Soil degradation factor identification and improvement decision-making method by fusing temporal memory network and graph centrality analysis

Abstract

The application discloses a soil degradation factor identification and improvement decision-making method fusing a time sequence memory network and graph centrality analysis, belongs to the field of soil quality evaluation, and comprises the following steps: acquiring fertilization evolution records and time node sequences of a target region, processing the fertilization evolution records and time node sequences through a long short-term memory network to obtain an accumulated effect feature vector; acquiring a soil physicochemical feature set according to the feature vector, and obtaining a comprehensive state matrix through principal component analysis dimension reduction; determining a degradation path branch according to the comprehensive state matrix, and extracting a linkage relationship graph corresponding to the degradation path containing negative succession values; determining key degradation factors according to node centrality in the graph and removing edge factors, and processing the degradation factors and the linkage relationship graph through a random forest algorithm to obtain a driving mechanism classification label; calculating cosine similarity of the label and historical improvement texts, and outputting a corresponding difference improvement scheme according to a similarity ranking list. The application realizes intelligent diagnosis of soil degradation and matching of a precise improvement scheme.

Description

Technical Field

[0001] This invention belongs to the field of soil quality assessment, and in particular relates to a method for identifying soil degradation factors and making improvement decisions by integrating temporal memory networks and graph centrality analysis. Background Technology

[0002] Soil quality is directly related to food security and sustainable agricultural development. Accurately assessing and diagnosing soil degradation is a crucial prerequisite for achieving precision agricultural management. Currently, the evolution of fertilization systems in agricultural production presents a complex process, from traditional organic fertilizers to chemical fertilizers as the dominant method, and then to a combination of organic and inorganic fertilizers. Different fertilization patterns have significantly different impacts on various soil nutrient indicators. Existing soil quality assessment methods are typically based on field sampling and laboratory physicochemical analysis. They involve single-point measurements or simple statistical analysis of soil indicators such as pH, organic matter, and available nutrients, combined with expert experience to manually determine the level of soil degradation. Some advanced solutions have introduced remote sensing monitoring and geographic information system (GIS) technologies to achieve spatial display and trend fitting of soil indicators. Other methods use principal component analysis or factor analysis to reduce the dimensionality of multidimensional soil indicators to extract a comprehensive quality index. In addition, a few studies have attempted to use time series analysis to perform regression modeling of fertilizer application and soil nutrient changes, and to construct a soil quality database for historical data query and comparative analysis.

[0003] However, current soil quality assessment methods generally suffer from the following problems. First, existing methods often focus on measuring indicators at a single point in time or on simple time series fitting, making it difficult to capture the cumulative effect of long-term fertilization regime evolution on soil quality, resulting in a lack of temporal depth in degradation diagnosis. Second, assessment systems often rely on single indicators or simple combinations of indicators, reflecting only the local state of soil quality. The complex interactions between soil physicochemical indicators lack an effective characterization of the overall soil state, making it difficult to fully reveal the degradation essence under the coupled effects of multiple indicators. Third, the degradation pathways caused by different fertilization patterns differ significantly. For example, long-term nitrogen fertilization alone may exacerbate soil acidification, while fertilization-free treatments may lead to nutrient imbalances. Current methods categorize different degradation types broadly as soil quality decline, lacking refined identification of degradation pathway branches. Fourth, due to the linkage effects between soil indicators, changes in one key indicator can trigger a chain reaction in other indicators. Relying solely on empirical judgment makes it difficult to accurately identify the core factors driving soil degradation from complex indicator correlations, resulting in unclear degradation blocking targets. Fifth, there is a lack of effective technical connection between degradation diagnosis results and improvement strategies. Existing methods cannot automatically match targeted improvement schemes for different degradation mechanisms, resulting in a lack of differentiation in improvement measures, low repair efficiency, and high costs. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for identifying and remediating soil degradation factors by integrating temporal memory networks and graph centrality analysis, comprising: Obtain fertilization evolution records and corresponding time node sequences for the target area, and process the fertilization evolution records and time node sequences through a long short-term memory network to obtain a cumulative effect feature vector; The set of soil physicochemical features of the target area is obtained based on the cumulative effect feature vector, and the set of soil physicochemical features is processed by the principal component analysis algorithm to obtain the dimension-reduced comprehensive state matrix. The degradation path branches are determined based on the comprehensive state matrix. If the first degradation path in the degradation path branch contains negative succession values, the linkage relationship map corresponding to the first degradation path is extracted. The node centrality value is obtained from the linkage relationship graph. If the node centrality value is greater than the centrality threshold, the first critical degradation factor corresponding to the node centrality value is determined. If the node centrality value is not greater than the centrality threshold, then the edge factor corresponding to the node centrality value is removed to obtain the updated linkage relationship graph. Obtain the first key degradation factor and the updated linkage relationship graph, and process the first key degradation factor and the updated linkage relationship graph using the random forest algorithm to obtain the driving mechanism classification label; Based on the driving mechanism classification label, obtain the historical improved text, calculate the cosine similarity value between the driving mechanism classification label and the historical improved text, and obtain a similarity ranking list; Obtain the similarity ranking list. If the first value in the list is greater than the matching threshold, extract the difference improvement output corresponding to the first value.

[0005] Optionally, the step of acquiring fertilization evolution records and corresponding time node sequences of the target area, and processing the fertilization evolution records and time node sequences through a long short-term memory network to obtain a cumulative effect feature vector includes: Obtain the record table of the target area, extract the fertilizer application amount and corresponding time points from the record table, and obtain the initial fertilizer application sequence set; Based on the initial fertilization sequence set, it is determined whether the fertilization amount is greater than the fertilization threshold. If the fertilization amount is greater than the fertilization threshold, the corresponding time point is retained to obtain the target fertilization sequence set. The target fertilization sequence set is processed by the Long Short-Term Memory Network, and the evolution values ​​in the target fertilization sequence set are extracted to obtain fertilization evolution feature quantities. Using the fertilization evolution characteristic, the cumulative amount between adjacent time points is calculated, and the initial effect value corresponding to the cumulative amount is determined. Obtain the initial effect value, determine whether the initial effect value is greater than the effect threshold, and if the initial effect value is greater than the effect threshold, extract the value corresponding to the initial effect value as the target feature quantity; A vector set is constructed using the target feature quantity, and the vector set is output as the cumulative effect feature vector.

