Section 2. In Section 3, the proposed approach

4436
M. Almutairi et al. Structures 34 (2021) 4435–4446

Fig. 1. Indicative architecture and hyperparameters of 1D CNNs; here as also used later in part of the study.

Fig. 2. Illustration of data generation, data pre-processing, and 1D CNNs training and testing.

Lieven [39] on a numerical cantilever beam, the variation of the Young’s support of a cantilever beam as a rotational spring. The spring stiffness
moduli was assumed to range between 118% and 75% of the design that produces similar dynamic behaviour, namely the frequency of the
value and 123% to 78% for the density. In another numerical study, Lu first bending mode, to the fixed cantilever beam was assumed as the
and Tu [40] modelled a six-storey frame as a lumped-mass stick and undamaged rotational spring-supported beam. Different levels of light
assumed 100% to 35% variation of the inter-storey sway stiffness rep­ damage were simulated by reducing the spring stiffness to different
resenting the range of undamaged to severely damaged structure. A values. Similarly, Hasançebi and Dumlupınar [42] simulated the
scaled suspension steel bridge was used by Chang et al. [20], and eight boundary conditions of a reinforced concrete bridge as linear springs in
FE model parameters were targeted; including rotational spring stiffness all three directions with different stiffness in each.
representing boundary conditions. They considered three levels for each To define an adequate range for the spring stiffness value, they
model parameter: 75%, 100% and 125% of their design values. The determined the first three natural frequencies while the value of each
work of Park et al. [41], on a steel girder bridge, resolved the issue of spring stiffness increased, until an insignificant change was observed in
unavailable design boundary conditions for the supports, which were the natural frequencies. The range of spring values was restricted to a
simulated as rotational springs. The spring values were assumed to vary range corresponding to a highly sensitive natural frequency to spring
between two support conditions, fixed and pin; values were acquired by changes. After selecting the levels of FE model parameters within the
increasing the spring value from zero and identifying the first natural defined range, the structure’s acceleration response was generated for
frequency before this converged to the fixed support case. Like Park each level forming the input data.
et al. [41], the numerical work of Almutairi et al. [26] simulated the Several data pre-processing techniques are typically used to improve

4437
M. Almutairi et al. Structures 34 (2021) 4435–4446

the performance of 1D CNNs, data splitting, data augmentation, data Once the algorithm is fed by the actual structure’s acceleration
scaling, and data shuffling. In data splitting, part of the training data response, in the form of frames, it estimates the scaled value of the FE
(hereafter called testing data) are not included in the training process. model parameter for each frame. Then, the scaled estimations are
Data splitting ensures a proper evaluation of the 1D CNNs’ learning averaged and de-scaled to evaluate the actual level of damage (i.e. the
process. Data augmentation aims to increase the number of training actual value of the FE model parameter) as shown in Fig. 3.
data. This is often used when the available training data are limited [33].
Elsewhere, the original time series can be sliced to several frames having 4. Numerical example
the same output value. This technique (called window slicing technique)
was applied in civil engineering applications by several researchers A numerical example is used in this section to evaluate the efficiency
[21,22,43]. In case of using multiple accelerometers, the data reshaping of 1D CNNs on determining the severity of simulated structural damage.
technique prepares the input data to conduct the 1D convolutional A simple 2D FE model of an I-section cantilever beam is created and
process. This technique converts the input data from single channel with simulated using MATLAB. It is made up of steel with Young’s modulus
multiple columns (e.g. 100 × 4 × 1) to multiple single column channels 210 GPa, density 7850 kg/m3, Poisson’s ratio 0.3, and 1% damping
(e.g. 100 × 1 × 4), as shown in Fig. 2. ratio. The cantilever beam’s length is chosen as 2.5 m to minimise the
The subsequent technique is scaling the output/input data to prevent shear deflection and better validate the FE model analytically using
the 1D CNNs of being biased towards the extreme data values [44]. Two standard Euler-Bernoulli beam theory [47,48].
methods of scaling the input acceleration data, of each frame, were re­
ported in the literature: the data normalisation, and the data stand­ 4.1. Validation of the FE model
ardisation method. The first method was used by Abdeljaber et al. [22],
and the acceleration data were scaled to range between 1 to − 1. In The FE model was validated in this section by comparing its static
contrast, Wang and Cha [45] used the data standardisation method, and dynamic behaviour against theoretical expectations in order to
which scales the data to a zero-mean with unit standard deviation. For investigate its reliability [26]. First, a static vertical load of 2 kN was
scaling the output data (i.e. FE model parameter), Lu and Tu [40] used a applied at the free end of the cantilever, and the maximum deflection
range of (0,1). Finally, the scaled frames have to be shuffled randomly to was determined numerically as 13.55 mm and analytically as 13.44,
improve the learning process of the 1D CNNs [46]. which shows only 0.8% error. Subsequently, the first five natural fre­
The next stage is designing the 1D CNNs in a process called hyper­ quencies of the numerical model were identified and compared with
parameter tuning [28]. In this process, different hyperparameters (e.g. their analytical counterparts, as presented in Table 1. The results
number of convolutional layers, type of optimisation function, and showed that the FE model is reliable and can be adopted in this study.
number of FC layers) are adjusted to improve the 1D CNN’s perfor­
mance. A trial-and-error approach of hyperparameter tuning was used
by several researchers in civil engineering [22,25,33,43]. There are 4.2. Generating training data and training the 1D CNNs
some “rules of thumb” that can be used therein. For example, the
number of kernel filters in the last convolution step should be more than The training data includes the input (i.e. acceleration responses) and
their number in the first. Also, it is recommended to use the ReLU output data (i.e. rotational spring stiffness). For the output data,
activation function. For the regression output layer, the linear activation different levels of spring stiffness were considered to simulate different
function and the MSE loss function are recommended [28]. The MSE in levels of structural damage at the support. The damping ratio was
this study is the mean squared difference between the scaled actual and
scaled estimated value of the FE model parameter for each input frame. Table 1
The MSE of training and testing data were calculated throughout the Comparison of the numerical and analytical solution of the cantilever natural
training process. These MSEs are widely used as a diagnostic tool for 1D frequencies.
CNNs and contribute to better hyperparameter tuning and data pre- Natural frequency (Hz) Error %
processing. For example, a large gap between training and testing Numerical solution Analytical solution
MSE, resulting possibly from overfitting, can be resolved using data
19.3 19.3 − 0.2
augmentation techniques [36]. After the algorithm is trained, it can be 121.3 121.5 − 0.2
used to estimate the FE model parameter representing the actual level of 339.5 340.4 − 0.2
damage. 665.3 667.5 − 0.3

