Intelligent matching method and system for multi-dimensional data

Abstract

The application discloses a kind of multi-dimensional data intelligent matching method and system, including the following steps: collection multi-source heterogeneous data set, and execute pre-processing;Perform feature extraction, and calculate the correlation strength between features, generate feature weight matrix;Calculate structural similarity, semantic similarity and statistical similarity, and execute weighted fusion according to feature weight matrix;Nonlinear resonance optimization is executed, and stable semantic resonance spectrum is obtained;Establish reinforcement learning matching strategy network, and train matching strategy network;Match strategy network is used after training in stable semantic resonance spectrum Dynamic matching decision is executed, and matching result set is generated.The application fuses semantic resonance spectrum and reinforcement learning method, realizes the intelligent matching of multi-dimensional heterogeneous data, with the advantages of high precision, strong self-adaptation and stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent matching, and in particular to an intelligent matching method and system for multi-dimensional data. Background Technology

[0002] Currently, matching techniques for multi-source heterogeneous data mainly rely on weighted fusion methods based on keyword similarity, semantic similarity, or statistical features. These methods achieve data association by calculating the similarity between different feature dimensions. However, these methods typically use static weights or fixed rules for matching, lacking the ability to dynamically model complex semantic relationships. They are also unable to handle semantic bias and feature noise between structured and unstructured data, resulting in insufficient matching accuracy and robustness in high-dimensional data environments.

[0003] Some existing technologies have introduced deep learning or graph neural network models to enhance semantic expression capabilities, but problems such as uninterpretable network training processes, single matching strategies, and rigid parameter update mechanisms still exist, making it impossible to achieve self-learning and self-correction of matching results. Summary of the Invention

[0004] One objective of this invention is to propose an intelligent matching method and system for multi-dimensional data. This invention integrates semantic resonance graphs and reinforcement learning methods to achieve intelligent matching of multi-dimensional heterogeneous data, and has the advantages of high accuracy, strong adaptability and excellent stability.

[0005] An intelligent matching method for multi-dimensional data according to an embodiment of the present invention includes the following steps:

[0006] Collect multi-source heterogeneous data sets and perform preprocessing to generate standardized data sets;

[0007] Feature extraction is performed on the standardized dataset to form a multidimensional feature vector set, and the correlation strength between features is calculated. A feature weight matrix is ​​generated based on the correlation strength.

[0008] Calculate structural similarity, semantic similarity, and statistical similarity, and perform weighted fusion based on the feature weight matrix to generate a preliminary fusion set;

[0009] Nonlinear resonance optimization is performed on the initial fusion set to construct a semantic resonance graph, establish energy mapping relationship and calculate semantic resonance intensity, dynamically adjust edge weights, and obtain a stable semantic resonance graph;

[0010] A reinforcement learning matching policy network is established. The multidimensional feature vector set is defined as the state space, and the action space and reward function are defined. The feature weight matrix is ​​set as the learning parameter, and the matching policy network is trained.

[0011] The trained matching policy network is used to perform dynamic matching decisions on a stable semantic resonance graph to generate a set of matching results.

[0012] Optionally, the multi-source heterogeneous data set includes structured data, semi-structured data, and unstructured data, and the preprocessing includes character format unification, time format normalization, encoding standardization, and noise removal.

[0013] Optionally, the generation of the feature weight matrix specifically includes:

[0014] Each record in the standardized dataset is represented as an input vector, and all input vectors form the input sample matrix.

[0015] The input sample matrix is ​​processed by feature extraction. The character field, time field, encoding field and text field are classified according to the field type, and character feature vector, time feature vector, encoding feature vector and text feature vector are generated.

[0016] The character feature vector, time feature vector, encoding feature vector and text feature vector are concatenated in the field order to form a multidimensional feature vector of a single sample. The multidimensional feature vectors of all samples constitute a multidimensional feature vector set.

[0017] Calculate the correlation strength between any two feature vectors in a multidimensional feature vector set;

[0018] The correlation matrix is ​​formed by combining the correlation strengths between all pairs of feature vectors.

[0019] The correlation matrix is ​​normalized to obtain the feature weight matrix.

[0020] Optionally, the correlation strength can be obtained by calculating the ratio of the inner product of two eigenvectors to the product of their magnitudes.

[0021] Optionally, the generation of the initial fusion set specifically includes:

[0022] Establish a mapping relationship between the weight coefficients in the feature weight matrix and the character feature vector, time feature vector, encoding feature vector and text feature vector in the multidimensional feature vector set;

[0023] Structural similarity is calculated based on character feature vectors to form a structural similarity matrix;

[0024] Semantic similarity calculation is performed based on text feature vectors, semantic space mapping is performed on text layer features, the semantic inclusion ratio between different text feature vectors is calculated, and a semantic similarity matrix is ​​generated.

[0025] Statistical similarity calculation is performed based on character feature vectors and encoded feature vectors. The co-occurrence frequency of features is statistically analyzed to construct a character co-occurrence distribution matrix. The joint occurrence probability is calculated and statistical similarity values ​​are obtained. A statistical similarity matrix is ​​formed according to feature index.

[0026] The structural similarity matrix, semantic similarity matrix, and statistical similarity matrix are weighted and fused according to the weight coefficients of the feature weight matrix to generate a fused similarity matrix;

[0027] The fusion similarity matrix is ​​normalized to form a normalized fusion similarity matrix;

[0028] In the normalized fusion similarity matrix, a similarity threshold is set. When the similarity value is greater than the similarity threshold, it is judged as a high matching pair, and when the similarity value is less than the similarity threshold, it is judged as a low matching pair.

[0029] Extract the feature vector indices and their fusion similarity values ​​contained in all high-matching pairs, form a feature matching mapping set in index order, and use this feature matching mapping set as the initial fusion set.

[0030] Optionally, the generation of the stable semantic resonance map specifically includes:

[0031] Feature indices and fusion similarity values ​​are extracted from the initial fusion set. The multi-dimensional feature vector corresponding to each feature index is set as a feature node, and the fusion similarity value is used as the edge weight between feature nodes to construct the initial semantic resonance graph structure.

[0032] In the initial semantic resonance graph, an energy mapping relationship is established based on the fusion similarity distribution and semantic association strength between feature nodes. The edge weights are mapped to energy transfer parameters. Energy aggregation calculation is performed on all node connection relationships to generate a global energy distribution vector.

