Near-infrared Adaptive Modeling Based On Just-in-time Learning

Near Infrared Spectroscopy(NIRS)analysis technology has great potential in real-time detection due to its nondestructive,efficient,fast,simple operation and low-cost,and has become an emerging online process analysis tool.However,in the actual production process,the initially established NIR model is often affected by changes in process operating conditions,instrument performance,production environment,processing materials or catalyst,which reduces the prediction accuracy of the target physical concentration,and thus cannot continue to meet the production needs.Therefore,it is necessary to update and maintain the model constantly to improve the adaptability of the NIR model.In this paper,aiming at the problems of poor stability and short time life of existing NIR models,just-in-time learning(JITL)method is adopted to improve the adaptability of NIR models by continuously establishing local models to track the changes of process characteristics.However,the problem of high dimension,redundant noise and irrelevant wavelength variables in NIR spectra restrict the model prediction performance.Therefore,in this paper,aiming at the shortcomings of NIR online modeling,combining the research dynamics of JIT learning and the characteristics of NIR spectral data,the NIR adaptive modeling based on JIT learning is studied and improved.The specific research contents are as follows:(1)Adaptive JIT-Lasso modeling based on time-space similarity.Lasso(Least Absolute Shrinkage and Selection Operator)algorithm is firstly selected as a local modeling technique in JIT learning,aiming at the characteristics of time-varying,nonlinear,high dimensional and large information of spectral wavelength variables in the industrial process.The algorithm can perform regression analysis and wavelength selection at the same time,so the spectral bands containing redundant noise or irrelevant information variables can be eliminated,which improves the model accuracy and model interpretation ability.In order to further overcome the process time-varying and nonlinear problems,the comprehensive time correlation and spatial correlation similarity measure method is studied.The experimental results of crude oil desalination and dehydration show that the proposed method has better performance.(2)Near infrared spectroscopic adaptive modeling based on local and global spectra features.On the basis of the determination of local modeling technology,the selection of similar spectral samples has a critical impact on the model accuracy.Therefore,an uncertain similarity measure criterion based on local-global spectral characteristics is proposed by further fusion of near-infrared spectral characteristics in the JIT learning framework..This criterion integrates the knowledge of spectroscopy,and comprehensively evaluates the similarity between NIR spectra samples through local spectral shape changes and global spectral information differences,then a JIT learning model based on NIR spectral characteristics is established.Using JIT-PLS and JIT-Lasso modeling techniques,combined with different similarity variable methods to compare the experimental prediction of physical property concentration of crude oil desalination and dehydration process,it is verified that the prediction performance of adaptive modeling based on local and global features of NIR spectra is better..(3)Near infrared spectroscopic adaptive modeling based on database update index.JIT learning is a memory-based approach,and model performance depends largely on the validity of database information.In view of this,this paper constructs the database updating index(DUI),and proposes the corresponding strategies for updating the database according to the types of changes in the actual process:(I)add the new spectral sample to the historical database,(II)replace the corresponding historical spectral sample with the new spectral sample.It realizes the update of the NIR spectra sample database while conducting JIT learning,which further improves the model prediction accuracy.Finally,the theoretical results are applied to the measurement of chloride content in the process of crude oil desalination and dehydration,which shows that the model prediction of JIT modeling framework based on DUI has better performance.