[0006] Optionally, the soil physicochemical feature set of the target area is obtained based on the cumulative effect feature vector, and the soil physicochemical feature set is processed by principal component analysis algorithm to obtain a dimension-reduced comprehensive state matrix, including: Based on the cumulative effect feature vector, the spatial coordinates of the target area are extracted, and the soil physicochemical sampling data of the target area are obtained from the spatial coordinates to obtain the initial physicochemical values. When performing standardization processing on the initial physicochemical values, standard physicochemical values ​​with dimensional differences are obtained, and the set of soil physicochemical characteristics is constructed based on the standard physicochemical values. The set of soil physicochemical characteristics is processed using a pre-established covariance calculation formula to determine the covariance matrix corresponding to the set of soil physicochemical characteristics. The covariance matrix is ​​decomposed by the principal component analysis algorithm to obtain at least one principal component and the variance contribution rate of each principal component. Determine whether the variance contribution rate is greater than the variance threshold. If the variance contribution rate is greater than the variance threshold, extract the corresponding principal component as the target principal component, and determine the dimensionality reduction based on the target principal component. The soil physicochemical feature set is projected and transformed according to the dimensionality reduction number to obtain the dimensionality-reduced comprehensive state matrix.

[0007] Optionally, degradation path branches are determined based on the comprehensive state matrix. If the first degradation path in the degradation path branch contains negative succession values, then the linkage relationship graph corresponding to the first degradation path is extracted, including: The initial state matrix is ​​obtained, and the dimensionality reduction process of the initial state matrix is ​​performed by the principal component analysis algorithm to obtain the comprehensive state matrix; Extract the state succession sequence based on the comprehensive state matrix, and determine the degenerate path branch through the state succession sequence; Obtain the first degradation path contained in the degradation path branch; if the first degradation path has the negative succession value, then extract the initial relationship graph corresponding to the first degradation path. For the initial relationship graph, the node feature vectors of the initial relationship graph are extracted using a graph convolutional network algorithm, and the linkage relationship graph is obtained based on the node feature vectors; Obtain the set of core nodes in the linkage graph, and determine the degradation blocking target point through the set of core nodes; Based on the aforementioned degradation blocking target, a target intervention strategy is generated.

[0008] Optionally, the node centrality values ​​are obtained from the linkage graph. If the node centrality value is greater than a centrality threshold, the first critical degradation factor corresponding to the node centrality value is determined, including: Obtain the relationship graph corresponding to the device's operating status, and obtain an adjacency table containing node connection relationships by parsing the relationship graph; Based on the adjacency list, the centrality of each node in the adjacency list is calculated using a random walk algorithm to obtain the node centrality value; Obtain the node centrality value; if the node centrality value is greater than the centrality threshold, then determine the first degenerate term corresponding to the node centrality value. Based on the first degenerate term, extract the weight value of the first degenerate term in the relation graph; Obtain the weight value; if the weight value is less than the weight threshold, then determine that the first degradation term is an isolated point. Based on the isolated points, the isolated points are removed to obtain the updated second degradation term.

[0009] Optionally, if the node centrality value is not greater than the centrality threshold, then the edge factors corresponding to the node centrality value are removed to obtain an updated linkage relationship graph, including: Obtain the initial linkage graph, and calculate the centrality of each target point in the linkage graph using a random walk algorithm to obtain the first set of centrality values; Based on the first set of centrality values, a pre-established threshold value is obtained. If the target centrality value in the first set of centrality values ​​is less than the threshold value, the target point corresponding to the target centrality value is determined to be a candidate to be removed. For the item to be removed, the edge item corresponding to the item to be removed in the linkage graph is obtained, and the edge item is anomaly detection is performed on the edge item by the isolated forest algorithm to obtain the abnormal edge factor value; Based on the abnormal edge factor value, the connection edge corresponding to the abnormal edge factor value is removed from the linkage graph to obtain the first updated graph; By traversing the remaining number of nodes in the first updated graph, the connected paths between the remaining nodes are calculated to obtain the second updated graph; Based on the second updated graph, extract the graph values ​​from the second updated graph, and determine whether the graph values ​​are greater than a preset value. If they are greater than the preset value, the final relationship graph is obtained.

[0010] Optionally, the first key degradation factor and the updated linkage graph are obtained, and the first key degradation factor and the updated linkage graph are processed by a random forest algorithm to obtain a driving mechanism classification label, including: Obtain the first key degradation factor and the updated linkage relationship graph, and obtain the graph feature matrix by extracting the node set and edge weights of the updated linkage relationship graph; The initial sample set is obtained by concatenating the spectral feature matrix with the first key degradation factor. The initial sample set is processed by the random forest algorithm, and the nodes are split by the decision tree inside the random forest algorithm to obtain the initial prediction value of each decision tree. Obtain the initial predicted value. If the initial predicted value is greater than the driving force threshold, then determine that the initial predicted value is a positive driving force, and obtain a set of positive driving forces. The target mechanism class is determined by applying a majority voting mechanism to the set of positive driving forces. The target mechanism class is encoded and converted to obtain the driving mechanism classification label.

[0011] Optionally, the similarity ranking list is obtained. If the first value in the list is greater than the matching threshold, the difference improvement output corresponding to the first value is extracted, including: Obtain an initial text set, process the initial text set using the bag-of-words model algorithm, and obtain an initial vector set; Based on the initial vector group, the spatial distance between each vector in the initial vector group is calculated using the cosine similarity algorithm to obtain the initial similarity. Obtain the initial similarity, sort the initial similarity in descending order of value, and determine the similarity ranking table; Extract the first value from the similarity ranking table. If the first value is greater than the matching threshold, then the first value is determined to be the target matching degree. Based on the target matching degree, compare it with the initial text set corresponding to the target matching degree, and extract the difference items in the initial text set; Obtain the difference item, perform a text replacement operation on the difference item, and obtain the difference improvement value; Obtain the difference improvement value and output the difference improvement value as the final output item.

