	US 12,450,316 B1
	Detection method, system, medium and device based on multi-spectral data fusion
Ying Kan, Beijing (CN); Ke Li, Beijing (CN); Zhengdong Zhang, Beijing (CN); Dan Song, Beijing (CN); Xin Zhang, Beijing (CN); Chaomin Ding, Beijing (CN); Fan Liu, Beijing (CN); and Qi Li, Beijing (CN)
Assigned to National Institute of Metrology China, Beijing (CN)
Filed by National Institute of Metrology China, Beijing (CN)
Filed on Jun. 18, 2025, as Appl. No. 19/241,593.
Application 19/241,593 is a continuation of application No. PCT/CN2025/076911, filed on Feb. 12, 2025.
Claims priority of application No. 202410697313.0 (CN), filed on May 31, 2024.
Int. Cl. G06N 20/00 (2019.01); G06F 18/25 (2023.01)

CPC G06F 18/251 (2023.01) [G06N 20/00 (2019.01)]

17 Claims

1. A detection method based on multi-spectral data fusion, comprising:

collecting spectrum: collecting near infrared spectrum and mid infrared spectrum of a substance to be detected to obtain a near infrared spectrum matrix and a mid infrared spectrum matrix of the substance to be detected;

constructing a detection model: constructing the detection model by a following formula,

Y=XP^TBQ

wherein X is an input matrix, Y is an output matrix, B is a coefficient matrix, and P and Q are load matrices of X and Y respectively; and T is a score matrix of an independent variable X;

training the detection model; and

predicting detection indexes;

wherein steps of training the detection model comprise:

collecting training samples to construct training sets, wherein the training sets comprise near infrared spectra and mid infrared spectra of a plurality of the training samples and index values of detection indexes of the plurality of the training samples;

dividing the training sets into a calibration set and a verification set to obtain a near infrared spectrum matrix, a mid infrared spectrum matrix and an index matrix of the calibration set and a near infrared spectrum matrix, a mid infrared spectrum matrix and an index matrix of the verification set;

performing spectrum preprocessing on near infrared spectra and mid infrared spectra of the calibration set and the verification set respectively to obtain a near infrared spectrum preprocessing matrix and a mid infrared spectrum preprocessing matrix of the calibration set, and a near infrared spectrum preprocessing matrix and a mid infrared spectrum preprocessing matrix of the verification set, wherein the spectrum preprocessing comprises derivative processing or/and vector normalization processing, and the derivative processing comprises first-order derivative processing or/and higher-order derivative processing;

performing spectrum variable screening processing on the near infrared spectra, the mid infrared spectra, near infrared spectra after the spectrum preprocessing and mid infrared spectra after the spectrum preprocessing of the calibration set and the verification set respectively to obtain a near infrared spectrum variable screening matrix, a mid infrared spectrum variable screening matrix, a preprocessing near infrared spectrum variable screening matrix and a preprocessing mid infrared spectrum variable screening matrix of the calibration set, and a near infrared spectrum variable screening matrix, a mid infrared spectrum variable screening matrix, a preprocessing near infrared spectrum variable screening matrix and a preprocessing mid infrared spectrum variable screening matrix of the verification set, wherein the spectrum variable screening processing comprises competitive adaptive reweighted sampling processing or/and variable importance projection processing;

inputting the near infrared spectrum matrix and the index matrix of the calibration set into the detection model, and training the detection model to obtain a first coefficient matrix;

inputting the mid infrared spectrum matrix and the index matrix of the calibration set into the detection model, and training the detection model to obtain a second coefficient matrix;

inputting the near infrared spectrum preprocessing matrix and the index matrix of the calibration set into the detection model, and training the detection model to obtain a third coefficient matrix;

inputting the mid infrared spectrum preprocessing matrix and the index matrix of the calibration set into the detection model, and training the detection model to obtain a fourth coefficient matrix;

inputting the near infrared spectrum matrix of the calibration set into the detection model corresponding to the first coefficient matrix to obtain a calibration set first index prediction matrix composed of predicted values of detection indexes of the calibration set;

inputting the mid infrared spectrum matrix of the calibration set into the detection model corresponding to the second coefficient matrix to obtain a calibration set second index prediction matrix composed of the predicted values of the detection indexes of the calibration set;

inputting the near infrared spectrum preprocessing matrix of the calibration set into the detection model corresponding to the third coefficient matrix to obtain a calibration set third index prediction matrix composed of the predicted values of the detection indexes of the calibration set;

inputting the mid infrared spectrum preprocessing matrix of the calibration set into the detection model corresponding to the fourth coefficient matrix to obtain a calibration set fourth index prediction matrix composed of the predicted values of the detection indexes of the calibration set;

inputting the near infrared spectrum matrix of the verification set into the detection model corresponding to the first coefficient matrix to obtain a verification set first index prediction matrix composed of predicted values of detection indexes of the verification set;

inputting the mid infrared spectrum matrix of the verification set into the detection model corresponding to the second coefficient matrix to obtain a verification set second index prediction matrix composed of the predicted values of the detection indexes of the verification set;

inputting the near infrared spectrum preprocessing matrix of the verification set into the detection model corresponding to the third coefficient matrix to obtain a verification set third index prediction matrix composed of the predicted values of the detection indexes of the verification set;

inputting the mid infrared spectrum preprocessing matrix of the verification set into the detection model corresponding to the fourth coefficient matrix to obtain a verification set fourth index prediction matrix composed of the predicted values of the detection indexes of the verification set;

