CN102135496A - Infrared spectrum quantitative analysis method and infrared spectrum quantitative analysis device based on multi-scale regression - Google Patents

Abstract

The invention relates to an infrared spectrum quantitative analysis method and device based on multi-scale regression, including a spectrometer connected to a data signal line, a preprocessor, a wavelet decomposition and reconstruction processor, and a partial least squares regression model integrator; the infrared spectrum Including mid-infrared and near-infrared spectra, that is, the wavelength range is: 780nm-50000nm. Through wavelet decomposition and reconstruction transformation, the construction of multiple models is realized, which overcomes the difficulty of extracting spectral signal information by the single model method; by separately determining the number of factors for different sub-models, the effective information is fully extracted and the infrared spectrum is improved. Analyze the predictive accuracy and robustness of the model.

Description

Method and device for quantitative analysis of infrared spectroscopy based on multi-scale regression

technical field

The invention relates to an infrared spectrum quantitative analysis method and device, in particular to an infrared spectrum quantitative analysis method and device based on multi-scale regression.

Background technique

Because the spectral signal obtained by the spectrometer contains both useful information and other random errors (background and noise). Therefore, it is difficult to obtain a model with high prediction accuracy when using the partial least squares regression method for quantitative analysis. To improve the prediction accuracy of the partial least squares regression model, a lot of research work has been carried out, mainly including spectral preprocessing methods and variable screening methods, these methods have been successfully used in spectral background correction, noise removal, Elimination of non-informative variables. However, when using the above methods for information extraction, we often face the problem of insufficient extraction of useful information or noise in the extracted information. Therefore, it is necessary to invent a method that does not need to extract the spectral information, but classifies the information contained in the spectrum, and then models and predicts the classified spectral information respectively, and finally performs model results for different types of spectral information. Integrating, this can not only avoid inaccurate extraction of spectral information, but also improve the prediction accuracy of the final result of the spectral model.

Wavelet transform has been successfully applied in the preprocessing of spectral signals, including data compression, smoothing and noise filtering, baseline correction, overlapping signal analysis and image processing. Unlike other preprocessing methods, wavelet transform has a "time-frequency advantage". Through wavelet transform processing, a spectral signal can be divided into several sub-signals of different frequencies. After these sub-signals are reconstructed, the information contained in the reconstructed spectrum is almost identical to the original spectrum. Finally, the reconstructed subspectral signals at different frequencies are modeled and predicted separately, and the predicted results are integrated.

In this patent, an infrared spectrum quantitative analysis method and device based on wavelet decomposition and reconstruction combined with partial least squares regression method are invented. It overcomes the difficulty of extracting spectral information by a single model method. When analyzing complex spectral data, the method and device of the present invention can directly use the reconstructed spectra at each scale to establish multiple partial least squares sub-models. These sub-models The number of factors can be selected more flexibly and effectively according to the amount of information contained in the sub-spectrum, so as to achieve the purpose of fully extracting useful information at each scale.

Contents of the invention

In order to solve the problem of inaccurate extraction of spectral information in the process of rapid analysis of components to be measured by infrared spectral analysis technology. The invention provides an infrared spectrum quantitative analysis method and device based on multi-scale regression. The method and device make full use of the characteristics of multi-resolution analysis of wavelet transform, and realize the full use and reasonable distribution of infrared spectrum information.

The device for realizing the technical solution includes: a spectrometer connected via a data signal line, a preprocessor, a wavelet decomposition and reconstruction processor, and a partial least square regression model integrator.

The infrared spectrum includes mid-infrared and near-infrared spectra, that is, the wavelength range is: 780 nm-50000 nm.

The preprocessor adopts centralization and vector normalization to process the original signal collected by the spectrometer.

The wavelet decomposition and reconstruction processor specifically includes the following processing steps: first, in the wavelet decomposition process, two parameters, the decomposition scale and the wavelet base, need to be set; secondly, the approximate or The detailed spectral components are reconstructed, so that the original spectral signal data array is transformed into several sub-spectral signal data arrays with the same dimensions as the original spectral signal data array.

