Aprendizado de máquina para predição de resistência à compressão de argamassa

Abstract

This study aims to evaluate the performance of Machine Learning algorithms in predicting the compressive stregth of mortars. The dataset evaluated underwent a pre-processing to verify outliers, which could not be removed as it would make the analysis unfeasible, due to the small sample size that resulted. In addition to the raw data, it was decided to test the normalized and standardized data. Then, the algorithms: k – Nearest Neighbour (k-NN), Support Vector Machine (SVM) and Random Forest (RF) were used to predict the compressive strength, evaluating the performance of the model, where the RF algorithm presented a better performance when compared to the others. The parameters of the RF algorithm were adjusted and selected the values that dissipate the lowest RMSE for the division of the dataset in training and testing, control of subset division, number of tree and number of variables evaluated per node. After determining the best parameter settings for the selected algorithm, the model was obtained again to verify its performance with adjustments, obtaining R2 of 86.94%, MAE of 4.64 MPa and RMSE of 7.42MPa, when using test data from the final model. Finally, it is concluded that Machine Learning (ML) is a practical calculation tool in predicting the compressive strength of mortars.