Structural Damage Detection Using Deep Learning and Fe Model Updating Techniques
Structural Damage Detection Using Deep Learning and Fe Model Updating Techniques
com/scientificreports
The structural condition can be estimated by various methods. Damage detection, as one of those
methods, deals with identifying changes in specific features within structural behavior based on
numerical models. Since the method is based on simulation for various damage conditions, there are
limitations in applicability due to inevitable discrepancies between the analytical model and the actual
structure. Finite element model updating is a technique for establishing a finite element model that
can reflect the current state of a target structure based on the measured responses. It is performed
based on optimization for various structural parameters, but the final output can converge differently
depending on the initial model and the characteristics of the algorithm. Although the updated model
may not faithfully replicate the target structure as it is, it can be considered equivalent in terms of
the relationship between the structural properties and behavioral characteristics of the target. This
allows for the analysis of changes in the mechanical relationships established for the target structure.
The change can be related to structural damage, and artificial intelligence technology can provide
an alternative solution in such complex problems where analytical approaches are challenging.
Taking practical aspects from the aforementioned methods, a novel structural damage detection
methodology is presented in this study for identifying the location and extent of the damage. Model
updating is used to establish a reference model that reflects the structural characteristics of the
target. Training data for various damage conditions based on the reference model allows the artificial
intelligence networks to identify damage to the target structure.
There are various methods for assessing the as-is condition of structures. Most popular methods that are being
utilized in the sites are on-site manual inspections and vibration-based structural health monitoring. On-site
inspections involve visual inspection and manual assessment of structural elements to detect visible damages or
anomalies1–3. While it provides a useful means of assessing the condition of structure, the accuracy and reliability
of the assessment can be subjective and prone to human errors. Additionally, on-site inspections are typically
conducted at specific intervals, which means that the assessment may not capture the real-time behavior or
deterioration of the structure continuously. Vibration-based structural health monitoring systems, on the other
hand, can continuously monitor the behavior and response of structures using sensors and data acquisition
techniques4–8. Structural health monitoring systems analyzes the collected data and performs post processing
to obtain useful information that could refer to the condition of the structure9–13. These methods offer a more
comprehensive and objective assessment of the structural condition, can provide preventive maintenance and
timely responses to maintain structural integrity and safety.
The structural response can be analyzed using various methods to evaluate their condition. One of the most
commonly employed approaches is vibration-based structural system identification, which aims to extract modal
parameters and dynamic characteristics of structures through the analysis of measured vibrations. By measuring
the dynamic characteristics, valuable insights into the structural integrity can be obtained. The advantages of this
method include its non-destructive nature, simplicity of application, and the ability to capture global responses.
There are a number of studies that attempted to estimate the change of structural conditions through changes
in natural frequencies using long-term measurement data14–18. While some studies have shown that the natural
frequency can provide valuable information related to the stiffness of structures, it was not sufficient to estimate
1
School of Civil Engineering, College of Engineering, Chungbuk National University, Cheongju, Republic of
Korea. 2Future and Fusion Lab of Architectural, Civil and Environmental Engineering, Korea University, Seoul,
Republic of Korea. 3School of Civil, Environmental, and Architectural Engineering, Korea University, Seoul,
Republic of Korea. *email: [email protected]
the condition of structures directly. For most studies, the natural frequency was rather utilized as the input
parameter (feature) of a system that can be used to evaluate the structural integrity indirectly.
Another method that are being widely studied for vibration-based structural health monitoring is damage
detection. Damage detection methods not only finds the location of the damages, but also estimates the mag-
nitude of the damage. Most of the methods utilize the differences in structural responses under damaged and
intact conditions to detect the damage. Various responses can be used to identify the behavioral characteristics
of the structure including acceleration19–21, strain22–24, displacement25–27. Once the response of the structure is
measured, advanced techniques including signal processing a lgorithms28–30, and machine learning31–33, can be
utilized to detect the damage. While dynamic characteristics such as natural frequency and mode shape are gen-
erally being used as input parameters for these post processing procedure, many researchers have used natural
frequencies as a basis for identification of structural i ntegrity34–36. In conventional methods, a specific damage
index was proposed so that the characteristics of the change in natural frequency for the damaged condition can
be emphasized37–39. However, these methods were difficult to derive an index that can sufficiently represent the
magnitude and location of damage from the natural frequency through a kind of numerical calculation process.
Moreover, since the method is performed based on simulation, there were limitations to its application due to
discrepancies in the behavior of actual structures.
Finite element (FE) model updating is an additional way to assess the current state of s tructures40–44. This
method utilizes FE models to refine and update the numerical model, aiming to achieve a more accurate rep-
resentation of the systematic behavior of the target structure. Parameters for the model can include material
properties, boundary conditions, or geometric features. This process typically employs an optimization algorithm
that minimizes the discrepancy between the simulated responses of the model and the measured responses
from sensors or experimental tests. By iteratively adjusting the model, a closer alignment with the behavior
of the target structure is achieved. The natural frequency of the structure serves as a prominent input feature
within this method as well. Since the natural frequency is one of the most effective factors that can represent
the relationship between the stiffness and mass of a structure, it can allow the FE model to converge towards
the mechanical characteristics of the target structure. However, the updated model finally converged depends
on the initial model and the characteristics of the optimization algorithm, as well as the assumptions within the
updating process. In addition, there is a limitation for constructing a model that can replicate all the behavioral
characteristics of the structure, including static behavior, solely based on dynamic characteristics. Nevertheless,
it can still present a possibility of being used to identify changes in such established relationships by optimizing
the relationship between the considered structural properties and behavioral characteristics.
