Classifying iron ore with water or dust adhesion combining differential feature and random forest using hyperspectral imaging

Abstract

Hyperspectral imaging (HSI), a promising technique integrating imaging and spectroscopy, can help sort iron ores with different total iron (TFe) contents. However, the adhesion of dust (caused by crushing) or water can affect the sorting process. Currently, the mechanisms underlying this influence and methods to conveniently mitigate it remain unclear, hindering the practical application of HSI-based sorting. This study aimed to investigate this issue. For the experimental materials, 300 ore samples (particle size: 20–40 mm) with different TFe contents were prepared. Subsequently, three sample conditions were prepared (“No dust, no water”, “With dust, no water” and “No dust, with water”) through washing and drying measures, and their hyperspectral images were acquired (953–2517 nm). Finally, the TFe content of each ore sample was measured. After preprocessing, the effects of water and dust on the spectra and sorting process were initially analyzed. Subsequently, a new spectral differential feature considering dust and water (DFDW) was proposed to mitigate this influence. Then, using the spectral and calculated proportion features as input, different grades of iron ore were classified into four classes using a machine learning classifier. For validation, models using several different input features and machine learning classifiers were tested for classification accuracy (the ratio of correctly predicted instances to the total number of predictions). On “No dust, no water”, “With dust, no water” and “No dust, with water” data, the model DFDW-random forest (RF) achieved accuracies of 87.7 %, 85.0 %, and 85.3 %, respectively, which was optimal. Overall, the results enhance the universality of HSI-based iron ore sorting and provide technical support for its practical implementation.