Prediction of billet loosening and central segregation based on wavelet transform weighted twin support vector regression

Abstract Billet loosening and central segregation are critical factors that significantly affect billet quality. Traditional neural networks often suffer from slow convergence, rigid parameter adjustment, and a tendency to converge to local optima when predicting billet loosening and central segregation, which limits both model accuracy and computational efficiency. To overcome these challenges, this study proposes the WTWTSVR prediction model, which integrates the wavelet transform-based weighted algorithm (WTW) with an optimized twin support vector machine for regression (TSVR). The WTW algorithm modifies the weight values in the TSVR-derived objective function, thereby converting it into an unconstrained optimization problem. This optimization enhances both computational speed and prediction accuracy. To validate the predictive performance of the model, the WTWTSVR was compared with the barrel theory-based optimizer (BTO) and SVR algorithms in experiments. The experimental results show that the WTWTSVR model achieved hit rates of 92.6%, 91.2%, and 95.4%, 96.1% for billet loosening and central segregation within ±5% and ±10%, respectively. The WTWTSVR model offers an effective method for predicting billet loosening and central segregation, providing practical value and guidance for billet production.