Navigation aided decision-making method based on large model

Abstract

The invention relates to a navigation aid decision-making method based on a large model, which comprises the following steps of: constructing a scene knowledge model of real-time information such as intelligent navigation situation awareness, situation cognition and intelligent navigation design operation conditions, and designing entities, attributes and relationships between the scene knowledge model and navigation knowledge to form a navigation knowledge aid decision-making graph mode; collecting and preprocessing navigation knowledge data, and constructing a user-defined navigation knowledge data set; performing fine tuning on the pre-trained general large model through the knowledge graph extraction data set, and constructing a knowledge graph knowledge extraction large model; performing fine tuning on the pre-trained general large model through the knowledge question-answer fine tuning data set, and constructing a navigation knowledge question-answer large model; constructing a navigation knowledge aided decision-making knowledge graph; based on the navigation knowledge question and answer large model and the navigation knowledge aid decision knowledge graph, constructing a navigation aid decision generation model based on reinforcement learning; and designing the navigation aid decision-making device based on the intelligent voice interaction mode. The method can provide auxiliary decision support for ship navigation.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence large models, and in particular relates to a navigation auxiliary decision-making method based on a large model Background Art

[0002] With the continuous development of intelligent navigation technology and the increasing intelligence of ships, in order to meet the needs of intelligent navigation in different coastal and inland waters, as well as the application requirements of different navigation business and mission operation scenarios, it is necessary to tap into the navigation knowledge required for the current scenario to provide a professional basis for decision-making support. Due to the complexity and lack of regularity of the navigation knowledge system, and the different navigation rules under the jurisdiction of different waters and regions, it is difficult for professionals to extract entities, establish relationships, and reason about knowledge from the vast amount of navigation knowledge. Manually extracting the required navigation knowledge from the vast amount of navigation knowledge and establishing a knowledge base would be a tedious and huge project, and it requires a threshold of professional navigation knowledge. How to automatically and efficiently build the navigation knowledge database required for intelligent navigation decision-making support and automatically extract the navigation knowledge based on the perceived information in the current scenario is of great research significance.

[0003] To achieve intelligent navigation decision-making support for ships, simply having a navigation knowledge base is insufficient. It requires integrating the navigation knowledge base, nautical chart databases, and real-time situational awareness scenario information. Using a large text-based generative model, it can generate navigation decision-making support recommendations that conform to navigation knowledge rules. However, existing general-purpose pre-trained large models lack specialized applicability for ship navigation in the maritime field, and they also suffer from shortcomings such as hallucinations in generative reasoning. Improving the accuracy of large-scale model-generated decision-making support is crucial.

[0004] Existing voice interaction technology is widely used in areas such as smart homes and autonomous driving. However, in the field of marine navigation, it is still mainly based on simple voice broadcasts. Therefore, when selecting a large-scale model-assisted decision-making application, active or passive voice interaction methods can be designed to provide ship operators with navigation assistance decision-making suggestions that meet navigation knowledge rules in the current scenario. Summary of the Invention

[0005] In response to the above technical problems, the present invention proposes a navigation auxiliary decision method based on a large model.

[0006] The above-mentioned purpose of the present invention is achieved through the following technical solutions:

[0007] A navigation decision-making assistance method based on a large model includes the following steps:

[0008] Step 1: Based on entity, attribute, and relationship triples, a scenario knowledge model for real-time information such as intelligent navigation situation awareness, situation cognition, and intelligent navigation design and operation conditions is constructed. Based on entity, attribute, and relationship triples, a navigation knowledge model is designed. The navigation knowledge model is then connected to the navigation scenario knowledge model through attributes and relationships. The navigation knowledge includes navigation rules, collision avoidance rules, and typical navigation terminology, forming a navigation knowledge decision-making support graph model.

