Вы находитесь на странице: 1из 8

9th International Congress on Civil Engineering, May 8-10, 2012

Isfahan University of Technology (IUT), Isfahan, Iran

Approximation Frequency of Dominant Modes in Double Layer


Grids by Using Fuzzy Logic and Genetic Programming

Mohamad Hadi Bagherinejad1, Javad Salajegheh2


1- Post Graduate, Department of Civil Engineering, Shahid Bahonar University of Kerman
2- Associate Professor, Department of Civil Engineering, Shahid Bahonar University of Kerman
mhb.civil@gmail.com
Abstract
Dynamic analysis of space structures is very complex and time consuming due to high degrees of
freedom and the large number of elements. In conventional buildings, the dominant modes appear in the
first modes of vibration, while in space structures they may be appeared in the higher modes so, in these
types of structures the dominant modes cannot be achieved easily. In this article the frequency of
dominant modes of double layer grids is approximated based on the percentage mass participation by
means of Fuzzy Logic and Genetic Programming. For this purpose, the frequency variation of dominant
modes in double layer grids induced by the effects of changes in the ratios of height to span length and
variations of gravity loads has been studied. In continue, the frequency of dominant modes is estimated
by means of fuzzy logic and genetic programming. Finally results of both systems have been compared.
Keywords: Double Layer Grids, Dominant Modes, Fuzzy Logic, Genetic Programming.

1.

Introduction

In the past two decades, modal analysis has become a major method in the quest for determining dynamic
characteristics of engineering structures. Modal analysis is used to determine the frequencies and vibration
modes of a structure. These characteristics are use full to recognize the behavior of the structure. The
participating mass ratio for a mode provides a measure of how important the mode is for computing the
response to the acceleration loads. In modal analysis, the dominant modes have more participating mass ratio
than other modes. Therefore, obtaining the dominant modes is useful for determining the accuracy of
response spectrum analysis and seismic time- history analyses because the dominant modes have the most
participation in the dynamic response.
Dynamic analysis of space structures due to high degrees of freedom and the large number of
elements, is very complex and time consuming. These structures are used to covering large spans. Therefore,
the seismic analysis of these structures under the vertical component of earthquake is very important. This
issue is more important in regions close to the fault. In conventional buildings, the dominant modes appear in
the first modes of vibration, while in space structures they may appear in the higher modes. In addition,
solving the eigenproblems for the space structures is very time consuming due to the high degrees
of freedom. Therefore, using approximation systems that would estimation frequency of the dominant modes
without solving the eigenproblems is useful. Recently, new systems for the approximation of different
problems have been created using the training data. During the last few years, a number of researchers have
employed some techniques to determine the frequencies and mode shapes of structures. Wasfy and Noor [1]
use the fuzzy set to transient analysis of space structures. Chrysanthakopoulos et al. [2] proposed a formula to
approximate natural periods of plane steel frames. Gao [3] proposed a new method called the interval factor
method (IFM) for the natural frequency and mode shape analysis of truss structures with interval parameters.
But so far, to achieve frequency of dominant modes in structures has not been investigated.
Fuzzy logic began with the 1965 proposal of fuzzy set theory by Lotfi Zadeh [4]. Though fuzzy logic
has been applied to many fields, from control theory to artificial intelligence, it still remains controversial
among most statisticians, who prefer Bayesian logic, and some control engineers, who prefer traditional twovalued logic. Adaptive Neuro Fuzzy Inference System (ANFIS) is a fuzzy mapping algorithm that is based
on Tagaki fuzzy inference system. ANFIS is integration of neural networks and fuzzy logic and have the
potential to capture the benefits of both these fields in a single framework. ANFIS utilizes linguistic
information from the fuzzy logic as well learning capability of an ANN for automatic fuzzy if-then rule
generation and parameter [5].

