Near infrared spectroscopy combined with interval selective ensemble learning for analysis of cement raw meal...

Detecting the component contents of raw meal is an important step in the quality control of the cement industry. Near infrared spectroscopy (NIRS) has demonstrated a significant application potential in the quality control of cement raw meal. This study proposes an interval selective ensemble learning (ISEL) method for quantitative analysis of CaO, SiO₂, Al₂O₃ and Fe₂O₃ in cement raw meal using NIRS. Firstly, the spectral data are divided into several sub-intervals to train individual learners. Then the K-means algorithm is used to cluster and select dominant individual learners from trained individual learners. The stacking integration strategy is used to combine the outputs of the advantageous individual learners. The cross-validation outputs of dominant individual learners are combined to construct a meat-training set, which is used to train a multilayer perceptron (MLP) for nonlinear integration of dominant individual learners. ISEL is validated on two NIRS datasets obtained from different cement plants and measured with different types of spectrometers. Experimental results show that the proposed ISEL outperforms widely used chemometric methods as well as ensemble and selective ensemble approaches. For dataset A, R²_p are 0.9721, 0.9476, 0.9383 and 0.9255, with RMSEP of 0.181%, 0.2015%, 0.0698% and 0.056%. For dataset B, the R²_p and RMSEP are 0.9247, 0.8544, 0.8474, 0.8363 and 0.1868%, 0.2238%, 0.0781%, 0.0328%, respectively. This study demonstrates that NIRS couples with the ISEL is a novel and reliable approach for the quantitative analysis of cement raw meal.