Port knowledge graph construction system based on large language model

Abstract

The invention relates to the technical field of port knowledge graph construction, in particular to a port knowledge graph construction system based on a large language model. Comprising the following modules: a port multi-dimensional data dynamic acquisition and intelligent label generation module, a heterogeneous data fusion preprocessing module, a situational entity relationship dynamic extraction and knowledge modeling module, an efficiency evaluation index and model module, and an intelligent label generation module. The system has the beneficial effects that various kinds of heterogeneous data inside and outside a port can be integrated in a unified manner, and a comprehensive knowledge graph can be formed; automatic conflict detection and resolution can be carried out on different data source knowledge, and dynamic updating of a graph is supported; a multi-round interactive decision support function can be realized; a newly developed algorithm can be integrated into the platform and verified in the atlas environment.

Description

Technical Field

[0001] The present invention relates to the technical field of port knowledge graph construction, and specifically to a port knowledge graph construction system based on a large language model. Background Art

[0002] With the rapid development of globalization and international trade, ports, as important hubs in the global logistics network, are becoming increasingly complex in terms of management and operation. Traditional port management systems rely on manual data entry and decentralized information systems, resulting in inefficient data acquisition and processing, difficulties in information sharing, and affecting port operational efficiency and scientific decision-making. Port operations involve a variety of heterogeneous data sources, including logistics, equipment, weather, and shipping, and data diversity and unstructured characteristics further increase processing complexity. Given that about 90% of global trade is carried out by sea, port operational efficiency is directly related to the stability and efficiency of the global supply chain. In response to the growth in international trade demand, especially after the epidemic, big data and artificial intelligence technology have gradually been introduced into port management to improve operational efficiency, promote port modernization, and promote sustainable development of global trade. This transformation will not only help enhance the competitiveness of my country's ports, but also contribute positively to global economic recovery and development.

[0003] The development of big language models brings new opportunities for port knowledge management and intelligent operation. By combining multi-source heterogeneous data fusion and knowledge graph construction, big language models can not only deeply understand the complex data of ports, but also support the reasoning and optimization of port business processes. However, port management still faces severe challenges, including heterogeneity of data sources, real-time requirements for data processing, and high-quality extraction and application of knowledge. Specifically, there are the following technical defects:

[0004] (1) Insufficient ability to process complex semantics: Existing technologies are obviously insufficient in processing complex semantic relationships and unstructured data. Traditional knowledge graph construction methods often rely on rules and templates, which makes it difficult to accurately understand and process complex semantics in multiple fields and languages, which limits their application in port business scenarios.

[0005] (2) Limited real-time update capability: The construction and update process of the existing knowledge graph is usually slow and cannot respond to real-time data changes in port operations in a timely manner. This delay causes information lag in the knowledge graph, making it difficult to meet the needs of dynamic business scenarios.

[0006] (3) Insufficient performance in large-scale data processing: Existing knowledge graph construction technologies usually exhibit low computing efficiency and excessive consumption of storage resources when facing large-scale data in the port field. This technical bottleneck leads to slow system response when processing massive data, making it difficult to achieve efficient knowledge extraction and analysis.

[0007] (4) Difficulty in integrating multi-source data: Existing technologies have problems with poor compatibility and difficulty ensuring data consistency when integrating and processing data from different data sources. This results in incomplete information or data conflicts in the knowledge graph after data integration, affecting the overall effectiveness of the system.

[0008] Based on the above problems, a port knowledge graph construction system based on a large language model is proposed. Summary of the invention

[0009] In order to overcome the technical defects in the prior art and effectively solve the technical problems in the background technology, the present invention provides the following technical solutions:

[0010] A port knowledge graph construction system based on a large language model includes the following modules: dynamic acquisition of port multi-dimensional data and intelligent label generation module, heterogeneous data fusion preprocessing module, contextual entity relationship dynamic extraction and knowledge modeling module, efficiency evaluation index and model module, adaptive knowledge graph fusion and conflict resolution module, and graph visualization and multi-round interactive intelligent decision support module:

[0011] The port multi-dimensional data dynamic collection and intelligent label generation module is responsible for collecting various data inside and outside the port in real time to ensure the timeliness and accuracy of the data, while providing contextual information for the data to improve the availability of subsequent analysis;

[0012] The heterogeneous data fusion preprocessing module cleans, fuses and extracts features from the collected data to ensure data consistency and high quality;

[0013] The contextual entity relationship dynamic extraction and knowledge modeling module automatically identifies and extracts entities and their relationships from the data, and uses TF-IDF technology to perform weighted evaluation on words in the text, helping the model to identify high-importance terms in the port text, and finally constructing a port operation knowledge graph to achieve in-depth analysis of complex relationships;

[0014] The efficiency evaluation index and model module is used to define and calculate key indicators of port operation efficiency, use appropriate mathematical models for evaluation, and provide decision support;

[0015] The adaptive knowledge graph fusion and conflict resolution module ensures the dynamic update and consistency of the knowledge graph and resolves conflicts between different data sources;

[0016] The graph visualization and multi-round interactive intelligent decision support module provides decision makers with intuitive visualization tools, enabling them to monitor and analyze port operations in real time, support multi-round interactions, and improve decision flexibility and accuracy.

[0017] As a preferred technical solution of the present invention, the port multi-dimensional data dynamic collection and intelligent label generation module includes multi-dimensional data collection, intelligent label generation, data dynamic update, data preprocessing and cleaning, and multi-source data integration.

[0018] As a preferred technical solution of the present invention, the heterogeneous data fusion preprocessing module includes data source identification and access, data cleaning and preprocessing, feature extraction and selection, data fusion and integration, and data conversion and normalization.

[0019] As a preferred technical solution of the present invention, the contextualized entity relationship dynamic extraction and knowledge modeling module includes entity recognition, relationship extraction, text clustering, knowledge modeling and dynamic updating and reasoning.

[0020] As a preferred technical solution of the present invention, the efficiency evaluation index and model module include efficiency evaluation mathematical model construction and weight adjustment (by adjusting the model weights and reconstructing a new scheduling strategy according to different scene requirements within the model).

[0021] As a preferred technical solution of the present invention, the adaptive knowledge graph fusion and conflict resolution module includes conflict detection and resolution, deduplication and simplification, adaptive update and consistency maintenance.

