A ship repair level analysis method and system based on a knowledge graph

Abstract

The application discloses a ship repair level analysis method and system based on a knowledge graph, which comprises the following steps: collecting multi-source heterogeneous data for ship repair level analysis and preprocessing; performing knowledge extraction, constructing an entity set and a relationship set in the field of ship maintenance; constructing a ship maintenance knowledge graph; generating an associated entity set for repair level analysis; using an improved CompGCN model for combined embedding representation learning to obtain deep semantic correlation features between fault entities and repair level entities; performing repair level reasoning calculation to generate repair level analysis results; and writing the repair level analysis results into a maintenance record database and feeding back updates to the ship maintenance knowledge graph, so that intelligent reasoning and determination of the ship repair level are realized, the accuracy and consistency of the repair level analysis results are improved, the dependence on manual experience is reduced, and good engineering application value is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis of ship maintenance, and in particular to a method and system for ship repair level analysis based on knowledge graphs. Background Technology

[0002] As typical large and complex pieces of equipment, ships operate in environments characterized by high humidity, high salt spray, high loads, and long-term continuous operation. Ship equipment is prone to various types of malfunctions during navigation, berthing, and operation. Ship maintenance typically involves multiple subsystems, including main engines, auxiliary engines, propulsion systems, electrical systems, piping systems, and deck machinery. Faults in different equipment exhibit diverse manifestations, complex impacts, and significantly different repair processes. Therefore, in ship maintenance management, accurately analyzing faults and rationally determining repair levels is a crucial technical aspect for ensuring safe ship operation, reducing maintenance costs, and improving maintenance efficiency.

[0003] Current methods for determining ship repair levels typically rely on manual decision-making based on maintenance personnel experience, maintenance specifications, and historical repair cases. However, ship fault information comes from complex sources, including equipment inspection data, alarm messages, maintenance records, maintenance specifications, and historical repair case data. This data is often heterogeneous and multi-sourced, with inconsistent formats, significant redundancy, and frequent missing values. Existing technologies lack a unified processing mechanism for this multi-source, heterogeneous data in the ship maintenance field, making it difficult to effectively integrate fault information and resulting in a lack of data support and structured representation in the repair level analysis process. Summary of the Invention

[0004] One objective of this invention is to propose a knowledge graph-based method and system for ship repair level analysis. This invention fully utilizes multi-source heterogeneous data from the ship maintenance field, knowledge graph construction technology, and an improved CompGCN model to perform entity mapping and relational reasoning on ship fault input information. It systematically realizes deep semantic association feature learning and repair level inference calculation between fault entities and repair level entities, thereby generating repair level analysis results. This invention effectively integrates scattered ship maintenance data, reduces reliance on manual experience, and possesses advantages such as high accuracy in repair level determination, strong consistency of inference results, and strong adaptability to complex fault scenarios.

[0005] A knowledge graph-based method for analyzing ship repair levels according to an embodiment of the present invention includes the following steps: Collect and preprocess multi-source heterogeneous data for ship repair level analysis; Knowledge extraction is performed on the preprocessed multi-source heterogeneous data to construct an entity set and a relation set in the field of ship repair. Construct a ship maintenance knowledge graph based on entity sets and relation sets; Receive fault input information from the ship to be analyzed, map the fault input information to the corresponding fault entity and equipment entity in the ship maintenance knowledge graph, and extract the associated maintenance process entity and repair level entity based on the fault entity and equipment entity to generate a set of associated entities for repair level analysis. By inputting the set of associated entities into the improved CompGCN model for combined embedding representation learning, deep semantic association features between fault entities and repair level entities are obtained. Repair level inference calculations are performed based on deep semantic association features to generate repair level analysis results; The repair level analysis results are written into the maintenance record database, and the repair level analysis results are fed back to update the ship maintenance knowledge graph.

[0006] Optionally, the multi-source heterogeneous data includes basic information of ship equipment, fault detection data, maintenance record data, repair specification data, and historical repair case data. The preprocessing includes data cleaning, noise removal, format unification, and missing value completion.

[0007] Optionally, the construction of the entity set and relation set in the ship repair field specifically includes: The preprocessed multi-source heterogeneous data is parsed one by one to obtain equipment information, fault information, maintenance process information and repair level information related to ship maintenance; Based on equipment information, fault information, maintenance process information, and repair level information, maintenance entity extraction is performed to obtain a candidate set of entities. Perform entity normalization and deduplication on the candidate entity set to construct an entity set for the ship repair domain; Based on the entity set, maintenance relationship extraction is performed to generate a set of relationship instances, forming a set of relationships in the field of ship maintenance. The set of relationship instances includes the association between equipment entities and fault entities, the applicability relationship between fault entities and maintenance process entities, and the correspondence between maintenance process entities and repair level entities.

[0008] Optionally, the construction of the ship maintenance knowledge graph specifically includes: Each entity in the entity set is taken as a node set, and the equipment entity, fault entity, maintenance process entity and repair level entity in the entity set are respectively taken as different types of nodes. Each relation in the relation set is taken as an edge set, and the association, applicable, and corresponding relations in the relation set are respectively taken as different types of edges. For each node in the node set, a node feature vector is constructed. The node feature vector includes the entity type identifier and entity attribute information of the corresponding entity. The node feature vectors are then aggregated to form a node feature matrix. For each edge in the edge set, construct a relation type identifier, and associate the relation type identifier with the corresponding head entity and tail entity to form a relation type set; A ship maintenance knowledge graph is formed based on the set of nodes, the set of edges, the node feature matrix, and the set of relationship types.