9th International Congress on Civil Engineering, May 8-10, 2012


Isfahan University of Technology (IUT), Isfahan, Iran

In artificial intelligence, genetic programming (GP) is an evolutionary algorithm based methodology inspired
by biological evolution to find computer programs that perform a user-defined task. It is a specialization of
genetic algorithms (GA) where each individual is a computer program. It is a machine learning technique
used to optimize a population of computer programs according to a fitness landscape determined by a
program's ability to perform a given computational task.

2.

Adaptive Neuro Fuzzy Inference System

Figure 1(a) illustrates the reasoning mechanism for the Sugeno model discussed above while the
corresponding Adaptive Neuro Fuzzy Inference System (ANFIS) architecture is as shown in Figure 1(b),
where nodes of the same layer have similar functions [6]. The output of ith node in layer l is denoted as Ol,i.
L ay e r 1

L ay e r 2

L ay e r 3

La yer 4
ln 1

Rule1 : if ln1is A1 and ln2 is B1 then f12 p11ln1 q11 ln2 r11

A1
A2

Rule 2 : if ln1is A1 and ln2 is B2 then f12 p12ln1 q12 ln2 r12

ij

11

12

1 1 f 11

12 f 12

2 1 f 21

22 f 22

ln 1

Rule3 : if ln1is A2 and ln2 is B1 then f 21 p21ln1 q21 ln2 r21


Rule 4 : if ln1is A2 and ln2 is B2 then f22 p22ln1 q22 ln2 r22

ln 1

ln 2

B1

B2

(a)

21

22

La yer 5

ln 2

ln 2

O ut

(b)

Figure 1: (a) A two-input first-order Sugeno fuzzy model with four rules; (b) Equivalent ANFIS architecture.

Layer 1: Every node i in this layer is an adaptive node with a node function.
O 1Ai Ai ln1 ,i 1,2 , OB1 j B j ln2 , j 1, 2

(1)

Where lni is the input to node i and Ai is a linguistic label (small or large) associated with the node. Here
the membership function for A can be any parameterized membership function. In this paper, generalized
Gaussian membership function is taken as follows:

0.5 ln a 2
(ln,a,b ) exp

b2

(2)

Where {a, b} is the parameter set. These are called premise parameters.
Layer 2: Every node in this layer is a fixed node labeled , whose output is the product of all the incoming
signals.

Oij2 Wij Ai ln1 Bj ln2 ,ij 1,2

(3)

Layer 3: Here, the ith node calculates the ratio of the ith rules firing strength to the sum of all rules firing
strengths.
Wij
Oij3 W
,i, j 1, 2
2
2
(4)
i 1 j 1Wij
Layer 4: Every node i in this layer is an adaptive node with a node function.

Oij4 W ij f ij W ij pij ln1 qij ln2 rij ,i, j 1, 2

(5)

Where W ij is a normalized firing strength from layer 3 and {pij, qij , rij} is the parameter set of the node.
These parameters are referred to as consequent parameters.
Layer 5: The single node in this layer is a fixed node labeled , which computes the overall output as
the summation of all incoming signals [7]:

9th International Congress on Civil Engineering, May 8-10, 2012


Isfahan University of Technology (IUT), Isfahan, Iran

Out Oij5 i 1 j 1W ij f ij i 1 j 1W ij pij ln1 qij ln2 rij

3.

(6)

Determine the Effective Inputs

The values of the some structural features of the generated structures and frequency of dominant modes are
treated as inputs and outputs of the ANFIS, respectively. Then, an exhaustive search is performed within the
involved inputs to select the set of inputs that have the most influence on the output. Essentially, exhaustive
search technique builds an ANFIS model for each combination of input vector components and trains it for a
little epoch and reports the performance achieved [8]. The steps of exhaustive search algorithm to determine
the ni influential inputs from nm candidates is shown in below.
1. The number of influential inputs is initialized ni=1
2. For each combination of ni variables from nd candidates one ANFIS model is built, this leads to:
nd
nd !
na
ni ni ! nd ni !

(7)

3. The RMSE for training and testing sets are calculated as:
RMSEt

1
mt

mt

ac

i 1

pri

RMSE c

1
mc

mc

ac
i 1

pri

(8)

4. Find the combination with minimal training and testing errors.


5. Check the convergence.
6. Save the optimal combination of ni input from nd candidates.

4.