[0033] The semantic resonance intensity value of each feature node is calculated based on the global energy distribution vector, and a set of node resonance intensities is generated.

[0034] In the semantic resonance graph structure, nonlinear resonance adjustment is performed on any adjacent feature nodes. When the difference in semantic resonance intensity between adjacent nodes exceeds a preset threshold, the corresponding edge weights are dynamically adjusted.

[0035] Global normalization is performed on the semantic resonance graph after nonlinear resonance adjustment to unify all edge weights to the standardized interval, and the global energy change rate is calculated based on the normalization result to generate the energy balance matrix.

[0036] The semantic responsibility values ​​of the feature nodes are calculated based on the energy balance matrix, and a set of node response vectors is generated.

[0037] An adaptive energy iteration process is performed based on the set of node response vectors. In each iteration, the node response vectors and edge weights are updated synchronously to continuously adjust the energy distribution state of the semantic resonance graph, so that the global energy gradually tends to stabilize and eventually converges to the minimum energy state.

[0038] Once the energy convergence is complete, a stable semantic resonance spectrum is generated.

[0039] Optionally, the training matching policy network specifically includes:

[0040] A reinforcement learning matching policy network is established by defining the multidimensional feature vector set as the state space, defining the action space and the reward function to form the structure of the reinforcement learning matching policy network.

[0041] The parameters of the reinforcement learning matching policy network are initialized using the feature weight matrix, and the values ​​of each element in the feature weight matrix are set as learning parameters.

[0042] The multidimensional feature vector set is input into the reinforcement learning matching policy network. Action selection and matching prediction are performed for each state. The matching confidence value corresponding to the current state and the selected action is calculated. The matching confidence value and the value evaluation result are differentially calculated to obtain the immediate reward difference. The immediate reward difference is multiplied by the learning rate to obtain the parameter update amount.

[0043] The feature weight matrix is ​​corrected based on the parameter update amount by adding each weight parameter with its corresponding update amount to generate the updated feature weight matrix.

[0044] The updated feature weight matrix is ​​written back into the reinforcement learning matching policy network;

[0045] Multiple rounds of training iterations are performed on the reinforcement learning matching policy network to update the network parameters, and the difference in matching confidence between adjacent rounds is calculated as the network convergence index.

[0046] When the network convergence index is less than the preset convergence threshold, the training of the matching policy network is stopped, a stable matching policy network is output, and the training process is completed.

[0047] Optionally, the reinforcement learning matching policy network adopts a dual-network structure based on the Actor-Critic architecture, including a policy sub-network and an evaluation sub-network. The policy sub-network takes a multi-dimensional feature vector set as input and outputs a matching action probability distribution. The evaluation sub-network takes matching confidence and semantic consistency as input and calculates the value evaluation result.

[0048] Optionally, the generation of the matching result set specifically includes:

[0049] The trained reinforcement learning matching policy network is loaded into the inference environment, and the stable semantic resonance graph is used as the input structure.

[0050] For each feature node in the stable semantic resonance graph, read the corresponding multidimensional feature vector and node response vector, concatenate the multidimensional feature vector and node response vector to form a node composite state vector, and input it into the policy sub-network of the reinforcement learning matching policy network to output the matching confidence value.

[0051] Receive the matching confidence value output by the policy sub-network, read the corresponding weight from the feature weight matrix, perform a weighted operation on the two, and generate a set of state matching scores using the weight as multiplicative coefficients;

[0052] Normalize the state matching score set to form a standardized matching score set;

[0053] Based on the standardized matching score set, the feature nodes in the semantic resonance graph are matched. When the standardized matching score of a node is higher than the preset matching score threshold, it is determined to be a valid matching node, and its corresponding matching node index is recorded.

[0054] Combine all valid matching nodes and their corresponding matching node indices to generate a set of matching results.

[0055] An intelligent matching system for multi-dimensional data according to an embodiment of the present invention includes:

[0056] The multi-source data acquisition module is used to acquire heterogeneous data from multiple sources and perform preprocessing.

[0057] The feature extraction module is used to extract features from a standardized dataset, form a multidimensional feature vector set, and generate a feature weight matrix.

[0058] The similarity fusion module is used to calculate structural similarity, semantic similarity, and statistical similarity, and performs weighted fusion based on the feature weight matrix to generate a preliminary fusion set;

[0059] The semantic resonance construction module is used to perform nonlinear resonance optimization on the initial fusion set to construct and stabilize the semantic resonance graph.

[0060] The reinforcement learning matching policy module is used to build and train the reinforcement learning matching policy network;

[0061] The dynamic matching decision module is used to perform dynamic matching decisions on a stable semantic resonance graph using the trained matching policy network, and generate a set of matching results.

[0062] The beneficial effects of this invention are:

[0063] This invention achieves deep fusion and dynamic optimization of multi-source heterogeneous data at the structural, semantic, and statistical layers by constructing a semantic resonance graph and a reinforcement learning matching strategy network. It overcomes the shortcomings of existing technologies, such as fixed weights, rigid matching strategies, and lack of adaptive semantic modeling capabilities. By introducing a dynamic learning mechanism for feature weight matrices into the method, the system can automatically allocate matching weights according to the correlation strength between features, achieving adaptive balance of features in multi-dimensional space. This ensures that structured data, semi-structured data, and unstructured data have a unified expression standard and a basis for mutual correlation during the matching process.

[0064] At the semantic level, this invention establishes a semantic resonance graph through a nonlinear resonance optimization process, mapping the semantic relationships between feature nodes into an energy propagation mechanism. By utilizing the energy mapping relationship and semantic resonance intensity calculation, dynamic energy transfer and balance between nodes are achieved, enabling semantic information to adaptively diffuse and stabilize globally, thereby improving the accuracy and consistency of semantic matching. The introduction of the energy balance matrix further enables global monitoring of the semantic resonance state. Through an adaptive energy iteration process, the node response vector and edge weight distribution are continuously corrected, allowing the system energy in the graph structure to gradually stabilize and converge to the optimal state.

