A method for constructing and dynamically evolving a product service network knowledge graph

Abstract

The application discloses a kind of product service network knowledge graph construction and dynamic evolution method, belong to dynamic knowledge graph and product service system field.Method includes: first, the knowledge graph of multiple product service networks is collected as data source and the ontology of each is extracted;Then attribute, concept and relation similarity function are sequentially constructed;Based on the calculation result of multi-granularity similarity, the ontology of each data source and the corresponding knowledge graph are merged into the unified ontology and the fusion knowledge graph constructed;When data source changes, first determine whether to carry out full update, select full update or incremental update according to the determination result, and finally extract the dynamic evolution information of the fusion knowledge graph through the type-oriented attention mechanism.The method of the application effectively improves the explainability, logical consistency and credibility of product service cross-source data, and overcomes the absurd or contradictory prediction that may be produced by a pure black box model.

Description

Technical Field

[0001] This invention belongs to the field of dynamic knowledge graphs and product service systems, specifically relating to a method for constructing and dynamically evolving a product service network knowledge graph. Technical Background With the advancement of Industry 4.0 and service-oriented transformation, Product Service Networks (PSNs) are becoming increasingly complex, involving multi-source heterogeneous data (such as sensor data, user feedback, and service orders). Knowledge graphs, as a core technology for semantic knowledge representation, are widely used in knowledge integration and intelligent decision-making within PSNs. However, existing knowledge graph methods are mostly designed for static environments and struggle to adapt to the dynamic and heterogeneous nature of PSNs, leading to knowledge fragmentation, delayed updates, and low inference reliability.

[0002] In terms of ontology construction for knowledge graphs, OntoLogX uses a pre-trained language model (LLM) to automatically generate structured knowledge graphs from log events and their contexts, and guides the model to follow predefined ontology constraints through a small number of examples. Regarding the evolution of knowledge graphs, DyGKT proposes a continuous-time dynamic graph network that distinguishes between long-term stable preferences and short-term behavioral fluctuations through a dual-temporal encoding mechanism, and introduces multi-set indicators to characterize the evolution of complex relationships. In terms of integrating heterogeneous data and enhancing reasoning capabilities, fusing representation learning and logical reasoning can effectively improve these capabilities. Related methods mainly focus on two aspects: first, improving knowledge graph embedding models, such as Iterlogic-E, which improves high-uncertainty predictions by introducing deterministic logical rules, outperforming pure embedding methods in link prediction tasks; second, automatically mining logical rules, such as TLogic, which can automatically mine rules containing temporal logic from large-scale knowledge graphs.

[0003] However, existing technologies still have the following drawbacks: (1) Fragmented entity relationships and lack of a unified semantic framework in product service networks: Existing methods mostly rely on domain-specific ontologies, but the cross-domain nature of PSN makes ontology design complex and prone to semantic conflicts. For example, different data sources may have different definitions of "service status" (such as "running" and "active"), resulting in inconsistent relationship mapping. In addition, manual ontology maintenance is costly and difficult to cover all entity types in PSN, leading to fragmentation of the knowledge graph.

[0004] (2) Static knowledge graphs are difficult to adapt to the dynamic updates and evolution of product service networks: PSNs are dynamic, such as product status updates and new service requests. Existing knowledge graphs mostly adopt periodic updates, which cannot respond to changes in real time and lack evolutionary mechanisms to handle long-term trends (such as service model evolution). This may lead to outdated or contradictory information in the graph, affecting decision support.

[0005] (3) Cross-source heterogeneous product service knowledge and data are difficult to effectively integrate and reason: Existing integration methods are mostly based on static rules or shallow machine learning, which cannot handle the spatiotemporal heterogeneity of multi-source data in PSN. For example, the integration of sensor data and user feedback may generate noise due to different formats, and the lack of consideration for dynamic relationships during reasoning leads to unreliable recommendation or prediction results. Summary of the Invention

[0006] To address the existing technical problems, this invention proposes a method for constructing and dynamically evolving a product service network knowledge graph. The specific technical solution is as follows: S1. Collect multiple product service networks and construct corresponding knowledge graphs. Each knowledge graph serves as a data source. Further extract the ontology of each data source. The ontology includes the entity types of all product services in the knowledge graph, the relationships between product services, and the product service attributes. S2, sequentially construct the attribute similarity function, concept similarity function, and relation similarity function for products and services; and construct an initial unified ontology and the corresponding initial fused knowledge graph; S3, traverse each data source, calculate the attribute similarity, concept similarity and relation similarity between the data source ontology and the unified ontology, merge the attributes, concepts and relations in the data source ontology into the unified ontology, and further merge the vertices and edges of the knowledge graph corresponding to the data source into the fused knowledge graph; S4, Entering the dynamic evolution stage, when the entities or relationships in the data source change, first determine whether the change at the current moment meets the full update trigger condition. If the full update trigger condition is met, then a full update is performed based on the data source at the current moment to obtain the unified ontology and fused knowledge graph at the next moment. If the full update trigger condition is not met, then an incremental update is triggered to synchronize the changes in the data source to the unified ontology and fused knowledge graph at the current moment to obtain the unified ontology and fused knowledge graph at the next moment. S5 introduces type-oriented attention weights and uses a feature fusion network to extract the relationships between products and services in the fused knowledge graph, thereby obtaining the predicted state of the fused knowledge graph at future moments and realizing intelligent scheduling of product and service resources.

