A method for analyzing a standard correlation path of marine investigation based on a graph algorithm

Abstract

The application provides a marine survey standard associated path analysis method based on a graph algorithm, and belongs to the technical field of marine survey standard associated path technology.The application constructs a restricted random walk path based on a core hub node set, counts visit frequencies, extracts a high-frequency path skeleton, applies a bidirectional breadth-first search algorithm to find a shortest associated path, inputs the shortest associated path into a path semantic consistency model to detect conflicts and screen an effective associated path set, clusters and groups the effective associated path set to form a semantic cluster, and generates an associated path analysis report containing a core hub node set, a high-frequency path skeleton, semantic cluster characteristics and potential conflict path early warning information, thereby solving the technical problem of low associated path analysis efficiency caused by insufficient accuracy of core hub node identification in a marine survey standard knowledge graph.

Description

Technical Field

[0001] This invention belongs to the technical field of marine survey standard association path technology, and specifically relates to a method for marine survey standard association path analysis based on graph algorithms. Background Technology

[0002] The construction of a marine survey standards system relies on the accurate identification of reference relationships and semantic dependencies between standards. Traditional methods establish standard associations using text similarity matching and keyword retrieval, constructing a directed graph structure to represent the reference network between standards, and extracting association paths using graph traversal algorithms. In the application of existing technologies, standards system analysts need to manually sort out the reference relationships between standard documents, rely on expert experience to identify core standard nodes, and then use depth-first search or breadth-first search algorithms to find association paths between standards. However, when processing large-scale standards knowledge graphs, traditional methods rely solely on local topological features such as node degree to identify core nodes, ignoring the betweenness centrality of nodes in the entire graph and the importance of neighboring nodes. This results in low accuracy in identifying core hub nodes, leading to an excessively large search range for association paths or the omission of key paths, resulting in low efficiency in path analysis. In other words, existing technologies suffer from the technical problem of insufficient accuracy in identifying core hub nodes in marine survey standards knowledge graphs, leading to low efficiency in association path analysis. Summary of the Invention

[0003] In view of this, the present invention provides a method for analyzing the association path of marine survey standards based on graph algorithms, which can solve the technical problem that the accuracy of core hub node identification in the knowledge graph of marine survey standards is insufficient, resulting in low efficiency of association path analysis.

[0004] This invention is implemented as follows: It provides a method for analyzing the association paths of marine survey standards based on graph algorithms. This includes constructing a knowledge graph of marine survey standards, transforming standard texts into node-edge-attribute triple structures; calculating the initial topological potential energy of each standard node in the knowledge graph, obtained by weighted summation of node degree normalization, betweenness centrality normalization, and average importance normalization of neighboring nodes; performing iterative propagation of topological potential energy, where standard nodes receive potential energy contributions from neighboring nodes and update their own potential energy values; stopping iteration when the rate of change of the overall graph potential energy is less than a set threshold for two consecutive iterations; calculating the potential energy distribution concentration coefficient and determining whether it is greater than a concentration threshold; if it is greater, extracting the standard nodes with the highest potential energy ranking as the core hub node set; otherwise, calling the standard propagation influence assessment module. The block recalculates node importance and updates the core hub node set; based on the core hub node set, it constructs a restricted random walk path, performs Monte Carlo simulation starting from the query starting standard node, and counts the access frequency of each node; it extracts nodes with an access frequency greater than a set number to form a high-frequency path skeleton, and applies a bidirectional breadth-first search algorithm to the node pairs in the high-frequency path skeleton to find the shortest associated path; it inputs the shortest associated path into the path semantic consistency model for conflict detection, and selects the shortest associated path with a logical consistency score higher than a set threshold as the effective associated path set; it clusters the effective associated path set, uses a distance metric function based on path length and edge type similarity, and uses a hierarchical clustering algorithm to divide the shortest associated path into semantic clusters; and generates an associated path analysis report.

[0005] The normalized value of node degree is obtained by dividing the actual degree of the standard node by the maximum degree value in the knowledge graph. The normalized value of betweenness centrality is obtained by calculating the proportion of the number of shortest paths passing through the standard node to the total number of shortest paths in the entire graph. The normalized value of the average importance of neighboring nodes is obtained by dividing the average degree of all first-order neighboring nodes of the standard node by the average degree of the entire graph.

[0006] In the iterative propagation of topological potential energy, the potential energy contribution received by the standard node is equal to the sum of the potential energy values ​​of all neighboring nodes in the previous round multiplied by the edge weights, multiplied by the attenuation coefficient, and finally added to the initial topological potential energy value of the standard node multiplied by the retention coefficient to obtain the new potential energy value for the current round.

[0007] The edge weight is determined by weighting the basic weight of the edge type and the frequency of historical calls. The basic weight of the reference relationship edge is 0.8, the basic weight of the semantic dependency edge is 0.6, and the edge weight increases by 0.05 for every 100 increases in the frequency of historical calls.

[0008] The concentration coefficient of potential energy distribution is calculated by dividing the sum of the potential energy of the top 20% of standard nodes by the total potential energy of the entire graph, subtracting 0.2, and then dividing by 0.8. The concentration threshold is 0.6.

[0009] The standard dissemination influence assessment module simulates standard knowledge diffusion through restricted random walks. Starting from the standard node to be evaluated, the next hop node is selected at each step according to the probability distribution after the outgoing edge weights are normalized. The walk step size follows a truncated geometric distribution. After performing a set number of walks, the cumulative number of visits to each standard node is counted, and the normalized value of the number of visits is used as the dissemination influence score.

[0010] Among them, the restricted random walk path sets the upper limit of the walk step size to 8 hops, the restart probability to 0.15, runs 10,000 walk trajectories, and extracts nodes with a visit frequency greater than 500 times to form a high-frequency path skeleton.

[0011] Among them, the bidirectional breadth-first search algorithm initiates breadth-first traversal from both the starting node and the target node. When the visited node sets in both directions intersect, the two paths are merged to form a complete shortest association path. The search depth is limited to 4 layers. The algorithm records the type label of each edge on the shortest association path and the difference between the standard publication year of the first and last nodes and the number of standard level conversions.

[0012] The path semantic consistency model's input layer receives the node sequence embedding, edge type sequence embedding, and attribute change vector of the path. The encoding layer uses a four-layer graph attention convolution module. The fusion layer concatenates the node representations output by the graph attention convolution module through global average pooling and max pooling, and then reduces the dimensionality through two layers of fully connected networks. The output layer is a single neuron sigmoid activation that outputs a logical consistency score.

[0013] The number of attention heads in the graph attention convolution module is dynamically adjusted based on three parameters: the normalized value of the average path length, the normalized value of the number of unique edge types in the path, and the normalized value of the standard year span involved in the path. The adjustment formula is 8 multiplied by the normalized value of the average path length to the power of 0.4, then multiplied by the normalized value of the number of unique edge types in the path to the power of 0.3, then multiplied by the normalized value of the standard year span involved in the path to the power of 0.3, and the result is rounded up and does not exceed 16.

[0014] The training dataset for the path semantic consistency model extracts verified and valid associated paths from historical oceanographic survey standard revision records as positive samples, and constructs conflicting paths containing circular dependencies or mutual exclusion requirements as negative samples through a forced path generation algorithm. The dataset is divided into training set, validation set and test set in a ratio of 7:2:1.

[0015] The path semantic consistency model training uses a binary cross-entropy loss function, the optimizer uses an adaptive moment estimation algorithm, the initial learning rate is set to 0.001, the learning rate decays to 0.9 times the original value after every 20 iterations, the batch size is set to 32, and the early stopping mechanism is triggered to terminate training when there is no performance improvement after 15 consecutive rounds of validation.

[0016] The distance metric function is defined as the absolute value of the path length difference divided by the longer path length plus the Jaccard distance of the edge type sequence, with each of the two weights accounting for 50%. When calculating the Jaccard distance, the number of elements in the intersection of the edge type sets of the two shortest associated paths is divided by the number of elements in the union, and then the ratio is subtracted from 1.

[0017] The hierarchical clustering algorithm uses the average linking method, with the clustering tree cutting threshold set to 0.4. The final number of clusters is controlled between 3 and 7. If the automatic division result is less than 3 semantic clusters, the cutting threshold is lowered to 0.35 and re-clustering is performed. If there are more than 7 semantic clusters, the cutting threshold is increased to 0.5 and re-clustering is performed.

[0018] Among them, the potential conflict path warning information is achieved by detecting abnormal paths within semantic clusters. The average intra-cluster distance of paths is calculated for each semantic cluster. The shortest associated path with a distance greater than twice the average intra-cluster distance is marked as an abnormal path. At the same time, highly similar path pairs across clusters are detected. If the distance between the shortest associated paths of two semantic clusters belonging to different semantic clusters is less than 0.25, they are marked as semantically ambiguous paths.

[0019] The structured knowledge graph visualization file adopts a graphical exchange format. Nodes are mapped to color depth according to their topological potential values. Standard nodes in the core hub node set are marked with star-shaped labels. Edges are marked with different line types according to their types. Semantic clusters are identified with different background color areas. Access paths in the high-frequency path skeleton are displayed in bold red, and abnormal paths are marked with orange dashed boxes.