[0065] In the matching decision layer, this invention innovatively introduces a reinforcement learning matching policy network based on the Actor-Critic architecture. The multi-dimensional feature vector set is defined as the state space, and the reward function is constructed with matching confidence and semantic consistency. This enables the matching policy to have self-learning and self-correction capabilities. Through multiple rounds of training iteration, the policy sub-network and the evaluation sub-network achieve synchronous optimization of parameters, enabling the system to dynamically adjust the matching behavior according to changes in the semantic environment and achieve continuous evolution of the matching policy. Attached Figure Description

[0066] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0067] Figure 1 This is a flowchart of an intelligent matching method for multi-dimensional data proposed in this invention;

[0068] Figure 2 This is a schematic diagram illustrating the dynamic matching decision-making process of the intelligent matching method for multi-dimensional data proposed in this invention. Detailed Implementation

[0069] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0070] refer to Figure 1 and Figure 2 A method for intelligent matching of multi-dimensional data includes the following steps:

[0071] Collect multi-source heterogeneous data sets and perform preprocessing to generate standardized data sets;

[0072] Feature extraction is performed on the standardized dataset to form a multidimensional feature vector set, and the correlation strength between features is calculated. A feature weight matrix is ​​generated based on the correlation strength.

[0073] Calculate structural similarity, semantic similarity, and statistical similarity, and perform weighted fusion based on the feature weight matrix to generate a preliminary fusion set;

[0074] Nonlinear resonance optimization is performed on the initial fusion set to construct a semantic resonance graph, establish energy mapping relationship and calculate semantic resonance intensity, dynamically adjust edge weights, and obtain a stable semantic resonance graph;

[0075] A reinforcement learning matching policy network is established. The multidimensional feature vector set is defined as the state space, and the action space and reward function are defined. The feature weight matrix is ​​set as the learning parameter, and the matching policy network is trained.

[0076] The trained matching policy network is used to perform dynamic matching decisions on a stable semantic resonance graph to generate a set of matching results.

[0077] In this embodiment, the multi-source heterogeneous data set includes structured data, semi-structured data, and unstructured data, and the preprocessing includes character format unification, time format normalization, encoding standardization, and noise removal.

[0078] In this embodiment, the generation of the feature weight matrix specifically includes:

[0079] Each record in the standardized dataset is represented as an input vector. Each record contains multiple field values, and each field value is used as an input element. All input vectors form the input sample matrix.

[0080] Feature extraction processing is performed on the input sample matrix. Character fields, time fields, encoding fields, and text fields are classified according to field type to generate character feature vectors, time feature vectors, encoding feature vectors, and text feature vectors. Specifically, character fields are vectorized into fixed-dimensional character feature vectors; time fields are decomposed into periodic components, time interval components, and time index components; encoding fields are embedded into dense vectors in continuous space; and text fields are semantically segmented and word vectorized to generate text feature vectors.

[0081] The character feature vector, time feature vector, encoding feature vector and text feature vector are concatenated in the field order to form a multidimensional feature vector of a single sample. The multidimensional feature vectors of all samples constitute a multidimensional feature vector set.

[0082] Calculate the correlation strength between any two feature vectors in a multidimensional feature vector set;

[0083] The correlation matrix is ​​formed by combining the correlation strengths between all pairs of feature vectors.

[0084] The correlation matrix is ​​normalized to obtain the feature weight matrix.

[0085] In this embodiment, the correlation strength is obtained by calculating the ratio of the inner product of two feature vectors to the product of their magnitudes, and is used to characterize the similarity between features.

[0086] In this embodiment, the generation of the preliminary fusion set specifically includes:

[0087] Establish a mapping relationship between the weight coefficients in the feature weight matrix and the character feature vector, time feature vector, encoding feature vector and text feature vector in the multidimensional feature vector set;

[0088] Structural similarity is calculated based on character feature vectors to form a structural similarity matrix;

[0089] The generation of the structural similarity matrix specifically includes: extracting all character feature vectors and combining them pairwise to form a feature pair set; performing character sequence alignment for each feature pair, scanning the sequence elements of the two character feature vectors character by character, and recording the number of insertion, deletion, or replacement operations when mismatched characters are found during the alignment process, generating insertion count, deletion count, and replacement count values ​​respectively; summing the three operation count values ​​to form the edit distance value of the corresponding feature pair, which is used to represent the degree of difference between the two character feature vectors at the character structure layer; further normalizing the edit distance values ​​of all feature pairs to obtain the structural similarity value of each feature pair, which is used to characterize the matching strength of the corresponding character feature pair at the structure layer; and arranging all structural similarity values ​​in the order of the character feature vector pairs' indices to generate a structural similarity matrix.

[0090] Semantic similarity calculation is performed based on text feature vectors, semantic space mapping is performed on text layer features, the semantic inclusion ratio between different text feature vectors is calculated, and a semantic similarity matrix is ​​generated.

[0091] The generation of the semantic similarity matrix specifically includes: performing semantic space mapping based on text feature vectors, mapping each text feature vector to a high-dimensional semantic representation vector; performing semantic inclusion degree operation on the high-dimensional semantic representation vectors between any two text feature vectors, calculating the semantic inclusion ratio based on the overlap of the two semantic representation vectors, and obtaining the semantic inclusion ratio value of each pair of text feature vectors; then arranging all semantic inclusion ratio values ​​according to the index order of the text feature vectors to construct the row and column index structure of the semantic similarity matrix; and filling the corresponding positions with the semantic inclusion ratio values ​​as matrix elements to generate the semantic similarity matrix.

[0092] Statistical similarity calculation is performed based on character feature vectors and encoded feature vectors. The co-occurrence frequency of features is statistically analyzed to construct a character co-occurrence distribution matrix. The joint occurrence probability is calculated and statistical similarity values ​​are obtained. A statistical similarity matrix is ​​formed according to feature index.

[0093] The generation of the statistical similarity matrix specifically includes: performing joint analysis on character feature vectors and encoded feature vectors, calculating the co-occurrence frequency of each character feature and encoded feature in the multidimensional feature vector set, and generating a co-occurrence frequency statistics table; constructing a character co-occurrence distribution matrix based on the co-occurrence frequency statistics table, mapping the co-occurrence frequency value of each character feature and encoded feature to matrix elements, forming a complete co-occurrence relationship representation; then calculating the joint occurrence probability value based on the occurrence frequency of each element in the character co-occurrence distribution matrix, which reflects the probability relationship of character features and encoded features co-occurring in the dataset; subsequently calculating the statistical similarity value between each character feature and encoded feature based on the joint occurrence probability value, generating a statistical similarity vector set; arranging the statistical similarity vector set according to the index order of character feature vectors and encoded feature vectors, and filling it into the corresponding positions of the statistical similarity matrix to form the statistical similarity matrix;

[0094] The structural similarity matrix, semantic similarity matrix, and statistical similarity matrix are weighted and fused according to the weight coefficients of the feature weight matrix to generate a fused similarity matrix, which is used to comprehensively reflect the fusion relationship of the three types of features: structural, semantic, and statistical.

