Keywords: patch load, steel structures, neural networks, fuzzy inference system, web buckling, web crippling, structural design and behaviour.

This work presents a neuro-fuzzy system applied to the prediction of steel beam patch load resistance. A neural network system, developed in preceding studies [1,2], have been compared with experimental data and existing design formulae, providing good results. The performance of neural networks was actually significantly more accurate than patch load prediction formulae [3]. Despite the accuracy of the obtained results, the system architecture did not explicitly considered the fundamental difference in the beam ultimate limit state associated with the structural collapse (web and flange yield, web buckling and web crippling).

This paper proposes a neuro-fuzzy system, Figure 44.1, that can take into account the patch load response according to the related physical phenomena. The System architecture is composed of one neuro-fuzzy classification network and one patch load prediction neural network.

The neuro-fuzzy network is used to classify the evaluated beams according to its pertinence to a specific structural response (yielding, buckling and crippling). Then, the pertinence values established by the neuro-fuzzy classification network are used by a neural network to produce the beam ultimate patch load resistance.

The data used in the training and testing phases of the Neuro Fuzzy classification system uses geometric material properties as input and the associated physical phenomenon as the output. This strategy required the previous knowledge of the physical phenomenon related to each patch load case in the database, Roberts, [3].

The software was developed at ICA Lab (Applied Computational Intelligence), PUC-Rio, as part of a MSc dissertation, Gonçalves, [4], conducted at the ICA Laboratory, Electrical Engineering Department of the Pontifical Catholic University of Rio de Janeiro, PUC-RIO. The program not only acts as a classifier tool but also identifies the membership level associated to different physical phenomena.

The database was divided into three groups: 70% for training, 10% for validation and 20% to test the adopted model. The input variables used in the first training phase were the eight geometrical and material properties. After several training procedures were performed using the above mentioned data additions, the error was reduced to 5.75% and 7.41% in the train and validation groups.

The Neuro-fuzzy classification network results provide some interesting conclusions. As the flange thickness is reduced, crippling collapse becomes more frequent, while beams with thicker flanges presented a web buckling collapse. Another interesting conclusion is related to the web thickness parameter. As expected, the web yielding class degree for thin web beams was close to zero while the web buckling, and specially web crippling class degrees were very significant.

The yield collapse is clearly associated with beams with compact webs while the crippling and buckling collapses tend to be more frequent in slender web beams. This investigation is proceeding with the train and test phases of the neural patch load prediction network with the obtained collapse load class degrees and the experimental ultimate loads. This will be followed by an extensive parametric analysis of rolled and welded beams performed with the Neuro-Fuzzy system.

1

E.T. Fonseca; M.M.B.R. Vellasco; P.C.G. da S. Vellasco; S.A.L. de Andrade; M.A.C. Pacheco; "A Neural Network System For Patch Load Prediction", J. of Int. and Rob. Systems, V. 31, (1/3), pp.185-200, (2001). doi:10.1023/A:1012027726962

2

E.T. Fonseca; M.M.B.R. Vellasco; P.C.G. da S. Vellasco; S.A.L. de Andrade; M.A.C. Pacheco; "A Unified Neural Network Approach for Steel Beams Patch Load Capacity", Proc. of the 6th Int. Conf. on the Application of Artificial Intell. to Civil and Struct. Eng., Civil-Comp Press, Stirling, 14 (2001). doi:10.4203/ccp.74.14

3

T.M. Roberts; A.C.B. Newark; "Strength of Webs Subjected to Compressive Edge Loading", J. Str. Eng., ASCE, 123(2), 176-183, (1997). doi:10.1061/(ASCE)0733-9445(1997)123:2(176)

4

L.B. Gonçalves; "Modelos Neuro-fuzzy Hierárquicos BSP para Classificação de Padrões e Extração de Regras Fuzzy em Banco de Dados", MSc Dissertation, Electrical Eng. Dept., PUC-Rio, 2001.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £82 +P&P)