[0012] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0013] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: The method provided by this invention effectively captures the nonlinear cumulative effect of long-term fertilization regime changes on soil quality by processing fertilization evolution records and time node sequences through a long short-term memory network, solving the problem that existing methods struggle to track the historical impact of fertilization. Based on this, principal component analysis (PCA) is used to reduce the dimensionality of the soil physicochemical feature set, achieving a comprehensive representation of multi-dimensional indicators while retaining key information, thus overcoming the limitations of traditional single-indicator assessments. Furthermore, degradation path branches are determined based on the dimensionality-reduced comprehensive state matrix, identifying degradation paths containing negative succession values, achieving refined differentiation of different degradation types. By constructing a linkage relationship graph and calculating node centrality, key degradation factors with the most significant impact on soil degradation are automatically screened, and a driving mechanism classification model is established using a random forest algorithm, clarifying degradation blocking targets. Finally, cosine similarity matching of historical improvement texts automatically extracts the differential improvement output corresponding to the current degradation mechanism, achieving fully automated connection from degradation diagnosis to precise improvement, effectively improving the targeting and remediation efficiency of soil improvement schemes. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 This embodiment provides a method for soil degradation factor identification and improvement decision-making that integrates temporal memory networks and graph centrality analysis, including: Obtain fertilization evolution records and corresponding time node sequences for the target area, and process the fertilization evolution records and time node sequences through a long short-term memory network to obtain a cumulative effect feature vector; The set of soil physicochemical features of the target area is obtained based on the cumulative effect feature vector, and the set of soil physicochemical features is processed by the principal component analysis algorithm to obtain the dimension-reduced comprehensive state matrix. The degradation path branches are determined based on the comprehensive state matrix. If the first degradation path in the degradation path branch contains negative succession values, the linkage relationship map corresponding to the first degradation path is extracted. The node centrality value is obtained from the linkage relationship graph. If the node centrality value is greater than the centrality threshold, the first critical degradation factor corresponding to the node centrality value is determined. If the node centrality value is not greater than the centrality threshold, then the edge factor corresponding to the node centrality value is removed to obtain the updated linkage relationship graph. Obtain the first key degradation factor and the updated linkage relationship graph, and process the first key degradation factor and the updated linkage relationship graph using the random forest algorithm to obtain the driving mechanism classification label; Based on the driving mechanism classification label, obtain the historical improved text, calculate the cosine similarity value between the driving mechanism classification label and the historical improved text, and obtain a similarity ranking list; Obtain the similarity ranking list. If the first value in the list is greater than the matching threshold, extract the difference improvement output corresponding to the first value.

[0019] The specific process is as follows: Figure 1 As shown, it includes: Step S101: Obtain the fertilization evolution record and corresponding time node sequence of the target area, and process the fertilization evolution record and time node sequence through the Long Short-Term Memory Network algorithm to obtain the cumulative effect feature vector.

[0020] Optionally, a record table of the target region is obtained, and the fertilization amount and corresponding time point are extracted from the record table to obtain an initial fertilization sequence set. Based on the initial fertilization sequence set, it is determined whether the fertilization amount is greater than a preset fertilization threshold. If the fertilization amount is greater than the fertilization threshold, the corresponding time point is retained to obtain a target fertilization sequence set. The target fertilization sequence set is processed using a Long Short-Term Memory (LSTM) network algorithm to extract evolution values ​​from the target fertilization sequence set, obtaining fertilization evolution feature quantities. Using the fertilization evolution feature quantities, the cumulative amount between adjacent time points is calculated to determine the initial effect value corresponding to the cumulative amount. The initial effect value is obtained, and it is determined whether the initial effect value is greater than a preset effect threshold. If the initial effect value is greater than the effect threshold, the value corresponding to the initial effect value is extracted as a target feature quantity. A vector set is constructed using the target feature quantities, and the vector set is output as a cumulative effect feature vector.

[0021] Specifically, Long Short-Term Memory (LSTM) networks capture long-short-term dependencies in time series sequences through forgetting gates, input gates, and output gates. For the t-th time step in the input sequence, the forgetting gate is calculated using the following formula: ; in, The output value of the forget gate is σ, where σ is the sigmoid activation function and W is the output value of the forget gate. f Let h be the weight matrix of the forget gate. t-1 The hidden state of the previous time step, x t Let b be the input vector at the current time step. f This is the bias term for the forget gate. The input gate calculation formula is: ; Among them, i t W is the output value of the input gate. i Let b be the weight matrix of the input gate. i This is the bias term for the input gate. The formula for calculating the candidate cell state is: ; Among them, C t For the cell state at the current time step, C t-1 This represents the cell state at the previous time step, and ⊙ denotes element-wise multiplication. The output gate calculation formula is: ; ; Among them, o t W is the output value of the output gate. o Let b be the weight matrix of the output gate. o h is the bias term for the output gate. tOutput the hidden state at the current time step. After processing by the above gating mechanism, extract the hidden state h of the last time step. T As a cumulative effect feature vector.

[0022] In one specific embodiment, the fertilization evolution records and corresponding time node sequences of a planting area over the past 60 months are automatically retrieved from an agricultural IoT database. Specifically, the 15th of each month is extracted as the time node, and the application rates of nitrogen fertilizer (120 kg / ha) and phosphate fertilizer (45 kg / ha) at that time node are simultaneously acquired to form a multidimensional time series matrix. This matrix is ​​then input into a Long Short-Term Memory (LSTM) network algorithm model for processing. This model has three hidden layers, each containing 128 neurons. The fertilization data from 60 time steps are mapped to input tensors. A forgetting gate mechanism within the model is used to set a retention weight of 0.6, automatically filtering out ineffective fertilizer residue data caused by rainfall washout. The model is then trained 1000 times using a time-term backpropagation algorithm with a learning rate of 0.01 to accurately capture the dynamic temporal relationship between fertilizer degradation and residue in the soil. After the network forward propagation calculation, the hidden layer state of the last time step of the model is extracted, and a cumulative effect feature vector with a dimension of 256 is directly output. The feature dimension with a value greater than 0.8 in the vector accurately quantifies the degree of soil phosphorus accumulation caused by five consecutive years of phosphate fertilizer application. Finally, the 256-dimensional cumulative effect feature vector is directly written into the node attribute field of the distributed graph database for storage.

[0023] Step S102: Based on the cumulative effect feature vector, obtain the set of soil physicochemical features of the target area, and use the principal component analysis algorithm to process the set of soil physicochemical features to obtain the dimension-reduced comprehensive state matrix.

