Authors Accepted Manuscript

Seismic Reliability Assessment of Structures Using

Artificial Neural Network

Sayyed Mohsen Vazirizade, Saeed Nozhati,

Mostafa Allameh Zadeh

www.elsevier.com/locate/jobe

PII: S2352-7102(16)30316-3
DOI: http://dx.doi.org/10.1016/j.jobe.2017.04.001
Reference: JOBE239
To appear in: Journal of Building Engineering
Received date: 21 November 2016
Accepted date: 2 April 2017
Cite this article as: Sayyed Mohsen Vazirizade, Saeed Nozhati and Mostafa
Allameh Zadeh, Seismic Reliability Assessment of Structures Using Artificial
Neural Network, Journal of Building Engineering,
http://dx.doi.org/10.1016/j.jobe.2017.04.001
This is a PDF file of an unedited manuscript that has been accepted for
publication. As a service to our customers we are providing this early version of
the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting galley proof before it is published in its final citable form.
Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal pertain.
 Seismic Reliability Assessment of Structures Using Artificial Neural Network

Sayyed Mohsen Vazirizade

Department of Civil and Environmental Engineering,
M.Sc. Graduate of Civil Engineering,
Sharif University of Technology,
Azadi st, Tehran, P.O.B 11155-9313, Iran email :sayyed.mohsen.vazirizade@iiees.ac.ir

Saeed Nozhati
Department of Mathematics, Statistics and Computer Science
Marquette University, Milwaukee, 53051, WI, U.S.A. email :saeed.nozhati@marquette.edu

Mostafa Allameh Zadeh, Corresponding Author

Department of Seismology
International Institute of Earthquake Engineering and Seismology
No. 21, Arghavan St., North Dibajee, Farmanieh, Tehran, P.O.B 19537-14453, Iran email: mallam@iiees.ac.ir

Abstract.
Localization and quantification of structural damage and estimating the failure probability are
key outputs in the reliability assessment of structures. In this study, an Artificial Neural Network
(ANN) is used to reduce the computational effort required for reliability analysis and damage
detection. Toward this end, one demonstrative structure is modeled and then several damage
scenarios are defined. These scenarios are considered as training data sets for establishing an
ANN model. In this regard, the relationship between structural response (input) and structural
stiffness (output) is established using ANN models. The established ANN is more economical
and achieves reasonable accuracy in detection of structural damage under a set of ground
motions. Furthermore, in order to assess the reliability of a structure, five random variables are
considered. These are columns area of the first, second, and third floor, elasticity modulus, and
gravity loads. The ANN is trained by suing the Monte Carlo Simulation (MCS) technique.
Finally, the trained neural network specifies the failure probability of the proposed structure.
Although MCS can predict the failure probability for a given structure, the ANN model helps
simulation techniques to receive an acceptable accuracy and reduce computational effort.
KeyWords: Seismic Reliability; Artificial Neural Network; Monte Carlo Simulation; Failure Probability

1. Introduction
The main reason for structural failure is a sudden damage. In the past decades, special attention was given
to avoid the unexpected failure of structural components by damage detection in structures in the early
states. To this end, in recent years, various developments of non-destructive techniques based on changes
in the structural responses have been widely published. They can not only detect the presence of damage

