A method for constructing a solid-state liquor authenticity identification model

Abstract

The present application relates to the technical field of liquor authenticity identification, and discloses a method for constructing an authenticity identification model of solid-state liquor, which comprises the following steps: first, obtaining absolute concentration data of cyanide, anions and trace metal elements in samples; determining edge weights in combination with thermodynamic coordination constants and shielding factors calculated based on anions, and constructing a reference weighted adjacency matrix; injecting virtual noise variables into metal concentration data to cause perturbation, and extracting network centrality drift variance; and finally, fusing the extracted static topological features and the drift variance to generate a high-order feature vector. By using a chromatograph and a spectrometer to obtain absolute concentration data of non-aroma substances such as cyanide, anions and heavy metals in liquor samples, the problem that aroma substances are easily changed in absolute concentration by artificially adding corresponding chemical reagents in the existing identification technology, thereby leading to model misjudgment, is solved, and classification and identification of solid-state liquor and edible alcohol blending liquor are realized.

Description

Technical Field

[0001] This invention relates to the field of liquor authenticity identification technology, specifically a method for constructing a model for identifying the authenticity of solid-state fermented liquor. Background Technology

[0002] Baijiu is a distilled spirit made primarily from grains through processes such as saccharification, fermentation, and distillation. Based on production processes, it is mainly divided into solid-state fermented baijiu and baijiu blended using edible alcohol and additives. Because solid-state fermented baijiu has a longer fermentation cycle and higher production costs than baijiu blended with edible alcohol, there are instances in the market where baijiu blended with edible alcohol is labeled and sold as solid-state fermented baijiu. Therefore, to confirm the actual brewing process of baijiu, it is necessary to use instrumental analysis and data processing techniques to identify the process category of baijiu samples.

[0003] In existing technologies, methods for constructing baijiu (Chinese liquor) identification models typically employ gas chromatography and other detection equipment to obtain the absolute concentration data of aroma compounds such as esters and alcohols in the baijiu sample. This concentration data is then used as classification feature parameters and input into a data analysis model for classification calculations to achieve identification. However, in practice, aroma compounds such as esters and alcohols are chemical reagents that can be obtained through conventional chemical processes. It is easy to artificially add corresponding aroma-producing chemical reagents directly to edible alcohol, thereby artificially altering the absolute concentration data of aroma compounds in the sample. When the artificial addition of reagents causes the absolute concentration of aroma compounds in the sample to reach the parameter range of naturally fermented solid-state baijiu, the identification model that uses aroma compound data as classification parameters will output incorrect classification results, leading to model misjudgment. Summary of the Invention

[0004] This invention provides a method for constructing a model to identify the authenticity of solid-state fermented baijiu, which solves the problem of misjudgment caused by the model when artificially added reagents make the absolute concentration of aroma substances in the sample reach the parameter range of naturally fermented solid-state baijiu.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing a model for identifying the authenticity of solid-state fermented liquor, comprising the following steps: S1. Establish a standardized sample set and obtain the absolute concentration data of core markers, anions and trace metal elements in each sample; S2. Map the absolute concentration data into a directed star network graph, define the core marker as the core node, define the trace metal element as the peripheral node set, and establish directed edges from the core node to each node in each peripheral node set. S3. Introduce the thermodynamic coordination stability constant and the dynamic shielding factor calculated based on the anion, calculate the weight of each directed edge, and construct a benchmark weighted adjacency matrix characterizing the intensity of static chemical interaction of the sample. S4. Inject virtual noise variables into the absolute concentration data of the trace metal elements, reconstruct the perturbation-enhanced weighted adjacency matrix, and extract the centrality drift variance by comparing the principal eigenvectors before and after the perturbation. S5. Extract the static topological features of the benchmark weighted adjacency matrix and fuse them with the centrality drift variance to generate a higher-order topological chemical feature vector. Input the higher-order topological chemical feature vector into the principal component analysis model for dimensionality reduction. Input the dimensionality-reduced principal component data into the random forest classification model for training to establish the classification decision boundary and generate a true / false identification model.