[0020] This invention proposes a graph-based algorithm for analyzing association paths in marine survey standards. It constructs a knowledge graph of marine survey standards and calculates initial values ​​of node topological potential. The method then uses a weighted sum of normalized node degree, normalized betweenness centrality, and normalized average importance of neighboring nodes to comprehensively evaluate the importance of nodes within the overall graph topology. The invention employs an iterative propagation mechanism of topological potential, allowing potential values ​​to diffuse and converge within the graph structure. By calculating the concentration coefficient of potential distribution, the method determines the concentration of core node distributions. When the concentration coefficient exceeds a threshold, high-potential-energy nodes are directly extracted as the core hub node set; otherwise, the method calls the standard propagation influence assessment module to supplement and identify important nodes missed by local topological features. Based on the core hub node set, the method constructs restricted random walk paths and counts visit frequencies. After extracting the high-frequency path skeleton, a bidirectional breadth-first search is applied to find the shortest association path, significantly reducing the search space and improving path extraction efficiency. In summary, this invention solves the technical problem of low efficiency in association path analysis due to insufficient accuracy in identifying core hub nodes in the marine survey standard knowledge graph. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a diagram illustrating the convergence process of the iterative propagation of topological potential energy.

[0023] Figure 3 This is a schematic diagram of a bidirectional search for the shortest path association.

[0024] Figure 4 A tree diagram for effectively associating paths.

[0025] Figure 5 This is a visualization layout diagram for a standard knowledge graph. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0027] like Figure 1 The diagram shown is a flowchart of a method for analyzing the correlation path of marine survey standards based on graph algorithms, provided by this invention. This method includes the following steps:

[0028] S01. Construct a knowledge graph of marine survey standards, transforming standard texts into a node-edge-attribute triple structure. Nodes represent standard clauses and terminology entities, edges represent citation relationships and semantic dependencies, and attributes record the standard number and publication time.

[0029] S02. Calculate the initial value of the topological potential energy of each standard node in the knowledge graph. The initial value of the topological potential energy is obtained by weighted summation of the node degree normalization value, the betweenness centrality normalization value, and the average importance normalization value of the neighboring nodes, with weight coefficients of 0.4, 0.35, and 0.25, respectively.

[0030] S03. Perform topological potential energy iterative propagation. In each iteration, the standard node receives potential energy contributions from neighboring nodes and updates its own potential energy value. When the rate of change of the total potential energy in two consecutive iterations is less than 0.01, stop the iteration, record the converged potential energy distribution, and calculate the potential energy distribution concentration coefficient.

[0031] S04. Determine whether the potential energy distribution concentration coefficient is greater than the concentration threshold, where the concentration threshold is 0.6. If the potential energy distribution concentration coefficient is greater than the concentration threshold, extract the top 15% of standard nodes by potential energy value as the core hub node set and execute step S05. If the potential energy distribution concentration coefficient is not greater than the concentration threshold, call the standard propagation influence assessment module to recalculate the node importance and update the core hub node set before executing step S05.

[0032] S05. Based on the core hub node set, construct a restricted random walk path, set the upper limit of the walk step size to 8 hops, the restart probability to 0.15, execute Monte Carlo simulation starting from the query starting standard node, run 10,000 walk trajectories and count the access frequency of each node.

[0033] S06. Extract nodes with a visit frequency greater than 500 times to form a high-frequency path skeleton. Apply a bidirectional breadth-first search algorithm to the node pairs in the high-frequency path skeleton to find the shortest associated path. At the same time, record the edge type sequence and attribute change vector on the path.

[0034] S07. Input the shortest associated path in the high-frequency path skeleton into the path semantic consistency model for conflict detection. The path semantic consistency model outputs the logical consistency score of each shortest associated path. Select the shortest associated path with a logical consistency score higher than 0.75 as the set of valid associated paths.

[0035] S08. Cluster the effective association path set, and use a distance metric function based on path length and edge type similarity to divide the shortest association path in the effective association path set into 3 to 7 semantic clusters using a hierarchical clustering algorithm. Each semantic cluster represents a standard association pattern.

[0036] S09. Generate an association path analysis report. The association path analysis report includes a node list of the core hub node set, access path statistics of the high-frequency path skeleton, feature description of the semantic cluster, and early warning information of potential conflict paths. The output format is a structured knowledge graph visualization file.

[0037] The normalized degree value of a node is calculated by dividing the actual degree of the standard node by the maximum degree value in the knowledge graph. The normalized betweenness centrality value is obtained by calculating the proportion of the number of shortest paths passing through the standard node to the total number of shortest paths in the entire graph. The normalized average importance value of neighboring nodes is obtained by dividing the average degree of all first-order neighboring nodes of the standard node by the average degree of the entire graph. The actual degree is the number of edges connected to the standard node. The maximum degree value is the maximum actual degree of all standard nodes in the knowledge graph. The number of shortest paths passing through the standard node is the total number of shortest paths between any two standard nodes in the knowledge graph that pass through the standard node. The total number of shortest paths in the entire graph is the total number of all shortest paths between any two standard nodes in the knowledge graph. The first-order neighboring node is the adjacent standard node directly connected to the standard node. The average degree of the entire graph is the sum of the actual degrees of all standard nodes in the knowledge graph divided by the total number of standard nodes.

[0038] The specific steps of the topological potential energy iterative propagation are as follows: In the t-th iteration, the potential energy contribution received by the standard node i is equal to the sum of the potential energy values ​​of all neighboring nodes j in the (t-1)-th iteration multiplied by the edge weights, then multiplied by the attenuation coefficient of 0.85, and finally added to the initial topological potential energy value of the standard node i multiplied by the retention coefficient of 0.15 as the new potential energy value for the t-th iteration. The overall graph potential energy change rate is defined as the sum of the absolute values ​​of the potential energy differences between all standard nodes in the t-th and (t-1)-th iterations, divided by the sum of the potential energy values ​​of all standard nodes in the (t-1)-th iterations. The neighboring node j is the adjacent standard node directly connected to the standard node i. The edge weights are determined by weighting the basic edge type weights with the historical call frequency. The basic edge type weights are: reference relationship edges have a basic weight of 0.8, and semantic dependency edges have a basic weight of 0.6. The historical call frequency is the cumulative number of times the edge between standard nodes has been used in historical queries; for every 100 uses, the edge weight increases by 0.05. The converged potential energy distribution is the potential energy distribution of each standard node when the rate of change of the full graph potential energy is less than 0.01 in two consecutive iterations.

[0039] The formula for calculating the potential energy distribution concentration coefficient is as follows: Divide the sum of the potential energy of the top 20% of standard nodes by the total potential energy of the entire graph, subtract 0.2, and then divide by 0.8. The result is the potential energy distribution concentration coefficient. This coefficient reflects the non-uniformity of the potential energy distribution; a larger value indicates a more prominent core node. The top 20% of standard nodes are those ranked from highest to lowest potential energy value. The total potential energy of the entire graph is the sum of the potential energy values ​​of all standard nodes in the knowledge graph.

[0040] The specific steps for obtaining the concentration threshold of 0.6 include: selecting 50 samples of the marine survey standard knowledge graph with labeled core node distributions; calculating the potential energy distribution concentration coefficient for each sample; and statistically determining that the potential energy distribution concentration coefficients for samples whose core node distributions conform to expert judgment range from 0.58 to 0.82. The lower bound of this range, 0.58, is adjusted upwards to 0.6 as the concentration threshold, ensuring that when the potential energy distribution concentration coefficient is higher than the concentration threshold, the overlap between the extracted core hub node set and the expert judgment results reaches over 85%. The labeled core node distributions are a set of important standard nodes manually labeled by domain experts according to the standard system structure. The core node distribution conforming to expert judgment means that the overlap rate between the calculated extracted core hub node set and the expert-labeled core nodes exceeds 80%.

[0041] The standard propagation influence assessment module simulates standard knowledge diffusion through restricted random walks. Starting from the standard node to be evaluated, each step selects the next hop node based on the probability distribution after normalizing the outgoing edge weights. The walk step size follows a truncated geometric distribution. A single walk terminates when the maximum step size is reached or the restart probability is triggered, and after restarting, it returns to the starting node to walk again. After 5000 walks, the cumulative number of visits to each standard node is counted, and the normalized visit count value is used as the propagation influence score of the standard node. Standard nodes with a propagation influence score higher than 0.08 are added to the core hub node set. The standard nodes to be evaluated are all standard nodes in the knowledge graph. The probability distribution after normalizing the outgoing edge weights is the probability value obtained by dividing the weight of each outgoing edge of the standard node by the sum of the weights of all outgoing edges. The next hop node is the adjacent standard node reachable from the current standard node along the outgoing edge. The truncated geometric distribution is a geometric distribution with parameter p equal to 0.15 that is truncated when the step size reaches 8. The maximum step size is 8 hops. The normalized value of the number of visits is the cumulative number of times the standard node is visited in 5000 walks divided by 5000.

[0042] The specific steps of the bidirectional breadth-first search algorithm are as follows: A breadth-first traversal is initiated simultaneously from the starting node and the target node. At each level of expansion, the set of visited nodes is recorded. When the sets of visited nodes in both directions intersect, the two paths are merged to form a complete shortest path. The search depth is limited to four levels. If no path is found beyond this limit, it is determined that there is no valid association between the two nodes. The edge type sequence records the type label of each edge on the shortest path. The attribute change vector records the difference in the standard publication year between the first and last nodes of the shortest path and the number of standard level conversions. The starting node and the target node are the two standard nodes in the high-frequency path skeleton for which an association path needs to be found. The set of visited nodes is the set of standard nodes already visited during the breadth-first traversal. The search depth is limited to the maximum number of levels in the breadth-first traversal. The type labels include two categories: reference relationships and semantic dependencies. The difference in the standard publication year is the absolute value of the publication year of the standard corresponding to the terminating node of the shortest path minus the publication year of the standard corresponding to the starting node. The number of standard level conversions refers to the cumulative number of times the standard level is converted from a national standard to an industry standard or vice versa along the shortest path.