[0022] As a preferred technical solution of the present invention, the graph visualization and multi-round interactive intelligent decision support module includes graph visualization, multi-round interactive analysis and decision support functions.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. Comprehensive data integration and modeling: This invention builds a port knowledge graph based on a large language model to integrate various heterogeneous data inside and outside the port (such as equipment status, cargo information, environmental data, etc.) to form a comprehensive knowledge graph. This graph can intuitively display the entities and their relationships in port operations, provide managers with a clear data structure and knowledge framework, and facilitate comprehensive monitoring and efficient management of port operations.

[0025] 2. Intelligent conflict resolution and adaptive update: The present invention adopts adaptive knowledge graph fusion and conflict resolution technology, which can automatically detect and resolve conflicts from different data sources and support dynamic update of graphs. The adaptive adjustment mechanism based on machine learning model can optimize the graph structure in real time, ensure that the knowledge graph maintains accuracy and consistency under the condition of continuous data update, and help to timely reflect the actual operation status of the port.

[0026] 3. Multi-round interactive decision support: Based on graph visualization and intelligent question-answering system, the present invention realizes multi-round interactive decision support function. Managers can mine data layer by layer through multi-round dialogue in the system to obtain in-depth analysis and optimization suggestions. This module not only improves the user-system interaction experience, but also helps managers quickly find key information and optimize operational decisions.

[0027] 4. Efficient algorithm testing and optimization platform: The present invention provides an extensible knowledge graph testing platform, where R&D personnel can integrate newly developed algorithms (such as scheduling optimization, path planning, etc.) into the platform and verify them in a graph environment. The system supports real-time updates and multiple iterations, and adjusts algorithm parameters through a closed loop of graph data feedback, thereby significantly improving the verification efficiency and implementation speed of new algorithms. This efficient and low-cost testing method helps to quickly iterate and optimize port management strategies.

[0028] 5. Intelligent conflict resolution and dynamic knowledge update mechanism: The system adopts a weighted comprehensive model to flexibly adjust weight parameters according to port operation needs. For example, in the resource allocation priority scenario, the weight of the data envelopment analysis (DEA) model is increased to improve resource utilization by optimizing input resources (such as equipment quantity, personnel distribution) and output (such as loading and unloading efficiency, cargo transportation volume); in the scheduling optimization scenario, based on the queuing theory efficiency model, the ratio of task arrival rate and service rate is adjusted to alleviate the task backlog bottleneck. The model formula includes the DEA efficiency evaluation formula and the M / M / 1 queuing model. Through continuous monitoring of real-time data (such as equipment status, task completion rate), the system can dynamically adjust the model parameters, such as giving priority to optimizing the immediate efficiency index and reallocating resources when the equipment frequently fails. During operation, the system establishes a feedback loop, identifies inefficient links and optimizes parameters through real-time analysis (such as using the Apache Flink data stream processing framework), and verifies the effect by comparing the indicators before and after the adjustment. For example, in the scenario of low efficiency of quay crane loading and unloading, after adjusting the immediate efficiency weight, the equipment utilization rate is increased by 10%, and the loading and unloading task time is shortened by 20%, thereby significantly improving the port operation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A simplified diagram of the framework of the system for constructing the present invention.

[0030] Figure 2 It is a schematic diagram of the overall coordination scheme of an embodiment of the present invention.

[0031] Figure 3 This is a diagram of the port task management and large model support architecture of the present invention.

[0032] Figure 4 The overall architecture diagram of the system deployment of the present invention.

[0033] Figure 5This is a flowchart of the contextualized entity relationship dynamic extraction and knowledge modeling of the present invention.

[0034] Figure 6 The present invention provides a flow chart for efficiency evaluation indicators and model construction.

[0035] Figure 7 This is a flow chart of the adaptive knowledge graph fusion and conflict resolution of the present invention.

[0036] Figure 8 This is the flowchart of the graph visualization and multi-round interactive intelligent decision-making of the present invention.

[0037] Fig. 9 It is the overall workflow diagram of the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] See also Figure 1-9 , the present invention provides a technical solution: a port knowledge graph construction system based on a large language model:

[0040] 1. The system consists of:

[0041] The port knowledge graph construction system of the present invention consists of six modules, each of which plays an important role in port management and operation.

[0042] First, the port multi-dimensional data dynamic collection and intelligent label generation module is responsible for real-time collection of various data inside and outside the port, including equipment status, transportation conditions, weather information, etc. It also automatically labels the port data in the initial data processing stage to improve information availability and intelligent reasoning capabilities. This module uses the Internet of Things technology and intelligent label generation technology to ensure the timeliness and accuracy of data, while providing contextual information for the data to improve the availability of subsequent analysis.

[0043] Secondly, the heterogeneous data fusion preprocessing module cleans, fuses and extracts features from the collected data to ensure data consistency and high quality. This process is achieved through ETL tools and data quality management technology. The module integrates data from different sources into a unified view, laying a solid foundation for subsequent data analysis and knowledge modeling. In addition, the contextual entity relationship dynamic extraction and knowledge modeling module uses natural language processing and graph neural network technology, and combines conditional random fields (CRF) with deep learning models (such as BERT) to improve entity recognition accuracy, and then uses dependency syntax analysis and triple extraction technology to build semantic associations, support real-time updates and knowledge discovery, and finally automatically identify and extract entities and their relationships from the data, and build a port operation knowledge graph to achieve in-depth analysis of complex relationships.

[0044] Finally, the efficiency evaluation index and model module in the system focuses on defining and calculating key indicators of port operation efficiency, and dynamically optimizes port resource allocation and task scheduling through DEA ​​models, queuing theory models and real-time data feedback mechanisms, improving operational efficiency and management intelligence. The adaptive knowledge graph fusion and conflict resolution module ensures dynamic updates and consistency of the knowledge graph and resolves conflicts between different data sources. The graph visualization and multi-round interactive intelligent decision support module provides decision makers with intuitive visualization tools, enabling them to monitor and analyze port operations in real time, support multi-round interactions, and improve decision flexibility and accuracy.