[0007] Furthermore, in S2, the concept similarity function is specifically: ; in, These are configurable weighting coefficients, and ; These are the lexical similarity function, attribute similarity function, and structural similarity function, respectively. and These are entity types from different data sources; For concept similarity function; The relation similarity function is specifically as follows: ; in, This is a relation similarity function; Representing the relationship The entity type corresponding to the starting point of the relationship and the entity type corresponding to the ending point of the relationship; These are weighted coefficients, and their sum is 1. and These represent product and service relationships from different data sources.

[0008] Furthermore, the attribute similarity function is calculated using the Jaccard similarity coefficient; the lexical similarity function is obtained by weighting the cosine similarity calculation result and the edit distance calculation result; and the structural similarity function is calculated using the Wu-Palmer similarity.

[0009] Furthermore, in S4, the full update trigger condition is specifically as follows: First, determine whether the cumulative change of all data sources between the current moment and the last full evolution exceeds a threshold. If it exceeds the threshold, a full update operation is triggered. If it does not exceed the threshold, further determine whether the current moment meets the preset full update time period. If it does, a full update operation is triggered. If it does not, exit the determination process.

[0010] Furthermore, in S4, the full update specifically refers to: Calculate the attribute similarity, concept similarity, and relation similarity between the current data source ontology and the current unified ontology. Based on the corresponding similarity calculation results, merge the attributes, concepts, or relations in the current data source ontology into the unified ontology to obtain the updated unified ontology. Based on the updated unified ontology, globally fuse the knowledge graphs of all data sources at the current time to obtain the fused knowledge graph.

[0011] Furthermore, the global fusion specifically refers to: For product services from two different data sources, the concept similarity between each product service and the updated unified ontology is calculated. The attribute similarity between the product services from the two data sources is further calculated. If both the concept similarity and attribute similarity are higher than the corresponding thresholds, the two product services are considered to be of equal value and cross-graph association edges are established, thereby achieving global fusion of the knowledge graphs corresponding to the two data sources. Otherwise, it is impossible to establish such a relationship.

[0012] Furthermore, in step S5, the feature fusion network includes a pre-training process, which specifically involves: S501 maps the entity types of the unified ontology during the training phase to semantic vectors, and encodes each feature information of the knowledge graph product service integrated during the training phase into a feature vector. S502, based on the semantic vector and feature vector, calculate the attention weight of each feature information, and further calculate the fused embedding representation of the product service; S503, based on the embedded representation of product services, calculates the energy scoring function of each directed edge in the fused knowledge graph during the training phase, and further calculates the joint loss function; S504, repeat S501-S504 and update all parameters in reverse according to the joint loss function until the joint loss function converges, finally obtaining the pre-trained feature fusion network.

[0013] Furthermore, the energy scoring function is: in, For Hadamard product, It is the L2 norm. To integrate directed edges in knowledge graphs, These represent the head and tail entities of a directed edge, respectively. The relationship represented by the directed edge. This serves as the timestamp for the establishment of the relationship. These are the fused embedding vectors representing the head and tail entities, respectively, for the product and service offerings. This is a diagonal rotation matrix converted from timestamps. This is the embedding vector of the relation.

[0014] Furthermore, the joint loss function is specifically as follows: ; in, For the joint loss function, For embedding loss, For logical rule loss based on Lukasiewicz T-fuzzy logic; For hyperparameters; This is the set of first-order temporal logic rules introduced, that is, the semantic constraints fused from the knowledge graph; For a single logical rule in the set; To integrate all product services and relationship sets in the knowledge graph at the current moment; This represents the set of all instantiated samples that satisfy the rules. Calculate the expected value; For the overall satisfaction of the rules, The true probability, Given the constraints, This is a conclusion drawn from the constraints.

[0015] Furthermore, the embedding loss is specifically: ; in, For the positive sample set, The negative sample set is generated by randomly replacing the head and tail entities of directed edges in the fused knowledge graph; and These are the replaced head and tail entities, respectively. For preset intervals, To integrate directed edges in knowledge graphs, These represent the head and tail entities of a directed edge, respectively. The relationship represented by the directed edge. This is the timestamp when the relationship was established.

