A ship supply and demand resource matching method and device based on a causal knowledge graph

Abstract

The application discloses a ship supply and demand resource matching method and device based on a causal knowledge graph, and the method comprises the following steps: S1, collecting a multi-source data set of ship supply and demand resources; the multi-source data set comprises a structured data subset, a semi-structured data subset and an unstructured data subset; S2, performing causal knowledge graph construction processing on the multi-source data set to obtain a causal knowledge graph; and S3, processing resource demand scene information input by a user based on the causal knowledge graph to obtain resource recommendation result information.

Description

Technical Field

[0001] This invention relates to the fields of supply chain management technology, knowledge graph technology, and big data mining technology, specifically to a method and apparatus for matching ship supply and demand resources based on causal knowledge graphs. Background Technology

[0002] As the scale of the shipping supply chain continues to expand and the operating environment becomes increasingly complex, traditional resource matching and allocation methods struggle to effectively address the challenges of integrating multi-source heterogeneous data and fail to accurately characterize the dynamic causal evolution mechanism in the supply and demand relationship of ships. Existing recommendation models based on collaborative filtering or rule matching largely rely on historical interaction data, neglecting the causal driving relationships between key factors such as ship operation, maintenance events, and environmental changes. This results in a lack of interpretability in the recommendation results, making it difficult to support scientific decision-making in high-risk, high-cost scenarios.

[0003] Current ship resource recommendation systems primarily rely on matching static attributes in structured data, lacking the ability to mine deep semantic information from unstructured text (such as accident reports, maintenance logs, and navigation notices) and semi-structured data (such as JSON-formatted equipment status messages). Furthermore, due to the lack of a unified knowledge representation framework, the system struggles to integrate entity relationships and dynamic evolution paths in multimodal data, failing to identify causal chains such as "main engine failure → shutdown → emergency spare parts allocation." Consequently, the recommendation logic remains at the level of superficial relevance, limiting recommendation accuracy and scenario adaptability.

[0004] The ship operating environment is highly dynamic, with supply and demand influenced by various factors such as weather, routes, and equipment health. Traditional recommendation methods typically employ fixed rules or thresholds for constraint processing, lacking the ability to perceive changes in causal relationships in real time. Furthermore, existing technologies often separate knowledge extraction, causal reasoning, and recommendation generation into independent stages, lacking an end-to-end intelligent decision-making loop. Therefore, there is an urgent need for an intelligent recommendation method that integrates AI agents, causal discovery algorithms, and hybrid causal knowledge graphs. This method should be able to automatically mine causal relationships from multi-source heterogeneous ship data, construct interpretable causal chains, and support the dynamic generation of accurate recommendations based on user needs, achieving intelligent, traceable, and adaptive matching of ship supply and demand resources. Summary of the Invention

[0005] This invention primarily addresses the problem of how to automatically mine causal relationships from multi-source heterogeneous ship data, construct interpretable causal chains, and support the dynamic generation of accurate recommendations based on user needs, thereby achieving intelligent, traceable, and adaptive matching of ship supply and demand resources. This invention discloses a method and apparatus for matching ship supply and demand resources based on causal knowledge graphs.

[0006] In a first aspect, this invention discloses a method for matching ship supply and demand resources based on causal knowledge graphs, comprising: S1, collect a multi-source dataset of ship supply and demand resources; the multi-source dataset includes a subset of structured data, a subset of semi-structured data, and a subset of unstructured data; S2, perform causal knowledge graph construction processing on the multi-source dataset to obtain a causal knowledge graph; S3. Based on the causal knowledge graph, process the resource demand scenario information input by the user to obtain resource recommendation result information.

[0007] The step of constructing a causal knowledge graph from the multi-source dataset to obtain the causal knowledge graph includes: S21, preprocess the multi-source dataset to obtain a preprocessed dataset; S22, Perform entity semantic representation on the preprocessed dataset to obtain a set of entity triples; S23, perform causal relationship inference on the entity triple set and the preprocessed dataset to obtain a causal triple set; S24. Based on the set of entity triples and the set of causal triples, a causal knowledge graph is constructed.

[0008] The preprocessed dataset is subjected to entity semantic representation to obtain a set of entity triples, including: S2201, based on the representation table of the structured data subset in the preprocessed dataset, the name of each row of data in the table is taken as an entity, the column headers of the table are taken as attribute names, and the cell values ​​of the table are taken as attribute values. S2202, Use the entity relationship identification method to identify the relationship of all entities and obtain the relationship information of each entity; S2203, using all entity, attribute value and relation information, construct entity triplet combinations and relation triplets; S2204: Using all entity triples and relation triples, a structured set of entity triples is constructed. S2205, for the semi-structured data subset in the preprocessed dataset, identify the object keys or label names of the semi-structured data as entities, and treat the nested content of the semi-structured data as attributes; perform contextual semantic analysis on the semi-structured data to establish relationship information between entities; S2206, using all entity, attribute and relationship information between entities, construct entity triplet combinations and relationship triplet combinations; S2207, using all entity triples and relation triples, constructs a semi-structured set of entity triples; S2208, For the unstructured data subset in the preprocessed dataset, a semantic parsing agent is used to perform named entity recognition, attribute extraction, and relation classification to obtain entities, attributes, and relations; S2209, using all entities, attributes and relations, constructs a set of unstructured entity triples; S2210: Using all the sets of structured entity triples, semi-structured entity triples, and unstructured entity triples, construct the entity triple set.

[0009] The step of performing causal relationship inference on the entity triple set and the preprocessed dataset to obtain a causal triple set includes: S231, perform causal relationship inference on the entity triple set and the preprocessed dataset to obtain an unstructured causal triple set; S232, Perform causal relationship inference on the entity triple set and the preprocessed dataset using semi-structured data to obtain a semi-structured causal triple set; S233, Perform causal relationship inference on the entity triple set and the preprocessed dataset to obtain a structured causal triple set; S234. Using the unstructured causal triple set, the semi-structured causal triple set, and the structured causal triple set, a causal triple set is constructed.