Fig. 3. Illustration of estimating damage level using a trained 1D CNNs including data pre-processing, and post-processing the output data.

4438
M. Almutairi et al. Structures 34 (2021) 4435–4446

simplistically assumed to be fixed throughout these scenarios. Graphics Processing Unit (GPU) which accelerates the training process
In the Almutairi et al. [26] study, the rotational spring value for the [49].
undamaged support case was defined as 1013 N.mm/rad; for this the first Selected architecture and hyperparameters of the 1D CNNs are
natural frequency closely approximates the corresponding natural fre­ indicatively presented in Fig. 1. The ReLU activation function was used
quency of the fixed cantilever beam. Then, different reduction levels of for all convolutional layers and FC layers, while the linear activation
this spring stiffness value were used to simulate different levels of function was used for the output layer. The error of the 1D CNNs’ esti­
damage, with the lowest spring stiffness value being 1011 N.mm/rad (see mations was determined by the MSE loss function, after feeding the al­
Fig. 4). The drawback of this approach is that it does not consider the gorithm with 64 frames (i.e. mini-batch size) a time. With a 0.0001
range where the natural frequency is highly sensitive to the reduction of learning rate, the Adam optimisation function was used to reduce the
spring stiffness and that represents the widest range of damage levels. In error by updating the learnable weights and biases. The training process
consequence, the range of spring stiffness is severely limited to light was conducted for 80 epochs. The training process was carried out ten
damage with little impact to dynamic properties. To determine the new times, and the CNN with the best performance (i.e. lowest MSE) over the
range of the spring stiffness, inspired by Hasançebi and Dumlupınar training/testing data was selected.
[42], the beam’s first natural frequency was evaluated as the stiffness While MSE was used as loss function, the Mean Absolute Percentage
values increased from 107 N.mm/rad to a point where convergence was Error (MAPE) was used in this study as a metric of evaluating the 1D
reached. As shown in Fig. 4, the first natural frequency is very sensitive CNNs’ performance, because it is more comprehensible than the MSE.
to spring stiffness change up to the value of 1011 N.mm/rad. For this This was conducted by averaging the scaled estimations of each testing
reason, 1011 N.mm/rad was set as the upper bound of the envisaged case, de-scaling the averaged values, and then converting them from the
range, while the lower bound was assumed to be 109 N.mm/rad. logarithmic scale to the linear scale to estimate spring stiffness values.
All levels of simulated damage were selected within this range. Fig. 4 The percentage error, for each testing case, was then calculated. Finally,
shows the nineteen values of stiffness, including the maximum and the the mean absolute of these percentages was determined to represent the
minimum, that were considered as training data. For testing data, eight MAPE.
cases were selected spread throughout the range to test the algorithm in
different levels of damage severity. The higher value of spring stiffness 4.3. Sensitivity analysis
corresponds to the lower level of the damage, and vice versa.
For the input data, vertical acceleration responses were measured at Different factors can influence the 1D CNNs’ performance on eval­
400 Hz by four virtual accelerometers located at equidistant locations, uating structural damage. Some of these factors can be categorised as
as shown Fig. 5. Dynamic excitation was applied at the free end of the excitation-related (e.g. type, location, and duration of excitation), data
cantilever beam from zero initial conditions. acquisition-related (e.g. number of sensors, and sensor’s noise), struc­
Pre-processing the training and testing data included applying data ture’s properties-related (e.g. natural frequency, and damping ratio),
augmentation, data reshaping, data scaling, and shuffling. First, data and data pre-processing related (e.g. size of frame, number of frames,
augmentation was implemented by dividing each matrix of input data to and scaling method).
frames with size 100 × 4 × 1, corresponding to 0.25 s. The four columns In this section, the sensitivity of 1D CNNs to different scenarios that
represent the four accelerometers. The frame row number (i.e. one include the abovementioned categories were investigated. The investi­
hundred) was selected by trial-and-error during the hyperparameter gated scenarios in this study are (1) influence of number of sensors, (2)
tuning task. This frame setting includes only one channel and subse­ influence of exciting natural frequency for harmonic load testing with
quently, applying data reshaping, each frame was converted to 100 × 1 different excitation durations and frame sizes, and (3) influence of
× 4. excitation types with different scaling methods.
Afterwards, the acceleration data (i.e. input data) of each frame were
standardised. The spring stiffness values (i.e. output data) were con­ 4.3.1. Influence of number of sensors
verted to the logarithmic scale and scaled down to the range of (0,1). The influence of the number of measurement points on the 1D CNNs’
Finally, the frames of training and testing data were shuffled performance was first investigated. The investigation was carried out by
independently. forming the input data from different number of measurement points
The main 1D CNN algorithm was designed by a trial-and-error and comparing these cases. In the first case, the input data consist only of
approach and was developed using the Python library Keras [35]. The the acceleration response of the tip of the cantilever beam. The second
algorithm was implemented in Google Colaboratory (Colab), a cloud- and third cases are using the acceleration responses from the first two,
based computing service provided by Google. Colab provides a and three sensors, starting from the beam’s tip, as shown in Fig. 6, while