The specific operation of the partial least squares regression model integrator is as follows: respectively establish corresponding partial least squares regression submodels for the subspectral signal data arrays obtained after processing by the wavelet decomposition and reconstruction processor (in each In the sub-model, the principal component scores can choose different values), use the established sub-models to predict the test set sample indicators respectively, and obtain the corresponding prediction results; finally, weight the prediction results of each sub-model to obtain the final The predicted value of the sample index of . By comparing the root mean square error (Root Mean Square Error of Prediction, RMSEP) value of the model under different decomposition scales and wavelet bases, determine the appropriate decomposition scale, wavelet base and model parameters and save them for subsequent infrared spectra of new samples predictive analysis.

Because the present invention adopts above technical scheme, obtain following effect:

Through wavelet decomposition and reconstruction transformation, the construction of multiple models is realized, which overcomes the difficulty of extracting spectral signal information by the single model method; by separately determining the number of factors for different sub-models, the effective information is fully extracted and the infrared spectrum is improved. Analyze the predictive accuracy and robustness of the model.

Description of drawings

Figure 1 Schematic diagram of the infrared spectroscopy quantitative analysis method and device for multi-scale regression;

Figure 2 near-infrared spectrum;

Figure 3 Schematic diagram of the operation of the multi-scale partial least squares regression method;

Figure 4. Effect of decomposition scale on multi-scale partial least squares regression method

Fig. 5 Schematic diagram of wavelet decomposition and reconstruction processor operation;

Fig. 6. The correlation diagram of the predicted value and the reference value of the infrared spectrum model.

Detailed ways

The specific embodiment will be described in conjunction with the following implementation examples. Taking the near-infrared spectrum of pears as an example, a multi-scale regression model was constructed for the sugar content index inside pears.

Figure 1 is a schematic diagram of the multi-scale regression infrared spectrum quantitative analysis method and device, and Figure 2 is a near-infrared spectrum diagram with a spectral range of 750-1800 nm, and each spectrum includes 1051 data points. All samples were divided into calibration set and test set according to the ratio of 2:1.

All samples are decomposed by wavelet. When decomposing, we choose db4 wavelet base here, and the decomposition scale is from 1 to 20. Figure 3 shows the operation diagram of the multi-scale partial least squares regression method. It can be seen from the figure that the spectrum is reconstructed by wavelet decomposition. After construction, a series of sub-spectral signal matrices with the same dimensions as the original spectrum are obtained. All sub-spectral spectra are quite different, and the useful information contained in each sub-spectral matrix is also different. Therefore, when modeling, only one principal component number can be selected for the original spectrum, while the multi-scale partial least squares regression method can select different principal component numbers for the subspectral matrix, making information extraction more flexible and sufficient.

Figure 4 shows the influence of the decomposition scale on the multi-scale partial least squares regression method. The dotted line in the figure represents the predicted root mean square error of the traditional partial least squares regression method, and the solid line represents the predicted root mean square error of the test set with The change curve of the wavelet decomposition scale, as can be seen from the figure, when the 5-scale wavelet decomposition is used for reconstruction, the prediction root mean square error of the test set is the smallest. Therefore, the decomposition scale and wavelet basis of the wavelet decomposition and reconstruction processor can be determined.

Figure 5 is a schematic diagram of the operation of the wavelet decomposition and reconstruction processor. Establish correction models for each sub-spectral signal matrix in the figure, and the corresponding principal component numbers are 7, 4, 5, 3, 2, 4 respectively. When not processed by wavelet decomposition and reconstruction processor, the principal components of the model The number is 4, it can be seen that the multi-scale partial least squares regression analysis method is more conducive to the extraction of spectral information.

Figure 6 shows the correlation diagram between the predicted value and the reference value of the infrared spectrum multi-scale partial least squares regression model. It can be seen from the figure that better prediction results are obtained.

The effect of the present invention is: by using the multi-scale partial least squares regression method to quantitatively analyze the infrared spectrum, it is helpful to extract useful information of spectral signals at different frequencies, and obtain more accurate and stable prediction results. Therefore, the method and device of the invention are expected to become a very promising infrared spectroscopic analysis method.

Claims (1)

1. The infrared spectrum quantitative analysis method and device based on multi-scale regression, characterized in that: comprising a spectrometer connected to a data signal line, a preprocessor, a wavelet decomposition and reconstruction processor, and a partial least squares regression model integrator; The infrared spectrum includes mid-infrared and near-infrared spectra, that is, the wavelength range is: 780 nm-50000 nm; The preprocessor adopts centralization and vector normalization to process the original signal collected by the spectrometer.

Images