近红外光谱分析的实木板材节子形态反演

The Inversion of Knots in Solid Wood Plates Based on Near-Infrared Spectroscopy

节子影响着实木板材力学性能,如何准确刻画出节子在实木板材内部的形态,进而计算出实木板材力学性能是一个具有应用价值的科学问题。目前,基于机器视觉的缺陷检测方法实现了实木板材表面缺陷检测与识别,超声波检测方法可以判断出实木板材中缺陷的存在, X-ray虽然可以全面的掌握实木信息,但其检测成本较高。近红外光谱分析具有结构丰富,测试方便、无损快速的特点,但是,光谱中存在的冗余与非线性信息影响建模精准度,提出一种基于Isomap和小波神经网络融合的节子倾角辨识方法,利用Isomap完成光谱信息非线性降维,运用小波神经网络建立节子边缘的物质成分与倾角间的非线性关联,通过边缘多点倾角反演出节子在实木板材内部的形态。首先,采用Pablo提出的节子斜圆锥模型,并结合图像处理提取实木板材表面的节子缺陷区域,计算出相应中心位置;提取节子边缘的多点位置,采集光谱信息并完成基线漂移和去噪处理;然后,利用K-S划分校正样本集,运用主成分与马氏距离结合剔除异常光谱;接着,运用Isomap方法设定降维数和邻近数,通过PLS完成不同光谱维度的快速建模,进而迭代出理想光谱特征;最后,应用具有局部信息优化能力的小波神经网络建立节子边缘光谱与该点倾角间的非线性关系,构建出1个12输入、 1输出的网络模型,并运用梯度修正网络参数;将节子倾角预测结果输入Solidworks软件完成节子椎体形态的三维呈现。实验采用落叶松实木板材作为对象,选取并采集了40个节子的160组光谱数据,通过测量上、下表面节子的相对空间位置,计算出边缘点倾斜角数值并进行建模分析,实验结果表明:采用S-G平滑与一阶导数进行光谱预处理,得到的光谱轮廓更清晰、吸收峰更明显;采用Isomap特征降维方法,选取非线性降维数 d =12、近邻数 k =19时, SECV最小,可以消除光谱信息的冗余数据;采用小波神经网络建立的节子倾角非线性模型,其预测相关系数为0.88,预测标准差为7.65,相对分析误差为2.14;可以实现节子在实木板材内部的形态反演,可以为力学性能预测提供定量化分析手段。

Knots affect the mechanical properties of solid wood plates.The accurate description of knots in wood plates and the calculation of wood board mechanical properties have been issues with great practical value.Nowadays,machine vision method is used to detect the defects on wood surface,ultrasonic testing is used to determine the existence of defects,and X-ray method can give a full description of solid wood,but the cost is high.The near infrared spectroscopy analysis technology has the characteristics of rich structure,convenient testing and being nondestructive,but the redundancy and nonlinear information in the spectrum affect the precision of the modeling.In this paper,a method of identifying the knots based on the fusion of Isomap and wavelet neural networks is proposed,and the nonlinear dimensionality reduction is completed by Isomap.The nonlinear relation is modelled by the wavelet neural network between the material and the angle of the knot edge,and the shape structure of the knot inside the wood plate is performed by the multi-point angle of the edge.First,the method uses the cone model to express the knot structure proposed by Pablo.Knots are extracted by machine vision method from the image,and their center positions are obtained by calculation.Then,the information of multi-point position about the edges of knots is extracted and processed by the baseline drift and denoising methods.After that,abnormal spectrums are eliminated by combining PCA and mahalanobis distance,the calibration sample sets are divided by K-S,and effective spectral information are extracted through Isomap,which set the dimensionality reduction and adjacent number,and the fast modeling of different spectral dimensions is completed through PLS,and then the ideal spectral feathers are iterated.Finally,wavelet network is used to establish the relationship between the edge spectrum and their inclination angel of the knots,and the 3D status of these knots is realized by Solidworks software.In this experiment,160 sets of spectral data of 40 knots were collected from Larix gmelinii plates.After measuring the relative spatial position of the upper and lower surfaces of knots,true inclination angles of every point were obtained.The result of the experiment reveals that S-G smoothing and first order derivative can give clear outline in spectral pre-processing and the absorption peak is more obvious.When Isomap method is for dimension reduction,with non-linear dimension reduction number d =12,the nearest neighbor number k =19,the SECV is the minimum and the redundant data of spectrum is eliminated.When wavelet neural network is used to build the model of nonlinear inclination angles of knots,the correlation coefficient is 0.88,the prediction standard deviation is 7.65,and the relative analysis error is 2.14.This method can realize the inversion of knot structure in wood plates,and can provide quantitative analysis means for the prediction of mechanical properties.