Knowledge graph-based ship repair multivariate recommendation system and method

Abstract

The invention provides a ship repair multi-element recommendation system and method based on a knowledge graph, and belongs to the field of ship repair intelligent decision making. The system comprises a data layer, an engine layer and an application layer, wherein the data layer stores a knowledge graph containing entities such as ships and repair projects in a Neo4j graph database; the engine layer is composed of a rule package engine, a statistical package engine, an intelligent package engine and an operation and maintenance package engine which are complementary in function, and compliance guarantee, high-frequency mode mining, personalized recommendation and dynamic response are achieved respectively. And the application layer provides a standardized API interface. According to the recommendation method, intelligent recommendation is achieved by building a knowledge graph, receiving request parameters, calling engines in parallel, collecting and de-weighting, and presenting results, algorithm models can be switched according to data scenes, and dynamic weight optimization can be introduced subsequently. According to the method, the problems of tedious compliance check, low experience utilization rate and insufficient individuation in the prior art are solved, the efficiency and accuracy of ship repair decision making are remarkably improved, and the workload of engineers is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of ship repair technology and intelligent recommendation, and particularly relates to a ship repair multi-element recommendation system and method based on a knowledge graph. BACKGROUND

[0002] The decision-making of the ship repair industry has long relied on the personal experience of engineers and manual consultation of technical manuals, which has significant technical limitations. In terms of compliance guarantee, engineers need to manually flip through complex specifications such as the Inland Waterway Vessel Inspection Regulations, which is not only time-consuming and laborious, but also prone to omissions or misjudgments. In terms of utilization of historical experience, a large amount of historical repair data exists in an unstructured form, making it difficult to systematically mine valuable repair patterns. In terms of personalized recommendation, existing recommendation algorithms are either too simple to adapt to complex scenarios or too complex to make the recommendation results interpretable. These problems result in low efficiency and low standardization of ship repair decision-making, which cannot meet the requirements of compliance, experience and personalization at the same time, and restricts the intelligent development of the ship repair industry. SUMMARY

[0003] The present application aims to provide a ship repair multi-element recommendation system and method based on a knowledge graph, which solves the problems of cumbersome compliance checking, low utilization of historical experience and insufficient personalized recommendation in the prior art, and realizes multi-dimensional collaborative recommendation of compliance, experience and personalization.

[0004] In order to achieve the above technical features, the purpose of the present application is achieved as follows: a ship repair multi-element recommendation system based on a knowledge graph, comprising a data layer, an engine layer and an application layer; The data layer takes a Neo4j graph database as the core and stores a knowledge graph containing ship entities, repair project entities and semantic relationships. The knowledge graph contains static relationships between ship entities and repair project entities, and also realizes dynamic association of knowledge through a semantic relationship network. The engine layer includes a rule package engine, a statistics package engine, an intelligent package engine and an operation and maintenance package engine. The rule package engine converts industry specifications into machine-readable knowledge and realizes compliance matching through Cypher queries. The statistics package engine mines high-frequency repair patterns based on historical data according to specified dimensions. The intelligent package engine integrates multiple recommendation algorithms and supports model switching. The operation and maintenance package engine responds to dynamic and real-time maintenance requirements of the ship. The application layer provides a standardized API interface to support intelligent generation of repair orders and query of recommendation results.

[0005] Preferably, each rule package in the rule package engine is a node in the knowledge graph, and the condition_config attribute stores the structured matching conditions corresponding to the inspection type and ship type in JSON format.

[0006] Preferably, the statistical dimensions of the statistical package engine include statistical units and statistical years, and the Top-K high-frequency items are screened by screening the execution times of each standard repair item in a specific statistical population.

[0007] Preferably, the recommendation algorithm integrated in the intelligent package engine includes a collaborative filtering (CF) algorithm and a graph attention network (GAT), the collaborative filtering algorithm includes a singular value decomposition (SVD) algorithm and a non-negative matrix factorization (NMF) algorithm, and the switching of the algorithm model is realized by the system global parameters, and the NMF algorithm is enabled by default.