1
but also identify the location and quantification of it. Additionally, the dire need to detect the presence of
damage in complex structures and infrastructures in the early stages has led to the increase of non-
destructive techniques and new developments [1-3]. During the past decades, many types of research have
been studying to propose different and efficient techniques. Friswell [4] presented a brief overview of the
use of inverse methods in damage detection and location from response data. A review based on the
detection of structural damage through changes in frequencies has been discussed by Salawu [5].
However, in the presence of complex structures, many of them are not applicable. Therefore, the methods
that are much more economical to achieve reasonable accuracy are always required. In recent years, there
has been a growing interest in using Artificial Neural Networks (ANNs), a computing technique that was
supposed to work in a way similar to that of biological nervous systems; however, nowadays, we know
biological nervous systems are far more complicated than ANNs. By the way, a large number of studies
corroborate this idea that in spite of simplicity, ANNs are a fruitful approach to solving the problems.
Many researchers [6], [7] used ANN to study a beam using multilayer perceptron (MLP) ANN.
Furthermore, another application of ANN is to the evaluatio of the failure probability and safety levels of
structural systems. Bakhshi and Vazirizade [8] used a radial network in order to predict the stiffness of
each member in a frame according to its response to a record. Although the ground motion records can be
reduced [9], the full-length records have been used.
In fact, they showed ANN can provide a mapping from the maximum story drifts to column stiffness.
Gomes et al. [10] and Bucher [11] used ANN for obtaining the failure probability for a cantilever beam
and compared ANN with other conventional methods. They found that ANN methods that approximate
the limit state function may decrease the total computational effort on the reliability assessment, but more
studies, including large systems with non-linear behavior, must be conducted. Elhewy et al. [12] studied
the ability of ANN model to predict the failure probability of a composite plate. They compared the
performance of the ANN-based RSM (Response Surface Methods) (ANN-based FORM and ANN-based
MCS) with that of the polynomial-based RSM. Their results showed that the ANN-based RSM was more
efficient and accurate than the polynomial-based RSM. It was shown that the RSM may not be precise
when the probability of failure is extremely small; and the RSM requires a relatively long computation
time as the number of random variables increases [13], [14]. Zhang and Foschi [15] employed ANN for
seismic reliability assessment of a bridge bent with and without seismic isolation, but in that case they
used explicit limit states. However, most of them utilized explicit and approximate limit states and more
focused on the reliability assessment of components by ANN. In this regard, this study is focused on two
separate parts; (1) localization and quantification of structural damages using ANN; (2) seismic reliability
assessment of one steel structure using ANN-based MCS.

2. Artificial Neural Network (ANN)

Artificial neural networks are comparatively crude electronic models. The advances in biological research
promise an initial understanding of the neural thinking mechanism [16]. The basic network includes
nodes and connections, which link the nodes. Each link and node are related to a weight and bias
properties, respectively, which are the principal mechanism by which a network stocks information.
Before a neural network can approximate complex unities, it has to be trained for the specific problem by
adjusting these weights and biases. One of the most widely used network types for approximation is the
feed-forward multi-layer Perceptron (MLP) trained by the back-propagation algorithm. Figure 1 shows
this schematic network type which is used in this study. This network consists of an input layer, one
hidden layer, and an output layer. The input and output layers contain three and two neurons, which
means three and two sets of data for input and output, respectively.

2
 1
1
2 1
2
3 2
3
4

Figure 1: Schematic structure of an artificial neural network

A neuron from the hidden layer is shown in Figure 2 with three inputs. Each input is weighted with an
appropriate w. The sum of the weighted inputs and the bias forms the input to the transfer function.
Transfer function prepares the data for the next layer. In this figure, the next layer is the output layer
which contains two neurons.

Weights Bias
b
w1
Transfer
Function

Output
Input

w2

w3

Figure 2: Schematic neuron from the hidden layer

The following equation simulates the mathematical relations between inputs and outputs in a
network.

m
a j i f (x ) f w j
k ,i a j 1k b j i (1)
k 1
Where aji is the output value of ith neuron in the jth layer, which is sent to the j+1th layer. aj-1k is the
output value of kth neuron in the j-1th layer, which is sent to the jth layer. m is the number of data as
inputs or the number of neurons in the j-1th layer, i is the number of the current neuron in the jth layer.
w j k ,i is the synaptic weight factor for the connection of the neuron i in the jth layer with the neuron k in
the j-1th layer. bji is the bias value of ith neuron in the jth layer, and f is the transfer function. For the
input layer j can be considered zero and m for this layer mean the number of network inputs.
Subsequently, m for the last layeroutput layeris the number of output data of the network.

3
It is mentioned [17] that the number of training samples n should be larger than the number of adjustable
parameters.

(m + 2)M + 1 < n (2)

Where m is the number of input values and M is the number of hidden neurons for a network with a single
hidden layer. This leads to a much smaller number of required samples compared to RSM for high-
dimensional problems if the number of hidden neurons taken is not too large. In Ref. [18] two other
approaches are discussed to avoid over-training for a large number of hidden neurons and a small number
of training samples: regularization and early stopping. For further details on ANN [19], [20] can be
referenced.