Fuzzy Clustering

Clustering of numerical data forms the basis of many classification and system modeling algorithms. The
purpose of clustering is to identify natural groupings of data from a large data set to produce a concise
representation of a system's behavior.
First we consider each data point as a potential cluster center and define a measure of the potential of data
point xi as [9]:
n

Pi e

xi x j

, = 4 ra

(9)

j 1

ra is a positive constant. We then revise the potential of each data point xi by the formula:

Pi Pi Pk* e

x i x *k

, = 4 rb
*

(10)
*

Where rb is a positive constant xk is the location of the kth cluster center and Pk is its potential value.
In Yager and Filevs procedure, the process of acquiring new cluster center and revising potential repeats
*
up *
until Pk P1 . We have therefore developed additional criteria for accepting or rejecting cluster center.
We use the following criteria:
*
up *
*
1. If Pk P1 ; accept xk as a cluster center and continue.
*
down *
*
2. Else if Pk P1 ; reject xk and end the clustering process.

3. Else; let
4. If

dmin =shortest distance between x*k and all previously found cluster center.

d min Pk*
* 1 ; accept x*k as a cluster center and continue.
ra
P1

9th International Congress on Civil Engineering, May 8-10, 2012


Isfahan University of Technology (IUT), Isfahan, Iran

*
*
5. Else; Reject xk and set the potential at xk to 0.
*

6. Select the data point with the next highest potential as the new xk and re-test.
Here specifies a threshold for the potential above which we will definitely accept the data point as a
cluster center; specifies a threshold below which we will definitely reject he data point. We use
down 0.15 and up 0.5 .

5.

Genetic Programming

Genetic programming (GP) creates computer programs in the lisp or scheme computer languages as the
solution. Genetic algorithms create a string of numbers that represent the solution. Genetic programming uses
four steps to solve problems:
1. Generate an initial population of random compositions of the functions and terminals of the problem
(computer programs).
2. Execute each program in the population and assign it a fitness value according to how well it solves
the problem.
3. Create a new population of computer programs.
i)
Copy the best existing programs
ii) Create new computer programs by mutation.
iii) Create new computer programs by crossover (sexual reproduction).
4. The best computer program that appeared in any generation, the best-so-far solution, is designated
as the result of genetic programming [10].

6.

Multigene Symbolic Regression

In contrast, in multigene symbolic regression each symbolic model (and each member of the GP population)
is a weighted linear combination of the outputs from a number of GP trees, where each tree may be
considered to be a gene. For example, the multigene model shown in Figure 2 predicts an output variable
using input variables x1, x2 and x3 [11].
+

+
tanh

*
0.41

X1

X2

X3

0.45

*
X2

sqrt

X3

y = d0 + d1 0.41x1 + tanh x2 x3 + d 2 0.45x3 + sqrt x2

Figure 2. Multigene model of a tree

This model structure contains non-linear terms (e.g. the hyperbolic tangent) but is linear in the parameters
with respect to the coefficients d0, d1 and d2. In practice, the user specifies the maximum number of genes
Gmax a model is allowed to have and the maximum tree depth Dmax any gene may have and therefore can
exert control over the maximum complexity of the evolved models. In particular, we have found that
enforcing stringent tree depth restrictions (i.e. maximum depths of 4 or 5 nodes) often allows the evolution of
relatively compact models that are linear combinations of For each model, the linear coefficients are
estimated from the training data using ordinary least squares techniques. Hence, multigene GP combines the
power of classical linear regression with the ability to capture non-linear behavior without needing to prespecify the structure of the non-linear model [12].

7.

