A soil pollutant tracing analysis method based on multivariate statistical analysis and RAG

Abstract

The application discloses a soil pollutant tracing analysis method based on multivariate statistical analysis and RAG, and relates to the technical field of environmental data analysis.The method comprises the following steps: collecting soil samples according to an industry standard and performing standardization pretreatment on pollutant concentration data; sequentially performing Spearman correlation analysis and principal component analysis (PCA) to extract pollutant correlation and contribution; constructing a similarity matrix and a contribution matrix and splicing to generate a pollutant feature matrix; based on the constructed literature vector knowledge base, using search enhancement to generate a RAG framework, combining statistical characteristics and text similarity to calculate a tracing score; generating a ternary group of 'pollutant-tracing score-literature' by the system, collating soil point sample metadata, and automatically generating an analysis result report according to a predetermined report template and providing the report to a user.

Description

Technical Field

[0001] This invention relates to the technical field of environmental data analysis, and more specifically, to a method for tracing the source of soil pollutants based on multivariate statistical analysis and RAG. Background Technology

[0002] The mainstream methods in the field of soil pollutant source tracing and analysis currently suffer from the following technical shortcomings:

[0003] Statistical modeling is disconnected from domain knowledge: Traditional multivariate statistical methods (such as principal component analysis and hierarchical clustering) can identify the correlation of pollutants, but only output numerical results. The analytical results such as the correlation of pollution source types rely on manual review of literature or experience judgment. There is a lack of an automated linkage mechanism with domain literature knowledge, resulting in a broken link of "data processing → manual decision-making".

[0004] The efficiency of multi-source data fusion is low: pollution feature analysis, spatial distribution data and literature evidence are scattered in independent systems. A matrix-based representation method for multi-dimensional features has not yet been established. Furthermore, the retrieval of literature knowledge is limited to unstructured patterns of manual search and comparison. There is a lack of encoded storage and automated matching mechanisms for pollution source fingerprint features, resulting in low efficiency of reusing historical case knowledge.

[0005] Poor interpretability of decision support: Conventional machine learning models have a "black box" problem, which cannot clearly present the logical basis for pollutant migration paths and source apportionment. Summary of the Invention

[0006] The purpose of this invention is to provide a soil pollution source tracing analysis method that integrates multivariate statistical modeling and literature knowledge enhancement. It combines pollution feature mining and domain knowledge retrieval generation technology to achieve intelligent diagnosis of pollution sources, and is not limited to manual search and comparison, but completes automatic search and matching of pollutant literature sources.

[0007] The technical solution of this invention is: to provide a soil pollutant source tracing analysis method based on multivariate statistical analysis and RAG, the method comprising:

[0008] S1. Soil sample collection and data preprocessing: Collect m soil samples from the area to be analyzed according to national or industry standards. Perform mean removal, variance standardization, outlier removal, and mean imputation of missing values ​​on the concentration data of n pollutants to form a standardized data matrix Z.

[0009] S2. Extracting multivariate statistical features: Spearman correlation analysis and principal component analysis (PCA) are performed on the data matrix Z in sequence to obtain the correlation coefficient ρ between pollutants, the eigenvalue λ corresponding to each principal component, and the principal component loading matrix V.

[0010] S3. Construct similarity and contribution matrices: Construct a similarity matrix S based on the correlation coefficient ρ between pollutants, construct a normalized contribution matrix P based on the eigenvalue λ corresponding to each principal component and the principal component loading matrix V, and concatenate the similarity matrix S and the contribution matrix P column by column to form the pollutant feature matrix X;

[0011] S4. Construct a literature knowledge base: Collect H articles related to soil pollution source tracing. After word segmentation and stop word removal, use the GTE text vectorization model to map each article into a fixed-dimensional vector u. h And build a vector index library that binds corresponding vectors and documents;

[0012] S5. Conduct source tracing analysis on pollutants: Input the pollutant feature matrix X into the retrieval enhancement generation technology RAG framework via API. Sequentially retrieve the K most relevant documents in the knowledge base for each column feature vector of the pollutant feature matrix X based on cosine similarity. Then, in the generation stage, calculate the source tracing score Score by combining statistical contribution and text similarity according to a weighted strategy, and output the results in the triple format of "pollutant - source tracing score Score - document D".