[0024] Optionally, spatial coordinates of the target area are extracted based on the cumulative effect feature vector, and soil physicochemical sampling data of the target area are obtained from the spatial coordinates to obtain initial physicochemical values. When standardizing the initial physicochemical values, standardized physicochemical values ​​eliminating dimensional differences are obtained, and a soil physicochemical feature set is constructed based on the standardized physicochemical values. The soil physicochemical feature set is processed using a pre-established covariance calculation formula to determine the covariance matrix corresponding to the soil physicochemical feature set. The covariance matrix is ​​decomposed using principal component analysis to obtain at least one principal component and the variance contribution rate corresponding to each principal component. It is determined whether the variance contribution rate is greater than a preset threshold. If the variance contribution rate is greater than the preset threshold, the corresponding principal component is extracted as the target principal component, and the dimensionality reduction is determined based on the target principal component. The soil physicochemical feature set is then projected and transformed based on the dimensionality reduction to obtain the dimensionality-reduced comprehensive state matrix.

[0025] Specifically, principal component analysis projects the original high-dimensional data into a low-dimensional space through linear transformation, achieving dimensionality reduction while preserving the main information. First, the original data matrix XX is standardized using the Z-score standardization formula: ; in, μ is the original value of the j-th feature of the i-th sample. j Let σ be the mean of the j-th feature. j Let be the standard deviation of the j-th feature. The standardized data matrix has a mean of 0 and a variance of 1, eliminating dimensional differences between different physicochemical indicators. Then, the covariance matrix is ​​calculated. The formula for calculating the sample covariance matrix is: ; Where S is the covariance matrix, n is the sample size, and X c This is the centered data matrix. The off-diagonal elements S of the covariance matrix S. jk This reflects the correlation between the j-th feature and the k-th feature. Next, eigenvalue decomposition is performed on the covariance matrix: ; The eigenvalues ​​λ1, λ2, ..., λp (arranged in descending order) and their corresponding eigenvectors v1, v2, ..., vp are obtained. The formula for calculating the variance contribution rate of the k-th principal component is: ; The formula for calculating the cumulative variance contribution rate of the first m principal components is: ; The eigenvectors corresponding to the top m principal components whose cumulative variance contribution rate reaches a preset threshold (e.g., 85%) are selected to form the projection matrix W=[v1,v2,…,vm]. The final formula for calculating the dimensionality-reduced comprehensive state matrix is ​​as follows: ; Where Y is the n×m dimensional comprehensive state matrix after dimensionality reduction.

[0026] In one specific embodiment, the received cumulative effect feature vector is parsed, for example, extracting feature arrays representing the accumulation rates of available phosphorus and available potassium. Based on the spatial coordinate information embedded in the vector, it is automatically connected to the soil monitoring database to accurately locate the target planting area. A set of soil physicochemical characteristics from 10,000 grid nodes within this area is extracted. This set constitutes an initial data matrix of 10,000 rows and 10 columns, covering ten dimensions of physicochemical indicators, including pH, organic matter content, moisture content, total nitrogen, available phosphorus, available potassium, and cation exchange capacity. Subsequently, a principal component analysis algorithm is called to process this initial data matrix. Zero-mean operation is used to eliminate dimensional differences between different physicochemical indicators. Then, the covariance matrix of the standardized matrix is ​​calculated, and its eigenvalues ​​and corresponding eigenvectors are solved. The variance contribution rates of the first three principal components are calculated to be 50%, 30%, and 10%, respectively. Since the cumulative variance contribution rate of these three principal components reaches 90%, exceeding the preset information retention threshold of 85%, the eigenvectors corresponding to these three largest eigenvalues ​​are selected to construct a transformation matrix. The original ten-dimensional physicochemical feature data is projected into a new three-dimensional orthogonal feature space, thereby obtaining a dimensionality-reduced comprehensive state matrix of 10,000 rows and three columns. This matrix eliminates redundant noise in the original data and is directly used as the bottom-level feature input to the deep neural network for automated inference of soil quality evolution trends.

[0027] Step S103: Determine the degradation path branches through the dimensionality-reduced comprehensive state matrix. If the first degradation path in the degradation path branch contains negative succession values, then extract the linkage relationship map corresponding to the first degradation path.

[0028] Optionally, an initial state matrix is ​​obtained, and principal component analysis (PCA) is used to perform dimensionality reduction on the initial state matrix to obtain a comprehensive state matrix. A state succession sequence is extracted from the comprehensive state matrix, and degenerate path branches are determined through the state succession sequence. The first degenerate path contained in the degenerate path branch is obtained. If the first degenerate path has a negative succession value, the initial relationship graph corresponding to the first degenerate path is extracted. For the initial relationship graph, a graph convolutional network algorithm is used to extract the node feature vectors of the initial relationship graph, and a linkage relationship graph is obtained based on the node feature vectors. The core node set in the linkage relationship graph is obtained, and degeneracy blocking targets are determined through the core node set. Target intervention strategies are generated based on the degeneracy blocking targets.

[0029] Specifically, a density-based spatial clustering algorithm (DBSCAN) is used to cluster the sample points in the comprehensive state matrix to determine the branches of the degenerate path. The Euclidean distance between sample points p and q is calculated using the following formula: ; Where m is the dimension number after dimensionality reduction, and pi and qi are the scores of sample points p and q on the i-th principal component, respectively. In the DBSCAN algorithm, the ε-neighborhood of point p is defined as: ; Where ε is the neighborhood radius threshold. When When p is a core point, MinPts is the threshold for the minimum number of sample points in the neighborhood. Starting from the core point, clusters are formed through density reachability relationships, and each cluster corresponds to a degenerate path branch.

[0030] After identifying the branches of the degenerate path, a linkage graph is extracted from the first degenerate path, and the initial linkage graph is processed using a graph convolutional network algorithm. The core operation of the graph convolutional network is to update the node representation by aggregating the features of neighboring nodes. Let the adjacency matrix be A and the feature matrix be X. First, self-connections are added: ; Where I is the identity matrix. Calculate the degree matrix. Its diagonal elements The propagation rule of the convolutional layer is as follows: ; in, Let l be the node feature matrix of the l-th layer. , Let be the learnable weight matrix of the l-th layer, and σ be the activation function. After multi-layer graph convolution, a graph feature representation that integrates the node's own features and information from its neighboring nodes is obtained.