[0016] Compared to existing methods, the advantages of this invention are: (1) The dynamic ontology alignment and fusion method based on multi-granularity similarity proposed in this invention can more accurately and robustly identify the equivalence of concepts and relationships in different data sources by comprehensively calculating the weighted similarity of concepts at the lexical layer (semantics), attribute layer (data pattern), and structural layer (ontology level). This method supports iterative fusion and dynamic updates, and can automatically resolve semantic conflicts caused by differences in terminology and changes in definitions, gradually building and maintaining a globally consistent and scalable unified semantic framework (PSN ontology). This fundamentally solves the problem of knowledge fragmentation, significantly reduces the cost and complexity of manually designing, maintaining, and adapting multi-domain ontology, and provides a solid and unified semantic foundation for subsequent data fusion and reasoning.

[0017] (2) The knowledge graph hybrid update mechanism based on triggering proposed in this invention creatively combines "real-time incremental update" and "condition-triggered full update". Its advantages are as follows: First, by processing incremental data in real time and accurately defining the "influencing subgraph", it can respond to local changes in the network with extremely low latency, ensuring the timeliness of the knowledge graph. Second, by intelligently triggering global reconstruction by monitoring the cumulative change or time threshold, it can periodically correct the cumulative errors or global inconsistencies that may be introduced by incremental updates, ensuring the long-term accuracy of the graph. Finally, this mechanism is systematic and automated, avoiding dependence on static update cycles, enabling the knowledge graph to adapt to the dynamic evolution rhythm of PSN, achieving an optimal balance between efficiency and quality.

[0018] (3) The knowledge reasoning method based on semantic-oriented fusion and temporal logic constraints proposed in this invention has multiple advantages: In terms of feature fusion, through a type-semantic-oriented cross-modal attention mechanism, the contribution weights of different modal features (such as text and temporal data) are dynamically adjusted using entity type information, achieving deeper and more discriminative entity representation learning, effectively overcoming heterogeneity and noise problems. In terms of dynamic relationship modeling, the complex spatial temporal rotation embedding is adopted, which can elegantly model the periodic or continuous evolution of entity states over time, accurately capturing dynamic patterns such as service cycles and device state transitions in PSN. In terms of reasoning reliability, the temporal logic rule constraints based on Lukasiewicz T-modulus are innovatively introduced, transforming domain knowledge into a differentiable form and jointly optimizing it with a data-driven embedding model. This makes the reasoning process not only dependent on data statistical regularities, but also required to conform to basic business logic and physical common sense, thereby greatly improving the interpretability, logical consistency and credibility of the reasoning results, and overcoming the absurd or contradictory predictions that may be generated by pure black-box models. Attached Figure Description

[0019] Figure 1 This is an overall framework diagram of the method of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0021] First, some of the terms used in this invention will be explained: Product service network: A service association network consisting of multiple product services as nodes and the relationships between services as edges.

[0022] To accurately describe the problem solved by this invention, we employ a formal method to define a product-service network knowledge graph and its evolution process. R represents a set of entities (such as products, services, users), and R represents a set of relations (such as "dependency", "provide", "use"). Representing the time domain, then the knowledge graph It can be defined as a directed graph ,in It is a vertex set. It is an edge set, where each directed edge can be represented as a triple, determined by the head node. Associative edge R and tail node composition, This is a timestamp function that indicates when the relationship was established.

[0023] In this context, the knowledge graph dynamic evolution problem addressed by this invention involves updating the graph as it changes over time: given the time interval for graph evolution within a given time domain (e.g., specifying a total time domain of 1 hour, requiring the knowledge graph to evolve every 10 minutes), let... This represents the spectral state at the t-th evolution, and the evolution process is a sequence. This represents the total number of evolutions within the time domain. Update operations include adding entities. Delete entity Adding relationships Deleting relationships Therefore, the map update can be represented as: in, It is the update operator.

[0024] Two major challenges can be identified during the update process. The first is the data fusion problem. This is particularly evident in the construction of the graph. When representing instances, the same entity is associated with multi-source heterogeneous instance content features (such as unstructured text and high-frequency time series sensor data). The challenge is that the modal heterogeneity and numerical noise of these instance content features make it difficult to fuse them into a unified entity vector representation with high-quality discriminative power. Furthermore, there is a lack of mechanisms to utilize the constructed ontology semantic type information to distinguish between key and non-key features in order to guide feature fusion.

[0025] Secondly, there is the problem of reasoning. Given a graph... and its time domain The goal of this study is to deduce the implicit, unrecorded dynamic relationships within the evolutionary data of the graph. The challenge lies in the fact that existing inference models struggle to capture the periodic evolution patterns and nonlinear dynamic trends of the graph. Furthermore, since the inference process is a black box and lacks domain logic constraints, the derived relationships lack interpretability and logical consistency, making it impossible to guarantee their high credibility.

[0026] Therefore, the objective of this invention is to design a method. , making It can build a unified semantic framework and achieve dynamic evolution. It completes feature fusion and logical reasoning, and finally outputs a consistent, real-time and rich knowledge graph.

