Method for reconstructing related components and extracting important information in seawater mixing spectrum

Abstract

The present invention discloses a method for reconstructing relevant components and extracting important information in the mixed spectrum of seawater, which relates to the technical field of marine detection. The method includes performing sparse representation on the total spectral information; calculating the optimal over-complete matrix; based on the optimal over-complete matrix, calculating the sparse coefficient matrix function and the overall function expression corresponding to different spectral component numbers, and iterating each variable until it converges to the global minimum value to obtain the optimal sparse coefficient matrix; multiplying the optimal over-complete matrix and the optimal sparse coefficient matrix to obtain the estimated values of each spectral component, and re-establishing a two-dimensional image to collect highly correlated data for analysis. The present invention removes the interference of irrelevant spectral information, highlights and excavates the spectral information related to the target substance, and while improving the robustness and accuracy of the detection ability of marine sensors, satisfies the requirement that the sensor can realize intelligent and real-time data analysis on site.

Description

Technical Field

[0001] The present invention relates to the technical field of marine detection, and in particular to a method for reconstructing relevant components and extracting important information in a seawater mixed spectrum. Background Art

[0002] Seawater is a complex system composed of multiple components. When using UV-VIS (ultraviolet-visible) spectroscopy to measure the spectral information of seawater, the full-band spectrum contains comprehensive information. However, using wavelength selection will inevitably lead to the loss of spectral information of target substances in the seawater system. Based on the results of spectral variable selection, a mathematical relationship model is established using the selected spectral variables and the corresponding target substance concentrations. Subsequently, the same variables in the new spectral signal can be substituted into the mathematical model to predict the corresponding concentrations. Therefore, the accuracy of spectral information related to the target substance is particularly crucial, directly affecting the accuracy and robustness of the established model. Therefore, how to exclude irrelevant information and remove noise and interference signals to isolate more effective relevant information for modeling is an important step in spectral water quality methods.

[0003] In related seawater quality detection, the generation of spectra follows the superposition law of Lambert-Beer: a parallel beam of monochromatic light is passed through a homogeneous and non-scattering solution, and the degree of absorption of the monochromatic light is proportional to the sample concentration and the optical path length of the solution passed through. Usually, the absorbances of different components in a multi-component solution simultaneously satisfy the linear superposition law. However, during the spectral sampling process of seawater with complex water quality components, it is not only affected by noise interference generated in spectrometers, pulsed light sources, and circuits, but also contains a large amount of chloride ions and bromide ions, forming non-linear spectral superpositions in the ultraviolet spectral band, and simultaneously interfering with the effective information spectral bands of target substances in the same region.

[0004] Facing the problem of interference from interference information to effective spectral signals, to improve the signal-to-noise ratio of the seawater spectral sensor system and maintain its stability, common methods for spectral smoothing and Fourier analysis can, to a certain extent, eliminate high-frequency random errors and improve the signal-to-noise ratio, thereby achieving the removal of spectral noise and interference. However, signal distortion and a decrease in spectral resolution are inevitable during the signal processing process. More importantly, common spectral processing means target the mixed overall spectral information of all components. The interference of irrelevant components in seawater to the analyte cannot be filtered out inside the spectrum, resulting in insufficient expression of the dynamic information of the spectral information components of the analyte to be measured or the relevant signals being submerged during the spectral processing process. Summary of the Invention

[0005] In order to overcome the above problems existing in the prior art, the present invention proposes a method for reconstructing relevant components and extracting important information in a seawater mixed spectrum.

[0006] The technical solution adopted by the present invention to solve its technical problems is: a method for reconstructing relevant components and extracting important information in a seawater mixed spectrum, including the following steps:

[0007] Step 1, introducing the spectra of individual components Perform sparse representation on the total spectrum ;

[0008] Step 2, initialize as any non-negative matrix, and alternately iterate and update the sparse coefficient matrix and the over-complete matrix to obtain the optimal over-complete matrix , where ;

[0009] Step 3, based on the optimal over-complete matrix obtained in Step 2, calculate the sparse coefficient matrix function corresponding to different spectral components k;

[0010] Step 4, add the sparse coefficient matrix functions corresponding to different spectral components k obtained in Step 3 to obtain the overall function expression , and perform iteration on each variable until it converges to the global minimum value to obtain the optimal sparse coefficient matrix ;