3. Methodology and ground motion record selection

In this study, a 3-story steel frame building is modeled by Open System for Earthquake Engineering
Simulation Software (OpenSees) [21], Figure 3. The steel constitutive behavior is modeled using the
elastic-perfectly plastic steel model. The initial design of all stories for columns and beams is the same. In
this study, a set of twenty earthquakes selected from FEMA440 [22] recorded on Site Class C, are used.
These ground motion records are listed in Table 1. The analyses have been done for the set of ground
motions and the mean and 95% confidence interval is computed.
3 @ 3m

0.183 m

0.1 m

4m
(a) (b)

Figure 3: a) Overview of the three-story frame b) Cross Section of Columns

Table 1: The suite of twenty ground motion records used

Magnitude
Earthquake Name Station Name Component(deg) PGA (cm/s2)
(Ms)
Imperial Valley, 1979 El Centro, Parachute Test Facility 6.8 315 200.2
San Fernando, 1971 Pasadena, CIT Athenaeum 6.5 90 107.9
San Fernando, 1972 Pearblossom Pump 6.5 21 133.4
Landers, 1992 Yermo, Fire Station 7.5 0 167.8

4
 Loma Prieta, 1989 APEEL 7, Pulgas 7.1 0 153
Loma Prieta, 1990 Gilroy #6, San YsidroMicrowavw Site 7.1 90 166.9
Loma Prieta, 1990 Saratoga, Aloha Ave 7.1 0 494.5
Loma Prieta, 1990 Gilroy, Gavilon College PhysSchBldg 7.1 67 349.1
Loma Prieta, 1990 Santa Cruz, University of California 7.1 360 433.1
Loma Prieta, 1990 San Francisco, Dimond Heights 7.1 90 110.8
Loma Prieta, 1990 Fremont, Mission San Jose 7.1 0 121.6
Loma Prieta, 1990 Monterey, City Hall 7.1 0 71.6
Loma Prieta, 1990 Yerba Buena Island 7.1 90 66.7
Loma Prieta, 1990 Anderson Dam, Downstream 7.1 270 239.4
Morgan Hill, 1984 Gilroy, Gavilon College PhysSciBldge 6.1 67 95
Morgan Hill, 1984 Gilroy #6, San YsidroMicrowavw Site 6.1 90 280.4
Palmsprings, 1986 Fun Valley 6 45 129
Northridge, 1994 Littlerock, Brainard Canyon 6.8 90 70.6
Northridge, 1994 Castaic, Old Ridge Route 6.8 360 504.2
Northridge, 1994 Lake Hughes #1, Fire station #78 6.8 0 84.9

3.1. ANN model for damage detection

In order to find the location and quantification of damages in the interested structure, two different data
sets are considered; (a) 64 different damage scenarios4 scenarios for each story(b) 729 different
damage scenarios9 scenarios for each story. It is noteworthy that these damage scenarios are based on
damages in the columns, which are presented as cross section reduction. The initial area of each column is
roughly equal to IPE20. According to the aforementioned damage scenarios, 64 and 729 different sets of
areas are defined less than IPE20. For each damage scenario, OpenSees analyzes the damaged structure,
and its outputs are the inputs of the ANN model. Subsequently, the maximum relative displacement of
each story is selected as an input for the ANN model, and cross section area of columns are considered as
an output data set. Thus, the ANN model has three Process Elements (PEs). Figure 4 clearly
demonstrates the arrangement of inputs and outputs in the network.

Figure 4 Inputs and Outputs for ANN

The number of the hidden nodes is specified based on equation (2). This procedure is performed through
the interaction between MATLAB and OpenSees, which is showed in Figure 5.