[0006] Preferably, in step S1, obtaining the absolute concentration data of core markers, anions, and trace metal elements in each sample includes the following steps: The absolute concentration of cyanide in the sample was determined using ion chromatography as the data for the core biomarker. The absolute concentration of a specific anion group, including chloride ions, nitrite ions, bromide ions, nitrate ions, sulfate ions, and phosphate ions, in a sample is determined using an ion chromatograph, and the data for the anions are obtained. The absolute concentration of a specific element group was determined by a combination of inductively coupled plasma mass spectrometry and inductively coupled plasma atomic emission spectrometry, and this concentration was used as data for the trace metal elements. The specific element group included lead, cadmium, chromium, arsenic, nickel, tin, copper, iron, manganese, and phosphorus.

[0007] Preferably, in step S2, mapping the absolute concentration data into a directed star network diagram includes the following steps: Cyanide is set as a globally unique node and designated as the core node of the network graph. Based on the material source mechanism of fermentation, a specific element group is divided into a heterogeneous, non-overlapping primary transition metal group and an exogenous trace impurity group, and all elements in these two groups are regarded as the peripheral node set, wherein the primary transition metal group includes copper, iron and manganese. Construct a one-way association path starting from the core node and pointing to each specific element in the set of peripheral nodes to generate the directed edges.

[0008] Preferably, in step S3, calculating the weight of each directed edge includes the following steps: The absolute concentrations of the core biomarker and the corresponding trace metal elements are extracted, and the thermodynamic coordination stability constants of the core biomarker and the trace metal elements under standard conditions are obtained. Multiply the absolute concentration of the core marker, the absolute concentration of the trace metal element, and the thermodynamic coordination stability constant together. After adding the obtained product result to the basic preset offset constant, perform a natural logarithm operation to obtain the initial weight for the directed edge; traverse the set of peripheral nodes to obtain the initial weights of all directed edges.

[0009] Preferably, in step S3, constructing the benchmark weighted adjacency matrix characterizing the intensity of static chemical interactions of the samples includes the following steps: Extract the absolute concentration of each anion and the absolute value of the charge number corresponding to each anion, calculate the product of the absolute concentration of each anion and the square of the absolute value of the corresponding charge number, sum the products of all anions and take half of the result to obtain the effective ionic strength of the anion. The effective ionic strength of the anion is multiplied by the preset system dielectric decay constant and the negative value is obtained. The exponent of the negative value is then calculated with the natural constant as the base to generate the dynamic shielding factor. The initial weights of each directed edge are multiplied by the dynamic shielding factor to obtain the baseline weights of each directed edge; the baseline weights of each directed edge are aggregated to form the baseline weighted adjacency matrix.

[0010] Preferably, in step S4, injecting virtual noise variables into the absolute concentration data of the trace metal elements includes the following steps: Construct a set of white noise random variables that follow a Gaussian distribution; The absolute concentrations of each element in the trace metal element are fused and superimposed with the extracted white noise random variable to generate the trace metal element perturbation concentration data.

[0011] Preferably, in step S4, reconstructing the perturbation-enhanced weighted adjacency matrix and extracting the centrality drift variance by comparing the principal eigenvectors before and after the perturbation includes the following steps: While keeping the value of the dynamic shielding factor unchanged, the perturbation concentration data is substituted into the edge weight calculation process to generate new weights for each edge after perturbation, and the weighted adjacency matrix after perturbation is reconstructed accordingly. Eigenvalue decomposition is performed on the baseline weighted adjacency matrix and the perturbation-enhanced weighted adjacency matrix respectively, and the pre-perturbation main eigenvector corresponding to the maximum eigenvalue of the baseline weighted adjacency matrix and the post-perturbation main eigenvector corresponding to the maximum eigenvalue of the perturbation-enhanced weighted adjacency matrix are extracted. The corresponding feature components in the pre-perturbation principal feature vector and the post-perturbation principal feature vector of each element in the original transition metal group are extracted respectively. The difference between the characteristic components of the same element before and after perturbation is calculated and squared. The squared results of all elements in the original transition metal group are summed and the average value is calculated. The average value is output as the centrality drift variance.

[0012] Preferably, in step S5, extracting the static topological features of the benchmark weighted adjacency matrix and fusing them with the centrality drift variance to generate a higher-order topological chemical feature vector includes the following steps: The graph weighted density is extracted by summing and averaging all the benchmark weights in the benchmark weighted adjacency matrix. The values ​​corresponding to the core nodes are extracted from the perturbation-pre-main feature vector and used as the baseline centrality components. The graph weighted density, the baseline centrality component, the centrality drift variance, and the baseline weights of each directed edge are numerically concatenated according to a preset dimensional order to generate the higher-order topological chemical feature vector.