[0009] Step 2: Collect and preprocess the navigation knowledge data to build a custom navigation knowledge dataset;

[0010] Step 3: Label the custom navigation knowledge dataset constructed in step 2, construct a knowledge graph extraction dataset, fine-tune the pre-trained general large model through the knowledge graph extraction dataset, and construct a knowledge graph knowledge extraction large model;

[0011] Step 4: Based on the custom navigation knowledge dataset constructed in step 2 and the designed navigation knowledge question and answer sentence structure, first construct a navigation knowledge question and answer fine-tuning dataset; then fine-tune the pre-trained general large model using the knowledge question and answer fine-tuning dataset to construct a navigation knowledge question and answer large model;

[0012] Step 5: Use the knowledge extraction model to extract entity, attribute, and relationship triples from the collected navigation knowledge documents. Disambiguate and align entities from different sources, unify different entity representations, and unify the representations of real-time variable names and graph entity names in the intelligent navigation system. The final knowledge graph is stored in Neo4j to build a knowledge graph for navigation knowledge-assisted decision-making.

[0013] Step 6: Based on the navigation knowledge question-answering model constructed in step 4 and the navigation knowledge decision-making assistance knowledge graph constructed in step 5, a navigation decision-making assistance generation model based on reinforcement learning is constructed.

[0014] Step 7. Design a navigation assistance decision-making device based on intelligent voice interaction, which includes a data acquisition unit, a command sending unit, a voice acquisition unit, a voice playback unit, a processor and a power supply; the navigation assistance decision-making generation model constructed in Step 6 is deployed in the processor to realize active / passive intelligent voice interaction of navigation assistance decision-making based on user voice input or real-time navigation, perception and other scene data input.

[0015] Moreover, in step 1, the scene knowledge model includes information such as: ship objects, object objects, water environment, ship attribute characteristics, object attribute characteristics, water environment attribute characteristics, the spatiotemporal relationship between ships and waters, the spatiotemporal relationship between ships and ships, the busyness of multiple ships and waters, and the tasks performed by ships; scene reasoning and recognition are achieved by arranging and combining some entities, attributes, and relationships in the scene model.

[0016] Furthermore, step 2 includes:

[0017] 2.1. Collect text data related to ship navigation tasks, including ship maneuvering terms, navigation instrument operating terms, smart ship specifications, coastal water navigation rules, inland water navigation rules, and collision avoidance rules; and collect expert experience and knowledge in the perception, decision-making, control, navigation monitoring, and alarm modules of intelligent navigation / assisted navigation systems;

[0018] 2.2. Preprocess the text data in Word and PDF formats collected in step 2.1, retain the title list, body text, complex tables, and formulas in the navigation knowledge text data, and convert the Word text and PDF text into Markdown format;

[0019] 2.3. Then, the text data is cleaned, noise, duplication and low-quality data are removed, and it is converted into a trainable format to complete the construction of the customized navigation knowledge dataset.

[0020] Furthermore, step 3 specifically includes:

[0021] 3.1 Based on the designed navigation knowledge decision-making support graph model, complete the entity, attribute, and relationship annotation of the text in the customized navigation knowledge dataset through manual and automatic methods, and complete the construction of the knowledge graph extraction dataset;

[0022] 3.2 According to the computing power of the hardware configuration, select the pre-trained general large model for knowledge extraction fine-tuning, and use the constructed knowledge graph extraction dataset to complete the fine-tuning based on the pre-trained general large model through full parameter fine-tuning or efficient parameter fine-tuning methods, thereby completing the construction of the knowledge extraction large model.

[0023] Furthermore, step 4 specifically includes:

[0024] 4.1. First, we constructed a fine-tuned dataset for maritime knowledge question and answering. We designed the structure of the maritime knowledge question and answer JSON statements, including instructions, input, output, source, and confidence level. Based on the titles and paragraphs of the preprocessed maritime knowledge documents from step 2, we input the different contents into the open-source big model and provided instructions to the big model, causing it to output the maritime knowledge question and answer JSON statements. We then manually proofread the generated JSON statements, modifying instructions, inputs, outputs, and supplementing the knowledge sources in the JSON statements, to improve the quality of the dataset.

[0025] 4.2. Based on the computing power of the hardware configuration, select a pre-trained general large model for fine-tuning the maritime knowledge question and answer problem. Use the constructed maritime knowledge question and answer fine-tuning dataset to complete fine-tuning based on the pre-trained general large model through full parameter fine-tuning or efficient parameter fine-tuning methods, thereby completing the construction of the maritime knowledge question and answer large model.