In an effort to address the challenges in a conventional method of structural damage detection, researchers
have investigated the application of artificial intelligence (AI) technology. AI technology can provide alternative
data-driven solutions for problems that involve hard-to-detect patterns or those that cannot be deterministically
defined45. The efficiency of the solutions is closely reliant on the quality of the training data. As technology for
generating and processing massive data has been developed, its usability has been significantly enhanced and
it is being applied to various fields. In terms of the architectural aspect of the network, it has been utilized in
various forms for structural damage detection. Deep neural network (DNN)46–48, as the fundamental form of a
neural network, explores the relationship between input and output through a feed-forward process. It is useful
when input and output can be relatively simply defined. Recurrent neural network (RNN)49–51, with its recurrent
structure capable of incorporating past information into the current state, is suitable for processing sequential
data such as time series data. Data such as the dynamic response of structures can be used as the input for RNN.
Convolutional neural network (CNN)52–54 is primarily specialized in processing spatial data such as images.
It is used for extracting spatial features from data, and is not only useful for image analysis but also facilitates
the extension of data features for the analysis. In terms of training methodology, there are supervised methods
where outputs are labeled for corresponding inputs to extract features for the relationship, and unsupervised
methods55–57 that discover patterns from data without labels and automatically train networks. Supervised meth-
ods, being based on clear labels, can provide excellent performance for accurate predictions. However, in cases
where it is challenging to specify clear labels, such as in the conditions of real structures, or when the goal is to
discover new features, unsupervised methods may be more appropriate.
This study adopts the practical aspects of the previously described techniques for efficient application to dam-
age detection in actual structures. An FE model updating technique was applied to construct an FE model capable
of appropriately representing the dynamic characteristics of the target structure. Subsequently, structural analysis
simulations were conducted for various arbitrary damage cases based on the updated FE model. Artificial intel-
ligence technology can be trained on various damage conditions that can occur in the target structure through
simulated data. Given that the reference FE model was designed to reflect the target structure faithfully, the
trained network can effectively identify the location and degree of damage in the target structure when it occurs.
The following chapter provides a description of the proposed methodology. The feasibility of the proposed
method has been verified through experimental tests, and the relevant comprehensive discussion is given in sec-
tion “Validation test”, including structural analysis simulations and the training of artificial intelligence network.
Section “Conclusion” covers a brief conclusion and future work for the study.
Methodology
A schematic description of the proposed method in this study is shown in Fig. 1. The proposed method can be
divided into three steps. The first step is to construct a reference FE model that can adequately represent the
current state of the target structure. This can be achieved through FE model updating using dynamic properties
obtained from system identification for the target structure. In the second step, numerous arbitrarily damaged
FE models are generated based on the reference FE model derived in the previous step. These models are used to
train the damage detection network for various damage conditions. System identification in the final stage, when
defects occur in the structure, captures the dynamic characteristics of the damaged structure. Subsequently, this
information is applied to the damage detection network to assess the current condition of the structure. Further
details of the methodology are described through following sections.
System identification
System identification refers to the process of determining a mathematical model or characteristics of a system
based on observed input and output data, as shown in Fig. 2. A system determines the relationship between input
and output. When it is difficult to know the system through analytical approaches, the relationships between
input and output data can be used to identify characteristics of the system and establish a corresponding math-
ematical system model. However, in case where the input to the system is not clear, such as in typical structures
with complexly acting loads, output-only system identification can be conducted by assuming the input as white
noise and using only the response of the structure.
System identification can be performed in several ways. In structural engineering, it is generally performed
through frequency domain analysis of the ambient vibration response of the structure. The dynamic response
data of the structure is transformed using techniques such as Fourier transform to derive frequency components,
and then applying the peak-picking method, dominant peaks or resonant frequencies can be identified. For the
processing of typical digital signals, the discrete Fourier transform as expressed in Eq. (1) is applied, allowing
for the extraction of frequency components from the dynamic response data of structures.
N−1
X(k) = x(n)e−i2πkn/N , (0 ≤ k < N) (1)
n=0
Here, x is the input data to be transformed to the frequency domain, N is the number of samples of the input,
and k denotes the frequency index and has values from 0 to N − 1.
Frequency Response Function (FRF) is one of the fundamental concepts in the field of structural dynamics,
which represents the relationship between the input excitation and the corresponding output response in the
frequency domain. It provides significant information for the dynamic behavior of a system and is widely used
in various engineering applications. Defining X(ω) and Y (ω) as the Fourier transforms of the input and output
signals, respectively, the FRF of the system can be expressed as Eq. (2).
FE model updating
The FE model is a numerical representation of a physical structure used in engineering analysis and simulations.
It divides the structure into smaller elements, enabling the complex structural behavior to be analyzed through
smaller and more easily comprehensible elements. However, the FE model cannot reflect the characteristics of
the target structure as it is, since there are inevitable differences between the actual structure and the theoretical
FE model in terms of geometrical properties, material properties, and boundary conditions. In this regard, an
FE model updating that minimizes the error between the target structure and the FE model is required.