[0009] Optionally, the generation of the associated entity set for the repair level analysis specifically includes: The system receives fault input information from the ship to be analyzed, including equipment identification, fault phenomenon, fault code and detection indicators. It obtains equipment entity set and fault entity set from the ship maintenance knowledge graph, and generates corresponding vector representations for equipment identification and fault phenomenon in the fault input information. The similarity between the vector representation and the vector representation of the equipment entity in the set of equipment entities and the vector representation of the fault entity in the set of fault entities is calculated to obtain the similarity results of the equipment entities and the similarity results of the fault entities. Based on a preset similarity threshold, the corresponding equipment entity is obtained from the equipment entity similarity results, and the corresponding fault entity is obtained from the fault entity similarity results; Starting with the corresponding equipment entity and the corresponding fault entity, the system traverses the relationship set along the association, applicability and correspondence relationships in the ship maintenance knowledge graph to extract the maintenance process entity and repair level entity associated with the corresponding equipment entity and the corresponding fault entity. The corresponding equipment entities, corresponding fault entities, maintenance process entities, and repair level entities are summarized to generate a set of related entities for repair level analysis.

[0010] Optionally, obtaining the deep semantic association features specifically includes: The set of related entities from the repair level analysis is input into the improved CompGCN model. The improved CompGCN model includes a relation enhancement combination embedding module, a fault-guided neighborhood aggregation module, and a repair level constraint inference output module. The improvements of the improved CompGCN model are to perform differentiated embedding modeling for different relation types, introduce a fault-guided mechanism into the neighborhood aggregation process, and apply hierarchical consistency constraints to the output results of repair level entities. In the relationship-enhanced composite embedding module, the entities and relationships in the associated entity set of the repair level analysis are subjected to composite embedding representation learning. Relationship-aware gating mechanisms are introduced for association, applicable and corresponding relationships respectively. By assigning independent gating coefficients to different relationship types, the composite representation learning process of entities under different relationship types is dynamically adjusted to generate relationship-enhanced entity embedding representations. In the fault-guided neighborhood aggregation module, a context attention mechanism based on fault entities is introduced. The relationship-enhanced entity embedding representation corresponding to the fault entity is used as the context vector. The neighborhood importance weight is calculated based on the correlation between the relationship-enhanced entity embedding representation corresponding to the adjacent entity and the relationship-enhanced entity embedding representation corresponding to the fault entity. The relationship-enhanced entity embedding representation corresponding to the adjacent entity is weighted and aggregated based on the neighborhood importance weight to obtain the entity embedding representation updated by fault guidance. In the repair level constraint reasoning output module, the entity embedding representation of the fault-guided update is checked for consistency according to the preset hierarchical order of the repair level, and a repair level entity embedding representation that satisfies the hierarchical relationship of the repair level is generated. Based on the fault-guided update entity embedding representation and the repair level entity embedding representation, the association feature calculation between the fault entity and the repair level entity is performed, and the deep semantic association feature is output. The deep semantic association feature includes association score feature, interaction feature and difference feature.

[0011] Optionally, the generation of the repair level analysis results specifically includes: For the deep semantic association features corresponding to each repair level entity, a repair level score is calculated based on preset scoring parameters to obtain the repair level score corresponding to each repair level entity. Normalize the repair level scores to obtain the repair level probability corresponding to each repair level entity. Repair level determination is performed based on repair level probability. The repair level probabilities corresponding to each repair level entity are sorted, the repair level entity with the highest repair level probability is determined, and the repair level identifier corresponding to the repair level entity is used as the repair level analysis result.

[0012] Optionally, the repair level analysis results are associated with the corresponding fault input information, the corresponding equipment entity identifier, and the fault entity identifier, and written into the maintenance record database according to a preset data storage format. After writing, the relationship between the repair level entity and the corresponding fault entity and equipment entity in the ship maintenance knowledge graph is updated based on the repair level analysis results. The repair level analysis results are added to the ship maintenance knowledge graph as new relation attributes or relation instances. The update process involves reading the repair level entity, fault entity, and equipment entity corresponding to the repair level analysis results in the ship maintenance knowledge graph, determining whether a corresponding relation record already exists between the repair level entity and the fault entity and equipment entity. If the relation record already exists, the repair level analysis results are written into the relation record as a relation attribute, and the time identifier or frequency identifier corresponding to the relation record is updated. If the relation record does not exist, a new relation instance is added to the ship maintenance knowledge graph with the equipment entity, fault entity, and repair level entity as nodes, and the repair level analysis results are stored as an attribute of the relation instance.

[0013] A knowledge graph-based ship repair level analysis system according to an embodiment of the present invention includes: The data processing module is used to collect and preprocess multi-source heterogeneous data for ship repair level analysis; The knowledge extraction module is used to extract knowledge from preprocessed multi-source heterogeneous data and construct entity sets and relation sets in the field of ship repair. The graph construction module is used to construct a ship maintenance knowledge graph based on entity sets and relation sets. The mapping and extraction module is used to receive fault input information of the ship to be analyzed, map the fault input information to the corresponding fault entities and equipment entities in the ship maintenance knowledge graph, and extract the associated maintenance process entities and repair level entities based on the corresponding fault entities and equipment entities to generate a set of associated entities for repair level analysis. The feature learning module is used to input the set of associated entities from repair level analysis into the improved CompGCN model for combined embedding representation learning, and to obtain deep semantic association features between fault entities and repair level entities. The inference and computation module is used to perform repair level inference and computation based on deep semantic association features and generate repair level analysis results. The knowledge graph update module is used to write the repair level analysis results into the maintenance record database and to update the repair level analysis results to the ship maintenance knowledge graph.

[0014] The beneficial effects of this invention are: This invention collects multi-source heterogeneous data for ship repair level analysis and performs unified preprocessing. Combined with a knowledge extraction mechanism, it constructs entity and relationship sets in the ship maintenance field, thereby forming a structured ship maintenance knowledge graph. This enables a systematic expression of the relationships, applicability, and correspondences between equipment entities, fault entities, maintenance process entities, and repair level entities. Compared to existing technologies that rely on manual experience or rule tables for repair level determination, this invention integrates scattered maintenance data with maintenance knowledge for modeling, providing clear knowledge support for the relationship between fault information and repair levels, thus improving the standardization and consistency of repair level analysis.