[0045] Specific:

[0046] (I) Port multi-dimensional data dynamic collection and intelligent label generation module:

[0047] 1. Multi-dimensional data collection: The module can collect a variety of data inside and outside the port in real time, including cargo transportation, equipment status, weather information, and traffic flow. Data collection is achieved through IoT technologies (such as sensors and RFID), connecting different port equipment (such as cranes, container yard management systems, transport vehicles, etc.) to ensure the timeliness and accuracy of data and form a comprehensive view of port operations. Use API interfaces to integrate different management systems (such as port management systems, transportation management systems) to obtain real-time and historical data. In addition, combined with large model capabilities, it is possible to deeply analyze and predict data trends, such as using machine learning algorithms (such as LSTM or CNN) to identify potential transportation delays or equipment failures. Use data stream processing frameworks such as Apache Kafka or Flink to achieve high-throughput data stream processing.

[0048] 2. Intelligent label generation: According to the characteristics of the collected data, intelligent labels with contextual information are automatically generated. These labels can identify the nature, source and relevance of the data, and contain specific information about port operations (such as equipment type, operation time, cargo characteristics, etc.). Key features in the data are identified and extracted through algorithms such as cluster analysis and principal component analysis. Through data labeling technology, the system can automatically label port data in the preliminary data processing stage, such as identifying key fields and attributes through predefined rules or trained models to improve information availability and intelligent reasoning capabilities. In addition, combined with knowledge graphs, these labels can be linked to existing knowledge to improve information intelligent reasoning capabilities. Finally, by applying natural language processing (NLP) technology, text data is analyzed and valuable information is extracted to enhance the descriptiveness of labels.

[0049] 3. Dynamic data update: The module can automatically update label information according to real-time data changes to ensure data timeliness and accuracy. This dynamic update mechanism makes port operation management more flexible and relies on big data processing frameworks (such as Hadoop or Spark) to support massive data processing. Combined with the adaptive capabilities of large models, data collection strategies can be automatically adjusted in a changing environment to optimize information flow. In addition, edge computing applications can reduce latency and improve real-time response capabilities.

[0050] 4. Data preprocessing and cleaning: During the data collection process, the module will perform preliminary cleaning and formatting of the raw data to remove redundant information and erroneous data and improve data quality. This process enables data interaction with other systems through API interface integration to ensure smooth and efficient data flow. At the same time, by combining with the knowledge graph, data accuracy can be verified in real time, and necessary adjustments can be made based on knowledge base feedback to improve data reliability.

[0051] 5. Multi-source data integration: Integrate data from different systems (such as logistics management systems, environmental monitoring systems, equipment management systems, etc.), use NLP technology to analyze text data, identify and extract relevant information, and achieve cross-system data interconnection. By building a knowledge graph, data from different sources can be semantically associated to improve data interoperability and value. In addition, visualization tools (such as D3.js or Tableau) are used to present the integrated data and labels in a graphical form to support decision makers in real-time monitoring and analysis.

[0052] (II) Heterogeneous data fusion preprocessing module:

[0053] 1. Data source identification and access: The module can identify and access heterogeneous data from multiple data sources, including structured data (such as relational databases), semi-structured data (such as JSON, XML) and unstructured data (such as text, images, etc.). For port data, especially historical data (such as historical transportation records, equipment maintenance records), real-time data (such as real-time sensor monitoring data) and container truck data (such as transportation vehicle scheduling information), the module will realize automatic data access and integration through standardized interfaces (such as REST API or GraphQL) to ensure that the system can flexibly handle diverse data sources. In addition, integrated ETL tools (such as Apache NiFi or Talend) can simplify the data access process and automate data extraction, conversion and loading.

[0054] 2. Data cleaning and preprocessing: Comprehensive cleaning is performed on the access data, including removing redundant and erroneous information, filling missing values, and format conversion to unify the data format. In view of the characteristics of port data, especially the fusion of historical and real-time data, the module will use data quality assessment tools (such as Apache Griffin) to ensure data consistency and accuracy. The process also includes the implementation of a data quality management framework to continuously monitor and evaluate the data to ensure the reliability of subsequent analysis.

[0055] 3. Feature extraction and selection: Extract key features from cleaned data to reduce data dimensions and improve model performance. For port data, especially equipment and truck data, feature selection algorithms (such as random forest, LASSO regression, etc.) are used to identify the most influential features for analysis. This process can help the system focus on the most relevant data, thereby improving the efficiency of subsequent analysis and modeling.

[0056] 4. Data fusion and integration: Through data fusion technology, data from different sources are integrated to form a unified view. For port history, real-time data and equipment data, the module will use semantic matching and relationship mapping technology to ensure semantic consistency and relevance between different data sources. For example, graph databases (such as Neo4j) are used to process complex relational data, and graph algorithms (such as PageRank or community discovery) are used to identify potential relationships between data, thereby enhancing the value of comprehensive data utilization.

[0057] 5. Data conversion and normalization: The fused data is converted and normalized to meet the needs of subsequent analysis. This includes data standardization, normalization, and timestamp unification to ensure comparability between different data formats. For port data, especially equipment and container truck data, the module will ensure that information from different data sources is effectively integrated by applying data fusion algorithms (such as weighted average, Bayesian fusion, or deep learning models) to improve the quality and reliability of the final data set.

[0058] (III) Contextualized entity relationship dynamic extraction and knowledge modeling module:

[0059] 1. Entity Recognition:

[0060] Through natural language processing (NLP) technology, key entities such as port equipment, transport vehicles, and cargo types are extracted from unstructured texts in the port field. First, the TF-IDF technology is used to perform weighted evaluation on the words in the text to help the model identify terms with high importance in the port text. Then, the named entity recognition (NER) algorithm is used to automatically identify specific entities by analyzing contextual information (such as equipment number, cargo type, geographic location, etc.). To further improve the accuracy of entity recognition, the conditional random field (CRF) model and deep learning model (such as BERT) are combined. CRF optimizes the sequence labeling effect by modeling contextual dependencies, while BERT captures more complex contextual relationships through deep semantic understanding, improving the entity recognition capability in specific port scenarios.

[0061] TF-IDF formula:

[0062]

[0063] TF-IDF is a commonly used weighting technique for information retrieval and data mining. TF stands for Term Frequency, IDF stands for Inverse Document Frequency, TF represents the frequency of word t in document d, DF represents the number of documents containing word t, and N is the total number of documents. This formula is used to measure the importance of a word in a specific document and help the model identify the most representative key entities in the text. In the entity recognition stage, the importance of a word in the text is evaluated by calculating the TFIDF value of the word to help the model identify key entities.