FE model updating is a process used to improve the accuracy of numerical FE models by adjusting their
parameters based on experimental data, as shown in Fig. 3. In FE model updating, the initial FE model is
compared with measured data obtained from sensors installed on the structure. By iteratively adjusting the
model parameters, such as material properties, boundary conditions, or geometric properties, the updated FE
model can better represent the dynamic behavior of the real structure. Model updating techniques generally
involve optimization algorithms that minimize the objective function, which quantifies the differences between
the FE model and the experimental measurements. The objective function is mainly set using the dynamic
characteristics of the structure, such as the natural frequency. Although the FE model updated using only the
natural frequencies cannot represent all the behaviors of the target structure as it is, it can be used to identify
the changes in dynamic characteristics due to changes in the structural system. Once an updated FE model that
can faithfully represent the dynamic characteristics of the target structure in its current state is established, the
dynamic characteristics of subsequent structural defects can also be expressed based on the reference FE model.
This study adopts the model updating technique to train a damage detection network by generating simulation
data for various potential defects based on the reference FE model that replicates the dynamic characteristics of
the target structural system.
In the above, z represents the weighted sum of inputs (x) and their corresponding weights (w), b is the bias
term, σ denotes the activation function, and h represents the output of the node.
By stacking multiple layers, DNNs can learn hierarchical representations of data, allowing for the extraction
of complex features. The information flow through the network can be expressed as Eq. (4)
h1 = σ (w1 ∗ x + b1 )
h2 = σ (w2 ∗ h1 + b2 )
.. (4)
.
hn = σ (wn ∗ hn−1 + bn )
Here, hi represents the output of layer i , wi denotes the weight matrix connecting layer i − 1 to layer i , and bi
is the bias vector of layer i . Finally, the output layer produces the desired output as Eq. (5) based on the processed
information from the previous layers.
y = σ (wn+1 ∗ hn + bn+1 ) (5)
Training a DNN involves optimizing its parameters (weights and biases) to minimize the discrepancy between
predicted outputs and ground truth labels. This is achieved through a process called backpropagation, which
computes the gradient of a defined loss function with respect to the parameters of the network. By applying an
optimization algorithm, such as stochastic gradient descent, the network iteratively updates its parameters to
improve its prediction accuracy.
Validation test
The feasibility of the proposed method was evaluated based on the experimental test. Vibrational tests were
conducted to extract the dynamic characteristics of the target structure, which were then used to establish a
reference FE model that adequately represents the dynamic characteristics of the target structure. Based on the
reference FE model, various FE models for arbitrary damage conditions were generated and used for training
the deep learning network. The performance of the trained network was evaluated by applying the dynamic
characteristics extracted through the test on intentionally damaged structures to the network.
Experimental test
Experimental setup
The experimental test was conducted on a three-story frame structure as shown in Fig. 5. The width and depth of
the structure are both 500 mm, and the total height is 1950 mm. Each floor consists of a steel plate with a thick-
ness of 50 mm, which is supported by wall plates with a height of 600 mm and a thickness of 5 mm. Structural
damaged condition was considered by reducing the thickness on one side of the first-floor wall plate by 20%.
The dynamic behavior of the structure was induced by applying ground motion using a shaking table, so
that the dynamic characteristics of the structure can be extracted. When ground motion has characteristics in
a specific frequency range, it can disturb identifying the dynamic characteristics of the structure. Therefore, the
applied ground motion was designed to have characteristics of white noise as shown in Fig. 6. The excitation
lasted for 60 s, and the peak ground acceleration was applied as about 0.1 g.
The structural responses were measured by applying Yoon’s method50 of vision-based displacement measure-
ment, using a camera installed in front of the structure. The lateral displacement of the steel plate of each floor
was extracted using a camera device with a frame rate of 30 fps.
0.15
0.1
Acceleration [g]
0.05
-0.05
-0.1
-0.15
0 10 20 30 40 50 60
Time [sec]
Experiment results
Figure 7 shows the displacements of each floor of the structure in intact condition for the applied ground motion.
The structure fluctuates for about 60 s where the ground motion is applied, and thereafter, it shows a gradual
dissipation of the excitation due to the inherent damping of the structure itself. The maximum displacements
in each floor were 11.56 mm, 17.70 mm, and 20.23 mm, respectively. Since the response is caused by the input
applied to the structure and the mechanical characteristics of the structure, the structural properties can be
identified by analyzing the ambient vibration responses.
Among the various methods for analyzing time series data in the frequency domain, this study used the FRF
of the responses for identifying the frequency domain characteristics of the structure. Figure 8 shows the FRFs
for the responses of each floor. The peak points in FRFs can represent the natural frequency of the structure, and
three points can be identified in the figure. The first peak point, that means the first natural frequency, appeared
most clearly. Although the second point, which represents the second vibration mode of the structure, was not
well-recognized in the FRF of the second-floor response, it could be identified from the other responses. There
is a relatively large amount of noise included near the third natural frequency, but it was possible to identify the
peak points from all the responses.
Figure 9 shows the responses of each floor when ground motion is applied to the structure in which the dam-
age is represented by reducing the wall thickness of one side of the first floor. The maximum displacements of each
floor appeared 11.18 mm, 16.31 mm, and 18.19 mm, respectively. Compared to the responses of the structure in
an intact condition as in Fig. 7, it is considered that the characteristics of the damaged condition are difficult to
be identified through the time series responses. Rather, the structural response in the intact condition exhibited
a larger magnitude compared to the damaged condition.
20
Displacement [mm]
10
-10
-20
0 10 20 30 40 50 60 70 80 90 100
Time [sec]
(a) Displacement at 1st floor
20
Displacement [mm]
10
-10
-20
0 10 20 30 40 50 60 70 80 90 100
Time [sec]
(b) Displacement at 2nd floor
20
Displacement [mm]
10
-10
-20
0 10 20 30 40 50 60 70 80 90 100
Time [sec]
(3) Displacement at 3rd floor
Figure 10 shows the FRFs of the responses to analyze the frequency domain characteristics of responses of
the structure in the damaged condition. Similar to the results of FRFs for responses of the structure in the intact
condition, it was observed that as the frequency range increased, the FRFs contained multiple noise components.