[0011] [[ID=StartFragment]]Step 5, multiply the optimal over-complete matrix obtained in Step 2 and the optimal sparse coefficient matrix obtained in Step 4 to obtain the estimated values of each spectral component, and evaluate the obtained estimated values of each spectral component;

[0012] Step 6, perform SVD decomposition on the estimated values of each spectral component obtained in Step ⑤, take the eigenvector with the largest eigenvalue in the load matrix obtained by the decomposition as the first principal component, and the eigenvector with the second largest eigenvalue as the second principal component. Select the first principal component and the second principal component to form a new two-dimensional characteristic variable space, use the sensitive wavelength variable point of the target detection substance as a labeled vector, and collect the points within the range of ±10% of the included angle between the origin and the labeled vector as the data for the next analysis and modeling.

[0013] For the above method for reconstructing relevant components and extracting important information in a seawater mixed spectrum, the formula for sparse representation in Step 1 is:

[0014]

[0015] where represents the over-complete matrix, ; represents a sparse coefficient matrix, ; as a constraint condition of sparsity, it represents the number of zero elements in each column is not less than , represents the spectrum of individual components.

[0016] For the method for reconstructing related components and extracting important information in a seawater mixed spectrum as described above, the calculation process of the optimal overcomplete matrix in step 2 is as follows:

[0017] Step 2.1, fix the overcomplete matrix unchanged, and use non - negative basis pursuit to act on the spectrum matrix of individual components and the overcomplete matrix , according to take the optimal solution of and constrain each column in to have no more than L non - zero terms;

[0018] Step 2.2, update the overcomplete matrix column by column. When updating the j - th column of the overcomplete matrix , calculate the residual matrix , is the j - th column of the overcomplete matrix , is the sparse coefficient matrix the j - th row of, and form the residual constraint matrix from the columns of the non - zero coefficient positions in the corresponding terms of the residual matrix ;

[0019] Step 2.3, use SVD decomposition to update the matrix obtained in step 2.2 to the product of the new vector representation and , so as to realize the iterative update of the j - th column of the overcomplete matrix and the j - th row of the sparse coefficient matrix ;

[0020] Step 2.4, after the overcomplete matrix is updated, perform normalization processing , so that the sum of the squares of each column of the elements in the overcomplete matrix is 1. Repeat steps 2.1 - 2.3, and take out the optimal overcomplete matrix<> .

[0021] For the method for reconstructing related components and extracting important information in a seawater mixed spectrum as described above, the evaluation of the obtained estimated values of each spectral component in step 5 specifically includes: finding the estimated value of the individual component The sum of the differences from the actual values of the corresponding spectral components is used as the decomposition error :

[0022]

[0023] where k represents the spectral component.

[0024] For the method for reconstructing relevant components and extracting important information in a seawater mixed spectrum described above, the sparse coefficient matrix function corresponding to the different numbers of spectral components obtained in step 3 has the expression:[[]]

[0025]

[0026] where represents the total spectral matrix for decomposition, represents the optimal over-complete matrix, is the sparsity of the sparse coefficient matrix, represents the sum of the reconstruction matrices of the remaining components except the k-th component, represents the over-complete matrix of the k-th component, represents the sparse coefficient matrix of the k-th component, represents the matrix and represents the sum of all elements in the matrix.

[0027] For the method for reconstructing relevant components and extracting important information in a seawater mixed spectrum described above, the overall function in step 4 has the expression:[[]]

[0028]

[0029] where k represents the spectral component, represents the optimal over-complete matrix, represents the total spectral matrix for decomposition, is the sparsity of the sparse coefficient matrix, represents the sparse coefficient matrix, represents the sum of all elements in the matrix.

[0030] For the method for reconstructing relevant components and extracting important information in a seawater mixed spectrum described above, the iterative formula for iteratively obtaining is:[[]]

[0031]

[0032] where represents the total spectral matrix for decomposition, is the sparsity of the sparse coefficient matrix, represents the optimal over-complete matrix, and t represents the number of iterations. represents the sum of the reconstruction matrices of the components other than the k-th component. represents the optimal over-complete matrix of the k-th component. represents the sparse coefficient matrix of the k-th component. represents the transposed matrix of.