Approximation of Multiple Output Using ANFIS and GP

ANFIS and GP are only able to approximate one output. Training in neural networks for multi-output
approximation is performed simultaneously and parallel. In this paper is presented a method to approximate
multiple output using ANFIS and GP. For this propose should be converted simultaneously and parallel

9th International Congress on Civil Engineering, May 8-10, 2012


Isfahan University of Technology (IUT), Isfahan, Iran

training to none simultaneously and series training. This means that train the data using first output. Then the
output from the previous Training is selected as an additional input for next training (Figure 3). This method
is useful when the outputs have same dimensions.
O utpu1

Outpu2

Outpu3

Output n

Input5
Input4
Input3
Input2
Input1

Input5
Input4
Input3
Input2
Input1

Input5
Input4
Input3
Input2
Input1

O utpu3
O utpu2
Outpu1
Input5
Input4
Input3
Input2
Input1

Outpu3
Outpu2
O utpu1
...

Training n

Outpu2
O utpu1

Training 3

Outpu1

Training 2

Training 1

Outpu n-1

Input5
Input4
Input3
Input2
Input1

Figure 3. Multi output model for ANFIS and GP

8.

Approximation Frequency of Dominant Modes

In this paper, one of the most popular models of double layer grids has been studied (figure 4). In these
model variations of gravity loads, Variations in the ratio of height to span length and changes in the number
of internal spans have been applied while distance between the columns is fixed. Then the frequencies of first
three dominant modes are determined for each changing. For the above changes, the following steps have
been taken.
For internal spans =4 to 9 , step=1
For height between two layers=60cm to 180cm , step=10cm
For Gravity load on all upper nodes =5000kg to 15000kg , step=100kg
Structural analysis and design to be done and determinate first three dominant modes based
on the percentage mass participation
End
End
End

180.00

150.00

900.00

1080.00

900.00

900.00

900.00

900.00

internal span = 4

internal span = 6

internal span = 5
1012.50

1000.00

128.57

100.00

900.00

1012.50

900.00
112.50

900.00

1028.57

1028.57

1000.00

1125.00

225.00

900.00

1050.00

1080.00

1050.00

1125.00

900.00

900.00

900.00

internal span = 7

internal span = 8
Figure 4. Double layer grids plan

internal span = 9

9th International Congress on Civil Engineering, May 8-10, 2012


Isfahan University of Technology (IUT), Isfahan, Iran

Now input data for training is determined. For this purpose, six non-geometric and geometric parameters as
the primary input is selected which is as follows:

Potential energy strain of structures based on allowed stress criteria: tp1 : E1s
Potential energy strain of structures based on existing axial force:
tp 2 : E 2s

tp3 : E1s E2s

Mass structural: tp4 : M

Height to upper span ratio (Figure 5): tp5 : h / L


Internal span to total span ratio (Figure 5): tp6 : l / L

Where
ne

E s Eie Eie iT i dVi e i=1,2,...,ne


i 1

(11)

Vi

Then Potential energy strain of structures based on existing axial force computed as follows:

ne
ne P 2
i
P
2
; i i ; Vi Ai Li E2s i Ai Li i Li
E
Ai
i 1 E
i 1 EAi

(12)

Where Pi , Ai and Li are axial force, cross section and length of the ith element respectively. Potential energy
strain of structures based on allowed stress criteria computed as follows:

F orF 2

E Pi Ti Ai Li

E
i 1

ne

(13)

s
1

2
2


1 1

where Fti 0.6Fy , Fpi


1 Fy , F.S 1.67 0.375 0.125
F.S 2 Cc
Cc
Cc

Ki Li
2 2 E
, Cc
, Fy 2400 kg cm2 , E=2.1 106 kg cm2
ri
Fy

(14)

(15)

Where Fpi, Fti, Ki, ri are the allowable stress in compression , allowable stress in tension, effective length
factor, radius of gyration of ith element respectively. E is modules of elasticity and Fy is yield stress.