[0013] S6. Automatically generate soil pollutant analysis reports: Collect the output triplet and metadata of soil point samples, and automatically generate soil pollutant analysis reports for users according to the predefined report template.

[0014] In any of the above technical solutions, the data preprocessing performed in step S1 further includes:

[0015] Perform mean removal and variance standardization on the original data:

[0016]

[0017] Among them, z ik Let c be the standardized concentration of the i-th pollutant in the k-th sample. ik Let be the original concentration of the i-th pollutant in the k-th sample. Let σ be the sample mean of the i-th pollutant. i Here, m is the sample standard deviation, and m is the total number of samples.

[0018] Set a range threshold, remove outliers, and impute missing values ​​using mean imputation to form a standardized data matrix Z = [z ik ] n×m , where n is the number of pollutant types.

[0019] In any of the above technical solutions, Spearman correlation analysis is further used to calculate the correlation coefficient ρ between the i-th and j-th pollutants. ij :

[0020]

[0021] Where i,j=1,2,…,n,d k This represents the difference in the ranking values ​​of the two pollutant concentrations in the k-th sample.

[0022] In any of the above technical solutions, the steps of principal component analysis (PCA) to calculate the eigenvalues ​​and principal component loading matrices corresponding to each principal component further include:

[0023] Construct the covariance matrix C:

[0024]

[0025] Subscript T is the transpose symbol, C ij Let z be the covariance between the standardized concentration values ​​of pollutant i and pollutant j. jk To determine the standardized concentration of pollutant j in the k-th sample, perform eigenvalue decomposition on the covariance matrix C and solve the following equation:

[0026] Cv b =λ b v b b = 1, 2, ..., n;

[0027] Where λ1≥λ2≥…≥λ n The eigenvalues ​​represent the variances of the principal components; v b The orthogonal eigenvectors corresponding to the eigenvalues ​​are the load vectors.

[0028] Calculate the cumulative variance contribution rate CR(p) of the first p principal components:

[0029]

[0030] The variance contribution rate of the current p principal components accounts for more than 85% of the total variance. Using the loading vectors corresponding to the selected first p eigenvalues, we obtain the dimensionality-reduced eigenvector Y:

[0031]

[0032] Among them, V p =[v1,v2,…,v p [ ] is the principal component loading matrix, and the dimensionality-reduced eigenvector Y is the projection of each sample onto the principal component space.

[0033] In any of the above technical solutions, the similarity matrix constructed in step S3 is further represented as S:

[0034] S = [S ij ], S ij =ρ ij, i,j=1,2,…,n;

[0035] Among them, S ij Let be the similarity between pollutant i and pollutant j;

[0036] The constructed contribution matrix is ​​represented as P:

[0037] P = [P] lj ],

[0038] Among them, P lj λ represents the normalized contribution of the l-th potential pollution source to the total concentration load of pollutant j; where λ is the normalized contribution. l v is the eigenvalue corresponding to the l-th principal component. lj The loading of the l-th principal component on the j-th pollutant dimension is the loading matrix V. p The element in the l-th column and j-th row, λ q With v qj Similarly;

[0039] By merging the similarity matrix S and the contribution matrix P, the pollutant feature matrix X is defined:

[0040]

[0041] All elements in the j-th column are represented by vector x. j The pollutant characteristic matrix X can be simplified to: X = [x j ].

[0042] In any of the above technical solutions, further, the fixed-dimensional vector u h Represented as:

[0043] u h =f GTE (D h );

[0044] This formula represents the use of the GTE text vectorization model to transform the h-th document D h After input, through function f GTE (*) is mapped to a fixed-dimensional vector u h .

[0045] In any of the above technical solutions, the vector index library constructed in step S4 is further represented as follows: H represents the total number of documents.

