木材抗弯强度是评价木材力学性质的重要指标, 其快速准确预测具有工程应用价值和科学意义。 重点研究了使用近红外光谱分析光谱特征优选的卡尔曼滤波(KF)方法进行PLS建模, 完成木材抗弯强度的预测。 试验用126个蒙古栎无疵试样, 依据国家标准《木材物理力学性质试验方法》测量抗弯强度得到力学真值; 在900～1 700 nm波段进行近红外光谱采集, 一阶导数与S-G卷积结合进行光谱预处理; 然后, 将光谱及抗弯力学样本视为动态系统, 光谱冗余波长视为噪声信号, 通过KF迭代得到系数矩阵和标准方差, 并运用二者比值实现特征优选; 最后建立蒙古栎的偏最小二乘(PLS)抗弯强度近红外模型。 结果表明, 经过KF优选后, 光谱变量数由117减小到18个, 预测模型的相关系数ｒ=0.81、 预测误差均方根RMSEP=6.59; 为了进一步验证方法有效性, 与无信息变量消除法(UVE)、 连续投影方法(SPA)特征选择方法进行了对比, KF特征优选后的预测相关系数r分别提高了0.05和0.16, 预测误差均方根RMSEP降低了2.33和7.66, 采用KF特征选择建立的模型预测结果最佳。 KF作为特征方法可有效选择近红外光谱特征波长, 降低模型维度, 提高模型的适用性与准确性。

The bending strength is an important index to evaluate the mechanical properties of wood, and the rapid and accurate prediction of its nature is a scientific problem with engineering application value. In this paper, the wood bending strength is predicted by near infrared spectroscopy (NIR), combined with Kalman filter (KF) and partial least squares method (PLS). A total of 126 samples of Mongolian oak (Quercus mongolica) were used, and their bending strengths were measured according to the national standard “Wood physical and mechanical properties test method”. In addition, NIR spectra were collected in the wavelengths ranging from 900 to 1 700 nm, and a pretreatment for NIR was carried out by the first order derivative combined with S-G convolution. Then, the spectrum and bending strength samples were considered as a dynamical system, the redundancy spectrum wavelength points were considered as noise signals. Besides, coefficient matrix and standard deviation were acquired by means of KF iteration, and feature selection was achieved by the ratio of coefficient to standard deviation. Finally, the prediction model of wood bending strength was build based on PLS and the selected wavelength points. The result shows that the number of variables is reduced from 117 to 18 after the KF selection, and the correlation coefficient R of the prediction model is 0.81, the root mean square error of prediction (RMSEP) is 6.59. In order to validate the effectiveness of KF, UVE and SPA were used to make a comparison, the correlation coefficient r is improved by 0.05 and 0.16 and the RMSEP is reduced by 2.33 and 7.66 respectively, which can show that KF can be used to select effective spectrum points, reduce the model dimension, and improve the applicability and accuracy of the model.