[0027] Based on the above principles and analysis, this invention provides a systematic solution for the construction and dynamic evolution of knowledge graphs for product service networks. This solution aims to solve the entire chain of problems, from the construction of a unified semantic framework and dynamic evolution support to intelligent reasoning, through a set of interconnected and synergistic technical modules.

[0028] like Figure 1 As shown, the overall technical framework consists of three core parts: First, a unified and evolvable semantic foundation is constructed through a dynamic ontology alignment and fusion method; second, a hybrid update mechanism is designed to ensure that the knowledge graph can respond efficiently and consistently to dynamic changes in the network; finally, a reasoning method that integrates semantic guidance and logical constraints is proposed to achieve deep fusion of cross-source heterogeneous knowledge and highly reliable dynamic relationship deduction. These three parts successively address the three core challenges of semantic heterogeneity, dynamic evolution, and intelligent reasoning, and together constitute a complete, robust, and adaptive product service network knowledge graph management system.

[0029] The detailed plan is as follows: Step 1: Construction of a Unified Semantic Framework To address the issues of fragmented entity relationships and the lack of a unified semantic framework in Product Service Networks (PSNs), this invention proposes a method for constructing a unified semantic framework based on dynamic ontology alignment and fusion. The aim is to build a flexible and consistent PSN Ontology as a shared semantic framework. The core of this method lies in iteratively aligning and fusing multi-source ontologies through a multi-granularity, weighted similarity calculation framework.

[0030] First, this method collects data from multiple data sources describing the interaction relationships between different product and service entities. Each data source contains multiple modal information such as descriptive text and images of the products and services. For each data source... Extract its local knowledge graph and the corresponding local ontology ,in It is a set of concepts (i.e., the entity types contained in the vertex set of a knowledge graph). It is a set of relations (relation type) used to construct directed edges in a knowledge graph. It is the set of attributes corresponding to all entity concepts in the knowledge graph. The local ontology captures the semantic patterns of entities and relationships in the data source, providing a foundation for subsequent alignment with the unified ontology.

[0031] To align multi-source ontology, we define a series of similarity functions to quantify the semantic similarity between concepts, relations, and attributes. The specific definitions of the similarity functions are as follows: (1) Conceptual similarity For any two concepts from different data sources and Their conceptual similarity The result is obtained by weighted summation of similarity across the following three dimensions: in, These are configurable weighting coefficients, and ; These are word similarity, attribute similarity, and structural similarity. The specific calculations for each dimension of similarity are as follows: 1) Lexical similarity The approach employs a linear combination of word vector cosine similarity and edit distance. First, a pre-trained text processing model (such as Word2Vec or BERT) is used to extract labels for entity concepts. And convert them into pre-trained word vectors. Then, cosine similarity is calculated. Simultaneously, the normalized edit distance of the tag strings is calculated. The final formula for calculating word similarity is: in, To adjust the weight, it is usually set to 0.7 to give more emphasis to semantic information; It is an L2 norm.

[0032] 2) Attribute similarity The Jaccard similarity coefficient is used to calculate the attribute set of two concepts. and The similarity is calculated using the following formula: in, The number of elements in the set. The union represents the total number of attributes in the two concept sets; The intersection represents the number of common attributes between two sets of concepts. The intersection is determined by calculating the number of common attributes between the concepts. Each attribute and concept Lexical similarity between each attribute If the calculated result is higher than the predetermined threshold... If the attributes of two concepts are equal, then they are considered to be semantically equivalent and are considered to be an intersection; otherwise, they are considered to be semantically inequivalent.

[0033] 3) Structural similarity The similarity is calculated using Wu-Palmer similarity (WUP similarity) based on the concept's position in the ontology hierarchy. This algorithm determines similarity by calculating the path depth between two concepts and their nearest common ancestor (LCA) within the hierarchy. in, and Representing concepts respectively and The depth within the hierarchical structure of their respective ontologies.

[0034] (2) Relationship similarity Relationships from two different entities Their relationship similarity It considers not only the lexical similarity of the relation itself, but also the degree of matching between its domain and range: in, Calculate the word similarity between relationships. For concept similarity function; Representing the relationship The concepts of domain (i.e., the entity type corresponding to the head node of the relation) and range (i.e., the entity type corresponding to the tail node of the relation). These are the weighting coefficients, and their sum is 1.

[0035] Based on the similarity calculation results, we further perform an alignment operation, specifically: For concepts from two different entities, if (in (If it is a predefined threshold), then the concept They are marked as equivalent and mapped to the same concept. Alignment operations for relations and attributes are similar, using a threshold. In the specific application of this invention, after calculating the similarity result between the ontology of each data source and the unified ontology, if the similarity is higher than the corresponding threshold, the two are considered equivalent, and the concepts, relations or attributes of the data source are mapped to the original concepts, relations and attributes in the unified ontology.