[0033] The beneficial effect of the present invention is that the present invention discloses a method for reconstructing relevant components and extracting important information in seawater mixed spectra. Without the requirement of mastering the spectral signals of individual components and the system characteristics in the mixed spectra, the interference spectra are reversely decomposed and stripped, and spectra with a relatively high signal-to-noise ratio of the target detection substance information in the key ultraviolet band are obtained. Then, through the vector included angle direction of the loaded signal in PCR, the final effective key spectral information is obtained, so as to realize the removal of the interference of irrelevant spectral information when the marine spectral sensor faces the complex and mixed seawater spectral information, highlight the excavation of the spectral information related to the target substance, and while improving the robustness and accuracy of the detection ability of the marine sensor, meet the requirement that the sensor can realize intelligent and real-time data analysis on site. Brief Description of the Drawings

[0034] The present invention will be further described below in conjunction with the drawings and embodiments.

[0035] Figure 1 is the total mixed seawater spectral absorbance curve in the embodiment of the present invention;

[0036] Figure 2 is to Figure 1 an example of splitting the mixed spectrum in into a matrix and a coefficient matrix;

[0037] Figure 3 is the schematic diagram of the core steps of the reverse decomposition of the total seawater spectrum in the embodiment of the present invention;

[0038] Figure 4 is the schematic diagram of spectral recombination in the embodiment of the present invention;

[0039] Figure 5 is the schematic diagram of the recombination of the component spectra in the embodiment of the present invention;

[0040] Figure 6 [[ID=X]] is the schematic diagram of the relevant payload of the target detection substance in the embodiment of the present invention. Specific Embodiments

[0041] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below in conjunction with the drawings and specific embodiments.

[0042] The absorption spectrum generated by measuring seawater can be regarded as a mixed spectral signal formed by weighting the original spectral signals of multiple components. When facing unknown and complex seawater, the weight coefficients of these mixed weightings cannot be known in advance, and the spectral information of individual component signals is also unknown without being recorded. Therefore, without the requirement of mastering the spectral signals of individual components and the system characteristics in the mixed spectrum, the interference spectrum is reversely decomposed and peeled off, and a spectrum with a relatively high signal-to-noise ratio of the target detection substance information in the key ultraviolet band is obtained. Then, through the vector included angle direction of the signal loaded in PCR, the final effective and key spectral information is obtained.

[0043] The initial seawater absorbance spectrum is formed by the superposition of different components in seawater. The problem of single-channel blind source separation for the mixed total absorption spectrum can be expressed as:

[0044]

[0045] In the formula, represents the initial spectrum, that is, the total spectral data of all components of seawater; represents the component spectrum, that is, the spectral signal generated by each component; n represents the number of terms of the component spectral signals.

[0046] In this embodiment, the absorption spectra with strong influence on the absorption spectrum in the ultraviolet band are identified and separated. Specifically, it can be manifested as the COD spectrum, nitrite spectrum with strong absorption peaks in the ultraviolet band and the remaining spectral information. The total spectrum can be disassembled into the following four different component spectra (the spectra required can be decomposed according to the actual situation):

[0047]

[0048] That is, it represents the initial total component initial spectrum, and the four items on the right side of the equal sign represent the four different component spectra decomposed from the initial spectrum. Among them, represents the component spectrum formed by COD-related components; represents the component spectrum formed by nitrite; represents the spectrum composed of other spectral components with relatively small influence information including the remaining component spectra, containing interference signals carried by the measurement environment and the instrument itself, simply referred to as the remaining spectrum. As Figure 1 shown, Figure 1This is the total mixed seawater spectral absorbance curve of the present invention example. The spectral intensities and curves generated by different contents of molecules in the seawater to be measured are different. According to the Lambert-Beer law and the superposition principle, they are combined into a complete spectrum. The abscissa is the spectral wavelength points scanned by the spectrometer, and the ordinate is the absorbance. Due to the linear superposition mechanism among components and the interference and inhibition of unknown components on other components, it is possible to select the spectral maps that have a greater impact on the analyte for separation research to exclude the interference of other factors.

