Study on Talc Grade Identification Based on LIBS Combined with Machine Learning

As a high-value industrial mineral, the accurate classification of talc quality grades directly impacts the quality of downstream industrial products and the added value of resource utilization. Currently, traditional talc classification primarily relies on manual visual inspection or conventional optical sorting techniques, which suffer from strong subjectivity, low efficiency, and high misjudgment rates. Existing online detection technologies such as X-ray fluorescence spectroscopy (XRF), near-infrared spectroscopy (NIRS), and neutron activation analysis (NAA) exhibit certain application advantages in specific scenarios but are generally plagued by high equipment costs, response delays, and elevated detection expenses. These limitations render them inadequate to meet the demand for rapid online detection in modern production lines. Therefore, there is an urgent need to develop an accurate, efficient online classification technology for the precise identification of talc quality grades. Laser-induced breakdown spectroscopy (LIBS) technology, characterized by the elimination of complex sample pretreatment, remote operability, and rapid real-time online analysis capabilities, offers a novel and effective solution for talc classification. In this study, a self-designed online LIBS experimental setup was constructed to achieve rapid quality detection of talc samples moving on a conveyor belt. To address LIBS spectral fluctuations caused by the irregular shape of talc samples, effective spectral data at the detection focal point were screened using the relative standard deviation (RSD) of characteristic element spectral intensities, thereby reducing data variability induced by surface irregularities. Characteristic spectral lines of key elements (Al, Ca, Mg, Fe, Na, and Si) were selected and dimensionally reduced via Principal Component Analysis (PCA). Subsequently, three machine learning classification models—support vector machine (SVM), K-nearest neighbors (KNN), and random forest (RF)—were established based on the PCA-processed data. Comparative analysis of the results demonstrated that the SVM model outperformed the others across four key metrics: accuracy, precision, recall, and F1-score, achieving a classification accuracy of 99.4%. The kernel function mechanism of SVM effectively captures nonlinear features in spectral data. In conclusion, the integration of LIBS technology with PCA dimensionality reduction and SVM machine learning provides a rapid, high-accuracy technical scheme for the real-time online classification of talc grades. This approach holds significant practical application value in promoting the efficient utilization of industrial talc resources.