[0031] In one specific embodiment, after obtaining a comprehensive state matrix containing indicators such as soil pH, organic matter, and available nutrient content, principal component analysis (PCA) is used to reduce its dimensionality to three-dimensional space, yielding corresponding feature vectors. Subsequently, a density-based spatial clustering algorithm is employed to calculate the Euclidean distance between each feature vector, and a neighborhood radius and minimum sample size for core points are set, thereby dividing the spatial data into multiple independent degradation path branches. These branches are traversed, and time-series nodes on the first degradation path are extracted, calculating the state vector difference between adjacent nodes. During the comparison process, it is found that the change in organic matter from the initial node to the secondary node is negative, directly exceeding the preset negative succession threshold, thus determining that the path contains negative succession values. Upon triggering this condition, an association rule mining algorithm is immediately initiated to calculate the covariance matrix of various soil physicochemical indicators under this path. By screening strongly correlated index pairs with covariance values ​​greater than a preset threshold, a linkage graph containing multiple entity nodes and corresponding associated edges, including soil pH, organic matter content, total nitrogen content, available phosphorus content, available potassium content, cation exchange capacity, microbial biomass carbon, and soil bulk density, is extracted from the graph database. The corresponding adjacency matrix data stream is then directly transmitted to the degradation intervention decision module.

[0032] Step S104: Based on the linkage relationship graph, obtain the node centrality value in the graph. If the node centrality value is greater than the preset centrality threshold, then determine the first critical degradation factor corresponding to the node centrality value.

[0033] Optionally, a soil degradation correlation graph is obtained, and an adjacency list containing node connection relationships is obtained by parsing the graph. Based on the adjacency list, a random walk algorithm is used to calculate the centrality of each node in the adjacency list. The centrality is obtained; if the centrality is greater than a preset threshold, a first degradation factor candidate corresponding to the centrality is determined. Based on the first degradation factor candidate, the weight value of the first degradation factor candidate in the correlation graph is extracted. The weight value is obtained; if the weight value is less than a preset weight threshold, the first degradation factor candidate is determined to be an isolated factor. Based on the isolated factors, the isolated factors are removed to obtain an updated second degradation factor candidate.

[0034] Specifically, the eigenvector centrality algorithm is used to calculate the importance of each node in the graph. For a graph G, its adjacency matrix is ​​A, and the eigenvector centrality of the nodes is x. i Defined as: ; Where λ is the largest eigenvalue of the adjacency matrix A, and n is the total number of nodes. The matrix form of this equation is: ; That is, x is the eigenvector corresponding to the largest eigenvalue λ of the adjacency matrix A. The core idea of ​​eigenvector centrality is that the centrality of a node is obtained by weighted summation of the centralities of its neighboring nodes, and nodes connected to important nodes have higher centrality.

[0035] In addition, degree centrality can be used as a supplementary metric. The formula for calculating the degree centrality of node ii is: ; Where deg(i) is the degree of node i (i.e., the number of edges connected to node i), and n-1 is the maximum possible degree in the graph. Degree centrality measures the direct influence of a node within its local area.

[0036] In one specific embodiment, the graph database file of the soil linkage map is read and the corresponding adjacency matrix is ​​constructed. Then, the eigenvector centrality algorithm in the graph computing engine is called to perform multiple iterations of calculation on the matrix to obtain the centrality value of each node in the map. For example, after matrix multiplication and damping coefficient smoothing, the centrality value of the node with decreasing soil pH is calculated to be 85, and the centrality value of the node with organic matter loss is 42. Next, the preset centrality threshold of 60 is pulled from the global configuration center, and the centrality values ​​of all nodes are traversed and compared through conditional branching logic. During the comparison process, the value of 85 for the node with decreasing soil pH is identified as being strictly greater than the preset threshold of 60. After this condition is triggered, the mapping retrieval program is immediately executed, and the primary key identifier of the high centrality node is input into the degradation factor feature library for association query, thereby determining that the first key degradation factor corresponding to the node is the soil acidification factor caused by long-term nitrogen fertilizer application. Then, the first key degradation factor is processed by one-hot encoding and written into the high-dimensional feature vector space, which is directly used as the primary input weight parameter of the soil degradation trend prediction model in the next stage.

[0037] Step S105: If the node centrality value is less than or equal to the preset centrality threshold, the edge factors corresponding to the node centrality value are removed to obtain the updated linkage relationship graph.

[0038] Optionally, an initial linkage graph is obtained, and a random walk algorithm is used to calculate the centrality of each target point in the linkage graph to obtain a first set of centrality values. Based on the first set of centrality values, a pre-established threshold is obtained. If the target centrality value in the first set of centrality values ​​is less than the threshold, the target point corresponding to the target centrality value is determined as a target to be removed. For the target to be removed, the edge items corresponding to the target to be removed in the linkage graph are obtained, and anomaly detection is performed on the edge items using the isolated forest algorithm to obtain anomaly edge factor values. Based on the anomaly edge factor values, the connecting edges corresponding to the anomaly edge factor values ​​are removed from the linkage graph to obtain a first updated graph. By traversing the remaining number of nodes in the first updated graph, the connected paths between the remaining nodes are calculated to obtain a second updated graph. Based on the second updated graph, the graph values ​​in the second updated graph are extracted, and it is determined whether the graph values ​​are greater than a preset value. If they are greater than the preset value, the final relationship graph is obtained.

[0039] Specifically, the Isolation Forest algorithm is used for anomaly detection of nodes with low centrality. The Isolation Forest constructs multiple isolated trees by randomly selecting features and split values. For a sample point x in dataset D, its path length h(x) in the isolated tree T is defined as the number of edges traversed from the root node to an external node. The anomaly score for sample x is calculated using the following formula: ; Where E(h(x)) is the average path length of x in all isolated trees, n is the total number of samples, and c(n) is the normalization constant, calculated by the following formula: ; Where H(k) = ln(k) + γ is the harmonic number, and γ ≈ 0.5772 is Euler's constant. The closer the anomaly score s(x,n) is to 1, the more anomalous the sample; the closer it is to 0, the more normal the sample.

[0040] After removing edge factors, the connectivity of the updated graph is verified. The number of connected paths in the graph can be evaluated by exponentiation of the adjacency matrix. Let the updated adjacency matrix be A′. Whether there exists a path of length k between node i and node j can be determined by... Whether it is greater than 0 is used to determine this. The structural compactness of a graph can be measured by its graph density. ; Where |E′| represents the updated number of edges, and |V′| represents the updated number of nodes. A higher graph density ρ indicates a more compact graph structure and a more concentrated cluster of core factors.