5
 Figure 5: ANN training for Damage detection through the interaction between Matlab and OpenSees

Each the aforementioned data set is divided into three different subsets; (a) training, (b) cross-validation
and (c) test. The ANN model accuracy is verified by Root Mean Square Error (RMSE), i.e. the difference
between the responses predicted by ANN model and actual data, and is calculated as
n
(y i y i )2
RMSE = MSE i 1
(3)
n

In the above expression, y 1 , y 2 ,, y n are instances of response values in the dataset, y 1 , y 2 ,, y n
are predicted values and n are the total numbers of points in the dataset. If the error with respect to this
subset is not acceptable, the training may be repeated. Indeed, this test is critical to ensure that the
network has successfully learned the correct functional relationship within the whole set of data.
The maximum relative displacements of each story for each data set scenario with 95% confidence
interval around the mean value are plotted in Figure 5 and Figure 6. RMSEs calculated for each scenario
are plotted on the left parts of Figure 7 and Figure 8. Box and Whisker plot, as shown on the right sides of
Figure 7 and Figure 8, is used to more properly show a central tendency, departure from symmetry, and
variability in values of RMSE in different stories.

Figure 6: The maximum relative displacements of each story with 95% confidence interval around the mean for 64 scenarios

6
 Figure 7: The maximum relative displacements of each story with 95% confidence interval around the mean for 729 scenarios

Figure 8: RMSEs calculated and its 95% confidence interval around the mean for 64 scenarios for train data set and its box plot

Figure 9: RMSEs calculated and its 95% confidence interval around the mean for 729 scenarios for train data set and its box plot

As it is shown, the very small RMSE means that the ANN model is trained base on data set. Thereafter,
the trained data is able to determine the location of damages in columns and check whether the structure
is reliable or not. In this regard, 10 different random data are considered as a test data. The performance
of ANN model is checked for these data sets in Figure 9 and Figure 10. Moreover, the final RMSEs for
the whole of trained and test data for 64 and 729 scenarios are presented in Table 2.

7
 Figure 10: RMSEs calculated and its 95% confidence interval around the mean for 64 scenarios for random data set

Figure 11: RMSEs calculated and its 95% confidence interval around the mean for 729 scenarios for random data set

Table 2: Final RMSEs for train and random data sets for each scenario

64 Scenarios 729 Scenarios

Story RMSE RMSE_TEST RMSE RMSE_TEST
1 0.0059 0.0106 0.01820 0.014924
2 0.0033 0.0079 0.016804 0.011904
3 0.0051 0.0182 0.01792 0.018246

With reference to the plots and final results of two different scenarios, it can be perceived that there is not
much difference between these data scenarios. In other words, the trained ANN model does not care about
the increment of input data. This is one of the most important benefits of ANN methods that can reduce
the computational effort without any significant decrease in accuracy. This feature is helpful for structural
damage detection as well as seismic reliability.

3.2. Linking ANN model to seismic reliability assessment

Once the ANN model is created, it will be used as a reliability method. Five random variables are
considered for the purposed steel structure; namely, columns areas of the first, second and third floor,
elasticity modulus, and gravity loads. The distribution, lower and upper bound, of each random variable,
are selected based on Ref. [23, 24] and are shown in Table 3. All stories have the same distribution

8
parameters, but this does not necessarily mean that all structures have the same columns sections.
Generally speaking, the procedure in the reliability assessment is the same as damage detection with this
difference that the inputs and outputs of the ANN model should be the same as OpenSees. Subsequently,
the ANN model has five PEs related to each random variables in the input layer and three PEs
corresponding to the maximum relative displacement of each story in the output layer. This procedure is,
exactly the same as before, performed through the interaction between MATLAB and OpenSeesa
network and a finite element model, Figure 11.

Figure 12: ANN training for Seismic Reliability assessment through the interaction between Matlab and OpenSees

Table 3: Statistical Distribution and Moments of Random Variables

Random Variables Distribution Mean Standard Deviation

Column Sections Normal 25 cm2 2.5 cm2
Elasticity Modulus Normal 2x1011 kN/m2 2x109 kN/m2
Gravity Loads Normal 50 kN 10 kN

