JOURNAL OF MATERIALS AND ENGINEERING STRUCTURES 9 (2022) 403–410 403

Research Paper

Damage detection of structural based on indirect vibration

measurement results combined with Artificial Neural Network

Nguyen Thi Cam Nhung a, Tran Quang Minh b,*, Hélder S. Sousa b, Ngo Van Thưc c, José C.
Matos b
a
University of Transport and Communications, Vietnam
b
University of Minho, ISISE, Department of Civil Engineering, Guimarães, Portugal
c
Mien Tay Construction University, Viet Nam

ARTICLE INFO ABSTRACT

Article history: In Structural Health Monitoring (SHM), damage detection and maintenance are among
the most critical factors. For surface damage, damage detection is simple and easy to
Received : 15 November 2022
perform. However, detecting and repairing is difficult for damage hidden deep in the
Revised : 16 December 2022 structure. Using the structure's dynamic features, damage can be detected and repaired in
time. With the development of sensor technology, indirect vibration measurement
Accepted : 24 December 2022
solutions give accurate results, minimizing errors by infinitely increasing the number of
measurements. This solution offers a great opportunity to reduce the cost of structural
health monitoring. Based on the large amount of data obtained from indirect monitoring,
Keywords: artificial intelligence technologies can be used to obtain a more comprehensive model of
SHM. In this paper, the dynamic responses of the structure will be extracted and
Indirect monitoring
determined through a vehicle crossing the bridge. Based on the results of structural
Damage detection dynamic response, a finite element model is built and updated so that this model can
represent the real structure. Damage cases will be analyzed and evaluated as input to train
Drive-by
the Artificial neural network. The trained network can detect damage through regular
Artificial neural network health monitoring by indirect methods.
F. ASMA & H. HAMMOUM (Eds.) special issue, 4th International Conference on Sustainability in
Civil Engineering ICSCE 2022, Hanoi, Vietnam, J. Mater. Eng. Struct. 9(4) (2022)

1 Introduction

The indirect monitoring technique is based on the dynamic response collected by sensors installed on the vehicle during
movement. This method brings more benefits than the traditional method, so has been interested and developed by many
researchers. Many authors have studied the feasibility and data processing of indirect monitoring [1-3]. Yang et al. [4, 5] first

* Corresponding author. Tel.: +351933312076

E-mail address: [email protected] / [email protected]

e-ISSN: 2170-127X,
404 JOURNAL OF MATERIALS AND ENGINEERING STRUCTURES 9 (2022) 403–410

proposed to use dynamic data extracted from vehicles crossing the bridge structure to determine the natural frequency of the
structure. By collecting vehicle data multiple times over the bridge, some damage was detected. Yang et al found this data
can be used to detect structural damage. McGetrick et al [6, 7] used a number of sensors fixed on the vehicle over the bridge.
Thereby, the natural frequencies of the bridge structure are determined. Although there are many problems with accuracy,
this can be compensated for by performing multiple monitoring, with lots of data. The detection of damage of the structure
will be detected if there is a large difference between many measurements. At the same time, this helps to reduce the cost of
monitoring equipment installed directly on the constructions without being used on other buildings. This indirect method is
capable of commercializing and attracting investment for many projects. With just one equipment investment, it can be used
for many different projects.
Today, with the great development of information technology, especially in the field of artificial intelligence and the
success of artificial neural networks in intellectual training for computers, there have been many studies incorporating
artificial intelligence in predicting and diagnosing events [8, 9]. It can be said that artificial intelligence has opened a new era
of development for industries, including structural health monitoring. In this article, by using indirect vibration measurement
results combined with artificial neural networks; the authors propose a new approach to detecting structural damage.