[0095] The fusion similarity matrix is ​​normalized to form a normalized fusion similarity matrix;

[0096] In the normalized fusion similarity matrix, a similarity threshold is set. When the similarity value is greater than the similarity threshold, it is judged as a high matching pair, and when the similarity value is less than the similarity threshold, it is judged as a low matching pair.

[0097] Extract the feature vector indices and their fusion similarity values ​​contained in all high-matching pairs, form a feature matching mapping set in index order, and use this feature matching mapping set as the initial fusion set.

[0098] In this embodiment, the generation of the stable semantic resonance spectrum specifically includes:

[0099] Feature indices and fusion similarity values ​​are extracted from the initial fusion set. The multidimensional feature vector corresponding to each feature index is set as a feature node, and the fusion similarity value is used as the edge weight between feature nodes to construct an initial semantic resonance graph structure to represent the fusion similarity relationship between multidimensional feature vectors.

[0100] In the initial semantic resonance graph, an energy mapping relationship is established based on the fusion similarity distribution and semantic association strength between feature nodes. The edge weights are mapped to energy transfer parameters. Energy aggregation calculation is performed on all node connection relationships to generate a global energy distribution vector, which is used to characterize the overall energy state of the semantic resonance graph.

[0101] The generation of the global energy distribution vector specifically includes: traversing all edge weights in the semantic resonance graph, extracting the fusion similarity value of each edge and the semantic association strength value of the two feature nodes it connects, multiplying the two values ​​to calculate the energy mapping value, and establishing a correspondence between the energy mapping value and the corresponding feature node index to generate an energy mapping lookup table; performing cumulative calculation on all energy mapping values ​​in the energy mapping lookup table to obtain the energy transfer value of each edge, which is used to represent the energy exchange strength between the feature nodes connected by the edge, forming an energy transfer set; then performing weighted summation on all energy transfer values ​​involved in each feature node in the energy transfer set, adding the energy transfer values ​​of the adjacent edges of each node according to the edge weights to obtain the energy summary value of each feature node, generating a node energy summary vector; performing standardization and normalization on the node energy summary vector; and arranging the normalized node energy summary vector according to the feature node index order to generate the global energy distribution vector.

[0102] The semantic resonance intensity value of each feature node is calculated based on the global energy distribution vector. This value is used to describe the response capability and semantic influence of the feature node during semantic propagation and to generate a set of node resonance intensities.

[0103] The generation of the node resonance intensity set specifically includes: traversing the global energy distribution vector, extracting the energy summary value corresponding to each feature node as the node's basic energy input value; calculating the difference between the node and the energy summary values ​​of all its neighboring feature nodes, and performing a weighted average of all differences according to the corresponding edge weights to obtain the node's energy fluctuation value; performing a weighted summation of the basic energy input value and the corresponding energy fluctuation value of each feature node, and using the weighted result as the node's semantic response energy; then normalizing the semantic response energy to obtain the semantic resonance intensity value of each feature node; and arranging the semantic resonance intensity values ​​of all feature nodes in node index order to generate the node resonance intensity set.

[0104] In the semantic resonance graph structure, nonlinear resonance adjustment is performed on any adjacent feature nodes. When the difference in semantic resonance intensity between adjacent nodes exceeds a preset threshold, the corresponding edge weights are dynamically adjusted so that the energy of high-intensity nodes diffuses to low-intensity nodes, thereby achieving local energy balance in the semantic resonance graph.

[0105] The dynamic adjustment of corresponding edge weights specifically includes: traversing all adjacent feature node pairs in the semantic resonance graph, extracting the semantic resonance intensity value and corresponding edge weight of each pair of nodes, calculating the semantic resonance intensity difference between the two nodes, and obtaining the semantic intensity difference; comparing the semantic intensity difference with a preset threshold, keeping the current edge weight unchanged when the difference is less than or equal to the threshold, and performing dynamic energy diffusion adjustment when the difference is greater than the threshold; during the adjustment, calculating the energy transfer coefficient according to the ratio of the semantic intensity difference to the threshold, multiplying the semantic resonance intensity value of the high-intensity node by the energy transfer coefficient to obtain the energy transfer amount, and dividing the energy transfer amount by the sum of the current edge weights to obtain the unit energy transfer gain value; then updating the edge weights based on the unit energy transfer gain value, subtracting the energy transfer gain value multiplied by the energy transfer coefficient from the original edge weight as the new edge weight, and simultaneously increasing the semantic resonance intensity value of the low-intensity node by the same amount of energy transfer amount to achieve numerical balance of energy between the two nodes;

[0106] Global normalization is performed on the semantic resonance graph after nonlinear resonance adjustment to unify all edge weights to the standardized interval, and the global energy change rate is calculated based on the normalization result to generate an energy balance matrix, which is used to reflect the overall energy stability of the semantic resonance graph.

[0107] The generation of the energy balance matrix specifically includes: traversing the normalized semantic resonance intensity, extracting the semantic resonance intensity value of each feature node at the current time and the previous time, calculating the difference between the two and taking the absolute value to obtain the set of energy change amplitudes for each node; performing a weighted average operation on the set of energy change amplitudes, using the weight coefficient of each node in the feature weight matrix as the weighting coefficient to obtain the global energy change rate, which is used to reflect the overall energy change degree of the semantic resonance graph; based on this, using the feature node index as the row and the adjacent node index as the column, calculating the difference in energy change amplitude between any two nodes, and filling each difference into the corresponding position in the matrix according to the node index position to generate the energy balance matrix;

[0108] The semantic responsibility values ​​of the feature nodes are calculated based on the energy balance matrix, and a set of node response vectors is generated.