[0043] The specific structure of the path semantic consistency model is as follows: The input layer receives the node sequence embedding, edge type sequence embedding, and attribute change vector of the path. The node sequence embedding is generated as a 256-dimensional vector through a pre-trained standard text encoder. The edge type sequence embedding uses a 64-dimensional learnable embedding matrix. The attribute change vector is a 32-dimensional numerical feature. The encoding layer uses a four-layer graph attention convolutional module, each layer containing an attention head with a hidden dimension of 128 and an activation function of rectified linear units. The number of attention heads in the graph attention convolutional module is dynamically adjusted based on three parameters: the normalized value of the average path length, the normalized value of the number of unique edge types in the path, and the normalized value of the standard year span involved in the path. The adjustment formula is as follows: the number of attention heads equals 8 multiplied by the normalized value of the average path length to the power of 0.4, multiplied by the normalized value of the number of unique edge types in the path to the power of 0.3, and multiplied by the normalized value of the standard year span involved in the path to the power of 0.3, rounded up and not exceeding 16. The fusion layer concatenates the node representations output by the graph attention convolutional module using global average pooling and max pooling, and then reduces the dimensionality to 64 dimensions through two fully connected layers. The output layer is activated by a single neuron with sigmoid activation, outputting a logical consistency score between 0 and 1. A higher logical consistency score indicates that the shortest path conforms more closely to the logical rules of the standard system. The node sequence embedding of the path is a vector sequence of standard clause texts corresponding to all standard nodes on the shortest path, transformed by a standard text encoder. The standard text encoder is a pre-trained language model based on a bidirectional encoder representation. The edge type sequence embedding is an embedding vector sequence corresponding to each type label in the edge type sequence. The learnable embedding matrix is ​​an embedding lookup table that can update parameters during model training. The graph attention convolutional module aggregates neighbor node information to update the node representation through an attention mechanism. The attention head consists of multiple sets of attention weights computed in parallel in the graph attention convolutional module. The corrected linear unit is the activation function that takes the maximum value between the input value and 0. The normalized average path length is the average length of all shortest paths in the current batch divided by the upper limit of the walk step size, 8. The normalized value for the number of unique edge types in the path is the number of non-repeating edge types in the shortest association path divided by the total number of edge types in the knowledge graph, which is 12. The normalized value for the standard year span involved in the path is obtained by dividing the difference in the publication years of the standards corresponding to the starting and ending nodes of the shortest association path by 50 years. When the difference in the standard publication years exceeds 50 years, the normalized value for the standard year span involved in the path is 1. Each node is represented as the feature vector of each standard node output by the graph attention convolution module. The global average pooling is the average of the representations of all nodes in the shortest association path along its dimension. The max pooling is the maximum value of the representations of all nodes in the shortest association path along its dimension. The fully connected network is a network layer in the neural network where each neuron is fully connected to all neurons in the previous layer.The single-neuron sigmoid activation means that the output layer contains one neuron and uses the sigmoid activation function to map the output to the interval between 0 and 1. The logical rules of the standard system include no circular dependencies, no mutual exclusion requirements, and transitive reference relationships.

[0044] The steps for establishing the training dataset for the path semantic consistency model specifically include: extracting 6000 verified valid association paths from historical marine survey standard revision records as positive samples. These valid association paths have been reviewed and confirmed by domain experts to be free of logical conflicts. When constructing negative samples, non-adjacent standard node pairs are randomly selected from the knowledge graph, and 5000 conflicting paths containing circular dependencies or mutual exclusion requirements are constructed using a forced path generation algorithm. For each path, node sequences, edge type sequences, and attribute change features are extracted. Each standard clause text in the node sequence is converted into an embedding vector using the standard text encoder. The edge type sequence is mapped to the embedding space after one-hot encoding. The attribute change features include path length normalization, year span normalization, standard level conversion number normalization, and reference depth normalization. The dataset is divided into a training set, a validation set, and a test set in a 7:2:1 ratio. The training set is used for model parameter optimization, the validation set is used for hyperparameter tuning, and the test set is used for final performance evaluation. The historical marine survey standard revision records are document records of version updates and changes in reference relationships in the marine survey standard system over the past 10 years. The domain experts are technical personnel with more than 5 years of experience in marine survey standard setting. The logical conflicts include circular dependencies, mutual exclusion requirements, and violations of transitivity of reference relationships. The non-adjacent standard node pairs are two standard nodes in the knowledge graph that do not have a direct connection edge. The forced path generation algorithm constructs conflicting paths by inserting intermediate nodes and edges that violate logical rules between non-adjacent standard nodes. A circular dependency occurs when standard node A references standard node B in the path, and standard node B directly or indirectly references standard node A. A mutual exclusion requirement occurs when the parameter value ranges specified by two standard clauses in the path do not overlap. The node sequence is an ordered arrangement of standard nodes on the path. The one-hot encoding is the conversion of type labels into vector representations with only one position set to 1 and the rest to 0. The embedding space is the vector space mapped by the learnable embedding matrix using the one-hot encoding. The path length normalization value is the actual path length divided by the upper limit of the walk step size, 8. The year span normalization value is the normalized value of the standard year span involved in the path. The standard level transformation count normalization value is the number of standard level transformations divided by the actual path length. The reference depth normalization value is the deepest reference level in the path divided by the maximum reference level in the knowledge graph, which is 5. The deepest reference level is the maximum number of hops that a given standard node in the path can reach along the reference relationship edges starting from the root standard node. The root standard node is the top-level standard node in the knowledge graph that is not referenced by any other standard.

[0045] The specific steps for training the path semantic consistency model include: using a binary cross-entropy loss function, with positive samples labeled as 1 and negative samples labeled as 0. The optimizer uses an adaptive moment estimation algorithm, with an initial learning rate of 0.001, which decays to 0.9 times its original value after every 20 iterations. The batch size is set to 32, and the total number of training rounds is 150. In each training round, the order of the training set samples is randomly shuffled, and the loss is calculated by forward propagation of the batches into the model. The model parameters are then updated via backpropagation. Every 5 rounds, the model performance is evaluated on the validation set, recording the accuracy and area under the receiver operating characteristic (ROC) curve. Training is terminated when there is no improvement in validation performance for 15 consecutive rounds. After training, the model's logistic consistency discrimination ability is evaluated on the test set, requiring an accuracy of 0.82 or higher and a false positive rate of less than 0.15. The binary cross-entropy loss function measures the difference between the model's output probability distribution and the true labels. The adaptive moment estimation algorithm is an optimization algorithm combining momentum and adaptive learning rate. The learning rate is the step size control parameter for updating model parameters. The batch size refers to the number of samples processed in each forward and backward propagation. The total number of training rounds refers to the number of times the entire training set is traversed. The forward propagation is the process of calculating the model output sequentially from the input layer to the output layer. The backward propagation is the process of updating the model parameters from the output layer to the input layer based on the gradient of the loss function. The accuracy rate is the proportion of samples correctly classified by the model out of the total number of samples. The area under the receiver operating characteristic (ROC) curve is a comprehensive indicator for evaluating the performance of the binary classification model; the closer the value is to 1, the better the model performance. The early stopping mechanism is a strategy to terminate training early to prevent overfitting when validation performance does not improve for several consecutive rounds. The false positive rate is the proportion of negative samples that the model misclassifies as positive samples.

[0046] The distance metric function is defined as the absolute value of the path length difference divided by the longer path length, plus the Jaccard distance of the edge type sequence, with each component weighted 50%. The Jaccard distance is calculated by dividing the number of elements in the intersection of the edge type sets of the two shortest associated paths by the number of elements in the union, and then subtracting the ratio from 1. The hierarchical clustering algorithm uses the average linking method, with a clustering tree cutting threshold of 0.4. Merging stops when the inter-cluster distance is greater than 0.4, and the final number of clusters is controlled between 3 and 7. If the automatic partitioning result is less than 3 semantic clusters, the cutting threshold is lowered to 0.35 for re-clustering; if there are more than 7 semantic clusters, the cutting threshold is increased to 0.5 for re-clustering. The path length difference is the absolute value of the difference between the lengths of the two shortest associated paths. The longer path length is the path length of the larger of the two shortest associated paths. The edge type set is the set of all unique type labels in the edge type sequence. The number of intersection elements is the number of type labels commonly contained in the two edge type sets. The number of elements in the union set is the total number of unique type labels after merging the two edge type sets. The average linking method is a clustering method that calculates the average distance between all sample pairs between two clusters as the inter-cluster distance. The clustering tree is a tree structure constructed by a hierarchical clustering algorithm to represent the hierarchical relationship between samples. The cutting threshold is a distance threshold that determines where to cut the clustering tree to form the final cluster. The inter-cluster distance is the distance metric function value between two semantic clusters.