obtaining a first verification value, a second verification value, a third verification value and a fourth verification value respectively according to verification indexes through the verification set first index prediction matrix, the verification set second index prediction matrix, the verification set third index prediction matrix, the verification set fourth index prediction matrix and the index matrix, wherein the verification indexes are used for representing prediction performance of the detection model;

taking an optimal value of the first verification value, the second verification value, the third verification value and the fourth verification value as an optimal verification value, taking the optimal verification value as a threshold value, and taking a range of the threshold value in a direction of improving the prediction performance of the detection model as a threshold value range;

performing data-level fusion on the near infrared spectrum matrix and the mid infrared spectrum matrix of the calibration set to form a calibration set first data-level fusion matrix, and inputting the calibration set first data-level fusion matrix and the index matrix of the calibration set into the detection model to train the detection model to obtain a fifth coefficient matrix;

performing the data-level fusion on the near infrared spectrum preprocessing matrix and the mid infrared spectrum preprocessing matrix of the calibration set to form a calibration set second data-level fusion matrix, and inputting the calibration set second data-level fusion matrix and the index matrix of the calibration set into the detection model to train the detection model to obtain a sixth coefficient matrix;

taking the calibration set first data-level fusion matrix as the input matrix, inputting the detection model corresponding to the fifth coefficient matrix, and obtaining a calibration set fifth index prediction matrix composed of the predicted values of the detection indexes of the calibration set;

taking the calibration set second data-level fusion matrix as the input matrix, inputting the detection model corresponding to the sixth coefficient matrix, and obtaining a calibration set sixth index prediction matrix composed of the predicted values of the detection indexes of the verification set;

taking a verification set first data-level fusion matrix formed by data-level fusion of the near infrared spectrum matrix and the mid infrared spectrum matrix of the verification set as the input matrix, and inputting the detection model corresponding to the fifth coefficient matrix to obtain a verification set fifth index prediction matrix composed of the predicted values of the detection indexes of the verification set;

taking a verification set second data-level fusion matrix formed by data-level fusion of the near infrared spectrum preprocessing matrix and the mid infrared spectrum preprocessing matrix of the verification set as the input matrix, and inputting the detection model corresponding to the sixth coefficient matrix to obtain a verification set sixth index prediction matrix composed of the predicted values of the detection indexes of the verification set;

respectively obtaining a fifth verification value and a sixth verification value according to the verification indexes through the verification set fifth index prediction matrix, the verification set sixth index prediction matrix and the index matrix;

performing feature-level fusion on the near infrared spectrum variable screening matrix and the mid infrared spectrum variable screening matrix of the calibration set to form a calibration set first feature-level fusion matrix, inputting the calibration set first feature-level fusion matrix and the index matrix of the calibration set into the detection model, and training the detection model to obtain a seventh coefficient matrix;

performing the feature-level fusion on the preprocessing near infrared spectrum variable screening matrix and the preprocessing mid infrared spectrum variable screening matrix of the calibration set to form a calibration set second feature-level fusion matrix, inputting the calibration set second feature-level fusion matrix and the index matrix of the calibration set into the detection model, and training the detection model to obtain an eighth coefficient matrix;

taking the calibration set first feature-level fusion matrix as the input matrix, inputting the detection model corresponding to the seventh coefficient matrix, and obtaining a calibration set seventh index prediction matrix composed of the predicted values of the detection indexes of the calibration set;

taking the calibration set second feature-level fusion matrix as the input matrix, inputting the detection model corresponding to the eighth coefficient matrix, and obtaining a calibration set eighth index prediction matrix composed of the predicted values of the detection indexes of the calibration set;

taking a verification set first feature-level fusion matrix formed by feature-level fusion of the near infrared spectrum variable screening matrix and the mid infrared spectrum variable screening matrix of the verification set as the input matrix, and inputting the detection model corresponding to the seventh coefficient matrix to obtain a verification set seventh index prediction matrix composed of the predicted values of the detection indexes of the verification set;

taking a verification set second feature-level fusion matrix formed by data-level fusion of the preprocessing near infrared spectrum variable screening matrix and the preprocessing mid infrared spectrum variable screening matrix of the verification set as the input matrix, and inputting the detection model corresponding to the eighth coefficient matrix to obtain a verification set eighth index prediction matrix composed of the predicted values of the detection indexes of the verification set;

respectively obtaining a seventh verification value and an eighth verification value according to the verification indexes through the verification set seventh index prediction matrix, the verification set eighth index prediction matrix and the index matrix;

screening the first verification value to the eighth verification value according to the threshold value range, and screening out a verification value group within the threshold value range; and

carrying out decision-level fusion on a calibration set index matrix corresponding to the verification value group to obtain a calibration set decision-level fusion matrix, inputting the calibration set decision-level fusion matrix and the calibration set index matrix into the detection model, and training the detection model to obtain a ninth coefficient matrix, wherein the ninth coefficient matrix is a trained coefficient matrix;

wherein steps of predicting the detection indexes comprise:

performing the data-level fusion on the near infrared spectrum matrix and the mid infrared spectrum matrix of the substance to be detected to obtain a spectral fusion matrix; and

inputting the spectral fusion matrix of the substance to be detected as the input matrix into a trained detection model to obtain a first index matrix as the output matrix; and taking the first index matrix as a detection result of the detection indexes of the substance to be detected; and

wherein the index matrix is a matrix composed of index values of detection indexes determined by measurements; the first index matrix is a matrix composed of index values of detection indexes predicted by the detection model; and the detection indexes comprise one or more of a flash point, a pour point, density and kinematic viscosity.