[0013] Preferably, in step S5, the higher-order topological chemical feature vector is input into the principal component analysis model for dimensionality reduction, and the dimensionality-reduced principal component data is input into the random forest classification model for training to establish the classification decision boundary and generate a true / false identification model, including the following steps: The high-order topological chemical feature vectors of samples known to be pure grain solid-state baijiu and samples known to be blended baijiu made with edible alcohol in the standardized sample set are aggregated to generate a global feature matrix. The global feature matrix is ​​centered and the covariance matrix is ​​calculated. Then, principal components that meet the preset cumulative variance contribution rate threshold are extracted by eigenvalue decomposition, and a low-dimensional principal component analysis space is constructed. The high-order topological chemical feature vectors of each sample are projected into the low-dimensional principal component analysis space, and the classification decision boundary is delineated in the space according to the clustering rules of the projected coordinates.

[0014] This invention provides a method for constructing a model for identifying genuine and counterfeit solid-state fermented liquor. It has the following beneficial effects: 1. This invention obtains the absolute concentration data of non-aroma substances such as cyanide, anions, and heavy metals in baijiu samples using chromatographs and spectrometers, and uses principal component analysis to perform dimensionality reduction and classification boundary delineation on the absolute concentration data. The non-aroma substance data is used to replace esters and alcohols as aroma substances as classification parameters, which solves the problem that the absolute concentration of aroma substances can be easily changed by adding corresponding chemical reagents, leading to model misjudgment in existing identification technologies. This invention realizes the classification and identification of solid-state baijiu and baijiu blended with edible alcohol.

[0015] 2. This invention establishes a multi-dimensional feature data sample database by collecting absolute concentration data of cyanide, specific anion groups, and trace metal elements from solid-state fermented liquor and blended liquor samples made with edible alcohol. This database provides training sample inputs for feature extraction and decision boundary construction of the principal component analysis model, enabling the model parameters to include the numerical distribution characteristics of the above-mentioned trace substances under different fermentation and blending environments. This reduces the interference of fluctuations in a single absolute concentration index caused by regional water quality differences on the model's identification output results and improves the classification accuracy of the model when processing test samples with different backgrounds.

[0016] 3. This invention maps the absolute concentration data of trace substances into a directed star network graph with cyanide as the core node and trace metal elements as the peripheral nodes. It uses the thermodynamic coordination stability constant and the dynamic shielding factor based on anion calculation to set the edge weights of the graph. In the algorithm, virtual noise variables are added to the absolute concentration data to extract the centrality drift variance before and after network reconstruction. The static absolute concentration of substances is converted into a high-order feature vector containing graph topology and structural perturbation variables, so that the discrimination model has the ability to distinguish samples with similar absolute concentration values ​​but different element coordination correlation structures. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] To better understand the above technical solutions, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please refer to Figure 1. This embodiment of the invention provides a method for constructing a model for identifying the authenticity of solid-state fermented liquor, including the following steps: S1. Establish a standardized sample set and obtain the absolute concentration data of core markers, anions and trace metal elements in each sample; Furthermore, in step S1, the absolute concentration data of core biomarkers, anions, and trace metal elements in each sample are obtained, including the following steps: The absolute concentration of cyanide in the sample was determined using ion chromatography as the data for the core biomarker. The absolute concentrations of specific anion groups, including chloride ions, nitrite ions, bromide ions, nitrate ions, sulfate ions, and phosphate ions, in a sample are determined using an ion chromatograph, and the data are used as anion data. The absolute concentrations of specific element groups were determined by a combination of inductively coupled plasma mass spectrometry and inductively coupled plasma atomic emission spectrometry as data for trace metal elements. The specific element groups included lead, cadmium, chromium, arsenic, nickel, tin, copper, iron, manganese, and phosphorus. By employing the above technical solution, and utilizing the precision analytical instrument ion chromatography to detect the liquid phase system of baijiu, the absolute concentration of cyanide, a characteristic endogenous product of pure grain fermentation, is captured as the core biomarker data. This core biomarker data is then digitized and defined as a scalar through instrument analysis. Simultaneously, this instrument comprehensively measures the absolute concentrations of a specific anion group, including chloride, nitrite, bromide, nitrate, sulfate, and phosphate ions, as anion data. This anion data is defined as an anion concentration column vector. ; in, Representing the absolute concentration variables of specific anion groups, this study further employs inductively coupled plasma mass spectrometry (ICP-MS) and inductively coupled plasma atomic emission spectrometry (ICP-AES) to perform in-depth quantitative analysis of specific element groups in the liquid phase system. This yields the absolute concentrations of elements including lead, cadmium, chromium, nickel, tin, copper, iron, manganese, and phosphorus as trace metal element data. This trace metal element data is defined as a metal concentration column vector. This method can extract non-significant trace metal fingerprints from the extremely complex matrix of baijiu (Chinese liquor), which are often overlooked by counterfeiters and difficult to accurately replicate artificially. It avoids the problem of traditional detection methods where aroma substances are easily confused with chemical flavorings. The technical blind spot of imitation through proportioning.