L
h
l

Figure 5

Now we want using fuzzy logic and genetic programming make mapping between inputs and output.
For mapping using fuzzy logic, we must first determine the effective inputs with using exhaustive search
algorithm. For this purpose, one Hundred data has been randomly selected.
Table 1. Results of exhaustive search algorithm
ni

Optimal Combination

Na

{tp5)

2
3
4
5

RMSE

Training
1.05

Testing
1.12

{tp3 tp5}
{tp4 tp5 tp6}
{ tp3 tp4 tp5 tp6 }

15
0
5

0.32
0.09
0.07

0.33
0.17
0.11

{ tp2 tp3 tp4 tp5 tp6 }

480

720

9th International Congress on Civil Engineering, May 8-10, 2012


Isfahan University of Technology (IUT), Isfahan, Iran

So combined inputs of the fourth row as effective inputs are selected. Using effective inputs and fuzzy
clustering the Data are prepared for training with ANFIS. In fuzzy clustering the parameters are as follows:
down 0.15 , up 0.5 , ra 0.7 & rb 1.25ra
For approximation frequency of the dominant modes, 5% of the data for training and 95% of data for
testing have been considered. Epoch training is equal to 100. The run took on a PC with core i5 processor
running at 2.40GHz with 3.79GB of RAM. Error curve of training and testing data to approximate the
frequency of the first three dominant modes are given in Figure 6.
0.3

RMSE

0.6

0.2

0.8
0.6

0.15

0.4

0.05

0
0

50

0.4

Training Data
Testing Data

0.2
0

100

Epoch

Training Data
Testing Data

0.1

0.2

RMSE

Training Data
Testing Data

0.8

1.2

0.25

RMSE

50

100

Epoch

50

100

Epoch

(a)
(b)
(c)
Figure 6. Error curve of frequency in ANFIS. (a) first mode, (b) second mode, (c) third mode

For approximation frequency of the dominant modes using genetic programming, we dont require
determining the effective inputs. Genetic programming automatically uses form effective inputs because of
evolutionary nature of the system.
Genetic programming run with the following settings was performed: Population size = 100, Number
of generations = 100, Tournament size = 2, Dmax = 4, Gmax 7, Elitism = 0.01 % of population, function
node set = {plus, minus, times, tanh, absolute square root, sine, cosine, exponential function}, ephemeral
random constants in the range [-40 40]. The multigene symbolic regression function was used in order to
minimize the root mean squared prediction error on the training data. The following recombination operator
event probabilities were used: Crossover events = 0.85, mutation events = 0.1, direct reproduction = 0.05.
The following sub-event probabilities were used: high level crossover = 0.2, low level crossover = 0.8,
subtree mutation = 0.9, replace input terminal with another random terminal = 0.05, Gaussian perturbation of
randomly selected constant = 0.05 (with standard deviation of Gaussian = 0.1). 5% of the data for training
and 95% of data for testing have been considered. Total number of data = 8484, epoch training = 100. The
run took on a PC with core i5 processor running at 2.40GHz with 3.79GB of RAM. Error curve of training
and testing data to approximate the frequency of the first three dominant modes using GP are given in
Figure 7:
5
4

Training Data
Testing Data

Testing Data

10
8
6
4
2
0

1
0

0
0

50

100

20

40

Epoch

60

Training Data
Testing Data

RMSE

Training Data

RMSE

RMSE

80

100

Epoch

50
Epoch

100

(a)
(b)
(c)
Figure 7. Error curve of frequency in GP. (a) first mode, (b) second mode, (c) third mode

Formulation the frequency of dominant modes by GP:


1 7.427

7.516exp sin tp 6
log tp 5

4.218 tp

tp 6

35.29 cos log log tp 5

tp
tp
8905 sin
2256
tp
4

5
4

tp
6.052 sin tp 3 0.8445 sin log 6

tp1

2 10.37 12.74tp 4tp 5 sin tp 6 58 .59tp 6 0.04102tp 612 345 .2tp 5 sin tp 6 114.8tp 5 15.49tp 4tp 5tp 6
20188tanh tp
5
3 158.3
7.9632 527

2 tp 5

0.1352coth tp 6

tp 4

tp 6

tp 5

0.5

0.03328

2 tp 4
tanh tp 6

336.4

tp 6

0.5

35.37 tanh tp 5
sin tp 6

9th International Congress on Civil Engineering, May 8-10, 2012


Isfahan University of Technology (IUT), Isfahan, Iran

Here, ANFIS and GP are compared in terms of time training and the error value. Comparative results are
given in table 2.
Table 2. Compare ANFIS and GP
Time (sec)

