基于核主成分分析和粒子群优化神经网络的充填体强度预测

Prediction of compressive strength of backfill based on analysis of principal components of kernel and PSO artificial neural network

以磷石膏、粉煤灰、磷渣等为主要原材料制备磷石膏胶结充填体,分析了主要原材料的理化特性,测定了磷石膏胶结充填体28d单轴抗压强度。通过核主成分分析对磷石膏胶结充填体单轴抗压强度影响因子进行非线性特征提取,基于获得的主成分构建粒子群优化BP人工神经网络(PSO-BP-ANN)预测模型。结果表明,核主成分分析能较好地实现充填体单轴抗压强度影响因子非线性特征的提取和降维的目的,同时,PSO-BP-ANN模型训练和预测值可决系数分别为0.995和0.991,均方根误差分别为3.660E-4和5.805E-2,平均相对误差分别为1.699%和3.602%,总体性能表现优于传统BP-ANN模型,对矿山充填体的配比设计具有指导意义。

The particle swarm optimization based BP artificial neu ral network （PSO-BP-ANN） model was made to predict the 28 days compressive strength of phosphogypsum backfill ,which con sists of phosphogypsum, fly ash, phosphorus slag and other mate rials. Meanwhile, the physicochemieal property research and uniaxial compression test were conducted. The PSO-BP ANN model was constructed based on the principal components obtained by the analysis of principal components of kernel to the impact factors of compressive strength of phosphogypsum backfill. The results indi cate that KPCA can extract nonlinear features of the impact factors of compressive strength of phosphogypsum backfill well, and the coefficients of determination of training and prediction values with PSO-BP-ANN model are 0. 995 and 0. 991, root-mean-square er rors are 3. 660E-4 and 5. 805E-2, average relative errors are 1. 699 and 3. 602 respectively. Above all, the PSO-BP-ANN model per forms better than BP-ANN model, that has guning significance to the matching design of backfill in mines.