[0008] Preferably, the project sources of the operation and maintenance package engine include daily inspection left-over projects, early warning suggestions generated by real-time monitoring data of the ship state, and special projects supplemented by experts on the spot, and the system continuously tracks high-frequency operation and maintenance projects and uses them as standard project library expansion candidates.

[0009] Preferably, the engine layer adopts a "compliance priority, hierarchical fusion" strategy, the recommendation result of the rule package engine has absolute priority, the statistical package, the intelligent package and the operation and maintenance package engine are sorted according to business importance, and the recommendation result is optimized through a deduplication mechanism.

[0010] Another aspect of the present application provides a ship repair multi-element recommendation method based on a knowledge graph, which is realized by using the ship repair multi-element recommendation system based on the knowledge graph, and includes the following steps: Step 1, constructing a ship repair knowledge graph and storing it in a Neo4j graph database, the knowledge graph contains ship, repair item entities and semantic relationships, and is associated with industry specifications, historical repair data information; Step 2, receiving the request parameters of the repair category and the inspection type input by the user; Step 3, calling the rule package engine, the statistical package engine, the intelligent package engine and the operation and maintenance package engine in parallel, respectively generating compliance recommendation items, high-frequency pattern recommendation items, personalized recommendation items and dynamically supplemented recommendation items; Step 4, collecting and deduplicating the recommendation items output by each engine, and merging the project source information; Step 5, presenting the deduplicated recommendation result to the user through the application layer API interface.

[0011] Preferably, in step 1, when constructing the knowledge graph, the Inland Vessel Inspection Regulations is converted into a machine-readable knowledge representation, and historical repair documents, ship technical archives and inspection and maintenance records are structured and organized to form a semantic relationship network.

[0012] Preferably, in step 3, the rule package engine dynamically generates matching logic through Cypher query, the statistical package engine is grouped and counted according to "statistical unit + statistical year", the intelligent package engine selects the corresponding algorithm according to the data sparsity, and the operation and maintenance package engine responds to the unplanned situation in real time.

[0013] Preferably, in step 4, when the same project is recommended by multiple engines at the same time, a single entry is retained and multiple sources are marked, and the deduplication processing is accurate matching according to the project name. It also includes a dynamic weight optimization step: as user interaction data accumulates, the optimal dynamic weight is assigned to the recommendation results of each engine through offline simulation calculation, realizing the evolution from "comprehensive display" to "precise sorting".

[0014] The present application has the following beneficial effects: 1. More reliable compliance guarantee: The present application converts industry standards into machine-readable knowledge, replaces manual review through automatic matching, avoids rule omission and misjudgment, and ensures that the recommended results meet 100% of industry standards.

[0015] 2. More sufficient use of historical experience: The present application uses the statistical package engine to mine group high-frequency repair patterns, realizes the systematic inheritance and promotion of industry collective wisdom.

[0016] 3. Stronger personalization and flexibility of recommendation: The intelligent package engine of the present application supports algorithm model switching, adapts to different data scenarios, and the operation and maintenance package engine responds to dynamic needs, meeting the unique and immediate requirements of a single ship.

[0017] 4. High engineering practicability: The system of the present application adopts modular design and can be seamlessly integrated into existing workflows, significantly reducing the workload of engineers and improving repair decision-making efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0018] The present application will be further described below in conjunction with the drawings and examples.

[0019] Figure 1 The present application is a system overall architecture diagram.

[0020] Figure 2 The present application is a recommendation method flowchart.

[0021] Figure 3 The present application is a simplified diagram of the system core architecture. DETAILED DESCRIPTION

[0022] The application will be further described in detail below through specific examples. The following examples can enable a person skilled in the art to have a more comprehensive understanding of the application, but in no way limit the application.

[0023] Example 1 I. Overall architecture of the system Referring to Figure 1 The embodiment provides a ship repair multi-element recommendation system based on a knowledge graph, which comprises a data layer, an engine layer and an application layer; the data layer takes a Neo4j graph database as a core and stores a knowledge graph comprising ship entities, repair item entities and semantic relationships; the knowledge graph comprises static relationships between the ship entities and the repair item entities and realizes dynamic association of knowledge through a semantic relationship network; the engine layer comprises a rule package engine, a statistics package engine, an intelligent package engine and an operation and maintenance package engine; the rule package engine converts industry specifications into machine-readable knowledge, realizes compliance matching through Cypher query, the statistics package engine mines high-frequency repair modes based on historical data according to specified dimensions, the intelligent package engine integrates various recommendation algorithms and supports model switching, and the operation and maintenance package engine responds to dynamic and real-time maintenance requirements of the ship; and the application layer provides a standardized API interface, supports intelligent generation of repair orders and query of recommendation results.