2 Research methods and approach

2.1 Indirect monitoring

In indirect monitoring techniques, the vehicle can be considered as both exciter and receiver (Fig. 1)[10]:

Fig. 1 – Indirect monitoring of a railway bridge

The interaction between the bridge and vehicle leads to a second order differential equation. This equation is written in
matrix form as follows:

[ M ].{q}  [C ].{q}  [ K ]{q}  [ F ] (1)

[M ] is the mass matrix, [C ] is the damping matrix; [ K ] represent the stiffness matrix, {q} is the displacement vector
and [ F ] denotes the external force vector. In the road bridge, because it is greatly affected by the flatness of the road surface,
indirect monitoring has not proved effective. However, for railway bridges, the dynamic responses of the structure can be
easily extracted through the dynamic response collected from the train crossing the bridge. This comes from the fact that the
railway bridge is not affected much by the contact surface. At the same time, the operating speed of the train is relatively
stable, there is not much fluctuation like that of road vehicles. One of the important data that can be extracted from the
dynamic response of the train when crossing the bridge is the frequency of the structure and the displacement at the sensor
mounting position over time.

2.2 Artificial neural network

An Artificial Neural Network (ANN) (Fig. 2) is a network (include many components) inspired by the Human nervous
system. ANN is one of the breakthroughs in artificial intelligence thanks to its ability to learn from experience and improve
itself by creating different paths to connect the neurons of the network. Applications of ANN include classification, pattern
recognition, control systems, and image processing. ANN consists of the input, hidden, and output layers. Each layer consists
of neurons that contain information and are linked by training parameters (weight and bias). At each neuron, a processing
element will be built and connected with other synapses based on the number of neurons in front of it.
 JOURNAL OF MATERIALS AND ENGINEERING STRUCTURES 9 (2022) 403–410 405

Fig. 2 – Artificial neural network architecture [11]

In this paper, based on indirect monitoring results, the finite element model is built and updated. This model has the same
behaviour with real bridge structure. Then, damage data will be generated corresponding to the data that can be obtained
from indirect monitoring data (here, natural frequency and displacement at a certain position). Finally, these data are used to
train the network.

3 Case study

3.1 Cat Linh – Ha Dong railway bridge

Cat Linh-Ha Dong is an urban railway line belonging to the Hanoi Urban Railway network connecting Cat Linh and Ha
Dong (Fig. 3). The total length of the route is 13021.48m, the whole route is designed on high. The span structure of the line
is a combination of simple spans. The span length is one of 18.5, 20, 26, 29, 30, 32m. Span structure with box cross section
(Fig. 4)

Fig. 3 – Cat Linh-Ha Dong railway bridge Fig. 4 – Cross section of span

3.2 Indirect monitoring of Cat Linh Ha Dong railway bridge

Using 4accelerometer sensors installed vertically and horizontally of the bridge. The sensors are mounted on the deck of
the train, at the position corresponding to the 4 wheels (Fig. 5).

Fig. 5 – Sensor installation diagram

406 JOURNAL OF MATERIALS AND ENGINEERING STRUCTURES 9 (2022) 403–410

For this field investigation, monitoring was carried out by collecting the data from train through bridge structure. During
the bridge traverse, the time was recorded and marked the timestamp when crossing the spans of the bridge. Based on the
marking times, the data will be broken down, extracting the data of the structure to be monitored, other data was discarded.
CompactDAQ Chassis (cDAQ‑9178) is a signal transmitter to a computer. The data received from the sensor will go through
the Chassis and store on the computer (Fig. 6).The data is considered and analyzed for a 32m.

Fig. 6 – Indirect monitoring data collection

All data was stored in the laptop for processing using MACEC toolbox [12]. Data is processed in steps to ensure accuracy.
The Fast Fourier Transform (FFT) was employed to convert the obtained dynamic signal from the time domain to a
representation in the frequency domain. The frequency range of interest was limited to an interval 0-12Hz, to reduce analysis
and identification time. The dynamic characteristics are estimated by a stabilization diagram. Simultaneously, by integrating
twice the accelerometer data from the accelerometer, the sensor displacement data on the wheel is recorded. The frequency
and displacement plot in time are shown in the Fig. 7.
After processing vibration measurement data in the time domain by parametric method through the program to identify
vibration characteristics by MACEC tool. The natural frequency results are f1=1.371, f2=15.5Hz. Displacement plot shows
the vertical displacement of the sensor. This is also the displacement of the wheel and the span structure. Based on these two
results, perform model update and generate data to train the network

a. b.