[0046] In any of the above technical solutions, step S5 further includes:

[0047] The pollutant feature matrix X constructed in step S3 is input into the RAG framework via the API interface. The RAG framework, combined with the GTE vector knowledge base, performs cross-modal retrieval on the input pollutant feature matrix: during the retrieval phase, all vectors x of the pollutant feature matrix X are queried sequentially. j Retrieve the K most relevant documents from the knowledge base; specifically, use cosine similarity sim(x) j ,u h )Measure the feature vector x j and literature D h The dimension vector u of the mapping h Correlation between them:

[0048]

[0049] The top K documents, sorted in descending order of similarity, are denoted as the document set {D}. h1 D h2 ,…,D hK}, and obtain the set of dimension vectors {u} mapped to each document in the collection. h1 ,u h2 ,…,u hK The collection of texts and the set of dimension vectors of the mapping are packaged together as a candidate knowledge set;

[0050] During the generation phase, the pollutant feature matrix X and the candidate knowledge set are read, and a source tracing report is generated according to the weighted fusion strategy. First, the source tracing score is calculated:

[0051] Score s,j =αP lsj +(1-α)sim(x j ,u hs ), α∈[0,1];

[0052] Score s,j The Score represents the source tracing score obtained by comprehensively considering the statistical contribution and text similarity between the j-th pollutant and the s-th retrieved document. s,j The score measures the degree of support and credibility of the literature for the potential source of the pollutant. The higher the score, the more likely the literature is to be evidence of the true source of the pollutant.

[0053] Wherein, the subscript s represents the s-th document taken after sorting by similarity from high to low, 1≤s≤K; α is the weighting coefficient that balances the statistical contribution and text similarity information. l in the subscript s This represents the principal component number corresponding to the s-th document. The lth element in the contribution matrix P s The element in row j represents the element in column l. sThe normalized contribution of each principal component to the j-th pollutant; u hs This is the sth document D hs The dimension vector of the mapping; cosine similarity sim(x) j ,u hs )Measure the feature vector x j and literature D hs The dimension vector u of the mapping hs The degree of correlation between them;

[0054] The system provides candidate literature for each pollutant j. Press Score s,j Sort in descending order, and by "pollutant j - source traceability score". s,j -Document D hs The output result is in the form of a triplet.

[0055] The beneficial effects of this invention are:

[0056] The technical solution in this invention utilizes Spearman correlation analysis and PCA to extract the correlation and contribution between pollutants, thereby achieving quantitative characterization of pollutant sources. This is significantly superior to traditional source tracing methods that rely solely on experience, reducing human error and improving the reliability of the results.

[0057] By combining statistical feature matrices with a document vector knowledge base, both contribution and text similarity are considered during the generation stage, and a "pollutant-score-document" triple is output. All conclusions can be traced back to specific documents and statistical parameters, meeting users' needs for verifiable results. Attached Figure Description

[0058] The advantages of the above and additional aspects of the present invention will become apparent and readily understood in the description of the embodiments in conjunction with the following drawings, wherein:

[0059] Figure 1 This is a schematic flowchart of a soil pollutant source tracing analysis method based on multivariate statistical analysis and RAG according to an embodiment of the present invention. Detailed Implementation

[0060] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0061] In the following description, many specific details are set forth in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0062] like Figure 1 As shown in this embodiment, a soil pollutant source tracing analysis method based on multivariate statistical analysis and RAG is provided. The method includes:

[0063] S1. Soil Sample Collection and Data Preprocessing: Within the area to be analyzed, collect m soil samples according to national or industry standards. Detect the concentration of n pollutants in each sample. Denote the original concentration of the i-th pollutant in the k-th sample as c. ik The original concentration unit is mg / kg, i = 1, 2, ..., n, k = 1, 2, ..., m.

[0064] Perform mean removal and variance standardization on the original data:

[0065]

[0066] Among them, z ik Let be the standardized concentration of the i-th pollutant in the k-th sample. Let σ be the sample mean of the i-th pollutant. i denoted as the sample standard deviation, and m as the total number of samples.