The failure probability of each iteration is obtained by MCS, Table3. The failure probability is the
likelihood of passing through the Immediate Occupancy limit state (IO), which is equal to the maximum
relative displacement of 1%. This limit state and relative displacement value can be varied; In this study,
however, according to FEMA 350[25] it is considered 1%, and there are considerable other studies that
assume this value for the drift limit state [26], [27]. Although MCS can calculate the failure probability
readily, it can achieve to acceptable accuracy in a high number of iterations. In other words, the failure
probability in the high number of iterations is time-consuming and sometimes MCS is not applicable for
complex structures, such as airplanes, helicopters, bridges, and so forth. Therefore, in these cases, ANN
models can help MCS have a decent accuracy in the failure probability estimation. With reference to
Table 4, the ANN model is trained by MCS. In fact, a certain number1000 in this studytraining pairs
is needed to train the network. The training data can be obtained by MCS with a certain number of
iteration. Afterward, the ANN learns this process and try to imitate the relationship between input and
output data. Therefore, a combination of ANN and MCS with only a certain iteration can take the place of
MCS in a large number of iterations and address the problem of time-consuming in MCS although this
replacement has an approximation. Figure 13 depicts the schematic view of this ANN.

9
 Figure 13 Arrangement of inputs and outputs for calculation of probability of failure

In this study the number of training pairs is 1000; thus, it is called ANN1000. The ANN accuracy is
acceptable after 1000 iterations (ANN1000 model). The failure probability for each iteration is calculated
by the ANN1000 model. Owing to RMSE, the ANN1000 model works accurately. This model can be
applied to calculate the failure probability for any arbitrary iteration number that MCS could not be
applied (for instance, the failure probability for 10000 iterations in this case).

Table 3: The mean failure probability based on MCS

No. Run 10 50 100 200 500 1000 1500 2000 4000

No. Failure 1 11 15 39 87 170 264 352 686
Time(min) 4.14 20.67 41.44 82 204 453 616 862 1806
Pf 0.100 0.220 0.150 0.195 0.174 0.170 0.176 0.176 0.1715

Table 4: The mean failure probability based on ANN

No. Run 10 50 100 200 500 1000 1500 2000 4000 10000
No. Failure 2 13 16 40 88 171 262 351 689 1745
Pf (ANN) 0.2 0.26 0.16 0.2 0.176 0.171 0.174 0.175 0.1725 0.1745
RMSE 0.071 0.0115 0.0054 0.0016 0.0002 7.3E-5 8.4E-5 2.4E-5 4.6E-5

4. Conclusions
In this study, an inverse problem approach is adopted by using the artificial neural network to
determine the quantification and location of damages in the columns. In addition to damage
detection, the ANN model can evaluate the failure probability after being trained. In fact, in
many cases the relationship between input and output of the system is either complicated or

10
implicit while MCS needs a simple or explicit formulation; otherwise, it might be very time-
consuming. In this regard, apart from many methods to estimate the limit state function, using
ANN can be both fast and precise. In other words, the ANN model learns and imitates the
relationship between inputs and outputs of the system in order to reduce the computational effort
required for reliability analysis of complex structures in MCS method. It has shown that the
ANN model, which is trained after only1000 iterations can calculate the failure probability for
any arbitrary iteration number. Although in this research shows just 1000 iterations are sufficient
for the purposed frame structure, this value may increase according to the complexity of the
problem and required accuracy, and this limit can be determined by assorted methods such as
RMSE. This approach can be applied to the realistic models with implicit limit state functions.
Furthermore, the application of this approach for large structures and infrastructures reduces the
time and computational efforts.

References

[1] [1] Abdollahzadeh G.R., Faghihmaleki H., Esmaili H. (2016). Comparing Hysteretic
Energy and inter-story drift in steel frames with V-shaped brace under near and far fault
earthquakes. Alexandria Engineering Journal. http://dx.doi.org/10.1016/j.aej.2016.09.015
[2] Abdollahzadeh G.R., Faghihmaleki H., (2017). A method to evaluate the risk-based
robustness index in blast-influenced structures. Earthquakes and Structures, 12(1): 47- 54.
[3] Khaloo, A. R., et al. "Influence of diaphragm opening on seismic response of rectangular
RC buildings with end shear walls." Scientia Iranica. Transaction A, Civil Engineering
23.4 (2016): 1689.
[4] M. I. Friswell and J. E. T. Penny, A simple nonlinear model of a cracked beam, in
PROCEEDINGS OF THE INTERNATIONAL MODAL ANALYSIS CONFERENCE, 1992, p. 516.
[5] O. S. Salawu, Detection of structural damage through changes in frequency: a review, Eng.
Struct., vol. 19, no. 9, pp. 718723, 1997.
[6] E. zkaya and H. R. z, Determination of natural frequencies and stability regions of axially
moving beams using artificial neural networks method, J. Sound Vib., vol. 252, no. 4, pp. 782
789, 2002.
[7] S. Suresh, S. N. Omkar, R. Ganguli, and V. Mani, Identification of crack location and depth in a
cantilever beam using a modular neural network approach, Smart Mater. Struct., vol. 13, no. 4,
pp. 907915, Aug. 2004.
[8] A. Bakhshi and S. M. Vazirizade, Structural Health Monitoring by Using Artificial Neural
Networks, in 7th International Conference of Seismology and Earthquake Engineering (SEE7),
2015.
[9] A. Khaloo, S. Nozhati, H. Masoomi, and H. Faghihmaleki, Influence of earthquake record
truncation on fragility curves of RC frames with different damage indices, J. Build. Eng., vol. 7,
pp. 2330, 2016.
[10] H. M. Gomes and A. M. Awruch, Comparison of response surface and neural network with other