Fig. 7–Result of indirect monitoring: a. Stabilization diagram; b measuring point displacement

 JOURNAL OF MATERIALS AND ENGINEERING STRUCTURES 9 (2022) 403–410 407

3.3 Build a finite element (FE) model and update the model

With the bridge structure considered, a FE model of the bridge is constructed. The built model consists of 64 beam
elements. Each element node contains 6 degrees of freedom (DOF) corresponding to translational and rotational
displacements in the X, Y and Z axes.

Fig. 8–FE model of span 32m

The train specifications are also shown in the figure below:

Fig. 9 –Train specifications

Accurate determination of material parameters requires many experiments; moreover, the results will also give a range
of values that are not exactly a specific number. Similar to the case under consideration, the parameters for concrete have not
been determined at all, but only available studies have been used to determine the limits of values of these parameters,
parameters of the material, the stiffness of the bearing also need to be considered. In order for the FEM model to have the
same behavior with the real structure, it is necessary to select the value of uncertainty parameters so that the difference
between the experimentally measured values and the model results is minimum. An objective function based on the natural
frequency and the displacement is used to evaluate the similarity of the FEM model and the real structure. After updating,
the uncertain parameters are optimized and presented in Table 1:

Table 1-Uncertain parameters

No Uncertain parameters Upper Lower Initial Updated

bound bound value value

1 Young’s modulus of concrete - E (GPa) 40.00 45.00 42.00 43.45

2 Weight density of concrete - ρ (kg/m3) 2500 2800 2500 2651.8

3 Stiffness of bearing- k (N/m) 1 × 1010 4 × 1010 1 × 1010 1.568 × 108

3.4 Generate data and train the ANN model

In this study, the architecture of ANN includes 1 input layer, 1 hidden layer, and 1output layer. Input data used in training
the network includes natural frequencies and displacements with different damage scenarios. The output data used is the level
of damage of the considering structure. Modal analysis of updated model is performed to to generate input and output data
for the ANN. The stiffness of the elements is gradually reduced to create damage in the structure. Consider failures from 0-
100% of the cases. The stiffness of the elements considered will decrease from 0-100% with a step of 1%. The load model
(train) is simulated as for real train specifications. The total number of samples used to train the network is 6400 samples.
408 JOURNAL OF MATERIALS AND ENGINEERING STRUCTURES 9 (2022) 403–410

Fig. 10 –Regression values of ANN

After generating the data, the ANN is set to the parameters for training the network. A loop is performed to select the
most suitable number of hidden layers. Setting the number of hidden layers that will be looped within the value range from
1 to 50. To ensure accuracy, some noise effects are also added to the input data with a noise level of 2%.ANN employs the
Levenberg-Marquardt back propagation algorithm to train the network. Data split in the training process 70%-15%-15% is
used for damage identification. The maximum number of epochs is set to 1000.

Fig. 11 – Error histogram of ANN Fig. 12 –Tolerance of network

 JOURNAL OF MATERIALS AND ENGINEERING STRUCTURES 9 (2022) 403–410 409

Fig. 10, Fig. 11, Fig. 12, present the regression values of training value set, validation value set and test value set. All of
the training cases are higher 0.99. The training dataset, validation dataset and test dataset are stick to the target line (45degree
line). This shows that the predicted value and the actual value are almost similar. The calculated regression value (R) in the
above training case is close to the value 1, which represents a high match between the predicted results and the target results.
Fig. 11 shows the error distribution between the predicted and target values. The difference between the target and the output
is very small. The network tolerance is very low, which shows that the network is working fine, the values in the sets are not
overfitting. Through the above results, it is possible to use indirect vibration measurement results in combination with
artificial neural networks to detect structural damage.

