Fig. 1. Preparation flowchart of accelerated thermal aging samples
Fig. 2. Original spectra of 140 insulating oil samples
Fig. 3. Correlation between furfural content in oil and aging time
Fig. 4. The iPLS plot of original spectra
Fig. 5. Contribution rates of seven principal components of NIR spectra of insulating oil samples
Fig. 6. Prediction results of four aging time models
Fig. 7. The iPLS plot of original spectra
Fig. 8. Contribution rates of four principal components of NIR spectra of insulating oil samples
Fig. 9. Prediction results of four furfural content models
样品组别 | 老化时间/h | 糠醛含量/(mg·L-1) |
---|
1 | 48 | 0.227 8 | 2 | 72 | 0.271 4 | 3 | 96 | 0.533 9 | 4 | 144 | 0.818 2 | 5 | 216 | 1.066 7 | 6 | 288 | 1.178 4 | 7 | 360 | 1.444 8 | 8 | 432 | 1.771 2 | 9 | 528 | 1.850 6 | 10 | 648 | 2.021 1 | 11 | 768 | 2.880 6 | 12 | 888 | 2.957 4 | 13 | 1 032 | 3.575 1 | 14 | 1 176 | 4.192 8 |
|
Table 1. Furfural content and aging time of insulating oil samples
吸收峰位/cm-1 | 归属 |
---|
8 373 | 甲基C—H伸缩振动的二级倍频 | 8 264 | 亚甲基C—H伸缩振动的二级倍频 | 5 855 | 甲基C—H伸缩振动的一级倍频 | 5 799 | 亚甲基C—H对称伸缩振动和反对称伸缩振动的组合频 | 5678 | 亚甲基C—H伸缩振动的一级倍频 | 7 181, 7 076双峰 | 亚甲基伸缩振动和弯曲振动的组合频 |
|
Table 2. Attribution of absorption near-infrared spectral peaks of insulating oil
BP网络学习算法 | 隐含层神 经元数目 | 校正集 | 预测集 |
---|
RMSEC | R2 | RMSEP | R2 |
---|
L-M优化算法 | 2 | 16.47 | 0.997 9 | 20.35 | 0.996 8 | Quasi-Newton算法 | 2 | 16.47 | 0.997 9 | 20.35 | 0.996 8 | 贝叶斯正则化的L-M算法 | 2 | 19.37 | 0.997 1 | 23.39 | 0.995 7 | 共轭梯度算法 | 2 | 15.59 | 0.998 1 | 18.67 | 0.997 3 | 弹性梯度下降法 | 2 | 17.05 | 0.997 7 | 20.81 | 0.996 7 | 梯度下降法 | 3 | 18.75 | 0.997 3 | 23.06 | 0.995 9 |
|
Table 3. Predicting results of correction model with different BP learning algorithms
建模方法 | 校正集 | 预测集 |
---|
RMSEC | R2 | RMSEP | R2 |
---|
PLSR(None) | 41.89 | 0.986 9 | 43.49 | 0.986 3 | PCR | 24.16 | 0.995 5 | 27.23 | 0.994 6 | PLSR | 22.84 | 0.996 1 | 22.21 | 0.996 2 | PCA-BP-ANN | 15.59 | 0.998 1 | 18.67 | 0.997 3 |
|
Table 4. Predicting results of models with different correction methods
BP网络学习算法 | 隐含层神 经元数目 | 校正集 | 预测集 |
---|
RMSEC | R2 | RMSEP | R2 |
---|
L-M优化算法 | 3 | 0.161 0 | 0.981 7 | 0.175 9 | 0.978 6 | Quasi-Newton算法 | 2 | 0.171 9 | 0.979 2 | 0.172 5 | 0.979 4 | 贝叶斯正则化的L-M算法 | 2 | 0.155 0 | 0.983 1 | 0.194 0 | 0.973 7 | 共轭梯度算法 | 3 | 0.116 0 | 0.990 5 | 0.134 4 | 0.987 7 | 弹性梯度下降法 | 2 | 0.164 0 | 0.981 0 | 0.187 2 | 0.975 4 | 梯度下降法 | 2 | 0.162 7 | 0.981 3 | 0.185 4 | 0.975 9 |
|
Table 5. Predicting results of correction models with different BP learning algorithms
建模方法 | 校正集 | 预测集 |
---|
RMSEC | R2 | RMSEP | R2 |
---|
PLSR(None) | 0.187 0 | 0.975 3 | 0.206 6 | 0.969 9 | PCR | 0.148 7 | 0.984 4 | 0.181 0 | 0.977 2 | PLSR | 0.140 1 | 0.986 2 | 0.148 7 | 0.984 6 | PCA-BP-ANN | 0.116 0 | 0.990 5 | 0.134 4 | 0.987 7 |
|
Table 6. Predicting results of models with different correction methods