NIRS-XRF fusion spectroscopy for coal calorific value prediction using data deficient learning

Abstract

Accurate and timely prediction of the coal calorific value holds immense significance in ensuring efficient and effective utilization of energy resources. In this manuscript, a novel analyzer for the rapid and precise determination of coal calorific value is developed by integrating near-infrared spectroscopy (NIRS) with X-ray fluorescence spectroscopy (XRF). Previous studies primarily rely on traditional linear regression methods to establish the mapping between NIRS-XRF fusion spectroscopy and the calorific value of coal, given the predominance of a linear relationship between the two. However traditional linear regression analysis methods fail to predict the accurate calorific value of coal, because of non-linear system biases and other non-linear issues that exist in real applications. Due to the limited open-source dataset and the difficulties of labeled data collection, few approaches utilize the non-linear neural networks to predict the calorific value of coal due to the risk of over-fitting. Improving the performance with such a small training set is crucial for the adjustment of the NIRS-XRF analyzer system. In this manuscript, we propose a novel approach named PLS-assisted neural fine-tuning, which consists of two stages: a PLS pre-training stage and a fine-tuning stage. The PLS pre-training stage employs data augmentation to generate simulated inputs with pseudo labels obtained by PLS. These synthesis samples are used to transfer the linear priori into neural networks from PLS to eliminate the risk of over-fitting issues. The proposed PLS-assisted neural fine-tuning method combines the linear priori and the advantages of neural networks to outperform conventional machine learning methods and vanilla neural networks. We perform comprehensive experiments to validate the effectiveness of the PLS-assisted neural fine-tuning method. Notably, the proposed PLS-assisted neural fine-tuning approach improves the performance without bells and whistles in data deficient scenarios, and especially brings significant improvements within 50 training samples. Furthermore, compared to the widely used PLS algorithm, our algorithm demonstrates superior robustness across data with varying crushing particle sizes.