[0064] 2. Relationship extraction:

[0065] After identifying the entities, the module analyzes the relationships between entities, such as the relationship between equipment and goods, the relationship between transportation paths, etc. The semantic relationship between entities is extracted through technologies such as dependency syntax analysis and graph convolutional networks, and the graph neural network (GNN) is used to capture the relationship between entities.

[0066] (1) Triplet extraction and annotation:

[0067] Triplet labeling judgment formula:

[0068] Among them, K represents the number of keywords recognized in the current sentence that overlap with the existing keyword list, which represents the total number of keywords recognized in the current sentence. This formula is used to determine whether a triple in a source split sentence needs to be labeled. If p is greater than the preset threshold, the triple is not labeled; if it is less than or equal to the preset threshold, it is labeled. The purpose of this formula is to improve the labeling efficiency and reduce unnecessary labeling operations.

[0069] Vector similarity judgment formula:

[0070]

[0071] Among them, represents the values ​​of two vectors A and B on the i-th feature. represents the weight of each feature. In the port knowledge graph, vectors A and B represent two different triples. Each triple vector can be represented by its involved entity and relationship feature vectors. These vectors can be extracted from the text through a large language model or other embedding methods; feature weights can reflect the importance of each feature (such as a specific attribute or relationship) in port operations. For example, some features may represent key port operations (such as loading and unloading, shipping time), and these features can be given higher weights. Feature weights can be set based on factors such as the importance, frequency, or decision-making impact of port operations. The weighted cosine similarity of two vectors is calculated to determine their similarity, ensuring the accuracy of identifying entities and relationships during the knowledge extraction process.

[0072] 3. Text Clustering

[0073] The pre-processed text data is clustered using machine learning algorithms such as support vector machines (SVM). The purpose of clustering is to classify different types of port information (such as ship information, cargo information, logistics dynamics, etc.) into different knowledge text data sets. The following formulas are mainly used in this module:

[0074] Text similarity calculation formula:

[0075] Among them, are two text vector representations, and the dot product is used to calculate the cosine similarity between the two texts. This formula is used to measure the similarity between two texts so that the texts can be assigned to the corresponding categories based on the similarity.

[0076] Contextual analysis:

[0077] Combining real-time data streams and historical data, the module can dynamically adjust entity and relationship contexts to ensure that the extraction results match the current environment. It improves the accuracy of extraction results through contextual information such as time and location. It uses context-based embedding representations (such as ELMo) to enhance the model's understanding of the situation.

[0078] 4. Knowledge Modeling:

[0079] The extracted entities and relationships are constructed into a knowledge graph to form a structured knowledge model. During the knowledge modeling process, a graph database (such as Neo4j) is used to store and query entities and their relationships, providing flexible data management and analysis capabilities.

[0080] Graph Representation: G=(V,E)

[0081] In the knowledge modeling stage, the knowledge graph infrastructure is constructed, G represents the knowledge graph, V is the entity set, and E is the relationship set.

[0082] 5. Dynamic Update and Reasoning:

[0083] The module supports dynamic updating of knowledge graphs, automatically updating entities and relationships based on the introduction of new data. At the same time, it performs reasoning based on inference engines (such as OWL2, RDF) to discover implicit relationships and new knowledge.

[0084]

[0085] PR(A) represents the importance of node A, d is the damping factor, and C(i) is the out-degree of node i. This formula is used to evaluate the importance of nodes in the knowledge graph to identify key entities and ensure effective management and utilization of knowledge.

[0086] (IV) Efficiency evaluation indicators and model modules:

[0087] This module is used to define key efficiency indicators of port operations, such as equipment utilization, operating time, transportation efficiency and container turnover rate. These indicators are calculated by combining port operation data (such as equipment operating time, cargo transportation information, loading and unloading volume, etc.), which can fully reflect the port operation status and provide decision-making basis for managers.

[0088] 1. Construction of mathematical model for efficiency evaluation:

[0089] In order to more accurately reflect the actual situation of port operations, a weighted comprehensive model is used to integrate different efficiency evaluation models to form an overall efficiency evaluation formula

[0090] E total =w1·E DEA +w2·E instant +w3·E queue

[0091] a. Among them, E total It represents the comprehensive evaluation value of the overall port efficiency, which is used to evaluate the overall performance of port operation efficiency.

[0092] B E DEAThe DEA model is used to calculate relative efficiency values ​​and evaluate the efficiency of resource allocation in different departments or equipment in the port. Its input can be resources such as the number of equipment, the number of personnel, and capital investment, and its output can be cargo handling volume, transportation efficiency, etc. Through the DEA model, port managers can identify the efficiency of resource allocation and provide a benchmark for the overall efficiency of each department.

[0093]

[0094] λ r : Weight parameter, which indicates the relative importance of the output indicator in the overall efficiency. The weight can be set according to the key performance indicators (KPI) of port operations, such as cargo handling volume, transportation efficiency, etc.

[0095] y r : The rrr-th output indicator value, such as the number of container loading and unloading, transportation completion rate, etc., indicates the performance or output obtained by inputting resources.

[0096] λ i : Input indicator weight parameter, indicating the weight of each input contribution to efficiency.

[0097] x i :iii. input indicator value, such as number of equipment, labor input, energy consumption, etc.

[0098] m: The total number of input indicators.

[0099] s: total number of output indicators.

[0100] E instant It is the real-time operating efficiency of equipment such as gantry cranes, reflecting the working condition of the equipment within a given time. N represents the number of tasks completed per unit time (such as the number of loaded and unloaded containers), and T represents the total operating time of the equipment. It is suitable for real-time monitoring of the real-time working efficiency of equipment (such as gantry cranes, quay cranes, etc.), and can quickly reflect the operating condition of the equipment within a specific period of time. For example, when the operating frequency of the gantry crane decreases, managers can quickly identify potential problems or inefficiencies in the equipment through this value.

[0101]

[0102] dE queue It is a scheduling efficiency value based on queuing theory, which is used to measure the relationship between task arrival and equipment service rate and reveal possible scheduling bottlenecks. λ is the arrival rate of tasks per unit time (such as the demand rate for loading and unloading containers), and μ is the equipment service rate (such as the number of containers processed per hour).

[0103]

[0104] 2. Weights w1, w2, w3 adjustment

[0105] a. Resource allocation priority: When the port hopes to optimize the overall resource allocation, the weight of w1 can be increased to let the DEA model results dominate the comprehensive evaluation.