[0109] The generation of the node response vector set specifically includes: traversing the row vectors corresponding to each feature node in the energy balance matrix, extracting the energy change rate values ​​between the node and all adjacent feature nodes, performing weighted summation on the rate values, using the corresponding edge weights as weighting coefficients to obtain the total energy response value of the node; calculating the difference between the total energy response value of the node and the global energy change rate to obtain the node semantic offset; then performing normalization on the node semantic offset to obtain the semantic responsiveness value, reflecting the relative strength of the node's semantic response; subsequently, arranging the semantic responsiveness values ​​of all feature nodes in node index order to generate a node response vector set, which is used to characterize the degree of response of each feature node to energy changes in the semantic resonance graph;

[0110] An adaptive energy iteration process is performed based on the set of node response vectors. In each iteration, the node response vectors and edge weights are updated synchronously to continuously adjust the energy distribution state of the semantic resonance graph, gradually stabilizing the global energy and eventually converging to the minimum energy state. Specifically, the response value of each node in the set of node response vectors is used as the initial energy input of the semantic resonance graph. All node response values ​​are arranged in node index order and stored in the energy distribution table to form the initial energy distribution state. All adjacent feature nodes are traversed, and the difference in response values ​​between each pair of nodes is extracted. The absolute value of the difference is taken as the energy difference amplitude and compared with a preset threshold. If the energy difference amplitude is greater than the threshold, the energy difference amplitude is multiplied by the proportional coefficient of the corresponding edge weight to obtain the energy transfer amount. The energy transfer amount is subtracted from the energy value of the high-response node and added to the energy value of the low-response node by the same amount. The process involves incrementally generating a node energy update table. A weighted average is calculated for the updated values ​​of all nodes in the table, using edge weights as a weighting factor to smooth the energy update table and obtain a new set of node responsivity. Next, the difference between the new set of node responsivity and the previous set is calculated, and the absolute value of the difference is taken and averaged to obtain the global energy difference. If the global energy difference is greater than the convergence threshold, all edge weights are synchronously updated by multiplying each edge weight by a scaling factor, decreasing the weights of edges with larger energy transfer and increasing the weights of edges with smaller energy transfer, resulting in a new set of edge weights. The new set of node responsivity and the new set of edge weights are used as input for the next iteration, and the above steps are repeated until the global energy difference is less than the convergence threshold. The final semantic resonance graph energy distribution result is then output, achieving adaptive balance and convergence of system energy in the graph structure.

[0111] Once the energy convergence is complete, a stable semantic resonance spectrum is generated.

[0112] In this embodiment, the training matching policy network specifically includes:

[0113] A reinforcement learning matching policy network is established, and a multi-dimensional feature vector set is defined as the state space. Each state in the state space corresponds to a multi-dimensional feature vector. An action space and a reward function are defined to form the structure of the reinforcement learning matching policy network.

[0114] The reward function is obtained by weighted calculation of matching confidence and semantic consistency;

[0115] The action space consists of a set of matching behaviors that the reinforcement learning matching policy network can execute, including confirming matching, rejecting matching, and fuzzy matching. These behaviors guide the network to make dynamic matching decisions under different states of the multidimensional feature vector set. Confirming matching is used to confirm the valid correspondence between feature pairs, rejecting matching is used to eliminate invalid feature pairs, and fuzzy matching is used to postpone the decision and record context information.

[0116] The parameters of the reinforcement learning matching policy network are initialized using the feature weight matrix, and the values ​​of each element in the feature weight matrix are set as learning parameters.

[0117] A multidimensional feature vector set is input into a reinforcement learning matching policy network. For each state, action selection and matching prediction are performed. The matching confidence value between the current state and the selected action is calculated. The matching confidence value and the value evaluation result are then subtracted to obtain the immediate reward difference. This immediate reward difference is multiplied by the learning rate to obtain the parameter update amount. Specifically, the multidimensional feature vector set is input into the policy sub-network. For each action in the action space, the execution probability distribution is calculated, and the action with the highest probability is selected for matching prediction, outputting the matching confidence value for the corresponding state. The matching confidence value and the corresponding semantic consistency value are input into the evaluation sub-network. The value evaluation result corresponding to the state and action is calculated based on the reward function. The matching confidence value output by the policy sub-network is subtracted from the value evaluation result output by the evaluation sub-network to obtain the immediate reward difference. This immediate reward difference is multiplied by the learning rate to generate the parameter update amount.

[0118] The generation of the semantic consistency value specifically includes: extracting the character feature vector, time feature vector, encoding feature vector, and text feature vector corresponding to the current state from a multi-dimensional feature vector set, and combining the four types of feature vectors into a composite semantic feature vector; calculating the semantic matching ratio between the composite semantic feature vector and the target feature vector, wherein the semantic matching ratio is obtained by statistically analyzing the proportion of the number of matching feature components in the four types of features to the total number of feature components; reading the weight coefficients corresponding to the current state from the feature weight matrix, performing a weighted multiplication operation on the semantic matching ratio to generate a weighted semantic relevance value; then performing a difference calculation between the weighted semantic relevance value and a preset semantic threshold to obtain the semantic deviation; then performing a linear normalization mapping on the semantic deviation, limiting the result to the range of 0 to 1, with the result closer to 1 indicating higher semantic consistency; and using the normalized result as the semantic consistency value, wherein the target feature vector is the feature vector corresponding to the feature node with the highest semantic resonance intensity with the current state node in the semantic resonance graph, and this feature vector is used as a reference vector for the reinforcement learning matching policy network to perform semantic consistency calculation and value evaluation.

[0119] The feature weight matrix is ​​corrected based on the parameter update amount by adding each weight parameter with its corresponding update amount to generate the updated feature weight matrix.

[0120] The updated feature weight matrix is ​​written back into the reinforcement learning matching policy network;

[0121] Multiple training iterations are performed on the reinforcement learning matching policy network to update the network parameters, and the difference in matching confidence between adjacent rounds is calculated as the network convergence index.

[0122] When the network convergence index is less than the preset convergence threshold, the training of the matching policy network is stopped, a stable matching policy network is output, and the training process is completed.

[0123] In this embodiment, the reinforcement learning matching policy network adopts a dual-network structure based on the Actor-Critic architecture, including a policy sub-network and an evaluation sub-network. The policy sub-network takes a multi-dimensional feature vector set as input and outputs a matching action probability distribution. The evaluation sub-network takes matching confidence and semantic consistency as input and calculates the value evaluation result. The two sub-networks share a feature weight matrix and achieve synchronous parameter updates through reward function feedback, which is used to dynamically optimize the matching policy.

