RCC BEAM
ABSTRACT
The aim of this paper is to investigate the potential of feedforward Artificial Neural Networks for the cost
optimum design of concrete structures.The paper presents and discusses the application of three Aritificial
Neural Network models for the cost optimum design of a singly reinforced simply supported RCC beam .
Backpropagation network is chosen for the proposed network using the Neural Network Toolbox of MATLAB.The
training and test data have been generated by the solution of the non linear constrained optimization problem
using the Solver Add In of Microsoft Excel. The results obtained from the neural nets are compared with those
INTRODUCTION
Optimal design of framed structures is always a goal of engineers. Most building designs are optimized by trial
and error combined with the experience of the designer. The trial-and-error process is continued until an
acceptable design is attained. The efficiency of the trial and error procedure is greatly dependent on the
experience and ability of the designer to select a good initial design and perturb the design in a better
direction towards an optimal design. As a result, techniques of optimization have been developed to automate
and mathematically orchestrate the trial and error procedure. In general, optimum design problems seek to
minimize a function (usually cost) using a set of design variables subjected to constraints. These optimization
techniques are usually lengthy and complicated to implement, hence putting more burden on the designer.
One of the techniques to reduce the computational efforts required in the aforementioned regard is to train a
suitable neural network with a number of optimum designs and simulate the same later on for achieving the
results. The neural network in this case comes up with a design based on its training rather than conducting a full
The cost function for a RCC beam typically includes the cost of concrete, the cost of reinforcement and the cost
FORMWORK
d D
CONCRETE
STEEL
The objective function is the cost per unit length of the beam and is given by the following relationship: -
Total Cost = Cc ( Ag Ast ) x 10-6 + Cs Ast x 7850 x 10-6 + Cf (b + 2D) x 10-3 (1)
The primary constraints for the optimization problem are based on the provisions furnished in IS 456:2000
which take into account the flexural strength of the member and the lower & upper limits of steel.
A side constraint can also be incorporated on the basis of the limiting value of the depth of the beam.
D-Dmax<= 0 (6)
The notations used in the above expressions have the following descriptions:-
The
Cc = Cost of concrete per unit volume (Rs/m3)
(6). The parameters Cc , Cs and Cf in equation (1) are considered as constants with values of 2500 (Rs/m 3 ), 35
(Rs/kg) and 140 (Rs/m2 ) respectively. The value of effective cover (d') is also taken as a constant equal to 20mm.
The problem is solved for different values of f y, fck, M, b and Dmax to build a database using which the neural
Network topologies with one and two hidden layers were considered for the present study. The number of nodes
in each of the hidden layers was varied (between 8 to 25 in the first hidden layer and between 4 to 10 in the
second hidden layer) as shown in Fig.2 and the resulting models were tested for convergence. The impact of the
size of the training data set on the convergence to the desired goal for each of the considered models was also
fy
doptimum
b
Ast optimum
Dmax
C optimum
M
Output Layer
Fig. 2 Neural Network Topologies considered for the cost optimum design of singly reinforced beam section
The other network parameters were kept constant and are stated below:-
For the formation of the training data set, two values of fck were used i.e 20 N/ mm2 and 25 N/ mm2, two values of
fy were used i.e. 250 N/ mm2 and 415 N/ mm2 , the magnitude of design moment was varied between 5 kNm to
100 kNm, values of 200mm, 230mm, 250mm and 300mm were adopted as the width of the beam and three
values 300mm, 350mm and 400mm of permissible maximum depth of the beam were used.