However, it appeared that each peak point could be identified from all responses. The natural frequencies of the
structure obtained from FRFs for responses of the structure in intact and damaged conditions are summarized
in Table 1. The natural frequencies of the structure in the damaged condition appeared with 7.89%, 5.29%, and
1.75% reduced results, respectively, compared to those in the intact condition.
3
10
Amplitude [-]
101
0
10
-1
10
10-2
0 5 10 15
Frequency [Hz]
(a) FRF curve for the 1st floor response
103
2
10 2.547 7.313 10.399
Amplitude [-]
101
100
10-1
10-2
0 5 10 15
Frequency [Hz]
(b) FRF curve for the 2nd floor response
103
1
10
100
10-1
10-2
0 5 10 15
Frequency [Hz]
(c) FRF curve for the 3rd floor response
20
Displacement [mm]
10
-10
-20
0 10 20 30 40 50 60 70 80 90 100
Time [sec]
(a) Displacement at 1st floor
20
Displacement [mm]
10
-10
-20
0 10 20 30 40 50 60 70 80 90 100
Time [sec]
(b) Displacement at 2nd floor
20
Displacement [mm]
10
-10
-20
0 10 20 30 40 50 60 70 80 90 100
Time [sec]
(3) Displacement at 3rd floor
different from those in the actual structure. Due to such difference, the FE model is not suitable for analyzing
the characteristics of damaged conditions for the target structure, even though it was designed for the structure
as a target. The FE model needs to be applied with model updating that optimizes it so that it can represent the
characteristics of the target structure.
Reference FE model
To determine the reference FE model that is the basis for analyzing characteristics of damaged conditions for the
target structure, a basic optimization process was applied to the initial FE model. The reference model is deter-
mined as the FE model that best represents the dynamic characteristics of the target structure through a general
model updating process. Note that the reference model is a kind of pseudo model for estimating possible damages
through changes in structural dynamic characteristics, and it does not represent the target model itself which can
represent other behaviors such as a static behavior of the target structure. The structural dynamic characteristics
are dependent on its stiffness and mass. Even for the model to represent a certain dynamic characteristic, there
3
10
2
10 2.346 6.926 10.217
Amplitude [-]
101
0
10
10-1
-2
10
0 5 10 15
Frequency [Hz]
(a) FRF curve for the 1st floor response
3
10
2
10 2.346 6.926 10.217
Amplitude [-]
1
10
100
10-1
-2
10
0 5 10 15
Frequency [Hz]
(b) FRF curve for the 2nd floor response
3
10
2
10 2.346 6.926 10.217
Amplitude [-]
1
10
100
-1
10
-2
10
0 5 10 15
Frequency [Hz]
(c) FRF curve for the 3rd floor response
Natural frequency
1st mode 2nd mode 3rd mode
Intact condition 2.547 Hz 7.313 Hz 10.399 Hz
Damaged condition 2.346 Hz 6.926 Hz 10.217 Hz
Difference (− 7.89%) (− 5.29%) (− 1.75%)
Mode shape
Table 3. Mode shapes and natural frequencies of the FE model with reference properties.
can be numerous possible models with different combinations of stiffness and mass that satisfy it. In order to
refer that the reference model is not uniquely determined and can be applied as a variety of models, this study
used several reference models for evaluating the applicability of the proposed method.
In this study, the reference models were determined by two procedures. At the first step, the approximate
properties of the reference model are determined based on the first natural frequency by regarding the elastic
modulus and mass of the entire model as single values. A total of three approximate reference models were
determined in this step: the two models were determined by considering one of the initial estimates of the elastic
modulus and the mass as constant and adjusting the other variables, and the other was determined as a model
having a compromise property between the two models. Table 4 shows the structural properties of the three
reference models determined by the first step.
The second step is the fine-tuning step for the approximately determined reference models. The variables are
expanded for each floor, and they are determined to minimize the differences in the 1st to 3rd natural frequen-
cies based on a series of optimization process. The objective function for optimization is as shown in Eq. (6).
m
1 2
({E}, {M}) = argmin (ωi − ωi ) (6)
m
i=1
Here, {E} and {M} are the elastic modulus and mass density of each floor, and ωi and ω i represent the natural
frequencies extracted through the experiment of the target structure and based on FE analysis, respectively. m is
the order of the natural frequencies to be considered, which was considered up to the 3rd order in this study. The
optimization problem at this stage for structural properties with natural frequencies is not much complicated, so
there are various methods that can be applied relatively simply. For the fine-tuning of the reference model, the
Nelder-Mead method58 was applied in this study, and the results of the optimization process for each reference
model are shown in Fig. 12.
Although the optimization starts from the initial values that are not different for each floor derived in the
previous step, the final values obtained through the process appear differently for the floors. The properties of
the reference models derived from the fine-tuning process are shown in Table 5. While the models each have
a different distribution of elastic modulus and mass density, they all represent quite identical characteristics in
the natural frequencies. Hereafter, the reference model determined in this stage would be then used as a basis
that represents an intact condition for training changes in dynamic characteristics caused by damages of the
target structure.
