Research On The Establishment And Transfer Of Near-infrared Spectroscopy Quantitative Analysis Model Based On Convolutional Neural Network

With the development of computers and the in-depth study of chemometrics,as well as the continuous development of spectroscopic instrument manufacturing technology,Near infrared reflectance spectroscopy（NIRS）has also been continuously improved.As an emerging component analysis technology,NIRS has the characteristics of low cost,no pollution and high efficiency,and is widely used in various fields.However,its absorption intensity is weak,spectral peaks overlap,and interference information exists,which makes the establishment of near-infrared spectroscopy models.Spectral preprocessing（denoising,derivation,baseline correction,etc.）and characteristic wavelength screening are usually required before the operation,which requires very high requirements for modeling analysts.Therefore,it is of great significance to study a simple and convenient near-infrared spectroscopy modeling method for lowering the threshold of modeling analysts.In addition,in the process of spectrum collection,the phenomenon of"the same substance collected under different spectrometers is different,but different substances are collected under the same spectrometer"phenomenon often occurs,which makes the models established under different spectrometers unable to Shared,it would be time-consuming and laborious to model each spectrometer individually.Therefore,research on a universal model transfer method for quantitative analysis of near-infrared spectroscopy also has important application value.In view of the above problems,this paper proposes a new idea and method for the establishment and transfer of a quantitative analysis model of near-infrared spectroscopy based on convolutional neural network.The main research work is as follows:（1）This paper proposes an"end-to-end"（End-to-End）near-infrared spectroscopy modeling method,that is,using the Convolutional Neural Networks（CNN）method to establish a near-infrared spectroscopy quantitative analysis model.First,according to the distribution law of the data set,the data set is divided into training set and test set,and the topology and hyperparameters of the convolutional neural network model are designed;After the training process converges,the near-infrared spectroscopy quantitative analysis model can be obtained;finally,it is constructed with the traditional Partial Least Squares（PLS）and Support Vector Regression（SVR）algorithms.model for comparison.The experimental results on the corn dataset show that the quantitative analysis model established by the convolutional neural network has the best performance,and the coefficient of determination R~2 can reach 0.97844.（2）This paper proposes a transfer method based on the transfer learning algorithm for quantitative analysis of near-infrared spectroscopy models,which takes the established convolutional neural network quantitative analysis model as the source domain pre-training model,uses a small number of samples in the target domain,and fine-tunes it.The transfer training method completes the transfer of the near-infrared spectroscopy quantitative analysis model from the source domain to the target domain.Further,this paper compares and analyzes the impact of three different transfer schemes（only training the fully connected layer,training the fully connected layer and the last convolution block,and training all layers of the network structure）on the performance of the target domain model.The experimental results on the corn dataset show that the generalization performance of the target domain NIR quantitative analysis model trained by the transfer learning algorithm is significantly better than that of the untransferred model（directly applying the source domain model to the target domain for prediction）and the traditional Tr Ada Boost Model transfer method,the coefficient of determination R~2 can reach 0.93.In this paper,an"end-to-end"near-infrared spectroscopy quantitative analysis model is established based on convolutional neural network.At the same time,the model transfer between different near-infrared spectrometer devices is realized based on the transfer learning algorithm.The results of this paper are of great significance and value for lowering the threshold of near-infrared spectroscopy modeling analysts and improving the applicability of quantitative analysis models.