[0036] Among them, Unified Ontology Constructed through an iterative fusion process: Initial It is constructed based on a pre-defined core ontology, and all data sources are iterated and merged one by one. Specifically, for each data source... We calculate the similarity of the above three dimensions to define its local ontology. With the current moment Perform alignment to generate an alignment set. Then update based on the alignment set During the update process, if the aligned sets and include new entity concepts, new relationships, and new attributes, ... In the event of a conflict, it will be resolved through a voting mechanism or by prioritizing based on the reliability of the data source.

[0037] The fusion operation can be formalized as: in, It is the fusion operator, responsible for merging equivalent concepts and relations and handling conflicts, where g is the g-th iteration of the knowledge graph.

[0038] To handle dynamic evolution, the unified ontology's fusion operation continuously monitors changes in data sources. When a new data source is added, the alignment and fusion process is retried to ensure that the unified ontology adapts to the dynamics of the product service network.

[0039] After obtaining the unified ontology, we use it to guide the fusion of entities and relations at the instance level: for data sources Each product / service entity in Calculate the entity type of each entity and The similarity of concepts and attributes between entity types is calculated if the concepts and attributes of the entity types corresponding to the entities are similar. If all are aligned, then... As a unified entity, it is merged into the entity set corresponding to the unified ontology, and denoted as the unified entity. .

[0040] For data source The interaction relationships between each product / service entity in the unified ontology are determined by first calculating the similarity between all relationships in the data source and the relationships in the unified ontology for alignment purposes. Secondly, for the unified entity... Relationships existing in the original data source are compared with existing relationships in the unified ontology. If the similarity is higher than a threshold, the relationship is directly merged into the original relationships in the fused knowledge graph, and the timestamps are merged into a timestamp sequence. If the similarity is lower than the threshold, it is considered new knowledge unique to the data source. This part of the interaction relationship of that entity is treated as a brand-new relationship instance, and the relationship and related head or tail nodes are added to the fused knowledge graph and the unified ontology, retaining the corresponding timestamps. Ultimately, this method outputs a consistent, real-time, and rich fused knowledge graph. This solves the problem of semantic fragmentation and provides a solid foundation for subsequent reasoning.

[0041] Step 2: Dynamic Evolution and Update To address the challenge of static knowledge graphs adapting to the dynamic updates and evolution of product service networks, this invention proposes a trigger-based hybrid dynamic evolution update strategy. This strategy combines real-time incremental updates with periodic full updates to achieve real-time responsiveness and long-term consistency of the knowledge graph. The strategy monitors changes in data sources, triggering updates when they occur. It prioritizes local updates to minimize latency and performs global reconstruction when a certain amount of change accumulates or a time threshold is reached, ensuring the accuracy and adaptability of the graph. The overall approach relies on formalized update operations and impact assessments to efficiently manage the evolution of the knowledge graph.

[0042] (1) Incremental update During the real-time incremental update phase, the system focuses on handling local changes in the data sources. Each data source... The incremental update is formally defined as: in These represent the sets of newly added and deleted entities, respectively. This corresponds to the set of added and deleted relationships. When such an update is detected, the system immediately initiates the incremental processing procedure.

[0043] The core of the update process lies in the precise definition of the scope of impact. The system first identifies all entities in the fused knowledge graph directly affected by the update, forming the affected entity set M: Based on this set, from the currently established fusion knowledge graph Extracting affected subgraphs Within this local scope, the system performs strict consistency checks to ensure that updated content... With unified ontology The semantic constraints are consistent, including the semantic constraint information of the data source itself and the semantic constraints generated in the integrated knowledge graph.

[0044] If the local consistency verification of the affected entity set fails, the data source change will not be updated in the merged knowledge graph. If the local consistency verification passes, the system will apply the incremental update operation. Specifically, the entity set is updated to The edge set is updated to Timestamp function Synchronous updates are performed, recording the current time for newly added relationships. Simultaneously, the unified ontology is updated synchronously.

[0045] This refined local processing ensures the system's high-speed response to dynamic changes.

[0046] (2) Full update During the full update phase, the system aims to resolve global inconsistencies that may accumulate during incremental updates. The update trigger mechanism is based on two independent conditions: when the cumulative changes exceed a threshold or a preset time period is reached, the system initiates the full update process.

[0047] The calculation of cumulative change is formalized as follows: in, Let be the cumulative change in the t-th evolution, and 'a' be the change in the a-th evolution. This is the last full update.

[0048] After entering the full update phase, the system first re-collects the complete state of all data sources at the current moment to obtain the latest local knowledge graph set. Based on this, the system re-executes the similarity calculation, ontology alignment, and fusion process in step one with the current unified ontology to generate an updated unified ontology. .