[0067] The above-mentioned mean removal refers to subtracting the average value of each variable (in the above formula, the concentration of each pollutant) from all the original values ​​of each variable, so that the mean of the processed data in the dimension of that variable is zero. This processing helps to eliminate the influence of different pollutant concentration magnitudes or overall shifts, allowing subsequent analysis to focus more on the relative changes and structural characteristics between variables.

[0068] Set a range threshold, remove outliers, and impute missing values ​​using mean imputation to form a standardized data matrix Z = [z ik ] n×m , where n is the number of pollutant types.

[0069] S2. Extracting Multivariate Statistical Features: The preprocessed pollutant concentration data is imported into the statistical analysis module, and the following steps are performed in sequence: Correlation analysis, calculating the Spearman correlation coefficient matrix to identify potential associations between pollutants; Principal component analysis (PCA), extracting the main influencing factors and calculating the principal component loading coefficients to achieve dimensionality reduction while retaining key information; The results of the above analyses form the feature vector and distribution pattern of pollutants, which is helpful for subsequent source tracing and analysis.

[0070] Correlation analysis was performed to calculate the correlation coefficient ρ between the i-th and j-th pollutants. ij :

[0071]

[0072] Where i,j=1,2,…,n,dk This represents the difference in the ranking values ​​of the two pollutant concentrations in the k-th sample.

[0073] Principal component analysis (PCA) is used to construct the covariance matrix C.

[0074]

[0075] Subscript T is the transpose symbol, C ij Let z be the covariance between the standardized concentration values ​​of pollutant i and pollutant j. jk To determine the standardized concentration of pollutant j in the k-th sample, perform eigenvalue decomposition on the covariance matrix C and solve the following equation:

[0076] Cv b =λ b v b b = 1, 2, ..., n;

[0077] Where λ1≥λ2≥…≥λ n The eigenvalues ​​represent the variances of the principal components; v b The standard orthogonal eigenvectors corresponding to the eigenvalues ​​are the load vectors; the subscript T is always a transpose symbol.

[0078] Calculate the cumulative variance contribution rate CR(p) of the first p principal components:

[0079]

[0080] The variance contribution rate of the current p principal components accounts for more than 85% of the total variance. Using the loading vectors corresponding to the selected first p eigenvalues, we obtain the dimensionality-reduced eigenvector Y:

[0081]

[0082] Among them, V p =[v1,v2,…,v p [ ] is the principal component loading matrix, and the dimensionality-reduced eigenvector Y is the projection of each sample onto the principal component space.

[0083] S3. Construct a data matrix, using the ρ obtained in step S2. ij Principal component loadings v li (Load of the l-th principal component on pollutant i)(v1,v2,…,v p To form a matrix V p The vector, v li For matrix V p By integrating the elements in the cluster and the clustering label information, a similarity matrix S between different pollutants and a contribution matrix P of each pollution source to each pollutant are constructed:

[0084] S = [S ij , S ij = ρ ij , i, j = 1, 2, …, n;

[0085] P = [P lj ,

[0086] The similarity matrix reflects the correlation strength between pollutants, and the contribution matrix quantitatively describes the relative contribution of each potential pollution source to the total pollutant load. S ij is the similarity between the i-th pollutant and the j-th pollutant; P lj represents the normalized contribution of the l-th potential pollution source (principal component) to the total concentration load of pollutant j; where λ l is the eigenvalue corresponding to the l-th principal component, and v lj is the load of the l-th principal component in the dimension of the j-th pollutant, that is, the element in the l-th column and j-th row of the load matrix V p , and λ q and v qj are the same by analogy.

[0087] Stitch and merge the above similarity matrix S and contribution matrix P, and define the pollutant feature matrix X:

[0088]

[0089] Use the vector x j to represent all elements in the j-th column, and the representation of the pollutant feature matrix X is simplified to: X = [x j .

