CABLE TENSION CONTROL OF TSUNEYOSHI BRIDGE

USING MULTI-OBJECTIVE GENETIC ALGORITHM

Hitoshi Furuta ‘, Yukio Kawamura 2, Hideki Arimura 3, Kazuo Takase3,
‘Faculty of Informatics, Kansai University
2-l-l Ryozenjicho, Takatsuki City, Osaka 569-1095, Japan
2Public Works Bureau, Osaka City Government
l-2-2-500 Umeda, Kita-ku, Osaka City, Osaka 530-0001, Japan
Downloaded from ascelibrary.org by University College London on 05/30/16. Copyright ASCE. For personal use only; all rights reserved.

3Bridge Designing Department, Komai Tekko Inc.

2-5-l Nakajima, Nishiyodogawa-ku, Osaka City, Osaka 5554041, Japan

This paper describes the application of GA (Genetic Algorithm) to the shim adjustment in the
erection accuracy control for the Tsuneyoshi Bridge ,which is a cable-stayed bridge and
constructed in Osaka Bay, Japan. Some methods have such a problem that weghting for various
factors is difficult in a job site. Therefore, an attempt is made to present pareto-optimal
solutions and apply GA as an optimum shim adjustment method. It is noted that the proposed
method can provide us with discrete combination values of shim plates instead of continuous
values. Furthermore, it is expected that even inexperienced engineers can easily select an
appropriate solution suitable for individual situations without parametric examination of various
weighting factors, because the method can give a set of solutions (called parato-optimal
solution) with different characteristics

KJWWORDS

Genetic Algorithm, Multi-Objective Problem, Pareto-Optimal Solution, Shim Adjustment,

Ranking & Sharing

INTRODUCI'ION

In the case of cable system bridges and cable-stayed bridges in particular, prestressing is
effective to achieve mechanical rationality and to minimize self-weight of the steel girders at the
design stage. Because the tensions of the cables are subtly effected by the stressof the structure,
the shape and tension of the cables are measured and the cables length or tension are adjusted in
order that the stressstate and the shape at the design stage are as close as possible to the design
values.
Shim adjustment requires advanced erection management technology and considerable time.
And when actually performing shim adjustment, solutions appropriate to erection conditions
must be obtained rapidly. In sum, it is important to quickly and rationally determine the
quantity of shim adjustment.
This research project adopts GA (Genetic Algorithm) that is a typical AL (artificial life)
technology method in order to develop the optimum shim adjustment method. Characteristics of
GA include an easy acceptance of the basic concept and its ability to solve combinatorial
optimization problems I! These characteristics of GA were taken advantage of to calculate the
optimal solution based on three objective functions including the number of cable adjustments,

Copyright ASCE 2004 Structures 2000

Advanced Technology in Structural Engineering
 2

that has not been directly handled as an objective function in the past, by number of adjustments
as discrete combination values. Because it presents the final results by means of a pareto-
optimal solution that is a compound solution of the optimum solution, the comparison and
selection of solutions can be performed easily.
In order to further simplify and speed up the solution process,where the goal is the minimization
of the error of the tension and shape, a system was constructed that can be used to rationally
select the solution with the small set number of cables whose shims are to be adjusted.
Downloaded from ascelibrary.org by University College London on 05/30/16. Copyright ASCE. For personal use only; all rights reserved.

Furthermore, as an example of application to a real bridge, we present our analysis result report
regarding erection control at Tsuneyoshi Bridge.

RANKINGANDSHARING

When GA is applied to a multi-objective optimization problem, the key point is to correctly

evaluate and select the pareto-optimal solution and leave it for the next generation. In order to
optimize the fitness functions, the sharing method was introduced to the ranking method by
Goldberg and Richardson in 1987 *).
To perform the multi-objective optimization, it is better for the candidate solutions to be obtained
as a dispersed combination @areto-optimum solution combination) than a solution (individual)
concentrated at a certain location. And if a pareto-optimum solution combination is obtained,
the decision-maker uses only one solution among these by accounting for the trade-off. The
sharing method is explained below.
An appropriate classification is made with the i-axis as the standard and the number in the
population included in each class is considered to be n(i). Then, based on the fitness function
possessedby the population included in each class, the shared fitness function is obtained by
dividing the fitness of individuals included in each class by the value obtained by normalizing
the number of individuals in that class.
fit(i)= l (1)
m(i)
fit(i)
sharedfit( - (2)
num(i)

where ;j%(Q):fitness function , RANK(i): number of ranks

In the caseof three objective functions, the distribution of individuals is three-dimensional, but in
order to relate the degree of scattering of the individuals to the fitness functions, it is necessq to
represent it with one evaluation value as in the case of two functions. However, it is impossible
to evaluate the distribution of two axes with a single value. Then, as shown in Figure 1, the
classification is done by focusing on one of the three axes in order to calculate the number of
individuals in each class.