In the following sections the performance of four different network models: 5-20-3, 5-10-5-3, 5-25-3 and 5-8-6-3
have been analyzed. The decrease in mse for the proposed models during training have been represented in
Fig. 3 Decrease in mse for model 5-20-3 during Fig. 4 Decrease in mse for model 5-10-5-3
training during training
Fig. 5 Decrease in mse for model 5-25-3 during Fig. 6 Decrease in mse for model 8-6-3 during
training training
TEST PROBLEMS
The performance of all the neural network models proposed above has been tested using five different test
problems. The theoretical results obtained by solving the optimization problem and those obtained from the
Network Output
S.N Inputs ActualResult
5-25-3 5-20-3 5-10-5-3 5-8-6-3
d optimum = d optimum = d optimum = d optimum = d optimum =
fck = 25
280 277.8894 280.1127 282.0403 280.5236
fy = 415 Ast optimum = Ast optimum = Ast optimum = Ast optimum = Ast optimum =
677.0407 699.8586 666.7170 665.6755 680.8054
b = 230
1
C optimum = C optimum = C optimum = C optimum = C optimum =
D max =
300 473.0243 476.1252 468.8748 469.7399 471.7825
M=
56.5
d optimum = d optimum = d optimum = d optimum = d optimum =
fck = 20
140.6888 142.1051 140.1266 141.9983 140.6096
fy = 250 Ast optimum = Ast optimum = Ast optimum = Ast optimum = Ast optimum =
329.532 315.1070 320.3145 335.1902 313.1932
b = 230
2
C optimum = C optimum = C optimum = C optimum = C optimum =
D max =
300 259.304 260.2668 261.1172 259.0890 255.1663
M = 8.8
3 fck = 20
fy = 250
d optimum = d optimum = d optimum = d optimum =
b = 250 d optimum = 281.6 277.3 282.8 282.2
D max = 280 Ast optimum = Ast optimum = Ast optimum = Ast optimum =
300 Ast optimum = 1834.2 1908.5 1881.5 1841.6
1852.65
M=
C optimum = C optimum = C optimum = C optimum =
75.5 C optimum =
813 826.8 817.4 818.3
fck = 25 d optimum =
810.8841 d optimum = d optimum = d optimum = d optimum =
293.3395 300.6635 298.0142 299.8487 300.6101
fy = 415
Ast optimum = Ast optimum = Ast optimum = Ast optimum = Ast optimum =
b = 200 569.547 548.3442 561.1210 552.5007 551.0221
4
D max = C optimum = C optimum = C optimum = C optimum = C optimum =
400 428.1671 428.9059 426.3959 427.6137
427.464
M=
50.6
For the sake of ease of comparisons, the theoretical and the network outputs for effective depth and steel have
350
300
250
Outputs
200
150
100
0 1 2 3 4
Sample Num ber
Actual d Model 1 Model 2 Model 3 Model 4
1700
Outputs
1200
700
200
0 1 2 3 4
Sam ple Num ber
Actual Model 1 Model 2 Model 3 Model 4
been presented graphically in Fig.7 and Fig.8 respectively. From Fig.7 and Fig.8 it is evident that all the four
models considered for study have produced outputs for effective depth and amount of steel, which are close to
CONCLUSIONS
The paper considered the optimization problem of minimizing the total cost of a rectangular RCC beam section.
The theoretical results obtained for different material combinations, limiting depths, section widths and design
moments were used to train different neural network models. The impact of the size of the training data sets on
the convergence to the desired goal during training of the models was also studied. The network models with
one/two hidden layer(s) and with different number of nodes in the hidden layer(s) as shown in Fig.2 were not
found to converge when the training data set consisted of 100,150 ,200 and 250 samples. However convergence
was achieved with the training data set consisting of 500 samples.
Four different models were then considered and were tested with five different problems. The deviations in the
outputs generated by the artificial neural network models relative to the theoretical results were found to be
within tolerable limits as indicated by Figs. 7 and 8. Improvement in results can be further increased by
increasing the size of the training data set or by further adjusting the network topology.
An artificial neural net with a suitable topology and weight adjusting algorithm can be considered to be a reliable
means for the solution of engineering problems involving function approximation. The main advantage lies in
the flexibility offered by the network and the savings made in the computational efforts, since the problem in
REFERENCES
ii. Muhammad N.S.Hadi. Neural networks applications in concrete structures. Computers and Structures
81 (2003) 373-381.
iii. Kishan Mehrotra, Sanjay Ranka, Chilukuri K. Mohan. Elements of Artificial Neural Networks. Penram
International.
iv. Steven C. Chapra, Raymond P. Canale. Numerical Methods for Engineers. TMH.