NIRS中联用EEMD与SPA建立变压器油中微水的LS-SVM回归模型

A joint EEMD and SPA approach for LS-SVM modeling of moisture content in transformer oil using NIRS

采集不同微水含量的变压器油的近红外光谱,利用集合经验模分解(EEMD)与连续投影算法(SPA),建立变压器油中微量水分的最小二乘支持向量机(LS-SVM)回归模型。结果表明,原始求导光谱经EEMD分解后得到8个本征模态函数(IMF),在去掉第一个IMF后重构数据比原始求导光谱数据直接建模具有较好的效果,而利用去掉第一个IMF后重构数据经SPA筛选出的4个特征光谱(只占全谱的0.78%)来建模则具有更好的预测效果,预测均方根误差为1.04776×10-3,预测相关系数为0.9840,说明EEMD与SPA联用具有比EEMD及SPA单独运用更好的效果,且最优模型应用于实际油品的检测同样具有很好的效果,对实现油中水分的高精度检测以及低维度变量建模具有实际意义。

By collecting the NIRS（near-IR-spectrum） of transformer oils with different moisture contents,a joint ensemble empirical mode decomposition（EEMD） and successive projections algorithm（SPA） approach is proposed for least squares support vector machine（LS-SVM） modeling of moisture contents in the oils.The results show that,the modeling precision using the reconstructed data by removing IMF1（Intrinsic Mode Functions） from 8 IMFs of the EEMD is better than that using the 1st-derivative spectra.Moreover,superb predict effect（RMSEP = 1.04776 × 10-3 and R = 0.9840） appears at modeling process using 4 characteristic spectral data selected by SPA（Successive Projections Algorithm） from the former reconstructed data.The results also indicate that the present joint method is better than the sole EEMD and SPA approach.What is more,good effect also shows in actual oil sample tests.The suggested approach has practical significance in terms of both better precision and lower dimension for modeling the moisture contents in transformer oils.