[0041] In one specific embodiment, in the analysis of soil degradation factor correlation graphs, a degree centrality algorithm is used to assess the importance of each soil indicator node in the graph and calculate the node centrality value. Traversing the indicator correlation relationships in the graph database, the degree centrality value of the soil bulk density node is calculated to be 0.15. This value is then compared with a preset centrality threshold in the system configuration table, which is a baseline value calculated based on a historical soil degradation propagation model and set to 0.25. After analysis by the logic judgment module, it is identified that the centrality value of the soil bulk density node (0.15) is less than the preset threshold of 0.25, triggering the graph pruning mechanism. Then, the operation of removing edge factors corresponding to the node centrality value is automatically executed. Specifically, the filtering function of the graph calculation engine is called to define the soil bulk density node and its associated weakly correlated indicator edges as edge factors and physically delete them from the in-memory graph structure. This removal process effectively filters out noisy data with a small impact on global soil degradation. Finally, the remaining nodes and edges in memory are re-indexed to obtain an updated linkage graph. At this point, the number of nodes in the graph is reduced from the initial one hundred to sixty-five, allowing the subsequent degradation mechanism discovery algorithm to focus on the core degradation factor cluster.

[0042] Step S106: Obtain the first key degradation factor and the updated linkage relationship graph, and use the random forest algorithm to process the first key degradation factor and the updated linkage relationship graph to obtain the driving mechanism classification label.

[0043] Optionally, a first key degradation factor and an updated linkage relationship graph are obtained. A graph feature matrix is ​​obtained by extracting the node set and edge weights of the updated linkage relationship graph. The graph feature matrix and the first key degradation factor are concatenated to obtain an initial sample set. The initial sample set is processed using a random forest algorithm. Nodes are split using the decision trees within the random forest algorithm to obtain initial prediction values ​​for each decision tree. The initial prediction values ​​are obtained. If the initial prediction value is greater than a preset threshold, it is determined that the initial prediction value is a positive driving force, and a set of positive driving forces is obtained. A majority voting mechanism is applied to the set of positive driving forces to determine the target mechanism class. The target mechanism class is encoded and converted to obtain a driving mechanism classification label.

[0044] Specifically, the Random Forest algorithm achieves classification by constructing multiple decision trees and aggregating the predictions from each tree. During the node splitting process of each decision tree, the Gini index is used as a measure of impurity. For a K-class classification problem, the Gini index of node t is calculated using the following formula: ; Where, p kThe Gini coefficient represents the proportion of samples of class k in node t. A smaller Gini coefficient indicates higher purity of the node. When a node is classified according to feature X... j The threshold s splits into left child nodes t L and right child node t R At that time, the Gini coefficient after splitting is: ; Where, n L and n R , where represents the number of samples in the left and right child nodes, and n is the total number of samples in the parent node. Feature X j The decrease in the Gini coefficient due to splitting at node t is: ; The random forest algorithm selects the feature that maximizes ΔGini and the splitting threshold for node splitting.

[0045] For all TT decision trees in the random forest, the final classification result is determined by a majority voting mechanism: ; Where ht(x) is the predicted class of sample x by the t-th decision tree, and l(⋅) is the indicator function.

[0046] In one specific embodiment, principal component analysis (PCA) is used to perform dimensionality reduction on multi-year soil monitoring data from multiple fields, calculating the variance contribution rate of each soil indicator. Soil acidification rate, with a contribution rate reaching 72%, is extracted as the first key degradation factor, currently showing a pH decrease of 0.15 units per year. Subsequently, a graph attention network model is invoked to inject this factor as a node feature into the initial soil indicator correlation graph. By calculating the attention coefficients between nodes, the correlation weights between various soil indicators in the graph are dynamically adjusted. For example, the correlation weight between soil pH and available heavy metal content is automatically updated from the initial 40 to 85, thereby generating an updated correlation graph. Next, the numerical features of the first key degradation factor are concatenated with the adjacency matrix of the updated graph to construct a feature vector of the corresponding dimension, which is then input into a pre-deployed random forest algorithm model. This model contains 200 decision trees, using the Gini impurity index to randomly select three features at each node for splitting and evaluation. Each decision tree independently performs depth-first traversal and inference calculations on the input feature vector. Finally, the prediction results of 200 trees were summarized and a majority voting mechanism was used to output the long-term nitrogen fertilizer-driven acidification and degradation type with the highest number of votes as the final driving mechanism classification label.

[0047] Step S107: Based on the driving mechanism classification label, obtain the historical improved text, and obtain the similarity ranking list by calculating the cosine similarity value between the driving mechanism classification label and the historical improved text.

[0048] Optionally, at least one driving mechanism classification label and its corresponding historical improved text set are obtained. The historical improved text set is segmented using a Chinese word segmentation tool to obtain a first segmented word set. The first segmented word set is vectorized using a term frequency-inverse document frequency algorithm to obtain a first text vector value corresponding to the historical improved text set. The same vectorization is performed on the driving mechanism classification label to determine a second label vector value corresponding to the driving mechanism classification label. The cosine similarity between the first text vector value and the second label vector value is calculated to obtain the initial similarity between the historical improved text set and the driving mechanism classification label. If the initial similarity is greater than a preset threshold, the corresponding historical improved text is determined to be a candidate improved text, and a candidate text set consisting of multiple candidate improved texts is obtained. The candidate improved texts in the candidate text set are sorted in descending order according to the initial similarity to generate a similarity ranking list.

[0049] Specifically, the Term Frequency-Inverse Document Frequency (TF-IDF) algorithm is used to convert the text into a vector representation. The term frequency TF(t,d) represents the frequency of term t in document d: ; in, Let be the number of times term t appears in document d, and let the denominator be the sum of the occurrences of all terms in document d. Inverse document frequency. Indicate the inverse frequency of term t in the entire document set D: ; Where |D| represents the total number of documents in the document collection. Let be the number of documents containing term tt. The TF-IDF weight of term t in document d is: ; Through the above calculations, each document is represented as a vector, with each dimension of the vector corresponding to the TF-IDF weight of a term. Then, a cosine similarity calculation mechanism is used to determine the similarity between the classification label vector and each historical improved text vector: ; Here, A and B are the TF-IDF vectors of the two texts, and n is the vector dimension. The cosine similarity ranges from -1 to 1, with values ​​closer to 1 indicating greater semantic similarity between the two texts.

