Uso de redes neurais de funções de base radial e regressão linear múltipla para a previsão de propriedades mecânicas de vergalhões de aço

Abstract

With the advancement of technology, it is becoming increasingly necessary to use advanced tools in the pursuit of maximizing production and cost reduction in the steel industries. The present work studies two data prediction algorithms: a Radial Base Function Neural Network (RBFN) and a Multiple Linear Regression (MLR) to predict 4 main mechanical properties resulting from the production of steel rebar: yield strength (LE), ultimate tensile strength (LR), ultimate tensile strength and yield strength (LR / LE) and elongation ratio (AL). Through data obtained from the steelmaking and rolling, ranging from the chemical composition of the material to temperatures throughout the process, production rate and water pressure and flow in the final heat treatment, we seek to find a suitable mathematical model that can the final mechanical properties of the product resulting from this process. The coefficient of determination (R2) was used to determine the quality of the model found in the various configurations tested during 100 test rounds each for the two algorithms, randomly dividing the available data into one training set and one test set. For the multiple linear regression, for the LE, LR, LR / LE and AL, respectively, an average R2 0.5146, 0.2388, 0.7633 and 0.5099 was found. For the neural network, 0.8437, 0.7403, 0.8691 and 0.7015. The results indicate that RBFN was able to more accurately predict the properties when compared to MLR. The R2 averages of 100 network rounds, considering a real and complex problem of a steel mill, can be considered high, representing an important and easy implementation tool.