[0026] Furthermore, step 6 includes:

[0027] 6.1. Use a graph embedding model to encode the entities and relationships in the navigation knowledge decision-making support knowledge graph from step 5 into graph embedding vectors. After receiving the user's question input or the current scene knowledge model information input, search the graph embedding vectors through direct search or multi-hop search via graph traversal.

[0028] 6.2. Enhance the retrieval results based on the large-scale navigation knowledge question-answering model constructed in step 4, and design a multi-factor reward function for auxiliary decision-making suggestion generation. By adjusting the attention weight of the navigation auxiliary decision-making generation model to control the confidence threshold of the generated content, the generation results are optimized to achieve the generation of navigation auxiliary decision-making suggestions under the current input.

[0029] Moreover, the reward function R in step 6.2 includes accuracy reward R1, manual feedback reward R2, and fluency penalty R3;

[0030] R1=α1·BLEU+α2·ROUGE

[0031] R=β1·R1+β2·R2-β3·R3

[0032] Among them, BLEU is an indicator for evaluating the accuracy of text generation; ROUGE is an indicator for evaluating the coverage of text generation; α1, α2, β1, β2, and β3 are the weights of each indicator.

[0033] The advantages and positive effects of the present invention are:

[0034] 1. The present invention can address the difficulties such as the complexity and poor regularity of the navigation knowledge system, and the different navigation rules under the jurisdiction of different waters and different areas. It combines the scenario knowledge model to construct a typical navigation knowledge auxiliary decision-making knowledge graph, and opens up the auxiliary decision-making link from rule text knowledge to real-time scenario navigation suggestions for ship navigation. By proposing navigation auxiliary decision-making suggestions that meet the constraints of navigation knowledge rules based on real-time scenario information when the ship is sailing, it provides auxiliary decision-making support for assisted navigation of manned ships or autonomous navigation of unmanned ships.

[0035] 2. The present invention improves the accuracy of complex decision-making by fine-tuning the pre-trained large model with professional navigation knowledge and introducing a reward mechanism for auxiliary decision-making suggestion generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a composition diagram of the navigation auxiliary decision-making device based on intelligent voice interaction constructed by the present invention. DETAILED DESCRIPTION

[0037] The structure of the present invention will be further described below with reference to the accompanying drawings and through examples. It should be noted that the present examples are descriptive rather than restrictive.

[0038] A navigation decision-making assistance method based on a large model, the invention comprises the following steps:

[0039] Step 1: Design a schema for navigation knowledge-assisted decision-making

[0040] In order to realize real-time information such as intelligent navigation situation awareness information and intelligent navigation design and operation conditions, and combine it with the ship's intelligent navigation knowledge, realize closed-loop navigation auxiliary decision-making from perception to decision-making of intelligent navigation, construct a scenario knowledge model of real-time information such as intelligent navigation situation awareness, situation cognition, and intelligent navigation design and operation conditions, design a navigation knowledge model based on entity, attribute, and relationship triples, connect the navigation knowledge model with the navigation scenario knowledge model through attributes and relationships, design the entity, attributes and relationship of the scenario knowledge model with navigation knowledge such as navigation rules, collision avoidance rules, typical navigation terminology, and form a navigation knowledge auxiliary decision-making graph model (schema).

[0041] The intelligent navigation scenario knowledge model is designed based on entity, attribute and relationship triples. The scenario knowledge model consists of information such as ship objects, object objects, water environment, ship attribute characteristics, object attribute characteristics, water environment attribute characteristics, ship-water spatiotemporal relationship, ship-to-ship spatiotemporal relationship, multi-ship and water congestion level, and ship execution tasks. Among them:

[0042] Ship objects include different ships classified according to executive agencies, purposes, etc.

[0043] Objects include buoys, lighthouses, shipwrecks and other navigation aids marked on nautical charts.

[0044] The aquatic environment includes geographical waters such as inland rivers, coastal areas, seas, anchorages, waterways, docks, narrow waterways, and restricted areas.

[0045] The attribute characteristics of landmarks include name, location, color, characteristics, etc.