[0050] In one specific embodiment, upon receiving the driving mechanism classification label "long-term nitrogen fertilizer-driven acidification and degradation," a database retrieval mechanism is triggered. An inverted index algorithm scans and extracts 150 relevant historical improvement texts from the historical soil improvement archive. To accurately match the best improvement scheme, a term frequency-inverse document frequency algorithm is used to transform the classification label and each historical improvement text into a high-dimensional feature vector. For example, the label is mapped to a numerical vector containing 500 dimensions, where the weight of the core word "acidification" is calculated as 0.85, and the corresponding dimension weight of a certain historical improvement text is 0.72. Then, the cosine similarity calculation formula is called, quantifying the semantic distance between texts by calculating the ratio of the dot product of two vectors to their modulus product. After matrix operations, the cosine similarity value between the historical improvement text and the label is found to be 0.89. Subsequently, a quick sorting algorithm is applied to all calculated similarity values ​​for descending order, prioritizing the 45 records with similarity values ​​greater than 0.80 at the front of the memory array, thereby generating a highly relevant similarity ranking list for direct use by the subsequent differentiated improvement scheme matching module.

[0051] Step S108: Obtain the similarity ranking list. If the first value in the similarity ranking list is greater than the preset matching threshold, then extract the difference improvement output corresponding to the first value.

[0052] Optionally, an initial text set is obtained and processed using a bag-of-words model algorithm to obtain an initial vector group. Based on the initial vector group, a cosine similarity algorithm is used to calculate the spatial distance between each vector within the initial vector group to obtain an initial similarity. The initial similarities are obtained and sorted in descending order of value to determine a similarity ranking table. The first value is extracted from the similarity ranking table; if the first value is greater than a preset value, it is determined to be the target matching degree. Based on the target matching degree, the initial text set corresponding to the target matching degree is compared, and the difference items in the initial text set are extracted. The difference items are obtained, and a text replacement operation is performed on the difference items to obtain a difference improvement value. The difference improvement value is obtained and output as the final output item.

[0053] Specifically, a bag-of-words model combined with cosine similarity is used for text matching. The bag-of-words model represents text as vectors, with each dimension corresponding to a word in the vocabulary. Dimension values ​​are weighted using term frequency (TF). ; in, Let i be the value of the vector in the dimension of the i-th term. For term t iThe frequency of occurrence in document d. The similarity between the target degradation description text Q and the historical improvement case text C is calculated using cosine similarity: ; when When θ is the preset matching threshold, it is determined that there is a matching historical improvement case.

[0054] When extracting the difference, a text comparison algorithm is used to identify the differences between the target text and the matching historical text. For two text sequences T1 and T2, the difference metric is defined as: ; This involves comparing characters or words one by one, recording all discrepancies and their corresponding content. Then, a text replacement operation is performed on the differences to generate a revised output. The replacement operation can be formalized as follows: ; This means that, based on the differences between the current text and the historical text, the corresponding parts of the historical improvement plan are replaced with content that is suitable for the current scenario.

[0055] In one specific embodiment, the system receives degraded text describing the target field submitted by the end user, calls a pre-trained deep bidirectional word vector representation model to transform the text into a 768-dimensional dense semantic vector, and simultaneously extracts semantic vectors of all known improvement cases from a historical soil improvement knowledge base. The semantic similarity is measured by calculating the cosine angle between the current vector and each vector in the knowledge base. Then, a fast sorting algorithm is used to sort all the calculated similarity scores in descending order, thus obtaining a complete similarity ranking list. For example, the scores at the top of the list might be 95%, 82%, and 71%. Next, the system automatically reads the first value in the ranking list, 95%, and compares it with a pre-set matching threshold of 90% in memory. After determining that the first value is 95% strictly greater than the preset threshold of 90%, the information extraction and reconstruction module is triggered. Based on the unique identifier associated with the first value, the corresponding historical soil improvement scheme text is retrieved from the database. At this time, the named entity recognition algorithm is further run to compare the differences between the current degradation description and the historical case in terms of degradation type description. It is found that the current text is acidification degradation caused by long-term single application of nitrogen fertilizer, while the historical text is nutrient imbalance degradation caused by long-term single application of phosphate fertilizer. The system then selectively retains the lime application improvement measures and the organic fertilizer application measures in the historical scheme, and replaces the improvement strategy for phosphorus fixation in the scheme with improvement measures for aluminum toxicity mitigation. Finally, the differentiated improvement text is output and directly written into the execution log of the agricultural terminal device.

[0056] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0057] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0058] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying soil degradation factors and making improvement decisions by integrating temporal memory networks and graph centrality analysis, characterized in that, include: Obtain fertilization evolution records and corresponding time node sequences for the target area, and process the fertilization evolution records and time node sequences through a long short-term memory network to obtain a cumulative effect feature vector; The set of soil physicochemical features of the target area is obtained based on the cumulative effect feature vector, and the set of soil physicochemical features is processed by the principal component analysis algorithm to obtain the dimension-reduced comprehensive state matrix. The degradation path branches are determined based on the comprehensive state matrix. If the first degradation path in the degradation path branches contains negative succession values, the linkage relationship map corresponding to the first degradation path is extracted. The node centrality value is obtained from the linkage relationship graph. If the node centrality value is greater than the centrality threshold, the first critical degradation factor corresponding to the node centrality value is determined. If the node centrality value is not greater than the centrality threshold, then the edge factor corresponding to the node centrality value is removed to obtain the updated linkage relationship graph. Obtain the first key degradation factor and the updated linkage relationship graph, and process the first key degradation factor and the updated linkage relationship graph using the random forest algorithm to obtain the driving mechanism classification label; Based on the driving mechanism classification label, obtain the historical improved text, calculate the cosine similarity value between the driving mechanism classification label and the historical improved text, and obtain a similarity ranking list; Obtain the similarity ranking list. If the first value in the list is greater than the matching threshold, extract the difference improvement output corresponding to the first value.

2. The method according to claim 1, characterized in that, The process involves acquiring fertilization evolution records and corresponding time-node sequences for the target region, processing these records and time-node sequences using a Long Short-Term Memory (LSTM) network, and obtaining a cumulative effect feature vector, including: Obtain the record table of the target area, extract the fertilizer application amount and corresponding time points from the record table, and obtain the initial fertilizer application sequence set; Based on the initial fertilization sequence set, it is determined whether the fertilization amount is greater than the fertilization threshold. If the fertilization amount is greater than the fertilization threshold, the corresponding time point is retained to obtain the target fertilization sequence set. The target fertilization sequence set is processed by the Long Short-Term Memory Network, and the evolution values ​​in the target fertilization sequence set are extracted to obtain fertilization evolution feature quantities. Using the fertilization evolution characteristic, the cumulative amount between adjacent time points is calculated, and the initial effect value corresponding to the cumulative amount is determined. Obtain the initial effect value, determine whether the initial effect value is greater than the effect threshold, and if the initial effect value is greater than the effect threshold, extract the value corresponding to the initial effect value as the target feature quantity; A vector set is constructed using the target feature quantity, and the vector set is output as the cumulative effect feature vector.