[0049] This process makes full use of the multi-granularity similarity calculation method in Part 1 to ensure the timeliness and accuracy of the semantic framework.

[0050] Subsequently, the system performs a global fusion operation, establishing cross-graph association edges between the knowledge graphs corresponding to all current data sources to obtain a fused knowledge graph. : Where n is the number of data sources, This is a global fusion function, specifically for any two entities from different data sources. If and only if their entity types are all in the unified ontology Center-aligned and satisfying attribute similarity Only when these are merged into a unified entity can an alignment relationship be established, which is the cross-graph association edge.

[0051] After completing the full reconstruction, the system will replace the current knowledge graph with... It also resets the cumulative changes. This two-tiered update architecture enables the knowledge graph to effectively balance real-time requirements with long-term consistency requirements, providing a reliable knowledge foundation for the dynamic characteristics of the product service network.

[0052] Step 3: Integration and Reasoning Although the foregoing technical solutions of this invention construct a unified semantic framework through ontology alignment, It has completed the association of entity identities and solved the "semantic heterogeneity" problem. However, at the instance content level, cross-source data still exhibits high modal heterogeneity (such as the coexistence of unstructured text and time-series sensor data) and numerical noise. Furthermore, static semantic connections cannot automatically infer implicit causal relationships in dynamic scenarios.

[0053] To this end, this invention further proposes a type-semantic-oriented cross-modal attention fusion and temporal neural symbolic reasoning method. When the fused knowledge graph is updated, the attributes of each entity in the fused knowledge graph are updated synchronously. This aims to deepen the understanding of entity attributes from two dimensions: "feature representation learning" and "logical reasoning." The specific steps are as follows: (1) Semantic-guided fusion This step aims to address the deep fusion problem of multi-source heterogeneous data (such as text and time-series signals) related to the same entity. Unlike simple vector concatenation, this invention utilizes the entity's concept type information to guide the allocation of weights for different modal features.

[0054] For the fusion knowledge graph constructed based on the unified ontology any entity in It is associated with a set of m heterogeneous features. .

[0055] This invention first utilizes the unified ontology constructed in the first part. , obtain entity Concept Types This is then mapped to a learnable type semantic vector (Type Query Vector). Simultaneously, a modality-specific encoder (e.g., BERT for text, LSTM for time series) is used to process each heterogeneous feature. Mapped to the corresponding feature vector Next, type-oriented attention weights are calculated. The calculation formula is as follows: in, For vector concatenation, Let be the projection matrix of the learnable k-th heterogeneous feature. This is a learnable attention parameter vector. Let k be the feature vector of the k-th heterogeneous feature. This is the activation function.

[0056] This mechanism allows the model to focus on the most relevant modalities based on entity type (e.g., "precision instrument" type focuses more on sensor data, "service order" type focuses more on text description).

[0057] Ultimately, the entity Fusion Embedded Representation The weighted summation method was used to calculate: .

[0058] (2) Temporal logical reasoning This step aims to capture the dynamic evolution patterns within the product service network. This invention models the directed edge interactions fused with knowledge graphs as temporal quadruples. ,in These represent the head entity and the tail entity, respectively. For the relationship, For timestamps.

[0059] To effectively represent the evolution of periodic maintenance and service states, this invention embeds entities and relations into a complex vector space. In the middle, the fused embedding representations corresponding to the head entity and the tail entity are denoted as follows: And randomly initialize the embedding vector of the relationship as At the same time, timestamp Mapped to a diagonal rotation matrix , representing the change in the entity's state over time, where, For a predefined sequence of parameters, it represents the specific frequency of the product / service entity in the d-th dimension.

[0060] Define the energy scoring function for a quadruple. for: in, This is the Hadamard product (element-by-element multiplication). It is an L2 norm.

[0061] The smaller the value of this function, the faster it represents in time. Under the influence of this, the closer the head entity is to the tail entity after the relation transformation, the higher the probability that the relation is valid.

[0062] (3) Introducing regularization constraints and joint optimization based on first-order temporal logic rules To address the problem that inference models based solely on vector embedding are prone to overfitting due to data sparsity, and that inference results often lack physical meaning (e.g., deriving conclusions that violate causality), this invention introduces temporal logic rules formalized from domain expert knowledge, using them as regularization terms to constrain the model's training process. The core of this step lies in leveraging the rigor of logical rules to compensate for the blindness of data-driven methods, achieving a balance between "data fitting" and "logical consistency" through joint optimization.

[0063] This invention constructs a joint loss function The aim is to minimize two types of errors: in, For embedding loss, Loss due to logical rules; This is a regularization hyperparameter used to balance the weights between data fitting and logistic loss. This refers to the set of first-order temporal logic rules introduced, i.e., the semantic constraints in the knowledge graph; For a single logical rule in the set; This refers to the set of all instantiated samples in the dynamic graph within the current time window (i.e., the combination of specific entities and relationships that satisfy the rule conditions). This represents the set of all instantiated samples that satisfy the rules. Calculate the expected value; For the overall satisfaction of the rules, The true probability, Given the constraints, This is a conclusion drawn from the constraints.

