Static and dynamic neural network modeling for reinforced concrete slab / Seyed Vahid Razavi Tosee

Seyed Vahid, Razavi Tosee (2012) Static and dynamic neural network modeling for reinforced concrete slab / Seyed Vahid Razavi Tosee. PhD thesis, University of Malaya.

Preview

PDF (Thesis PhD) Download (7Mb) | Preview

Abstract

Ttraditional analysis models for reinforced concrete (RC) structures are reliable and the behavior of structural elements can be successfully determined by solving several numerical equations. Another alternative analytical modeling method is Artificial Neural Networks (ANNs), which capture the numerical equations between its nodes and no formal formula is observable within the network generation. Previous researches of ANNs in Structural Engineering mainly focused on Feed-forward Back-propagation Neural Network (FBNN) using sufficient data for network generation. The key objective of this research is to train ANNs to predict mid-span deflection of RC one-way slabs for situation where insufficient data is available for network generation. This research also considered a network modeling with internal dynamic space and taped-delay line for load defection analysis which inherently could memorize the input data while training process. It involves the prediction of load deflection of 19 non-strengthened RC slabs under mid-span punching load and 7 Carbon Fiber Reinforced Polymer (CFRP) strengthened RC slabs under four point line loads. The results of experimental data were compared with finite element analysis using LUSAS software. The results were also validated with BS and ISIS code for non-strengthened and CFRP strengthened RC slab respectively. Generalized Regression Neural Network (GRNN) as Static Neural Network (SNN) was generated from the experimental results while there were insufficient data for network generation. To predict the mid-span deflection of RC slab, two types of Dynamic Neural Network (DNN) namely Focused Feed-forward Time-delay Neural Network (FFTDNN) and Recurrent Neural Network (RNN) were generated with sufficient data. This study also compared together the GRNN, FBNN, FFTDNN, and RNN for situations where data is sufficient for network generation. The results showed that the generated GRNN using insufficient data solve the problems in suitable techniques with mean error of 8 and 11.3% for non-strengthened and CFRP strengthened RC slab respectively. The generated FFTDNN, RNN, FBNN, and GRNN using sufficient data predicted the deflection with mean error of 8, 9.7, 10.5, and 14.9% respectively for non-strengthened RC slab and 7.3, 8.4, 9.3, 14.4% respectively for CFRP strengthened RC slab. It is clear that using FFTDNN and RNN modelling provided outstanding performance over the FBNN and GRNN for load-deflection analysis of RC slab.

Actions (For repository staff only : Login required)