遗传神经网络-X射线荧光光谱法测定铁矿石中铅砷

摘要

In order to determine the content of trace lead and arsenic in iron ore,the sample was grinded by planetary ball mill.Then it was pre-fused at 800 ℃ followed by fusion at 1 050 ℃ for 5 min.After cooling,the sample was re-fused for 8 min to prepare the stable fuse pieces with low dilution ratio (the mass ratio of sample and flux was 1 ∶ 2).The fluorescence intensity of lead and arsenic in fuse pieces of synthetic sample was scanned with standard-less analysis software.The X-ray fluorescence intensity data obtained under different scanning angles were used as the input of neural network.Meanwhile,the content of lead and arsenic was used as output.The weight and threshold values of network were optimized by genetic algorithm.The spectral overlapping of lead and arsenic in sample of test set was corrected,solving the problem of high background intensity due to low dilution ratio.The root-mean-square error of calibration (RMSEC) for lead and arsenic content in prediction set was 0.39 and 0.42,respectively.The correlation coefficients were both 0.98,which had no significant difference with the theoretical α coefficient regression equation method.%为测定铁矿石中微量的铅和砷,采用行星球磨机研磨铁矿石样品,经800℃预熔融后,于1 050℃熔融5 min,冷却后再次熔融8 min的方法制备了稳定的低稀释比熔片(样品与熔剂的质量比为1∶2).采用无标样分析软件对合成样品熔片中铅和砷元素荧光强度进行扫描,不同扫描角度下的X射线荧光光谱强度数据作为神经网络的输入,铅、砷元素含量作为输出,用遗传算法对网络的权值和阈值进行优化,对测试集中样片的铅、砷谱线重叠进行校正,克服了低稀释比导致的背景强度高的缺点.对预测集中铅、砷元素含量模型预测的均方根误差(RMSEC)分别为0.39和0.42,相关系数均为0.98.可见实验方法与理论α系数回归方程法没有明显区别.