[0064] The loss function consists of two parts: Embedding loss To learn the distribution patterns of entities and relationships from observed data, this invention employs an interval-based ranking loss. in, For the positive sample set, The negative sample set (generated by randomly replacing entities). and These are the replaced head and tail entities, respectively. This is a preset interval. Its purpose is to widen the gap between reality and noise, ensuring the model can accurately identify known spectral structures.

[0065] Logical rule loss: In order to endow the model with logical reasoning ability and correct unreasonable predictions, this invention uses Lukasiewicz T-fuzzy logic to transform discrete symbolic rules into continuously differentiable numerical constraints.

[0066] To transform domain prior knowledge into constraints on the vector embedding process, this invention instantiates rules based on predefined abstract logical rule templates (e.g., part A fails ∧ part A is a component of machine B ⇒ machine B fails). For each specific instantiated rule encountered... The model does not directly generate conclusions, but rather evaluates its antecedents. With the successor Logical consistency in the current vector space. The specific numerical constraint calculation process is as follows: First, the atomic formulas are truth-valued. The truth probability is defined. The calculation formula is: Where Q is a parameter, Let be the energy scoring function for the quaternion, and tanh be the hyperbolic tangent function. This is achieved by transforming the distance scoring into... The truth probability establishes a bridge between the vector space and the logical space.

[0067] This then aggregates the rule antecedent (Body). If ,calculate: in, and These are two predefined constraints. This formula ensures that the antecedent is activated only when all preconditions are met with high confidence (i.e., the truth probability is higher than a threshold), thus proceeding to the next step of calculation. .

[0068] Finally, the implication relation (Head) constraint is calculated using the Lukasiewicz implication operator to compute the logical rule loss. : This loss forces the model to satisfy the consequent rule with a higher probability when satisfying the antecedent rule, thus strongly and consistently following the fault propagation and service dependency logic in the product service network.

[0069] Within the defined time domain, by minimizing the joint loss mentioned above and updating all learnable parameters in step three, the finally trained model can maintain its data fitting ability while strongly and consistently following the logical constraints in the product service network (such as fault propagation logic and service dependency logic).

[0070] During the model inference phase, multiple dynamically updated data sources to be built are input. Within a specified time domain, the model determines whether to trigger a full update mechanism or whether a specified full update interval has been reached based on the updated content of the data sources. It then selects incremental update or full update accordingly, thereby generating a dynamically updated fusion knowledge graph. Based on the probability P of each relation in the fusion knowledge graph calculated in real time, if the probability of the relation is high, it is injected into the fusion knowledge graph at the current moment. The modified fusion knowledge graph is the future state of the fusion knowledge graph inferred by the system.

[0071] In summary, this invention constructs an end-to-end reasoning framework through the aforementioned three-level linkage mechanism of "feature fusion - temporal embedding - logical constraints." This scheme not only utilizes a type-oriented attention mechanism to address the modal heterogeneity and noise issues of cross-source data, but also overcomes the uninterpretability and logical violations inherent in pure black-box models during dynamic reasoning by introducing fuzzy logic rules. Ultimately, it achieves accurate and reliable inference of the implicit relationships within the product-service network.

[0072] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for constructing and dynamically evolving a product service network knowledge graph, characterized in that, include: S1. Collect multiple product service networks and construct corresponding knowledge graphs. Each knowledge graph serves as a data source, and the ontology of each data source is further extracted. The ontology includes the entity types of all products and services in the knowledge graph, the relationships between products and services, and the attributes of products and services. S2, sequentially construct the attribute similarity function, concept similarity function, and relation similarity function for products and services; and construct an initial unified ontology and the corresponding initial fused knowledge graph; S3, traverse each data source, calculate the attribute similarity, concept similarity and relation similarity between the data source ontology and the unified ontology, merge the attributes, concepts and relations in the data source ontology into the unified ontology, and further merge the vertices and edges of the knowledge graph corresponding to the data source into the fused knowledge graph; S4, Entering the dynamic evolution stage, when the entities or relationships in the data source change, first determine whether the change at the current moment meets the full update trigger condition. If the full update trigger condition is met, then a full update is performed based on the data source at the current moment to obtain the unified ontology and fused knowledge graph at the next moment. If the full update trigger condition is not met, then an incremental update is triggered to synchronize the changes in the data source to the unified ontology and fused knowledge graph at the current moment to obtain the unified ontology and fused knowledge graph at the next moment. S5 introduces type-oriented attention weights and uses a feature fusion network to extract the relationships between products and services in the fused knowledge graph, thereby obtaining the predicted state of the fused knowledge graph at future moments and realizing intelligent scheduling of product and service resources.