[0106] b. Real-time efficiency first: In scenarios where equipment frequently fails or requires real-time monitoring, the weight of w2 can be increased to ensure that immediate efficiency becomes the basis for decision-making.

[0107] c. Scheduling optimization priority: When the task arrival rate is high or the volatility is large, the w3 weight can be increased to make the scheduling efficiency more influential in the comprehensive evaluation.

[0108] Through the above comprehensive model, the efficiency evaluation index and model module can provide more comprehensive efficiency evaluation results for the port. By dynamically adjusting the weights, the port management can flexibly configure the influence of different models according to actual operational needs to achieve more effective efficiency monitoring and optimization.

[0109] (V) Adaptive knowledge graph fusion and conflict resolution module:

[0110] This module is mainly responsible for maintaining the consistency, simplicity and dynamic adaptability of the knowledge graph. It uses a variety of technical means to achieve knowledge fusion, conflict resolution, duplication removal and consistency maintenance from different sources to ensure the stability and efficiency of the knowledge graph in complex environments.

[0111] 1. Conflict detection and resolution

[0112] a. Similarity calculation: The module uses weighted cosine similarity calculation to detect the similarity of entities and relations from different sources and identify possible conflicts. For example, if the entities in two sources are similar but the attributes are inconsistent, there may be a conflict.

[0113] b. Resolution strategy: Conflict resolution is performed through strategies such as priority, timestamp, and data source credibility:

[0114] i. Priority strategy: Set the priority of data sources, and give priority to retaining data sources with higher credibility.

[0115] ii. Time strategy: For data with high real-time requirements, you can choose the latest update.

[0116] iii. Weighted strategy: Assign different weights according to the importance of feature business and automatically select the optimal conflict resolution method.

[0117] 2. Deduplication and simplification

[0118] a. Triple deduplication: In the knowledge graph, the weighted cosine similarity formula is used to determine the similarity between triples. If the similarity is higher than the threshold, the triple is considered to be repeated, and deduplication is performed to ensure the simplicity of the knowledge graph. Weighted cosine similarity is used for relation extraction in the contextualized entity relationship dynamic extraction module, and is used for duplicate triple detection in this module.

[0119] b. Weight configuration: When deduplicating, weights are configured based on feature importance. For example, shipping routes or time similarity may have higher weights, while non-critical attributes (such as manufacturer information) may have lower weights.

[0120] 3. Adaptive Update

[0121] a. Dynamic update: The module uses incremental update technology to intelligently adjust the graph structure when new data comes in. Through periodic detection, it ensures that only the affected nodes and edges are adjusted during the update, improving processing efficiency.

[0122] b. Machine learning prediction: The module uses LSTM (Long Short-Term Memory Network) as a prediction model to analyze data change trends and automatically adjust conflict resolution and deduplication rules. The LSTM model is particularly suitable for nonlinear and dynamically changing port business environments, and can optimize rules based on historical trends to ensure the adaptability and accuracy of the knowledge graph.

[0123] 4. Maintain consistency

[0124] a. Logical consistency check algorithm: The module uses logical rule check algorithms such as the transitive closure algorithm to ensure that the relationships within the graph conform to the logical consistency rules. For example, the transitive closure algorithm is used to maintain transitive relationships (such as A is related to B, B is related to C, then A is related to C), and logical rules such as symmetry and antisymmetry are verified.

[0125] b. Graph structure optimization and automatic repair: Regularly run graph structure optimization algorithms (such as the shortest path algorithm) to ensure that the graph has no redundant paths or isolated nodes. When a logical conflict is detected (such as non-compliance with antisymmetric relationships), it is automatically repaired according to the rules to ensure the integrity and logical coherence of the knowledge graph.

[0126] (VI) Graph visualization and multi-round interactive intelligent decision support module

[0127] The module aims to provide intelligent decision support for port managers through intuitive graph display and multi-round interactive analysis. This module can not only present the knowledge graph information structure in a visual way, but also support managers to interact with the system in multiple rounds, helping to mine deep information and optimize port business decisions.

[0128] 1. Graph Visualization

[0129] a. Knowledge graph display: By graphically displaying entities, relationships and attributes, managers can intuitively view the associations and relationship paths between different entities (such as ships, cargo, equipment, etc.). Using front-end visualization tools such as D3.js, the knowledge graph is presented in a node-edge manner.

[0130] b. Information hierarchical presentation: Supports information display at different levels, and managers can quickly switch from a global view to local details. The module allows certain relationship types or entity categories to be displayed or hidden as needed, so that managers can focus on key nodes and core relationships.

[0131] c. Dynamically update the graph: As the knowledge graph content is updated, the graph display will automatically refresh. Real-time updates ensure that managers can always make decisions based on the latest data.

[0132] 2. Multi-round interaction analysis

[0133] a. Intelligent question-answering function: The module integrates intelligent question-answering function, which can retrieve relevant information from the graph according to the user input question and answer it in a concise and clear form. Based on the large language model NLP technology, the module can understand complex questions and perform multiple rounds of information query and path search in the graph.

[0134] b. Multiple rounds of dialogue and information tracking: Managers can gradually explore the information in the graph through multiple rounds of interaction with the system. For example, users can gradually refine their queries, from "viewing the cargo transportation route" to "analyzing specific equipment transportation tasks", and the system will automatically record the conversation context to ensure that each query step is based on the previous operation.

[0135] c. In-depth analysis and recommendation: The module can identify the needs of decision makers during the interaction process and provide optimization suggestions for specific tasks or scenarios based on historical data and relationships in the graph. For example, for equipment scheduling problems, the system can recommend efficient equipment configuration solutions or task allocation strategies.

[0136] 3. Decision support function

[0137] a. Real-time data analysis and early warning: The module integrates real-time data flow analysis function. When certain indicators (such as equipment utilization, transportation efficiency) approach or exceed the preset threshold, the system will generate an early warning and indicate potential risks. This helps managers intervene in time to avoid business interruption or waste of resources.

[0138] b. Scenario simulation and deduction: Supports "hypothesis-result" scenario simulation function. Managers can set different scenario variables in the system (such as task load changes, resource allocation adjustments, etc.) and observe their impact on port operations. This function helps port management evaluate decision-making effects in advance and optimize actual plans.

[0139] Intelligent recommendation and decision support: The system can provide personalized intelligent recommendations to decision makers in complex and changing business environments. Based on in-depth analysis of the port knowledge graph, the module can identify the optimal path or solution and provide optimization suggestions for scheduling, logistics, resource allocation and other fields to ensure scientific and efficient decision-making.

