Intelligent Determination of Oxide Contents in Cement Raw Meal Based on Near-Infrared Spectroscopy and Deep Learning Techniques

The composition analysis of cement raw meal is central to quality control in cement production where accurate determination of major oxide contents directly impacts clinker quality and cement performance. However, traditional detection methods are incompatible with dynamic quality control demands in modern industrial production because of some inherent limitations such as prolonged analysis cycles and destructive sampling. Recently, Near-Infrared (NIR) spectroscopy has emerged as a promising technology due to its rapid and nondestructive detection capabilities. The contribution of this paper is to develop a digital-intelligent system integrating NIR with Deep Learning (DL) for mineral content determination in cement raw meal. A full-process analytical framework comprising spectral acquisition, data preprocessing, feature wavelength selection and DL model construction was established to quantify SiO2, Fe2O3, Al2O3, and CaCO3. Standard Normal Variate (SNV) transformation and Savitzky-Golay (SG) smoothing were applied to mitigate spectral noise and baseline drift, while the Successive Projection Algorithm (SPA) identified key wavelengths to reduce data dimensionality. A hybrid Convolutional Neural Networks (CNN)-Bidirectional Long Short-Term Memory (BiLSTM)-Attention model was subsequently developed to perform deep feature extraction and nonlinear modeling. For cement raw material samples from the Qufu Factory test set, the model achieved robust prediction performance for major oxide contents with determination coefficients (R2) exceeding 0.86 and Root Mean Square Errors (RMSE) below 0.0038 for CaCO3. The model also maintains stable performance when applied for the Linyi factory. This NIR-DL approach enables rapid, nondestructive, and precise oxide contents monitoring, providing an efficient solution for automated quality control in cement production.