[0024] Among them, the system adopts a knowledge graph as a core technical base and constructs a multi-element recommendation architecture comprising four complementary recommendation engines. The entire system can be divided into three main levels: a data layer, an engine layer and an application layer.

[0025] The data layer takes a Neo4j graph database as a core and stores a knowledge graph comprising 479 ships, 8152 historical repair items and 253 standard repair items. The knowledge graph not only comprises static relationships between entities, but also realizes dynamic association of knowledge through a semantic relationship network.

[0026] The engine layer is a core innovation of the system and comprises four specially designed recommendation engines: the rule package engine is responsible for ensuring that the recommendation results completely meet the requirements of industry specifications; the statistics package engine focuses on mining high-frequency modes in historical repair data; the intelligent package engine realizes personalized recommendation by using advanced machine learning algorithms; and the operation and maintenance package engine is used to respond to real-time operation and maintenance requirements and emergency situations.

[0027] The application layer provides a standardized API interface, supports intelligent generation of repair orders and flexible query of recommendation results, and ensures that the system can be seamlessly integrated into the existing ship repair workflow.

[0028] II. Technical functions of each component Rule package: compliance guarantee based on graph query Each rule package in the rule package engine is a node in the knowledge graph, and its condition_config attribute stores the structured matching conditions corresponding to the inspection type and ship type in JSON format. By searching for active rule packages and matching the request parameters, the compliance items that must be performed are determined.

[0029] Among them, the rule package is the core guarantee of system compliance. It converts industry standards such as "Inland Waterway Ship Inspection Rules 2024" into machine-readable knowledge. This paper discards the traditional hard-coded rule engine and innovatively uses the native query capability of the knowledge graph. Each rule package is a node in the graph, and its condition_config attribute stores structured matching conditions (such as inspection type, ship type, etc.) in JSON format. When the recommendation request is triggered, the system will execute a Cypher query to retrieve all active rule packages, and match each rule package's condition_config with the request parameters (such as ship name, inspection type) one by one. The "standard repair items" associated with the matched rule packages are the compliance items that must be performed.

[0030] Statistical package: pattern mining based on historical frequency The statistical dimensions of the statistical package engine include statistical units and statistical years. By counting the number of times each standard repair item is executed within a specific group, the Top-K high-frequency items are selected to form a statistical package.

[0031] Among them, the statistical package aims to discover the high-frequency repair patterns of a specific group from historical data. Unlike the association rule mining that explores the "co-occurrence relationship" between items, the statistical package of this system adopts a more direct statistical recommendation strategy based on historical frequency. Based on historical repair data, it groups according to dimensions such as "statistical unit + statistical year", and counts the total number of times each standard repair item is executed within the group. Finally, the Top-K items with the highest frequency are selected to form a statistical package. For example, the system found that "stern shaft disassembly and inspection" was one of the highest frequency items in the repair records of all ships in the Yangtze River Yueyang waterway in 2023. This pattern will be stored as a statistical package, providing data reference for future recommendations for ships in this unit, effectively making up for the lag and incompleteness of rules.

[0032] Intelligent package: personalized recommendation based on optional models The recommendation algorithms integrated in the intelligent package engine include collaborative filtering (CF) algorithms and graph attention networks (GAT). The collaborative filtering algorithms include singular value decomposition (SVD) and non-negative matrix factorization (NMF) algorithms. The algorithm model is switched through the system global parameters, and the NMF algorithm is enabled by default.

[0033] To meet the demand of personalized recommendation for a single ship, the intelligent package integrates a series of advanced recommendation algorithms. Considering that different algorithms perform differently under different conditions such as data sparsity and graph quality, the system designs a switchable model mechanism based on system parameters rather than fixed weighted fusion. Administrators can flexibly switch between collaborative filtering (CF) and graph attention network (GAT) according to the current data status and recommendation strategy.