0.12
Reference model 1
Reference model 2
0.1 Reference model 3
0.08
Function value
0.06
0.04
0.02
0
0 20 40 60 80 100
Iteration
Network configuration
A schematic configuration of the training network used in this study is shown in Fig. 13. The input layer of the
network consists of natural frequencies from the 1st to the 3rd order, and the stiffness integrities of the wall plates
on the 1st to 3rd floors are derived from the output layer. Here, the stiffness integrity represents the ratio of the
elastic modulus, expressed as a percentage relative to its distribution in the reference model, with 100% being
the maximum value. The relationship between input and output data would be trained by iterative operations
passing through hidden layers between the input and output layers. All layers are fully connected to each other
to identify the relationships through successive operations. The required number of hidden layers and nodes to
serve the network performance depends on the characteristics of the input and output data. In this study, it has
been determined through the parametric study in the following section.
Parametric study
The training performance of the network depends on various factors. To determine the optimal network suitable
for the purpose of the study, this section analyzes the parameters for the hidden layer, which are the significant
parameters of the network used in this study. One is the number of hidden layers, which determines the depth
and complexity of the network. Increasing the number of hidden layers allows for capturing more intricate
relationships in the data, but it also increases the risk of overfitting. Another is the number of nodes within each
hidden layer. The number of nodes affects the capacity of the network to learn and represent patterns in the
data. More nodes can potentially capture more nuanced features, but an excessive number may lead to overfit-
ting or increased computational complexity. Table 6 shows the values considered in the parametric study. The
Parameter Parameter
1 2
=N =N
Node Layer
N =3 N =3
Input Out
Natural frequency Stiffness
1 1
Natural frequency2 Stiffness2
Natural frequency Stiffness
3 3
values were determined as ranges in which the parameters represented sufficient performance and their effects
tended to converge.
In the training process, an epoch represents a single iteration in which the model processes all the dataset
once, and there is a requirement to set a constraint on the maximum epoch for proper training termination. Once
the maximum epoch is reached, the training process typically stops, regardless of whether the model has fully
converged or not. The choice of the maximum epoch value depends on various factors, such as the complexity
of the problem, dataset size, and computational resources available. Setting it too low may result in underfitting,
where the model fails to capture the underlying patterns in the data. On the other hand, setting it too high can
lead to overfitting, where the model becomes too specific to the training data and performs poorly on new data.
In this study, the maximum epoch was set to 10,000 by default, which was found to be sufficient for the training
performance of the network to converge in the preliminary analysis. Instead, the patience option was applied for
an early stop to prevent excessive repetition of the training process. The option refers to waiting for a few more
epochs before determining that there is no further improvement in the training process. When the validation
accuracy for the validation data applied in the training process does not improve for several epochs, the training
would be early-stopped, and the value was set to 200 in this study.
A batch is the number of training data that are processed together in a single forward and backward pass
during each iteration. The choice of batch size is a trade-off between computational efficiency and model conver-
gence. A larger batch size can make better use of parallel computing capabilities and speed up training but may
require more memory. Conversely, a smaller batch size may provide a more accurate estimate of the gradient but
can be computationally slower. In this study, the batch size of 32 was applied, which is a relatively small value,
because the problem to be solved through network training is not that complicated.
During the training process, the weights of the model are updated step by step according to the learning
rate, which controls the size of the weight adjustment based on the computed gradient. If the learning rate is too
high, the model may overshoot the optimal solution and fail to converge. On the other hand, if the learning rate
is too low, the training process may be slow, or the model may get stuck in a suboptimal solution. It can also be
adjusted during progress using the learning rate schedule option. However, since the dataset and model com-
plexity do not exhibit significant variations in this study, a fixed learning rate of 0.001 was adopted to maintain
a stable and consistent update scheme throughout the procedure. It has been widely used as a default value in
various deep learning applications, allowing the process to be simplified by eliminating the need for additional
hyperparameter tuning and scheduling mechanisms.
When the available data is limited, exhibits imbalance, or shows significant variability that may result in biased
training, cross-validation59 can be considered to provide a generalized performance evaluation of the model. This
study used structural analysis simulation data for training, which were evenly distributed across various condi-
tions. Additionally, the training process was verified by applying a validation dataset that was distinct from the
training data, and it was confirmed that overfitting due to bias between the training and validation datasets did
not occur. Accordingly, considering that the data was applied relatively clearly to avoid biased network training,
cross-validation was not considered in this study.
The reference models have only differences in their material properties, and changes in their dynamic char-
acteristics to the decrease in stiffness appear similar. Therefore, the parametric study was conducted based only
on the reference model 1.
Figure 14 shows the performance of the network for the considered parameters. As shown in Fig. 14a, the
performance of the network tends to converge when the number of hidden layers is about 5, although it varies
to some extent depending on the number of nodes in the hidden layer. Figure 14b shows that, while there is a
tendency for the network performance to improve to some extent with an increase in the number of nodes, most
networks represent satisfactory performance, except for cases where the number of nodes is too small as 10.
According to the results of Fig. 14c, this study determined the number of hidden layers and nodes to be 5 and
2000, respectively, for structural damage detection by training the network on arbitrarily damaged FEA data.
Training results
The network determined in the previous section was applied to each reference model and trained for the natural
frequencies of FE models with the stiffness of each floor arbitrarily reduced from 50 to 100%. The range of stiff-
ness reduction can be adjusted for a comprehensive evaluation of various damage scenarios according to the
purpose. A total of 1000 damaged models were generated for each reference model, providing a diverse set of
8 8
N node =10 N layer=1
7 N node =100 7 N layer=2
3 3
2 2
1 1
1 2 3 4 5 6 7 0 500 1000 1500 2000 2500 3000
Number of layers [-] Number of nodes [-]
(a) RMSE for the number of layers (b) RMSE for the number of nodes
7
10 7.130 4.227 2.861 2.761 2.360 2.470 2.731
6
100 2.571 2.191 1.976 2.076 1.877 1.341 1.565
training and testing data. To evaluate the training of the network, 80% of the generated models were used for
training the network, while the remaining 20% were reserved for testing its capabilities.
