A Spectral Preprocessing Method

Abstract

The present invention relates to a spectrum preprocessing method. Firstly, the spectrum is denoised: the spectrum is input into the convolution denoising layer C1 for denoising, and then the spectrum is baseline corrected: the spectrum denoised by C1 is input into the baseline calibration Layer C2. In the method of the present invention, the denoising and baseline correction are integrated into the method by convolution, so that the preprocessing process and the identification problem are transformed into a unified model framework for solution, and the self-adaptive processing of data is realized, which makes up for the shortcomings of the traditional method; In the process of noise and baseline correction, there is only one convolution kernel in each convolution layer, which is more interpretable than traditional networks, and it is easier to monitor the output of the convolution layer to see whether the expected effect is achieved.

Description

technical field [0001] The invention relates to the field of spectral analysis methods, in particular to a spectral preprocessing method. Background technique [0002] The original spectrum measured by the instrument is usually accompanied by serious noise and baseline drift, which limits the extraction of useful information in the spectral data during identification. Therefore, the process of spectral analysis usually involves cumbersome denoising and baseline correction processes. [0003] Most of the current denoising and baseline correction methods are independent of the model, which increases the risk of the model's predictive ability being reduced due to improper preprocessing. At the same time, affected by the physical properties of the substance to be measured (such as particle size, packing density and uniformity, etc.), the ambient temperature and the nonlinear response of the detector, there is a certain nonlinear relationship between the spectrum and the properti...

AI Technical Summary

Problems solved by technology

[0002] The original spectrum measured by the instrument is usually accompanied by serious noise and baseline drift, which limits the extraction of useful information in the spectral data during identification. Therefore, there are usually cumbersome denoising and baseline correction processes in the spectral analysis process.

[0003] Most of the current denoising and baseline correction methods are independent of the model, which increases the risk of the model's predictive ability being reduced due to improper preprocessing

At the same time, affected by the physical properties of the substance to be measured (such as particle size, packing density and uniformity, etc.), the ambient temperature and the nonlinear response of the detector, there is a certain nonlinear relationship between the spectrum and the properties and composition of the substance to be measured. However, most of the spectral qualitative or quantitative correction methods at this stage are linear models, which cannot be well represented for this nonlinear relationship.

[0004] In the prior art, when establishing a classification model, most of the model parameters are preset based on experience. On the one hand, this method limits the accuracy of spectral analysis. On the other hand, according to the characteristics of the substances to be classified, the parameters of the classification model also need to be adjusted. , so the prediction model parameters set according to empirical values ​​can only be used to classify some specific substances, and the parameters of the classification model need to be reset every time different substances are measured, which not only leads to the generalization of the classification model The performance is poor, and more importantly, the classification accuracy of the classification model is also difficult to guarantee

Images