4 Conclusions

Through the combined approach of indirect monitoring and artificial intelligence network, this study presents a new
solution for structural damage detection. The method is practically applied at the Cat Linh - Ha Dong railway project, with
positive results. The main conclusions drawn are as follows:
The indirect monitoring method has potential in structural health monitoring, especially in the field of railway bridges.
By collecting data during mining, it is possible to predict structural health. This method allows large amounts of data to be
collected by allowing an infinite number of measurements during service operation, and it is not necessary to stop serving
during data collection.
The combination of indirect monitoring and artificial intelligence network can accurately predict the location of damage
on the structure. In this paper, using indirect measurement results, a FE model is updated and data generated. From the
generated data, the artificial neural network model is trained and makes future predictions
The values of the trained artificial neural network are achieved at very high (greater than 0.99).The model using indirect
monitoring data is successfully trained. In further studies, research is needed to increase training performance and network
training speed

Acknowledgements

The authors acknowledge the financial support of the project research “B2022-GHA- 03” of the Ministry of Education
and Training. This work was partly financed by FCT/MCTES through national funds (PIDDAC) under the R&D Unit Institute
for Sustainability and Innovation in Structural Engineering (ISISE), under reference UIDB/04029/2020. Tran Quang Minh
was supported by the doctoral Grant reference PRT/BD/154268/2022 financed by Portuguese Foundation for Science and
Technology (FCT), under MIT Portugal Program (2022 MPP2030-FCT). The third author acknowledges the funding by FCT
through the Scientific Employment Stimulus - 4th Edition.

REFERENCES

[1]- Yang, Y., Y. Li, and K. Chang, Effect of road surface roughness on the response of a moving vehicle for
identification of bridge frequencies. Interact. Multisc. Mech., 5(4) (2012) 347-368.
doi:10.12989/imm.2012.5.4.347.
[2]- Nhung, N.T.C., T.Q. Minh, J.C. e Matos, and H.S. Sousa. Research and application of indirect monitoring methods
for transport infrastructures to monitor and evaluate structural health. in 2nd International Conference on Structural
Health Monitoring and Engineering Structures. Select Proceedings of SHM&ES 2022. (2022).
[3]- NHUNG, N.T.C. and Q.M. TRAN, The effects of ground vibration induced by construction activities of urban
railways in Hanoi. J. Mat. Eng. Str. «JMES», 7(4) (2020) 703-710.
[4]- Yang, Y., Y. Lee, and K. Chang, Effect of road surface roughness on extraction of bridge frequencies by moving
vehicle, in Mechanics and model-based control of advanced engineering systems. Springer: Vienna, Austria. (2014),
295-305. doi:10.1007/978-3-7091-1571-8_32.
[5]- Yang, Y.-B., C. Lin, and J. Yau, Extracting bridge frequencies from the dynamic response of a passing vehicle. J.
Sound Vibrat., 272(3-5) (2004) 471-493. doi:10.1016/S0022-460X(03)00378-X.
[6]- McGetrick, P.J. The use of an instrumented vehicle to monitor transport infrastructure. University College Dublin,
2012.
[7]- McGetrick, P.J. and C.W. Kim. A parametric study of a drive by bridge inspection system based on the Morlet
410 JOURNAL OF MATERIALS AND ENGINEERING STRUCTURES 9 (2022) 403–410

wavelet. in Key Engineering Materials. Trans Tech Publ. (2013), 262-269.

[8]- Tran, M.Q., A.M. Bento, E. Teixeira, H. Sousa, and J.C. e Matos. Application of Artificial Intelligence (AI) in Flood
Risk Forecasting. in Book of Abstracts. 275.
[9]- Carden, E.P. and P. Fanning, Vibration based condition monitoring: a review. Struct. Health Monit., 3(4) (2004)
355-377. doi:10.1177/1475921704047500.
[10]- Malekjafarian, A., P.J. McGetrick, and E.J. OBrien, A review of indirect bridge monitoring using passing vehicles.
Shock Vibrat., 2015 (2015). doi:10.1155/2015/286139.
[11]- Peter Galan Introduction to artificial neural networks in control applications. A Control Engineering / CFE Media
and Technology article 2022.
[12]- Reynders, E., M. Schevenels, and G. De Roeck, Macec 3.3. A Matlab toolbox for experimental and operational
modal analysis. User Manual, Leuven, (2014).