[0020] S2. Map the absolute concentration data into a directed star network graph, define the core markers as core nodes, define trace metal elements as peripheral node sets, and establish directed edges from the core nodes to each node in each peripheral node set. Furthermore, in step S2, mapping the absolute concentration data into a directed star network diagram includes the following steps: Cyanide is set as a globally unique node and designated as the core node of the network graph. Based on the material source mechanism of fermentation, a specific element group is divided into a heterogeneous, non-overlapping primary transition metal group and an exogenous trace impurity group, and all elements in these two groups are regarded as a set of peripheral nodes. The primary transition metal group includes copper, iron, and manganese. Construct a unidirectional path starting from the core node and pointing to each specific element in the set of peripheral nodes to generate directed edges; By adopting the above technical solution and structuring the physicochemical data through graph theory mapping algorithms, the mathematical structure of the directed star network graph is defined as follows: ,in, Represents a directed star network diagram. Represents the core set of graph theory, Representing a set of directed edges, oxides with a strong tendency to form complexes are defined as the core nodes of the global network topology. Meanwhile, a deeper analysis of specific element groups detected during solid-state fermentation reveals that, based on chemical properties, they are strictly classified into the primary transition metal group, which includes copper, iron, and manganese elements naturally introduced from traditional brewing vessels and fermented grains. And the group of exogenous trace impurities mainly introduced by humans from inferior blended water is defined as The aforementioned groups together constitute the set of peripheral nodes surrounding the core, defined as... The formula for its node mapping algorithm is: ; Then, a unidirectional path is constructed from the core node to each specific element in the set of peripheral nodes to generate directed edges. The algorithm formula for this edge mapping is as follows: ; in, This represents a single directed edge. Representing each independent and specific element node in the peripheral node set, the originally discrete and isolated concentration data of various trace elements are transformed into a structured spatial topology algorithm map that can intuitively map the complex colloidal equilibrium state inside the solid-state fermentation system, so that the element symbiotic relationship and network coordination tendency originally hidden behind non-primary aroma trace substances can be expressed mathematically in a concrete way.

[0021] S3. Introduce the thermodynamic coordination stability constant and the dynamic shielding factor based on anion calculation, calculate the weight of each directed edge, and construct a benchmark weighted adjacency matrix to characterize the intensity of static chemical interaction of the sample. Furthermore, in step S3, calculating the weight of each directed edge includes the following steps: The absolute concentrations of the core biomarkers and the corresponding trace metal elements were extracted, and the thermodynamic coordination stability constants of the core biomarkers and trace metal elements under standard conditions were obtained. Multiply the absolute concentrations of the core markers, the absolute concentrations of trace metal elements, and the thermodynamic coordination stability constants together. After adding the obtained product to the basic preset offset constant, perform a natural logarithm operation to obtain the initial weight for the directed edge; traverse the outer node set to obtain the initial weights of all directed edges. In step S3, constructing the benchmark weighted adjacency matrix characterizing the intensity of static chemical interactions in the samples includes the following steps: Extract the absolute concentration of each anion and the absolute value of the charge number corresponding to each anion. Calculate the product of the absolute concentration of each anion and the square of the absolute value of the corresponding charge number. Sum all the products of anions and take half of the result to obtain the effective ionic strength of the anion. The effective ionic strength of the anion is multiplied by the preset system dielectric decay constant and the negative value is obtained. The exponent of this negative value is then calculated with the natural constant as the base to generate the dynamic shielding factor. The initial weights of each directed edge are multiplied by the dynamic shielding factor to obtain the baseline weights of each directed edge; the baseline weights of each directed edge are aggregated to form a baseline weighted adjacency matrix. By adopting the above technical solution, the absolute concentrations of the core biomarker and the corresponding trace metal elements are first extracted, and the thermodynamic coordination stability constants of the core biomarker and the trace metal elements under standard conditions are obtained. The absolute concentrations of the core biomarker and the trace metal elements, along with the thermodynamic coordination stability constants, are multiplied together. The result is then added to a pre-set offset constant, and the initial weights are obtained by performing a natural logarithmic operation. The algorithm formula for this calculation logic is as follows:

[0022] ;

[0023] in, Represents the initial weights. Represents the thermodynamic coordination stability constant. This represents a basic preset offset constant. Then, the absolute concentration of each anion and the absolute value of its corresponding charge number are extracted. The product of the absolute concentration of each anion and its corresponding absolute charge number is calculated. This product of all anions is summed and halved to obtain the effective ionic strength of the anion. The algorithm formula for this calculation logic is as follows: ; in, Represents the effective ionic strength of anions. This represents the halving constant used in halving operations. The absolute value of the charge number corresponding to the anion is represented by the effective ionic strength of the anion multiplied by the preset system dielectric decay constant, and the negative value is used to calculate the dynamic shielding factor by exponentiation of this value with the natural constant as the base. The algorithm formula for this calculation logic is as follows: ; in, Represents the dynamic masking factor. Representing the dielectric decay constant of the system, the initial weights of each directed edge are multiplied by the dynamic shielding factor to obtain the baseline weights of each directed edge. Finally, the baseline weights are aggregated to construct a baseline weighted adjacency matrix. This matrix aggregation algorithm simulates the ion atmosphere shielding effect of complex anion groups on the core cyanide coordination framework due to electrostatic interaction in the real liquor solution environment, and directly hard-codes the objective thermodynamic laws into the underlying computational logic of the computer graph theory algorithm. S4. Inject virtual noise variables into the absolute concentration data of trace metal elements, reconstruct the perturbation-enhanced weighted adjacency matrix, and extract the centrality drift variance by comparing the principal eigenvectors before and after the perturbation. Furthermore, in step S4, dummy noise variables are injected into the absolute concentration data of trace metal elements, including the following steps: Construct a set of white noise random variables that follow a Gaussian distribution; The absolute concentrations of each element in the trace metal element are fused and superimposed with the extracted white noise random variables to generate trace metal element perturbation concentration data. In step S4, the perturbation-reconstructed weighted adjacency matrix is ​​reconstructed, and the centrality drift variance is extracted by comparing the principal eigenvectors before and after the perturbation. This includes the following steps: While keeping the dynamic shielding factor value unchanged, the perturbation concentration data is substituted into the edge weight calculation process to generate new weights for each edge after perturbation, and the weighted adjacency matrix after perturbation is reconstructed accordingly. Eigenvalue decomposition is performed on the baseline weighted adjacency matrix and the perturbation-enhanced weighted adjacency matrix respectively. The main eigenvector before perturbation corresponding to the largest eigenvalue of the baseline weighted adjacency matrix and the main eigenvector after perturbation corresponding to the largest eigenvalue of the perturbation-enhanced weighted adjacency matrix are extracted. The corresponding feature components in the principal eigenvectors before and after perturbation of each element in the original transition metal group are extracted respectively. Calculate the difference between the eigencomponents of the same element before and after perturbation and perform a square operation. Sum the square operation results of all elements in the original transition metal group and calculate the average value. Output the average value as the centrality drift variance. By adopting the above technical solution, a set of white noise random variables following a Gaussian distribution is constructed within the algorithm system as a virtual test interference source. The absolute concentrations of each element in the trace metal are fused and superimposed with the extracted white noise random variables to generate trace metal perturbation concentration data. The algorithm, while keeping the dynamic shielding factor value unchanged, substitutes the perturbation concentration data into the edge weight calculation process to generate new weights for each perturbed edge and reconstructs the perturbed weighted adjacency matrix accordingly. The baseline weighted adjacency matrix and the perturbed weighted adjacency matrix are then compared. The matrix undergoes eigenvalue decomposition to extract the pre-perturbation principal eigenvector corresponding to the largest eigenvalue of the baseline weighted adjacency matrix and the post-perturbation principal eigenvector corresponding to the largest eigenvalue of the weighted adjacency matrix. Corresponding eigencomponents of each element in the original transition metal group are extracted from the pre-perturbation and post-perturbation principal eigenvectors. The square of the product of the differences between the eigencomponents of the same element before and after perturbation is calculated. The squares of all elements in the original transition metal group are summed, and the average is calculated. This average is output as the centrality drift variance. The algorithm formula for this calculation is as follows: ; in, Represents the variance of centrality drift. This represents the average constant used to calculate the within-group average. This represents the feature components in the principal eigenvector after perturbation. The eigencomponents in the pre-perturbation principal eigenvectors were used to perform a digital anti-interference virtual stress test on the chemical coordination network.

