Ding Li, Xiang Yu-Hong, Huang An-Min, Zhang Zhuo-Yong
Department of Chemistry, Capital Normal University, Beijing 100037, China.
Guang Pu Xue Yu Guang Pu Fen Xi. 2009 Jul;29(7):1784-7.
The amount of holocellulose, lignin, and microfibril angle of Chinese fir was predicted by using back-propagation neural network (BP-ANN) combined with near infrared (NIR) spectrometry. First, the data of original spectra were pretreated by Savitzky-Golay smoothing algorithm and the second derivative, then the data of near infrared spectrometry with 171 points were compressed to 86 points by using wavelet transform, and finally, the models were established by using BP-ANN. The models were validated using leave-n-out cross-validation approach, and the influences of the number of hidden neurons, learning rate, momentum, and epochs were discussed in the present paper. The prediction samples, which were not used in the model generation, were predicted by using the obtained models, the correlation coefficients (R2) of holocellulose, lignin and microfibril angle were 0.91, 0.90 and 0.87, respectively. The root mean square errors of prediction (RMSEP) of the established models were 0.86%, 0.33%, and 4.9%, respectively. The obtained results showed that the method is fast and nondestructive and can basically satisfy the requirement of quantitative analysis.
采用反向传播神经网络（BP-ANN）结合近红外（NIR）光谱法预测杉木的综纤维素、木质素含量及微纤丝角。首先，利用Savitzky-Golay平滑算法和二阶导数对原始光谱数据进行预处理，然后采用小波变换将171个点的近红外光谱数据压缩至86个点，最后利用BP-ANN建立模型。采用留一法交叉验证方法对模型进行验证，并讨论了隐藏神经元数量、学习率、动量和训练轮次的影响。利用所建模型对未参与模型建立的预测样本进行预测，综纤维素、木质素和微纤丝角的相关系数（R2）分别为0.91、0.90和0.87。所建模型的预测均方根误差（RMSEP）分别为0.86%、0.33%和4.9%。结果表明，该方法快速、无损，基本能满足定量分析要求。
