论文部分内容阅读
分别采用卷积平滑法、小波变换法对蒙古栎木材近红外光谱(NIRS)做去噪处理,并讨论两者混合去噪时,处理顺序变化对光谱去噪效果的影响,最后应用偏最小二乘法(partial least squares regression,PLS)和主成分回归法建立蒙古栎木材气干密度近红外定标模型。结果表明,当平滑点数为3,db5小波分解层数为2时,以平滑+小波方式去噪效果最好,其信噪比(SNR)为18.546,均方根误差为0.04。平滑+小波去噪后,基于PLS的蒙古栎木材密度近红外校正模型决定系数由0.767提高到0.902,校正均方根误差降低了35.32%,预测集决定系数为0.860,内部交叉验证和预测均方根误差分别达到最低,剩余预测偏差为2.67。因此,近红外光谱技术可实现蒙古栎木材气干密度快速预测,合理选择处理参数和建模方法可以有效提高模型精度。
The convolution smoothing method and wavelet transform were used to denoise the NIRS of Quercus mongolica and the effect of the change of processing sequence on the spectral denoising effect was discussed when the two were mixed de-noising. Finally, Multiplicative partial least squares regression (PLS) and principal component regression were used to establish a near infrared calibration model of dry density of Q. mongolica wood. The results show that when the number of smoothing points is 3 and db5 wavelet decomposition level is 2, the smoothing + wavelet method has the best performance of denoising. The SNR is 18.546 and the root mean square error is 0.04. After the smoothed + wavelet denoising, the determination coefficient of wood density near-infrared calibration model based on PLS was increased from 0.767 to 0.902, the corrected root mean square error was reduced by 35.32%, the prediction coefficient of prediction set was 0.860, the internal cross-validation and prediction mean square The root errors are respectively the lowest and the residual forecast deviation is 2.67. Therefore, near-infrared spectroscopy can quickly predict the dry density of Q. mongolica wood. The reasonable selection of processing parameters and modeling methods can effectively improve the accuracy of the model.