基于近红外光谱技术的针叶材板材表面节子缺陷检测

摘要

为了实现木板材依据节子进行自动化分级,采用近红外光谱技术研究了多种针叶材表面节子缺陷的检测方法.采用SmartEye 1700近红外光谱仪获取北美黄杉Pseudotsuga menziesii,铁杉Tsuga chinensis,云杉Picea aspera-ta,白云杉Picea glauca-英格曼云杉Picea engelmannii-扭叶松Pinus contorta-冷杉Abies laciocarpa(SPF)等4种板材的近红外光谱(1000~1650 nm),比较了光谱预处理方法、 建模方法对节子识别的影响,并首次对多种针叶树材进行了节子识别的适应性研究,随后引入一种新的变量选择方法即随机青蛙算法用于优选节子检测的特征波长,在此基础上建立了板材节子识别的最小二乘-支持向量机(LS-SVM)模型.结果显示:一阶导数光谱预处理结合LS-SVM所建混合树种板材节子识别模型性能最优.随机青蛙算法提取了8个特征波长变量,仅占全波段变量的1.23%,所建简化模型效果最好.该模型对测试集的敏感性 、 特异性和识别准确率分别为98.49%,93.42%和96.30%.近红外光谱技术结合化学计量学方法可以对针叶材树种板材的表面节子进行快速准确检测,随机青蛙算法是提取板材表面节子缺陷特征的有效方法.该结果可为下一步搭建木材节子快速检测系统提供技术支撑.%To develop a calibration model for rapid, accurate, and nondestructive grading of wood on the basis of knots, near infrared spectroscopy (NIRS) technology was used on coniferous boards from Douglas fir (Pseu-dotsuga menziesii), Chinese hemlock (Tsuga chinensis), Dragon spruce (Picea asperata), and Spruce (Picea glauca and Picea engelmannii)-Pine (Pinus contorta)-Fir (Abies lasiocarpa) (SPF). Altogether 1056 spec-trums of samples were obtained in the wavelength range of 1000-1650 nm by SmartEye1700. Spectral pre-treatment methods, including standard normal variate (SNV) and first derivative (FD) as well as modelling methods such as principal component analysis-linear discriminant analysis (PCA-LDA), partial least squares-linear discriminant analysis (PLS-LDA), and least squares-support vector machine (LS-SVM) were used and compared. The experiments also explored the ability of using a model built for one species to predict samples from other species. Then, a random frog algorithm was applied to select effective wavelengths (EWs). Finally, a LS-SVM model was established to detect knot defect on board surfaces based on eight effective wavelengths (EWs) or only 1.23% of the full wavelengths with results compared based on sensitivity, specificity, and accu-racy. Results of the validation set of mixed boards were: sensitivity--98.49%, specificity--93.42%, and accu-racy--96.30%. Thus, near infrared spectroscopy combined with chemometric methods could be used to detect surface knots on boards for different species of wood with the random frog algorithm being a powerful tool for selecting efficient variables to optimize the model and improve detection efficiency.