[0047] The potential conflict path warning information is implemented by detecting abnormal paths within the semantic clusters. For each semantic cluster, the average intra-cluster distance of paths is calculated, and the shortest associated path with a distance greater than twice the average intra-cluster distance is marked as an abnormal path. Simultaneously, highly similar path pairs across clusters are detected. If the distance between two shortest associated paths belonging to different semantic clusters is less than 0.25, they are marked as semantically ambiguous paths, requiring manual review to confirm their logical consistency. The potential conflict path warning information includes the node sequence of the abnormal path, the relevant standard clause text, and a suggested manual review priority. The manual review priority is determined by a comprehensive score based on path length, the number of core nodes involved, and historical call frequency. An abnormal path is the shortest associated path that differs significantly from other paths within the same semantic cluster. The average intra-cluster distance is the average of the distance metric function values ​​between all pairwise shortest associated paths within the semantic cluster. A highly similar path pair is a combination of paths where the distance metric function value between two shortest associated paths is less than a preset threshold. A semantically ambiguous path is a shortest associated path belonging to different semantic clusters but with high similarity. The manual review involves domain experts manually determining the logical consistency of the semantically ambiguous paths. The standard clause text refers to the standard specification text content corresponding to the standard nodes on the abnormal path. The number of core nodes involved refers to the number of standard nodes belonging to the core hub node set on the abnormal path. The comprehensive score is calculated by multiplying the path length normalized value by 0.3, adding the number of core nodes involved normalized value by 0.4, and adding the historical call frequency normalized value by 0.3; the larger the value, the higher the priority. The number of core nodes involved normalized value is the number of core nodes involved divided by the actual path length. The historical call frequency normalized value is the cumulative number of times the abnormal path was used in historical queries divided by the maximum historical call frequency of all paths in the high-frequency path skeleton.

[0048] The structured knowledge graph visualization file uses a graph exchange format. Nodes are mapped to color depth according to their topological potential values; the higher the topological potential value, the darker the color. Standard nodes in the core hub node set are marked with star symbols. Edges use different line types according to their type: reference edges are solid lines, semantic dependency edges are dashed lines, and edge thickness is proportional to the edge weight. Semantic clusters are identified using different background color areas. Access paths in the high-frequency path skeleton are displayed in bold red, and abnormal paths are marked with orange dashed boxes. The structured knowledge graph visualization file supports interactive zooming and node detail queries. Clicking on a node expands the text summary and associated path list of the corresponding standard clause. The graph exchange format is a graph data exchange standard based on Extensible Markup Language (XML). The color depth refers to the brightness of the node's displayed color, represented by a grayscale value normalized to 0 to 255 based on the topological potential value. The star symbol is a pentagram graphic superimposed on the core hub node. The different line types include solid lines, dashed lines, dotted lines, and dashed-dot lines. The edge thickness refers to the width of the displayed line, normalized to a range of 1 to 5 pixels using edge weights. Different background color areas assign a semi-transparent background rectangle of a different color to each semantic cluster. The bold red indicates a display style with a line width of 3 pixels and a red color. The orange dashed border is a rectangular border with an orange color and a dashed line style. The interactive zoom allows users to adjust the display ratio of the visualization using the mouse wheel or touch gestures. The node details query is an interactive function that triggers a pop-up window displaying relevant node information when a node is clicked. The text summary is the first 200 characters of the standard clause text. The associated path list is a display list containing all the shortest associated paths to the node.

[0049] As an optional implementation, the present invention also provides a method for forming a marine survey standard association path analysis system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions execute the above-described method when the computer is run.

[0050] The specific implementation methods of the above steps are described in detail below.

[0051] The specific implementation of step S01 is as follows: First, standard text is extracted from the marine survey standard document library and preprocessed. Natural language processing technology is used to identify structured content such as standard clauses, terminology definitions, and parameter specifications. The identified standard clauses and terminology entities are used as nodes in the knowledge graph. Citation relationships are identified by analyzing citation expressions such as "must conform to a certain standard" in the standard text. Semantic dependencies such as parameter dependencies and method dependencies are extracted using dependency parsing technology to construct semantic dependency edges. Attribute information such as standard number, publication time, and standard level is added to each node to form a complete node-edge-attribute triple structure. The node-edge-attribute triple structure is the basic representation of the knowledge graph, which facilitates subsequent graph algorithm processing and path analysis.

[0052] The specific implementation of step S02 involves traversing all standard nodes in the knowledge graph, counting the number of connecting edges for each standard node to obtain its actual degree, finding the maximum actual degree in the knowledge graph, calculating the ratio of the current node's actual degree to the maximum degree value to obtain the node's degree normalization value, calculating the shortest path between any pair of nodes in the knowledge graph using the Floyd-Warshall full-source shortest path algorithm, counting the number of shortest paths passing through the current node, calculating the ratio of this number to the total number of shortest paths in the entire graph to obtain the betweenness centrality normalization value, extracting all first-order neighbor nodes of the current node, calculating the average actual degree of these first-order neighbor nodes, and dividing the average by... The average importance value of neighboring nodes is obtained by using the average degree of the entire graph. Finally, the normalized value of the node degree is multiplied by a weight coefficient of 0.4, the normalized value of betweenness centrality is multiplied by a weight coefficient of 0.35, and the normalized value of the average importance of neighboring nodes is multiplied by a weight coefficient of 0.25. The initial value of the topological potential is obtained by weighted summing of the three terms. The Floyd-Warshall algorithm is a dynamic programming algorithm used to efficiently calculate the shortest path between all pairs of nodes in the graph. The betweenness centrality measures the pivotal role of a node in information propagation. The weighted summation of the three terms comprehensively considers the local connectivity strength, global path importance, and neighborhood influence of a node, which can comprehensively evaluate the topological status of a standard node in the knowledge graph.

[0053] The specific implementation of step S03 is as follows: Initialize the iteration round counter to 1. For each standard node, traverse all its neighboring nodes, extract the potential energy value of the neighboring node in the previous iteration and the edge weight of the connecting edge, multiply the potential energy value and the edge weight, and then sum them to obtain the total potential energy contribution. Multiply the total potential energy contribution by the attenuation coefficient 0.85 to simulate the energy loss of potential energy during propagation. At the same time, multiply the initial value of the current node's topological potential energy by the retention coefficient 0.15 to maintain the inherent importance of the node. Add the two to obtain the new potential energy value of the current node in this iteration. After completing the potential energy update of all nodes, calculate the sum of the absolute values ​​of the potential energy differences between the current and previous rounds. Divide the sum of the absolute values ​​by the sum of the potential energy values ​​of all nodes in the previous round to obtain the overall graph potential energy change rate. Determine the overall graph potential energy change rate. If the convergence rate is less than the convergence threshold of 0.01, and the iteration round counter is incremented by 1 and the above process is repeated, if convergence has occurred, the potential energy value of each node is recorded as the converged potential energy distribution. The potential energy distribution is traversed and sorted from high to low potential energy value. The top 20% of standard nodes are extracted, and the sum of the potential energy of the standard nodes is divided by the sum of the potential energy of the entire graph to obtain the concentration ratio. The concentration ratio is subtracted by 0.2 and then divided by 0.8 to obtain the potential energy distribution concentration coefficient. The design of the attenuation coefficient and the retention coefficient is based on the principle of the confined diffusion model, so that the potential energy propagation can reflect the topological structure of the graph and maintain the initial characteristics of the nodes. The convergence threshold of 0.01 is an empirical value to ensure the stability of the iteration process. The potential energy distribution concentration coefficient quantifies the prominence of the core nodes.

[0054] The specific implementation of step S04 involves reading the potential energy distribution concentration coefficient calculated in step S03, comparing it with a concentration threshold of 0.6. If the potential energy distribution concentration coefficient is greater than 0.6, it indicates that the potential energy distribution exhibits a clear core-edge structure. In this case, the top 15% of standard nodes are extracted according to their potential energy values ​​from high to low to form a core hub node set, and step S05 is executed directly. If the potential energy distribution concentration coefficient is not greater than 0.6, it indicates that the potential energy distribution is relatively uniform and lacks a clear core. In this case, the standard propagation influence assessment module is activated. This module performs 5000 restricted random walks on each standard node in the knowledge graph. Each walk starts from the current node and randomly selects the next hop node according to the probability distribution after the outgoing edge weights are normalized. The walk step size follows a parameter. The walk uses a truncated geometric distribution of 0.15 and a maximum step size of 8 hops. It terminates and returns to the starting node when the walk reaches the maximum step size or triggers a restart probability of 0.15. The total number of visits to each node in 5000 walks is divided by 5000 to obtain a normalized visit count as the propagation influence score. Nodes with a propagation influence score higher than 0.08 are added to the core hub node set. The concentration threshold of 0.6 is determined based on statistical analysis of 50 labeled samples. This restricted random walk simulates the propagation process of standard knowledge in the graph. The truncated geometric distribution controls the walk length distribution, causing most walks to concentrate in local areas. The restart probability prevents walks from getting trapped in local clusters. The propagation influence score threshold of 0.08 is used as a reference value to screen nodes with significant propagation effects.