Fig. 4. Sensitivity of the first natural frequency of the cantilever beam model to change in rotational spring stiffness value.

4439
M. Almutairi et al. Structures 34 (2021) 4435–4446

Fig. 5. (a) Locations of measuring acceleration response on the spring-supported cantilever beam (not to scale), (b) beam’s cross-section, and (c) location
of excitation.

Fig. 7. Learning curve of one sensor case.

Fig. 6. Sensors’ location for the different cases of sensor numbers.

the fourth case is considering the acceleration responses of all the four
sensors.
The acceleration response was generated by exciting the cantilever
beam using monoharmonic excitation of 19 Hz, which is near the beam’s
first natural frequency (see Table 1). The acceleration response was
measured for 100 s. After dividing the signal to frames with size 100,
each frame was scaled by using the earlier explained standardisation
approach.
The case of one sensor shows the highest MSE (0.019), while the four
sensors case shows the non-proportionally lowest MSE (0.00011). The
learning curve of both cases is shown in Figs. 7 and 8, respectively. The
learning curves show good fit learning process, in which the training and
testing MSEs are quickly decreasing to a point of stability, and the gap
between them is minimal.
Fig. 9 shows that increasing the number of sensors can reduce the 1D
Fig. 8. Learning curve of four sensor case.
CNNs’ performance error from 51% for the one sensor case to 2.4% for
the four sensors case. From the same figure, it can be seen that using
multiple sensors has a significant effect on the 1D CNNs’ performance. depth evaluation was conducted by comparing the estimated spring
This observation is consistent with that of Chao et al. [50]. By using stiffness values of the testing cases with their actual counterparts. The
multi-channel input data, more filter weights contribute in the feature estimated value of spring stiffness is the averaged estimation over the
extraction stage. As such, more features can be extracted that improve frames of each testing case. The upper bound of the error band is the
the 1D CNNs’ learning process [50]. maximum estimation, while the lower band is the minimum estimation.
After studying the overall error of the 1D CNNs’ estimations, an in- The estimations with their error bands are presented in Fig. 10. The

4440
M. Almutairi et al. Structures 34 (2021) 4435–4446

Fig. 9. Influence of sensors numbers on 1D CNNs’ performance on estimating

damage severity.