[0010] The step of performing semi-structured causal relationship inference on the entity triple set and the preprocessed dataset to obtain a semi-structured causal triple set includes: S2321, A structure-aware causal reasoning agent is used to extract attribute data from a subset of semi-structured data in the preprocessed dataset to obtain an attribute time series. S2322, Based on the set of entity triples, an initialized causal graph network is constructed; S2323, Based on the attribute time series, a causal graph network optimization model is constructed; S2324, Solve the causal graph network optimization model to obtain the optimal causal graph. ; S2325, from the optimal causal graph In the process, an attribute-level causal relationship set is extracted; the attribute-level causal relationship set includes attribute-level causal relationships; the attribute-level causal relationships include the attributes of two entities that have a causal relationship. S2326, Perform correlation degree fusion calculation on each attribute-level causal relationship in the attribute-level causal relationship set to obtain the corresponding correlation degree value; S2327, using the entity information and correlation value corresponding to an attribute-level causal relationship, a causal triple is constructed; S2328. Using all the causal triples, a semi-structured set of causal triples is constructed.

[0011] The correlation fusion calculation includes: Calculate the first confidence level for attribute-level causal relationships to obtain the first confidence value; The second confidence level is calculated for the attribute-level causal relationship to obtain the second confidence value; The third confidence level is calculated for attribute-level causal relationships to obtain the third confidence value; The first confidence value, the second confidence value, and the third confidence value are fused together to obtain the correlation value corresponding to the attribute-level causal relationship; The expression for calculating the first confidence level is: in, This represents the maximum difference between the scores of intermediate causal graph networks containing attribute-level causal relationships and those not containing attribute-level causal relationships, among all intermediate causal graph networks obtained during the solution process of the causal graph network optimization model. Here, i represents the index of the intermediate causal graph network obtained during the solution process. The maximum value, and These represent the optimal cause-effect graphs. Scoring with and without attribute-level causality. This represents the first confidence level value; The expression for calculating the second confidence level is: , in, This represents the second confidence level. Represents attribute-level causal relationships The number of times it appears in the optimal cause-effect graph. The first attribute representing the attribute-level causal relationship The number of times it appears in the optimal cause-effect graph. The second attribute represents the attribute-level causal relationship; The expression for calculating the third confidence level is as follows: in, Represents attribute-level causal relationships The third confidence level; The expression for the fusion calculation is: in, This represents the sum of the first to third confidence levels. This represents the degree of association corresponding to attribute-level causal relationships.

[0012] The construction of a causal knowledge graph based on the entity triple set and the causal triple set includes: S241, extract the co-occurrence relation set, hierarchical relation set, and user preference set from the entity triple set; The set of co-occurrence relationships , represented as , Represents the i-th entity in the set of entity triples. and the j-th entity The frequency of their co-occurrence The preset frequency threshold; the hierarchical relationship set , represented as , Representing entities Type information; the user preference set , represented as , Let r represent the user entity in the entity triple set, and let r represent the user operation entity in the entity triple set. For entity combination User preference score, The preset score threshold; S242, using the co-occurrence relation set, hierarchical relation set, and user preference set, a non-causal edge set is constructed; S243, using all attribute-level causal relationships in the set of causal triples, a set of directed edges is constructed; S244, Extract all entity information in the entity triplet set; S245, using all the extracted entity information as nodes and the non-causal edge set and directed edge set as edges, a causal knowledge graph is constructed.

[0013] A second aspect of this invention discloses a ship supply and demand resource matching device based on a causal knowledge graph, the device comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the ship supply and demand resource matching method based on causal knowledge graph.

[0014] In a third aspect of this invention, a computer-storable medium is disclosed, wherein the computer-storable medium stores computer instructions, and when the computer instructions are invoked by a computer, they are used to execute the ship supply and demand resource matching method based on causal knowledge graph.

[0015] In a fourth aspect of this invention, an information data processing terminal is disclosed, which is used to implement the ship supply and demand resource matching method based on causal knowledge graph.

[0016] The beneficial effects of this invention are as follows: This invention innovatively integrates AI Agent technology, causal discovery algorithms, and hybrid causal knowledge graphs to propose an intelligent recommendation method for supply and demand resources in the ship supply chain. This method can automatically extract entity-relationship-attribute triples from multi-source heterogeneous data (structured, semi-structured, and unstructured), infer causal relationships between entities, construct a hybrid causal knowledge graph supporting interpretable reasoning, and generate causal transmission chains based on user demand scenarios. This enables accurate and dynamic recommendations for key resources such as ships, spare parts, maintenance services, and shipping routes. By introducing a Semantic Parsing Agent (SPA), a Structure-Aware Causal Inference Agent (SACIA), a Semantic-Driven Causal Discovery Agent (SDCDA), and a Causal Recommendation Agent (CRA), this invention constructs an end-to-end intelligent decision-making framework, significantly improving the accuracy, interpretability, and real-time response capabilities of ship resource matching.

[0017] This invention overcomes the limitations of traditional recommendation systems that rely solely on relevance matching by constructing a hybrid causal knowledge graph, achieving deep reasoning of entity-level causal relationships. The introduced AI Agent system possesses autonomous decision-making and multimodal processing capabilities, significantly improving the system's adaptability to complex ship scenarios. Furthermore, through a dynamic subscription mechanism, the system can perceive knowledge evolution in real time and proactively push strategy updates, compensating for the lag in response of static recommendation models. Using this method, accurate, explainable, and adaptive recommendations for ship supply and demand resources can be achieved, providing powerful intelligent decision support for ship operation, maintenance scheduling, and navigation management, significantly improving the operational efficiency and safety of the ship supply chain. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention. Detailed Implementation

[0019] To better understand the content of this invention, an embodiment is provided here.

[0020] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention.

[0021] In a first aspect, this invention discloses a method for matching ship supply and demand resources based on causal knowledge graphs, comprising: S1, collect a multi-source dataset of ship supply and demand resources; the multi-source dataset includes a subset of structured data, a subset of semi-structured data, and a subset of unstructured data; S2, perform causal knowledge graph construction processing on the multi-source dataset to obtain a causal knowledge graph; S3. Based on the causal knowledge graph, process the resource demand scenario information input by the user to obtain resource recommendation result information.