[0055] The specific implementation of step S05 is as follows: The user inputs the query starting standard node; the Monte Carlo simulation counter is initialized to 0; the maximum step size is set to 8 hops; and the restart probability is set to 0.15. For each simulation starting from the query starting standard node, all outgoing edges and their weights are extracted from the current node. The sum of all outgoing edge weights is calculated, and the weight of each outgoing edge is divided by the sum of the weights to obtain the transition probability distribution. Based on the transition probability distribution, an outgoing edge is randomly selected and the user moves to the corresponding next-hop node. The access frequency of the next-hop node is incremented by 1 in the node access record table. It is then determined whether the current step size has reached the maximum step size or whether the random number is less than the restart probability of 0.15. If either condition is met, the current traversal is terminated. The simulation counter is incremented by 1 and it is checked whether 10,000 visits have been reached. If not, a new traversal is started from the initial standard node of the query. After 10,000 visits, the node access record table is traversed, and standard nodes with an access frequency greater than 500 visits are selected to form a high-frequency path skeleton. The Monte Carlo simulation reflects the relevance of nodes in the query context by statistically analyzing the node access frequency through a large number of random walks. The transition probability distribution is based on edge weights, which makes the walk tend to select important associations. The upper limit of the walk step size of 8 hops is a reference value to balance the path exploration range and computational efficiency. The restart probability of 0.15 controls the backtracking frequency of the walk to enhance the robustness of the algorithm. The access frequency threshold of 500 visits is a reference value used to filter low-relevance nodes.

[0056] The specific implementation of step S06 is to traverse all node pairs in the high-frequency path skeleton. For each node pair, two queues are initialized, and a breadth-first traversal is initiated from the starting node and the target node, respectively. The current node is taken from the starting node queue, all its neighbor nodes are expanded, and the visited path is recorded. Simultaneously, the current node is taken from the target node queue and expanded in reverse. At each expansion layer, it is determined whether there is an intersection between the set of visited nodes in the forward traversal and the set of visited nodes in the reverse traversal. If there is an intersection, the intersection node is extracted. The forward path from the starting node to the intersection node and the reverse path from the intersection node to the target node are merged to form a complete shortest associated path, and the shortest path is recorded. The type label of each edge on the short association path constitutes the edge type sequence. The difference in the publication year of the standard corresponding to the first and last nodes of the path and the number of times the standard level on the path is converted from national standard to industry standard or from industry standard to national standard constitute the attribute change vector. If the traversal depth exceeds 4 layers and no intersection is found, it is determined that there is no effective association between the two nodes and they are skipped. The bidirectional breadth-first search algorithm reduces the search space from exponential to polynomial by exploring from both ends at the same time. The search depth limit of 4 layers is a reference value to control the path length and avoid excessively long paths. The edge type sequence and attribute change vector are the semantic features of the path used for subsequent consistency detection.

[0057] The specific implementation of step S07 involves using all the shortest association paths extracted in step S06 as input to the path semantic consistency model. For each shortest association path, its node sequence is extracted, and each standard clause text is converted into a 256-dimensional embedding vector using a standard text encoder. Edge type sequences are extracted, and each type label is mapped to an embedding vector using a 64-dimensional learnable embedding matrix. The difference in standard publication year and the number of standard level conversions in the attribute change vector are normalized to 32-dimensional numerical features. The node sequence embedding, edge type sequence embedding, and attribute change vector are concatenated and input into the encoding layer of the path semantic consistency model. The four-layer graph attention convolutional module in the encoding layer aggregates neighbor node information through an attention mechanism to update the representation of each node on the path. The number of attention heads is normalized according to the average path length of the current batch. The normalized value of the number of unique edge types in the path and the normalized value of the standard year span involved in the path are dynamically calculated by power product. The fusion layer performs global average pooling and max pooling on the node representation output by the graph attention convolution module and then concatenates them. The dimensionality is reduced to 64 dimensions through two fully connected network layers. The output layer maps the features to a logical consistency score between 0 and 1 through a single neuron sigmoid activation function. The logical consistency scores of all shortest associated paths are traversed, and paths with scores higher than 0.75 are selected to form a set of effective associated paths. The path semantic consistency model is based on a graph neural network architecture and can learn deep semantic dependencies between standard terms. The dynamic adjustment of the number of attention heads allows the model to adapt to path structures of different complexities. The logical consistency score threshold of 0.75 is used as a reference value to filter potential conflicting paths.

[0058] The specific implementation of step S08 involves traversing all path pairs in the effective associated path set. For each path pair, the absolute value of its length difference is calculated and divided by the length of the longer path to obtain the length distance component. The edge type sequences of the two paths are extracted and converted into edge type sets. The number of elements in the intersection of the two edge type sets is calculated and divided by the number of elements in the union to obtain the Jaccard similarity. The Jaccard similarity is subtracted from 1 to obtain the Jaccard distance. The length distance component and the Jaccard distance are each multiplied by a weight of 0.5 and then added together to obtain the distance metric for the path pair. Based on the distance metric, a distance matrix of the effective associated path set is constructed. The average link method of the hierarchical clustering algorithm is used to iteratively merge the two clusters with the smallest distance. After each merge, the average link between the new cluster and other clusters is calculated. The distance matrix is ​​updated using the average distance. Merging stops when the distance between clusters is greater than the cutting threshold of 0.4. The number of semantic clusters formed is counted. If the number of semantic clusters is less than 3, the cutting threshold is lowered to 0.35 and clustering is re-executed. If the number of semantic clusters is more than 7, the cutting threshold is raised to 0.5 and clustering is re-executed. For each semantic cluster, the common features of its contained paths are analyzed to extract standard association patterns. The distance metric function combines path structure similarity and semantic similarity to achieve multi-dimensional path comparison. The hierarchical clustering algorithm does not require a preset number of clusters through a bottom-up merging strategy. The cutting threshold of 0.4 is a reference value that can be adjusted to control the granularity of clusters. The range of 3 to 7 semantic clusters is a reference value to balance cluster fineness and interpretability.

[0059] The specific implementation of step S09 involves generating a node list of the core hub node set. This list includes the standard number, standard name, and topological potential value of each core node. The access frequency of each access path in the high-frequency path skeleton is statistically analyzed and sorted from high to low to generate an access path statistics table. Each semantic cluster is traversed to extract the average length of paths within the cluster, the distribution of main edge types, and the range of standard years involved. These statistical features are summarized into a semantic cluster feature description. For each semantic cluster, the average distance metric between all pairs of paths within the cluster is calculated as the average distance within the cluster. Paths with a distance greater than twice the average distance within the cluster are marked as abnormal paths. Simultaneously, path pairs belonging to different semantic clusters but with a distance less than 0.25 are marked as semantically ambiguous paths. A comprehensive score is obtained by multiplying the normalized path length value by 0.3, adding the normalized value of the number of core nodes involved by 0.4, and adding the normalized value of historical call frequency by 0.3 to the normalized path length value of the abnormal paths and semantically ambiguous paths, and finally obtaining a comprehensive score. The system prioritizes manual reviews based on comprehensive scores, ranking them from highest to lowest. It then compiles node sequences of abnormal paths, relevant standard clause texts, and manual review priorities to form a potential conflict path warning system. The system integrates the node list of the core hub node set, access path statistics of high-frequency path skeletons, semantic cluster feature descriptions, and potential conflict path warning information into a related path analysis report. A structured knowledge graph visualization file is output using a graphical exchange format. In the visualization file, node colors are mapped to topological potential values, core hub nodes are marked with star symbols, reference edges are represented by solid lines, semantic dependency edges are represented by dashed lines, edge thickness is proportional to edge weight, each semantic cluster is identified by a different background color area, high-frequency paths are displayed in bold red, and abnormal paths are marked with orange dashed boxes. The related path analysis report provides users with a panoramic view of the standard system architecture and potential problem diagnosis. The visualization file supports interactive zooming and node detail queries to enhance the user's analytical experience.

[0060] It should be noted that the key technical ideas of this invention include a core node identification mechanism based on topological potential energy iterative propagation. By simulating the potential energy diffusion process in the physical field and combining it with multi-dimensional evaluation of node degree, betweenness centrality, and neighbor importance, the core hub nodes in the standard system can be accurately located. Compared with the traditional single centrality index, the multi-dimensional evaluation overcomes the problem of nodes with dense local connections but low global importance being misjudged as core nodes. The potential energy iterative propagation, through the balanced design of attenuation coefficient and retention coefficient, makes the potential energy distribution reflect both the topological structure of the graph and maintain the inherent characteristics of the nodes. The adaptive judgment mechanism of the concentration coefficient can dynamically select the core node extraction strategy or activate the propagation influence evaluation module according to the potential energy distribution characteristics, which significantly improves the adaptability and accuracy of core node identification under different standard system architectures.

[0061] The second key technical approach is an efficient path construction method that integrates Monte Carlo simulation and bidirectional breadth-first search. It identifies and queries high-frequency path skeletons by statistically analyzing node access frequencies through large-scale simulations using restricted random walks. Based on these skeletons, bidirectional breadth-first search is applied to reduce the time complexity of path finding from exponential to polynomial time. Compared to traditional full graph traversal methods, this two-stage path construction strategy significantly reduces computational overhead while maintaining path quality by first narrowing the search space and then performing a precise search. The restart probability and step size constraints introduced by the restricted random walk prevent the algorithm from getting stuck in locally dense regions. The simultaneous expansion strategy of the bidirectional search halves the search depth. For standard knowledge graphs with millions of nodes, it can complete associated path analysis within seconds.