lighter levels of damage are shown in Fig. 10(a), while the more severe
damage levels are shown in Fig. 10(b).
The results indicate that the estimations, with their error bands, for
the more severe damage levels are more accurate than the lighter levels
of damage where the frequency sensitivity is the highest. What is very
interesting is that there is an intricate nonlinear coupling between the
damage level and the accuracy of stiffness estimation, whereby the more
sensors do not always yield the best results. It can be noticed from
Fig. 10, that the case of four sensors can evaluate all the eight levels of
damage quite robustly. On the other hand, the case of two and three
sensors can evaluate better medium and severe levels of damage (i.e.
starting from 1.2 × 1010 N.mm/rad to 9.5 × 109 and N.mm/rad,
respectively). Although the one sensor case shows the ability to evaluate
some severe damage levels, it lost performance for the most severe one
(i.e. 1.2 × 109 N.mm/rad). The results highlight that the 1D CNNs can
detect the existence of the simulated structural damage in most cases of
sensor numbers, refuting at points engineering intuition (e.g. see one
against three sensors for 6.6 × 109 N.mm/rad)

4.3.2. Influence of exciting frequency

The influence of exciting frequency on 1D CNNs’ performance was
investigated by exciting the cantilever beam with three different mon­
oharmonic loads: 5 Hz, 19 Hz, and 35 Hz representing below, near, and
above the beam’s natural frequency. To have a better understanding of
these excitations and investigating also frame size influence, different
frame sizes were considered, namely, 100, 200, 300, 400, and 500. To
ensure the number of frames is not affecting the investigation in this
section, same number of frames were produced for each case by
considering different signal durations. To illustrate, 100 s signals were
divided into frames with a size of 100, standardising the selected
number of frames of each signal to 400 frames. In the same manner,
signals of 200 s, 300 s, 400 s and 500 s long were selected for the 200,
300, 400 and 500 frame sizes, respectively, and this analysis was
extended to all the levels of simulated damage.
Another reason for selecting these frame sizes is considering the
selected frame sizes in the literature. For example, Zhang et al. [21] used
20,000 as frame size, which is equivalent to signal duration of 2 s, while
Abdeljaber et al. [22] used a frame size of 128 which, for their sampling
rate, represents 0.125 s long duration. In this study, the frame sizes were
selected so that they represent signal durations in between 1.25 s and Fig. 10. Estimated spring stiffness with error band for different testing cases
0.25 s. It worth mentioning that standardisation was applied in this and different monoharmonic excitations.
section to scale each frame of the input data.
The 1D CNNs algorithm was trained and tested for each of the fifteen

4441
M. Almutairi et al. Structures 34 (2021) 4435–4446

scenarios of excitations (i.e. three excitation types with five signal du­
rations). The case of 19 Hz excitation and 300 s signal duration has the
highest MSE value (0.00026) among the fifteen scenarios. The 19 Hz
excitation with 400 s long signal has the lowest MSE (0.000007). The
learning curves for both cases are presented in Figs. 11 and 12.The re­
sults in Fig. 13 show that the 19 Hz excitation outperforms the other two
excitations for the 200 s and 400 s signal duration, where the lowest
MAPE reaches 0.8%. On the other hand, the 35 Hz excitation produced
the lowest MAPE for the 100 s, 300 s, and 500 s signal durations. The
results demonstrate that exciting the beam at near its natural frequency,
where response is more lively, does not necessarily improve the 1D
CNNs’ performance. Another interesting observation is that the 19 Hz
case is very sensitive to the change in frame size as its MAPE varies in a
range of 0.8%–4.4%. On the other hand, the excitation loads below (i.e.
5 Hz) and above (i.e. 35 Hz) the beam’s natural frequency are less
sensitive to the frame size selection as their MAPE have a range of 2.9%–
2.1% and 2.4%–1.5%, respectively.
Like in the previous section, the distributions of error throughout the
testing cases were investigated as shown in Fig. 14. Only error bands of Fig. 12. Learning curve of 19 Hz excitation case for 400 s ling signal.
the 100 s and 400 s cases are presented in Fig. 14 because results of the
200 s and 400 s cases are relatively close to each other in terms of MAPE,
and the 500 s and 300 s cases have wider error bands which could
compromise the interpretation of the figure. Fig. 14(a) shows the lighter
levels, while Fig. 14(b) shows the more severe levels of simulated
damage.
Fig. 14 shows that again the higher errors occur proportionally at the
lower levels of the simulated damage. Interestingly, the 100 s and 5 Hz
case shows less accurate estimation for the second lowest level of
damage rather than for the lowest of simulated damages (i.e. 9.5 × 1010
vs 6.6 × 1010 N.mm/rad), while the comparison with the 400 s signal
case, throughout the frequency range, is quite complex with both signal
durations being able to outperform each other in different damage-
frequency scenarios. For support stiffness of 3.4 × 1010 N.mm/rad and
below the estimations of 100 s and 400 s seem to converge with the
actual simulated damage levels. The error bands show that the 100 s
cases are more accurate than the 400 s cases almost throughout the
comparison spectrum.

