The Study Of Improving Quantitative Accuracy Of OP/FT-IR Spectra

As a remote sensing technique,open path Fourier transform infrared(OP/FT-IR)spectrometer is applied widely for gas monitoring in various types of open environment,with advantages of no sampling and real-time measurement.The open-path mode complicates measured spectra with serious interferences,which result in seriously spectral noise and overlap,and consequent difficulties in quantitative analysis.Multivariate calibration models establish the quantitative relationship between concentration and spectra through chemometric methods and finish the analysis of complex spectra.There are mainly four steps to build a model,including the construction of calibration set,spectral pretreatment,wavelength selection,and model building.Every step has an effect on the model accuracy.Particularly,the representative calibration set improves the model applicability.Wavelength selection can greatly strengthen the linear relationship between concentration and spectral absorption,and improve the model robustness and accuracy.This paper focuses on the construction of calibration set and wavelength selection.1.The predictive ability of the model constructed by actual calibration set is investigated.Upon synthetic and actual calibration sets that are composed of synthetic spectra and actual spectra respectively,partial least squares(PLS)models are built and evaluated with the validation of OP/FT-IR spectra.Results show that root mean squared error of prediction(RMSEP)of the actual PLS model is decreased by 86.60%compared with that of the synthetic PLS model,suggesting that actual spectra are more suitable than synthetic spectra to build PLS models for OP/FT-IR spectra prediction.2.Mixed calibration sets composed of synthetic spectra and actual spectra are constructed.Although prediction accuracy of model built on actual calibration set is good,the relatively long acquirement of actual spectra degrades its practicality.For seeking a better solution,mixed calibration sets containing actual spectra of different percentage and concentration are constructed and corresponding PLS models are built.Results show that the predictive ability of mixed PLS models is improved yet not stable when the percentage is below 25%.With the percentage being 25%,RMSEP of the mixed PLS model is decreased by 80%compared with that of the synthetic PLS model and the performance is more robust.On the one hand,the above result demonstrates that OP/FT-IR spectra as calibration samples play an important role in model building.On the other hand,it is a high cost-effective choice to use mixed calibration sets before the actual calibration set is available.3.A method named double channel wavelength selection(DCWS)is proposed.The feasibility of wavelength selection utilizing modified reference spectrum and stability as information vectors is verified respectively.In order to estimate variables more comprehensively and reliably,DCWS adopts both the modified reference spectrum and the stability as information vectors to evaluate the importance of every variable from internal and external factors.And a series of PLS models are built by using more variables that are in order of importance.Results show that the predictive error of model processed by DCWS is the smallest compared with other wavelength selection methods in this paper,which indicates the potential of DCWS in the field of wavelength selection.