[0022] The step of constructing a causal knowledge graph from the multi-source dataset to obtain the causal knowledge graph includes: S21, preprocess the multi-source dataset to obtain a preprocessed dataset; S22, Perform entity semantic representation on the preprocessed dataset to obtain a set of entity triples; S23, perform causal relationship inference on the entity triple set and the preprocessed dataset to obtain a causal triple set; S24. Based on the set of entity triples and the set of causal triples, a causal knowledge graph is constructed.

[0023] The preprocessed dataset is subjected to entity semantic representation to obtain a set of entity triples, including: S2201, based on the representation table of the structured data subset in the preprocessed dataset, the name of each row of data in the table is taken as an entity, the column headers of the table are taken as attribute names, and the cell values ​​of the table are taken as attribute values. S2202, Use the entity relationship identification method to identify the relationship of all entities and obtain the relationship information of each entity; S2203, using all entity, attribute value and relation information, construct entity triplet combinations and relation triplets; S2204: Using all entity triples and relation triples, a structured set of entity triples is constructed. S2205, for the semi-structured data subset in the preprocessed dataset, identify the object keys or label names of the semi-structured data as entities, and treat the nested content of the semi-structured data as attributes or associated entities; perform contextual semantic analysis on the semi-structured data to establish relationship information between entities; S2206, using all entity, attribute and relationship information between entities, construct entity triplet combinations and relationship triplet combinations; S2207, using all entity triples and relation triples, constructs a semi-structured set of entity triples; S2208, For the unstructured data subset in the preprocessed dataset, a semantic parsing agent is used to perform named entity recognition, attribute extraction, and relation classification to obtain entities, attributes, and relations; S2209, using all entities, attributes and relations, constructs a set of unstructured entity triples; S2210: Using all the sets of structured entity triples, semi-structured entity triples, and unstructured entity triples, construct the entity triple set.

[0024] The step of performing causal relationship inference on the entity triple set and the preprocessed dataset to obtain a causal triple set includes: S231, perform causal relationship inference on the entity triple set and the preprocessed dataset to obtain an unstructured causal triple set; S232, Perform causal relationship inference on the entity triple set and the preprocessed dataset using semi-structured data to obtain a semi-structured causal triple set; S233, Perform causal relationship inference on the entity triple set and the preprocessed dataset to obtain a structured causal triple set; S234. Using the unstructured causal triple set, the semi-structured causal triple set, and the structured causal triple set, a causal triple set is constructed.

[0025] The step of performing semi-structured causal relationship inference on the entity triple set and the preprocessed dataset to obtain a semi-structured causal triple set includes: S2321, Using a Structure-Aware Causal Reasoning Agent (SACIA), attribute extraction processing is performed on the semi-structured data subset in the preprocessed dataset to obtain the attribute time series; S2322, Based on the set of entity triples, an initialized causal graph network is constructed; S2323, Based on the attribute time series, a causal graph network optimization model is constructed; S2324, Solve the causal graph network optimization model to obtain the optimal causal graph; S2325, From the optimal causal graph, an attribute-level causal relationship set is extracted; the attribute-level causal relationship set includes attribute-level causal relationships; the attribute-level causal relationships include the attributes of two entities that have causal relationships. S2326, Perform correlation degree fusion calculation on each attribute-level causal relationship in the attribute-level causal relationship set to obtain the corresponding correlation degree value; S2327, using the entity information and correlation value corresponding to an attribute-level causal relationship, a causal triple is constructed; S2328. Using all the causal triples, a semi-structured set of causal triples is constructed.

[0026] The correlation fusion calculation includes: Calculate the first confidence level for attribute-level causal relationships to obtain the first confidence value; The second confidence level is calculated for the attribute-level causal relationship to obtain the second confidence value; The third confidence level is calculated for attribute-level causal relationships to obtain the third confidence value; All confidence scores are fused together to obtain the correlation score corresponding to the attribute-level causal relationship.

[0027] The expression for calculating the first confidence level is: in, This represents the maximum difference between the scores of intermediate causal graph networks containing attribute-level causal relationships and those not containing attribute-level causal relationships, among all intermediate causal graph networks obtained during the solution process of the causal graph network optimization model. Here, i represents the index of the intermediate causal graph network obtained during the solution process. The maximum value, and These represent the optimal cause-effect graphs. Scoring with and without attribute-level causality. This represents the first confidence level value; The expression for calculating the second confidence level is: , in, This represents the second confidence level. Represents attribute-level causal relationships The number of times it appears in the optimal cause-effect graph. The first attribute representing the attribute-level causal relationship The number of times it appears in the optimal cause-effect graph. The second attribute represents the attribute-level causal relationship; The expression for calculating the third confidence level is as follows: in, Represents attribute-level causal relationships The third confidence level; The expression for the fusion calculation is: in, This represents the sum of the first to third confidence levels. This represents the degree of association corresponding to attribute-level causal relationships.

[0028] The causal graph network and the optimal causal graph The score was calculated using the GES algorithm. Attribute-level causality. Indirect association exists in relational triples within a set of entity triples, meaning... and They are linked through other relationships. A conflict in a relational triplet of an entity triplet set refers to the existence of a conflict within the relational triplet of the entity triplet set. .

[0029] The step of performing causal relationship inference on the entity triple set and the preprocessed dataset to obtain an unstructured causal triple set includes: This is achieved through a Structural Causal Model (SCM) in the Semantic Driven Causal Discovery Agent (SDCDA), and its expression is: in For the outcome variable, For structure functions, it describes How is it determined by its parent variable? for The set of parent variables, For noise terms, i.e., all terms that are not [discussed / received]. The variables contained in and their influence Other factors; SDCDA utilizes the theoretical framework of SCM to guide its identification of features from unstructured text data. and This allows us to infer causal relationships and ultimately generate causal triples for use in knowledge graphs. Calculating causal strength in text using attention mechanisms: in Let t be the BERT-encoded hidden state of text t, and q be the query vector representing the causal concept. This represents the scoring function used to calculate the hidden state. The relevance or similarity between the query vector q and the query vector q. This represents the attention weight at time step t of the text sequence, which, after weighting, yields the causal confidence score. ; All causal triples constitute a causal set. ,in Confidence level; The step of performing causal relationship inference on the entity triple set and the preprocessed dataset to obtain a structured causal triple set includes: Multi-table joins and information fusion are performed on structured data from multiple sources to obtain a complete set of attribute variables for causal inference. These attribute variables are then... Treating nodes as random variables, construct a joint probability distribution. and based on attribute variables Initialize the causal graph; based on the initialized causal graph, solve for the optimal causal graph. ; The optimal causal graph is obtained by solving based on the initial causal graph. ,include: The posterior distribution model of the causal graph G is as follows: Where D is an observation dataset containing N samples, each sample being... Observed values; To transform the graph discrete problem in traditional causal discovery into a continuously differentiable optimization problem, structure encoding is introduced. ,one A continuous matrix whose elements ; Simultaneously introduce variational distribution , Representing a discrete graph G as a graph composed of... Parameterized probability distribution Its function is to approximate complex true posterior distributions. This distribution It describes the probability of all possible causal graphs occurring; In order to learn the distribution To approximate the true posterior distribution as closely as possible. This is achieved by minimizing the KL divergence between the two: Directly calculating this KL divergence is very difficult. Through mathematical derivation, it can be transformed into an equivalent optimization objective with a computable gradient—the lower bound of evidence: When the lower bound of evidence is optimized to be sufficiently small, it means It becomes A very good approximation is that we can sample the graph with the highest probability from this distribution as the final output: the optimal causal graph. .