Figure 1: Classfication of Three Objective Functions

Copyright ASCE 2004 Structures 2000

Advanced Technology in Structural Engineering
 3

NUMERICALEXAMPLE

(1) Analysis Model

An analysis model is prepared for a cable-stayed bridge that is a 3-span continuous steel plate
deck box girder bridge shown in Figure 2. The structural type and dimension are given as;
Bridge length: 340.800 mm, Span: 24.000 + 65.000 t 248.850 m
Main tower: single reverse Y tower, Tower height: 80 m
Downloaded from ascelibrary.org by University College London on 05/30/16. Copyright ASCE. For personal use only; all rights reserved.

Main girder: steel plate deck box girder

Cable: fan type with two surfaces, six levels of cable (totally 24 cables)
TP

d
CA-6 1 CA-?

t
s

25800 65000 I 250 004 J

Figure 2 : Analysis Model

(2) Analysis Conditions
1) Cables adjusted; CAlL to CA12L, CAlR to CA12R for 24 cables
2) Shim tables adjusted; horn - 63 mm to 64 mm at 1 mm intervals, 128 types (7 bits)
3) Errors considered; Shape errors: Rf r i (main girder camber at 38 points, tower top
at 1 point), Tension errors: Rt , , (for 24 cables)
4) Initial errors; Maximum error of the shape: Rf - max = 115.8 mm,
Maximum error of the tension: Rt-max = 48.6 tf
5) Maximum repeated generations; 100 generations
6) Population size; 800 individuals
7) Other conditions; Crossover method: l-point crossover, crossover probability: 0.5,
mutation probability: 0.05
Figure 3 shows a distribution diagram of the initial error.
Population size is, in the case of a single objective problem, often set at about 10 times (240
individuals) of the design variables (number of cables: 24), but for this study, it was set at 800
individuals because that in a case where ranking is done to find the pareto-optimum solution,
more individuals are necessary at the trial calculation stage, to obtain a suitable solution. The
mutation probability ranges from 0.01 to 0.05.
160
120
error of shape
80 ,
Y- 40
‘1

E -4;
error of tention
-80
-120

Copyright ASCE 2004 Structures 2000

Advanced Technology in Structural Engineering
 4

(3) Jktmples with Two Objective Functions

The following is an example of the calculation of a problem with two objective functions for the
residual error of shape and tension. The fitness function considers the apportionment of the
error accounting for the stress balance of the total structural system, but this research project
focused on the absolute values of the maximum error of the shape and tension (Rt-max, Rf-max)
among the residual errors in tension and shape of each individual to perform ranking and sharing
using equations (1) to (2). Figure 4 shows the Rt-max/Rf-max distribution diagram for
Downloaded from ascelibrary.org by University College London on 05/30/16. Copyright ASCE. For personal use only; all rights reserved.

individuals after convergence. The solution obtained with the fuzzy regression method 3, is
shown as a solution for comparison purposes. In this example, even at the final generation, a
comparison of the convergence properties of the paretooptimal solution with those of the fuzzy
regression method reveals that the only solutions obtain able were those for which Rt-max, Rf-
max are not improved.
Increasing the population size is cited as one way to improve the precision of the solution. The
population size was set between 20,000 and 50,000 at the trial calculation stage,but this did not
improve the precision of the solution. Then, an attempt was made to improve the convergence
solution by focusing on the shape of the distribution of the first generation.
The initial error, quantity of shim adjustment, and the error after adjustment are represented by
the following equation through an influence matrix.
Ax+b, =b, (3)
where ;A: influence matrix , X: quantity of shim adjustment,
b o: initial error, b 1: error after adjustment
Nextx,,.<x,cxm,, was provided as constraints and the value of x was found to satisfy these
conditions in order to minimize b ,. Centered on x and based on the normal distribution, the
quantity of shim adjustment is produced to form the individual and to include only the needed
number in of individuals the initial generation. In this example, the total population of 800
individuals and 400 individuals were produced using the above method. Figure 5 shows the
calculation results. As these results indicate, the individuals are distributed near the final
solution even in the initial generation and the convergence of the pareto-optimal solution in the
final generation is improved,

400

350

300

0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Rt-max(tf) Rt-max(tfJ

Optimization
Figure 4 : F&m.lt.a of Tkvo Objective Figure 5 : Results of for a Case with
the improved Initial Generation

Copyright ASCE 2004 Structures 2000

Advanced Technology in Structural Engineering
 5

error of shape
? 70 I
2 60
fJ 50
EI 40

a! 30
20
10
Downloaded from ascelibrary.org by University College London on 05/30/16. Copyright ASCE. For personal use only; all rights reserved.