Figure 15 shows the results for the tested data. The results demonstrate that the network has been trained
adequately, as denoted by the achieved RMSE values of less than 0.85% for all reference models. The low RMSE
values indicate a strong correlation between the predicted and actual stiffnesses, further validating that the
network can represent generalized performance over a range of stiffness degradations.
This study adopts the reference FE models to simulate the dynamic characteristics of the structure subjected
to damages, as observed in the experimental test. The natural frequencies obtained from the experimental test
are applied as inputs for the trained neural networks corresponding to each reference model, and the results are
shown in Fig. 16.
90 90 90
Predicted [%]
Predicted [%]
Predicted [%]
80 80 80
RMSE =0.58 RMSE =0.84 RMSE =0.80
70 70 70
60 60 60
50 50 50
50 60 70 80 90 100 50 60 70 80 90 100 50 60 70 80 90 100
Ground truth [%] Ground truth [%] Ground truth [%]
(a) Network for reference model 1 (b) Network for reference model 2 (c) Network for reference model 3
Figure 16. Predicted results for the target structure in the experimental test.
The results indicate that all networks yield reasonably accurate estimations of stiffness when the natural
frequencies of the structure in an intact condition are applied. Although the network for reference model 2
exhibits the largest error, estimating the stiffness of the second-floor wall plate as 99.27%, it can be regarded as
acceptable error for the purpose of the study.
During the experimental test, the damaged condition of the structure was represented by reducing the thick-
ness of one of the first-floor wall plates by 20%. By neglecting the decrease in mass resulting from the reduced
thickness of the wall plate, it can be evaluated as equivalent to 75.6% of the stiffness of the first floor compared
to the intact condition. The results for the structure in damaged condition demonstrate that it also can exhibit
adequate performance in estimating the stiffness distribution of the structure under such damaged conditions.
Although the stiffness of the third-floor wall plate was estimated relatively lower than that of the intact condition,
it still signifies that the stiffness is sufficiently secured in comparison to the first floor, in which the actual dam-
age condition was applied. Furthermore, the estimated values for the stiffness of the first-floor exhibit relatively
accurate results when compared to their ideal values, indicating promising potential for practical application in
structural damage detection.
The results also confirmed that the reference models with different material properties can be used equiva-
lently to estimate structural damage based on the dynamic characteristics of a single target structure. Although
they cannot reflect the complete behavior of the structure, they can be used for the analysis of dynamic charac-
teristics dominated by the stiffness and mass of the structure. It indicates that there are various FE models that
can represent the dynamic characteristics of a single structure, unless it is to analyze the complete behavior of
the structure.
Conclusion
This study proposed a new damage detection methods using model updating, structural analysis simulation, and
deep learning. The deep learning network was used for reliable estimation of various damage conditions, and the
FE model updating was used for improving the applicability of simulation-based training data to actual target
structure. The proposed method was validated through experimental tests on a three-story frame structure, and
the following conclusions were drawn.
I. The reference FE model for structural damage detection derived from the FE model updating can have
various structural properties as initial values. This implies the presence of multiple FE models that can
effectively represent the dynamic characteristics of the target structure, rather than indicating a failure
of the model updating process to converge to a single model.
II. The reference FE model is intended to represent changes in the dynamic characteristics resulting from
changes in the structural system, rather than replicating all behaviors of the target structure. It has been
confirmed that the generated training data based on such reference model can sufficiently represent the
changes in dynamic characteristics caused by structural damages in the target structure.
III. The proposed method showed an average error of approximately 0.29% and 1.58% for the stiffness dis-
tribution of intact and damaged conditions of the target structure, respectively, providing evidence of
the effectiveness of the proposed method.
Since this study was conducted for a relatively simple three-story frame structure and an extensive level of
damage was applied, it was possible to sufficiently identify the damage by using only some natural frequencies as
input features. In addition to natural frequency, there are various behavioral characteristics such as mode shape
and modal energy that can be used as input features. Moreover, beyond the frequency domain methods applied
in this study, other advanced techniques like time-domain methods can be used for additional input features.
To address more complex target structures or achieve a higher level of precision in damage detection, it may be
necessary to consider employing such additional input features or advanced methods.
Data availability
The datasets used and/or analysed during the current study available from the corresponding author on reason-
able request.
References
1. Phares, B. M., Rolander, D. D., Graybeal, B. A. & Washer, G. A. Studying the reliability of bridge inspection. Public Roads 64, 15–19
(2000).
2. Rolander, D., Phares, B., Graybeal, B., Moore, M. & Washer, G. Highway bridge inspection: State-of-the-practice survey. Transp.
Res. Rec. 1749, 73–81 (2001).
3. Graybeal, B. A., Phares, B. M., Rolander, D. D., Moore, M. & Washer, G. Visual inspection of highway bridges. J. Nondestr. Eval.
21, 67–83 (2002).
4. Lynch, J. P. & Loh, K. J. A summary review of wireless sensors and sensor networks for structural health monitoring. Shock Vib.
Dig. 38, 91–130 (2006).
5. Shen, N. et al. A review of global navigation satellite system (GNSS)-based dynamic monitoring technologies for structural health
monitoring. Remote Sens. 11, 1001 (2019).