[0030] extract The nodes in the diagram represent causal directed edges between attributes. And mapped to entity-level causal triples: ,in, As an attribute, for The corresponding entity, It is an attribute.

[0031] By utilizing all entity-level causal triples, a set of structured causal triples is constructed.

[0032] The construction of a causal knowledge graph based on the entity triple set and the causal triple set includes: S241, extract the co-occurrence relation set, hierarchical relation set, and user preference set from the entity triple set; The set of co-occurrence relationships , represented as: , Represents the i-th entity in the set of entity triples. and the j-th entity The frequency of their co-occurrence This is a preset frequency threshold, which can take the value 0.2; The set of hierarchical relationships , represented as , Representing entities The type information mainly comes from the table type field of structured data. In database tables, there is usually an explicit "type" field. The set of user preferences , represented as , Let r represent the user entity in the entity triple set, and let r represent the user operation entity in the entity triple set. For entity combination User preference score, The preset score threshold can be 3. S242, using the co-occurrence relation set, hierarchical relation set, and user preference set, a non-causal edge set is constructed; S243, using all attribute-level causal relationships in the set of causal triples, a set of directed edges is constructed; S244, Extract all entity information in the entity triplet set; S245, using all the extracted entity information as nodes and the non-causal edge set and directed edge set as edges, a causal knowledge graph is constructed.

[0033] The co-occurrence of the i-th entity and the j-th entity means that they appear in the same relational information.

[0034] The structure-aware causal reasoning agent can be implemented using a SACIA agent. The semantic parsing agent can adopt models such as the Sima Zhuge model.

[0035] The unstructured data subset includes unstructured data; the unstructured data includes natural language text such as technical documents, laws and regulations, news reports, and accident reports. The semi-structured data subset includes semi-structured data; the semi-structured data includes configuration files, logs, or API response data stored in JSON, XML, or HTML format, which have tags, key-value pairs, or hierarchical structures. The entity relationship identification method can be implemented by using foreign key association method or by using preset semantic rules to identify entity relationships; The entity ternary combination includes (entity A, relation R, entity B). The relation triple includes (entity, attribute, value); The preprocessing includes: text data cleaning and processing, etc. The process of processing user-inputted resource demand scenario information based on the causal knowledge graph to obtain resource recommendation results includes: S31, Obtain the resource requirement scenario information input by the user. The resource requirement scenario information is represented as follows: , These represent (user, entity, task, and constraint scenario), respectively. S32, Entity e and task t are obtained from the resource demand scenario information; S33, using entity e as the starting entity and task t as the ending entity, a path to be queried is constructed. The causal knowledge graph is then used to query this path, resulting in a set P of causal path chains from e to t. , in, Indicates a causal path chain; S34, calculate the corresponding comprehensive score for each causal path chain in the causal path chain set P; The formula for calculating the comprehensive score is as follows: in, For the k-th entity in the causal path chain, For causal path chains and The causal confidence of each connected edge. As weight, The preset current resource availability level, This represents the user's preference for resources. The causal confidence level can be calculated using methods similar to those used in causal relationship inference for unstructured data. The current resource availability and preference levels are preset values, which can be obtained centrally from collected multi-source datasets.

[0036] S35, based on the comprehensive score of all causal path chains, output resource recommendation result information; the resource recommendation result information The expression is: in, The set of entities for the causal path chain with the maximum overall score.

[0037] After completing the step of processing the resource demand scenario information input by the user based on the causal knowledge graph to obtain resource recommendation result information, the method further includes: Dynamically subscribing to entity information and updating resource recommendation results based on that entity information specifically includes: The user subscribes to entity e. CRA continuously monitors map update events. When a change in a relevant causal edge is detected: And the change in confidence level exceeds the threshold. : If S3 is executed, a push notification is triggered, and the push frequency is dynamically adjusted by the feedback mechanism. in , To ignore the number of times, For an update event, C is the set of all causal relationships, representing the change of a causal edge in a hybrid causal knowledge graph; This is a boolean function used to determine whether a given update event will affect a specific entity. For push frequency, This is the user's ignored value.

[0038] In the calculation expressions of this invention, the variables involved have all been dimensionless before calculation.

[0039] A second aspect of this invention discloses a ship supply and demand resource matching device based on a causal knowledge graph, the device comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the ship supply and demand resource matching method based on causal knowledge graph.

[0040] In a third aspect of this invention, a computer-storable medium is disclosed, wherein the computer-storable medium stores computer instructions, and when the computer instructions are invoked by a computer, they are used to execute the ship supply and demand resource matching method based on causal knowledge graph.

[0041] In a fourth aspect of this invention, an information data processing terminal is disclosed, which is used to implement the ship supply and demand resource matching method based on causal knowledge graph.

[0042] A fifth aspect of this invention discloses a method for matching ship supply and demand resources based on causal knowledge graphs, comprising: Step 1: Multi-source data preprocessing, extracting entity-relationship-attribute triples. First, multi-source data is collected from ship management systems, maintenance logs, AIS records, ship accident reports, IMO regulatory texts, and port scheduling systems, covering structured, semi-structured, and unstructured data.