[0015] This invention further receives fault input information from the vessel to be analyzed, maps this information to corresponding fault entities and equipment entities in the ship maintenance knowledge graph, and extracts associated maintenance process entities and repair level entities to generate a set of associated entities for repair level analysis. This constructs an association structure for repair level reasoning at the knowledge graph level. This approach effectively solves the problems of inaccurate matching of fault information and maintenance knowledge, and the difficulty in automatically extracting maintenance relationships in existing technologies, making the repair level analysis process more adaptable and scalable.

[0016] This invention introduces an improved CompGCN model in the feature learning stage, performing combinatorial embedding representation learning on the set of associated entities. Through mechanisms such as relation-enhanced combinatorial embedding, fault-guided neighborhood aggregation, and repair level hierarchical constraint reasoning output, it obtains deep semantic association features between fault entities and repair level entities, significantly enhancing the modeling ability for complex relationship type differences and fault semantic information. Compared to traditional graph embedding or simple graph neural network reasoning methods, this invention can more accurately characterize the deep semantic association between faults and repair levels, improving the accuracy and stability of repair level reasoning calculations.

[0017] This invention performs repair level inference calculations based on deep semantic association features, generates repair level analysis results, writes the analysis results into a maintenance record database, and feeds them back to update the ship maintenance knowledge graph, achieving dynamic supplementation and continuous evolution of maintenance knowledge. This mechanism overcomes the shortcomings of existing technologies, such as lagging maintenance knowledge updates and difficulty in continuously optimizing repair level determination. It enables the system to continuously improve its analytical capabilities as maintenance data accumulates, thus achieving a higher level of intelligence and engineering application value in ship maintenance management. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is an overall flowchart of a knowledge graph-based ship repair level analysis method and system proposed in this invention; Figure 2 This is a schematic diagram illustrating the construction of a ship maintenance knowledge graph for a ship repair level analysis method and system based on knowledge graphs proposed in this invention. Figure 3 This is a schematic diagram of the structure of the improved CompGCN model of the ship repair level analysis method and system based on knowledge graph proposed in this invention; Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A knowledge graph-based method for analyzing ship repair levels includes the following steps: Collect and preprocess multi-source heterogeneous data for ship repair level analysis; Knowledge extraction is performed on the preprocessed multi-source heterogeneous data to construct an entity set and a relation set in the field of ship repair. Construct a ship maintenance knowledge graph based on entity sets and relation sets; Receive fault input information from the ship to be analyzed, map the fault input information to the corresponding fault entity and equipment entity in the ship maintenance knowledge graph, and extract the associated maintenance process entity and repair level entity based on the fault entity and equipment entity to generate a set of associated entities for repair level analysis. By inputting the set of associated entities into the improved CompGCN model for combined embedding representation learning, deep semantic association features between fault entities and repair level entities are obtained. Repair level inference calculations are performed based on deep semantic association features to generate repair level analysis results; The repair level analysis results are written into the maintenance record database, and the repair level analysis results are fed back to update the ship maintenance knowledge graph.

[0021] In this embodiment, the multi-source heterogeneous data includes basic information of ship equipment, fault detection data, maintenance record data, repair specification data, and historical repair case data. The preprocessing includes data cleaning, noise removal, format unification, and missing value completion.

[0022] In this embodiment, the construction of the entity set and relation set in the ship repair field specifically includes: The preprocessed multi-source heterogeneous data is parsed one by one to obtain equipment information, fault information, maintenance process information and repair level information related to ship maintenance; The specific parsing process is as follows: for each data record in the preprocessed multi-source heterogeneous data, field identification and structured splitting are performed according to the data source type; equipment identifier, equipment name, equipment model, and equipment location are parsed into equipment information; fault phenomenon, fault code, detection index, and alarm information are parsed into fault information; maintenance measures, process steps, spare parts replacement, and working hours records are parsed into maintenance process information; and repair level mark, repair category, and repair scope are parsed into repair level information. Based on equipment information, fault information, maintenance process information, and repair level information, maintenance entity extraction is performed to obtain a candidate set of entities. The specific steps of the maintenance entity extraction are as follows: equipment information, fault information, maintenance process information, and repair level information are segmented and labeled with words; candidate terms are initially screened based on a pre-defined entity dictionary and entity type rules for the ship maintenance field; and a sequence labeling model is used to identify entity boundaries and determine entity types from the initial screening results. The identified equipment name, equipment model, and equipment location are identified as equipment entities; fault phenomena, fault codes, and abnormal detection indicators are identified as fault entities; maintenance measure names, process step names, and spare parts replacement items are identified as maintenance process entities; and repair level markers and repair categories are identified as repair level entities. These are then summarized to form a candidate entity set containing equipment entities, fault entities, maintenance process entities, and repair level entities. Perform entity normalization and deduplication on the candidate entity set to construct an entity set for the ship repair domain; The construction process is as follows: For each entity candidate in the entity candidate set, a standard name mapping process is performed. This process includes mapping aliases, abbreviations, synonyms, and entity candidates with different writing formats to a preset standard entity name, and generating a unique identifier for each standard entity name. After completing the standard name mapping, the entity candidates are aggregated based on the unique identifier. Entity candidates with the same standard entity name and pointing to the same unique identifier are merged into the same entity, forming a deduplicated entity record. Entity type tags are added to the deduplicated entity record, and the attribute name, attribute value, and attribute unit corresponding to the entity are associated with the entity, thus summarizing to obtain the entity set in the ship repair field. Based on the entity set, maintenance relationship extraction is performed to generate a set of relationship instances, forming a set of relationships in the field of ship maintenance. The set of relationship instances includes the association between equipment entities and fault entities, the applicability relationship between fault entities and maintenance process entities, and the correspondence between maintenance process entities and repair level entities. The maintenance relationship extraction process uses equipment entities, fault entities, maintenance process entities, and repair level entities from the entity set as extraction objects. It performs relationship trigger word identification and relationship template matching on the field content, text description, and structured records corresponding to the entities in the preprocessed multi-source heterogeneous data. The process extracts the association between equipment entities and fault entities according to the co-occurrence position constraints and field association constraints of equipment entities and fault entities. It extracts the applicable relationship between fault entities and maintenance process entities according to the maintenance measure assignment field constraints and syntactic dependency constraints of fault entities and maintenance process entities. It extracts the correspondence between maintenance process entities and repair level entities according to the specification clause mapping constraints and case record corresponding field constraints of maintenance process entities and repair level entities. Finally, it records the unique entity identifier, tail entity unique identifier, and relationship type for each extracted relationship instance, generating a set of relationship instances.

