Spectrum wave number selection method

Abstract

The invention discloses a spectrum wave number selection method. Correction samples are randomly sampled many times according to wave numbers of spectra, a partial least square regression model is built, a variable projection importance coefficient of each wave number is calculated, the variable projection importance coefficients are sequenced in a descending mode, wave number sets corresponding to the distribution sequence are obtained, the re-sequenced wave number sets are subjected to wave number screening step by step, a result of each time of wave number selection is calculated, and a wave number primary selection set is obtained; then, absolute values of a partial least square regression coefficient of each wave number in the wave number primary selection sets is counted and correspondingly processed, the processed partial least square regression coefficients are sequenced in a descending mode, a corresponding wave number distribution sequence is recorded, then a strategy for reversely eliminating wake-correlation wave numbers is adopted, and the optimal feature wave number set is obtained. The method can fully mine effective information in wave numbers, effectively solves the subjectivity problem of wave number selection, extracts feature wave numbers to the largest extent, eliminates influences of weak correlation factors, and remarkably improves robustness and precision of the model.

Description

technical field [0001] The invention relates to the field of spectral analysis, in particular to a method for selecting spectral wavenumbers. Background technique [0002] As a fast and non-destructive quantitative analysis method, spectral analysis technology has been successfully applied in food, agriculture, petrochemical and other fields. However, the spectra obtained in spectral detection often overlap seriously, the spectral information is redundant, and the characteristic absorption regions are not obvious. In order to improve the prediction accuracy of the model and simplify the model, it is necessary to optimize the wave number, and select the characteristic wave number most relevant to the sample information to be tested for the establishment of the model. [0003] At present, research on wavelength selection methods in spectral analysis (wavelength is the reciprocal of wavenumber) mainly includes interval-based wavelength selection strategies, non-informative var...

AI Technical Summary

Problems solved by technology

Based on the interval wavelength selection strategy, it is difficult to determine the size of the wavelength interval, and the characteristic wavelength may be a single point, and the selected wavelength segment may contain irrelevant wavelengths, which cannot eliminate irrelevant wavelengths to the greatest extent; no information variable elimination Algorithm elimination of irrelevant variables is also a commonly used method of wavelength selection. It adopts the method of artificially generating a noise matrix for the correction spectral matrix and eliminating wavelengths whose information is smaller than the noise variable for feature wavelength extraction. However, this method also has certain subjectivity. And the choice of noise matrix affects the result of wavelength selection, and the final wavelength selection result usually has more redundancy

The limitations of the above existing methods affect the robustness and accuracy of the model

Images