铜精矿是冶炼铜及其合金的基础工业原料, 不同产地进口的铜精矿在元素组成、 含量上存在着差异, 进口铜精矿伪报、 掺杂、 有害元素超标案件多发, 危害国家经济安全。 因此建立入境铜精矿产地识别模型, 将有助于风险分级, 预警。 该研究对象为智利、 秘鲁、 菲律宾、 西班牙、 纳米比亚、 伊朗、 马来西亚和阿尔巴尼亚8个国家进口铜精矿的280批次铜精矿样品。 应用波长色散-X射线荧光光谱无标样分析法测定所有研究样品的元素组成及含量, 结果表明铜精矿样品的检出元素共计53种。 选择O, Mg, Al, Si, P, S, K, Ca, Ti, Fe, Cu, Zn, Mn, As, Mo, Ag和Pb共17种元素含量作为变量, 建立进口铜精矿国别的BP神经网络预测模型。 采用F-score筛选出O, Mg, Al, Si, P, S, K, Ca, Cu, Zn, Mo, Ag和Pb共13个元素的含量作为特征变量, 分别建立进口铜精矿国别的Fisher判别分析预测模型和BP神经网络预测模型。 3种预测模型的结果如下: 采用F-score筛选变量的Fisher判别分析模型对建模样品的识别准确率为94.2%, 交叉验证准确率为92.9%, 对预测样品的识别准确率为96.7%; 输入层为17与13个变量的BP神经网络的训练集, 校正集, 验证集, 建模集以及预测样品的准确识别率分别为: 100%, 97.1%, 94.1%, 98.2%, 100%与100%, 97.1%, 100%, 99.6%, 100%。 比较3次建模的结果可知, 经过F-score筛选变量后用BP神经网络所建模型的准确识别率最高, 该方法不仅可以减少建模的输入变量还可以提高识别准确度。 波长色散-X射线荧光光谱无标样分析虽是半定量分析方法, 但具有分析速度快和稳定性好的优点, 利用该方法结合F-score筛选变量用于BP神经网络模式识别可以实现对铜精矿的国别识别。

Copper concentrate is the basic industrial raw material for smelting copper and its alloys. Imported Copper concentrate with different origins varies in elemental composition and content. Cases of imported copper concentrate falsifying, adulterating and exceeding the standard of harmful elements frequently occur, which endangers national economic security. So it is necessary to establish a rapid identification model of the origin of imported copper concentrates in major importing countries, which is conducive to risk classification and early warning. The research objects of this paper are 280 imported copper concentrate samples from 8 countries in Chile, Peru, Philippines, Spain, Namibia, Iran, Malaysia and Albania. The elemental composition and content of all research samples were determined by wavelength dispersive X-ray fluorescence spectrum standard-less analysis method, and it turned out that elements detected from copper concentrate samples are 53 in total. Among them, we chose 17 elements and conducted a BP neural network prediction model, including O, Mg, Al, Si, P, S, K, Ca, Ti, Fe, Cu, Zn, Mn, As, Mo, Ag, Pb. Moreover, 13 elements including O, Mg, Al, Si, P, S, K, Ca, Cu, Zn, Mo, Ag, Pb were screened out as valid variables by F-score, and the Fisher discriminant analysis prediction model and BP neural network prediction model were established for importing copper concentrate countries respectively. The results of the three prediction models are as follows: the Fisher discriminant analysis model, which uses F-score to screen variables, the recognition accuracy of the model for the modeled sample was 94.2%, the one of cross-validation was 92.8%, and that of the predicted sample reached 96.8%. The accuracy rate of training set, calibration set, validation set, modeling set and prediction sample of BP neural network with input layer of 17 and 13 variables is: 100%, 97.1%, 94.1%, 98.2%, 100% and 100%, 97.1%, 100%, 99.6% and 100%, respectively. Comparing the results of three times of modeling can be seen that the model established by BP neural network has the highest accurate recognition degree after the variables are screened by f-score. This method can not only reduce the input variables of modeling, but also improve the recognition accuracy. Although the wavelength dispersion X-ray fluorescence spectrum standard-less analysis method is a semi-quantitative analysis method, it has the advantages of fast analysis speed and good stability. The country identification of copper concentrate can be realized by using this method combined with F-score screening variables for BP neural network pattern recognition.