为解决速生桉抽出物测定方法繁琐耗时, 木浆生产能耗居高不下等问题, 以引种的3种广西速生尾巨桉原料(DH32-29, DH32-26, DH33-27)为研究对象, 采集了144个样本的近红外光谱, 按国标方法测定全部样品的苯醇抽出物和1%NaOH抽出物含量。 在Matlab 8.0中采用信号平滑, 一阶、 二阶导数, 矢量归一化, 多元散射校正等方法预处理原始光谱, 用偏最小二乘法、 支持向量机法和人工神经网络法以及常用于宏观经济分析的LASSO法分别结合上述预处理方法建立模型, 筛选出最优建模方法。 运用遗传算法对波段进行选择, 提高了模型的精确度从而优化了模型。 确定了建立苯醇抽出物含量模型时, 可联用平滑、 MSC和一阶导数预处理光谱数据, 以1 345.0～1 821.4和2 127.8～2 241.3 nm区间波段参与建模, 建模方法为偏最小二乘法, 最佳主成分数为9时, 模型有最好的精确度。 其RMSEP值可达0.25%, 绝对偏差范围为-0.39%～0.38%。 筛选出的波段包含了如1 410和1 447 nm附近酚羟基伸缩振动的一级倍频, 2 133 nm处苯环上碳氢键的伸缩振动与碳碳双键伸缩振动的合频等苯醇抽出物的特征波段。 建立1%NaOH抽出物分析模型时, 可联用平滑、 矢量归一化和一阶导数预处理, 选择1 138.2～2 363.0 nm波段数据, 建模方法为LASSO, 选取的调整参数值μ为12.61, 此时模型精确度最高。 RMSEP值为0.37%, 绝对偏差范围为-0.56%～0.53%。 筛选出的波段包含了1 158和1 170 nm附近乙酰脂基团CH3中C—H的伸缩振动二级倍频吸收, 1 666, 1 681和1 790 nm附近CH3中C—H伸缩振动的一级倍频吸收等特征吸收。 模型的预测能力从组分结构角度得到了解释。 模型的RPD值分别为4.67和5.77, 模型性能均可满足实际需求, 有望应用于制浆造纸生产线上的速生桉抽出物含量分析。 研究结果表明, 通过预处理方法选择和建模方法选择, 结合遗传算法的应用, 可以建立并优化广西速生桉木抽出物含量的近红外测定模型; 同时, LASSO算法作为一种新兴算法, 在近红外光谱分析中表现出了较好的处理共复线性数据的能力, 可以建立准确性较好的分析模型。

In pulping and papermaking industry, extractives of wood chips influence the impregnation efficiency, pulp energy consumption and pulp yield. But traditional analysis methods for the content of extractives are not applicable for industrial online monitoring because of being time consuming and costly. Therefore, the present study used near infrared (NIR) spectroscopy to predict rapidly extractives content of three species of fast-growing Eucalyptus urophylla×E.grandi chips (DH32-29, DH32-26, DH33-27) grown in China’s Guangxi Province. NIR spectra of 144 fast-growing Eucalyptus were collected using a holographic grating spectrometer equipped with a halogen illumination and array detector. The benzene-alcohol extractives and 1% NaOH extractives content of 144 samples were gravimetrically determined according to the Chinese national standard test method respectively. The near-infrared spectrum were pretreated using smoothing, first derivative, second derivative, vector normalization and multivariate scattering correction in Matlab 8.0, and the models were developed for various pretreatment methods by loading PLS, LASSO, SVR and ANN algorithm. The optimal modeling methods were selected. Genetic algorithm was used to select the bands, which improved the accuracy of the models and optimized the models. In conclusion, in order to develop analysis model of benzene-alcohol extractives, smoothing, MSC and first derivative methods should be used to preprocess the original spectrum, the bands of 1 345.0~1 821.4 and 2 127.8~2 241.3 nm were selected, meanwhile, the partial least squares algorithm was used with the optimal factor 9. The model had the best accuracy for the RMSEP value as low as 0.25%, and the absolute deviation range was -0.39%~0.38%. The optimal bands between 1 345.0~1 821.4 and 2 127.8~2 241.3 nm have been associated with O—H stretching (1st overtone) of phenolic compound (1 410 and 1 447 nm), as well as C—H stretching and CC stretching group frequencies of benzene ring (2 133 nm) and other characteristic absorption. In order to establish the content analysis model of 1% NaOH, smoothing, vector normalization, first derivative should be used to pretreat the original data, the bands between 1 138.2~2 363.0 nm were picked and LASSO was adopted. The model had the best accuracy when the μ value was 12.61, the independent verification show the RMSEP value was 0.37%, and the absolute deviation range was -0.56%~0.53%. The optimal bands between 1 138.2~2 363.0 nm have been associated with C—H stretching (2nd overtone) of COCH3 (1 158 and 1 170 nm), as well as C—H stretching (1st overtone) of —CH3 (1 666, 1 681 and 1 790 nm) and other characteristic absorption. The characteristic absorption of benzene-alcohol extractives and 1% NaOH extractives on the optimal bands was analyzed from the point of view of molecular structure, and the performance of models was explained theoretically. The models can meet the actual demand and can be applied to the analysis of the content of Eucalyptus extractives in pulping and papermaking industry. The results showed that performance of near-infrared models can be developed and optimized by the selection of pretreatment and modeling methods combined with the genetic algorithm for the prediction of Eucalyptus extractives. At the same time, as an emerging algorithm, LASSO algorithm has a good ability to process co-complex linear data in near-infrared spectroscopy, and can establish models with good analysis performance.