CAJ | 학술논문

根据汽油辛烷值预测体系本身的非线性特点，提出主成分回归残差神经网络校正算法（principal component regressionresidual artificial neural network,PCRRANN)用于近红外测定汽油辛烷值的预测模型校正。该方法结合了主成分回归算法（PC)，与经典的线性校正算法PLS（Partial Least Square）,PCR,以及非线性PLS（NPLS,Non-linearPLS）等相比，预测能力有明显的改善。文中还讨论了PCR主成分数及训练参数对预测模可能的影响。
근거기유신완치예측체계본신적비선성특점，제출주성분회귀잔차신경망락교정산법（principal component regressionresidual artificial neural network,PCRRANN)용우근홍외측정기유신완치적예측모형교정。해방법결합료주성분회귀산법（PC)，여경전적선성교정산법PLS（Partial Least Square）,PCR,이급비선성PLS（NPLS,Non-linearPLS）등상비，예측능력유명현적개선。문중환토론료PCR주성분수급훈련삼수대예측모가능적영향。
A novel calibration algorithm, PCRRANN (principal comp onentregression residual artificial neural network) method, was proposed based o n the intrinsic non-linearity of the prediction of gasoline octane number, and then applie d to the calibration of the prediction model of the near infra-red measurement of gasoline octane number. The method combined the linear calibration abilit y of the pricipal component regression (PCR) method and the excellent non-line ar approximating ability of artificial neuralnetwork using the residual of P CR calibration as target signal and the PCR scores as input signal of the neuralnetwork respectively. Compared with the classical linear algorithms such as the PLS (partial least squares), PCR and NP LS (Non-linear PLS), the proposed method showed obvious improvement in predicti on ability. The effects of the number of principal components of PCR part and s ome training parameters on the prediction model were also discussed.