[0140] 2. The overall architecture of system deployment is as follows:

[0141] The port large model knowledge graph construction system consists of multiple layers and modules, which cooperate with each other through the application layer, collaboration layer, construction layer and support layer to jointly realize the intelligent and efficient port management and operation. The application layer includes multiple functional modules such as entity understanding, text understanding, search recommendation, answering, domain data annotation and knowledge base management. It relies on the knowledge graph to provide basic data and models, and serves specific business applications in the port through AVP retrieval, Type reasoning, intelligent dialogue and complex task guidance, covering the whole process from equipment status monitoring to intelligent operation recommendation. The collaboration layer is the core of the system, providing a variety of models and reasoning capabilities to ensure data flow and functional collaboration between modules at all levels. The knowledge graph deeply supports various operations in the port through prompt engineering, knowledge reasoning interpretation, and construction of specific operation task models, and uses small sample adaptive learning technology to achieve system flexibility to adapt to the ever-changing business scenarios in the port. The construction layer is mainly responsible for the construction and maintenance of the knowledge graph, including data standardization, entity extraction, semi-structured data processing and multi-source data fusion, etc. It relies on high-performance data processing frameworks and databases to achieve rapid processing and integration of large-scale heterogeneous data and ensure the accuracy and timeliness of the knowledge graph. The support layer provides the entire system with infrastructure such as data collection, intelligent crawlers, and graph management, supports the dynamic update and expansion of the port knowledge graph, and ensures the comprehensiveness of data collection, efficient storage, and data access security.

[0142] The system adopts a distributed architecture, and its core consists of multiple modules and services to ensure that the system has good scalability and fault tolerance. The data acquisition and processing module dynamically acquires and preprocesses data through distributed crawlers and ETL tools, and stores the processed data in a graph database. The collaborative layer uses high-performance computing nodes to run various reasoning tasks and machine learning models to provide real-time support for the application layer. The application layer is connected to the user interface through a RESTful API to implement various port business applications. The system isolates and manages each module through containerization technology (such as Docker and Kubernetes) to ensure flexible deployment and expansion in different port environments, and designs and improves the monitoring and operation and maintenance mechanism to monitor the operating status of each module in real time to ensure high reliability and availability of the system.

[0143] 3. The system workflow is as follows:

[0144] Step 1: Dynamic collection of multi-dimensional data and generation of intelligent labels

[0145] 1. Data Collection

[0146] (1) Historical data collection

[0147] Historical data is used for long-term trend analysis and forecasting, and mainly comes from various port management systems, including:

[0148] a. Port Management System (PMS): records long-term cargo entry and exit, loading and unloading history, equipment maintenance, resource allocation and task scheduling data.

[0149] b. Equipment Management System (EMS): covers information such as equipment work logs, historical fault records, maintenance records, etc., which helps to analyze equipment performance and predict future maintenance needs.

[0150] c. Logistics Management System (LMS): Provides historical records of goods in and out, storage location changes, transportation plans, etc., to facilitate understanding of historical trends in goods flow.

[0151] (2) Real-time data collection

[0152] Real-time data supports dynamic port operations, coming from a variety of smart devices and IoT systems:

[0153] a. Sensor equipment: installed on various types of port equipment (such as cranes, forklifts, loading and unloading equipment) to monitor the equipment's working status, temperature, vibration, etc. in real time.

[0154] b. RFID reader: used to track the location of containers and cargo in real time. By scanning RFID tags, the port management system can locate the real-time location of containers and cargo in the yard or port.

[0155] c.GPS positioning system: widely used to track trucks and cargo transportation routes, providing real-time location information to understand the exact location of cargo and vehicles.

[0156] d. Video surveillance system: installed in key areas (such as storage yards, docks, and port entrances), it uses image analysis to obtain real-time dynamics of people, vehicles, and equipment in the scene. Combined with computer vision technology, it can identify cargo and vehicle queues and potential safety hazards.

[0157] (3) Truck data collection

[0158] The dispatching information of container trucks (container trucks in ports) is very important, and its data mainly comes from the following equipment and systems:

[0159] a. Vehicle Dispatch System (VDS): records the task status (such as loading, unloading, standby), driving route, empty and full status of each truck in real time.

[0160] b. Vehicle-mounted GPS and sensors: GPS devices are installed on container trucks to provide real-time positioning information, and vehicle sensors monitor fuel consumption, driving speed, vehicle condition, etc., so that the port can control the status and location of each vehicle in real time when dispatching container trucks.

[0161] c. Road monitoring and navigation system: Utilize road cameras and traffic light status analysis to monitor the driving routes of container trucks and optimize traffic flow within the port area.

[0162] 2. Intelligent label generation

[0163] (1) Based on the collected data, the system generates intelligent tags to more effectively describe and use the data in the knowledge graph. Tag generation is divided into the following steps:

[0164] a. Label key features: Extract key attributes such as equipment status, cargo type, and operation time from the collected data, and assign contextual labels, such as "cargo type: refrigerated cargo" and "equipment status: fault".

[0165] b. Data layer annotation: According to the data source, the data layer (such as historical data, real-time data, sensor data) and device or vehicle ID are automatically annotated to ensure the consistency of data source and time context.

[0166] c. Dynamic update: The system adjusts the label in real time according to the port dynamics to keep the label information up to date. Edge computing and other technologies are used to reduce data transmission delays and achieve real-time label updates.

[0167] Step 2: Heterogeneous data fusion and preprocessing

[0168] 1. Data cleaning and normalization

[0169] After data collection, it needs to be cleaned and normalized to ensure data quality and consistency.

[0170] a. Redundant and erroneous data processing: Data obtained from different devices and systems may be redundant or inconsistent. The system uses data cleaning algorithms (such as anomaly detection) to eliminate invalid data, such as repeated device status updates or erroneous sensor readings.

[0171] b. Format standardization: Data formats may be inconsistent across devices and systems. Through unified formatting operations (such as date and time format standardization), data can be standardized to ensure consistency in subsequent processing.

[0172] 2. Feature extraction and selection

[0173] The system extracts key features from the cleaned data to support efficient knowledge graph construction. Feature extraction methods include:

[0174] a. Feature extraction algorithm: cluster analysis, principal component analysis and other algorithms are used to extract key features such as transportation efficiency and queue waiting time for truck data. For equipment data, time series analysis is used to extract features such as operating frequency and failure rate.