3. The method according to claim 1, characterized in that, The soil physicochemical feature set of the target region is obtained based on the cumulative effect feature vector. This set is then processed using principal component analysis to obtain a dimension-reduced comprehensive state matrix, including: Based on the cumulative effect feature vector, the spatial coordinates of the target area are extracted, and the soil physicochemical sampling data of the target area are obtained from the spatial coordinates to obtain the initial physicochemical values. When performing standardization processing on the initial physicochemical values, standard physicochemical values ​​with dimensional differences are obtained, and the set of soil physicochemical characteristics is constructed based on the standard physicochemical values. The set of soil physicochemical characteristics is processed using a pre-established covariance calculation formula to determine the covariance matrix corresponding to the set of soil physicochemical characteristics. The covariance matrix is ​​decomposed by the principal component analysis algorithm to obtain at least one principal component and the variance contribution rate of each principal component. Determine whether the variance contribution rate is greater than the variance threshold. If the variance contribution rate is greater than the variance threshold, extract the corresponding principal component as the target principal component, and determine the dimensionality reduction based on the target principal component. The soil physicochemical feature set is projected and transformed according to the dimensionality reduction number to obtain the dimensionality-reduced comprehensive state matrix.

4. The method according to claim 1, characterized in that, Based on the comprehensive state matrix, degradation path branches are determined. If the first degradation path in the degradation path branch contains negative succession values, the linkage relationship graph corresponding to the first degradation path is extracted, including: The initial state matrix is ​​obtained, and the dimensionality reduction process of the initial state matrix is ​​performed by the principal component analysis algorithm to obtain the comprehensive state matrix; Extract the state succession sequence based on the comprehensive state matrix, and determine the degenerate path branch through the state succession sequence; Obtain the first degradation path contained in the degradation path branch; if the first degradation path has the negative succession value, then extract the initial relationship graph corresponding to the first degradation path. For the initial relationship graph, the node feature vectors of the initial relationship graph are extracted using a graph convolutional network algorithm, and the linkage relationship graph is obtained based on the node feature vectors; Obtain the set of core nodes in the linkage relationship graph, and determine the degradation blocking target point through the set of core nodes; Based on the aforementioned degradation blocking target, a target intervention strategy is generated.

5. The method according to claim 1, characterized in that, The node centrality values ​​are obtained from the linkage graph. If the node centrality value is greater than the centrality threshold, the first critical degradation factor corresponding to the node centrality value is determined, including: Obtain the relationship graph corresponding to the device's operating status, and obtain an adjacency table containing node connection relationships by parsing the relationship graph; Based on the adjacency list, the centrality of each node in the adjacency list is calculated using a random walk algorithm to obtain the node centrality value; Obtain the node centrality value; if the node centrality value is greater than the centrality threshold, then determine the first degenerate term corresponding to the node centrality value. Based on the first degenerate term, extract the weight value of the first degenerate term in the relation graph; Obtain the weight value; if the weight value is less than the weight threshold, then determine that the first degradation term is an isolated point. Based on the isolated points, the isolated points are removed to obtain the updated second degradation term.

6. The method according to claim 1, characterized in that, If the node centrality value is not greater than the centrality threshold, then the edge factor corresponding to the node centrality value is removed to obtain the updated linkage relationship graph, including: Obtain the initial linkage graph, and calculate the centrality of each target point in the linkage graph using a random walk algorithm to obtain the first set of centrality values; Based on the first set of centrality values, a pre-established threshold value is obtained. If the target centrality value in the first set of centrality values ​​is less than the threshold value, the target point corresponding to the target centrality value is determined to be a candidate to be removed. For the item to be removed, the edge item corresponding to the item to be removed in the linkage graph is obtained, and the edge item is anomaly detection is performed on the edge item by the isolated forest algorithm to obtain the abnormal edge factor value; Based on the abnormal edge factor value, the connection edge corresponding to the abnormal edge factor value is removed from the linkage graph to obtain the first updated graph; By traversing the remaining number of nodes in the first updated graph, the connected paths between the remaining nodes are calculated to obtain the second updated graph; Based on the second updated graph, extract the graph values ​​from the second updated graph, and determine whether the graph values ​​are greater than a preset value. If they are greater than the preset value, the final relationship graph is obtained.

7. The method according to claim 1, characterized in that, Obtain the first key degradation factor and the updated linkage graph, process the first key degradation factor and the updated linkage graph using the random forest algorithm to obtain the driving mechanism classification label, including: Obtain the first key degradation factor and the updated linkage relationship graph, and obtain the graph feature matrix by extracting the node set and edge weights of the updated linkage relationship graph; The initial sample set is obtained by concatenating the spectral feature matrix with the first key degradation factor. The initial sample set is processed by the random forest algorithm, and the nodes are split by the decision tree inside the random forest algorithm to obtain the initial prediction value of each decision tree. Obtain the initial predicted value. If the initial predicted value is greater than the driving force threshold, then determine that the initial predicted value is a positive driving force, and obtain a set of positive driving forces. The target mechanism class is determined by applying a majority voting mechanism to the set of positive driving forces. The target mechanism class is encoded and converted to obtain the driving mechanism classification label.

8. The method according to claim 1, characterized in that, Obtain the similarity ranking list. If the first value in the list is greater than the matching threshold, extract the difference improvement output corresponding to the first value, including: Obtain an initial text set, process the initial text set using the bag-of-words model algorithm, and obtain an initial vector set; Based on the initial vector group, the spatial distance between each vector in the initial vector group is calculated using the cosine similarity algorithm to obtain the initial similarity. Obtain the initial similarity, sort the initial similarity in descending order of value, and determine the similarity ranking table; Extract the first value from the similarity ranking table. If the first value is greater than the matching threshold, then the first value is determined to be the target matching degree. Based on the target matching degree, compare it with the initial text set corresponding to the target matching degree, and extract the difference items in the initial text set; Obtain the difference item, perform a text replacement operation on the difference item, and obtain the difference improvement value; Obtain the difference improvement value and output the difference improvement value as the final output item.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images