针对实木板材表面存在的活节、 死节、 裂纹与虫眼4类缺陷, 提出了基于近红外光谱分析的定性识别模型。 随机选取50个样本组成训练集, 30个样本组成测试集, 在室内温度20 ℃、 相对平均湿度50%环境下, 采用900~1 700 nm的近红外光谱仪采集样本表面光谱, 并利用SNV方法进行光谱数据预处理, 以消除固体颗粒大小、 表面散射及光程变化对漫反射光谱的影响； 然后, 采用偏二叉树双支持向量机(PBT-SVM)构建缺陷分类模型, 运用模拟退火算法(SA)对4类核函数、 参数及波长特征进行全局寻优； 寻优过程以97个波长吸收度为输入特征, 运用顺序前向法依次加入新特征, 当分类器准确率达到90%时, 得到核参数及波长特征； 最后, 通过确定的核函数、 参数与波长构建了缺陷分类模型, 并对测试样本集进行了分类验证。 实验结果表明, SNV预处理方法使相同缺陷的近红外光谱具有较好的一致性, 其中, 活节与死节光谱差异显著, 但死节、 裂纹与虫眼的光谱趋势相近； 当PBT-SVM分类器采用多项式核函数、 参数在γ=28.63, coef=18.69, d=1, C=12.03时, 缺陷识别效果最好, 裂纹和活节的识别率达到了100%, 虫眼为93.33%, 死节为93.33%, 平均准确率达到了96.65%, 平均识别时间仅为0.002 s。 利用近红外光谱分析的方法能够快速、 有效地完成4类实木板材缺陷的识别。

In this paper, near infrared spectroscopy was applied to build an identification model to predict four types of defects on the surface of wood boards. A calibration set and a prediction set made of 50 and 30 samples were built randomly and respectively. In addition, a near infrared spectrometer, ranging from 900 to 1 700 nm was used to collect the spectra of the surface of the boards. The original spectra were pre-treated by SNV algorithm to eliminate the influence of solid particle size, surface scattering, and the change of optical path of diffused reflectance spectra. Afterwards, a training model was built by partial binary tree of support vector machine (PBT-SVM), and parameters were optimized by simulated annealing (SA) algorithm to find the optimal parameters and band characteristics. Then an identification model was built based on optimal parameters, band characteristics, and the identification of prediction set. The results showed that the performance of polynomial kernel function was obtained with the parameters setting as γ=28.63, coef=18.69, d=1 and, C=12.03. The recognition rate of crack and live knot was 100%, while the recognition rate of dead knot and wormhole was 93.33%. The mean accuracy of identification reached 96.65% with an average recognition time of 0.002 s. The approach was feasible to classify the four types of defects on the surface of solid wood effectively.