[0034] Collaborative filtering (matrix factorization: SVD / NMF) Collaborative filtering is a classic recommendation algorithm that discovers potential associations by analyzing the historical interaction matrix of ships and repair items. The system implements two mainstream collaborative filtering algorithms based on matrix factorization: SVD (TruncatedSVD): Through singular value decomposition, the high-dimensional sparse interaction matrix X ∈ R^(M×N) (ship × item) is approximately decomposed into two low-dimensional dense factor matrices: ship factor matrix P ∈ R^(M×K) and item factor matrix Q ∈ R^(N×K). The recommendation score function for ship u and item i is the inner product of their hidden vectors: s(u,i)=p_u^⊤q_i.

[0035] NMF (Non-negative MF): Non-negative matrix factorization adds a non-negative constraint (X≈WH,W≥0,H≥0) to SVD, making the decomposed factors more interpretable, especially suitable for processing non-negative interaction data such as maintenance records.

[0036] Graph Attention Network (GAT) When the knowledge graph is well constructed, GAT can more deeply utilize the topological structure information of the graph. It assigns different weights to the neighbors of a node through an attention mechanism, aggregates neighbor information to update the node representation.

[0037] Core idea: The core of GAT is the self-attention mechanism. When node i aggregates information from its neighbors j ∈ N(i), the attention coefficient α_ij is dynamically calculated through a learnable function, which allows the model to focus on more important neighbor nodes.

[0038] Model selection mechanism: Instead of fusion output, different algorithms in the intelligent package are switched through the system global parameter. System administrators can configure the currently effective model to be CF or GAT, and can specify the specific model file ID. The current system defaults to the collaborative filtering model (NMF) that performs more stably in data sparse scenarios. This design balances the effectiveness of current recommendations and the flexibility of future technology evolution. When the graph data is more abundant and of higher quality in the future, the GAT model can be smoothly switched to obtain better performance.

[0039] Operation and maintenance package: immediate response to dynamic scenarios The project source of the operation and maintenance package engine includes daily inspection left-over projects, early warning suggestions generated by real-time monitoring data of ship state, and special projects supplemented by experts on the spot, and the system continuously tracks high-frequency operation and maintenance projects and serves as a standard project library expansion candidate.

[0040] The operation and maintenance package is a flexible supplementary engine for the dynamic situation of a single ship. Its content sources are diverse, mainly including: 1) left-over projects found in daily inspection and maintenance that are not included in the plan; 2) early warning maintenance suggestions generated according to real-time monitoring data of ship state (such as vibration, temperature); 3) special projects supplemented by maintenance experts on the spot according to specific ship conditions and historical maintenance records. Unlike the rule package which faces the general specifications of a certain "type" of ship, the operation and maintenance package highly focuses on the uniqueness and immediate needs of the "individual" ship, ensuring that the recommendation system can respond to unexpected situations and dynamic risks outside the plan, and serving as a bridge connecting standardized knowledge and individual actual operation and maintenance. The system will continuously track high-frequency operation and maintenance projects and use them as candidates for expanding the standard project library and rule package in the future.

[0041] Further, the engine layer adopts a "compliance first, hierarchical fusion" strategy, and the recommendation results of the rule package engine have absolute priority. The statistical package, intelligent package, and operation and maintenance package engines are sorted by business importance, and the recommendation results are optimized through a deduplication mechanism.

[0042] Embodiment 2: Referring to Figure 2 The embodiment provides a ship repair multi-element recommendation method based on a knowledge graph, which is implemented by using a ship repair multi-element recommendation system based on a knowledge graph and includes the following steps: Step 1, constructing a ship repair knowledge graph and storing it in a Neo4j graph database, the knowledge graph containing ship, repair project entity and semantic relationship, associating industry specifications, historical repair data information; Step 2, receiving user input repair category and inspection type request parameters; Step 3, calling the rule package engine, statistical package engine, intelligent package engine and operation and maintenance package engine in parallel, respectively generating compliance recommendation projects, high-frequency pattern recommendation projects, personalized recommendation projects and dynamically supplemented recommendation projects; Step 4, collecting and deduplicating the recommendation projects output by each engine, and merging the project source information; Step 5, presenting the deduplicated recommendation results to the user through the application layer API interface.