[0023] In this embodiment, the construction of the ship maintenance knowledge graph specifically includes: Each entity in the entity set is taken as a node set, and the equipment entity, fault entity, maintenance process entity and repair level entity in the entity set are respectively taken as different types of nodes. Each relation in the relation set is taken as an edge set, and the association, applicable, and corresponding relations in the relation set are respectively taken as different types of edges. For each node in the node set, a node feature vector is constructed. The node feature vector includes the entity type identifier and entity attribute information of the corresponding entity. The node feature vectors are then aggregated to form a node feature matrix. The construction process is as follows: For each node in the node set, extract the entity type identifier of the entity corresponding to the node and encode it. The entity type identifier corresponds to the equipment entity, fault entity, maintenance process entity, or repair level entity. At the same time, extract the entity attribute information of the entity corresponding to the node. The entity attribute information includes the attribute name, attribute value, and attribute unit. After the attribute value is processed with a unified unit, it is numerically or discretized and encoded. The encoding result of the entity type identifier and the encoding result of the entity attribute information are concatenated in a preset dimension order to obtain the node feature vector. For each edge in the edge set, construct a relation type identifier, and associate the relation type identifier with the corresponding head entity and tail entity to form a relation type set; The construction process is as follows: for each edge in the edge set, the relationship type is determined according to the relationship category of the edge in the relationship set. The edge between the equipment entity and the fault entity is identified as the association relationship type, the edge between the fault entity and the maintenance process entity is identified as the applicable relationship type, and the edge between the maintenance process entity and the repair level entity is identified as the corresponding relationship type. A unique relationship type code or label is assigned to each relationship type to generate a relationship type identifier. A ship maintenance knowledge graph is formed based on the set of nodes, the set of edges, the node feature matrix, and the set of relationship types.

[0024] In this embodiment, the generation of the associated entity set for repair level analysis specifically includes: The system receives fault input information from the ship to be analyzed, including equipment identification, fault phenomenon, fault code and detection indicators. It obtains equipment entity set and fault entity set from the ship maintenance knowledge graph, and generates corresponding vector representations for equipment identification and fault phenomenon in the fault input information. The specific process for obtaining the equipment entity set and the fault entity set is as follows: In the ship maintenance knowledge graph, entity retrieval is performed using entity type identifiers as filtering conditions. All entity records identified as equipment entities are filtered and aggregated to form the equipment entity set. Similarly, all entity records identified as fault entities are filtered and aggregated to form the fault entity set. When the fault input information contains an equipment identifier, a precise matching retrieval of equipment entities is performed in the ship maintenance knowledge graph based on the equipment identifier. Adjacency traversal is then performed along the association relationships centered on the matched equipment entity to extract associated equipment entities and aggregate them to form the equipment entity set. When the fault input information contains a fault code or fault phenomenon, a precise matching retrieval of fault entities is performed in the ship maintenance knowledge graph based on the fault code or fault phenomenon. Adjacency traversal is then performed along the association relationships centered on the matched fault entity to extract associated fault entities and aggregate them to form the fault entity set. The specific process for generating the vector representation is as follows: Text normalization is performed on both the equipment identifier and the fault phenomenon, including case unification, symbol cleanup, synonym replacement, and word segmentation. The normalized equipment identifier is indexed and mapped in a preset identifier vocabulary to obtain a corresponding identifier sequence, and the identifier sequence is embedded and lookup-based to obtain the initial vector of the equipment identifier. The normalized fault phenomenon is segmented into words to obtain a word term sequence, and each word term is embedded and lookup-based to obtain a word vector sequence. The word vector sequence is aggregated using a weighted summation method to obtain the initial vector of the fault phenomenon, where the weights in the weighted summation are determined by the co-occurrence frequency of the word term in the ship maintenance knowledge graph. Vector normalization is performed on the initial vectors of the equipment identifier and the fault phenomenon to obtain the corresponding vector representations of the equipment identifier and the fault phenomenon. The similarity between the vector representation and the vector representation of the equipment entity in the set of equipment entities and the vector representation of the fault entity in the set of fault entities is calculated to obtain the similarity results of the equipment entities and the similarity results of the fault entities. The specific similarity calculation is as follows: For each device entity in the device entity set, obtain the device entity vector representation of the device entity, and calculate the vector dot product after dimensionally aligning the vector representation corresponding to the device identifier in the fault input information with the vector representation of the device entity. Calculate the vector magnitude of the vector representation corresponding to the device identifier and the vector magnitude of the vector representation of the device entity respectively. Divide the vector dot product by the product of the two vector magnitudes to obtain the similarity score of the device entity. Record the similarity score of each device entity according to the unique identifier of the device entity to form the device entity similarity result. For each fault entity in the fault entity set, obtain the fault entity vector representation of the fault entity, and calculate the vector dot product and vector magnitude of the vector representation corresponding to the fault phenomenon in the fault input information and the vector representation of the fault entity in the same way to obtain the similarity score of the fault entity. Record the similarity score of each fault entity according to the unique identifier of the fault entity to form the fault entity similarity result. Based on a preset similarity threshold, the corresponding equipment entity is obtained from the equipment entity similarity results, and the corresponding fault entity is obtained from the fault entity similarity results; Starting with the corresponding equipment entity and the corresponding fault entity, the system traverses the relationship set along the association, applicability and correspondence relationships in the ship maintenance knowledge graph to extract the maintenance process entity and repair level entity associated with the corresponding equipment entity and the corresponding fault entity. The extraction process starts with the corresponding equipment entity and the corresponding fault entity as the traversal starting point. In the ship maintenance knowledge graph, the fault entities that are directly connected to the corresponding equipment entity through the association relationship are retrieved respectively, and the retrieved fault entities are added to the traversal queue. For each fault entity in the traversal queue, the maintenance process entity that is directly connected to it is retrieved along the applicable relationship, and the repair level entity that is directly connected to it is retrieved along the correspondence relationship of the maintenance process entity. The corresponding equipment entities, corresponding fault entities, maintenance process entities, and repair level entities are summarized to generate a set of related entities for repair level analysis.