0
0 IO 20 30 40 50
Rt-max (K)

(a) Wmax/Rt-max and number of cables adjusted (b) hitial error

(c) 14 cables adjusted (d) 7 cables adjusted

Figure 6 : Error Distribution for Number of Cables Adjusted
For this research, one purpose in adopting the GA was to present a wide variety of diverse
solutions instead of a single solution. In this program, varying the random number causes
fluctuations in the convergence solution. So diverse results were obtained by concluding the
solutions at the 100th generation where the convergence peaked in order to perform the final
ranking based on 10 times of analysis results.
Figure 6 shows the results obtained under the conditions in (2) for the number of cables adjusted.
This analysis obtained pareto-optimal solutions from 24 to 7 cables,but in Figure6, only solution
candidates for 18, 14, 10, and 7 cables are presented for comparison purposes. Moreover,
Figure 6 (b) to (d) show the error distribution by the number of cables adjusted. The analysis
time in this example was 330 seconds(cpu:pentium MMX 233MIIz).
Judging fkom Figure 6, within a range of 18 to 7 cables adjusted, a comparison of 14 cable
adjustment case with 7 cable adjustment case reveals that in 7 cable adjustment case, the
maximum shape error is about 7 times that in 14 cable adjustment case. It is possible to judge
the boundary where the errors in shape or tension appear based on the number of cables adjusted
by performing the optimization for the three objective functions in this way. In other words, it is
easy to decide if it is possible to obtain results rationally and effectively depending on the
number of cables to be adjusted. In fact, the number to be adjusted is minimized and the
solution where the residual error satisfies the control values is selected.

APPLICATION TO THE ERECTION CONTROL OF TSUNEYOSJXI BRIDGE

(1) Outline of erection control
We executed erection control of this bridge with the following steps.
STEP 1 Adjustments at erection time
We conducted the measuring in this stage to confnm the measurements of the tension and shape
of the cables in each erection step. Gable erection was divided into 6 stages, which were

Copyright ASCE 2004 Structures 2000

Advanced Technology in Structural Engineering
 6

conducted in order starting with the first stage. Each stage was made up of one cycle of 4 cables.
In each stage of erection, we measured the tension and shape of all completely erected cables
and also measured the bridge temperature. We adjusted the shim plate value whenever
necessary.
STJZP2 Adjustments at final stage
We conducted the measurements in this stage as the final stage of tension adjustments to confirm
the adjustment results at erection time and to confirm the final tension and shape. Measurements
Downloaded from ascelibrary.org by University College London on 05/30/16. Copyright ASCE. For personal use only; all rights reserved.

were conducted in the same manner as in STEP 1.

Figure 7 shows the items we observed with erection control. Table 1 shows the target values for
control, which were derived from examinin g the results of each kind of error analysis and past
pelfOll?XlIl~.

Tower Section

Girder Section

Flgunz 7 Control items

Table 1 Contil target values
Control item Control target value
+ H/moo (mm> H: tower height (mm)
Tower collapse 6=+80,000/2000
=t4.omm
+ {25+025(L-50)) (mm) L span length (m)
Main girder camber P2-P3 : 6C=f {25+0.25(24.00-50))= fl9mm
P3-P4 : 6C=t {25tO.25(65.00-50)) = +29mm
P4-P5 : 6G-t {25+0.25(248.85-50)}= -t75mm
Cable tension (D+PS)XO.15 D: dead load tension
PS: prestress

(2) Erection control results

The results in Figures 8 and 9 show the shape and tension error at the time of final adjustments
of this bridge. The maximum error for each caseis as follows.
Before adjustment Maximum tension error (Rt-max) : -45.9%
Before adjustment Maximum shape error (Rf-max) : 39.6mm
We conducted 11 series of analyses at 100 generations per series. We finished analysis of the
whole cable at 382 seconds without limiting the number of shim adjustments. Figure 10 shows