[0049] Therefore, component spectral separation can be regarded as blind source separation. A total observed spectrum is composed of the mixed original spectral signals of multiple components, and its mixing model can be expressed as:

[0050]

[0051] In the formula, is the original total spectral matrix; S is the source matrix, which contains all the source signals to be solved; the matrix B` defines the contribution degree of the source signals to the total spectrum through the weight coefficients; E is the noise matrix or model error. Except for the total spectral matrix , other matrices including the mixing matrix A, the source matrix S, and the noise matrix E are all unknown.

[0052] The spectral feature bases of different component information are represented by different overcomplete matrices, and their weight coefficients are represented by sparse coefficients, so that the reverse decomposition of the total spectral components is in the form of a linear sum of the products of different component matrices and the corresponding sparse coefficient matrices.

[0053] Given the training component spectral matrix , it represents that the i-th component data set contains M input components and N dimensions. In this embodiment, , the overcomplete matrix is satisfied, and ; the sparse coefficient matrix needs to satisfy the condition that few elements in the matrix are non-zero and non-negative, and satisfies under the constraint conditions of sparsity and non-negativity. As shown in Figure 2 , Figure 2 is an example of decomposing the mixed spectrum into the matrix and the coefficient matrix . For the total spectrum formed by mixing that needs to be split, it is necessary to explore the component spectra representing each component information from its matrix information, and then reverse decompose it from its total waveform load data into component load components with practical significance. Specifically, based on the differences in the absorbances of each component at different wavelengths, the total spectrum is reversely analyzed and disassembled into a matrix representing each component that can be sparsified and its sparse coefficient matrix .

[0054] The method for reconstructing relevant components and extracting important information in the seawater mixed spectrum disclosed in this embodiment. The core steps of the seawater total spectrum back decomposition are as Figure 3 shown. Solve the overcomplete matrix of each spectral component and the sparse coefficient matrix by using the original total matrix and the training component spectral matrix. And multiply each optimal overcomplete matrix by its corresponding sparse coefficient matrix to obtain the separated and reconstructed component spectral matrix of the estimated component spectra. In this example, includes the COD component spectrum, the nitrite component spectrum, and the background component spectrum. Use the trained overcomplete matrix and the relevant sparse matrix to reconstruct the component spectra. The specific steps are as follows:

[0055] Step 1, introduce the individual component spectrum to perform sparse representation on the total spectrum :

[0056]

[0057] Among them, represents the overcomplete matrix, ; represents the sparse coefficient matrix, ; is used as the constraint condition of sparsity to represent that the number of zero elements in each column is not less than , represents the individual component spectrum;

[0058] It is stipulated that in the optimization problem expression, represents the F-norm, is the number of non-zero terms in the vector, expressed as the 0-norm. Under this constraint condition, make approach the minimum value to approximate the optimal solution.

[0059] Step 2, initialize as any non-negative matrix, and alternately iterate and update the sparse coefficient matrix and the overcomplete matrix to obtain the optimal overcomplete matrix , where ;

[0060] The calculation process of the optimal overcomplete matrix is as follows:

[0061] Step 2.1, fix the overcomplete matrix unchanged, and use the non-negative basis pursuit to act on the individual component spectral matrix With an overcomplete matrix , according to take the optimal solution of and constrain each column in to have no more than L non-zero terms;

[0062] Step 2.2, update the overcomplete matrix column by column. When updating the j-th column of the overcomplete matrix , calculate the residual matrix , , is the j-th column of the overcomplete matrix , is the j-th row of the sparse coefficient matrix . The columns corresponding to the non-zero coefficient positions in the corresponding terms of the residual matrix form the residual constraint matrix ;

[0063] Step 2.3, use SVD decomposition to update the result obtained in Step 2.2 to the product of the new vector representation and , thereby realizing the iterative update of the j-th column of the overcomplete matrix and the j-th row of the sparse coefficient matrix ;

[0064] Step 2.4, perform normalization processing on the updated overcomplete matrix , satisfying that the sum of the squares of each column of the elements in the overcomplete matrix is 1. Repeat Steps 2.1 - 2.3 to obtain the optimal overcomplete matrix .