Training RMSE

Testing RMSE

Frequency of:

ANFIS

GP

ANFIS

GP

ANFIS

GP

ANFIS

GP

ANFIS

GP

First dominant mode

5.48

16.6

0.09

0.14

0.12

0.36

3.04%

3.31%

0%

0%

Second dominant mode

7.27

17.4

0.18

0.27

0.213

0.55

2.76%

3.68%

1E-3%

1E-4%

Third dominant mode

10.39

17.7

0.55

1.37

0.735

1.94

9.45%

11%

3E-3%

1E-3%

9.

Maximum error

Minimum error

Conclusions

As shown in this paper, the frequency of the dominant modes in double layer grids can be predicted using the
approximation systems. In eigenvalue problems, the equation must be solved to determine the dominant
modes but in this type of systems based on the apparent structural features, the frequencies of dominant
modes are approximated directly. Hence using these systems for space structures is very useful and efficient.
Because solving the eigenvalue problems for space structures is very time consuming due to high degrees of
freedom and the large number of elements. In addition to the dominant modes in these structures appear in
higher modes.
In this paper, ANFIS and GP have been used to predict the frequency of the dominant modes. Few
data have been used for training. Performance of ANFIS in training time, training error and testing error is
better than GP. In contrast GP dont require determining the effective inputs because of evolutionary nature.
Also the output of GP is a formula and this makes the system easier to use.

10.

Reference

1. Wasfy, T.M and Noor, A.K. (1998), Application of Fuzzy Sets to Transient Analysis of Space
Structures, Finite Elements in Analysis and Design 29, pp. 153-171.
2. Chrysanthakopoulos, C. Bazeos, N. Beskos, D.E. (2006), Approximate Formulae for Natural Periods
of Plane Steel Frames, Journal of Constructional Steel Research 62, pp. 592604.
3. Gao, W. (2006), Interval Natural Frequency and Mode Shape Analysis for Truss Structures with
Interval Parameters, Finite Elements in Analysis and Design 42, pp. 471-477.
4. Zadeh, L.A. (1965), Fuzzy sets, Information and Control 8 (3), pp. 338353.
5. Nedjah, Nadia, ed. Studies in Fuzziness and Soft Computing, Germany: Springer Verlag, pp. 53-83.
6. Ross, T. (1997), Fuzzy Logic with Engineering Applications, McGraw-Hill International Editions,
Electrical Engineering Series.
7. Jang, Jsr. (1993), ANFIS: Adaptive-Network-Based Fuzzy Inference Systems, IEEE Transactions on
Systems Man and Cybernetics, 23, pp. 665685.
8. Seyedpoor, S.M. Salajegheh, J. Salajegheh, E. Gholizadeh, S. (2009), Optimum Shape Design of Arch
Dams for Earthquake Loading Using Fuzzy Inference System and Wavelet Neural Networks,
Engineering Optimization, 41, pp. 473-479.
9. Chiu, S.L. (1994), Fuzzy Model Identification Based Pn Cluster Estimation, Journal of Intelligent and
Fuzzy Systems, 2, pp. 267278.
10. Koza, John R. (1992), Genetic Programming: On the Programming of Computers by Means of Natural
Selection, Cambridge, MA: The MIT Press.
11. Searson, D.P. Leahy, D.E. Willis, M.J. (2010), GPTIPS: An Open Source Genetic Programming
Toolbox for Multigene Symbolic Regression, Proceedings of the International Multi Conference of
Engineers and Computer Scientists, Vol I, IMECS, March 17-19, Hong Kong.
12. Hinchliffe, M.P. Willis, MJ. Hiden, H. Tham, M.T. McKay, B. & Barton, GW. (1996), Modeling
chemical process systems using a multi-gene genetic programming algorithm, In Genetic
Programming: Proceedings of the First Annual Conference, pp. 56-65. The MIT Press, USA.

Вам также может понравиться