Copyright ASCE 2004 Structures 2000

Advanced Technology in Structural Engineering
 7

the solution distribution of the analyses results. We observed possible solutions of 12

adjustments in which all control items were within target values.
Among those with tension error, 6 cables, which had comparatively low target tension on the
main tower side, surpassed the control target value. They were included with the cables that
were subject for adjustment. The other cables became subjects for adjustment in order to lower
the error of the main tower, which had maximum error in shape. But even if we did adjust these
cables,the error in shape of the main tower would hardly be corrected at all. In fact, we surmised
Downloaded from ascelibrary.org by University College London on 05/30/16. Copyright ASCE. For personal use only; all rights reserved.

that the tension error in the cables would actually become worse than before the adjustments. We
also surmised that the shape error on the principal span side, which was starting to rise slightly,
would not improve.
For that, we excluded the cables with shape errors that tended to worsen and the cables with
comparatively little adjustment value. Considering the fluctuation capacity of the main girder
and main tower through cable adjustments until that time, we decided to finally adjust 7 cables.
The final measurement results are asfollows. The results satisfied the control target values.
After adjustment Maximum tension error (Rt-max): 9.9%
After adjustment Maximum shape error (Rt-max): -252mm
With the various optimum shim decision techniques including GA, results are presented in
which sacrifice is made within the allowed limitations as a tradeoff in order to improve certain
items. We believe that technological judgment that considers the efficacy of improvement and
error distribution is indispensable in the process of deciding the final adjustment value.

(a) Before final adjustment (b) After final adjustment

Tension measurement result Tension measurement result
Figure 8 At time of final adjustment Tension measurement result

shape error shape error

k------j /“---;
----r-----55-

P2 P3 P5
---L---z-25 j----

; :,‘----L‘Tz?::----
________
1 -
-
50 i
75 ----
t
control value
--------------------i
i

___- 100 I

(a) Before tinal adjustment (b) After final adjustment

Shape measurement result Shape measurement result
F&yre 9 At time of final adjustment Shape measurement result

Copyright ASCE 2004 Structures 2000

Advanced Technology in Structural Engineering
 8
E

Downloaded from ascelibrary.org by University College London on 05/30/16. Copyright ASCE. For personal use only; all rights reserved.

Rt - rnax (%)

Figure10 GA analysis result Rf-Rt plot chart

CONCLUSIONS
The following conclusions were obtained from this study.
(1) The application of genetic algorithm is an effective way to solve shim adjustment problems
for cable system bridges such as a cable stayed bridge, because it can account for discrete
combination values that could not be easily obtained using other methods. The program
method project succeeded in obtaining an adequately usable solution to a shim adjustment
problem with the cable-stayed bridge in a short period of time.
(2) The ranking and sharing method that has been used in casesof conventional two objective
functions was used for a three objective functions problem including the number of cables
adjusted allowing the presentation of a pareto-optimal solution for the residual error of the
shape and tension and for the number of cables adjusted.
(3) By focusing on the form of the distribution of the initial population and mixing the
population close to the final solution, it was possible to improve the convergence precision
of the pareto-optimal solutions that had peaked in the past.
(4) Using conventional methods, it was necessaryto perform trial calculations to vary the given
conditions such as weighting and aspiration levels when making decisions, but with this
method, it is not necessary to repeat the analysis when conditions have changed, and
therefore a solution suitable to the conditions may be selected based on the results of the first
analysis. When performing erection control, it is extremely convenient to use GA that can
present pareto-optimal solutions including the number of cables to be adjusted.
(5) With the GA analysis results in the case of Tsuneyoshi Bridge, control target values were
satisfied because possible solutions of 12 out of 24 adjustments were utilized as an
application example of actual erection construction. We brought GA into the erection control
system this time. This is the tirst time applied to a real bridge but we have confirmed its high
reliability.
REFERENCES
1) H. Furuta and H. Sugimoto (1997): Application of Genetic Algorithms to Structural
Engineering, Morikita Shuppan. (in Japanese)
2) D.E. Goldberg and J. Richardson: Genetic Algorithms with Sharing for Multimodal
Function Optimization, in Proceedings of the Second International Conference on Genetic
Algorithms, Lawrence Erlbaum Associates, Hillsdale, NJ, 1987.
3) M. Kaneyoshi, H. Furuta and H. Tanaka: Practical Cable Tension Adjustment Based upon
the Satisfaction Consept of Bridge Designers, Proceeding of the Japan Society of Civil
Engineers, No. 598/I-44,1998,381-390.

Copyright ASCE 2004 Structures 2000

Advanced Technology in Structural Engineering