[0043] For structured ship resource data, such as ship information and voyage records, stored in databases, tables, CSV files, or other structured formats, the field names and record content have a clear semantic correspondence. When extracting entity-relationship-attribute triples, the table structure is parsed, treating each row of data as an entity instance, the column header as the attribute name, and the cell value as the attribute value. Foreign key associations or preset semantic rules are used to identify the relationships between entities, thereby automatically generating triples in the form of (entity A, relation R, entity B) or (entity, attribute, value). This achieves high-precision, automated triple extraction. Taking the "Ship Equipment Status Table" in a shipping company's database as an example, the fields include MMSI, equipment type, operating status, timestamp, etc. Treating each record as an entity instance and the column header as the attribute name generates triples in the following form: in, Represents the i-th entity. This represents the j-th attribute. The corresponding value of the attribute; For semi-structured data, including configuration files, logs, or API response data stored in JSON, XML, or HTML formats, which have tags, key-value pairs, or hierarchical structures, the parser traverses the data structure, identifies object keys or tag names as entities or attributes, uses nested content as attribute values ​​or associated entities, and combines path matching and contextual semantic analysis to establish relationships between entities, generate standardized triples, and achieve effective semantic extraction of loosely structured data. For unstructured data, including technical documents, laws and regulations, news reports, and accident reports, the information is expressed in free text form and lacks explicit structure. By introducing a Semantic Parsing Agent (SPA), utilizing its integrated natural language understanding module, it performs named entity recognition, relation classification, and attribute extraction based on a pre-trained language model. Combined with contextual semantic understanding capabilities, it autonomously parses the latent semantic structure in the text. The SPA improves triple extraction accuracy by dynamically invoking prompting engineering, knowledge base retrieval, or external tools through task planning and reasoning mechanisms. It also supports coreference resolution and semantic normalization, enabling intelligent and autonomous generation of structured triples from unstructured text. Let the input text be... The SPA first performs named entity recognition: Then, the semantic relationships between entity pairs are determined using a relation classification model: in These are the parameters for the pre-trained language model. For sentences containing causal conjunctions, such as "because A leads to B", the model outputs... Simultaneously, SPA performs coreference resolution: in As a pronoun (such as "it"), Sim Perform semantic normalization on the semantic similarity function (based on BERT embedding): in The mapping function unifies "Main Engine Failure" and "Host failure" into the standard term "Host failure". All triples are cleaned and stored as an RDF triple set. This serves as input for subsequent causal reasoning.

[0044] Step 2: Infer causal relationships and generate causal triples.

[0045] Multi-table joins and information fusion are performed on structured data from multiple sources to obtain a complete set of attribute variables for causal inference. These attribute variables are then... Treating nodes as random variables, construct a joint probability distribution. and based on attribute variables Initialize the causal graph; based on the initialized causal graph, solve for the optimal causal graph. ; The optimal causal graph is obtained by solving based on the initial causal graph. ,include: The posterior distribution model of the causal graph G is as follows: Where D is an observation dataset containing N samples, each sample being... Observed values; To transform the graph discrete problem in traditional causal discovery into a continuously differentiable optimization problem, structure encoding is introduced. ,one A continuous matrix whose elements ; Simultaneously introduce variational distribution , Representing a discrete graph G as a graph composed of... Parameterized probability distribution Its function is to approximate complex true posterior distributions. This distribution It describes the probability of all possible causal graphs occurring; In order to learn the distribution To approximate the true posterior distribution as closely as possible. This is achieved by minimizing the KL divergence between the two: Directly calculating this KL divergence is very difficult. Through mathematical derivation, it can be transformed into an equivalent optimization objective with a computable gradient—the lower bound of evidence: When the lower bound of evidence is optimized to be sufficiently small, it means It becomes A very good approximation is that we can sample the graph with the highest probability from this distribution as the final output: the optimal causal graph. .

[0046] extract The nodes in the diagram represent causal directed edges between attributes. And mapped to entity-level causal triples: ,in, As an attribute, for The corresponding entity, It is an attribute.

[0047] For semi-structured data, which typically contains timestamps and multiple attribute values, a Structure-Aware Causal Inference Agent (SACIA) is used to process semi-structured data from multiple sources and extract attribute time series data. , represents a set of ordered observations formed by the change of attribute values ​​over time. This represents the observed value of this attribute at time point t, where t represents the time index and T represents the total number of records for the entire observation period. Time series of attribute variables Using the observed data, a causal graph G is initialized as the starting point for the search, and the GES algorithm is adopted (this criterion integrates the likelihood function of the data). (And model complexity penalty term) Search for the optimal causal network based on the BIC criterion: in, For the observation dataset, N is the number of samples in the semi-structured data. This is the parameter dimension of the graph (each node (variable) has a conditional probability distribution, which describes the probability that the node (effect) will take different values ​​given its parent nodes (causes). The more parent nodes there are, the more complex the conditional probability distribution becomes, and the more parameters are required). This term is used to penalize overly complex graph structures and prevent overfitting. From the optimal causal graph... Extracting attribute-level causal relationships And further calculate the overall confidence level of the causal edge. It is obtained by weighted fusion of the following three dimensions: Model fit confidence ( (Maximum score difference): Based on the score difference between edges containing and not containing the edge, obtained using the GES algorithm: The normalized result reflects the relative contribution of the edge to the overall model improvement.

[0048] Data support frequency ( ): The co-occurrence ratio of this causal pattern in the log sequence: Domain knowledge matching degree ( ): Determine by consulting ship fault knowledge bases (such as FMEA, maintenance manuals). Is it a known causal pattern? Finally, SACIA weights and fuses the three scores to generate the overall confidence score of the causal edge: Among them, weight satisfy It can be configured according to the application scenario.

[0049] After completing attribute-level causal identification and confidence assessment, SACIA performs semantic enhancement: in That is, the entity to which the attribute belongs.

