Research On Data Analysis And Feature Modeling Methods Of Near Infrared Spectroscopy

In recent years,non-destructive detection technology based on the physical properties of materials such as sound,light,and electromagnetic has been developed rapidly.As a typical analysis method,near infrared reflectance spectroscopy(NIRS)technology has been widely used in key production areas,such as biopharmaceutical,agriculture,and chemical industry,due to its advantages of fast,non-destructive and high-efficiency.However,the NIRS data collected in the actual process has the problem of baseline shift,feature overlap,highdimension,and noise,and cannot be used for analysis and modelling directly.To this end,under the funding of national key research and development program “Research on Key Technologies for New Types of Identification and Testing of Special High-value Agricultural Products”(Grant: No.2017YFF0211301),penalized least squares,sparse representation and just-in-learning(JITL)are combined and used to perform the research of NIRS data processing and feature modeling.The specific content is concluded as follows:(1)For the problem of baseline drift of NIRS data,an adaptive derivatively reweighted penalized least squares baseline correction method is proposed.The derivative weighting strategy and balance factor were introduced to iteratively calculate the correction error,which can perform the repaid baseline correction in the feature overlap area.The simulation and experimental data results confirmed that the proposed method improve the accuracy and calculation efficiency of NIRS baseline correction.(2)For the problems of high dimensionality,variable redundancy,and collinearity of NIRS data,the application of Elastic net method is explored in NIRS feature modeling.For small samples of high-dimensional NIRS data,this method can perform variable selection and regression analysis simultaneously,thus it can effectively improve the accuracy and interpretation of the model.In order to further analyze the high collinearity of NIRS data,the group feature extraction capability of Elastic net methods is studied.The performance of the NIRS feature modeling based on the sparse representation of Elastic net is verified through the experimental analysis of the key indicators of Tricholoma Matsutake.(3)In view of the low prediction performance of the global model caused by the timevarying,non-linear,high-dimensional and other characteristics of the actual analysis process,the JITL framework and Elastic net were utilized to update the NIRS model.In order to further improve the interpretation ability of the NIRS local model,the shape and information characteristics of NIRS data were combined and a sample similarity measurement criterion based on spectral feature fusion was proposed in this work.The prediction performance of different local modeling strategies and similarity measurement criteria were verified by the NIRS data of the tea drying process.The analysis results confirmed that the performance of the local model based on spectral feature fusion is better.The above research results provide technical support for the construction of an integrated non-destructive testing system for precise attribute screening,multi-dimensional quality identification and accurate characteristic analysis in the production process,which is conducive to promoting the application of NIRS analysis technology in government quality inspection center,custom inspection and quarantine and other regulatory systems.It is of great significance to further improve the process intelligence,information,and digitization of key fields such as food medicine,and chemical industry,and to solve the technical problems of intelligent manufacturing.