[0043] Further, in step 1 of constructing the knowledge graph, the "Rules for Inspection of Inland Ships" is converted into a machine-readable knowledge representation, and historical repair documents, ship technical archives, and inspection and maintenance records are structured and organized to form a semantic relationship network.

[0044] Further, in step 3, the rule package engine generates matching logic dynamically through Cypher query, the statistics package engine groups statistics according to "statistical unit + statistical year", the intelligent package engine selects corresponding algorithm according to data sparsity, and the operation and maintenance package engine responds to unplanned conditions in real time.

[0045] Further, in step 4, when the same project is recommended by multiple engines at the same time, a single entry is retained and multiple sources are marked, and the deduplication process is performed according to the exact match of the project name. Further, it also includes a dynamic weight optimization step: as user interaction data accumulates, the optimal dynamic weight is assigned to the recommendation results of each engine through offline simulation calculation, realizing the evolution from "comprehensive display" to "precise sorting".

[0046] Embodiment 3: Referring to Figure 3 , the system workflow is as follows: In order to integrate the outputs of the four engines into a coordinated and consistent recommendation list, the application designs a practical fusion mechanism, and plans different strategies at different stages of project development.

[0047] In the early stage of the project, in order to maximize the recall rate and reduce the initial use threshold of the user, a fusion strategy based on merging and deduplication is adopted. The core process of this strategy is as follows: 1) parallel calling: the system calls all applicable recommendation packages in parallel according to the repair category and inspection category input by the user. 2) result collection: all recommended projects generated by the rule package, statistics package, intelligent package and operation and maintenance package are collected into a list. 3) merging and deduplication: the projects in the list are deduplicated according to the name, and the source information is merged. For example, if a project is recommended by the statistics package and the intelligent package at the same time, it will finally appear as an entry and mark its multiple sources. 4) unified presentation: the final result after deduplication is completely displayed to the user.

[0048] Practice has proved that through reasonable setting of the recommendation number (Top-K) of each package, this simple and efficient strategy can maintain acceptable precision rate while ensuring high recall rate, which meets the application goal of the early stage of the project. With the continuous accumulation of user feedback and interaction data, the system will introduce a dynamic weight mechanism. At that time, the optimal dynamic weight will be found for the recommendation results of different engines through offline simulation calculation of historical adoption data, evolving from "comprehensive display" to "precise sorting" to further improve the accuracy of the recommendation.

[0049] Main technical advantages of the application: 1. Innovative application of knowledge graph technology.

[0050] The system realizes important innovation in the application of knowledge graph technology. Traditional ship repair knowledge exists in the form of scattered technical manuals and engineer experience, lacking a unified knowledge representation framework, which directly leads to low knowledge utilization efficiency and inconsistent decision-making standards. The system creatively unifies industry standards, historical data, and algorithm models under the knowledge graph framework, building a complete ship repair knowledge network containing 10116 nodes and 27884 relationships.

[0051] This innovative application is reflected in three key aspects: First, the system converts complex regulatory documents such as the "Rules for Inspection of Inland Waterway Ships" into machine-readable knowledge representation, solving the rigid problem of traditional rule engines and enabling compliance checks to be automatically executed instead of manually reviewed. Second, the semantic relationship network enables the structured organization of historical repair data, providing a solid data foundation for experience mining and enabling collective wisdom to be systematically inherited and utilized. Finally, the graph structure of the knowledge graph provides rich contextual information for intelligent recommendation algorithms, significantly improving the accuracy and interpretability of recommendations. This fine-grained knowledge representation fundamentally changes the status quo of insufficient knowledge representation in traditional technology.

[0052] 2. Design innovation of multi-element recommendation architecture.

[0053] The four-element recommendation architecture is the core innovation of the system. Traditional recommendation systems often use a single algorithm architecture, which cannot meet multiple constraint conditions in the ship repair scenario, limiting the practical application value of the recommendation system in industrial scenarios. The system realizes multi-dimensional satisfaction of complex business requirements through the collaborative work of four functionally complementary engines: rule package, statistical package, intelligent package, and operation and maintenance package.