[0024] S5. Extract the static topological features of the benchmark weighted adjacency matrix and fuse them with the centrality drift variance to generate a higher-order topological chemical feature vector. Input the higher-order topological chemical feature vector into the principal component analysis model for dimensionality reduction. Input the dimensionality-reduced principal component data into the random forest classification model for training and establish the classification decision boundary to generate a true / false identification model.

[0025] Furthermore, in step S5, the static topological features of the benchmark weighted adjacency matrix are extracted and fused with the centrality drift variance to generate a higher-order topological chemical feature vector, including the following steps: The graph weighted density is extracted by summing and averaging all the baseline weights in the baseline weighted adjacency matrix. The values ​​of the corresponding core nodes are extracted from the pre-perturbation principal eigenvectors and used as the baseline centrality components. The graph weighted density, baseline centrality components, centrality drift variance, and baseline weights of each directed edge are numerically concatenated according to a preset dimensional order to generate a high-order topological chemical feature vector. In step S5, the high-order topological chemical feature vectors are input into the principal component analysis model for dimensionality reduction. The dimensionality-reduced principal component data is then input into the random forest classification model for training. The classification decision boundary is established to generate a true / false identification model, including the following steps: The high-order topological chemical feature vectors of samples known to be pure grain solid-state baijiu and samples known to be blended baijiu made with edible alcohol in the standardized sample set are aggregated to generate a global feature matrix. The global feature matrix is ​​centered and the covariance matrix is ​​calculated. Then, principal components that meet the preset cumulative variance contribution rate threshold are extracted by eigenvalue decomposition, and a low-dimensional principal component analysis space is constructed. The high-order topological chemical feature vectors of each sample are projected into the low-dimensional principal component analysis space, and the classification decision boundary is delineated in the space according to the clustering rules of the projected coordinates. By adopting the above technical solution, the graph weighted density used to quantify the overall network density is first extracted by summing and averaging all the benchmark weights in the benchmark weighted adjacency matrix. The algorithm formula for this calculation is as follows: , ; in, Represents weighted density, This represents the average metric constant used to calculate density. This represents the weighted weight of each directed edge, and simultaneously extracts the values ​​of the corresponding core nodes from the pre-perturbation principal feature vector as the baseline centrality component. Then, the graph weighted density baseline centrality drift variance and the baseline weights of each directed edge are reassembled according to a preset dimensional order to generate a higher-order topological chemical feature vector. The algorithm formula is as follows: ; in, Represents higher-order topological chemical eigenvectors. Representing the baseline centrality component, the high-order topological chemical feature vectors of samples known to be pure grain solid-state fermented liquor and samples known to be blended liquor made with edible alcohol in the standardized sample set are aggregated to generate a global feature matrix. The global feature matrix is ​​then centered and its covariance matrix is ​​calculated. Principal components satisfying a preset cumulative variance contribution rate threshold are extracted through eigenvalue decomposition, constructing a low-dimensional principal component analysis space to obtain principal component data for each sample. The principal component data of each sample and its corresponding true / false category labels constitute a training set. This training set is input into a constructed random forest classification model for supervised training. Multiple decision trees are generated to perform branching and ensemble voting calculations on the principal component data to generate the classification decision boundary, thus generating a true / false identification model. The algorithm formula for this random forest ensemble voting calculation is as follows: ; in, This represents the final classification output. The operator represents the variable operator that takes the maximum value, and the label represents the candidate true / false category. Represents the accumulation operator. Represents a conditional indicator function. The classification function of a single decision tree is represented by the principal component data. The principal component analysis model is used to perform orthogonal transformation on the high-dimensional topological chemical feature vectors to reduce the dimensionality. The multi-decision tree ensemble calculation of the random forest classification model is combined to process the dimensionality-reduced feature data. This overcomes the limitation of relying on manual observation of projected coordinates to divide the boundaries and solves the problem that a single linear model is difficult to fit the nonlinear boundaries of high-order topological data.

