Near Infrared Spectral Modeling Based On Variational Autoencoder

Near Infrared Spectroscopy(NIRS)technology is widely used in food,industry,pharmaceuticals and other fields due to its fast,non-destructive,green,sensitive and other excellent properties.How to establish a mapping between spectrum and target parameters that can accurately fit The model of the relationship has become a key issue in the application of near-infrared technology.However,near-infrared spectroscopy has high-dimensional characteristics,and there is often a nonlinear relationship between its dimensions,and it is easily affected by environmental factors,and irrelevant information is incorporated into the spectrum.These properties restrict the traditional modeling algorithm for near-infrared spectroscopy.Learning ability and modeling effect.In this paper,in view of the characteristics of high dimensionality of near-infrared spectral features,nonlinear relationship between dimensions,low signal-to-noise ratio,and oversensitivity,a spectral model is constructed based on a variational autoencoder(VAE)to achieve accurate tag values.predict.However,the common problems of multi-working conditions and unbalanced sample distribution in industrial production will affect the model,expand the model error and even lead to model mismatch.The weight change of each working condition sample is studied.The specific content of the paper is as follows:(1)Near-infrared spectral modeling based on supervised variational autoencoder: using variational autoencoder as a feature extraction module,using its stacked network structure to solve the problem of high-dimensional and nonlinear correlation of spectra,and realizing deep analysis of spectral data feature learning.Different from the traditional model,VAE learns the distribution representation of data in the hidden space by injecting noise into the hidden layer of the network during the training process,so it has a strong resistance to noise and is suitable for solving the hypersensitivity problem of the near-infrared spectrum.However,the unsupervised property of VAE limits its effective learning of the mapping relationship between spectral data and label values.Therefore,this chapter adds a regression layer based on VAE,proposes a near-infrared spectral modeling method based on supervised variational autoencoders,and tests its superiority on the corn dataset.(2)Hybrid Variational Autoencoder near-infrared spectroscopy modeling based on similarity measure: Aiming at the problem of different distribution of data characteristics of each working condition in the process of multi-working conditions,VAE cannot independently learn the characteristics of each working condition due to the limitation of unimodal distribution.Condition feature distribution.In this paper,a hybrid variational autoencoder algorithm based on similarity measures is used to construct a spectral model and realize feature extraction for different working conditions.The proposed algorithm introduces the concept of mixed model,and optimizes its prior assumptions on the basis of VAE,making it obey the multimodal distribution to independently learn the distribution of hidden variables under each working condition and obtain the unique characteristics of each working condition.In addition,for the probability vector calculation criterion used in feature fusion,this paper introduces the spectral similarity measurement criterion for optimization,and takes the correlation between the sample and the spectrum under each working condition as the probability of belonging to the class for fusion.The method proposed in this paper was applied to the near-infrared data set of melamine formaldehyde,and the quality monitoring model was established with the cloud point of the process parameter as the label value,which proved the feasibility of the method.(3)Hybrid variational autoencoder near-infrared spectral modeling based on Adaboost: In the face of the problem of unbalanced sample distribution in the process of multi-working conditions,the near-infrared modeling method of hybrid variational autoencoder based on similarity measure The feature framework guarantees independent learning of the characteristics of each working condition,but its regression module is still affected by the number of samples,that is,it tends to fit working conditions with a large number of samples and pays insufficient attention to working conditions with few samples.Based on this,this paper introduces the Adaboost algorithm to optimize the regression module.This algorithm increases the influence of the sample error on the model by giving the sample with a large error a high weight,and improves the model’s attention to the sample.Therefore,when assigning values to the weight matrix of samples,the working condition with few samples will be given a larger weight,so as to increase the attention of the model to this working condition,so as to achieve the purpose of improving the prediction accuracy.