[0065] Subsequently, calculate the optimal sparse coefficient matrix based on the optimal overcomplete matrix obtained in Step 2. Based on the optimal overcomplete matrix , apply it to the decomposition of the total spectrum and obtain the optimal sparse coefficient matrix that can satisfy , and ensure the non-negativity and sparsity of the optimal sparse coefficient matrix . Combining the above constraints, the specific spectral component optimization problem is expressed as: where

[0066]

[0067] is the optimal overcomplete matrix obtained in Step 2, represents the F-norm, and the obtained with the minimum value ensures , and the reconstructed component spectrum and the original component spectrum obtain the maximum similarity;​​ Approximating the minimum value ensures the sparsity of the matrix; let the constraint matrix is non - negative; k is the number of decomposition components; is the matrix sparsity.

[0068] Step 3: Based on the optimal over - complete matrix obtained in Step 2 , calculate the sparse coefficient matrix function corresponding to different spectral components k ;

[0069] The sparse coefficient matrix functions corresponding to different numbers of spectral components obtained The expression is:

[0070]

[0071] Among them, represents the total spectral matrix for decomposition, represents the optimal over - complete matrix, is the sparsity of the sparse coefficient matrix, represents the sum of the reconstruction matrices of the remaining components except the k - th component, represents the over - complete matrix of the k - th component, represents the sparse coefficient matrix of the k - th component, represents the matrix The sum of the elements in the p - th row and q - th column (i.e., all elements) of.

[0072] Step 4: Add the sparse coefficient matrix functions corresponding to different spectral components k obtained in Step 3 to obtain the overall function expression , and iterate each variable until it converges to the global minimum value to obtain the optimal sparse coefficient matrix ;

[0073] The specific process includes:

[0074] For the number of spectral components k and its corresponding variable , when the non - negative matrix is fixed, the optimization objective is , according to the non - negative sparse coding algorithm formula:

[0075] It can be derived that

[0076]

[0077] Y is the non - negative representation by and , so that we can let , the following iterative recurrence is performed with a non - negative initial matrix, and the variable to be found has a positive initial value and is always non - negative and non - increasing during the iteration:

[0078]

[0079] Among them, represents the total spectral matrix for decomposition, is the sparsity of the sparse coefficient matrix, represents the optimal over - complete matrix, t represents the number of iterations, represents the sum of the reconstruction matrices of the other components except the k - th component, represents the optimal over - complete matrix of the k - th component, represents the sparse coefficient matrix of the k - th component, represents the transpose matrix of.

[0080] The overall function in step 4 is expressed as:

[0081]

[0082] Among them, k represents the spectral component, represents the optimal over - complete matrix, represents the total spectral matrix for decomposition, is the sparsity of the sparse coefficient matrix, represents the sparse coefficient matrix, represents the sum of the elements in the p - th row and q - th column of the matrix (i.e., all elements).

[0083] Step 5, using the optimal over - complete matrix obtained in step 2 and the optimal sparse coefficient matrix obtained in step 4 are multiplied to obtain the estimated values of each spectral component, and the obtained estimated values of each spectral component are evaluated; Figure 4 is a schematic diagram of spectral recombination, which expresses the main methods and ideas based on inverse spectral component decomposition, realizes the optimal total spectral decomposition and the solution of the sparse coefficient matrix. The decomposition method of the inverse cooling load is reflected in the successive iteration of matrices towards the optimal approximation. The main method of decomposing the reverse cooling load is: first, train for each component spectral component, and then solve the sparse matrix on the basis of satisfying sparsity and non - negativity, so as to realize the representation of the component information to be separated.

[0084] The evaluation of the obtained estimated values of each spectral component in step 5 specifically includes: finding the sum of the differences between the estimated value of a single component and the actual value of the corresponding spectral component as the decomposition error :

[0085]

[0086] Among them, k represents the spectral component.

[0087] Figure 5 It is a comparison schematic diagram of the reconstructed spectrum of COD-related component information and the original component spectrum. Among them, the blue solid line is the reconstructed spectrum, and the red dotted line is the original component spectrum. There is a certain drift and jitter in the reconstructed line, but the overall trend is consistent with the original line. After using the non-negative sparse representation algorithm for matrix decomposition and component information reconstruction, on the basis of solving sparsity and non-negativity, balancing the relationship between global search and optimal solution development can make the reconstructed spectrum more capable of information discrimination and the reconstruction error smaller.

