Model Transfer Strategy Of NIR Analysis For Pulpwood Lignin Based On DS And PDS

In order to optimize the pulp and paper process parameters,real-time and rapid detection of the physical and chemical properties of the pulp material is required in the pulp and paper production process,which is of great importance to ensure pulp quality.Near-infrared spectroscopy is suitable for online inspection of pulp material properties because of its fast and non-destructive advantages.With the rapid development of NIR spectroscopy technology,there is growing interest in the use of portable NIR spectrometers.The novel and diverse wavelength selection techniques make it impossible to share the established models among instruments,and it takes a lot of time,manpower and material resources to re-establish multivariate calibration models.Model transfer is an important technical tool to solve this critical problem.In this study,the NIR spectra of 81 mixed pulpwood samples and their lignin contents were studied based on four NIR spectrometers(two benchtop spectrometers of the same model: Lengguang 1 and Lengguang 2,two Micro-Electro-Mechanical System based portable spectrometers: Si-Ware and Insion).With Lengguang 1 as the master device and the other three instruments as target devices,the application strategies of Direct Standardization(DS)and Piecewise Direct Standardization(PDS)methods commonly used in model transfer were explored in depth and detail to lay the foundation for the eventual use of NIR spectroscopy for quality monitoring of pulp and paper processes.The main research of this paper is as follows:(1)A characterization method based on Principal Component Analysis(PCA)of the overall difference of spectra between instruments was proposed for the model transfer of pulpwood content from four NIR spectrometers.The method was compared with three evaluation methods:Average Difference of Spectra,PCA + Average Mahalanobis Distance and PCA + Average Euclidean Distance,and all of them are found to be consistent in characterizing the global differences in spectra between instruments.The differences between Lengguang 1 and three target spectrometers are in the following order: Si-Ware > Insion > Lengguang 2.(2)Based on the instrument differences evaluation of PCA,model transfer between instruments was achieved by DS and PDS methods.It is shown that the greater the difference between the master and slave instruments,the more the number of transfer set is required for DS,and PDS method tends to use a larger window width(greater than 10)and a larger number of Partial Least Squares Regression(PLSR)latent factors(greater than 2).For different types of NIR spectrometers with large differences,the DS method outperforms the PDS method,and when the number of latent factors of PDS gradually increases(generally should not exceed 3),the PDS prediction results will gradually improve but still will not be better than the DS method.(3)The influence of the sequence of relevant steps in the model transfer process on the effectiveness of model transfer was explored.The relevant steps are overlapping band interception and wavelength interpolation(both collectively referred to as wavelength alignment),spectral preprocessing,and model transfer.The comparison reveals that for NIR spectra obtained from powdered pulpwood samples,the best results are found with Multiplicative Scatter Correction spectral pre-processing of transfer and prediction sets first,followed by a DS model transfer method for the target prediction set.After model transfer,the Ratio of Prediction to Deviation value of Lengguang 2 improve to 5.1217,Si-Ware,which is more different from the master spectrometer(Lengguang 1),improve to 3.3853,while Insion,which is less different from the master spectrometer(Lengguang 1),improve to 2.8600(4)The model transfer strategy of DS method for different types of master spectrometer and target spectrometers with inconsistent wavelength points was investigated.The comparison of prediction results after target devices’ model transfer shows that when the differences between master and target spectrometer are large,wavelength alignment is required to achieve the optimal prediction results;when the differences between master and target spectrometer are small,the results of both wavelength alignment and wavelength unalignment meet the actual needs,and the DS method with wavelength unalignment can be used when fewer sample sets are needed,besides the number of transfer set can be reduced by 10.However,in terms of accuracy and universality,wavelength aligned preprocessing is preferred in practical applications.(5)A DS-PDS model transfer strategy was proposed,which is the joint use of DS and PDS methods.The method combines the advantages of global correction of the DS algorithm and local correction of the PDS.For target spectrometers(Si-Ware and Insion)with large differences from the master spectrometer(Lenguang 1),the prediction set spectra were first corrected for the global differences,and then for the local differences.After the DS-PDS method,compared with the optimal results obtained by DS,the number of transfer set doesn’t increase.The window width,latent factor’s number of the PLSR model,and the asymmetry are all reduced in the PDS algorithm.The prediction accuracy of the target spectrometers is improved from 3.3853 to4.0335 for Si-Ware,and from 2.8600 to 3.4462 for Insion.