[0054] The core value of this architectural innovation lies in the "divide and conquer" design philosophy. The rule package engine is designed to address compliance assurance issues, ensuring that the recommendation results meet industry standards 100%, effectively avoiding safety hazards caused by rule omissions. The statistical package engine focuses on the effective use of historical experience, leveraging group repair patterns to inherit industry wisdom and enable excellent repair practices to be systematically promoted and applied. The intelligent package engine uses advanced machine learning algorithms to achieve personalized recommendations, providing targeted repair suggestions based on specific ship characteristics and historical data. The operation and maintenance package engine is designed to handle dynamic business requirements, responding to unexpected situations and special requirements in a timely manner.

[0055] The system adopts the strategy of "compliance priority and hierarchical fusion", ensuring that the results of the rule package have absolute priority, which is of decisive significance in actual business scenarios. The statistical package, intelligent package, and operation and maintenance package are prioritized according to business importance, and the intelligent deduplication mechanism ensures the simplicity and effectiveness of the recommended results. This fusion mechanism not only solves the problem of single algorithm unable to balance multiple constraints, but also ensures the interpretability and business rationality of the recommended results, enabling engineers to clearly understand the decision logic of each recommended item.

[0056] 3. Innovation breakthrough in algorithm technology.

[0057] At the algorithm level, the system realizes multiple technical innovations, significantly improving the performance of the recommendation system. The graph attention network algorithm integrated in the intelligent package engine can fully utilize the graph structure information of the knowledge graph and capture the complex associations between different entities through a multi-head attention mechanism. Compared with traditional collaborative filtering algorithms, this graph neural network-based recommendation method has better interpretability, and engineers can clearly understand the decision basis of each recommended item, which is crucial for ship repair decisions that need to bear safety responsibilities.

[0058] Another important algorithm innovation is the switchable design of the model. The system allows flexible switching between collaborative filtering and graph attention network algorithms according to data characteristics and business needs. This design ensures the flexibility of algorithm selection and provides the possibility for performance optimization in different scenarios. In particular, in the case of data sparsity or cold start, the system can automatically select a more suitable algorithm strategy to ensure the stability of the recommendation quality, and this adaptive ability greatly enhances the robustness of the system in actual application.

[0059] 4. Technical value of engineering practice.

[0060] From the perspective of engineering practice, the technical value of this system is reflected in several aspects. First, the system realizes the organic unification of knowledge representation and algorithm calculation, providing a reusable technical framework for the intelligent transformation of similar industrial scenarios, and this architecture design has good promotion value. Second, the modular design of the system allows each component to be optimized and upgraded independently, with good maintainability and scalability, and can adapt to the changing business needs. Most importantly, the performance of the system in actual application fully proves the feasibility and effectiveness of its technical solution.

[0061] Experimental results show that the system has a significant improvement in key indicators compared to the optimal baseline model. In the actual application case, when the system helped "Channel 08109" to formulate the 2025 annual intermediate repair plan, the precision rate reached 64.3%, the recall rate reached 69.2%, and the engineers' work burden was reduced by nearly 70%. These data fully prove the actual value brought by the technical innovation of this system.

[0062] Although the preferred embodiments of the present application have been described above with reference to the accompanying drawings, the present application is not limited to the above-described specific embodiments, and the above-described specific embodiments are merely illustrative and are not restrictive. A person of ordinary skill in the art can make many specific modifications under the inspiration of the present application without departing from the spirit of the present application and the scope of protection of the claims, and these all belong to the scope of protection of the present application.

Claims

1. A knowledge graph-based multi-recommendation system for ship repair, characterized in that, This includes the data layer, engine layer, and application layer; The data layer uses the Neo4j graph database as its core, storing a knowledge graph containing ship entities, repair project entities, and semantic relationships. The knowledge graph contains static relationships between ship entities and repair project entities, and also realizes dynamic association of knowledge through a semantic relationship network. The engine layer includes a rule package engine, a statistics package engine, an intelligent package engine, and an operation and maintenance package engine. The rule package engine transforms industry standards into machine-readable knowledge and achieves compliance matching through Cypher queries. The statistics package engine mines high-frequency repair patterns based on historical data according to specified dimensions. The intelligent package engine integrates multiple recommendation algorithms and supports model switching. The operation and maintenance package engine responds to the dynamic and real-time maintenance needs of ships. The application layer provides standardized API interfaces to support intelligent generation of repair orders and query of recommended results.

