基于卷积神经网络的近红外光谱多组分定量分析模型研究OA北大核心CSTPCD

Near infrared spectroscopy(NIRS)has emerged as an indispensable analytical technology for quality monitoring in industrial and agricultural production.It is widely used in quantitative analysis in areas such as food,agriculture and medicine.To meet the requirements of industrial and agricultural production,it is particularly important to develop a NIRS quantitative analysis model that can predict the multicomponent of different samples.In this study,the multicomponent quantitative analysis model of NIRS based on convolution neural network(MulCoSpecNet)was proposed.MulCoSpecNet was composed of an encoding and decoding module,an expert module,a gate module,a multicomponent quantitative prediction module,and a hyperparameter optimizer.The spectral noise and random errors were mitigated,and the signal-to-noise ratio was enhanced through up-sampling and down-sampling in the encoding and decoding module.Diverse weightings were employed by the expert module and gate module to construct distinct sub-spectra.The model prediction accuracy and generalization ability were enhanced by the multicomponent quantitative prediction module,which employed convolutional and pooling operations.The hyperparameters in the hyperparameter space were synchronously optimized by the hyperparameter optimizer.By taking public NIRS datasets of grain and corn as examples,the prediction results of MulCoSpecNet were compared with partial least squares(PLS),extreme learning machine(ELM),support vector regression(SVM)and back propagation neural network(BP).The results showed that compared to PLS,the prediction accuracy of MulCoSpecNet to grain and corn were increased by 25.5%‒45.2%and 10.0%‒35.7%,respectively.Compared to ELM,the prediction accuracy of MulCoSpecNet were increased by 17.8%‒38.6%and 18.2%‒37.2%,respectively.Compared to SVM,the prediction accuracy of MulCoSpecNet were increased by 33.6%‒47.0%and 31.3%‒50.7%,respectively.Compared to BP,the prediction accuracy of MulCoSpecNet were increased by 2.0%‒58.5%and 29.6%‒48.6%,respectively.The issues of low prediction accuracy and poor generalization ability were effectively solved by the MulCoSpecNet,which was a NIRS multicomponent prediction model based on convolutional neural network.It provided a theoretical foundation for establishing non-destructive and high-precision NIRS multicomponent quantitative analysis model.