[0175] b. Feature selection: Identify the most relevant features through algorithms such as random forest and LASSO, reduce unnecessary data dimensions, and ensure efficient system processing.

[0176] 3. Data fusion and integration

[0177] The fused data forms a unified view and is applied to knowledge graph construction:

[0178] a. Semantic mapping: Use NLP technology to perform semantic matching on different data sources and unify synonymous entity names (such as "container" and "cargo box").

[0179] b. Data fusion technology: Combine real-time data and historical data, and use algorithms such as weighted average or Bayesian fusion to ensure the quality and consistency of the fused data.

[0180] Step 3: Dynamic extraction of contextual entity relationships and knowledge modeling

[0181] In this step, the system uses natural language processing (NLP) and deep learning technology to identify key entities, extract relationships, and build a knowledge graph from various data sources in the port.

[0182] 1. Entity Recognition

[0183] (1) Named Entity Recognition (NER)

[0184] Natural language processing techniques (such as BERT, RoBERTa, etc.) are used to automatically identify key entities in port operations from unstructured text, including ships, equipment (such as cranes, quay cranes), transportation routes, locations, etc. For example, BERT can capture text context information to identify unique entities such as "container number" and "equipment number".

[0185] (2) Contextual analysis

[0186] The system uses conditional random fields (CRF) and context embedding models to identify the specific meaning of entities in specific business scenarios. Through semantic understanding, data is annotated at multiple levels, such as determining the current operating status (loading or unloading) of a "crane" or the current load status (empty or fully loaded) of a "container truck" to obtain more accurate entity information.

[0187] 2. Relation extraction and triple generation

[0188] (1) Dependency parsing

[0189] The system parses the semantic structure in the text through dependency syntactic analysis and identifies the semantic relationship between entities. For example, in the description of "the container truck transports the container to the dock", the system can accurately identify the relationship between "container truck" and "container" as "transportation" and further associate it with the location of "dock".

[0190] (2) Graph Convolutional Network (GNN) Modeling

[0191] The system models complex entity relationships through graph convolutional networks. For example, through GNN, the system can model the operation process of "collecting trucks-transporting-container-unloading" as a coherent relationship path and expand the triples according to the business context.

[0192] (3) Triple generation and storage

[0193] The system extracts each relationship into triples (entity-relationship-entity), such as "truck-load-container", "crane-maintenance-date", etc., and gradually builds them into knowledge graph nodes and edges.

[0194] 3. Knowledge Modeling

[0195] (1) Graph database construction

[0196] Use graph databases (such as Neo4j) to build and manage knowledge graphs, mapping triples to nodes and edges. The knowledge graph contains entity sets (such as equipment, vehicles, and goods) and relationship sets (such as transportation, loading and unloading, and maintenance), providing a basis for subsequent analysis.

[0197] (2) Entity and relationship hierarchical modeling

[0198] In order to meet the complex business needs of ports, the system models the knowledge graph in layers. For example, the "equipment" node is further subdivided into types (such as cranes, quay cranes, and automated guided vehicles), and the "transportation" relationship is divided into operational levels (such as loading, unloading, transshipment, etc.), so as to achieve in-depth analysis of different scenarios.

[0199] Step 4: Efficiency evaluation indicators and model construction

[0200] In this step, the system builds multiple efficiency evaluation models based on the collected port operation data to provide real-time and comprehensive port operation efficiency feedback.

[0201] 1. Definition of key efficiency indicators

[0202] (1) Equipment utilization

[0203] The utilization rate is calculated based on the equipment working time and standby time to measure the actual use effect of the equipment. The system collects the working time, idle time and maintenance time of each equipment, and generates key indicators such as "crane daily utilization rate" and "automatic guided vehicle load rate".

[0204] (2) Container turnover rate

[0205] By counting the time each container enters and leaves the port, the average turnover time is calculated and the turnover rate is defined. This indicator is used to evaluate the efficiency of the port in handling cargo and help management evaluate the overall turnover speed of the port.

[0206] (3) Truck dispatch efficiency

[0207] The system evaluates the efficiency of container truck use by analyzing data such as container truck routes, transportation volume, and frequency of completed tasks. This indicator can help identify container truck scheduling bottlenecks and underutilization of resources within the port.

[0208] 2. Mathematical model construction and evaluation

[0209] (1) Data Envelopment Analysis (DEA) Model

[0210] Based on the DEA model, relative efficiency is calculated through multiple inputs (such as the number of equipment and labor) and multiple outputs (such as cargo handling volume and loading and unloading frequency). The DEA model helps management compare the efficiency of various resource allocations and identify efficient and inefficient equipment or departments.

[0211] (2) Queuing theory model

[0212] The queuing theory model is used to analyze the queueing of trucks and cargo at different operation links to determine the port scheduling efficiency and processing capacity. The model parameters include the task arrival rate (such as the arrival rate of truck tasks) and the service rate (such as the equipment processing rate). By calculating the system bottleneck (such as the backlog of tasks during peak hours), it provides a basis for port optimization.

[0213] (3) Weighted comprehensive model

[0214] The system weights the above indicators to form an overall efficiency evaluation score. Management can adjust the weights (such as equipment utilization, scheduling efficiency) according to different needs to adapt to real-time decision-making and long-term planning.

[0215] Step 5: Adaptive knowledge graph fusion and conflict resolution

[0216] The knowledge graph needs to maintain consistency as data is constantly updated. The system uses adaptive technology to achieve dynamic fusion and conflict resolution of the knowledge graph.

[0217] 1. Conflict detection and resolution

[0218] (1). Weighted cosine similarity detection

[0219] The system calculates the similarity of entities and relationships from different sources through weighted cosine similarity to identify and detect potential conflicts. For example, for the same device status record (such as device status data at different times), the system will judge its source credibility and data freshness.

[0220] (2) Conflict resolution strategies

[0221] The system resolves conflicts based on multiple strategies to ensure the consistency of graph data:

[0222] a. Priority strategy: prioritize data sources and retain data with high credibility first.

[0223] b. Time strategy: In data scenarios with high real-time requirements, the latest data is retained first, such as real-time updates of device status changes.

[0224] c. Weighted strategy: Set different weights according to specific features (such as data importance) to achieve adaptive conflict resolution.