4.3.3. Influence of excitation type

The influence of excitation type on the performance of 1D CNNs was
analysed by exciting the cantilever beam with wideband and mono­
harmonic excitations. For the wideband excitation types, white noise Fig. 13. Influence of three monoharmonic excitations for five different frame
and chirp signals with a nominal frequency range of 0–35 Hz were sizes on 1D CNNs’ performance on estimating damage severity.
considered. Furthermore, both of the input scaling approaches (i.e.
normalisation and standardisation approaches) were implemented and compared for each excitation type. To study the influence of number of
frames, an adopted signal duration as in the previous section (i.e. 500 s)
but with a different optimum frame size (i.e. 100 instead of 500) was
considered.
Ten scenarios (i.e. five excitation types with two scaling approaches)
were considered. The normalised case of 35 Hz excitation shows the best
performance among the ten scenarios with 0.000007 MSE. On the other
hand, the normalised chirp excitation case has 0.0007 MSE, which is the
highest MSE. Figs. 15 and 16 show that both cases have a good learning
curve.
The results in Fig. 17 show that the standardisation approach can
improve wideband excitation cases, particularly white noise excitation.
In monoharmonic excitations, normalisation reduces the MAPE in a
range of 0.8% to 0.2% compared to standardisation. The results
demonstrate the importance of selecting scaling approach for the input
data.
Different observations can be noticed from Fig. 17 and Fig. 13 by
comparing the same excitation frequency in three cases: (1) different
frame sizes but same signal duration, (2) same frame size but different
signal durations, and (3) different frame sizes and different signal
durations.
Fig. 11. Learning curve of 19 Hz excitation case for 300 s ling signal. In the first case, using the same signal duration with different frame

4442
M. Almutairi et al. Structures 34 (2021) 4435–4446

Fig. 15. Learning curve of standardised response of chirp excitation case.

Fig. 16. Learning curve of normalised response of 35 Hz excitation case.

sizes can highly affect the 1D CNN’s performance. To demonstrate this,

by reducing only the frame size from 500 to 100, the MAPE for 5 Hz, 19
Hz, and 35 Hz at 500 s durations were reduced from 2.5%, 4.2%, and
2.4% (Fig. 13) to 1.8%, 2.1% and 1.1%, respectively (Fig. 17).
In the second case, using the same frame size with different signal
durations can produce the same trend in the 1D CNNs’ performance. For
example, the 35 Hz excitation outperforms other monoharmonic exci­
tations when using the same frame size (i.e. 100) for both cases of 100 s
and 500 s signals. Also, the results show, as probably expected, that
increasing signal duration with the same frame size improves the algo­
rithm performance, yet not probably as substantially as it would have
been expected (see e.g. here considering five times more data). To
illustrate, by increasing only the signal durations from 100 s to 500 s, the
MAPE for 5 Hz, 19 Hz, and 35 Hz were reduced from 2.9%, 2.4%, and
1.5% (Fig. 13) to 1.8%, 2.1% and 1.1% (Fig. 17), respectively.
In the third case, increasing the signal duration with different frame
size does not ensure any improvement. For example, in the case of 19 Hz
excitation with 400 s duration and 400 frame size the MAPE is 0.8%
(Fig. 13), while the same excitation with longer duration (500 s) and
smaller frame size (1 0 0) has a higher MAPE (2.1%) (Fig. 17). In
contrast, the 5 Hz and 35 Hz show an improvement on their MAPE by
Fig. 14. Estimated spring stiffness with error band for different testing cases
increasing the signal duration and reducing frame size: from 2.3% to
and different monoharmonic excitations.
1.8% and from 1.8% to 1.1%, respectively. The results highlight the
importance of frame size as 1D CNNs’ hyperparameter.

4443
M. Almutairi et al. Structures 34 (2021) 4435–4446

Fig. 17. Influence of different excitation types with different input scaling
techniques on 1D CNNs’ performance on estimating damage severity.

The distributions of error were recovered as shown in Fig. 18. The

error bands in Fig. 18 are relatively large. For this reason, only error
bands for four levels of simulated damage are shown.
It can be noticed within Fig. 18 that once more the light levels of
simulated damage show higher estimation errors. The highest error is
observed at the estimation of the normalised white noise case, at the
lowest level of simulated damage (i.e. 9.5 × 1010 N.mm/rad), in which
the estimation is 8.2 × 1010 N.mm/rad. In the other levels, white noise
and chirp excitations show, in general, similar error distributions.
The error bands in wideband excitation are larger than their coun­
terparts in the monoharmonic excitations something not really in-sync
with classical modal analysis perceptions. The results of Fig. 18,
coupled with Fig. 14, show that selecting the same frame size (i.e. 100)
for different signal durations (i.e. 100 s and 500 s) can produce similar
error bands for a monoharmonic excitations.

