Research on Prediction of Bubble Defects in Float Glass Based on LSTM-XGBoost

Float glass production is a highly complex continuous high-temperature physicochemical reaction process, and bubble defects are a key factor affecting the yield of finished glass. Due to the large thermal inertia of the kiln, strong time-series correlation, and complex nonlinear coupling relationships, traditional single physical modeling or shallow machine learning models face limitations in bubble prediction accuracy. This paper constructs a time-series fusion hybrid prediction model based on a combination of Long Short-Term Memory (LSTM) and Extreme Gradient Boosting (XGBoost). The model utilizes LSTM to perform sliding window modeling on multi-dimensional process parameters to extract deep time-series fusion feature vectors. Subsequently, these time-series features are concatenated with the cross-sectional process parameters at the current moment and input into an XGBoost regressor for defect prediction. This study conducts experimental validation based on actual industrial data from a float glass production line, and compares it with Linear Regression (LR), Support Vector Regression (SVR), and pure XGBoost models. Experimental results show that the proposed LSTM-XGBoost hybrid prediction model significantly outperforms traditional single baseline models, providing data support for the early prediction of glass bubbles and intelligent control of the melting furnace process.