[0025] In this embodiment, obtaining the deep semantic association features specifically includes: The set of related entities from the repair level analysis is input into the improved CompGCN model. The improved CompGCN model includes a relation enhancement combination embedding module, a fault-guided neighborhood aggregation module, and a repair level constraint inference output module. The improvements of the improved CompGCN model are to perform differentiated embedding modeling for different relation types, introduce a fault-guided mechanism into the neighborhood aggregation process, and apply hierarchical consistency constraints to the output results of repair level entities. In the relationship-enhanced composite embedding module, the entities and relationships in the associated entity set of the repair level analysis are subjected to composite embedding representation learning. Relationship-aware gating mechanisms are introduced for association, applicable and corresponding relationships respectively. By assigning independent gating coefficients to different relationship types, the composite representation learning process of entities under different relationship types is dynamically adjusted to generate relationship-enhanced entity embedding representations. In the fault-guided neighborhood aggregation module, a context attention mechanism based on fault entities is introduced. The relationship-enhanced entity embedding representation corresponding to the fault entity is used as the context vector. The neighborhood importance weight is calculated based on the correlation between the relationship-enhanced entity embedding representation corresponding to the adjacent entity and the relationship-enhanced entity embedding representation corresponding to the fault entity. The relationship-enhanced entity embedding representation corresponding to the adjacent entity is weighted and aggregated based on the neighborhood importance weight to obtain the entity embedding representation updated by fault guidance. The calculation process is as follows: A vector dot product is used, and the vector magnitude is normalized to obtain the relevance score of each adjacent entity; the relevance score of each adjacent entity is input into an exponential function to obtain a non-negative weight value; the non-negative weight values ​​of all adjacent entities corresponding to the same faulty entity are normalized so that the sum of the normalization results of each adjacent entity is one; the normalization result is used as the neighborhood importance weight of the corresponding adjacent entity. In the repair level constraint reasoning output module, the entity embedding representation of the fault-guided update is checked for consistency according to the preset hierarchical order of the repair level, and a repair level entity embedding representation that satisfies the hierarchical relationship of the repair level is generated. The consistency check specifically involves: extracting the entity embedding representation of the fault-guided update corresponding to the repair level entity from the entity embedding representation of the fault-guided update, and arranging the repair level entities in an ordered manner according to a preset hierarchical order of the repair level; for adjacent repair level entities in the ordered arrangement, calculating the hierarchical order consistency judgment value of their corresponding entity embedding representations in turn, and comparing the hierarchical order consistency judgment value with a preset consistency threshold; when the hierarchical order consistency judgment value of any two adjacent repair level entities does not meet the preset consistency threshold, it is determined that the entity embedding representation of the fault-guided update corresponding to the two adjacent repair level entities does not meet the preset hierarchical order of the repair level, and the repair level entity that does not meet the preset hierarchical order of the repair level is marked as an inconsistent repair level entity; when the hierarchical order consistency judgment values ​​of all two adjacent repair level entities meet the preset consistency threshold, it is determined that the entity embedding representation of the fault-guided update corresponding to the repair level entity meets the preset hierarchical order of the repair level, and is output as a repair level entity embedding representation that meets the hierarchical relationship of the repair level. Based on the fault-guided update entity embedding representation and the repair level entity embedding representation, the association feature calculation between the fault entity and the repair level entity is performed, and the deep semantic association feature is output. The deep semantic association feature includes association score feature, interaction feature and difference feature. The association feature calculation includes calculating the association score feature by performing vector dot product and vector magnitude normalization on the entity embedding representation of fault-guided update and the entity embedding representation of repair level; calculating the interaction feature by performing dimension-wise multiplication on the entity embedding representation of fault-guided update and the entity embedding representation of repair level; and calculating the difference feature by taking the absolute value of the dimension-wise difference between the entity embedding representation of fault-guided update and the entity embedding representation of repair level.