[0062] The third key technical approach is a path semantic consistency detection model based on graph attention convolution. This model dynamically adjusts the number of attention heads to adapt to the semantic learning needs of paths with different complexities. By integrating multimodal inputs such as node sequence embedding, edge type sequence embedding, and attribute change vectors, it can comprehensively capture the structural and semantic features of the path. Compared with traditional rule matching or pattern library methods, this deep learning model can learn the implicit logical dependencies and conflict patterns between standard clauses. The design of the correlation between the number of attention heads and the average path length, the number of unique edge types, and the standard year span allows the model parameters to be dynamically adjusted with the path features, significantly improving the model's generalization ability to heterogeneous standard texts and cross-year standard evolution. The logical consistency score provides a quantitative indicator for path quality to support automated conflict detection.

[0063] The synergistic effect of the three key technical approaches described above is to construct a complete analysis process from core node identification to path construction and semantic verification. The topological potential energy mechanism provides a high-quality starting point and skeleton nodes for path construction, the efficient path construction method quickly establishes association channels between core nodes, and the semantic consistency detection model intelligently filters and warns of conflicts in the constructed paths. The three form a progressive analysis system from coarse to fine and from structure to semantics. Compared with existing static standard indexes or simple association query methods, this invention can achieve intelligent deep association analysis in large-scale heterogeneous standard knowledge graphs. The multi-stage collaborative mechanism enables the output of each stage to provide optimized input for the next stage. The overall analysis accuracy and efficiency are significantly better than a simple combination of single algorithms.

[0064] It should be noted that this invention also solves the following technical problem: the low efficiency of logical conflict identification caused by the reliance on manual review in existing technologies for semantic consistency detection of associated paths. Traditional methods, after extracting standard associated paths, require domain experts to review each path for logical conflicts such as circular dependencies, mutual exclusion requirements, or violations of transitivity of reference relationships. Manual review is inefficient and prone to missing hidden conflicts. This invention constructs a path semantic consistency model. By embedding the node sequence, edge type sequence, and attribute change vector of the input path, a graph attention convolution module is used to aggregate neighbor node information and update the node representation. The node representation output from the encoding layer is concatenated by global average pooling and max pooling, and then dimensionality-reduced through a fully connected network. The output layer uses a single-neuron sigmoid activation to output a logical consistency score, automatically selecting paths with logical consistency scores higher than a threshold as the effective associated path set. This significantly improves the efficiency of logical conflict identification and reduces the workload of manual review.

[0065] Furthermore, this invention addresses the technical problem in existing technologies where the lack of semantic distinction in associated path grouping leads to difficulties in identifying standard association patterns. Traditional methods simply group extracted associated paths according to path length or the number of standards involved, ignoring the semantic differences in edge type sequences within the paths. This results in paths with different association patterns being mixed in the same group, making it difficult to identify typical association patterns within the standard system. This invention employs a distance metric function based on path length and edge type similarity. It calculates the absolute value of the path length difference divided by the longer path length, adds the Jaccard distance of the edge type sequences, and uses a hierarchical clustering algorithm to divide the effective associated path set into semantic clusters, with each semantic cluster representing a type of standard association pattern. Simultaneously, by detecting abnormal paths within clusters and highly similar path pairs across clusters, it generates early warning information for potential conflict paths, achieving semantic grouping of associated paths and automatic identification of standard association patterns.

[0066] Specifically, the principle of this invention is as follows: This invention calculates the initial potential energy of a node by integrating three dimensions of topological features: node degree, betweenness centrality, and average importance of neighboring nodes. This overcomes the limitations of traditional methods that rely solely on a single local feature, and can simultaneously capture the node's connectivity breadth, bridging ability, and neighborhood influence. The topological potential energy iterative propagation mechanism simulates the diffusion process of potential energy in the graph structure. Through potential energy transfer between adjacent nodes, the importance assessment results are globally converged, avoiding the trap of local optima. The potential energy distribution concentration coefficient can identify the non-uniformity of the core node distribution. When the distribution is too dispersed, the standard propagation influence assessment module is triggered. Through a restricted random walk, the standard knowledge diffusion process is simulated, supplementing the identification of nodes with strong propagation influence that were missed by the topological potential energy calculation, ensuring the integrity and accuracy of the core hub node set. Based on the accurately identified core hub node set, a restricted random walk path is constructed, which can focus the search on paths near high-potential-energy nodes. High-frequency path skeletons are extracted through access frequency statistics, and then a bidirectional breadth-first search algorithm is applied to find the shortest associated path within the skeleton, significantly reducing invalid searches and improving the efficiency of associated path analysis.

[0067] The following provides a specific embodiment 1 of the present invention. The specific implementation methods of steps S01, S05, S06, and S07 in this embodiment 1 are the same as those described above, and will not be repeated in detail here. The specific implementation methods of other steps are described in detail below.

[0068] The specific implementation of step S02 is to calculate the initial value of the topological potential of each standard node in the knowledge graph. This initial value of the topological potential is obtained by weighted summation of the node degree normalization value, the betweenness centrality normalization value, and the average importance normalization value of neighboring nodes, as shown below:

[0069] ;

[0070] In the formula, For standard nodes The initial value of the topological potential energy is dimensionless; For standard nodes The normalized value of the node degree is dimensionless; For standard nodes The normalized value of the betweenness centrality is dimensionless; For standard nodes The average importance normalized value of neighboring nodes, dimensionless; This refers to the index of standard nodes in the knowledge graph. Among them, The calculation formula is expressed as follows:

[0071] ;

[0072] In the formula, For standard nodes The actual degree, in units of stripes; This represents the maximum degree value in the knowledge graph, expressed in terms of records. The calculation formula is expressed as follows:

[0073] ;

[0074] In the formula, For standard nodes To standard node The total number of shortest paths, in units of paths; For passing through standard nodes slave node To the node The number of shortest paths, expressed in units of paths; This is the starting standard node in the knowledge graph; For target standard nodes in the knowledge graph; It refers to any standard node in the knowledge graph. The calculation formula is expressed as follows:

[0075] ;

[0076] In the formula, For standard nodes The set of first-order neighbor nodes; For set The neighbor node index in the data; For neighboring nodes The actual degree, in units of stripes; This represents the number of first-order neighbor nodes, expressed in units. The average degree of the entire graph, expressed in terms of nodes, is obtained by dividing the sum of the actual degrees of all standard nodes in the knowledge graph by the total number of standard nodes. The calculation formula is as follows:

[0077] ;

[0078] In the formula, For standard nodes The actual degree, in units of stripes; This represents the total number of standard nodes in the knowledge graph, expressed in units of nodes. This is the traversal index for all standard nodes in the knowledge graph.

[0079] The specific implementation of step S03 is to perform iterative propagation of topological potential energy, in the first... Standard nodes in round iteration The formula for updating the potential energy value is expressed as follows:

[0080] ;

[0081] In the formula, For standard nodes In the The potential energy of the wheel is dimensionless; This represents the current iteration round number; For standard nodes The neighbor node index; For standard nodes with neighboring nodes The edge weights between them are dimensionless. For neighboring nodes In the The potential energy of the wheel is dimensionless. Edge weights. The calculation formula is expressed as follows:

[0082] ;

[0083] In the formula, The basic weights for edge types are dimensionless; the basic weight for reference edges is 0.8, and the basic weight for semantic dependency edges is 0.6. For standard nodes With nodes The historical call frequency of the edges between them, in times; The function is rounded down. Rate of change of potential energy across the entire graph. The calculation formula is expressed as follows:

[0084] ;

[0085] In the formula, For the first The rate of change of potential energy of the wheel as a whole is dimensionless; For the first Wheel and the first The sum of the absolute values ​​of the potential energy differences at all standard nodes of the wheel, dimensionless; For the first The sum of the potential energy values ​​of all standard nodes in each round is dimensionless. The iteration stops when the rate of change of the total potential energy in two consecutive rounds is less than 0.01. The converged potential energy distribution is recorded and the concentration coefficient of the potential energy distribution is calculated.

[0086] The specific implementation of step S04 involves calculating the potential energy distribution concentration coefficient and comparing it with a concentration threshold. The calculation formula is expressed as follows:

[0087] ;

[0088] In the formula, is the potential energy distribution concentration coefficient, which is dimensionless; This is the set of standard nodes whose potential energy values ​​rank in the top 20%. The sum of potential energy of the top 20% of standard nodes, dimensionless; This represents the total potential energy of the entire graph, dimensionless. If the potential energy distribution concentration coefficient is greater than the concentration threshold of 0.6, the top 15% of standard nodes by potential energy value are extracted as the core hub node set. If the potential energy distribution concentration coefficient is not greater than the concentration threshold, the standard propagation influence assessment module is called to recalculate the node importance. The propagation influence score is located in the standard propagation influence assessment module. The calculation formula is expressed as follows:

[0089] ;

[0090] In the formula, For standard nodes The score for the dissemination influence is dimensionless; For standard nodes The cumulative number of visits during 5000 visits, expressed as visits. Nodes with a propagation influence score higher than 0.08 are added to the core hub node set.