[0124] In this embodiment, the generation of the matching result set specifically includes:

[0125] The trained reinforcement learning matching policy network is loaded into the inference environment, and the stable semantic resonance graph is used as the input structure.

[0126] For each feature node in the stable semantic resonance graph, read the corresponding multidimensional feature vector and node response vector, concatenate the multidimensional feature vector and node response vector to form a node composite state vector, and input it into the policy sub-network of the reinforcement learning matching policy network to output the matching confidence value.

[0127] Receive the matching confidence value output by the policy sub-network, read the corresponding weight from the feature weight matrix, perform a weighted operation on the two, and generate a set of state matching scores using the weight as multiplicative coefficients;

[0128] Normalize the state matching score set to form a standardized matching score set;

[0129] Based on the standardized matching score set, the feature nodes in the semantic resonance graph are matched. When the standardized matching score of a node is higher than the preset matching score threshold, it is determined to be a valid matching node, and its corresponding matching node index is recorded.

[0130] Combine all valid matching nodes and their corresponding matching node indices to generate a set of matching results.

[0131] An intelligent matching system for multi-dimensional data includes:

[0132] The multi-source data acquisition module is used to acquire heterogeneous data from multiple sources and perform preprocessing.

[0133] The feature extraction module is used to extract features from a standardized dataset, form a multidimensional feature vector set, and generate a feature weight matrix.

[0134] The similarity fusion module is used to calculate structural similarity, semantic similarity, and statistical similarity, and performs weighted fusion based on the feature weight matrix to generate a preliminary fusion set;

[0135] The semantic resonance construction module is used to perform nonlinear resonance optimization on the initial fusion set to construct and stabilize the semantic resonance graph.

[0136] The reinforcement learning matching policy module is used to build and train the reinforcement learning matching policy network;

[0137] The dynamic matching decision module is used to perform dynamic matching decisions on a stable semantic resonance graph using the trained matching policy network, and generate a set of matching results.

[0138] Example 1:

[0139] To verify the feasibility of this invention in practice, it was applied to the information integration process of heterogeneous data fusion and intelligent matching in a data management center. This data management center needs to process multi-source heterogeneous data from multiple departments such as government affairs, finance, healthcare, and transportation. These data have significant differences in source structure, encoding format, time record, and semantic expression. Traditional matching methods based on keywords or rule templates are difficult to ensure both accuracy and flexibility, often resulting in unstable matching results, low feature correlation, and inability to integrate cross-domain data, which seriously affects subsequent data analysis and intelligent decision-making applications. Therefore, the intelligent matching method for multi-dimensional data provided by this invention was deployed in the center's heterogeneous data integration system to achieve semantic fusion and adaptive matching between high-dimensional features.

[0140] In this application scenario, the system utilizes the multi-source data acquisition module proposed in this invention to uniformly collect and preprocess data provided by different departments. The original data with different formats are transformed into a standardized data set after character format unification, time format normalization, encoding standardization, and noise removal. Subsequently, the feature extraction module extracts character feature vectors, time feature vectors, encoding feature vectors, and text feature vectors, and generates a feature weight matrix based on the correlation strength between features. This enables different types of features to have adaptive weight allocation capabilities in fusion calculation, thereby ensuring the unified expression and quantification of heterogeneous features at the semantic layer.

[0141] In the data fusion stage, the system calculates structural similarity, semantic similarity, and statistical similarity based on the feature weight matrix, and performs weighted fusion to generate a preliminary fusion set. Unlike traditional methods that compare based on a single feature similarity, this invention achieves complementary enhancement between multi-level features by fusing matching information from multiple feature dimensions, making the matching judgment more robust and interpretable. Subsequently, the similarity fusion result is input into the semantic resonance construction module, and a semantic resonance graph is constructed through a nonlinear resonance optimization mechanism. In this process, an energy mapping relationship is established and the semantic resonance intensity is calculated, thereby dynamically adjusting the edge weights to form a stable semantic resonance graph with global energy balance. This graph can adaptively adjust the semantic propagation relationship between nodes between different semantic layers, avoiding the problems of incorrect clustering or association breakage caused by semantic drift in traditional static graph structures.

[0142] During the reinforcement learning phase, the system utilizes the reinforcement learning matching policy network of this invention to model the state space of the multidimensional feature vector set, and achieves dynamic decision optimization by defining the action space and reward function. The action space includes three behavioral forms: confirming matching, rejecting matching, and fuzzy matching, enabling the model to maintain self-learning and exploration capabilities in uncertain semantic regions. The reward function is generated by weighted calculation of matching confidence and semantic consistency, thereby balancing the accuracy and semantic rationality of the matching results. During multiple rounds of training iterations, the network parameters are continuously updated and converged, eventually forming a stable matching policy network, providing intelligent support for the dynamic decision-making phase.

[0143] In actual deployment, the trained matching strategy network is loaded into the dynamic matching decision module to perform dynamic matching inference on the stable semantic resonance graph. For each feature node, the system reads the corresponding multi-dimensional feature vector and node response vector, concatenates them into a composite state vector and inputs it into the network. The strategy sub-network outputs the matching confidence, and the evaluation sub-network generates the state matching score by combining the feature weight matrix. Finally, through normalization and threshold judgment, the effective matching nodes and their corresponding indices are automatically determined. After the complete process is run, the system can automatically output the matching result set, providing a reliable foundation for subsequent data mining, intelligent prediction and information decision-making.

[0144] To verify the performance of the present invention, it was compared with the traditional method. The comparison results are shown in Table 1.

[0145] Table 1. Performance Comparison of Multi-Dimensional Data Intelligent Matching Methods and Traditional Methods

[0146] Comparison indicators Traditional keyword matching methods Rule template-based method Method of the present invention Match accuracy (%) 82.4 87.1 93.8 Semantic consistency score (0~1) 0.68 0.73 0.94 Success rate of heterogeneous data fusion (%) 79.2 84.6 95.3 Matching response time (ms) 612 498 435 Stability index for dynamic data scenarios (0~1) 0.71 0.77 0.93 Feature weight adaptive adjustment capability (0~1) 0.32 0.46 0.89 Semantic drift suppression rate (%) 61.5 69.2 91.7 Overall system scalability score (0~1) 0.66 0.74 0.95

[0147] As can be seen from Table 1, the multi-dimensional data intelligent matching method proposed in this invention outperforms the traditional keyword matching method and rule template method in terms of overall performance.