[0026] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a model for identifying the authenticity of solid-state fermented liquor, characterized in that, Includes the following steps: S1. Establish a standardized sample set and obtain the absolute concentration data of core markers, anions and trace metal elements in each sample; S2. Map the absolute concentration data into a directed star network graph, define the core marker as the core node, define the trace metal element as the peripheral node set, and establish directed edges from the core node to each node in each peripheral node set. S3. Introduce the thermodynamic coordination stability constant and the dynamic shielding factor calculated based on the anion, calculate the weight of each directed edge, and construct a benchmark weighted adjacency matrix characterizing the intensity of static chemical interaction of the sample. S4. Inject virtual noise variables into the absolute concentration data of the trace metal elements, reconstruct the perturbation-enhanced weighted adjacency matrix, and extract the centrality drift variance by comparing the principal eigenvectors before and after the perturbation. S5. Extract the static topological features of the benchmark weighted adjacency matrix and fuse them with the centrality drift variance to generate a higher-order topological chemical feature vector. Input the higher-order topological chemical feature vector into the principal component analysis model for dimensionality reduction. Input the dimensionality-reduced principal component data into the random forest classification model for training to establish the classification decision boundary and generate a true / false identification model.

2. The method for constructing a model for identifying the authenticity of solid-state fermented liquor according to claim 1, characterized in that, In step S1, the absolute concentration data of core markers, anions, and trace metal elements in each sample are obtained, including the following steps: The absolute concentration of cyanide in the sample was determined using ion chromatography as the data for the core biomarker. The absolute concentration of a specific anion group, including chloride ions, nitrite ions, bromide ions, nitrate ions, sulfate ions, and phosphate ions, in a sample is determined using an ion chromatograph, and the data for the anions are obtained. The absolute concentration of a specific element group was determined by a combination of inductively coupled plasma mass spectrometry and inductively coupled plasma atomic emission spectrometry, and this concentration was used as data for the trace metal elements. The specific element group included lead, cadmium, chromium, arsenic, nickel, tin, copper, iron, manganese, and phosphorus.

3. The method for constructing a model for identifying the authenticity of solid-state fermented liquor according to claim 1, characterized in that, In step S2, mapping the absolute concentration data into a directed star network diagram includes the following steps: Cyanide is set as a globally unique node and designated as the core node of the network graph. Based on the material source mechanism of fermentation, a specific element group is divided into a heterogeneous, non-overlapping primary transition metal group and an exogenous trace impurity group, and all elements in these two groups are regarded as the peripheral node set, wherein the primary transition metal group includes copper, iron and manganese. Construct a one-way association path starting from the core node and pointing to each specific element in the set of peripheral nodes to generate the directed edges.

4. The method for constructing a model for identifying the authenticity of solid-state fermented liquor according to claim 1, characterized in that, In step S3, calculating the weight of each directed edge includes the following steps: The absolute concentrations of the core biomarker and the corresponding trace metal elements are extracted, and the thermodynamic coordination stability constants of the core biomarker and the trace metal elements under standard conditions are obtained. Multiply the absolute concentration of the core marker, the absolute concentration of the trace metal element, and the thermodynamic coordination stability constant together. After adding the obtained product result to the basic preset offset constant, perform a natural logarithm operation to obtain the initial weight for the directed edge; traverse the set of peripheral nodes to obtain the initial weights of all directed edges.

