Rapid Detection of Iron Ore Grades Based on Fractional-Order Derivative Spectroscopy and Machine Learning

The time-consuming nature of chemical testing techniques makes them lag behind mineral processing. Therefore, this article combines visible-infrared reflectance spectroscopy with machine learning (ML) algorithms to achieve rapid detection of iron ore grades and meet the requirements of mining production. First, the standard normal variate (SNV) and de-trending (DT) are used to eliminate noise and baseline drift in the original spectral data. Then, extraneous signals are removed using direct orthogonal signal correction (DOSC). In addition, fractional-order derivative (FOD) is performed on the DOSC spectrum to further amplify the spectral details. To extract spectral features and reduce the spectral dimension, a multilayer incremental extreme learning machine autoencoder (MIELM-AE) is proposed in this article. MIELM-AE can automatically match the optimal number of network nodes and network layers to minimize the reconstruction error. The experimental results show that the Pearson correlation coefficient (
${R}^{{2}}$
) of the extreme learning machine (ELM) built using MIELM-AE improves from 0.715 to 0.821, compared with the ELM built without the dimensionality reduction method. To increase the measurement accuracy, this article uses Tikhonov regularization and truncated singular value decomposition (TSVD) to alleviate the ill-conditioned matrix of the hidden layer of the ELM and uses the incremental method to match the optimal network nodes. Finally, double-regularization incremental ELM (DRIELM) is proposed in this article. Experiments show that DRIELM obtained the highest detection accuracy with an
${R}^{{2}}$
of 0.932 at an FOD of 0.4.