[0148] In terms of matching accuracy, this invention achieved 93.8%, which is 11.4 percentage points higher than keyword matching methods and 6.7 percentage points higher than rule template methods. This improvement is mainly due to the introduction of feature weight matrix, which enables the system to comprehensively calculate structural similarity, semantic similarity and statistical similarity at the multi-dimensional feature level, avoiding the matching omission problem caused by traditional methods that only rely on surface characters or rule templates.

[0149] In terms of semantic consistency score, the present invention achieved a level of 0.94, which is more than 0.2 higher than the traditional method. This indicates that the system has a stronger expressive ability in cross-domain data semantic association recognition. In particular, when text and coded fields are fused, dynamic association at the context level is achieved through semantic resonance graph, thereby effectively capturing potential semantic relationships and improving the reliability of cross-modal matching.

[0150] In terms of the success rate of heterogeneous data fusion, the present invention achieves 95.3%, which is far higher than the 79.2% and 84.6% of traditional methods. This is mainly due to the energy balance iterative model under the semantic resonance mechanism, which enables the association between feature nodes to be automatically adjusted according to the semantic energy difference, thereby enhancing the model's adaptability when fusing unstructured and semi-structured data.

[0151] The optimization of response time is also remarkable. The average response time of this invention is 435ms, which is about 29% shorter than that of traditional keyword matching. This demonstrates the efficiency of reinforcement learning matching strategy network in dynamic decision-making process. Through the Actor-Critic dual network structure, the system can complete the matching action selection and parameter update with minimal computational cost, reduce redundant computation paths, and thus improve overall operating efficiency.

[0152] In terms of dynamic scene stability, the stability index of this invention is 0.93, which is about 0.2 higher than that of traditional methods. This indicates that this invention can still maintain the continuity and consistency of matching results when facing data flow fluctuations or new feature fields. This performance advantage comes from the adaptive energy adjustment mechanism of the semantic resonance graph, which can automatically redistribute weights when significant semantic shifts occur between nodes, and achieve balanced convergence of the semantic layer.

[0153] The adaptive adjustment capability of feature weights improved from 0.46 to 0.89, indicating that the system can still automatically learn and adjust the importance ratio of different features even when there are significant differences in feature dimensions. The reinforcement learning strategy network continuously corrects the feature weight matrix during training, enabling the model to self-optimize when faced with complex feature combinations, thereby enhancing generalization ability and robustness.

[0154] Furthermore, the semantic drift suppression rate is as high as 91.7%, which is significantly better than traditional methods. This means that when the semantics of the data change dynamically or some features are missing, the system can still maintain the correctness of the matching decision. This result comes from the combined effect of the semantic resonance mechanism and the reward function, which enables the model to self-correct under energy perturbation.

[0155] The system achieves a scalability score of 0.95, demonstrating high portability in multi-source, cross-modal data fusion scenarios. This invention, through modular design and adaptive matching strategy network, enables continuous and efficient operation without the need for remodeling when new data is accessed, feature dimensions are expanded, or semantic domains are migrated, requiring only local parameter adjustments.

[0156] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent matching method for multi-dimensional data, characterized in that, Includes the following steps: Collect multi-source heterogeneous data sets and perform preprocessing to generate standardized data sets; Feature extraction is performed on the standardized dataset to form a multidimensional feature vector set, and the correlation strength between features is calculated. A feature weight matrix is ​​generated based on the correlation strength. Calculate structural similarity, semantic similarity, and statistical similarity, and perform weighted fusion based on the feature weight matrix to generate a preliminary fusion set; Nonlinear resonance optimization is performed on the initial fusion set to construct a semantic resonance graph, establish energy mapping relationship and calculate semantic resonance intensity, dynamically adjust edge weights, and obtain a stable semantic resonance graph; A reinforcement learning matching policy network is established. The multidimensional feature vector set is defined as the state space, and the action space and reward function are defined. The feature weight matrix is ​​set as the learning parameter, and the matching policy network is trained. The trained matching policy network is used to perform dynamic matching decisions on a stable semantic resonance graph to generate a set of matching results. The weighted fusion is a weighted fusion of structural similarity, semantic similarity, and statistical similarity based on the weight coefficients in the feature weight matrix; The generation of the multidimensional feature vector set specifically includes: Each record in the standardized dataset is represented as an input vector, and all input vectors form the input sample matrix. The input sample matrix is ​​processed by feature extraction. The character field, time field, encoding field and text field are classified according to the field type, and character feature vector, time feature vector, encoding feature vector and text feature vector are generated. The character feature vector, time feature vector, encoding feature vector, and text feature vector are concatenated in the order of the fields to form a multidimensional feature vector for a single sample. The multidimensional feature vectors of all samples constitute a multidimensional feature vector set.

2. The intelligent matching method for multi-dimensional data according to claim 1, characterized in that, The multi-source heterogeneous data set includes structured data, semi-structured data, and unstructured data. The preprocessing includes character format unification, time format normalization, encoding standardization, and noise removal.

3. The intelligent matching method for multi-dimensional data according to claim 1, characterized in that, The generation of the feature weight matrix specifically includes: Calculate the correlation strength between any two feature vectors in a multidimensional feature vector set; The correlation matrix is ​​formed by combining the correlation strengths between all pairs of feature vectors. The correlation matrix is ​​normalized to obtain the feature weight matrix.

4. The intelligent matching method for multi-dimensional data according to claim 3, characterized in that, The correlation strength is obtained by calculating the ratio of the inner product of two eigenvectors to the product of their magnitudes.

