Near-Infrared Spectroscopy Modeling For 2,6-Dimethylphenol Purity Based On Transfer Learning

The critical raw material for synthesizing polyphenylene oxide(PPO)is2,6-dimethylphenol(2,6-DMP).The realization of large-scale industrial production of2,6-DMP is significant in promoting the development of the engineering plastics industry in China.Online measurement of 2,6-DMP product concentration is essential to control and optimize product quality.The traditional detection methods are mainly obtained by measuring the physical and chemical properties that vary with the purity of the solute in the sample system,and online measurement cannot be made.As an advanced online process detection technique,near-infrared(NIR)spectroscopy has been widely applied in many fields,such as the chemical,pharmaceutical,biological,food,and medical industries.The NIR detection model needs chemometrics methods to establish the relationship between sample spectrum features and analytes’ contents.However,the accuracy of NIR detection results is closely related to the quality of the model.In the final stage of the 2,6-DMP distillation and purification process,the diversity of samples is insufficient,which will lead to the lack of correlation between the solute concentration and the spectrum.It is hard to establish a reliable and highly accurate near-infrared detection model.In this paper,we study the NIR modeling based on transfer learning and quantitatively analyze the relationship between the quality and quantity of transfer samples and the model performance.The specific research contents are as follows:(1)Sample-based transfer learning modeling of near-infrared spectroscopy.In the process of 2,6-DMP distillation,due to the high purity of the product purity of the 2,6-DMP product column,the dispersion degree of samples is low,resulting in the lack of correlation between the NIR spectrum and product purity.The established detection model is challenging to meet the accuracy requirements.The automatic updating strategy of spectral weights is established by the Two-Stage Tr Ada Boost.R2 algorithm.It can reduce the differences in distribution between the spectra of different detection points.The spectra with considerable similarities between other detection points and the product column detection point are retained,and those with significant differences from the product column detection points are eliminated,so that the spectra of other detection points can help to improve the accuracy of the product column detection point model.The effectiveness of the proposed method is verified by establishing a near-infrared online detection model for the product purity of the2,6-DMP distillation and purification process.(2)Feature-based transfer learning modeling of near-infrared spectroscopy.After applying the sample-based transfer learning method,considering that the near-infrared spectrum is high-dimensional data,a feature-based transfer learning modeling of near-infrared spectroscopy is proposed using the similarity between spectral features.The method creates subspaces and extracts features for the data sets of different detection points.Then,a mapping that aligns the other tower subspace into the product tower is learned by minimizing a Bregman matrix divergence function,which can reduce the feature distribution discrepancy between other columns and the product column.The advantage of the method is verified by analyzing the NIR spectral data sets of the 2,6-DMP distillation and purification process.(3)Performance evaluation of NIR models based on transfer learning.When using transfer learning for NIR modeling of 2,6-DMP purity,the product column detection model’s performance will be degraded because the spectra of other detection points are not similar to the product column detection point,which is a harmful transfer problem.A NIR model performance evaluation method based on transfer learning is proposed.The analytic relationship between the detection model performance and the spectral information to be transferred was developed using the influence function.According to the analytic relationship,the effect of the quantity and quality of NIR spectra on detection model performance can be studied.This helps provide guidance on whether to transfer source domain information knowledge and avoids negative transfer.Finally,the theoretical results are applied to the NIR data sets of the 2,6-dimethylphenol distillation and purification process,which shows that the method can transfer more favorable spectral data and make the accuracy of the detection model significantly improved.In summary,this paper establishes an online detection model for product purity in the2,6-DMP distillation and purification process using near-infrared spectroscopy and solves the significant problem of insufficient sample diversity in the product tower based on transfer learning.At the same time,the effectiveness of the proposed methods is verified through experiments.The research content of this paper helps to solve the problem of accurate modeling when using near-infrared spectroscopy for online detection of substance concentration,provides new ideas for researchers in the field of quality detection,and lays a foundation for future research work.