四种算法用于近红外测定制浆材材性的对比研究

采集了常见制浆材(桉木、相思木及杨木)样品的近红外光谱,测定了样品的基本密度、综纤维素、木质素和苯醇抽出物含量,用人为控制水分的方法测定了样品的水分含量。对原始光谱进行预处理后,分别运用偏最小二乘法(PLS)、LASSO算法、支持向量机法(SVR)和人工神经网络法(BP-ANN)建立基本密度、水分含量、综纤维素、木质素和苯醇抽出物含量的预测模型。对预测模型进行独立验证,结果显示:LASSO算法建立的基本密度和综纤维素模型性能最优,其预测均方根误差(RMSEP)分别为0.006 3 g/cm~3和0.49%,绝对偏差(AD)范围分别为-0.008 8~0.009 6 g/cm~3和-0.85%~0.87%;PLS建立的水分含量模型及苯醇抽出物模型最优,RMSEP值分别为1.21%和0.24%,AD范围分别为-1.99%~2.03%和-0.35%~0.38%;SVR建立的木质素模型最优,RMSEP值为0.43%,AD范围为-0.76%~0.74%,均满足制浆造纸工业中对误差的要求。

Four Kinds of Algorithms Used for the Determination of Pulpwood Properties by Near Infrared Spectroscopy

Near infrared( NIR) spectra of pulpwood species were collected. The basic density, holocellulose, lignin and benzene-alcohol extractive content of samples were analyzed by traditional methods. The moisture content under manual control was analyzed,too. After the pretreatment of the original spectra, partial least squares( PLS) algorithm, LASSO algorithm, support vector regression(SVR) algorithm and back propagation artificial neural network(BP-ANN) algorithm were used to build the prediction models for basic density, moisture content, holocellulose, lignin and benzene-alcohol extractive content. The independent verification of the prediction models showed that the optimal model for basic density was built by LASSO algorithm with the root mean square error(RMSEP) of 0. 006 3 g/cm3 and the absolute deviation(AD) of -0. 008 8-0. 009 6 g/cm3. The optimal model for moisture content was built by PLS algorithm with the RMSEP of 1. 21% and the AD of -1. 99% -2. 03%. The optimal model for holocellulose content was built by LASSO algorithm with the RMSEP of 0. 49% and the AD of -0. 85% -0. 87%. The optimal model for lignin content was built by SVR algorithm with the RMSEP of 0. 43% and the AD of-0. 76% -0. 74%. The optimal model for the benzene-alcohol extractive content was built by PLS algorithm with the RMSEP of 0. 24% and the AD of -0. 35% -0. 38%. The prediction performance of the models could meet the needs of pulping and papermaking industry. The detemination accuracy of pulpwood properties were promoted by algorithm selection.