Development and application of a rapid coal calorific value analyzer based on NIRS-XRF

Abstract

The combination of a rapid coal calorific value analyzer based on NIRS-XRF and a PLSR-based holistic-segmented modeling method greatly improves measurement accuracy and repeatability.

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

“…3, the range of calorific value for all coal samples is between 17.75 MJ kg −1 and 25.86 MJ kg −1 . Furthermore, to validate the stability of our proposed algorithm with regard to coal sample crushing particle size, we have deliberately collected NIRS-XRF signals from coal samples with a 0.2 mm 21,32 crushing particle size. Notably, the methodology for capturing the NIRS-XRF signals from both 0.2 mm and 6 mm crushing particle sizes is identical, and the samples are sourced from the same batch of coal, thus ensuring a direct correlation between the two sets of data.…”

“…However, this approach is associated with radioactive hazards, necessitating significant financial investments in equipment procurement and ongoing maintenance, thereby rendering it a costly endeavor. Recently, the spectral sensors laser-induced breakdown spectroscopy (LIBS), 2,3,8–17 near-infrared spectroscopy (NIRS), 16,18–21 and X-ray fluorescence spectroscopy (XRF) 21–24 are utilized for evaluating coal properties. LIBS leverages a high-energy laser beam to ablate coal samples that generate plasma exhibiting a unique radiation spectrum.…”

“…In the past, conventional machine learning methods such as partial least squares (PLS) regression 15,21,32–36 and random forest (RF) 37–40 are mostly used to establish the relationship between spectral signals and ground-truth calorific labels in this field. Due to the existence of complex non-linear system biases and other non-linear issues in real-world applications, such as the complex environment issue, coal particle size and so on, these traditional machine learning methods could not accurately predict the calorific value of coal.…”

“…• The proposed PLS-assisted neural fine-tuning framework leverages the robust generalization of PLS on small training samples and the excellent fitting ability of neural networks, resulting in significantly better performance on datasets of different sizes compared to the PLS algorithm commonly used by previous studies. 21,32…”

“…• Upon rigorous evaluation, our proposed algorithm demonstrates remarkable robustness across data sets featuring diverse crushing particle sizes, surpassing the performance of the PLS algorithm. This advancement allows us to relax the crushing particle size from the commonly used 0.2 mm in previous studies 21,32 to 6 mm, ultimately translating to a reduction in manufacturing costs of NIRS-XRF analyzers.…”