[0225] 2. Deduplication and simplification

[0226] (1) Deduplication of triples

[0227] Use weighted similarity to detect and clean up duplicate triplets. If the similarity between two triplets is higher than the set threshold, the system considers them to be duplicates and removes them to reduce graph redundancy.

[0228] (2) Weight configuration

[0229] The system sets weights based on the importance of data features. For example, a higher weight is given to key data fields (such as transportation routes) and a lower weight is given to unimportant fields (such as manufacturer information) to ensure that key information is retained first.

[0230] 3. Dynamic Updates

[0231] (1) Incremental update

[0232] The system uses incremental update technology to improve update efficiency by only adjusting the affected nodes and edges when new data flows in.

[0233] (2) Machine learning prediction

[0234] The system uses deep learning models such as LSTM to predict data trends, thereby automatically optimizing graph update rules. For example, for frequent equipment data in port operations, the system can automatically adjust conflict detection and deduplication rules based on historical trends.

[0235] Step 6: Graph visualization and multi-round interactive intelligent decision support module workflow

[0236] This module provides intelligent decision-making support for port managers through knowledge graph visualization and multi-round interactions.

[0237] 1. Graph Visualization

[0238] (1) Graphical display

[0239] The system displays the knowledge graph through a graphical interface, and users can intuitively view the relationship between entities. For example, the system displays the "truck-transport-container" relationship in the form of nodes and edges, allowing users to obtain the operation process and relationship path at a glance.

[0240] (2) Hierarchical presentation

[0241] The system supports hiding or displaying different levels of information as needed. For example, managers can view the overall transportation situation in a global view, or click on a node (such as a specific device) to obtain more detailed operation records.

[0242] (3) Dynamic graph update

[0243] After the map content is updated, the system automatically refreshes and displays the latest data to ensure that managers always obtain the most timely information.

[0244] 2. Multi-round interaction analysis

[0245] (1). Intelligent question-answering function

[0246] The system integrates intelligent question-answering functions based on natural language processing. Users can directly enter questions on the interface (such as "collection truck task status"), and the system automatically parses the questions and retrieves the corresponding entity relationships in the knowledge graph, and answers them in a concise and clear form.

[0247] (2) Multi-round conversation tracking

[0248] Managers can gradually explore the information in depth, and the system will record the context of each query. For example, from "querying the transportation route of a certain cargo" to "the time when the cargo is unloaded on the crane", the system automatically updates the context according to each query step, ensuring that each search is based on the results of the previous step.

[0249] (3). In-depth analysis and recommendations

[0250] The system can provide optimization suggestions based on the manager's interaction history and current scenario. For example, in terms of equipment scheduling, the system can recommend the optimal equipment configuration plan based on real-time data and graph relationships, or propose alternative solutions for transportation delays.

[0251] 3. Decision support function

[0252] (1) Real-time monitoring and early warning

[0253] The system integrates real-time data flow analysis functions and sets early warning thresholds. For example, when equipment utilization or transportation efficiency exceeds the preset threshold, the system will automatically generate early warning notifications to indicate potential operational risks and help managers make timely adjustments.

[0254] (2) Situation simulation and deduction

[0255] Supports "hypothesis-result" scenario simulation function. Managers can set different variables in the system (such as task load, resource allocation), and the system simulates the operational impact under different scenarios to help managers evaluate the effect of decisions in advance.

[0256] (3) Intelligent recommendation and decision support

[0257] The system provides decision suggestions based on in-depth analysis of the graph, such as recommending the best strategy when dispatching vehicles within the port and optimizing container routes.

[0258] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A port knowledge graph construction system based on a large language model, characterized by: It includes the following modules: dynamic collection of multi-dimensional data of ports and intelligent label generation module, heterogeneous data fusion preprocessing module, contextual entity relationship dynamic extraction and knowledge modeling module, efficiency evaluation index and model module, adaptive knowledge graph fusion and conflict resolution module, and graph visualization and multi-round interactive intelligent decision support module: The port multi-dimensional data dynamic collection and intelligent label generation module is responsible for collecting various data inside and outside the port in real time to ensure the timeliness and accuracy of the data, while providing contextual information for the data to improve the availability of subsequent analysis; The heterogeneous data fusion preprocessing module cleans, fuses and extracts features from the collected data to ensure data consistency and high quality; The contextual entity relationship dynamic extraction and knowledge modeling module automatically identifies and extracts entities and their relationships from data, constructs a port operation knowledge graph, and enables in-depth analysis of complex relationships; The efficiency evaluation index and model module is used to define and calculate key indicators of port operation efficiency, use appropriate mathematical models for evaluation, and provide decision support; The adaptive knowledge graph fusion and conflict resolution module ensures the dynamic update and consistency of the knowledge graph and resolves conflicts between different data sources; The graph visualization and multi-round interactive intelligent decision support module provides decision makers with intuitive visualization tools, enabling them to monitor and analyze port operations in real time, support multi-round interactions, and improve decision flexibility and accuracy.

2. The port knowledge graph construction system based on a large language model according to claim 1 is characterized by: The port multi-dimensional data dynamic collection and intelligent label generation module includes multi-dimensional data collection, intelligent label generation, data dynamic update, data preprocessing and cleaning, and multi-source data integration.

3. The port knowledge graph construction system based on a large language model according to claim 1 is characterized in that: The heterogeneous data fusion preprocessing module includes data source identification and access, data cleaning and preprocessing, feature extraction and selection, data fusion and integration, and data conversion and normalization.

4. The port knowledge graph construction system based on a large language model according to claim 1 is characterized by: The contextualized entity relationship dynamic extraction and knowledge modeling module includes entity recognition, relationship extraction, text clustering, knowledge modeling, and dynamic updating and reasoning.

5. The port knowledge graph construction system based on a large language model according to claim 1 is characterized by: The efficiency evaluation index and model module includes efficiency evaluation mathematical model construction and weight adjustment.

6. The port knowledge graph construction system based on a large language model according to claim 1 is characterized by: The adaptive knowledge graph fusion and conflict resolution module includes conflict detection and resolution, deduplication and simplification, adaptive updating and consistency maintenance.

7. The port knowledge graph construction system based on a large language model according to claim 1 is characterized by: The graph visualization and multi-round interactive intelligent decision support module includes graph visualization, multi-round interactive analysis and decision support functions.

Images