Data fusion of laser-induced breakdown spectroscopy and spectral reflectance techniques for estimating the mineralogical composition of copper concentrates.

Abstract

The pyrometallurgical copper industry faces some challenges in terms of the instrumenta tion for its processes. In this work, Laser-Induced Breakdown Spectroscopy (LIBS) data will be studied and combined with Diffuse Reflectance Spectroscopy (DRS) data and also with Hy perspectral Imaging (HSI) data to characterize the elemental and mineralogical composition in copper concentrates. This knowledge can be used to develop a sensor that replaces the current procedure used, which is risky, slow, and generates toxic waste and gaseous emissions. LIBS spectra are used for elemental characterization of samples, whereas DRS spectra can be used for molecular or mineral determination. HSI sensors provide a wider range of data for the sample material. The information from these sources can be fused to obtain a more reliable characterization. These spectroscopic techniques are high dimensional in terms of features or wavelengths. In order to process these datasets, it is essential to reduce their dimensionality, which can be done by using variable selection techniques. In LIBS, the expert selection is frequently used since there are peaks that are known to be associated with certain elemental species. For DRS and HSI data, it is less direct how to choose some wavelengths. Thus some automatic variable selection algorithms can be applied for this task. In this work, two variable selection methods are proposed for LIBS data. Both methods combine the use of expert knowledge to select the best wavelengths. Before fusing LIBS and HSI datasets, DRS is fused with LIBS data using a small dataset. LIBS and HSI data are finally fused using low-level and mid-level data fusion techniques. For each regression analysis, artificial neural networks (ANN) were used, which have gained attention for regression studies due to the flexibility in dealing with large amounts of nonlinear correlated data. The results show that by using mid-level data fusion, it is possible to outperform the performance of the individual sources, with root mean squared errors of prediction reductions ranging from 4% to 70% in the case of LIBS-DRS data fusion, and from 1% to 74% in the case of LIBS-HSI data fusion.