[0091] The specific implementation of step S08 involves clustering and grouping the effective associated path set, using a distance metric function. The calculation formula is expressed as follows:

[0092] ;

[0093] In the formula, For the shortest path and The distance between them is a dimensionless metric. and For two distinct shortest association paths in the set of valid association paths; and These are the shortest association paths. and The length, in units of jumps; This represents the maximum length of the two paths, expressed in jumps. and These are the shortest association paths. and A set of edge types; The number of elements in the intersection of two sets of edge types, expressed in units of 1; represents the number of elements in the union of two edge type sets, expressed in units of . The hierarchical clustering algorithm uses the average linking method, with the clustering tree cutting threshold set to 0.4. The final number of clusters is controlled between 3 and 7. If the automatic partitioning result is less than 3 semantic clusters, the cutting threshold is lowered to 0.35 for re-clustering; if there are more than 7 semantic clusters, the cutting threshold is increased to 0.5 for re-clustering.

[0094] The specific implementation of step S09 is to generate an association path analysis report, in which the manual review priority is given to the potential conflict path early warning information. The calculation formula is expressed as follows:

[0095] ;

[0096] In the formula, The score is a dimensionless, comprehensive evaluation of the priority of manual review of abnormal paths. This represents the actual length of the abnormal path, in jumps. This represents the number of core nodes involved in the abnormal path, in units of [number]. This represents the cumulative number of times the abnormal path was used in historical queries, expressed in times. This represents the maximum historical call frequency of all paths in the high-frequency path skeleton, expressed in times. Abnormal paths are determined by calculating the average intra-cluster distance for each semantic cluster. The calculation formula is expressed as follows:

[0097] ;

[0098] In the formula, semantic clusters The average intra-cluster distance is dimensionless; For a certain semantic cluster; and semantic clusters Different shortest paths within; and The index of the path within the cluster and Ensure that each path pair is calculated only once; semantic clusters The number of paths within a cluster, in units of individual paths. The shortest association path whose distance is greater than twice the average distance within the cluster is marked as an abnormal path, as described in the following criteria:

[0099] ;

[0100] In the formula, The shortest associated path to be determined; semantic clusters The central path; For path The distance metric to the center path, dimensionless. Center path Defined as the path with the minimum sum of distances to all other paths within the cluster, it is expressed by the following formula:

[0101] ;

[0102] In the formula, and semantic clusters Path within; and For path index; This represents the path that minimizes the objective function. The association path analysis report is output as a structured knowledge graph visualization file, using a graphical exchange format.

[0103] The number of attention heads in the graph attention convolutional module of the path semantic consistency model The calculation formula is expressed as follows:

[0104] ;

[0105] In the formula, The number of attention heads in the graph attention convolution module, expressed in units of 1; This represents the average length of all shortest associated paths in the current batch, in jumps. This represents the number of unique edge types in the shortest path, expressed in units of 1; This is the difference in the publication year of the standard corresponding to the starting and ending nodes of the shortest path, in years. This means that when the difference between the release years exceeds 50 years, 50 years will be used; It is a rounding function; This indicates that the result does not exceed 16.

[0106] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0107] A technical team undertook a marine ecological environment survey project, requiring the execution of comprehensive survey tasks involving hydrological observation, biological sampling, sediment analysis, and pollutant monitoring. This project needed to comply with multiple national and industry standards, including GB / T 12763.1 General Principles of Marine Survey Specifications, GB / T 12763.2 Marine Survey Specifications for Marine Hydrological Observation, HY / T147.1 Guidelines for Nearshore Marine Ecological Health Assessment, and HY / T 147.3 Specifications for Marine Sediment Quality Assessment, totaling 58 standards. The technical team needed to analyze the relationships between these standards, identify core standard nodes, and ensure that the survey plan conformed to the logical requirements of the standard system. Traditional methods rely on manually sifting through standard documents, which is time-consuming and prone to missing key reference relationships. Therefore, the technical team adopted the graph algorithm-based marine survey standard association path analysis method of this invention to solve this problem.

[0108] The technical team first constructed a knowledge graph of marine survey standards, transforming 58 standard texts into a node-edge-attribute triple structure. The knowledge graph contains 58 standard nodes, representing standard clauses and terminology entities. A total of 137 referencing edges and 89 semantic dependency edges were identified. Attribute records include the standard number, publication time, and standard level. For example, the GB / T12763.2 standard references the terminology definition section of the GB / T 12763.1 standard, establishing a referencing edge between them with a base edge weight of 0.8. The technical team calculated the initial topological potential of each standard node in the knowledge graph. For the GB / T12763.1 node, its actual degree is 15, the maximum degree value in the knowledge graph is 18, and the normalized degree value is 0.833. The normalized betweenness centrality value of this node was obtained by calculating the proportion of the number of shortest paths passing through this node to the total number of shortest paths in the entire graph, resulting in 0.267. The average degree of the first-order neighbors of this node is 9.2, the average degree of the entire graph is 7.8, and the normalized value of the average importance of the neighbors is 1.179. The initial value of the topological potential is obtained by weighted summation of the above three normalized values, with weight coefficients of 0.4, 0.35, and 0.25, respectively, and the calculated result is 0.622.

[0109] The technical team performed iterative propagation of topological potential energy. In the first iteration, the GB / T 12763.1 node received potential energy contributions from its 15 neighboring nodes. These neighboring nodes included GB / T 12763.2, GB / T 12763.6, and HY / T 147.1, whose potential energy values ​​in round 0 were 0.485, 0.412, and 0.538, respectively. Regarding edge weights, the base weight of the reference edge between GB / T 12763.1 and GB / T 12763.2 was 0.8. After 320 historical calls, the edge weight increased by 0.15, resulting in a final edge weight of 0.95. The sum of the potential energy values ​​of all neighboring nodes multiplied by their corresponding edge weights yields 6.873. Multiplying this by the attenuation coefficient of 0.85 gives 5.842. Finally, adding the initial topological potential energy value of the node (GB / T 12763.1) of 0.622 multiplied by the retention coefficient of 0.15 gives 0.093. The new potential energy value for the first round is 5.935. After iterative propagation, in the 12th iteration, the rate of change of the overall graph potential energy over two consecutive iterations is 0.008, which is less than the set threshold of 0.01, and the iteration stops. Figure 2 As shown, the technical team recorded the potential energy distribution after convergence and calculated the potential energy distribution concentration coefficient.

[0110] The technical team calculated the total potential energy of the top 20% of standard nodes (12 nodes) to be 68.47, and the total potential energy of the entire graph to be 152.83. The potential energy distribution concentration coefficient was calculated to be 0.548. This coefficient is less than the concentration threshold of 0.6, indicating that the potential energy distribution is relatively dispersed and the core nodes are not prominent enough. The technical team called the standard propagation influence assessment module to recalculate the importance of nodes. Starting from the standard node to be evaluated, HY / T 147.3, the next hop node was selected based on the probability distribution after normalization of the outgoing edge weights. This node has 4 outgoing edges with edge weights of 0.85, 0.72, 0.68, and 0.95, respectively, and the normalized probability distributions are 0.266, 0.225, 0.213, and 0.297. The walk step size follows a truncated geometric distribution with a parameter p equal to 0.15 and a maximum step size of 8 hops. After 5000 visits, the cumulative number of visits to each standard node was counted. The HY / T 147.1 node was visited 542 times, and the normalized value of the visit count was 0.108, which was used as the dissemination influence score. Seven standard nodes with a dissemination influence score higher than 0.08 were selected and added to the core hub node set, resulting in a final core hub node set containing 19 standard nodes.

[0111] The technical team constructed a restricted random walk path based on the core hub node set, setting the maximum walk step size to 8 hops and the restart probability to 0.15. Starting from the query starting standard node GB / T 12763.2, a Monte Carlo simulation was performed, running 10,000 walks and counting the access frequency of each node, as shown in Table 1.

[0112] Table 1. Statistics of High-Frequency Access Nodes

[0113]

[0114] The technical team extracted nodes accessed more than 500 times to construct a high-frequency path skeleton, containing 7 core nodes. A bidirectional breadth-first search algorithm was applied to the node pairs in the high-frequency path skeleton to find the shortest associated path, simultaneously initiating a breadth-first traversal from nodes GB / T12763.2 and HY / T 147.1. During the first-level expansion, node GB / T 12763.2 visited nodes GB / T 12763.1 and GB / T 17378.4, while node HY / T 147.1 visited nodes HY / T 147.3 and GB / T12763.6. During the second-level expansion, the visited node sets in both directions intersect at node GB / T 12763.6. The two paths are merged to form the complete shortest associated path GB / T 12763.2 to GB / T 12763.1 to GB / T 12763.6 to HY / T 147.1, with a path length of 3. For example... Figure 3 As shown, the technical team records the type label of each edge on the shortest association path. The edge type sequence is: reference relationship, semantic dependency, and reference relationship again. The attribute change vector records the difference in the standard release year between the first and last nodes of the path, which is 6 years, and the standard level conversion is 1 time.

[0115] The technical team input 21 shortest association paths into the path semantic consistency model for conflict detection. The input layer of the path semantic consistency model receives the node sequence embeddings, edge type sequence embeddings, and attribute change vectors of the paths. Node sequence embeddings are generated as 256-dimensional vectors using a pre-trained standard text encoder; edge type sequence embeddings use a 64-dimensional learnable embedding matrix; and attribute change vectors are 32-dimensional numerical features. The encoding layer uses a four-layer graph attention convolutional module. For paths GB / T12763.2 to GB / T 12763.1 to GB / T 12763.6 to HY / T 147.1, the average length of all shortest association paths in the current batch is 3.4, and the normalized average path length value is 0.425. The number of unique edge types in the paths is 2, the total number of edge types in the knowledge graph is 12, and the normalized value of the number of unique edge types in the paths is 0.167. The standard year span involved in the paths is 6 years, and the normalized value of the standard year span involved in the paths is 0.12. The number of attention heads is dynamically adjusted to 10. The fusion layer concatenates the node representations output by the graph attention convolutional module using global average pooling and max pooling, and then reduces the dimensionality to 64 dimensions through two fully connected layers. The output layer is activated by a single neuron with sigmoid activation, and the output logistic consistency score is 0.87. The technical team selected the shortest association paths with a logistic consistency score higher than 0.75 as the set of effective association paths, and a total of 18 effective association paths were selected.

