基于卷积神经网络的樟子松木材密度近红外预测模型优化

NIR Prediction Model Optimization Study of Pinus sylvestris Wood Density Based on Convolutional Neural Network

近红外光谱分析技术在木材密度的预测方面具有独特的优势,是一种方便且快速的无损检测技术。卷积神经网络作为经典的深度学习模型之一,能够利用卷积和池化操作提取数据中的特征映射进行学习,与传统的学习模型相比具有更强的模型表达能力。为此将卷积神经网络用于近红外光谱预测木材的气干密度,以樟子松为研究对象,获取样本木材横切面的近红外光谱数据,采用杠杆值与学生化残差t检验(HLSR)法剔除奇异样本,采用SGS+MC+Auto(Savitzky-Golay smoothing+mean centering+autoscaling)对光谱数据进行预处理,通过竞争性自适应重加权算法(competitive adaptive reweighted sampling method,CARS)对特征波长进行提取,构建卷积神经网络模型,预测樟子松的气干密度;并与偏最小二乘回归(partial least squares regression,PLSR)、支持向量机(support vector regression,SVR)和BPNN(backpropagation network)神经网络的预测结果进行对比。结果表明,当校正集比例小于0.65时,模型预测结果略低于PLSR模型。但当校正集比例大于0.7时,卷积神经网络(convolution neural network,CNN)模型的预测精度优于其他模型,且随着训练样本比例的增加,模型的性能和稳定性也随之提升。研究表明CNN可以显著提高近红外预测木材气干密度的模型精度,实现基于近红外技术的木材密度有效预测。为木材气干密度无损检测提供了理论基础和科学依据。

Near-infrared spectroscopy is uniquely suited to the prediction of wood density,and is a convenient and rapid non-destructive testing technique.Convolutional neural networks,as one of the classical deep learning models,is capable of extracting feature mappings from data for learning using convolutional and pooling operations,and has more powerful modeling capability compared to traditional learning models.In this study,CNN was used in near-infrared spectroscopy to predict the air-dry density of wood.Pinus sylvestris was used as the study object to obtain the near-infrared spectral data of the cross-section of the sample wood.The high leverage-studentized residual(HLSR)method was used to remove abnormal samples,and SGS+MC+Auto was used to pre-process the spectral data.The characteristic wavelengths were extracted by the CARS algorithm and a CNN model was constructed to predict the air-dry density of the samples.The prediction results were compared with the prediction results of the partial least squares(PLS),the support vector machine(SVR)and the BPNN neural network.The results showed:when the calibration set ratio was less than 0.65,the model prediction results were slightly lower than the PLSR model.However,when the calibration set proportion was greater than 0.7,the prediction accuracy of the CNN model was better than the other models,and the performance and stability of the model improved with the increase of the proportion of calibration set.The study showed that CNN can significantly improve the accuracy of the model for predicting wood air-dry density in near infrared(near infrared spectroscopy),and realize the effective prediction of wood density based on NIR technology.It provides a theoretical foundation and scientific basis for the non-destructive testing of wood air-dry density.