成品油混合浓度的预测对成品油顺序输送过程中的安全监控、混油段分割具有重要的意义.本研究配制92#汽油-3#航煤以及3#航煤-0#车柴两组包含不同浓度的混合样品,并对其进行拉曼光谱采集;依次采用归一化、多元散射校正、BaselineWavelet基线校正3种光谱预处理方法进行优化;之后采用改进的栈式稀疏自编码器(Stacked Sparse Autoencoder,SSAE)模型对预处理之后的拉曼光谱进行稀疏特征提取,并结合全连接层进行回归预测;最后根据均方根误差(Root Mean Square Error,RMSE)和决定系数(R2)两项评价指标,与偏最小二乘回归(Partial Least Square Regression,PLSR)、最小二乘支持向量回归(Least Square Support Vector Machine,LSSVR)以及SSAE 3种模型进行对比.结果表明:改进的SSAE-FC模型表现出更优的预测精度和稳定性,92#汽油-3#航煤混油测试集的R2和RMSEC指标分别为0.9952和0.8932,3#航煤-0#车柴混油测试集的R2和RMSEC指标分别为0.9837和1.1967,且学习得到的稀疏特征的可解释性强.
Prediction Method of Mixed Concentration of Refined Oil Based on Raman Spectroscopy Combined with Improved Stacked Sparse Autoencoder Feature Optimization
The prediction of mixed concentration of refined oil is of significant importance when considering safety monitoring and mixed oil section segmentation in the process of batch transportation of refined oil.In this study,two groups of mixed samples containing different concentrations were prepared,92 # gasoline-3 # jet fuel mixture and 3 # jet fuel-0 # diesel mixture respectively;then collected Raman spectra and applied three spectra preprocessing methods in sequence:normalization,multiple scattering correction,baseline wavelet correction;after that,the improved stacked sparse autoencoder(SSAE)was proposed to extract the sparse spectra features from preprocessed spectra and predict results combining with fully connected network;finally the improved model was compared with PLS,LSSVR and SSAE three other classical regression models based on RMSE and R2 evaluation indics.The results show that the improved SSAE model exhibits best accuracy and stability,the RMSE and R2 indexes of 92 # gasoline-3 # jet fuel validation set are 0.9952 and 0.8932 respectively,the RMSE and R2 indexes of 3 # jet fuel-0 # diesel validation set are 0.9837 and 1.1967 respectively,furthermore,the spectral features obtained from model were highly interpretable.
万方数据
维普