[0026] In this embodiment, the generation of the repair level analysis results specifically includes: For the deep semantic association features corresponding to each repair level entity, a repair level score is calculated based on preset scoring parameters to obtain the repair level score corresponding to each repair level entity. The specific implementation of the scoring calculation is as follows: For the deep semantic association features corresponding to each repair level entity, a preset scoring parameter is called to perform weighted summation on the association score features, interaction features, and difference features in the deep semantic association features, and the weighted summation results of each part are summarized to obtain the repair level score of the repair level entity; wherein, a first weight coefficient is assigned to the association score features, a second weight coefficient is assigned to the interaction features, and a third weight coefficient is assigned to the difference features, and the weighted results corresponding to the first weight coefficient, the second weight coefficient, and the third weight coefficient are added to a preset bias parameter to output the repair level score corresponding to the repair level entity; Normalize the repair level scores to obtain the repair level probability corresponding to each repair level entity. The specific implementation of the normalization process is as follows: obtain the repair level scores corresponding to all repair level entities, and perform exponential mapping processing on the repair level scores of each repair level entity, and use the result after exponential mapping as the non-negative score value of the repair level entity; sum the non-negative score values ​​of all repair level entities to obtain the normalization benchmark value; divide the non-negative score value of each repair level entity by the normalization benchmark value to obtain the repair level probability corresponding to the repair level entity. Repair level determination is performed based on repair level probability. The repair level probabilities corresponding to each repair level entity are sorted, the repair level entity with the highest repair level probability is determined, and the repair level identifier corresponding to the repair level entity is used as the repair level analysis result.

[0027] In this embodiment, the repair level analysis results are associated with the corresponding fault input information, the corresponding equipment entity identifier, and the fault entity identifier, and written into the maintenance record database according to a preset data storage format. After writing, the relationship between the repair level entity and the corresponding fault entity and equipment entity in the ship maintenance knowledge graph is updated based on the repair level analysis results. The repair level analysis results are added to the ship maintenance knowledge graph as new relationship attributes or relationship instances. The update process involves reading the repair level entity, fault entity, and equipment entity corresponding to the repair level analysis results in the ship maintenance knowledge graph, determining whether a corresponding relationship record already exists between the repair level entity and the fault entity and equipment entity, and if the relationship record already exists, the repair level analysis results are written into the relationship record as a relationship attribute, and the time identifier or frequency identifier corresponding to the relationship record is updated. If the relationship record does not exist, a new relationship instance is added to the ship maintenance knowledge graph with the equipment entity, fault entity, and repair level entity as nodes, and the repair level analysis results are stored as an attribute of the relationship instance.

[0028] A knowledge graph-based ship repair level analysis system includes: The data processing module is used to collect and preprocess multi-source heterogeneous data for ship repair level analysis; The knowledge extraction module is used to extract knowledge from preprocessed multi-source heterogeneous data and construct entity sets and relation sets in the field of ship repair. The graph construction module is used to construct a ship maintenance knowledge graph based on entity sets and relation sets. The mapping and extraction module is used to receive fault input information of the ship to be analyzed, map the fault input information to the corresponding fault entities and equipment entities in the ship maintenance knowledge graph, and extract the associated maintenance process entities and repair level entities based on the corresponding fault entities and equipment entities to generate a set of associated entities for repair level analysis. The feature learning module is used to input the set of associated entities from repair level analysis into the improved CompGCN model for combined embedding representation learning, and to obtain deep semantic association features between fault entities and repair level entities. The inference and computation module is used to perform repair level inference and computation based on deep semantic association features and generate repair level analysis results. The knowledge graph update module is used to write the repair level analysis results into the maintenance record database and to update the repair level analysis results to the ship maintenance knowledge graph.

[0029] Example 1: The application scenario is a 3,000-ton multipurpose cargo ship belonging to a coastal shipping company. This vessel operates primarily on near-shore routes and has been in service for over 10 years. Its main propulsion system, auxiliary machinery systems, and deck machinery all exhibit varying degrees of aging. During actual operation, the company found frequent equipment failures. However, the maintenance department heavily relied on experienced personnel to determine the repair level, leading to significant discrepancies in repair levels given by different maintenance engineers for the same fault. This resulted in over-repair of some equipment, increasing maintenance costs, while other equipment suffered secondary failures due to under-repair, impacting navigational safety.

[0030] In this application scenario, multi-source heterogeneous data related to ship repair level analysis is first collected. Data sources include basic equipment information from the ship's equipment ledger, fault detection data collected by the online monitoring system, maintenance record data from the historical maintenance record system, repair specification data published by classification societies, and historical repair case data accumulated internally by the company. These data differ significantly in source, structure, and representation. For example, fault detection data is stored in time-series format, maintenance record data is primarily text-based, and repair specification data is mostly in a clause-based structure. Through unified data cleaning, noise removal, format standardization, and missing value completion, data from different sources are given a consistent foundation for data representation.

[0031] After data preprocessing, knowledge extraction is performed on the multi-source heterogeneous data. The system parses each maintenance record text and repair specification clause, resolving equipment name, equipment model, and equipment location into equipment information; fault phenomena, fault codes, and detection indicators into fault information; maintenance measures, procedures, and spare parts replacement into maintenance process information; and repair category and repair scope into repair level information. Through entity extraction and entity normalization, unified equipment entities, fault entities, maintenance process entities, and repair level entities are constructed. Furthermore, the system extracts the relationships between equipment entities and fault entities, the applicability relationships between fault entities and maintenance process entities, and the correspondence between maintenance process entities and repair level entities, thereby forming a structured ship maintenance knowledge graph.

[0032] In practical applications, when a vessel experiences an alarm for an abnormal main engine fuel pump during a voyage, the system receives fault input information to be analyzed. This includes the equipment identifier "Main Engine Fuel Pump," the fault symptom "Abnormal Pressure Fluctuation," the fault code, and the corresponding detection index value. The system maps this fault input information to the corresponding equipment and fault entities in the ship maintenance knowledge graph. Based on the association between equipment and fault entities in the graph, it extracts the associated maintenance process and repair level entities, generating a set of associated entities for repair level analysis.