[0046] Ship attribute characteristics include inherent attributes such as ship number, ship name, actuator, sound signal, light signal, as well as motion attributes such as position (longitude, latitude), speed, heading, heading angle, roll angle, pitch angle, angular velocity, and angular acceleration.

[0047] The static attribute characteristics of the water environment include geographical attribute characteristics such as environment location, water boundary, environment name, water depth information, underwater reefs, as well as dynamic attributes such as wind speed and direction, visibility, flow speed and direction, temperature, humidity, wave height, wave direction, and sea condition level.

[0048] The temporal and spatial relationship between ships and waters includes approaching, moving away, entering, exiting, crossing, and being located.

[0049] The time and space relationship between ships includes encountering, left crossing, right crossing, overtaking, etc.

[0050] The busyness of multiple ships and waters includes busy, open, etc.

[0051] By arranging and combining entities, attributes, and relationships in the scene model, scene reasoning and recognition are performed. Typical scenarios include busy dock scenes, unberthing scenes, anchoring scenes, and channel crossing scenes.

[0052] Based on typical textual navigation knowledge such as navigation rules, international regulations for avoiding collisions, and typical navigation terms, a typical navigation knowledge schema is designed based on entity, attribute, and relationship triples. The navigation knowledge schema is then connected with the navigation scenario knowledge model schema through attributes and relationships to complete the design of a typical navigation knowledge decision-making support graph schema.

[0053] Examples of entity types and entity attributes in a typical navigation knowledge schema are as follows:

[0054]

[0055] Examples of relationship types in a typical navigation knowledge schema are as follows:

[0056] Header entity type Relationship Type Tail entity type condition Include condition rule source Rule File

[0057] Step 2: Navigation knowledge data collection and preprocessing

[0058] Collect text data related to ship navigation tasks, such as ship maneuvering terms, navigation instrument operating terms, smart ship specifications, coastal water navigation rules, inland water navigation rules, collision avoidance rules, etc., and collect expert experience knowledge in multiple modules such as perception, decision-making, control, navigation monitoring and alarm of intelligent navigation / assisted navigation systems.

[0059] Data preprocessing is performed on text data in Word, PDF and other formats, and Word text and PDF text are converted into Markdown format. Then, data cleaning is performed on the text data, including removing noise, duplication and low-quality data, and unifying the text encoding, converting it into a trainable format, and constructing a customized typical navigation knowledge dataset. The noise and low-quality data include text structure noise in the text (such as headers and footers, separators, typesetting residues, log information), character-level noise (such as encoding errors, full-width and half-width mixing, special symbols, redundant spaces, etc.), semantic noise (such as meaningless text, mixed Chinese and English, repeated content, logical confusion, etc.), format quality issues (paragraph breaks, punctuation errors, inconsistent capitalization, inconsistent date format, etc.), text content quality issues (such as low sentence information density, error messages, outdated rule versions, etc.) and compliance issues.

[0060] Step 3: Build a large knowledge extraction model

[0061] According to the designed navigation knowledge decision-making support graph model, the navigation knowledge text cleaned in step 2 is annotated with JSON statements of entities, attributes, and relationships to complete the construction of the knowledge graph extraction dataset.

[0062] According to the computing power of the hardware configuration, a pre-trained general large model is selected for knowledge extraction fine-tuning. The constructed knowledge graph extraction dataset is used to complete fine-tuning based on the pre-trained general large model through full parameter fine-tuning or efficient parameter fine-tuning methods, thereby completing the construction of the knowledge extraction large model.

[0063] Based on the constructed knowledge extraction model, knowledge extraction can be performed on the navigation knowledge text cleaned in step 2, and entity, attribute, and relationship triples can be extracted from the input navigation knowledge text. The extracted knowledge graph is output and stored in the Neo4j graph database. The model can convert the complex, irregular, and jurisdictional navigation rules in different waters and regions into a navigation knowledge graph database according to the designed navigation knowledge decision-making support graph schema, thereby reducing labor costs.