5. Conclusions

This study proposes a 1D CNN-based approach for evaluating

structural damage severity, at a single predefined location, using only
distributed raw acceleration data. The performance of the proposed
framework was evaluated for a simple numerical case study under
different scenarios of excitation types. different data pre-processing
techniques and for different levels of damage severity.
A 2D FE model of a cantilever beam with rotationally non-rigid/
elastic support was adopted. Different levels of damage severity were
simulated to be proportional to reduction of the support rotational
spring stiffness. The acceleration responses of several set measurement
points, for each damage level, were used to form the input data matrix.
1D CNN algorithms were trained to estimate the reduced values of the
rotational spring stiffness and were further converted to levels of dam­
age severity.
The results of this study show high accuracy in evaluating the
simulated structural damage, reaching as high as 99.2%. The results
include the realisation that the monoharmonic excitation cases outper­
form white noise and chirp excitation cases, typically used and pro­
moted within experimental dynamics works. Furthermore, the results
Fig. 18. Estimated spring stiffness with error band for different testing cases
suggest that the standardisation approach for scaling the input data and different excitations.
improves wideband excitation cases, particularly white noise excitation,
while the normalisation approach improves the monoharmonic

4444
M. Almutairi et al. Structures 34 (2021) 4435–4446

excitation cases. Machine Learning and Deep Learning applications. Mech Syst Signal Process 2021;
147:107077. https://doi.org/10.1016/j.ymssp.2020.107077.
This study also establishes different sensitivity to frame size selec­
[4] Erazo K, Sen D, Nagarajaiah S, Sun L. Vibration-based structural health monitoring
tion, depending on the different frequencies of the envisaged mono­ under changing environmental conditions using Kalman filtering. Mech Syst Signal
harmonic excitation load. The most sensitive case is the case with a Process 2019;117:1–15. https://doi.org/10.1016/j.ymssp.2018.07.041.
frequency that is closer to the beam’s first natural frequency, or in more [5] Magalhães F, Cunha A, Caetano E. Vibration based structural health monitoring of
an arch bridge: from automated OMA to damage detection. Mech Syst Signal
general closer to a resonance condition. Process 2012;28:212–28. https://doi.org/10.1016/j.ymssp.2011.06.011.
The distribution of error through the simulated damage scenarios [6] Wang H, Jing X. Vibration signal–based fault diagnosis in complex structures: a
shows, in general, that the higher levels of damage severity, within the beam-like structure approach. Struct Heal Monit 2018;17(3):472–93. https://doi.
org/10.1177/1475921717704383.
considered range, correspond to more accurate damage evaluation. [7] Entezami A, Shariatmadar H. Damage localization under ambient excitations and
However, the evaluation accuracy shows lack of perfect monotonicity non-stationary vibration signals by a new hybrid algorithm for feature extraction
over the damage levels. and multivariate distance correlation methods. Struct Heal Monit 2018.
1475921718754372.
The results also show that, on average, the more severe levels of [8] Kostić B, Gül M. Vibration-based damage detection of bridges under varying
damage, for the simple structure in-hand, can be evaluated with a temperature effects using time-series analysis and artificial neural networks.
limited number of sensors, while the lighter ones require more sensors to J Bridg Eng 2017;22(10):04017065. https://doi.org/10.1061/(ASCE)BE.1943-
5592.0001085.
be evaluated with sufficient accuracy. The effect of sensor numbers is [9] Jin S-S, Jung H-J. Vibration-based damage detection using online learning
more complex than it sounds, and to this one should also take into ac­ algorithm for output-only structural health monitoring. Struct Heal Monit 2017.
count that no sensor placement exquisite methodology was considered. 1475921717717310.
[10] Giagopoulos D, Arailopoulos A, Dertimanis V, Papadimitriou C, Chatzi E,
Since the proposed framework estimates the level of damage in the
Grompanopoulos K. Structural health monitoring and fatigue damage estimation
form of reduction of an FE model parameter within a continuous range using vibration measurements and finite element model updating. Struct Heal
of damage levels, rather than predefined discrete levels, this framework Monit 2019;18(4):1189–206. https://doi.org/10.1177/1475921718790188.
can also be used as a possible precursor for a CNN-based FE model [11] Toh G, Park J. Review of vibration-based structural health monitoring using deep
learning. Appl Sci 2020;10(5):1680. https://doi.org/10.3390/app10051680.
updating technique. [12] Vamvoudakis-Stefanou KJ, Sakellariou JS, Fassois SD. Vibration-based damage
All in all, this work uncovers that the performance of CNNs as tools detection for a population of nominally identical structures: unsupervised Multiple
within dynamic testing of structures is promising but not straight- Model (MM) statistical time series type methods. Mech Syst Signal Process 2018;
111:149–71.
forward, with CNNs’ predictive results sometimes challenging expecta­ [13] Alamdari MM. Vibration-Based Structural Health; 2016.
tions and classical dynamic testing intuition. The observations of this [14] Farrar CR, Doebling SW, Nix DA. Vibration–based structural damage identification.
