采用KPCA特征提取的近红外煤炭发热量预测模型

NIRS prediction model of calorific value of coal with KPCA feature extract

近红外光谱分析技术通过搭建基于GA-BP神经网络方法的定量分析模型,实现煤炭发热量的快速评估。为了提高模型的学习速度和精度,必须对光谱信息进行数据处理。该过程属于复杂的非线性问题,经典的线性主成分分析方法具有一定的局限性,因此采用了一种基于多项式核主成分分析特征提取方法。通过分析主成分的特征值筛选异常样本。实验结果表明,该方法提取的特征信息主成分集中度高、降维效果明显、与输出变量间的相关性好,且能够准确判断出异常样本,大幅度提高了模型的准确性,为近红外煤质分析模型提供了一种分析速度快、准确率高的有效数据处理方法。

The near-infrared spectroscopy（NIRS）with GA-BP neural network model was used for rapid prediction of gross calorific value of coal.The prediction model was non-linear,so the classical linear principal component analysis（LPCA）method was not applicable for processing spectral data.The nonlinear polynomial kernel principal component analysis（P-KPCA）method was proposed for extracting spectral feature and filtering abnormal samples in this paper.The extracted principal components had high feature concentration,obvious dimension reduction effect,and good correlation with output variable.After eliminating the abnormal spectra,prediction accuracy was greatly improved.The results showed that P-KPCA provided effectively processed spectral data for the rapid prediction model.