You are on page 1of 6

DEVELOPMENT OF ARTIFICIAL NEURAL NETWORKS BASED

HYSTERETIC MODEL
Waon-Ho Yi
Dr. Eng., Professor, Dept. of Arch. Engrg., Kwangwoon Univ., Seoul, Korea
whyi@daisy.kwangwoon.ac.kr
Hack-Soo Lee
Dr. Eng., Professor, Dept. of Civil & Env. Engrg., Hannam Univ., Taejeon, Korea
leehs@eve.hannam.ac.kr
Seung-Chang Lee
Dr. Eng., Manager, Hyundai Development Company, Seoul, Korea
sclee88@unitel.co.kr
Ho-Sung Kim
Graduate Student, Dept. of Arch. Engrg., Kwangwoon Univ., Seoul, Korea
hcastle1@hanmail.net
ABSTRACT
Artificial Neural Network(ANN) is a computational model inspired by the structure
and operations of human brains. It is a massively parallel system, consisting of a large
number of highly interconnected and simplified processing units. The purpose of this
paper is to verify the applicability of ANN to predict experimental results through the
use of measured experimental data. Although there have been accumulated data based
on hysteretic characteristics of structural element s under the cyclic loading, it is
difficult to directly apply them to the analysis of elastic and plastic response. Thus
simplified mathematical models, such as Bi-Linear Model, Ramberg-Osgood Model,
Degrading Tri Model, Takeda Model, Slip type Model, and etc, have been used. To
verify the practicality and capability of this study, ANN is adapted to several models
with mathematical formula using numerical data. To show the efficiency of ANN in
nonlinear analysis, it is important to determine the adequate input and output variables
of hysteretic models and to minimize error s in ANN process. For the verification,
beam-column joint test data were employed using the ANN in modeling of the linear
and nonlinear hysteretic behavior of structures.
KEYWORDS: neural networks, hysteretic behavior, beam-column joint test,
simulations, architecture, performance evaluation
INTRODUCTION
There are enormous experimental results or data involving the hysteretic
characteristics of structural element, which have been produced by cyclic load tests.
Since it is difficult to directly apply them in the analysis of the elastic and plastic
response, many researchers have proposed various hysteretic models with
mathematical formula, such as Bi-Linear model, Ramberg-Osgood model, Degrading
Tri model, Takeda model, Slip type model, and etc.
Soft computing with artificial neural networks as a parallel computational model is
similar to the human brains that can self -organize complex nonlinear relationships
without making assumptions. The concept that multi-layer feed-forward networks are
the class of universal approximators , can be introduced to the hysteretic modeling
because it deals with the problems involving incomplete or imprecise information.

This paper is composed of two parts. The first part of this paper presents the
architecture of neural networks in order to train the numerical data with hysteretic
characteristics. Various simulations are performed for determining the neural
networks architecture using the performance evaluation with training parameters.
Displacements as the input and loads as the output are considered in the architecture
of neural networks. The trained neural networks are recalled to predict the hysteretic
model with the different hysteretic rules, yield load, yield displacements, and etc.
The second part of this paper presents how to apply neural networks based model to
predict behaviors of the beam-column joint. In this part, neural networks predict loaddisplacement curve with the controlled displacements.
In this paper, we described the methodology to develop Neural Networks based
model for the prediction of hysteretic behaviors. The approach is verified by the
applied example, beam-column joint test, to predict its linear and nonlinear behaviors.
ARTIFICIAL NEURAL NETWORK (ANN)
Neural networks are massively parallel computational models for knowledge
representation and information processing. They are inspired by the neuronal
architectures and operations of human brains. Learning from examples is probably the
most important capability of neural networks. Neural networks are capable of learning
highly complicated nonlinear relationships and associations from a large body of data.
The information and knowledge obtained by the neural networks are encoded and
stored in the connection strengths of the networks. The retrieval of the stored
information is done routinely, providing the network with an input pattern, which acts
as a key. Neural networks are also very robust computational models; they are noise
and fault tolerant and they are truly adaptive system. The capabilities mentioned
above also make neural networks good candidates for the application to a wide range
of structural engineering problems, where the relationship within the data may be
complex and a sufficient amount of data is often not available or is difficult to obtain.
Whatever data may exist often contains some scatters.[1][2]
NEURAL NETWORKS-BASED HYSTERETIC MODEL (NNHM)
The basic scheme for developing a NNHM is to train a multi-layer feed-forward
neural networks using a back-propagation algorithm, which is the one of learning
algorithms on the result of mathematical hysteretic models.
Hysteretic Model
To examine the applicability of neural networks, the hysteretic models, such as BiLinear and Ramberg-Osgood models are employed in this paper. The properties and
types of each model are shown in Figure 1. The Bi-linear Model is described as lines
with Ky, stiffness in the elastic area and Km, stiffness in the inelastic area. If the
stiffness in the elastic area, Km is zero, it will be a perfect elastic-plastic model. This
model is used in this study. The Ramberg-Osgood model, proposed for the metal
material behaviors is composed of the yield strength, yield displacement and other
parameters. Since the Ramberg-Osgood model with a curve type can be used for the
hysteresis behaviors, such as steel and/or reinforced concrete structure behaviors, this
model with Bi-Linear model have the probability and applicability to construct neural
networks based hysteretic model.