study were limited to evaluating simulated damage at a single point in a Philos Trans R Soc London A Math Phys Eng Sci 2001;359(1778):131–49.
[15] Comanducci G, Magalhães F, Ubertini F, Cunha Á. On vibration-based damage
simple demonstrator FE model (e.g. linear constant damping, and linear detection by multivariate statistical techniques: application to a long-span arch
analysis despite the damage) preserving simplicity as means for better bridge. Struct Heal Monit 2016;15(5):505–24. https://doi.org/10.1177/
elaborating on outputs. In future works, the limitation is to be waived by 1475921716650630.
[16] Deraemaeker A, Reynders E, De Roeck G, Kullaa J. Vibration-based structural
studying the efficiency of 1D CNNs on identifying, evaluating, and health monitoring using output-only measurements under changing environment.
localising several continuously distributed levels and types of structural Mech Syst Signal Process 2008;22(1):34–56. https://doi.org/10.1016/j.
damage in more complex structures (e.g. possessing damping, non­ ymssp.2007.07.004.
[17] Worden K, Farrar CR, Haywood J, Todd M. A review of nonlinear dynamics
linearities, and closely spaced natural frequencies waiving the ability for
applications to structural health monitoring. Struct Control Heal Monit Off J Int
a SDOF type study as in here). Studying the influence of more load types, Assoc Struct Control Monit Eur Assoc Control Struct 2008;15(4):540–67.
particularly those of more realistic nature (see e.g. wind through a quasi- [18] Catbas FN, Gul M, Burkett JL. Conceptual damage-sensitive features for structural
steady approximation within the study of Duan et al. [51]) alongside health monitoring: laboratory and field demonstrations. Mech Syst Signal Process
2008;22(7):1650–69.
uncertainty contributions would further expand considerably the results [19] Mehrjoo M, Khaji N, Moharrami H, Bahreininejad A. Damage detection of truss
delivered herein. Finally, physical testing on structures will be practiced bridge joints using Artificial Neural Networks. Expert Syst Appl 2008;35(3):
to evidence whether similar complex trends identified here pertain. For 1122–31.
[20] Chang CC, Chang TYP, Xu YG, Wang ML. Structural damage detection using an
such tests, additional parameters like the temperature, see e.g. [52], iterative neural network. J Intell Mater Syst Struct 2000;11(1):32–42.
other environmental factors, uncertainty and the combination of them [21] Zhang Y, Miyamori Y, Mikami S, Saito T. Vibration-based structural state
could have a major impact on the performance achieved. identification by a 1-dimensional convolutional neural network. Comput Civ
Infrastruct Eng 2019;34(9):822–39.
[22] Abdeljaber O, Avci O, Kiranyaz S, Gabbouj M, Inman DJ. Real-time vibration-based
structural damage detection using one-dimensional convolutional neural networks.
Declaration of Competing Interest J Sound Vib 2017;388:154–70.
[23] Avci O, Abdeljaber O, Kiranyaz S, Hussein M, Inman DJ. Wireless and real-time
The authors declare that they have no known competing financial structural damage detection: a novel decentralized method for wireless sensor
networks. J Sound Vib 2018;424:158–72.
interests or personal relationships that could have appeared to influence
[24] Teng Z, Teng S, Zhang J, Chen G, Cui F. Structural damage detection based on real-
the work reported in this paper. time vibration signal and convolutional neural network. Appl Sci 2020;10(14):
4720. https://doi.org/10.3390/app10144720.
[25] Khodabandehlou H, Pekcan G, Fadali MS. Vibration-based structural condition
Acknowledgments assessment using convolution neural networks. Struct Control Heal Monit 2019;26:
e2308.
M.A. was sponsored with a scholarship by the Kuwait Institute for [26] Almutairi M, Avci O, Nikitas N. Efficiency Of 1D CNNs on Finite Element Model
Parameter Estimation Using Synthetic Dynamic Responses. EURODYN 2020. XI Int.
Scientific Research for conducting doctoral studies in the University of Conf. Struct. Dyn., Athens, Greece; 2020.
Leeds, UK. M.B. was supported by the Polish National Agency for Aca­ [27] Eickenberg M, Gramfort A, Varoquaux G, Thirion B. Seeing it all: convolutional
demic Exchange (NAWA) under the Polish Returns programme, grant network layers map the function of the human visual system. Neuroimage 2017;
152:184–94.
number PPN/PPO/2019/1/00036.
[28] Patterson J, Gibson A. Deep learning: A practitioner’s approach. “ O’Reilly Media,
Inc.”; 2017.
References [29] Yu J, Zhang C, Wang S. Multichannel one-dimensional convolutional neural
network-based feature learning for fault diagnosis of industrial processes. Neural
Comput Appl 2021;33(8):3085–104.
[1] Das S, Saha P, Patro SK. Vibration-based damage detection techniques used for
[30] Chen X, Kopsaftopoulos F, Wu Q, Ren H, Chang F-K. A self-adaptive 1D
health monitoring of structures: a review. J Civ Struct Heal Monit 2016;6(3):
convolutional neural network for flight-state identification. Sensors 2019;19(2):
477–507. https://doi.org/10.1007/s13349-016-0168-5.
275. https://doi.org/10.3390/s19020275.
[2] Fan W, Qiao P. Vibration-based damage identification methods: a review and
[31] Tang X, Li W, Li X, Ma W, Dang X. Motor imagery EEG recognition based on
comparative study. Struct Heal Monit 2011;10:83–111.
conditional optimization empirical mode decomposition and multi-scale
[3] Avci O, Abdeljaber O, Kiranyaz S, Hussein M, Gabbouj M, Inman DJ. A review of
vibration-based damage detection in civil structures: from traditional methods to

