Near-infrared Spectroscopy Modeling Technology Based On Manifold Learning

Near Infrared Spectroscopy(NIRS)analysis technology,as an efficient,rapid,nondestructive and low-cost online detection technology,can realize the quantitative analysis of product quality,and is widely used in the production process of modern process industry.However,due to the problems such as multicollinearity,high detection sensitivity and baseline drift of NIR spectrum,quantitative information about samples cannot be obtained directly from spectral signals.In this case,NIR spectra analysis technology can only be detected through indirect analysis,which relies on high-performance calibration model.Therefore,the key point of NIRS analysis technology is how to establish a high-precision calibration model.In this paper,due to the high and nonlinear problems in the NIR modeling,based on the manifold hypothesis of spectral data,the method of manifold extraction is used to help extract global and local structure information of the data,to improve the accuracy of modeling.However,the problems of multiple operating conditions and insufficient spectral data labels often occur in the process industry,which affect the actual modeling work.Therefore,the multi-manifold hypothesis is carried out on the spectral data under different historical operating conditions,and the research is carried out from the perspective of identifying multiple operating conditions and migrating information of multiple operating conditions respectively.The specific research contents are as follows:(1)NIRS modeling based on global-local preserving embedding.To solve the problem of high dimension and nonlinearity in the spectral data,the global and local structure information is used to build a high precision calibration model.The acquisition of local information can extract local invariant information of data manifold and help to deal with nonlinear relations between data.The extraction of global information can capture the overall structure information of data(i.e.,variance information)and alleviate the quality problem of data sampling.Based on the above advantages,this method can deal with high-dimensional and nonlinear problems while retaining more effective information,thus helping to build a more accurate correction model.The advantages and feasibility of the proposed method are demonstrated by the NIRS analysis of crude oil desalting and dehydration process.(2)Multi-manifold NIRS modeling via stacked contractive auto-encoders.In this paper,a multi-manifold modeling method based on deep autoencoder is proposed based on the multi-manifold hypothesis of spectral data.The method uses a contractive auto-encoder,which can capture the tangent plane direction of data manifold and help solve the nonlinear problem between data.To improve the robustness of the model,the deep network structure is used to enhance the contractive auto-encoder,which alleviates the high sensitivity of spectral detection to environmental changes.In addition,as for different operating conditions,the method extracts the corresponding sub-manifold structure information,and then carries out offline modeling based on this information.After the modeling is completed,the method can identify the conditions online and predict the measured attribute values with high performance.Compared with other traditional methods,the experiment of crude oil desalting and dehydration process verifies the superiority of this method.(3)NIRS modeling via multi-manifold migration.Under the background of multiple operating conditions in the production process,aiming at the problem of insufficient data labels,the data information of historical working conditions is used to help the current operating conditions establish a correction model meeting certain accuracy requirement.In this paper,a partial least squares manifold transfer modeling method is proposed,which preserves the local structural properties of the data manifold and the regression relationship between the spectral data and the sample attribute values while transferring the data information under the historical operating conditions,to ensure that the transferred calibration model has better prediction performance.In addition,a multi-manifold migration framework is proposed,which uses a universal feature extractor to remove redundant information from all spectral data,and then models are modeled for different historical conditions.Finally,the results of different models are weighted according to maximum mean discrepancy(MMD).Experiments on crude oil desalting and dehydration show that this method has good predictive performance.