Mix Proportion and Compressive Strength Prediction Models for High Performance Concrete based on Computational Intelligence

• Main Author: Chiew, Fei Ha
• Format: Thesis
• Language: English
• Published: Universiti Malaysia Sarawak, (UNIMAS) 2018
• Subjects: TA Engineering (General). Civil engineering (General)
• Online Access: http://ir.unimas.my/id/eprint/24829/1/Chiew%20Fei%20Ha%20ft.pdf; http://ir.unimas.my/id/eprint/24829/
• Institution: Universiti Malaysia Sarawak

This study proposed two new models based on computational intelligence approach to aid mix proportioning of high performance concrete. The models developed in this study are: (1) a data-driven fuzzy adaptive resonance theory (ART)-based model for predicting high performance concrete mix design, and (2) a data-driven monotonicity preserving single input rule modules (SIRMs)-connected fuzzy inference system (FIS) with harmony search (HS) model for prediction of high performance concrete compressive strength. Fuzzy ART is adopted because of its online learning and suitability for large and incremental datasets. The fuzzy adaptive resonance theory (ART)-based model receives a set of desired range of concrete performances, searches for a set of mix proportions which satisfies the desired range of concrete performances, classifies the mix proportions of selected data into clusters, measures the similarity between performances of deduced clusters with desired performances, and deduces a mix proportion. The model was used to estimate mix proportions of five batches of concrete based on the performance criteria of 7 th and 28th day compressive strengths. The generated mix proportions were used in an experimental work, and 4 out of 5 mix proportions achieved experimental 7th and 28th day compressive strengths that satisfy the desired performances, signifying the ability of the fuzzy ART-based model in estimating the mix proportion of HPC. The ability of the model to estimate mix proportions with different constituents is also demonstrated. The proposed high performance concrete mix proportion prediction model is a novel datadriven fuzzy ART-based model with an application in civil engineering. On the other hand, the monotonicity preserving HS-based SIRMs-connected FIS model is adopted because it considers the monotonic relationship between the input (i.e., water-binder ratio) and output v of the model (i.e., compressive strength). The proposed model is evaluated using experimental data. An evaluation was carried out to study the suitability of monotonicity preserving property of the HS-based SIRMs-connected FIS model, by comparing its results to a HS-based SIRMs-connected FIS model without monotonicity preserving property. Results showed that the predictions from the model with monotonicity preserving property for testing dataset are able to give lower root-mean-square error (RMSE) and mean absolute percentage error (MAPE) values. Results from the model were further compared with results from neural networks fitting tool in matrix laboratory (MATLAB). Although neural networks fitting tool gave better predictions for training data set, neural networks does not show any meaningful relationship between water-binder ratio with predicted compressive strength. The monotonicity preserving HS-based SIRMs-connected FIS model is found to give better predictions for testing dataset by having a lower RMSE value compared to neural networks fitting tool. Incorporating monotonicity preserving property into the compressive strength prediction model reflects the nature of relationship between water-binder ratio with compressive strength, and is able to correct noisy data during evaluation. Both models in this study serve the purpose as material models that can provide more economical and faster way in obtaining a satisfactory high performance concrete mixture with desired properties. The limitations of both models are that both models are not able to consider experimental data with missing values. Therefore, future works recommended is to extend the use of both models to consider experimental data with missing values.