Data Preprocessing Enhancement for Artificial Neural Networks in Predicting Mechanical Properties of Dual-Phase Steel

摘要

The mechanical properties of Dual-phase (DP) steel, including toughness and hardness, are influenced by the intercritical annealing process parameters, such as temperature, holding time, and cooling rate. Traditional regression models have limitations in accurately predicting these nonlinear relationships. This study addresses the challenge by introducing a data preprocessing method, the Regression-Based Data Preprocessing (RBDP) method, aimed at improving the accuracy of Artificial Neural Network (ANN) models in predicting toughness and hardness of DP steel. The RBDP method enhances the training dataset by generating synthetic data through regression analysis, allowing for more robust ANN performance. The mechanical properties of nine DP steel specimens were analyzed using RBDP combined with ANN models and compared with predictions from pure ANN models and traditional regression approaches. For hardness prediction, the PRH + ANN model emerged as the most accurate, achieving the lowest average error of 0.72%, performing particularly well for specimens 1, 3, 4, and 9. For toughness prediction, the PRT + ANN model was the most effective, delivering the lowest prediction errors across most specimens, with particularly low errors for specimens 1, 4, 5, 7, 8, and 9. Both PRH + ANN and PRT + ANN models outperformed traditional regression approaches and standalone ANN models. These findings demonstrate that the proposed RBDP method significantly improves the accuracy of ANN models for predicting the mechanical properties of DP steel. This study highlights the potential to enhance predictive modelling in materials science.