[0064] Step 4: Build a large model of navigation knowledge questions and answers

[0065] First, a fine-tuning dataset for maritime knowledge question and answer is constructed, and the structure of the maritime knowledge question and answer JSON statement is designed, including instructions, input, output, source, confidence, etc.; open source large models such as Qwen and Deepseek are used to semi-automatically annotate the maritime knowledge question and answer dataset. According to the title and paragraph division of the maritime knowledge document preprocessed in step 2, different contents are input into the open source large model, and instructions are prompted to the open source large model so that the open source large model outputs the maritime knowledge question and answer JSON statement. The generated JSON statement is manually proofread, including modifying the instructions, input, output, and supplementing the knowledge source in the JSON statement, to improve the quality of the dataset.

[0066] An example of a knowledge question and answer JSON dataset is as follows:

[0067] {

[0068] "instruction":"What should you do when two power-driven vessels meet on opposite or nearly opposite courses and there is a risk of collision?",

[0069] "input":"",

[0070] "output":"Each should turn to starboard so that each passes the other on the port side.",

[0071] "source":"International Regulations for Preventing Collisions at Sea, 1972, Chapter II, Section 2, Rule 14",

[0072] "confidence":"1.0"

[0073] }

[0074] According to the computing power of the hardware configuration, a pre-trained general large model is selected for fine-tuning of the navigation knowledge question and answer. The constructed navigation knowledge question and answer fine-tuning dataset is used to complete the fine-tuning based on the pre-trained general large model through full parameter fine-tuning or efficient parameter fine-tuning methods, thereby completing the construction of the navigation knowledge question and answer large model.

[0075] The constructed large model of maritime knowledge question and answer can learn the maritime knowledge in the maritime knowledge question and answer dataset, and can output answers that meet the rules of maritime knowledge based on the questions entered by the user.

[0076] Step 5: Constructing a knowledge graph for navigation knowledge-assisted decision-making

[0077] After constructing the graph knowledge extraction model, the model was used to extract entity, attribute, and relationship triples from the collected maritime knowledge documents. The extracted knowledge graph was then stored in Neo4j. Because the static attributes of the water environment in the scenario knowledge model are highly correlated with the electronic nautical chart database, a connection was established between the classic maritime knowledge decision-making support knowledge graph and the electronic nautical chart database to obtain attribute information such as waterway boundaries, dock attributes, and bridge areas. Considering that some entity relationships in texts such as navigation rules and collision avoidance rules share the same meaning as those in the scenario knowledge model, but differ in representation, entity disambiguation and cross-source alignment were required. The semantic similarity between entities in the scenario knowledge model and the extracted entities was calculated based on TF-IDF cosine similarity. When the semantic similarity exceeded a set threshold, the two entities were considered to have the same meaning and an "equal" relationship. The aligned knowledge graph was then stored in Neo4j.

[0078] Step 6: Build a navigation assistance decision generation model based on reinforcement learning

[0079] After completing the construction of the typical navigation knowledge-assisted decision-making knowledge graph in step five, the graph embedding model is used to encode the entities and relationships of the graph into graph embedding vectors.

[0080] After receiving the user's question input or the current scene knowledge model information input, the graph embedding vector is searched through direct retrieval or multi-hop retrieval via graph traversal.

[0081] Based on the large navigation knowledge question-answering model constructed in step 4, the retrieval results are enhanced and generated to generate navigation auxiliary decision suggestions under the current input.

[0082] In order to improve the accuracy of the generated navigation assistance decision suggestions, the reinforcement learning method is introduced to optimize the generation model, and the reward function is designed. By adjusting the attention weight of the generation model, the confidence threshold of the generated content is controlled to optimize the generation results and realize the output of navigation assistance decision suggestions.

[0083] When designing the reward function R, it includes accuracy reward R1, manual feedback reward R2, and fluency penalty R3.

[0084] R1=α1·BLEU+α2·ROUGE

[0085] R=β1·R1+β2·R2-β3·R3

[0086] BLEU (Bilingual Evaluation Understudy) is an indicator for evaluating text generation accuracy, and ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is an indicator for evaluating text generation coverage. α1, α2, β1, β2, and β3 are the weights of each indicator.

[0087] Step 7: Design a navigation assistance decision-making device based on intelligent voice interaction, and deploy the navigation assistance decision-making generation model constructed in step 6 in the device.

