【摘要】 金矿资源具有重要的经济和金融价值，不仅为国家提供了贵重的金属资源，推动经济增长，还在增强货币稳定性和国际金融市场中的避险能力方面具有现实意义。然而，当前矿山用于金矿品位测量的化学分析法尽管精确，但存在耗时长、成本高以及药剂污染等多种问题，无法实现基于实时品位信息的矿石品位与选矿方法的自动化调整。相比之下，可见光-近红外光谱分析法因其高效、绿色环保及原位测定等优势，逐渐成为估算矿区金属品位的有效替代方法。为此以中国辽宁省二道沟、凌源和排山楼三个金矿为研究区，共采集了389个金矿样本，以SVC便携式地物光谱仪测试的高光谱数据和化学分析数据为数据源。首先对原始光谱数据进行Savitzky-Golay平滑(SG)处理，并分析金矿的光谱特征，发现反射率与金品位具有一定相关性，且在455 nm处具有金的吸收特征，基于此，利用主成分分析法(PCA)、等距特征映射(ISOMAP)和局部线性嵌入(LLE)算法对原始光谱数据进行降维处理，对应降维结果的维数分别为6, 5, 5。最后基于随机森林(RF)、极端随机树(ET)、决策树(DT)、梯度提升树(GBDT)和自适应增强(Adaboost)、极端梯度提升树(XGBoost)和Stacking集成学习算法对降维后的数据建立了金品位预测模型。研究结果表明，Stacking集成学习方法在各方面性能均优于单一模型，其中LLE-Stacking组合模型的精度最高，预测值与真实值的R~2为0.972, RPD为5.935,平均相对误差为0.231。利用本方法可以快速准确预测矿粉中金的品位，相比于传统模型的品位反演精度有明显的提升，为矿山金品位的快速、原位测定提供了新的技术手段，对金矿的高效开采具有重要意义。更多还原

【Abstract】 The gold mining resource holds significant economic and financial value, providing precious metal resources for the country, driving economic growth, and enhancing currency stability and hedging capabilities in the international financial market. However, while precise, the current chemical analysis methods for measuring gold ore grades in mines face issues such as long processing times, high costs, and reagent pollution, hindering the automation of ore grade and beneficiation method adjustments based on real-time grade information. In contrast, due to its efficiency, eco-friendliness, and in-situ measurement advantages, visible-near infrared spectroscopy is gradually becoming an effective alternative for estimating metal grades in mining areas. First, the raw spectral data were processed using Savitzky-Golay(SG) smoothing to reduce noise, and the spectral characteristics of gold ores were analyzed. It was found that reflectance correlates with gold grade, and a gold absorption feature is present at 455 nm. Based on this finding, dimensionality reduction was performed on the raw spectral data using principal component analysis(PCA), isometric feature mapping(ISOMAP), and locally linear embedding(LLE), with the resulting dimensions reduced to 6, 5, and 5, respectively. Finally, prediction models for gold grade were established using random forest(RF), extremely randomized trees(ET), decision trees(DT), gradient boosting decision tree(GBDT), adaptive boosting(Adaboost), extreme gradient boosting(XGBoost), and stacking ensemble learning algorithms on the dimensionally reduced data.Results indicated that the Stacking ensemble learning method outperformed single models in all aspects. Among them, the LLE-Stacking combined model achieved the highest accuracy, withR~2 of 0.972, RPD of 5.935, and an average relative error of 0.231 between predicted and actual values. The method proposed in this study allows for rapid and accurate predictions of gold content in ore, significantly improving the inversion accuracy compared to traditional models, providing new technological means for the rapid and in-situ measurement of gold grades in mines, and holding great significance for efficient gold extraction.更多还原