Applications of Neural Network and Neuro-Fuzzy Network to Estimate the Parameters of Self-Compacting Concrete

Abstract

The paper presents the new application of two classical nonlinear estimators, which are the multi layer perceptron and the neuro-fuzzy networks, to approximate the workability parameters of fresh selfcompacting concrete based on the amount of input ingredients like cement, fly ash, water, additives or admixtures. The estimation of workability parameters is much needed to determine the quality of the fresh selfcompacting concrete before starting the production. A total of 360 real field tests of 30 types of self-compacting concrete were conducted and seven basic parameters were measured for each test. These samples will form the training and testing data sets for the nonlinear models. The numerical results showed that the MLP network could estimate the workability parameters with relative errors less than 3.6% and the TSK could estimate with te relative errors less than 2.7%. These results proved the ability to create high accuracy approximation models of the proposed solutions, where the neuro-fuzzy model would show a little better performance than the multilayer perceptron. Both of the models required only relatively simple structures, making them more promising to be used in practical applications.

The Application of Neural Networks to Predict the Water Evaporation Percentage and the Plastic Shrinkage Size of Self-Compacting Concrete Structure

“…The transfer function of hidden neurons is denoted as f1, and the transfer function of output neurons is f2. The popular selection of f1 is the 𝑡𝑎𝑛𝑠𝑖𝑔() defined as follow [22]:…”

“…In this paper, the classic Levenberg-Marquadrt algorithm was used to train the MLPs [25]. The selection of the optimal number of hidden neurons was determined using a trial-and-error approach, similar to the method outlined in Nguyen & Tran [22] and Haykin [25].…”

“…During the training process, it's needed to find the number of rules that enable the network to achieve a desirable low level of testing errors. We used a similar "trial and test" approach, starting with a network containing just one rule and increasing the number of rules during these trials [22]. To train the TSK networks, the hybrid algorithm described in Linh [27] was used.…”

“…In recent years, in the fields of technology in general and construction technology in particular, artificial neural networks (ANNs) have been studied and applied to model nonlinear relationships between object parameters [22][23][24] as well as the behavior of materials. According to Haykin [25], the advantage of using ANN is the capability to solve problems that do not have a known or given mathematical model of the object.…”