铁矿粉中全铁含量的SFIM-RFR高光谱预测模型  

Hyperspectral SFIM-RFR Model on Predicting the Total Iron Contents of Iron Ore Powders

作　　者：高伟 杨可明[1] 李孟倩 李艳茹 韩倩倩 GAO Wei;YANG Ke-ming;LI Meng-qian;LI Yan-ru;HAN Qian-qian(College of Geoscience and Surveying Engineering,China University of Mining&Technology(Beijing),Beijing 100083,China;North China University of Science and Technology,Tangshan 063210,China)

机构地区：[1]中国矿业大学(北京)地球科学与测绘工程学院,北京100083 [2]华北理工大学,河北唐山063210

出　　处：《光谱学与光谱分析》2020年第8期2546-2551,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(41971401)资助。

摘　　要：铁矿是全球储量最高的金属矿产之一。全铁含量是评价铁矿石、铁矿粉品质的重要指标,在铁矿开采、矿石精选、矿粉冶炼等生产环节中有特殊意义。传统的铁矿粉全铁含量化学分析方法存在耗时久、操作复杂、污染严重等缺点,因此,探寻一种快速、有效、无污染的检测方法越来越成为矿山环境的研究热点。高光谱技术具有光谱分辨率高、曲线连续、无损伤、无污染、可对物质特征或成分进行精确探测等特点。使用铁矿粉高光谱数据,通过建立用于光谱特征筛选的光谱特征重要性评分(SFIM)指标,并结合随机森林回归(RFR)方法构建铁矿粉全铁含量预测的SFIM-RFR模型。以河北省阳原县三义庄铁矿为研究区,于2018年11月与2019年3月在研究区收集铁精粉、铁尾砂原料,分别制作第一批次的训练组和验证组铁矿粉试样以及第二批次的二次验证组铁矿粉试样,并使用ASD Field Spec4型光谱仪测量试样的光谱反射率;然后使用第一批次的训练组光谱数据训练SFIM-RFR模型,对第一批次的验证组样本的全铁含量进行预测,同时采用常规RFR、线性回归(LR)预测模型来对比分析铁矿粉样本全铁含量预测结果;最后使用二次验证组光谱数据检验多模型鲁棒性。结果表明:SFIM-RFR, RFR和LR模型全铁含量预测结果与2018年11月采集的验证组样本全铁含量真实值的确定系数(R-Square)分别为0.991 8, 0.988 4和0.898 7,均方根误差(RMSE)分别为0.016 9, 0.020 1和0.059 6,多模型预测效果总体较好, SFIM-RFR模型预测结果误差最小,说明了SFIM-RFR模型用于预测铁矿粉中全铁含量的可行性和有效性,且SFIM-RFR模型预测效果优于常规的预测模型;SFIM-RFR, RFR和LR模型全铁含量预测结果与2019年3月采集的二次验证组样本全铁含量真实值的R-square分别为0.976 8, 0.974 5和0.914 0, RMSE分别为0.034 6, 0.036 2和0.071 9,证明了SFIM-RFR模型的预测效果�Iron ore is one of the most abundant metallic minerals in the world. Total iron contents is an important index to evaluate the quality of iron ore and iron ore powder, and it has a special significance in iron ore mining, ore dressing, ore smelting and other production links. The traditional chemical methods have the disadvantages of a time-consuming, complex operation, seriously pollution. Therefore, exploring a new method of rapid, effective and pollution-free detection has become a hot spot in mine environment research. Hyperspectral technology has the characteristics of high spectral resolution, continuous curve, no damage, no pollution and accurate detection of characteristics or components of materials. The purpose of this paper is to establisha data evaluation index of spectral feature importance measures(SFIM) and to screen spectral features based on the hyperspectral data of iron ore powder, and then combined with random forest regression(RFR) to establish the SFIM-RFR prediction model and predict the total iron contents of iron ore powder. First, taking Sanyizhuang iron mine in Yangyuan county, Hebei province as a research object, based on the iron concentrate and iron powder tail collected in the research area in November 2018 and March 2019, the first batch of iron ore powder samples in the training group and the testing group and the second batch of iron ore powder samples in the second testing group were made respectively. Spectral data of samples were measured by the ASD Field Spec4 spectrometer. Then, spectral data of the first batch of training group were used in the SFIM-RFR model training, and the total iron contents in the samples of the first batch of the testing group were predicted. Meanwhile, conventional methods, including RFR and linear regression(LR) prediction model, were used to compare and analyze the predicted results of total iron contents in iron ore powder samples. Finally, the spectral data of the second testing group were used to validatethe robustness of the multi-model. The r

关 键 词：高光谱 铁矿粉全铁含量 预测模型 光谱特征重要性评分 随机森林回归 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]