[0090] The elements in each column of the pollutant feature matrix X are all the characteristics of the same pollutant. Each column combines the correlation characteristics and source contribution characteristics of the pollutant, and serves as the standardized input for RAG retrieval and reasoning.

[0091] S4. Build a literature knowledge base, collect literature corpus related to soil pollution source tracing, and record the h-th literature text as D h . After word segmentation, removing stop words (such as "de", "le", "zai", "he", etc.), and denoising, use the GTE text vectorization model ("GTE text vector - Chinese - general domain - large" model provided by the ModelScope platform) to map each literature into a fixed - dimension vector u h :

[0092] u h = f GTE (D h );

[0093] The above formula means using the GTE text vectorization model to map the h-th literature D hAfter input, through function f GTE (*) is mapped to a fixed-dimensional vector u h .

[0094] Using a fixed-dimensional vector u h Building a vector index library This system enables structured storage of massive document vectors, where H represents the total number of documents. This index can efficiently support subsequent fast retrieval based on vector similarity.

[0095] S5. Source tracing analysis is performed using Retrieval Enhanced Generation (RAG) technology. The pollutant feature matrix X constructed in step S3 is input into the RAG framework via an API interface. The RAG framework, combined with the GTE vector knowledge base, performs cross-modal retrieval on the input pollutant feature matrix: during the retrieval phase, all vectors x of the pollutant feature matrix X are queried sequentially. j Retrieve the K most relevant documents from the knowledge base; specifically, use cosine similarity sim(x) j ,u h )Measure the feature vector x j and literature D h The dimension vector u of the mapping h Correlation between them:

[0096]

[0097] The top K documents, sorted in descending order of similarity, are denoted as the document set {D}. h1 D h2 ,…,D hK}, and obtain the set of dimension vectors {u} mapped to each document in the collection. h1 ,u h2 ,…,u hK The set of text collections and the set of dimension vectors of the mapping are packaged together as a candidate knowledge set.

[0098] During the generation phase, a source tracing report is generated according to a weighted fusion strategy. The source tracing score is calculated first.

[0099] Score s,j =αP lsj +(1-α)sim(x j ,u hs ), α∈[0,1];

[0100] Score s,j The source tracing score is obtained by comprehensively considering the statistical contribution and text similarity between the j-th pollutant and the s-th retrieved document. It is used to measure the degree of support and credibility of the document for the potential source of the pollutant. The higher the score, the more likely the document is to be true evidence of the pollutant's origin.

[0101] Where, the subscript s represents the sth document taken after sorting by similarity from high to low, s is a positive integer, 1≤s≤K; α is the weighting coefficient, used to balance the two parts of information: statistical contribution and text similarity; l in the subscript s This represents the principal component number corresponding to the s-th document. The lth element in the contribution matrix P s The element in row j represents the element in column l. s The normalized contribution of each principal component to the j-th pollutant; u hs This is the sth document D hs The dimension vector of the mapping; cosine similarity sim(x) j ,u hs )Measure the feature vector x j and literature D hs The dimension vector u of the mapping hs The correlation between them.

[0102] The system provides candidate literature for each pollutant j. Press Score s,j Sort in descending order, and by "pollutant j - source traceability score". s,j -Document D hs The output results are presented in a triplet format to ensure that each source conclusion can be traced back to the corresponding statistical parameters and literature.

[0103] S6. The system collects and organizes the "pollutant j-source traceability score" output from step S5. s,j -Document D hs "The data of the triplet and soil point samples are automatically used to generate a soil pollutant analysis report for the user according to the pre-defined report template."

[0104] In summary, this invention proposes a soil pollutant source tracing analysis method based on multivariate statistical analysis and RAG, comprising:

[0105] S1. Soil sample collection and data preprocessing: Collect m soil samples from the area to be analyzed according to national or industry standards. Perform mean removal, variance standardization, outlier removal, and mean imputation of missing values ​​on the concentration data of n pollutants to form a standardized data matrix Z.