[0050] For causal triples in unstructured data, this is achieved through the Structured Causal Model (SCM) in the Semantic Driven Causal Discovery Agent (SDCDA), and its expression is as follows: It provides a standard, computable mathematical expression for causal relationships, in which For the outcome variable, For structure functions, it describes How is it determined by its parent variable? for The set of parent variables, For noise terms, i.e., all terms that are not [discussed / received]. The variables contained in and their influence Other factors; SDCDA utilizes the theoretical framework of SCM to guide its identification of features from unstructured text data. and This allows us to infer causal relationships and ultimately generate causal triples for use in knowledge graphs. Calculating causal strength in text using attention mechanisms: in Let t be the BERT-encoded hidden state of text t, and q be the query vector representing the causal concept. This represents the scoring function used to calculate the hidden state. The relevance or similarity between the query vector q and the query vector q. This represents the attention weight at time step t of the text sequence, which, after weighting, yields the causal confidence score. ; All causal triples constitute a causal set. ,in Confidence level; Step 3: Construct a causal subgraph with entity set E as nodes and causal relationship C as directed edges. At the same time, non-causal edges are introduced, including: co-occurrence relations: in, Represents the i-th entity in the set of entity triples. and the j-th entity The frequency of their co-occurrence; Hierarchical relationship: in, Representing entities The type information mainly comes from the table type field of structured data. In database tables, there is usually a clear "type" field. User preferences: The user preference score is derived from the user's historical behavior data and explicit feedback. The system analyzes user u's historical operations to obtain the score, which represents the frequency of common occurrences. The final graph is represented as follows: in, It is a causal boundary set. For other associated edge sets, Each edge is appended with metadata: The graph is stored in Neo4j and supports Cypher queries.

[0051] Step 4: Generate causal chains and implement resource recommendations.

[0052] When the user inputs the requirement scenario (User, Entity, Task, Constraint), Causal Recommendation Agent (CRA) execution: 1) Analyze the key entity e and the target state t; 2) Query the causal path chain starting from e in the graph: 3) Calculate the overall score for each path: in, For the k-th entity in the causal path chain, For causal path chains and The causal confidence of each connected edge. As weight, The preset current resource availability level, This represents the user's preference for certain resources.

[0053] 4) Output recommendation results: Step 5: Dynamic Subscription and Real-time Push: Users subscribe to entity e. CRA continuously monitors map update events. When a change in a relevant causal edge is detected: And the change in confidence level exceeds the threshold. : If S4 is executed, a push notification is triggered, and the push frequency is dynamically adjusted by the feedback mechanism. in , To ignore the number of times, For an update event, C is the set of all causal relationships, representing the change of a causal edge in a hybrid causal knowledge graph; This is a boolean function used to determine whether a given update event will affect a specific entity. For push frequency, This is the user's ignored value.

[0054] In summary, this invention proposes an intelligent recommendation method for ship supply and demand resources that integrates AI Agent, causal discovery algorithm, and hybrid causal knowledge graph. It achieves a closed-loop technology for automatically mining causal relationships from multi-source ship data, generating explainable recommendations, and supporting a dynamic subscription mechanism. This invention not only realizes end-to-end intelligentization of the "data → knowledge → reasoning → recommendation → feedback" chain in its technical architecture, but also addresses practical pain points in the ship supply chain, such as difficulty in resource matching, ambiguous decision-making basis, and delayed response. The proposed method possesses high scalability and engineering applicability, and can be widely applied to key business scenarios such as ship maintenance scheduling, spare parts inventory optimization, navigation risk early warning, and emergency resource allocation, providing strong technical support for smart shipping and the digital transformation of ships.

[0055] A sixth aspect of this invention discloses a method for matching ship supply and demand resources based on causal knowledge graphs, comprising: Step 1) Multi-source data preprocessing: For structured, semi-structured and unstructured ship resource data, perform data cleaning, format parsing and semantic extraction to generate standardized entity-relationship-attribute triples; Step 2) Causal Relationship Inference: Based on the extracted triples, the causal discovery algorithm and AI Agent technology are used to infer the causal direction between entities and generate causal triples (cause entity, confidence level, result entity); Step 3) Construction of Hybrid Causal Knowledge Graph: Integrate causal relationships with various associations such as co-occurrence, hierarchy, attribute attribution, and user preferences to construct a knowledge graph that supports intervention simulation (applying intervention to variables to infer their impact) and counterfactual reasoning (reasoning about counterfactual assumptions); Step 4) Causal chain generation and resource recommendation: Based on the user's input demand scenario information (structured or textual information), query the causal path in the knowledge graph, generate a causal transmission chain from the initial motivation to the target state, identify the recommendable intermediate resource nodes in the chain (an intermediate step in knowledge graph processing), and output accurate and interpretable recommendation results. Step 5) Dynamic Subscription and Real-time Push: Establish a user subscription mechanism for causal chains, continuously monitor knowledge graph updates, and automatically trigger recommendation strategy adjustments and push notifications in real time when significant changes in causal relationships are detected.

[0056] Furthermore, the specific steps of step 1) above are as follows: Step 1)-1: For structured data, including ship registration forms, navigation records, and other data stored in database or CSV format, parse its table structure, treat each row as an entity instance, column headers as attribute names, and cell values ​​as attribute values, generate triples in the form of (entity, attribute, value), and identify the relationships between entities through foreign key associations (identify cases where foreign keys reference the primary keys of other tables, such as identifying MMSI in the table). Steps 1)-2: For semi-structured data, including device logs, API responses, or configuration files in JSON, XML, or YAML format, use a parser to traverse its hierarchical structure, identify object keys or tags as entities or attributes, nested content as attribute values ​​or associated entities, and combine path matching rules (predefined matching rules that use the hierarchical path of the data to map the found values ​​and their paths to "entities", "attributes", or "relationships" in the knowledge graph) to generate standardized triples; Steps 1)-3: For unstructured data, including natural language texts such as accident reports, legal provisions, and maintenance descriptions, a semantic parsing agent (SPA) is used (pre-trained + domain fine-tuning, requiring high-quality ship annotation data) to perform named entity recognition, relation classification, and attribute extraction. Combined with contextual understanding capabilities, it autonomously analyzes the potential semantic structure in the text. Steps 1)-4: SPA dynamically calls prompting engineering, domain knowledge base retrieval or external tools through task planning mechanism to optimize triple extraction accuracy, and performs coreference resolution and semantic normalization to unify different expressions into standard terms to ensure knowledge consistency.