[0088] The navigation auxiliary decision-making device based on intelligent voice interaction includes a data acquisition unit, a voice acquisition unit, a processor, a voice playback unit, a command sending unit and a power supply.

[0089] The data acquisition unit is used to collect real-time navigation data, situation awareness data, and navigation control data of the ship, thereby obtaining current scene information;

[0090] The voice collection unit is used to collect voice information sent by the ship operator to the navigation auxiliary decision device;

[0091] The processor is used for deploying navigation assistance decision algorithm software, including speech recognition, semantic understanding, dialogue management, and text calculation and generation of the navigation assistance decision generation model constructed in step six. In an active interaction scenario, the processor receives the current scene information collected by the data acquisition unit and the voice information emitted by the ship operator collected by the voice acquisition unit, and calculates and generates an answer result that conforms to the current scene or meets the operator's requirements; in a passive interaction scenario, the processor receives the current scene information collected by the data acquisition unit, and calculates and generates navigation assistance decision suggestions that meet the knowledge of navigation rules for the current scene, such as "The current ship is located in the main channel of Tianjin Port. According to the "Tianjin Maritime Safety Administration Vessel Traffic Management System Safety Supervision and Management Rules", the minimum speed of ships in the main channel shall not be less than 5 knots. The current ship speed is close to the minimum speed limit. Please drive with caution."

[0092] The voice playback unit is used to convert the navigation assistance decision suggestions or other answer results generated by the navigation assistance decision generation model into voice and voice broadcast;

[0093] When the answer result calculated by the navigation auxiliary decision generation model can directly interact with the intelligent navigation system, the instruction sending unit is used to convert the generated answer result into an interactive instruction for the intelligent navigation system and send it, such as sending the instruction "eliminate alarm";

[0094] The navigation auxiliary decision-making device based on intelligent voice interaction has two modes: active interaction and passive interaction.

[0095] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will understand that various replacements, changes and modifications are possible without departing from the spirit of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A navigation decision-making assistance method based on a large model, comprising the following steps: Step 1: Based on entity, attribute, and relationship triples, a scenario knowledge model for real-time information such as intelligent navigation situation awareness, situation cognition, and intelligent navigation design and operation conditions is constructed. Based on entity, attribute, and relationship triples, a navigation knowledge model is designed. The navigation knowledge model is then connected to the navigation scenario knowledge model through attributes and relationships. The navigation knowledge includes navigation rules, collision avoidance rules, and typical navigation terminology, forming a navigation knowledge decision-making support graph model. Step 2: Collect and preprocess the navigation knowledge data to build a custom navigation knowledge dataset; Step 3: Label the custom navigation knowledge dataset constructed in step 2, construct a knowledge graph extraction dataset, fine-tune the pre-trained general large model through the knowledge graph extraction dataset, and construct a knowledge graph knowledge extraction large model; Step 4: Based on the custom navigation knowledge dataset constructed in step 2 and the designed navigation knowledge question and answer sentence structure, first construct a navigation knowledge question and answer fine-tuning dataset; then fine-tune the pre-trained general large model using the knowledge question and answer fine-tuning dataset to construct a navigation knowledge question and answer large model; Step 5: Use the knowledge extraction model to extract entity, attribute, and relationship triples from the collected navigation knowledge documents. Disambiguate and align entities from different sources, unify different entity representations, and unify the representations of real-time variable names and graph entity names in the intelligent navigation system. The final knowledge graph is stored in Neo4j to build a knowledge graph for navigation knowledge-assisted decision-making. Step 6: Based on the navigation knowledge question-answering model constructed in step 4 and the navigation knowledge decision-making assistance knowledge graph constructed in step 5, a navigation decision-making assistance generation model based on reinforcement learning is constructed; Step 7. Design a navigation assistance decision-making device based on intelligent voice interaction, which includes a data acquisition unit, a command sending unit, a voice acquisition unit, a voice playback unit, a processor and a power supply; the navigation assistance decision-making generation model constructed in Step 6 is deployed in the processor to realize active / passive intelligent voice interaction of navigation assistance decision-making based on user voice input or real-time navigation, perception and other scene data input.