[0033] Subsequently, the system inputs the set of related entities into the improved CompGCN model for combined embedding representation learning. In the relation-enhanced combined embedding module, the system distinguishes the impact of association, applicability, and correspondence relationships on entity representations, and obtains relation-enhanced entity embedding representations through a relation-aware gating mechanism. In the fault-guided neighborhood aggregation module, the system uses the fault entity as context to perform weighted aggregation on entities associated with it, giving higher weights to maintenance process entities and repair level entities that are highly relevant to the current fault in the representation space. In the repair level constraint reasoning output module, the system applies hierarchical consistency constraints to the embedding representations of repair level entities to avoid repair level reasoning results that do not conform to actual maintenance logic.

[0034] Based on the aforementioned deep semantic association features, the system performs repair level inference calculations to obtain the scores and probabilities corresponding to each repair level, and finally outputs the repair level analysis results. In this embodiment, the system determines that the repair level of the main engine fuel pump failure is "intermediate repair," a result consistent with the subsequent maintenance decision jointly confirmed by the senior chief engineer and the classification society surveyor.

[0035] During the vessel's 12-month operation, the system analyzed 58 typical equipment failures. Comparison of maintenance data before and after the system's implementation revealed a significant improvement in the consistency and accuracy of repair level determination, a marked reduction in recurring maintenance incidents, and effective control of maintenance costs.

[0036] Table 1. Comparison of Ship Repair Level Analysis System Before and After Application (Statistical Table)

[0037] As shown in Table 1, before the system was implemented, repair levels relied heavily on manual experience, leading to significant differences in repair levels given by different maintenance personnel for the same fault. The consistency rate for repair level determination was only 68.5%, and repair levels were adjusted multiple times during the repair process, resulting in extended repair cycles. After the system was implemented, the repair level analysis method based on knowledge graphs and an improved CompGCN model increased the consistency rate to 91.3%, and the number of repair level adjustments was significantly reduced, indicating that the repair level analysis results output by the system are more stable and reliable.

[0038] Meanwhile, the system can more accurately match the relationship between faults, maintenance processes, and repair levels, making maintenance plans more reasonable. The average single maintenance cycle has been reduced from 9.6 days to 6.8 days, and the total annual maintenance cost has been reduced by approximately 19.4%. In addition, because the repair level determination is more in line with the actual equipment condition, it avoids under-maintenance caused by under-repairing the repair level, and the secondary failure rate has decreased from 14.5% to 5.2%, significantly improving the safety and reliability of ship operations.

[0039] In summary, this embodiment fully demonstrates that the knowledge graph-based ship repair level analysis method and system proposed in this invention can effectively solve the problems of existing technologies, such as reliance on experience in repair level determination, inconsistent results, and high maintenance costs, in real ship maintenance scenarios. It has significant engineering application value and promotion significance.

[0040] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based method for analyzing ship repair levels, characterized in that, Includes the following steps: Collect and preprocess multi-source heterogeneous data for ship repair level analysis; Knowledge extraction is performed on the preprocessed multi-source heterogeneous data to construct an entity set and a relation set in the field of ship repair. Construct a ship maintenance knowledge graph based on entity sets and relation sets; Receive fault input information from the ship to be analyzed, map the fault input information to the corresponding fault entity and equipment entity in the ship maintenance knowledge graph, and extract the associated maintenance process entity and repair level entity based on the fault entity and equipment entity to generate a set of associated entities for repair level analysis. By inputting the set of associated entities into the improved CompGCN model for combined embedding representation learning, deep semantic association features between fault entities and repair level entities are obtained. Repair level inference calculations are performed based on deep semantic association features to generate repair level analysis results; The repair level analysis results are written into the maintenance record database, and the repair level analysis results are fed back to update the ship maintenance knowledge graph.

2. The method for analyzing ship repair levels based on knowledge graphs according to claim 1, characterized in that, The multi-source heterogeneous data includes basic information on ship equipment, fault detection data, maintenance record data, repair specification data, and historical repair case data. The preprocessing includes data cleaning, noise removal, format standardization, and missing value completion.

3. The method for analyzing ship repair levels based on knowledge graphs according to claim 1, characterized in that, The construction of the entity set and relation set in the field of ship repair specifically includes: The preprocessed multi-source heterogeneous data is parsed one by one to obtain equipment information, fault information, maintenance process information and repair level information related to ship maintenance; Based on equipment information, fault information, maintenance process information, and repair level information, maintenance entity extraction is performed to obtain a candidate set of entities. Perform entity normalization and deduplication on the candidate entity set to construct an entity set for the ship repair domain; Based on the entity set, maintenance relationship extraction is performed to generate a set of relationship instances, forming a set of relationships in the field of ship maintenance. The set of relationship instances includes the association between equipment entities and fault entities, the applicability relationship between fault entities and maintenance process entities, and the correspondence between maintenance process entities and repair level entities.

4. The method for analyzing ship repair levels based on knowledge graphs according to claim 1, characterized in that, The construction of the ship maintenance knowledge graph specifically includes: Each entity in the entity set is taken as a node set, and the equipment entity, fault entity, maintenance process entity and repair level entity in the entity set are respectively taken as different types of nodes. Each relation in the relation set is taken as an edge set, and the association, applicable, and corresponding relations in the relation set are respectively taken as different types of edges. For each node in the node set, a node feature vector is constructed. The node feature vector includes the entity type identifier and entity attribute information of the corresponding entity. The node feature vectors are then aggregated to form a node feature matrix. For each edge in the edge set, construct a relation type identifier, and associate the relation type identifier with the corresponding head entity and tail entity to form a relation type set; A ship maintenance knowledge graph is formed based on the set of nodes, the set of edges, the node feature matrix, and the set of relationship types.