5. The method for constructing a model for identifying the authenticity of solid-state fermented liquor according to claim 1, characterized in that, In step S3, constructing the benchmark weighted adjacency matrix characterizing the intensity of static chemical interactions in the samples includes the following steps: Extract the absolute concentration of each anion and the absolute value of the charge number corresponding to each anion, calculate the product of the absolute concentration of each anion and the square of the absolute value of the corresponding charge number, sum the products of all anions and take half of the result to obtain the effective ionic strength of the anion. The effective ionic strength of the anion is multiplied by the preset system dielectric decay constant and the negative value is obtained. The exponent of the negative value is then calculated with the natural constant as the base to generate the dynamic shielding factor. The initial weights of each directed edge are multiplied by the dynamic shielding factor to obtain the baseline weights of each directed edge; the baseline weights of each directed edge are aggregated to form the baseline weighted adjacency matrix.

6. The method for constructing a model for identifying the authenticity of solid-state fermented liquor according to claim 1, characterized in that, In step S4, injecting virtual noise variables into the absolute concentration data of the trace metal elements includes the following steps: Construct a set of white noise random variables that follow a Gaussian distribution; The absolute concentrations of each element in the trace metal element are fused and superimposed with the extracted white noise random variable to generate the trace metal element perturbation concentration data.

7. The method for constructing a model for identifying the authenticity of solid-state fermented liquor according to claim 1, characterized in that, In step S4, the perturbation-reconstructed weighted adjacency matrix is ​​reconstructed, and the centrality drift variance is extracted by comparing the principal eigenvectors before and after the perturbation. This includes the following steps: While keeping the value of the dynamic shielding factor unchanged, the perturbation concentration data is substituted into the edge weight calculation process to generate new weights for each edge after perturbation, and the weighted adjacency matrix after perturbation is reconstructed accordingly. Eigenvalue decomposition is performed on the baseline weighted adjacency matrix and the perturbation-enhanced weighted adjacency matrix respectively, and the pre-perturbation main eigenvector corresponding to the maximum eigenvalue of the baseline weighted adjacency matrix and the post-perturbation main eigenvector corresponding to the maximum eigenvalue of the perturbation-enhanced weighted adjacency matrix are extracted. The corresponding feature components in the pre-perturbation principal feature vector and the post-perturbation principal feature vector of each element in the original transition metal group are extracted respectively. The difference between the characteristic components of the same element before and after perturbation is calculated and squared. The squared results of all elements in the original transition metal group are summed and the average value is calculated. The average value is output as the centrality drift variance.

8. The method for constructing a model for identifying the authenticity of solid-state fermented liquor according to claim 1, characterized in that, In step S5, the static topological features of the benchmark weighted adjacency matrix are extracted and fused with the centrality drift variance to generate a higher-order topological chemical feature vector, including the following steps: The graph weighted density is extracted by summing and averaging all the benchmark weights in the benchmark weighted adjacency matrix. The values ​​corresponding to the core nodes are extracted from the perturbation-pre-main feature vector and used as the baseline centrality components. The graph weighted density, the baseline centrality component, the centrality drift variance, and the baseline weights of each directed edge are numerically concatenated according to a preset dimensional order to generate the higher-order topological chemical feature vector.

9. The method for constructing a model for identifying the authenticity of solid-state fermented liquor according to claim 1, characterized in that, In step S5, the high-order topological chemical feature vector is input into the principal component analysis model for dimensionality reduction, and the dimensionality-reduced principal component data is input into the random forest classification model for training to establish the classification decision boundary and generate a true / false identification model. This includes the following steps: The high-order topological chemical feature vectors of samples known to be pure grain solid-state baijiu and samples known to be blended baijiu made with edible alcohol in the standardized sample set are aggregated to generate a global feature matrix. The global feature matrix is ​​centered and the covariance matrix is ​​calculated. Then, principal components that meet the preset cumulative variance contribution rate threshold are extracted by eigenvalue decomposition. A low-dimensional principal component analysis space is constructed to obtain the principal component data of each sample. The principal component data of each sample and its corresponding true / false category labels constitute the training set; The training set is input into the constructed random forest classification model for supervised training. Multiple decision trees are generated to perform branching and ensemble voting calculations on the principal component data, generating the classification decision boundary. The high-order topological chemical feature vectors of each sample are projected into the low-dimensional principal component analysis space. Based on the clustering rules of the projected coordinates, the classification decision boundary is divided in the space.

Images