[0106] S2. Extracting multivariate statistical features: Spearman correlation analysis and principal component analysis (PCA) are performed on the data matrix Z in sequence to obtain the correlation coefficients ρ between pollutants, the eigenvalues ​​λ corresponding to each principal component, and the principal component loading matrix V.

[0107] S3. Construct similarity and contribution matrices: Construct a similarity matrix S based on the correlation coefficient ρ between pollutants, construct a normalized contribution matrix P based on the eigenvalue λ corresponding to each principal component and the principal component loading matrix V, and concatenate the similarity matrix S and the contribution matrix P column by column to form the pollutant feature matrix X.

[0108] S4. Construct a literature knowledge base: Collect H articles related to soil pollution source tracing. After word segmentation and stop word removal, use the GTE text vectorization model to map each article into a fixed-dimensional vector u. h And build a vector index library that binds the corresponding vectors and documents.

[0109] S5. Conduct source tracing analysis on pollutants: Input the pollutant feature matrix X into the retrieval enhancement generation technology RAG framework via API. Sequentially retrieve the K most relevant documents in the knowledge base for each column feature vector of the pollutant feature matrix X based on cosine similarity. Then, in the generation stage, calculate the source tracing score by combining statistical contribution and text similarity according to a weighted strategy, and output the results in the triple format of "pollutant - source tracing score - document D".

[0110] S6. Automatically generate soil pollutant analysis reports: Collect the output triplet and metadata of soil point samples, and automatically generate soil pollutant analysis reports for users according to the predefined report template.

[0111] The steps in this invention can be adjusted, combined, or deleted according to actual needs.

[0112] The units in the device of the present invention can be merged, divided, or reduced according to actual needs.

[0113] Although the invention has been disclosed in detail with reference to the accompanying drawings, it should be understood that these descriptions are merely exemplary and not intended to limit the application of the invention. The scope of protection of the invention is defined by the appended claims and may include various modifications, alterations, and equivalents made to the invention without departing from the scope and spirit of the invention.

Claims

1. A method for tracing the source of soil pollutants based on multivariate statistical analysis and RAG, characterized in that, The method includes: S1. Collect m soil samples from the area to be analyzed, preprocess the concentration data of n pollutants, and form a standardized data matrix Z. S2. Spearman correlation analysis and principal component analysis (PCA) are performed sequentially on the data matrix Z to obtain the correlation coefficient ρ between pollutants, the eigenvalue λ corresponding to each principal component, and the principal component loading matrix V. S3. Construct a similarity matrix S based on the correlation coefficient ρ between pollutants, construct a normalized contribution matrix P based on the eigenvalue λ corresponding to each principal component and the principal component loading matrix V, and concatenate the similarity matrix S and the contribution matrix P column by column to form the pollutant feature matrix X. S4. Collect H articles related to soil pollution source tracing. After word segmentation and stop word removal, use the GTE text vectorization model to map each article into a fixed-dimensional vector u. h And build a vector index library that binds corresponding vectors and documents; S5. Input the pollutant feature matrix X into the retrieval enhancement generation technology RAG framework, and sequentially retrieve the most relevant literature in the knowledge base for each column feature vector of the pollutant feature matrix X according to the cosine similarity. Generate the source tracing score Score calculated by combining statistical contribution and text similarity according to the weighted strategy, and output the results in the triple format of "pollutant-source tracing score-literature D". S6. Collect the output triplet and metadata of soil point samples, and automatically generate a soil pollutant analysis report according to the predetermined report template.

2. The soil pollutant source tracing analysis method based on multivariate statistical analysis and RAG as described in claim 1, characterized in that, The Spearman correlation analysis calculates the correlation coefficient ρ between the i-th and j-th pollutants. ij : Where i,j=1,2,…,n,m is the total number of samples, and d k This represents the difference in the ranking values ​​of the two pollutant concentrations in the k-th sample.