[0088] Step 6: Select the corresponding target reconstructed spectrum. This component spectrum matrix is the target component spectrum separated and reconstructed from the total spectrum matrix through the above steps. Its spectral response to the target substance is clearer than the total spectrum, excluding the influence of the interfering substance spectrum. Perform SVD analysis on the concentration of the target analyte according to the component spectrum, so as to find the variable wavelength points with strong correlation with the target detection substance in this spectrum matrix for modeling.

[0089] Use SVD decomposition to decompose the target component spectrum matrix into the product of three matrices where U and V are respectively and orthogonal matrices; S is a matrix, and the diagonal elements satisfy . Take the square of the diagonal elements in the S matrix as the eigenvalues of matrix A, corresponding to the variances of the main factors.

[0090] Formula The first k columns of U, the first k rows and columns of S in the formula form the molar absorptivity matrix:

[0091]

[0092] Let \(n\) be the wavelength variable point, \(m\) be the number of samples, the \(E\) matrix be the score matrix, which represents the projection of the second - order spectral data on the new eigenvectors, that is, the projection in the new coordinate space; the \(V\) matrix is the loading matrix (principal component), which represents the projection direction of the eigenvectors in the new space. The variance of the data points in the loading matrix projected on the eigenvectors is the corresponding eigenvalue. Selecting the eigenvector with the largest eigenvalue means selecting the direction with the largest projection variance of the points, which is the first principal component with the highest information content; the second - best projection direction is in the orthogonal space of the best projection direction, which is the eigenvector corresponding to the second - largest eigenvalue and is the second principal component. Furthermore, select the first principal component and the second principal component to form a new two - dimensional characteristic variable space to achieve the dimensionality reduction and simplification of the original matrix. The new first principal component and the second principal component variable space carrying the most information can explain as much as possible the variance and differences of the matrix data. The score is the projection of the object on the new variable space. The loading plot drawn by the two explains the important relationships between different wavelength variable points of the samples in the new variable space.

[0093] Spectral matrix Draw a loading plot of the variable space composed of the first principal component and the second principal component. Each point in the loading plot corresponds to a wavelength variable point. Any variable point has a large positive correlation with the variable points in the same direction and a large negative correlation with the variable points in the opposite direction. Therefore, in the figure, the sensitive wavelength variable points of the target detection substance are used as a labeled vector to analyze and select the variable points with high correlation and importance with the target detection substance. The relationship between variables is defined using the projection of the new variable space.

[0094] The variables related to the target detection substance can be selected in the loading plot. The first principal component is used as the abscissa and the second principal component is used as the ordinate. Each wavelength variable is represented as a point in the two - dimensional loading plot. The wavelength variables closely related to its information are concentrated within the range of ±10% of the angle between the vector formed by the origin and the relevant variable point. The positive direction is the wavelength variable point with the strongest positive correlation, and the negative direction is the wavelength variable point with the strongest negative correlation. Selecting the wavelength variables within the range can be regarded as the vectors with the maximum positive and negative correlations with the information of the target detection substance. Select these variables and use them to build a model.

[0095] In this embodiment, the obtained loading plot is as Figure 6 shown. The \(x\) - axis is the first principal component and the \(y\) - axis is the second principal component. In this new variable interval, the orthogonal loadings composed of latent variables explain as much as possible the variance of the data matrix. The score is the projection of the object on the latent variables, which defines the sample relationship in the reduced variable space spanned by the latent variables. The variable importance in the spectral direction can be obtained through the bivariate loading plot respectively. The points corresponding to the target - related information will be within the range of a small angle on the same - direction straight line. Therefore, the variables in the region within the loading plot direction corresponding to the target detection substance - related information within the ±10° angle interval can explain the target information to the greatest extent.

[0096] The above embodiments are only exemplary embodiments of the present invention and are not intended to limit the present invention. The protection scope of the present invention is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions within the essence and protection scope of the present invention, and such modifications or equivalent substitutions should also be regarded as falling within the protection scope of the present invention.

