Contribuições ao problema de predição das propriedades mecânicas de vergalhões de aço usando regressão por máquinas de Kernel

Abstract

Aiming to ensure product quality and reduce the cost of steel manufacturing, nonlinear regression models are developed to predict the mechanical properties of steel rebar using machine learning techniques. Bearing this in mind, this problem is revisited by developing an architectural methodology that combines two key concepts in the construction of parsimonious models: (i) sparsity, a property in which a predictive model is built with fewer relevant elements (e.g., support vectors) and consequently with fewer associated parameters, and (ii) locality, a property in which several simpler models are fitted to smaller partitions of the data. In this context, two regression models based on the (least squares support vector regression, LSSVR) model are developed. The first is an enhanced sparse version, and the second is a new local regression model based on LSSVR. The task of interest is the prediction of four variables used to evaluate the mechanical properties of the produced steel rebar, namely: yield strength (YS), ultimate tensile strength (UTS), the UTS/YS ratio, and percent elongation (PE) based on information about its chemical composition (12 variables) and the parameters of the rolling heat treatment (6 variables). The proposed models are evaluated using real data collected from steel rebar manufacturing and compared with the global LSSVR model in terms of accuracy and efficiency. The local and sparse approach of the LSSVR model with the Cauchy kernel consistently outperforms the standard single regression model approach with the Gaussian kernel in the task of interest, achieving average improvements in R2 compared to previous studies: 6.49% for YS, 6.47% for UTS, 2.72% for UTS/YS, and 4.95% for PE. Additionally, the model sparsification and local regression approach significantly reduce the number of operations with support vectors, using an average of 36.25% of the total available support vectors for the estimation of YS, 38.75% for UTS, 31.83% for UTS/YS, and 36.73% for PE, reducing approximately 52% of the calculation time for predicting the model outputs.