P
d
p
p
= i = 1+
dy py
py

py
Km
Ky

py

d d i p pi
p pi
=
=1 +
2d y
2 py
2py

dy

Bi-linear model

Ramberg-Osgood model

Figure 1. Mathematical Hysteretic Models


The Consistency and Learning of NNHM
The NNHM is a multi-layer feed-forward neural networks consisting of three
layers; input, output, and hidden layers. The NNHM receives five (in Bi-Linear
model) or the six (in Ramberg-Osgood model) normalized displacements according to
the input patterns (Figure 2). Numbers of nodes in the hidden layers are determined
during the training process using dynamic node creation procedures. The processing
units or nodes use the standard sigma units of which output functions are sigmoid
functions with the range of 1 and 1. Each layer is fully connected to its sources and
target layers. The Generalized Delta Rule with momentum is used as learning rule.
The initial values for the connection weights are assigned by the uniform random
number generator. In each training cycle, the connection weights are updated after all
training patterns are represented in the networks. [3]
Hidden Layer
Input Layer

Consist of 1~20 units


Output Layer

Displacement,
etc.

Loads

Figure 2. Architecture of Neural Networks Based Hysteretic Model (NNHM)


The input data used in NNHM are displacement , maximum experience
displacement max, displacement of previous point r , yield displacement y , yield load
P y and parameter (in Ramberg-Osgood model only) in 3 cycles of hysteresis loops.
The details of the each model are tabulated in Table 1. Normally the data is divided
into two sets, namely the training set and the testing set. As the names imply, the
neural networks are trained with the training data set and the performance of the
trained neural networks is evaluated using the test data. The training set contains 32
input-output pairs for the Bi-linear model and 52 input-output pairs for the RambergOsgood model. The trained neural networks are tested at each relative cycle,
respectively. During the training process, the number of hidden units evolves from 1

to 20 and the learning rate is gradually decreased from the starting value of 0.9. The
neural networks have 10~11 hidden nodes in a single hidden layer. In this study, we
divide two parts to find the prediction of neural networks through the different
hysteretic histories; these are Type A ( y =1.0, Py =1.5) and Type B ( y =1.0, Py =1.5).
The control error value was determined with 2.010-3 as an average error.
Table 1. Details of Network for Each Model
Bi-Linear Model

Network parameter
No. of INPUT units
No. of OUTPUT units
No. of HIDDEN units
Learning Rate
Activation Function
No. of Training Cycles

Type A
Type B
5
5
1
1
10
10
0.9
0.9
Sigmoid Function
1983
2273

Ramberg-Osgood Model
Type A
Type B
6
6
1
1
11
11
0.9
0.9
Sigmoid Function
2321
2547

SIMULATION RESULTS

Error(X10E-2)

Error(X10E-2)

Learning iterations were repeated until increasing errors became equal to the
average error. Errors in both groups of Type A and Type B became tolerant after the
2000 to 2500 learning iterations. To verify the suitable node numbers in hidden layers,
the number of used nodes were varied from 1 to 20 with the number of 1000 learning
iterations. Figure 3 and Figure 4 show the errors in B i-Linear model and RambergOsgood model according to the number of units in hidden layers, respectively. Since
the errors in both models can be minimized when the unit number of hidden layers is
10 as shown in Figure 3 and Figure 4, the unit of 10 for the hidden layers was
employed in this study.