2. The knowledge graph-based multi-recommendation system for ship repair according to claim 1, characterized in that, Each rule package in the rule package engine serves as a node in the knowledge graph. Its condition_config attribute stores the structured matching conditions corresponding to the inspection type and ship type in JSON format. By retrieving the rule packages in the active state and matching the request parameters, the compliance items that must be executed are determined.

3. The knowledge graph-based multi-recommendation system for ship repair according to claim 1, characterized in that, The statistical dimensions of the statistical package engine include statistical units and statistical years. By statistically analyzing the number of times each standard repair item is executed within a specific group, a top-K number of high-frequency items are selected to form a statistical package.

4. The knowledge graph-based multi-recommendation system for ship repair according to claim 1, characterized in that, The recommendation algorithms integrated by the intelligent package engine include collaborative filtering (CF) and graph attention network (GAT). The collaborative filtering algorithm includes singular value decomposition (SVD) and nonnegative matrix factorization (NMF) algorithms. The algorithm model can be switched through global system parameters, with NMF algorithm enabled by default.

5. The knowledge graph-based multi-recommendation system for ship repair according to claim 1, characterized in that, The project sources for the maintenance package engine include routine inspection legacy items, early warning suggestions generated from real-time ship status monitoring data, and special items supplemented by experts on-site. The system continuously tracks high-frequency maintenance items and expands the standard project library as candidates.

6. The knowledge graph-based multi-recommendation system for ship repair according to claim 1, characterized in that, The engine layer adopts a "compliance-first, hierarchical integration" strategy. The recommendation results of the rule package engine have absolute priority, while the statistics package, intelligent package, and operation and maintenance package engines are sorted according to business importance, and the recommendation results are optimized through a deduplication mechanism.

7. A knowledge graph-based multi-recommendation method for ship repair, characterized in that, The method is implemented using the knowledge graph-based multi-recommendation system for ship repair as described in any one of claims 1-6, and includes the following steps: Step 1: Construct a ship repair knowledge graph and store it in the Neo4j graph database. The knowledge graph includes ship, repair project entities and semantic relationships, and is associated with industry standards and historical repair data. Step 2: Receive the request parameters for repair category and inspection type input by the user; Step 3: Invoke the rule package engine, statistics package engine, intelligent package engine and operation and maintenance package engine in parallel to generate compliance recommendation projects, high-frequency pattern recommendation projects, personalized recommendation projects and dynamically supplemented recommendation projects respectively; Step 4: Collect and deduplicate the recommended items output by each engine, and merge the source information of the items; Step 5: Present the deduplicated recommendation results to the user through the application layer API interface.

8. The knowledge graph-based multi-recommendation method for ship repair according to claim 7, characterized in that, In step 1, when constructing the knowledge graph, the "Rules for Inspection of Inland Waterway Vessels" are transformed into a machine-readable knowledge representation. At the same time, historical repair documents, vessel technical files, and inspection and maintenance records are structured and organized to form a semantic relationship network.

9. The knowledge graph-based multi-recommendation method for ship repair according to claim 7, characterized in that, In step 3, the rule package engine dynamically generates matching logic through Cypher queries, the statistics package engine groups statistics by "statistical unit + statistical year", the intelligent package engine selects the corresponding algorithm based on data sparsity, and the operation and maintenance package engine responds to unplanned situations in real time.

10. The knowledge graph-based multi-recommendation method for ship repair according to claim 7, characterized in that, In step 4, when the same project is recommended by multiple engines at the same time, a single entry is retained and multiple sources are marked. The deduplication process is performed by matching the project name precisely. It also includes a dynamic weight optimization step: as user interaction data accumulates, the optimal dynamic weight is assigned to the recommendation results of each engine through offline simulation calculation, realizing the evolution from "comprehensive display" to "precise ranking".

Images