4445
M. Almutairi et al. Structures 34 (2021) 4435–4446

convolutional neural network. Expert Syst Appl 2020;149:113285. https://doi. [43] Avci O, Abdeljaber O, Kiranyaz S, Inman D. Structural damage detection in real
org/10.1016/j.eswa.2020.113285. time: implementation of 1D convolutional neural networks for SHM applications.
[32] Wang J, Wang D, Wang X. Fault diagnosis of industrial robots based on multi- In: Struct. Heal. Monit. Damage Detect. Springer; 2017. p. 49–54.
sensor information fusion and 1D convolutional neural network. In: 39th Chinese [44] Hao L, Lewin P. Partial discharge source discrimination using a support vector
Control Conf. IEEE; 2020. p. 3087–91. machine. IEEE Trans Dielectr Electr Insul 2010;17(1):189–97.
[33] Jiao J, Zhao M, Lin J, Zhao J. A multivariate encoder information based [45] Wang Z, Cha Y. Automated damage-sensitive feature extraction using unsupervised
convolutional neural network for intelligent fault diagnosis of planetary gearboxes. convolutional neural networks. Sensors Smart Struct. Technol. Civil, Mech. Aerosp.
Knowledge-Based Syst 2018;160:237–50. Syst. International Society for Optics and Photonics; 2018. p. 105981J.
[34] Jing L, Zhao M, Li P, Xu X. A convolutional neural network based feature learning [46] Abdeljaber O, Avci O, Kiranyaz MS, Boashash B, Sodano H, Inman DJ. 1-D CNNs
and fault diagnosis method for the condition monitoring of gearbox. Measurement for structural damage detection: verification on a structural health monitoring
2017;111:1–10. benchmark data. Neurocomputing 2018;275:1308–17.
[35] Chollet F. keras; 2015. [47] Wang TM. Natural frequencies of continuous Timoshenko beams. J Sound Vib
[36] Goodfellow I, Bengio Y, Courville A, Bengio Y. Deep Learning. MIT press 1970;13(4):409–14.
Cambridge; 2016. [48] Blodgett OW. Design of welded structures. Clevel James F Lincoln Arc Weld Found
[37] Kingma DP, Ba J. Adam: A method for stochastic optimization. ArXiv Prepr 1966; 1966.
ArXiv14126980; 2014. [49] Google. Google Colab n.d.
[38] Nguyen D, Widrow B. Improving the learning speed of 2-layer neural networks by [50] Chao Q, Tao J, Wei X, Wang Y, Meng L, Liu C. Cavitation intensity recognition for
choosing initial values of the adaptive weights. In: IJCNN Int. Jt. Conf. Neural high-speed axial piston pumps using 1-D convolutional neural networks with
Networks. IEEE; 1990. p. 21–6. multi-channel inputs of vibration signals. Alexandria Eng J 2020;59(6):4463–73.
[39] Levin RI, Lieven NAJ. Dynamic finite element model updating using neural [51] Duan Y, Chen Q, Zhang H, Yun CB, Wu S, Zhu Q. CNN-based damage identification
networks. J Sound Vib 1998;210(5):593–607. method of tied-arch bridge using spatial-spectral information. Smart Struct Syst
[40] Lu Y, Tu Z. A two-level neural network approach for dynamic FE model updating 2019;23:507–20. https://doi.org/10.12989/sss.2019.23.5.507.
including damping. J Sound Vib 2004;275(3-5):931–52. [52] Fallahian M, Khoshnoudian F, Meruane V. Ensemble classification method for
[41] Park Y-S, Kim S, Kim N, Lee J-J. Finite element model updating considering structural damage assessment under varying temperature. SAGE 2018;17(4):
boundary conditions using neural networks. Eng Struct 2017;150:511–9. 747–62. https://doi.org/10.1177/1475921717717311.
[42] Hasançebi O, Dumlupınar T. Linear and nonlinear model updating of reinforced
concrete T-beam bridges using artificial neural networks. Comput Struct 2013;119:
1–11.

4446