The Unit Number of Hidden Layer

The Unit Number of Hidden Layer

Figure 3. Errors in Bi-Linear


Model

Figure 4. Errors in Ramberg-Osgood


Model

The simulated hysteretic loops using both models in cycle 1, 2, and 5 are compared
to those obtained by the mathematical models in Figure 5 and Figure 6, respectively.
The prediction of the trained neural networks in cycles 1 and 2 satisfies the required
desirable output results. However, it was not proper to predict the desired data in cycle
4 and cycle 5 for each hysteretic model.

1 cycle

2 cycle

5 cycle

Figure 5. Hysteresis Loops by Bi-Linear Model

1 cycle

2 cycle

5 cycle

Figure 6. Hysteresis Loops by Ramberg-Osgood Model


CASE STUDIES
To verify hysteretic models using neural networks proposed in this study, the
experimental data from the beam-column joint test[4][5] were adopted to predict the
hysteretic history and were compared to the predicted one.
Test Scheme
During the test, the specimens are held horizontally in a steel frame with pin
supports at near the end of the beam and column (Figure 7). The actuator is attached
to a steel frame prestressed to the structural floor of the laboratory. The constant axial
force corresponding to 0.2f' c A g (192 ton) is applied at the column prior to the
hysteretic beam loading.

Figure 7. Schematic drawing of


the testing frame

Figure 8. Loading history

Loads are applied using a 50ton servo-hydraulic actuator under the displacement
control. The beam is first loaded upward (positive bending) slightly beyond the yield
strain of the longitudinal reinforcement. The yield displacement, y is set when the
flexural reinforcement reaches its yield strain. The direction of loading is then
reversed. The loading history scheduled is shown in Fig ure 8. The test is proceeded
until the maximum cycle load, P m is 75% of its yield load P y, which is defined as the
collapse load.

Loads(ton)

Loads(ton)

Results of Analysis
The analysis results are shown in Figure 9 and Figure 10. These figures show the
comparison of experimental data and the prediction of a trained neural networks using
experimental data.

Displacement(mm)

Experience data

Experience data

The Prediction of
Neural Network

The Prediction of
Neural Network

Displacement(mm)

Figure 9. Comparison of experience data Figure 10. Comparison of experience data


and neural network data in cycle 1 to 4 and neural network data in cycle 1 to 8
CONCLUDING REMARKS
The methodology to develo p Artificial Neural Networks based hysteretic model for
the prediction of hysteretic behaviors is described. The training of the neural networks
can be accomplished with the input parameters suggested in this study. Also the
prediction of neural networks based hysteretic model is more efficient using the Bilinear model and Ramberg-Osgood model. Thus, it can be said that with suggested
input parameters, the neural networks based hysteretic model (NNHM) can be used as
a mathematical hysteretic model.
Based on the simulation results, it is shown that for the cycles 1 ~ 4 (elastic range)
NNHM can provide satisfactory results, but for cycles 4 ~ 8 (inelastic range) NNHM
represents unsatisfactory results. However, this can be improved by increasing the
numbers of training data.
REFERENCES
1.

J. Ghaboussi and M.R. Banan (1994), Neural Networks in Engineering


Diagnostics, Proceedings, Earthmoving Conference, Soc. of Automotive Engrs.
2. A. Joghataie and J. Ghaboussi, (1996), A Comparative Study of Learning
Methods and Mathematical Algorithms in Structural Control, Proceedings, 11th
World Conference on Earthquake Engineering, Acapulco, Mexico.
3. S.C. Lee, S.W. Han and B.H. Lee, (1999), Generation of Artificial Earthquake
Accelerograms and Response Spectra using Neural Networks , Singapore,
Fourth Asia-Pacific Conference on Computational Mechanics, pp.1335~1340.
4. Y.C. You, W.H. Yi and L.H. Lee, (1999) , Inelastic Analysis of RC BeamsColumn Subassemblages Under Various Loading Histories, Structural
Engineering and Mechanics, Vol. 7, No. 1, pp.69~ 80.
5. W.H. Yi and R. Peek, (1993). Posterior Timestep Adjustment In Pseudodynamic
Testing, Journal of Engineering Mechanics, ASCE, Vol. 119, No. 7, pp.1376~
1386.