5. The intelligent matching method for multi-dimensional data according to claim 1, characterized in that, The generation of the preliminary fusion set specifically includes: Establish a mapping relationship between the weight coefficients in the feature weight matrix and the character feature vector, time feature vector, encoding feature vector and text feature vector in the multidimensional feature vector set; Structural similarity is calculated based on character feature vectors to form a structural similarity matrix; Semantic similarity calculation is performed based on text feature vectors, semantic space mapping is performed on text layer features, the semantic inclusion ratio between different text feature vectors is calculated, and a semantic similarity matrix is ​​generated. Statistical similarity calculation is performed based on character feature vectors and encoded feature vectors. The co-occurrence frequency of features is statistically analyzed to construct a character co-occurrence distribution matrix. The joint occurrence probability is calculated and statistical similarity values ​​are obtained. A statistical similarity matrix is ​​formed according to feature index. The structural similarity matrix, semantic similarity matrix, and statistical similarity matrix are weighted and fused according to the weight coefficients of the feature weight matrix to generate a fused similarity matrix; The fusion similarity matrix is ​​normalized to form a normalized fusion similarity matrix; In the normalized fusion similarity matrix, a similarity threshold is set. When the similarity value is greater than the similarity threshold, it is judged as a high matching pair, and when the similarity value is less than the similarity threshold, it is judged as a low matching pair. Extract the feature vector indices and their fusion similarity values ​​contained in all high-matching pairs, form a feature matching mapping set in index order, and use this feature matching mapping set as the initial fusion set.

6. The intelligent matching method for multi-dimensional data according to claim 1, characterized in that, The generation of the stable semantic resonance map specifically includes: Feature indices and fusion similarity values ​​are extracted from the initial fusion set. The multi-dimensional feature vector corresponding to each feature index is set as a feature node, and the fusion similarity value is used as the edge weight between feature nodes to construct the initial semantic resonance graph structure. In the initial semantic resonance graph, an energy mapping relationship is established based on the fusion similarity distribution and semantic association strength between feature nodes. The edge weights are mapped to energy transfer parameters. Energy aggregation calculation is performed on all node connection relationships to generate a global energy distribution vector. The semantic resonance intensity value of each feature node is calculated based on the global energy distribution vector, and a set of node resonance intensities is generated. In the semantic resonance graph structure, nonlinear resonance adjustment is performed on any adjacent feature nodes. When the difference in semantic resonance intensity between adjacent nodes exceeds a preset threshold, the corresponding edge weights are dynamically adjusted. Global normalization is performed on the semantic resonance graph after nonlinear resonance adjustment to unify all edge weights to the standardized interval, and the global energy change rate is calculated based on the normalization result to generate the energy balance matrix. The semantic responsibility values ​​of the feature nodes are calculated based on the energy balance matrix, and a set of node response vectors is generated. An adaptive energy iteration process is performed based on the set of node response vectors. In each iteration, the node response vectors and edge weights are updated synchronously to continuously adjust the energy distribution state of the semantic resonance graph, so that the global energy gradually tends to stabilize and eventually converges to the minimum energy state. Once the energy convergence is complete, a stable semantic resonance spectrum is generated.

7. The intelligent matching method for multi-dimensional data according to claim 1, characterized in that, The training matching strategy network specifically includes: A reinforcement learning matching policy network is established by defining the multidimensional feature vector set as the state space, defining the action space and the reward function to form the structure of the reinforcement learning matching policy network. The parameters of the reinforcement learning matching policy network are initialized using the feature weight matrix, and the values ​​of each element in the feature weight matrix are set as learning parameters. The multidimensional feature vector set is input into the reinforcement learning matching policy network. Action selection and matching prediction are performed for each state. The matching confidence value corresponding to the current state and the selected action is calculated. The matching confidence value and the value evaluation result are differentially calculated to obtain the immediate reward difference. The immediate reward difference is multiplied by the learning rate to obtain the parameter update amount. The feature weight matrix is ​​corrected based on the parameter update amount by adding each weight parameter with its corresponding update amount to generate the updated feature weight matrix. The updated feature weight matrix is ​​written back into the reinforcement learning matching policy network; Multiple rounds of training iterations are performed on the reinforcement learning matching policy network to update the network parameters, and the difference in matching confidence between adjacent rounds is calculated as the network convergence index. When the network convergence index is less than the preset convergence threshold, the training of the matching policy network is stopped, a stable matching policy network is output, and the training process is completed.

8. The intelligent matching method for multi-dimensional data according to claim 7, characterized in that, The reinforcement learning matching policy network adopts a dual-network structure based on the Actor-Critic architecture, including a policy sub-network and an evaluation sub-network. The policy sub-network takes a multi-dimensional feature vector set as input and outputs the probability distribution of matching actions. The evaluation sub-network takes matching confidence and semantic consistency as input and calculates the value evaluation result.

9. The intelligent matching method for multi-dimensional data according to claim 1, characterized in that, The generation of the matching result set specifically includes: The trained reinforcement learning matching policy network is loaded into the inference environment, and the stable semantic resonance graph is used as the input structure. For each feature node in the stable semantic resonance graph, read the corresponding multidimensional feature vector and node response vector, concatenate the multidimensional feature vector and node response vector to form a node composite state vector, and input it into the policy sub-network of the reinforcement learning matching policy network to output the matching confidence value. Receive the matching confidence value output by the policy sub-network, read the corresponding weight from the feature weight matrix, perform a weighted operation on the two, and generate a set of state matching scores using the weight as multiplicative coefficients; Normalize the state matching score set to form a standardized matching score set; Based on the standardized matching score set, the feature nodes in the semantic resonance graph are matched. When the standardized matching score of a node is higher than the preset matching score threshold, it is determined to be a valid matching node, and its corresponding matching node index is recorded. Combine all valid matching nodes and their corresponding matching node indices to generate a set of matching results.

10. An intelligent matching system for multi-dimensional data, executing the intelligent matching method for multi-dimensional data according to any one of claims 1 to 9, characterized in that, include: The multi-source data acquisition module is used to acquire heterogeneous data from multiple sources and perform preprocessing. The feature extraction module is used to extract features from a standardized dataset, form a multidimensional feature vector set, and generate a feature weight matrix. The similarity fusion module is used to calculate structural similarity, semantic similarity, and statistical similarity, and performs weighted fusion based on the feature weight matrix to generate a preliminary fusion set; The semantic resonance construction module is used to perform nonlinear resonance optimization on the initial fusion set to construct and stabilize the semantic resonance graph. The reinforcement learning matching policy module is used to build and train the reinforcement learning matching policy network; The dynamic matching decision module is used to perform dynamic matching decisions on a stable semantic resonance graph using the trained matching policy network, and generate a set of matching results.

Images