[0057] Furthermore, the specific steps of step 2) above are as follows: Step 2)-1: For the attribute variables in the structured data, map them to random variable nodes, construct the joint probability distribution representation of the random variable nodes, and use the DiBS (Differentiable Bayesian Structure Learning) algorithm to transform causal structure learning into a continuously differentiable optimization problem. Search for the optimal Bayesian network structure through variational inference and gradient optimization to infer the causal direction between attributes. Step 2)-2: For semi-structured data, the value sequence, timestamp and change history of attribute variables are extracted by Structured-AwareCausal Inference Agent (SACIA), time-enhanced feature vectors are constructed, the BIC criterion is optimized by score-based GES algorithm, the optimal causal network is searched, and causal dependencies between variables are identified. Steps 2)-3: After completing attribute-level causal identification, SACIA performs semantic lifting operation based on the entity context to which the attribute belongs, mapping attribute causality such as "temperature rise → pressure rise" to entity-level causal triples such as "engine → cooling system". Steps 2)-4: For unstructured text, the Semantic-Driven Causal Discovery Agent (SDDA) calls the Structural Causal Model (SCM) and the Transformer-based Causal Recognition Model, combining causal connectives, counterfactual expressions and time cues to analyze the implicit causal logic in the text and generate triples with evidence sources (cause entity, confidence level, result entity).

[0058] Furthermore, the specific steps of step 3) above are as follows: Step 3)-1: Based on the causal triples generated in the previous steps, construct a causal subgraph with entities as nodes and causal relationships as directed edges, and attach confidence, evidence source, and time series window metadata to each edge; Step 3)-2: Introduce co-occurrence relationships (indicating frequent co-occurrence of two entities), hierarchical relationships (indicating class structure), attribute affiliation relationships (indicating that an entity has a certain attribute) and user preference relationships (indicating that a user's preference for a certain type of resource) into the same graph to form multi-type association edges; Step 3)-3: Label each edge with the "relationship type" field to clearly distinguish between causal and non-causal edges, and avoid the recommendation system from mistakenly treating relevance as causation; Steps 3)-4: Detect causal loops and logical conflicts through consistency verification mechanisms, and correct unreasonable paths by combining prior knowledge in the shipbuilding field; Steps 3)-5: Store the final graph in a graph database, supporting multi-hop queries, intervention simulations, and counterfactual reasoning to serve subsequent recommendation decisions.

[0059] Furthermore, the specific steps of step 4) above are as follows: Step 4)-1: The Causal Recommendation Agent (CRA) receives the user's input description of the demand scenario, which includes the user's identity, target task (e.g., "emergency repair"), environmental context (e.g., "during a typhoon"), and constraints (e.g., "available within 48 hours"). Step 4)-2: CRA performs semantic parsing on the requirement scenario to extract key entities (such as "a certain ship") and target states (such as "resuming navigation"). Step 4)-3: Execute a query operation on the hybrid causal knowledge graph, retrieve causal paths related to key entities, perform reverse attribution analysis (e.g., "unable to navigate" ← "host failure" ← "filter blockage"), and combine time window and confidence filtering and sorting to generate a set of candidate causal chains; Step 4)-4: CRA filters and reconstructs the optimal causal chain from the candidate chains, and identifies missing but recommendable intermediate resource nodes in the chain (such as "filter replacement service" and "spare parts allocation"). Steps 4)-5: Rank the candidate resources by combining user preferences, resource availability and causal path confidence, output the recommendation results and provide an explanation (e.g., "The host failure is caused by filter blockage, it is recommended to replace the filter first"). Steps 4)-6: CRA continuously optimizes the query strategy, causal chain generation model, and recommendation ranking algorithm through an iterative feedback mechanism to achieve adaptive recommendation.

[0060] Furthermore, the specific steps of step 5) above are as follows: Step 5)-1: The CRA establishes a dynamic subscription relationship between user demand scenarios and causal chains in the knowledge graph, taking the entities, target states, or recommended paths that the user is interested in as subscription topics. Step 5)-2: CRA continuously monitors update events related to the subscribed topic in the hybrid causal knowledge graph, including new causal relationships, changes in causal strength, confidence adjustments, or counterexample denials; Step 5)-3: When a significant change in the causal chain is detected (such as a decrease in confidence exceeding a threshold), CRA triggers a real-time push process to generate a notification message containing updated content, scope of impact, and recommendations for adjusting the recommendation strategy. Step 5)-4: Notification messages are distributed in real time through user terminals, API interfaces or message queues, supporting dynamic optimization and interpretability feedback of recommendation results; Step 5)-5: Based on user feedback (such as ignore, click, confirm) and subscription effectiveness metrics, CRA dynamically adjusts the listening granularity and push frequency to achieve a low-latency, highly relevant personalized subscription service.

[0061] In all embodiments of the present invention, all computational expressions or mathematical functions have undergone dimensionless processing of the variables involved before calculation.

[0062] In all embodiments of the present invention, the values ​​of the independent variables in the input of all computational expressions or mathematical functions meet the reasonable requirements of the input range of the computational expressions or mathematical functions, and can ensure that the computational expressions or mathematical functions can be calculated smoothly without violating physical laws or mathematical rules.

[0063] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for matching ship supply and demand resources based on causal knowledge graphs, characterized in that, include: S1, collects multi-source datasets of ship supply and demand resources; The multi-source dataset includes a subset of structured data, a subset of semi-structured data, and a subset of unstructured data; S2, perform causal knowledge graph construction processing on the multi-source dataset to obtain a causal knowledge graph; S3. Based on the causal knowledge graph, process the resource demand scenario information input by the user to obtain resource recommendation result information.

2. The ship supply and demand resource matching method based on causal knowledge graph as described in claim 1, characterized in that, The step of constructing a causal knowledge graph from the multi-source dataset to obtain the causal knowledge graph includes: S21, preprocess the multi-source dataset to obtain a preprocessed dataset; S22, Perform entity semantic representation on the preprocessed dataset to obtain a set of entity triples; S23, perform causal relationship inference on the entity triple set and the preprocessed dataset to obtain a causal triple set; S24. Based on the set of entity triples and the set of causal triples, a causal knowledge graph is constructed.