11
 methods for structural reliability analysis, Struct. Saf., vol. 26, no. 1, pp. 4967, 2004.
[11] C. Bucher and T. Most, A comparison of approximate response functions in structural reliability
analysis, Probabilistic Eng. Mech., vol. 23, no. 2, pp. 154163, 2008.
[12] A. H. Elhewy, E. Mesbahi, and Y. Pu, Reliability analysis of structures using neural network
method, Probabilistic Eng. Mech., vol. 21, no. 1, pp. 4453, 2006.
[13] X. L. Guan and R. E. Melchers, Effect of response surface parameter variation on structural
reliability estimates, Struct. Saf., vol. 23, no. 4, pp. 429444, 2001.
[14] J. Cheng and R.-C. Xiao, Serviceability reliability analysis of cable-stayed bridges, Struct. Eng.
Mech., vol. 20, no. 6, pp. 609630, 2005.
[15] J. Zhang and R. O. Foschi, Performance-based design and seismic reliability analysis using
designed experiments and neural networks, Probabilistic Eng. Mech., vol. 19, no. 3, pp. 259267,
2004.
[16] T. Most, Approximation of complex nonlinear functions by means of neural networks,
Weimarer () ptimierungs-und Stochastiktage, vol. 2, pp. 117, 2005.
[17] M. T. Hagan, H. B. Demuth, and M. H. Beale, Neural network design. Pws Pub. Boston, 1996.
[18] H. Demuth and M. Beale, Neural network toolbox for use with MATLAB, 1993.
[19] S. S. Haykin, Neural networks and learning machines, vol. 3. New York: Prentice Hall/Pearson,
2009.
[20] C. M. Bishop, Neural Networks for Pattern Recognition. Oxford University Press, Inc., 1995.
[21] S. Mazzoni, F. McKenna, M. H. Scott, and G. L. Fenves, OpenSees command language manual,
Pacific Earthq. Eng. Res. Cent., 2006.
[22] FEMA 440 (2005). Improvement of Nonlinear Static Seismic Analysis Procedures, Federal
Emergency Management Agency, Washington, DC, USA.
[23] S. Mahadevan and A. Haldar, Probability, reliability and statistical method in engineering design.
John Wiley & Sons, 2000.
[24] D. G. Lu, P. Y. Song, and X. H. Yu, Analysis of global reliability of structures, in Safety,
Reliability, Risk and Life-Cycle Performance of Structures and Infrastructures, CRC Press, 2014,
pp. 11851190.
[25] FEMA-350, Recommended Seismic Design Criteria for New Steel Moment Frame Buildings,
Washington DC, USA, 2000.
[26] D. Vamvatsikos, F. Jalayer, and C. A. Cornell, Application of incremental dynamic analysis to an
RC-structure, Proc. FIB Symp. Concr. Struct. Seism. Reg. Athens, 2003.
[27] J. Huh, Dynamic reliability analysis for nonlinear structures using stochastic finite element
method. The University of Arizona.

12
Highlights
Apart from Damage detection, the failure probability of a structure is calculated by
artificial neural networks.
Artificial neural networks are used to reduce computational efforts.
By defining the acceptable error value and index, artificial neural networks enhance
Monte Carlo Simulation technique.

13