6. Colombani, I. A. & Andrawes, B. A study of multi-target image-based displacement measurement approach for field testing of
bridges. J. Struct. Integr. Maint. 7, 207–216 (2022).
7. Lee, Y., Lee, G., Moon, D. S. & Yoon, H. Vision-based displacement measurement using a camera mounted on a structure with
stationary background targets outside the structure. Struct. Control Health Monit. 29, e3095 (2022).
8. Sharma, S., Dangi, S. K., Bairwa, S. K. & Sen, S. Comparative study on sensitivity of acceleration and strain responses for bridge
health monitoring. J. Struct. Integr. Maint. 7, 238–251 (2022).
9. Zhang, Q., Zhang, J., Duan, W. & Wu, Z. Deflection distribution estimation of tied-arch bridges using long-gauge strain measure-
ments. Struct. Control Health Monit. 25, e2119 (2018).
10. Ye, C., Butler, L. J., Elshafie, M. Z. & Middleton, C. R. Evaluating prestress losses in a prestressed concrete girder railway bridge
using distributed and discrete fibre optic sensors. Constr. Build. Mater. 247, 118518 (2020).
11. Lee, Y., Park, W. J., Kang, Y. J. & Kim, S. Response pattern analysis-based structural health monitoring of cable-stayed bridges.
Struct. Control Health Monit. 28, e2822 (2021).
12. Nazimko, V. Synchronization of the yield as a way to increase integrity and bearing capacity of the frame supports. J. Struct. Integr.
Maint. 7, 1–14 (2022).
13. Šedková, L., Vlk, V. & Šedek, J. Delamination/disbond propagation analysis in adhesively bonded composite joints using guided
waves. J. Struct. Integr. Maint. 7, 25–33 (2022).
14. Pandey, A. & Biswas, M. Damage detection in structures using changes in flexibility. J. Sound Vib. 169, 3–17 (1994).
15. Farrar, C. & James Iii, G. System identification from ambient vibration measurements on a bridge. J. Sound Vib. 205, 1–18 (1997).
16. Vestroni, F. & Capecchi, D. Damage detection in beam structures based on frequency measurements. J. Eng. Mech. 126, 761–768
(2000).
17. Siringoringo, D. M. & Fujino, Y. System identification of suspension bridge from ambient vibration response. Eng. Struct. 30,
462–477 (2008).
18. Soyoz, S. & Feng, M. Q. Long-term monitoring and identification of bridge structural parameters. Comput. Aided Civ. Infrastruct.
Eng. 24, 82–92 (2009).
19. Xu, Z.-D. & Wu, Z. Energy damage detection strategy based on acceleration responses for long-span bridge structures. Eng. Struct.
29, 609–617 (2007).
20. Hester, D. & González, A. A wavelet-based damage detection algorithm based on bridge acceleration response to a vehicle. Mech.
Syst. Signal Process. 28, 145–166 (2012).
21. Chen, Z., Pan, C. & Yu, L. Structural damage detection via adaptive dictionary learning and sparse representation of measured
acceleration responses. Measurement 128, 377–387 (2018).
22. Zhang, J., Guo, S., Wu, Z. & Zhang, Q. Structural identification and damage detection through long-gauge strain measurements.
Eng. Struct. 99, 173–183 (2015).
23. Cui, H., Du, H., Zhao, F. & Hong, M. Damage identification in a plate structure based on a cross-direction strain measurement
method. Measurement 158, 107714 (2020).
24. Qureshi, Y., Tarfaoui, M., Lafdi, K. K. & Lafdi, K. Real-time strain monitoring and damage detection of composites in different
directions of the applied load using a microscale flexible Nylon/Ag strain sensor. Struct. Health Monit. 19, 885–901 (2020).
25. Lee, E.-T. & Eun, H.-C. Damage detection of damaged beam by constrained displacement curvature. J. Mech. Sci. Technol. 22,
1111–1120 (2008).
26. Xu, B., Song, G. & Masri, S. F. Damage detection for a frame structure model using vibration displacement measurement. Struct.
Health Monit. 11, 281–292 (2012).
27. He, W.-Y., Ren, W.-X. & Zhu, S. Damage detection of beam structures using quasi-static moving load induced displacement
response. Eng. Struct. 145, 70–82 (2017).
28. Staszewski, W. Intelligent signal processing for damage detection in composite materials. Compos. Sci. Technol. 62, 941–950 (2002).
29. Li, H., Deng, X. & Dai, H. Structural damage detection using the combination method of EMD and wavelet analysis. Mech. Syst.
Signal Process. 21, 298–306 (2007).
30. Dao, P. B. & Staszewski, W. J. Lamb wave based structural damage detection using cointegration and fractal signal processing.
Mech. Syst. Signal Process. 49, 285–301 (2014).
31. Figueiredo, E., Park, G., Farrar, C. R., Worden, K. & Figueiras, J. Machine learning algorithms for damage detection under opera-
tional and environmental variability. Struct. Health Monit. 10, 559–572 (2011).
32. Ghiasi, R., Torkzadeh, P. & Noori, M. A machine-learning approach for structural damage detection using least square support
vector machine based on a new combinational kernel function. Struct. Health Monit. 15, 302–316 (2016).
33. Zhang, Z. & Sun, C. Multi-site structural damage identification using a multi-label classification scheme of machine learning.
Measurement 154, 107473 (2020).
34. Messina, A., Williams, E. & Contursi, T. Structural damage detection by a sensitivity and statistical-based method. J. Sound Vib.