3. The ship supply and demand resource matching method based on causal knowledge graph as described in claim 2, characterized in that, The preprocessed dataset is subjected to entity semantic representation to obtain a set of entity triples, including: S2201, based on the representation table of the structured data subset in the preprocessed dataset, the name of each row of data in the table is taken as an entity, the column headers of the table are taken as attribute names, and the cell values ​​of the table are taken as attribute values. S2202, Use the entity relationship identification method to identify the relationship of all entities and obtain the relationship information of each entity; S2203, using all entity, attribute value and relation information, construct entity triplet combinations and relation triplets; S2204: Using all entity triples and relation triples, a structured set of entity triples is constructed. S2205, for the semi-structured data subset in the preprocessed dataset, identify the object keys or label names of the semi-structured data as entities, and treat the nested content of the semi-structured data as attributes; perform contextual semantic analysis on the semi-structured data to establish relationship information between entities; S2206, using all entity, attribute and relationship information between entities, construct entity triplet combinations and relationship triplet combinations; S2207, using all entity triples and relation triples, constructs a semi-structured set of entity triples; S2208, For the unstructured data subset in the preprocessed dataset, a semantic parsing agent is used to perform named entity recognition, attribute extraction, and relation classification to obtain entities, attributes, and relations; S2209, using all entities, attributes and relations, constructs a set of unstructured entity triples; S2210: Using all the sets of structured entity triples, semi-structured entity triples, and unstructured entity triples, construct the entity triple set.

4. The ship supply and demand resource matching method based on causal knowledge graph as described in claim 2, characterized in that, The step of performing causal relationship inference on the entity triple set and the preprocessed dataset to obtain a causal triple set includes: S231, perform causal relationship inference on the entity triple set and the preprocessed dataset to obtain an unstructured causal triple set; S232, Perform causal relationship inference on the entity triple set and the preprocessed dataset using semi-structured data to obtain a semi-structured causal triple set; S233, Perform causal relationship inference on the entity triple set and the preprocessed dataset to obtain a structured causal triple set; S234. Using the unstructured causal triple set, the semi-structured causal triple set, and the structured causal triple set, a causal triple set is constructed.

5. The ship supply and demand resource matching method based on causal knowledge graph as described in claim 4, characterized in that, The step of performing semi-structured causal relationship inference on the entity triple set and the preprocessed dataset to obtain a semi-structured causal triple set includes: S2321, A structure-aware causal reasoning agent is used to extract attribute data from a subset of semi-structured data in the preprocessed dataset to obtain an attribute time series. S2322, Based on the set of entity triples, an initialized causal graph network is constructed; S2323, Based on the attribute time series, a causal graph network optimization model is constructed; S2324, Solve the causal graph network optimization model to obtain the optimal causal graph. ; S2325, from the optimal causal graph In the process, an attribute-level causal relationship set is extracted; the attribute-level causal relationship set includes attribute-level causal relationships; the attribute-level causal relationships include the attributes of two entities that have a causal relationship. S2326, Perform correlation degree fusion calculation on each attribute-level causal relationship in the attribute-level causal relationship set to obtain the corresponding correlation degree value; S2327, using the entity information and correlation value corresponding to an attribute-level causal relationship, a causal triple is constructed; S2328. Using all the causal triples, a semi-structured set of causal triples is constructed.

6. The ship supply and demand resource matching method based on causal knowledge graph as described in claim 5, characterized in that, The correlation fusion calculation includes: Calculate the first confidence level for attribute-level causal relationships to obtain the first confidence value; The second confidence level is calculated for the attribute-level causal relationship to obtain the second confidence value; The third confidence level is calculated for attribute-level causal relationships to obtain the third confidence value; The first confidence value, the second confidence value, and the third confidence value are fused together to obtain the correlation value corresponding to the attribute-level causal relationship; The expression for calculating the first confidence level is: in, This represents the maximum difference between the scores of intermediate causal graph networks containing attribute-level causal relationships and those not containing attribute-level causal relationships, among all intermediate causal graph networks obtained during the solution process of the causal graph network optimization model. Here, i represents the index of the intermediate causal graph network obtained during the solution process. The maximum value, and These represent the optimal cause-effect graphs. Scoring with and without attribute-level causality. This represents the first confidence level value; The expression for calculating the second confidence level is: ， in, This represents the second confidence level value. Represents attribute-level causal relationships The number of times it appears in the optimal cause-effect graph. The first attribute representing the attribute-level causal relationship The number of times it appears in the optimal cause-effect graph. The second attribute represents the attribute-level causal relationship; The expression for calculating the third confidence level is as follows: in, Represents attribute-level causal relationships The third confidence level; The expression for the fusion calculation is: in, This represents the sum of the first to third confidence levels. This represents the degree of association corresponding to attribute-level causal relationships.

7. The ship supply and demand resource matching method based on causal knowledge graph as described in claim 2, characterized in that, The construction of a causal knowledge graph based on the entity triple set and the causal triple set includes: S241, extract the co-occurrence relation set, hierarchical relation set, and user preference set from the entity triple set; The set of co-occurrence relationships , represented as , Represents the i-th entity in the set of entity triples. and the j-th entity The frequency of their co-occurrence The preset frequency threshold; the hierarchical relationship set , represented as , Representing entities Type information; the user preference set , represented as , Let r represent the user entity in the entity triple set, and let r represent the user operation entity in the entity triple set. For entity combination User preference score, The preset score threshold; S242, using the co-occurrence relation set, hierarchical relation set, and user preference set, a non-causal edge set is constructed; S243, using all attribute-level causal relationships in the set of causal triples, a set of directed edges is constructed; S244, Extract all entity information in the entity triplet set; S245, using all the extracted entity information as nodes and the non-causal edge set and directed edge set as edges, a causal knowledge graph is constructed.

8. A ship supply and demand resource matching device based on causal knowledge graph, characterized in that, The device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the ship supply and demand resource matching method based on causal knowledge graph as described in any one of claims 1 to 7.

9. A computer-storable medium, characterized in that, The computer storage medium stores computer instructions, which, when invoked by the computer, are used to execute the ship supply and demand resource matching method based on causal knowledge graph as described in any one of claims 1 to 7.

10. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the ship supply and demand resource matching method based on causal knowledge graph as described in any one of claims 1 to 7.

Images