2. The navigation decision-making assistance method based on a large model according to claim 1, characterized in that: In step 1, the scene knowledge model includes the following information: ship objects, object objects, water environment, ship attribute characteristics, object attribute characteristics, water environment attribute characteristics, ship and water spatiotemporal relationship, ship and ship spatiotemporal relationship, multiple ships and water busyness, and ship execution tasks; scene reasoning and recognition are achieved by arranging and combining some entities, attributes, and relationships in the scene model.

3. The navigation decision-making assistance method based on a large model according to claim 1, characterized in that: Step 2 includes: 2.

1. Collect text data related to ship navigation tasks, including ship maneuvering terms, navigation instrument operating terms, smart ship specifications, coastal water navigation rules, inland water navigation rules, and collision avoidance rules; and collect expert experience and knowledge in the perception, decision-making, control, navigation monitoring, and alarm modules of intelligent navigation / assisted navigation systems; 2.

2. Preprocess the text data in Word and PDF formats collected in step 2.1, retain the title list, body text, complex tables, and formulas in the navigation knowledge text data, and convert the Word text and PDF text into Markdown format; 2.

3. Then, the text data is cleaned, noise, duplication and low-quality data are removed, and it is converted into a trainable format to complete the construction of the customized navigation knowledge dataset.

4. The navigation decision-making assistance method based on a large model according to claim 1, characterized in that: Step 3 specifically includes: 3.1 Based on the designed navigation knowledge decision-making support graph model, complete the entity, attribute, and relationship annotation of the text in the customized navigation knowledge dataset through manual and automatic methods, and complete the construction of the knowledge graph extraction dataset; 3.2 According to the computing power of the hardware configuration, select the pre-trained general large model for knowledge extraction fine-tuning, and use the constructed knowledge graph extraction dataset to complete the fine-tuning based on the pre-trained general large model through full parameter fine-tuning or efficient parameter fine-tuning methods, thereby completing the construction of the knowledge extraction large model.

5. The navigation decision-making assistance method based on a large model according to claim 1, characterized in that: Step 4 specifically includes: 4.

1. First, we constructed a fine-tuning dataset for maritime knowledge question and answering. We designed the structure of the maritime knowledge question and answer JSON statements, including instructions, input, output, source, and confidence level. We then input the different contents into the open-source big model based on the titles and paragraphs of the maritime knowledge documents preprocessed in step 2. We then prompted the open-source big model with instructions, causing it to output the maritime knowledge question and answer JSON statements. We then manually proofread the generated JSON statements, including modifying instructions, inputs, outputs, and supplementing the knowledge sources in the JSON statements. 4.

2. Based on the computing power of the hardware configuration, select a pre-trained general large model for fine-tuning the maritime knowledge question and answer problem. Use the constructed maritime knowledge question and answer fine-tuning dataset to complete fine-tuning based on the pre-trained general large model through full parameter fine-tuning or efficient parameter fine-tuning methods to complete the construction of the maritime knowledge question and answer large model.

6. The navigation decision-making assistance method based on a large model according to claim 1, characterized in that: Step 6 includes: 6.

1. Use a graph embedding model to encode the entities and relationships in the navigation knowledge decision-making support knowledge graph from step 5 into graph embedding vectors. After receiving the user's question input or the current scene knowledge model information input, search the graph embedding vectors through direct search or multi-hop search via graph traversal. 6.

2. Enhance the retrieval results based on the large-scale navigation knowledge question-answering model constructed in step 4, and design a multi-factor reward function for auxiliary decision-making suggestion generation. By adjusting the attention weight of the navigation auxiliary decision-making generation model to control the confidence threshold of the generated content, the generation results are optimized to achieve the generation of navigation auxiliary decision-making suggestions under the current input.

7. The navigation decision-making assistance method based on a large model according to claim 6, characterized in that: The reward function R in step 6.2 includes the accuracy reward R1, the manual feedback reward R2, and the fluency penalty R3; R1=α1·BLEU+α2·ROUGE R=β1·R1+β2·R2-β3·R3 Among them, BLEU is an indicator for evaluating the accuracy of text generation; ROUGE is an indicator for evaluating the coverage of text generation; α1, α2, β1, β2, β3 are the weights of each indicator.

Images