Objective From iron ore to the final steel processing, accurate mineral content data is essential to maximize raw materials and energy accurately control the manufacturing. Mineral flotation is a beneficiation method in which target minerals and impurities are separated based on the physical and chemical properties of target minerals and impurities and then extracted from the original ore slurry. Content of iron ore slurry directly affects the flotation effect and quality and output benefit of the final product. Therefore, conducting an accurate quantitative analysis of the iron ore slurry composition is essential. Laser-induced breakdown spectroscopy (LIBS) has been widely used to detect material composition owing to its advantages such as online, in situ, and simultaneous measurement of multiple elements. However, self-absoprtion and matrix effects in LIBS affect the accuracy of the analysis. Simultaneously, with the continuous improvement of the spectrometer’s resolution, the data dimension is increasing, including a large amount of redundant information that is unnecessary for component analysis. When using PLS and LIBS for quantitative analysis, the existing research uses spectral line feature selection to reduce dimensionality and nonlinear correction to make improvements separately. To simultaneously reduce the data dimension and correct the nonlinear problem of the data itself, we build a nonlinear PLS model to reduce the influence of self-absorption and matrix effects on the accuracy of quantitative analysis. In addition, the characteristic variables are cyclically filtered to reduce the modeling complexity.
Methods PLS is widely used in the quantitative analysis of material components, but as a linear processing method, it cannot resolve the nonlinear effects of self-absorption and matrix effects on the spectrum, reducing the accuracy of quantitative analysis. The characteristic spectrum line n-order polynomial form was proposed to be added to the PLS model. Thus, we can reduce the dimensionality of the data to extract the most useful information and reduce the complexity of the model by filtering feature variables. Taking the iron (Fe) element in the iron ore concentrate slurry as the analysis object, to reduce the influence of self-absorption on the quantitative analysis of the element to be analyzed, 10 characteristic spectral lines of Fe were selected. Simultaneously, to reduce the interference of other elements, 5 characteristic spectral lines of silicon (Si) were selected, and their three-order polynomial form was added to the modeling of PLS to correct the nonlinear influence caused by self-absorption and matrix effect. The regression coefficients of the variables were sorted according to the absolute value, and the optimal variables were determined by cyclically filtering the variables to reduce the interference of redundant information of the variables and reduce the model’s complexity.
Results and Discussions Using the training set to build the model and determining the optimal variables and the number of principal components according to the root mean square error (RMSE) of the validation set, we made predictions on the prediction set and compared the traditional PLS model, the nonlinear PLS model with the characteristic spectrum lines three-order polynomial form, and the model proposed in this paper. The RMSE of the traditional PLS model is 1.15%, and the coefficient of determination R2 is only 0.51 (Fig.6). However, the RMSE of the nonlinear PLS model is reduced by 0.85%, and the coefficient of determination R2 is 0.73 (Fig.8). Furthermore, the RMSE of the cyclic filtering variable nonlinear PLS model proposed in this paper is reduced to 0.70%, and the coefficient of determination R2is increased to 0.86 (Fig.11).
Conclusions We propose a nonlinear PLS model based on cyclic variable filtering to address the problem that LIBS is used for composition analysis, which is often affected by self-absorption and matrix effects and data redundancy caused by excessively high spectral data dimensions. The analysis object is the Fe element in the iron ore concentrate slurry, compared with the traditional PLS modeling method (Table 2). As a result, the RMSE of validation set is reduced from 1.15% to 0.70%, and the coefficient of determination R2 increased from 0.51 to 0.86. The result shows that the nonlinear PLS model based on cyclic variable filtering can significantly improve the analysis accuracy of Fe in iron concentrate slurry, indicating that this method has evident effects on the quantitative analysis of elements that are greatly affected by the matrix effect and self-absorption.