5. A method for analyzing ship repair levels based on knowledge graphs according to claim 1, characterized in that, The generation of the associated entity set for the repair level analysis specifically includes: The system receives fault input information from the ship to be analyzed, including equipment identification, fault phenomenon, fault code and detection indicators. It obtains equipment entity set and fault entity set from the ship maintenance knowledge graph, and generates corresponding vector representations for equipment identification and fault phenomenon in the fault input information. The similarity between the vector representation and the vector representation of the equipment entity in the set of equipment entities and the vector representation of the fault entity in the set of fault entities is calculated to obtain the similarity results of the equipment entities and the similarity results of the fault entities. Based on a preset similarity threshold, the corresponding equipment entity is obtained from the equipment entity similarity results, and the corresponding fault entity is obtained from the fault entity similarity results; Starting with the corresponding equipment entity and the corresponding fault entity, the system traverses the relationship set along the association, applicability and correspondence relationships in the ship maintenance knowledge graph to extract the maintenance process entity and repair level entity associated with the corresponding equipment entity and the corresponding fault entity. The corresponding equipment entities, corresponding fault entities, maintenance process entities, and repair level entities are summarized to generate a set of related entities for repair level analysis.

6. A method for analyzing ship repair levels based on knowledge graphs according to claim 1, characterized in that, The acquisition of the deep semantic association features specifically includes: The set of related entities from the repair level analysis is input into the improved CompGCN model. The improved CompGCN model includes a relation enhancement combination embedding module, a fault-guided neighborhood aggregation module, and a repair level constraint inference output module. The improvements of the improved CompGCN model are to perform differentiated embedding modeling for different relation types, introduce a fault-guided mechanism into the neighborhood aggregation process, and apply hierarchical consistency constraints to the output results of repair level entities. In the relationship-enhanced composite embedding module, the entities and relationships in the associated entity set of the repair level analysis are subjected to composite embedding representation learning. Relationship-aware gating mechanisms are introduced for association, applicable and corresponding relationships respectively. By assigning independent gating coefficients to different relationship types, the composite representation learning process of entities under different relationship types is dynamically adjusted to generate relationship-enhanced entity embedding representations. In the fault-guided neighborhood aggregation module, a context attention mechanism based on fault entities is introduced. The relationship-enhanced entity embedding representation corresponding to the fault entity is used as the context vector. The neighborhood importance weight is calculated based on the correlation between the relationship-enhanced entity embedding representation corresponding to the adjacent entity and the relationship-enhanced entity embedding representation corresponding to the fault entity. The relationship-enhanced entity embedding representation corresponding to the adjacent entity is weighted and aggregated based on the neighborhood importance weight to obtain the entity embedding representation updated by fault guidance. In the repair level constraint reasoning output module, the entity embedding representation of the fault-guided update is checked for consistency according to the preset hierarchical order of the repair level, and a repair level entity embedding representation that satisfies the hierarchical relationship of the repair level is generated. Based on the fault-guided update entity embedding representation and the repair level entity embedding representation, the association feature calculation between the fault entity and the repair level entity is performed, and the deep semantic association feature is output. The deep semantic association feature includes association score feature, interaction feature and difference feature.

7. A method for analyzing ship repair levels based on knowledge graphs according to claim 1, characterized in that, The generation of the repair level analysis results specifically includes: For the deep semantic association features corresponding to each repair level entity, a repair level score is calculated based on preset scoring parameters to obtain the repair level score corresponding to each repair level entity. Normalize the repair level scores to obtain the repair level probability corresponding to each repair level entity. Repair level determination is performed based on repair level probability. The repair level probabilities corresponding to each repair level entity are sorted, the repair level entity with the highest repair level probability is determined, and the repair level identifier corresponding to the repair level entity is used as the repair level analysis result.

8. A method for analyzing ship repair levels based on knowledge graphs according to claim 1, characterized in that, The repair level analysis results are associated with the corresponding fault input information, the corresponding equipment entity identifier, and the fault entity identifier, and written to the maintenance record database according to a preset data storage format. After writing, the relationship between the repair level entity and the corresponding fault entity and equipment entity in the ship maintenance knowledge graph is updated based on the repair level analysis results. The repair level analysis results are added to the ship maintenance knowledge graph as new relation attributes or relation instances. The update process involves reading the repair level entity, fault entity, and equipment entity corresponding to the repair level analysis results in the ship maintenance knowledge graph, determining whether a corresponding relation record already exists between the repair level entity and the fault entity and equipment entity, and if the relation record already exists, writing the repair level analysis results as a relation attribute into the relation record and updating the time identifier or frequency identifier corresponding to the relation record. If the relation record does not exist, a new relation instance is added to the ship maintenance knowledge graph with the equipment entity, fault entity, and repair level entity as nodes, and the repair level analysis results are stored as an attribute of the relation instance.

9. A knowledge graph-based ship repair level analysis system, executing the knowledge graph-based ship repair level analysis method according to any one of claims 1 to 9, characterized in that, include: The data processing module is used to collect and preprocess multi-source heterogeneous data for ship repair level analysis; The knowledge extraction module is used to extract knowledge from preprocessed multi-source heterogeneous data and construct entity sets and relation sets in the field of ship repair. The graph construction module is used to construct a ship maintenance knowledge graph based on entity sets and relation sets. The mapping and extraction module is used to receive fault input information of the ship to be analyzed, map the fault input information to the corresponding fault entities and equipment entities in the ship maintenance knowledge graph, and extract the associated maintenance process entities and repair level entities based on the corresponding fault entities and equipment entities to generate a set of associated entities for repair level analysis. The feature learning module is used to input the set of associated entities from repair level analysis into the improved CompGCN model for combined embedding representation learning, and to obtain deep semantic association features between fault entities and repair level entities. The inference and computation module is used to perform repair level inference and computation based on deep semantic association features and generate repair level analysis results. The knowledge graph update module is used to write the repair level analysis results into the maintenance record database and to update the repair level analysis results to the ship maintenance knowledge graph.

Images