3. The soil pollutant source tracing analysis method based on multivariate statistical analysis and RAG as described in claim 2, characterized in that, The steps of calculating the eigenvalues ​​and principal component loading matrices for each principal component using principal component analysis (PCA) include: Construct the covariance matrix C: Subscript T is the transpose symbol, C ij Let z be the covariance between the standardized concentration values ​​of pollutant i and pollutant j. ik Let z be the standardized concentration of the i-th pollutant in the k-th sample. jk To determine the standardized concentration of pollutant j in the k-th sample, perform eigenvalue decomposition on the covariance matrix C and solve the following equation: Cv b Zλ b v b ,b61,2,...,n6 Where λ1≥λ2≥…≥λ n The eigenvalues ​​represent the variances of the principal components; v b The orthogonal eigenvectors corresponding to the eigenvalues ​​are the load vectors. Assuming the cumulative variance contribution of the first p principal components accounts for more than 85% of the total variance, the principal component loading matrix V is obtained. p =[v1,v2,…,v p ].

4. The soil pollutant source tracing analysis method based on multivariate statistical analysis and RAG as described in claim 3, characterized in that, The similarity matrix constructed in step S3 is represented as S: S=[S ij ],S ij =ρ ij ,i,j=1,2,…,n; Among them, S ij Let be the similarity between pollutant i and pollutant j; The constructed contribution matrix is ​​represented as P: Among them, P lj λ represents the normalized contribution of the l-th potential pollution source to the total concentration load of pollutant j; where λ is the normalized contribution. l v is the eigenvalue corresponding to the l-th principal component. lj The loading of the l-th principal component on the j-th pollutant dimension is the loading matrix V. p The element in the l-th column and j-th row, λ q With v qj Similarly; By merging the similarity matrix S and the contribution matrix P, the pollutant feature matrix X is defined: All elements in the j-th column are represented by vector x. j The pollutant characteristic matrix X can be simplified to: X = [x j ].

5. The soil pollutant source tracing analysis method based on multivariate statistical analysis and RAG as described in claim 1, characterized in that, The fixed-dimensional vector u h Represented as: u h =f GTE (D h ); This formula represents the use of the GTE text vectorization model to transform the h-th document D h After input, through function f GTE (*) is mapped to a fixed-dimensional vector u h .

6. The soil pollutant source tracing analysis method based on multivariate statistical analysis and RAG as described in any one of claims 4 or 5, characterized in that, Step S5 includes: The pollutant feature matrix X constructed in step S3 is input into the RAG framework. The RAG framework, combined with the GTE vector knowledge base, sequentially queries all vectors x of the pollutant feature matrix X. j Retrieve from the knowledge base using cosine similarity sim(x) j ,u h )Measure the feature vector x j and literature D h The dimension vector u of the mapping h The relevance between the documents is determined by sorting them in descending order of cosine similarity, and the top K documents are denoted as the document set {D}. h1 D h2 ,…,D hK }, and obtain the set of dimension vectors {u} mapped to each document in the collection. h1 ,u h2 ,…,u hK The collection of texts and the set of dimension vectors of the mapping are packaged together as a candidate knowledge set; Calculate the source tracing score: Score s,j =αP lsj +(1-α)sim(x j ,u hs ),α∈[0,1]; Score s,j The Score represents the source tracing score obtained by comprehensively considering the statistical contribution and text similarity between the j-th pollutant and the s-th retrieved document. s,j To assess the degree of support and credibility of this literature for the potential source of pollution of this pollutant; Wherein, the subscript s represents the s-th document taken after sorting by similarity from high to low, 1≤s≤K; α is the weighting coefficient that balances the statistical contribution and text similarity information. l in the subscript s This represents the principal component number corresponding to the s-th document. The lth element in the contribution matrix P s The element in row j represents the element in column l. s The normalized contribution of each principal component to the j-th pollutant; u hs This is the sth document D hs The dimension vector of the mapping; cosine similarity sim(x) j ,u hs )Measure the feature vector x j and literature D hs The dimension vector u of the mapping hs The correlation between them; The system provides candidate literature for each pollutant j. Press Score s,j Sort in descending order, and by "pollutant j - source traceability score". s,j -Document D hs The output result is in the form of a triplet.

Images