216, 791–808 (1998).
35. Xu, G., Zhu, W. & Emory, B. Experimental and numerical investigation of structural damage detection using changes in natural
frequencies (2007).
36. Zhao, B., Xu, Z., Kan, X., Zhong, J. & Guo, T. Structural damage detection by using single natural frequency and the corresponding
mode shape. Shock Vib. 2016 (2016).
37. Kim, J.-T., Ryu, Y.-S., Cho, H.-M. & Stubbs, N. Damage identification in beam-type structures: Frequency-based method vs mode-
shape-based method. Eng. Struct. 25, 57–67 (2003).
38. Yang, Z. & Wang, L. Structural damage detection by changes in natural frequencies. J. Intell. Mater. Syst. Struct. 21, 309–319 (2010).
39. Paral, A., Singha Roy, D. K. & Samanta, A. K. Application of a mode shape derivative-based damage index in artificial neural
network for structural damage identification in shear frame building. J. Civ. Struct. Health Monit. 9, 411–423 (2019).
40. Wu, J. & Li, Q. Finite element model updating for a high-rise structure based on ambient vibration measurements. Eng. Struct. 26,
979–990 (2004).
41. Jaishi, B. & Ren, W.-X. Structural finite element model updating using ambient vibration test results. J. Struct. Eng. 131, 617–628
(2005).
42. Ren, W.-X. & Chen, H.-B. Finite element model updating in structural dynamics by using the response surface method. Eng. Struct.
32, 2455–2465 (2010).
43. Park, G., Hong, K.-N. & Yoon, H. Vision-based structural FE model updating using genetic algorithm. Appl. Sci. 11, 1622 (2021).
44. Park, G., Jung, J. & Yoon, H. Structural finite element model updating considering soil-structure interaction using ls-dyna in loop.
Sci. Rep. 13, 4753 (2023).
45. Goodfellow, I., Bengio, Y. & Courville, A. Deep Learning (MIT Press, 2016).
46. Truong, T. T., Dinh-Cong, D., Lee, J. & Nguyen-Thoi, T. An effective deep feedforward neural networks (DFNN) method for
damage identification of truss structures using noisy incomplete modal data. J. Build. Eng. 30, 101244 (2020).
47. Le, H. Q., Truong, T. T., Dinh-Cong, D. & Nguyen-Thoi, T. A deep feed-forward neural network for damage detection in function-
ally graded carbon nanotube-reinforced composite plates using modal kinetic energy. Front. Struct. Civ. Eng., 1–27 (2021).
48. Fernandez-Navamuel, A. et al. Supervised deep learning with finite element simulations for damage identification in bridges. Eng.
Struct. 257, 114016 (2022).
49. Choe, D.-E., Kim, H.-C. & Kim, M.-H. Sequence-based modeling of deep learning with LSTM and GRU networks for structural
damage detection of floating offshore wind turbine blades. Renew. Energy 174, 218–235 (2021).
50. Iannelli, P., Angeletti, F., Gasbarri, P., Panella, M. & Rosato, A. Deep learning-based Structural Health Monitoring for damage
detection on a large space antenna. Acta Astronaut. 193, 635–643 (2022).
51. Min, S., Lee, Y., Byun, Y.-H., Kang, Y. J. & Kim, S. Merged LSTM-based pattern recognition of structural behavior of cable-
supported bridges. Eng. Appl. Artif. Intell. 125, 106774 (2023).
52. Lin, Y. Z., Nie, Z. H. & Ma, H. W. Structural damage detection with automatic feature-extraction through deep learning. Comput.
Aided Civ. Infrastruct. Eng. 32, 1025–1046 (2017).
53. Sony, S. Towards Multiclass Damage Detection and Localization Using Limited Vibration Measurements (The University of Western
Ontario, 2021).
54. Vy, V. et al. Damage localization using acoustic emission sensors via convolutional neural network and continuous wavelet trans-
form. Mech. Syst. Signal Process. 204, 110831 (2023).
55. Cha, Y.-J. & Wang, Z. Unsupervised novelty detection-based structural damage localization using a density peaks-based fast
clustering algorithm. Struct. Health Monit. 17, 313–324 (2018).
56. Rastin, Z., Ghodrati Amiri, G. & Darvishan, E. Unsupervised structural damage detection technique based on a deep convolutional
autoencoder. Shock Vib. 2021, 1–11 (2021).
57. Wang, Z. & Cha, Y.-J. Unsupervised deep learning approach using a deep auto-encoder with a one-class support vector machine
to detect damage. Struct. Health Monit. 20, 406–425 (2021).
58. Nelder, J. A. & Mead, R. A simplex method for function minimization. Comput. J. 7, 308–313 (1965).
59. Refaeilzadeh, P., Tang, L. & Liu, H. Cross-validation. Encyclopedia of Database Systems, 532–538 (2009).
Acknowledgements
This work was supported by National Research Foundation of Korea (NRF) grants funded by the Korean govern-
ment (MSIT) (NRF-2022R1C1C1003012 and NRF-2022R1A6A3A01086405).
Author contributions
Y.L. managed the overall research progress and wrote the main manuscript. H.K. carried out finite element analy-
sis and data analysis. S.M. developed and improved the deep learning algorithms. H.Y. directed and supervised
the overall research. All authors reviewed the manuscript.
Competing interests
The authors declare no competing interests.
Additional information
Correspondence and requests for materials should be addressed to H.Y.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the
Creative Commons licence, and indicate if changes were made. The images or other third party material in this
article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from
the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at