2. The method for constructing and dynamically evolving a product service network knowledge graph according to claim 1, characterized in that, In S2, the concept similarity function is specifically as follows: ； in, These are configurable weighting coefficients, and ; These are the lexical similarity function, attribute similarity function, and structural similarity function, respectively. and These are entity types from different data sources; For concept similarity function; The relation similarity function is specifically as follows: ； in, This is a relation similarity function; Representing the relationship The entity type corresponding to the starting point of the relationship and the entity type corresponding to the ending point of the relationship; These are weighted coefficients, and their sum is 1. and These represent product and service relationships from different data sources.

3. The method for constructing and dynamically evolving a product service network knowledge graph according to claim 2, characterized in that, The attribute similarity function is calculated using the Jaccard similarity coefficient; the lexical similarity function is obtained by weighting the cosine similarity calculation result and the edit distance calculation result; and the structural similarity function is calculated using the Wu-Palmer similarity.

4. The method for constructing and dynamically evolving a product service network knowledge graph according to claim 1, characterized in that, In S4, the full update trigger condition is specifically as follows: First, determine whether the cumulative change of all data sources between the current moment and the last full evolution exceeds a threshold. If it exceeds the threshold, a full update operation is triggered. If it does not exceed the threshold, further determine whether the current moment meets the preset full update time period. If it does, a full update operation is triggered. If it does not, exit the determination process.

5. The method for constructing and dynamically evolving a product service network knowledge graph according to claim 1, characterized in that, In S4, the full update specifically refers to: Calculate the attribute similarity, concept similarity, and relation similarity between the current data source ontology and the current unified ontology. Based on the corresponding similarity calculation results, merge the attributes, concepts, or relations in the current data source ontology into the unified ontology to obtain the updated unified ontology. Based on the updated unified ontology, globally fuse the knowledge graphs of all data sources at the current time to obtain the fused knowledge graph.

6. The method for constructing and dynamically evolving a product service network knowledge graph according to claim 5, characterized in that, The global fusion specifically refers to: For product services from two different data sources, the concept similarity between each product service and the updated unified ontology is calculated. The attribute similarity between the product services from the two data sources is further calculated. If both the concept similarity and attribute similarity are higher than the corresponding thresholds, the two product services are considered to be of equal value and cross-graph association edges are established, thereby achieving global fusion of the knowledge graphs corresponding to the two data sources. Otherwise, it is impossible to establish such a relationship.

7. The method for constructing and dynamically evolving a product service network knowledge graph according to claim 1, characterized in that, In step S5, the feature fusion network includes a pre-training process, which specifically involves: S501 maps the entity types of the unified ontology during the training phase to semantic vectors, and encodes each feature information of the knowledge graph product service integrated during the training phase into a feature vector. S502, based on the semantic vector and feature vector, calculate the attention weight of each feature information, and further calculate the fused embedding representation of the product service; S503, based on the embedded representation of product services, calculates the energy scoring function of each directed edge in the fused knowledge graph during the training phase, and further calculates the joint loss function; S504, repeat S501-S504 and update all parameters in reverse according to the joint loss function until the joint loss function converges, finally obtaining the pre-trained feature fusion network.

8. The method for constructing and dynamically evolving a product service network knowledge graph according to claim 7, characterized in that, The energy scoring function is: in, For Hadamard product, It is the L2 norm. To integrate directed edges in knowledge graphs, These represent the head and tail entities of a directed edge, respectively. The relationship represented by the directed edge. This serves as the timestamp for the establishment of the relationship. These are the fused embedding vectors representing the head and tail entities, respectively, for the product and service offerings. This is a diagonal rotation matrix converted from timestamps. This is the embedding vector of the relation.

9. The method for constructing and dynamically evolving a product service network knowledge graph according to claim 7, characterized in that, The joint loss function is specifically as follows: ； in, For the joint loss function, For embedding loss, For logical rule loss based on Lukasiewicz T-fuzzy logic; For hyperparameters; This is the set of first-order temporal logic rules introduced, that is, the semantic constraints fused from the knowledge graph; For a single logical rule in the set; To integrate all product services and relationship sets in the knowledge graph at the current moment; This represents the set of all instantiated samples that satisfy the rule. Calculate the expected value; For the overall satisfaction of the rules, The true probability, Given the constraints, This is a conclusion drawn from the constraints.

10. The method for constructing and dynamically evolving a product service network knowledge graph according to claim 9, characterized in that, The embedding loss is specifically: ； in, For the positive sample set, The negative sample set is generated by randomly replacing the head and tail entities of directed edges in the fused knowledge graph; and These are the replaced head and tail entities, respectively. For preset intervals, To integrate directed edges in knowledge graphs, These represent the head and tail entities of a directed edge, respectively. The relationship represented by the directed edge. This is the timestamp when the relationship was established.

Images