Claims

1. A method for reconstructing relevant components and extracting important information in the seawater mixing spectrum, characterized in that: Including the following steps: Step 1, introduce the spectra of individual components For the total spectrum perform sparse representation; Step 2, initialization is an arbitrary non-negative matrix, the sparse coefficient matrix and the over-complete matrix are alternately iteratively updated to obtain the optimal over-complete matrix , where ; Step 3, based on the optimal over-complete matrix obtained in Step 2 , calculate the sparse coefficient matrix function corresponding to different spectral components k ; Step 4: Add the sparse coefficient matrix functions corresponding to different spectral components k obtained in Step 3 to obtain the overall function expression , and iterate each variable until converges to the global minimum value to obtain the optimal sparse coefficient matrix ; Step 5: Use the optimal over-complete matrix obtained in Step 2 and the optimal sparse coefficient matrix obtained in Step 4 to multiply and obtain the estimated values of each spectral component , and evaluate the obtained estimated values of each spectral component; Step 6: Use SVD decomposition to process the estimated values of each spectral component obtained in Step 5 Perform decomposition. Take the eigenvector with the largest eigenvalue in the obtained loading matrix as the first principal component, and the eigenvector with the second largest eigenvalue as the second principal component. Select the first principal component and the second principal component to form a new two-dimensional characteristic variable space. Use the sensitive wavelength variable points of the target detection substance as a labeled vector, and collect the points within the range of ±10% of the angle between the origin and the labeled vector as the data for the next analysis and modeling; The formula for sparse representation in step 1 is as follows: Among them, represents an overcomplete matrix, ; represents a sparse coefficient matrix, ; As a constraint condition for sparsity, it means that the number of zero elements in each column is not less than , represents the spectrum of a single component.

2. The method for reconstructing relevant components and extracting important information in a seawater mixing spectrum according to claim 1, characterized in that, The optimal over-complete matrix in step 2 The calculation process is as follows: Step 2.1, fix the overcomplete matrix Remain unchanged, and use non - negative basis pursuit to act on the individual component spectral matrix and the overcomplete matrix , according to Take the optimal solution of and constrain each column in not to exceed L non - zero terms; Step 2.2, update the overcomplete matrix column by column When updating the j-th column of the overcomplete matrix , calculate the residual matrix , is the j-th column of the overcomplete matrix , is the j-th row of the sparse coefficient matrix . The columns at the non-zero coefficient positions of the corresponding terms in the residual matrix form the residual constraint matrix ; Step 2.3, using SVD decomposition to update the result obtained in Step 2.2 to a new vector representation and the product of, so as to realize the iterative update of the j-th column of the overcomplete matrix and the j-th row of the sparse coefficient matrix ; Step 2.4, overcomplete matrix Normalize after updating , such that the sum of squares of each column of the elements in the overcomplete matrix is 1. Repeat steps 2.1 - 2.3 to obtain the optimal overcomplete matrix .

3. A method for reconstructing relevant components and extracting important information in a seawater mixing spectrum, according to claim 1, characterized in that The evaluation of the obtained spectral component estimates in step 5 specifically includes: obtaining the sum of the differences between the individual component estimates and the actual values of the corresponding spectral components as the decomposition error : where k represents spectral components.

4. A method for reconstructing relevant components and extracting important information in a seawater mixing spectrum, characterized in that, The sparse coefficient matrix function corresponding to the different spectral component numbers obtained in the step 3 The expression is as follows: Among them, represents the total spectral matrix for decomposition, represents the optimal overcomplete matrix, is the sparsity of the sparse coefficient matrix, represents the sum of the reconstruction matrices of the remaining components except the k-th component, represents the overcomplete matrix of the k-th component, represents the sparse coefficient matrix of the k-th component, represents the matrix the sum of all elements in.

5. A method for reconstructing relevant components and extracting important information in a seawater mixing spectrum, characterized in that, The overall function expression in step 4 is expressed as: where k represents spectral components, represents the optimal over-complete matrix, represents the total spectral matrix for decomposition, is the sparsity of the sparse coefficient matrix, represents the sparse coefficient matrix, represents matrix the sum of all elements of.

6. A method for reconstructing relevant components and extracting important information in a seawater mixed spectrum according to claim 1, characterized in that, The iterative calculation in step 4 has the following iterative formula: Among them, represents the total spectral matrix for decomposition, is the sparsity of the sparse coefficient matrix, represents the optimal over-complete matrix, t represents the number of iterations, represents the sum of the reconstruction matrices of the other components except the k-th component, represents the optimal over-complete matrix of the k-th component, represents the sparse coefficient matrix of the k-th component, represents the transpose matrix of.

Images