用人工神经网络-近红外光谱法测定毛竹中木质素与综纤维素的含量

A Determination of Lignin and Holocellulose Content of Bamboo by Artificial Neural Network and Near-infrared Spectroscopy

近红外光谱结合反向传播的人工神经网络(back propagation artificial neural network,BP-ANN)技术预测了毛竹中木质素与综纤维素的含量.用常规湿化学方法测定了54株毛竹样品的木质素含量以及53株毛竹样品的综纤维素含量.用近红外光谱仪采集相应的光谱,为了提高信噪比和计算速度,对原始近红外光谱进行平滑、压缩、归一化预处理.利用预处理后的近红外光谱数据建立BP-ANN模型.在模型建立过程中采用Leave-n-out交叉验证法优化了隐含层神经元的个数,学习率,动量因子和学习次数.优化的BP-ANN模型用于预测测试集中9个毛竹样品中木质素与综纤维素的含量,预测均方根误差分别为0.88%、1.40%.结果表明,应用毛竹的近红外光谱数据和BP-ANN技术,可以用于预测木质素和综纤维素的含量,基本能满足定量分析的要求.

Lignin and holocellulose content of bamboo were predicted by near-infrared spectroscopy and back propagation artificial neural network （ BP-ANN）. Lignin content of 54 bamboo samples and holocellulose content of 53 bamboo samples was measured according to wet-chemical method. The spectra of bamboo samples were recorded by near infrared spectrometer. In order to improve the signal-noise ratio and accelerate computation speed of neural network, the raw spectral data were pretreated by smoothing, compress and scaling. The BP-ANN models were established based on pretreated near-infrared spectroscopy of bamboo. The number of hidden neurons, learning rate, momentum, and epochs were optimized by using leave-n-out cross-validation approach during the process of establishment of neural network. The prediction root mean square errors of lignin and holocellulose contents of 9 bamboo samples test set were 0.88% , 1.40% , respectively. These results demonstrated near-infrared spectroscopy of bamboo can be used to predict the content of lignin and holocellulose based on BP-ANN method, and the predicted results can basically satisfy the requirement of quantitative analysis.