品位估值的自适应径向基神经网络构建技术

在简要分析常用储量计算方法与BP神经网络预测方法存在缺陷的基础上，分析了径向基神经网络隐层节点参数在映射机理上与地质统计学方法理论上的一致性，以及其权系数能解析方式求解、可避免网络训练过程陷入局部最优乃至不收敛现象的特征，提出了构建径向基函数神经网络进行矿床品位估值模型的研究思路。通过多方案分析，得出了待估点三维坐标及周围样品点个数是影响径向基函数神经网络模型估值精度的主要因素，给出了输入节点变量空间的基本配置方式--3个坐标加周边8个样品点品位。针对实际工程中样品空间较大的特征，分析了隐层中心、宽度等参数需根据输入变量自适应构造的必要性，以及利用正则化正交最小二乘的前向选择法的可行性。利用开发的具备用户自定义和交互式输入参数的计算机软件，构造了两种不同的品位估计模型。验证试验表明：基于样本空间自适应构建的径向基函数神经网络，建模速度快、可靠性强，平均估值误差最大为309%，且正则化参数对模型的估值精度影响较大，考虑了该参数的模型估值效果更好。

RESEARCH ON THE ADAPTIVE RBF NEURAL NETWORK MODELING FOR THE DEPOSIT GRADE ESTIMATION

In this paper the limitation and deficiency of the conventional reserve methods and the BP neural networks for reserves prediction were briefly analyzed. Then based on the analysis results shown that the parameters of hidden RBF nodes of RBF neural network were coincident with geostatistics techniques on the architecture of input-output, and the weights of RBF model could be solved though analytical approach, therefore, the training process can avoid being trapped into local optimal and non-convergence, a self-adaptive radius basis function (RBF) neural network for grade estimation was presented. After multi schemes analysis, it was found that 3D coordinates of estimating location and amount of neighbor samples were the primary factors for the estimation accuracy of RBF network. Accordingly, the basic configuration of input node variables was proposed, that was, three coordinates and eight neighbor samples’ grade values. For the sample space was relatively big in practical application, the necessity for adaptively constructing the center and width of RBF in the hidden layer and the feasibility of forward selection utilizing regularized orthogonal least squares were analyzed. Computer software package with both user defined and interactive input parameters was developed, and two grade estimation models were constructed by using the package. The results shown that the adaptive RBF neural network based on sample space was fast and robust. The maximum average estimation error was 3.09%. Moreover, the regularization parameter had more effect on the estimation accuracy and brought on better estimation result.