[0116] The technical team clustered the set of valid associated paths, using a distance metric based on path length and edge type similarity. The distances between paths GB / T 12763.2 to GB / T 12763.1 to GB / T 12763.6 to HY / T 147.1 and HY / T 147.3 to GB / T 17378.4 to GB / T 12763.1 were calculated, with path lengths of 3 and 2 respectively. The absolute value of the path length difference was 1, the longer path length was 3, and the first term was 0.333. The edge type sets of the two paths were reference relation and semantic dependency, respectively. The intersection had 1 element, the union had 2 elements, and the Jaccard distance was 0.5. The distance metric value was 0.417. Figure 4 As shown, the technical team used a hierarchical clustering algorithm to divide the 18 shortest association paths into 5 semantic clusters, with the clustering tree's cutting threshold set to 0.4. Semantic cluster 1 contains 7 paths, representing the association pattern of terminology definition references in the standard system. Semantic cluster 2 contains 5 paths, representing the association pattern of observation method dependence. Semantic cluster 3 contains 3 paths, representing the association pattern of quality control requirement transmission. Semantic cluster 4 contains 2 paths, representing the association pattern of data processing flow. Semantic cluster 5 contains 1 path, representing the association pattern of special application scenarios.

[0117] The technical team detected abnormal paths within semantic clusters. The average intra-cluster distance for paths in semantic cluster 1 was calculated to be 0.28. Within semantic cluster 1, the average distance between paths GB / T 12763.1 to HY / T 081 to GB / T 17378.4 to GB / T 12763.6 to HY / T147.1 and other paths was 0.64, exceeding the threshold of 0.56 (twice the average intra-cluster distance), and was therefore marked as an abnormal path. The team also detected highly similar path pairs across clusters. The distance between paths HY / T 147.1 to GB / T 12763.2 to GB / T 12763.1 in semantic cluster 2 and paths HY / T 147.3 to GB / T 12763.6 to GB / T 12763.1 in semantic cluster 3 was 0.22, less than 0.25, and was therefore marked as semantically ambiguous paths. The technical team generated a potential conflict path warning. The path length normalization value for abnormal paths GB / T12763.1 to HY / T 081 to GB / T 17378.4 to GB / T 12763.6 to HY / T 147.1 is 0.625, involving 4 core nodes, with a normalized value of 0.8 for the number of core nodes involved. The historical call frequency is 78 times, and the maximum historical call frequency across all paths in the high-frequency path skeleton is 245 times, with a normalized value of 0.318 for historical call frequency. The overall priority score for manual review is 0.622, requiring priority for manual review.

[0118] like Figure 5 As shown, the technical team generated a correlation path analysis report. The report includes a list of 19 standard nodes in the core hub node set, access path statistics for 7 core nodes in the high-frequency path skeleton, feature descriptions of 5 semantic clusters, and early warning information for potential conflict paths, including one abnormal path and one pair of semantically ambiguous paths. The output format is a structured knowledge graph visualization file using a graphical exchange format. Nodes are mapped to color depth according to their topological potential value. According to GB / T 12763.1, the node topological potential value is 8.73, and the color depth grayscale value is 223. Standard nodes in the core hub node set are marked with a star. Edges use different line types according to their type: reference edges are solid lines, semantic dependency edges are dashed lines, and edge thickness is proportional to edge weight; an edge weight of 0.95 corresponds to a line width of 4.75 pixels. Semantic clusters are identified using different background color areas; access paths in the high-frequency path skeleton are displayed in bold red, and abnormal paths are marked with orange dashed boxes. The visualized document supports interactive zooming and node detail querying. The technical team can click on the GB / T 12763.1 node to expand the text summary and associated path list of the corresponding standard clauses.

[0119] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

[0120] It should be noted that the user data involved in the embodiments of this application have all been authorized, acquired, processed, and transmitted in accordance with legal and regulatory requirements.

Claims

1. A method for analyzing the association paths of marine survey standards based on graph algorithms, characterized in that, This includes constructing a knowledge graph of marine survey standards, transforming standard texts into node-edge-attribute triple structures; calculating the initial topological potential energy of each standard node in the knowledge graph, obtained by weighted summation of node degree normalization, betweenness centrality normalization, and average importance normalization of neighboring nodes; performing iterative propagation of topological potential energy, where standard nodes receive potential energy contributions from neighboring nodes and update their own potential energy values, stopping iteration when the rate of change of the overall graph potential energy is less than a set threshold for two consecutive iterations; calculating the potential energy distribution concentration coefficient and determining whether it is greater than a concentration threshold, if greater, extracting the standard nodes with the highest potential energy ranking as the core hub node set, otherwise calling the standard propagation influence assessment module to recalculate node importance and update the core hub node set. The process involves: constructing a restricted random walk path based on the core hub node set; performing Monte Carlo simulations starting from the initial standard node and counting the access frequency of each node; extracting nodes with access frequencies greater than a set number to form a high-frequency path skeleton; applying a bidirectional breadth-first search algorithm to find the shortest associated path for node pairs in the high-frequency path skeleton; inputting the shortest associated path into a path semantic consistency model for conflict detection; selecting the shortest associated path with a logical consistency score higher than a set threshold as the effective associated path set; clustering the effective associated path set; using a distance metric function based on path length and edge type similarity, and employing a hierarchical clustering algorithm to divide the shortest associated paths into semantic clusters; and generating an associated path analysis report.

2. The method according to claim 1, characterized in that, The normalized value of node degree is obtained by dividing the actual degree of the standard node by the maximum degree value in the knowledge graph. The normalized value of betweenness centrality is obtained by calculating the proportion of the number of shortest paths passing through the standard node to the total number of shortest paths in the entire graph. The normalized value of the average importance of neighboring nodes is obtained by dividing the average degree of all first-order neighboring nodes of the standard node by the average degree of the entire graph.

3. The method according to claim 2, characterized in that, In the iterative propagation of topological potential energy, the potential energy contribution received by the standard node is equal to the sum of the potential energy values ​​of all neighboring nodes in the previous round multiplied by the edge weights, multiplied by the attenuation coefficient, and finally added to the initial topological potential energy value of the standard node multiplied by the retention coefficient to obtain the new potential energy value for the current round.

4. The method according to claim 3, characterized in that, The edge weight is determined by the weighted average of the edge type base weight and the historical call frequency. The base weight of the reference relationship edge is 0.8, the base weight of the semantic dependency edge is 0.6, and the edge weight increases by 0.05 for every 100 increases in historical call frequency.

5. The method according to claim 4, characterized in that, When calculating the potential energy distribution concentration coefficient, the sum of the potential energy of the top 20% of standard nodes is divided by the total potential energy of the entire graph, then 0.2 is subtracted, and then divided by 0.

8. The concentration threshold is 0.

6.

6. The method according to claim 5, characterized in that, The standard dissemination influence assessment module simulates the diffusion of standard knowledge through restricted random walks. Starting from the standard node to be evaluated, the next hop node is selected at each step according to the probability distribution after the outgoing edge weights are normalized. The walk step size follows a truncated geometric distribution. After performing a set number of walks, the cumulative number of visits to each standard node is counted, and the normalized value of the number of visits is used as the dissemination influence score.

7. The method according to claim 6, characterized in that, The restricted random walk path sets the maximum step size to 8 hops and the restart probability to 0.

15. After running 10,000 walks, nodes with a visit frequency greater than 500 times are extracted to form a high-frequency path skeleton.

8. The method according to claim 7, characterized in that, The bidirectional breadth-first search algorithm initiates breadth-first traversal from both the starting node and the target node. When the visited node sets in both directions intersect, the two paths are merged to form a complete shortest path. The search depth is limited to 4 layers. The algorithm records the type label of each edge on the shortest path and the difference between the standard publication year of the first and last nodes and the number of standard level conversions.

9. The method according to claim 8, characterized in that, The input layer of the path semantic consistency model receives the node sequence embedding, edge type sequence embedding, and attribute change vector of the path. The encoding layer uses a four-layer graph attention convolution module. The fusion layer concatenates the node representations output by the graph attention convolution module through global average pooling and max pooling, and then reduces the dimensionality through two layers of fully connected network. The output layer is a single neuron sigmoid activation that outputs a logical consistency score.

10. The method according to claim 9, characterized in that, The number of attention heads in the graph attention convolution module is dynamically adjusted based on three parameters: the normalized value of the average path length, the normalized value of the number of unique edge types in the path, and the normalized value of the standard year span involved in the path. The adjustment formula is 8 multiplied by the normalized value of the average path length to the power of 0.4, then multiplied by the normalized value of the number of unique edge types in the path to the power of 0.3, then multiplied by the normalized value of the standard year span involved in the path to the power of 0.3, and the result is rounded up and does not exceed 16.

Images