A self-learning-based maritime collision avoidance rule knowledge graph construction method

Abstract

This invention discloses a method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning, relating to the field of intelligent shipping technology. The main steps include: first, forming a standardized feature sequence by fusing multi-source navigation data; constructing an initial knowledge graph based on collision avoidance rules and adding dynamic attributes such as confidence levels to knowledge elements; using a self-learning engine to analyze real-time and historical data, updating the graph incrementally, and extracting empirical rules and risk knowledge to inject into the graph; for the current scenario, comprehensively reasoning about various types of knowledge in the graph and generating decision suggestions with multi-source interpretations; finally, dynamically adjusting the confidence levels and weights of knowledge in the graph based on feedback from actual operations, forming a continuously optimized closed-loop system. This invention transforms a static collision avoidance rule knowledge graph into a continuously self-optimizing dynamic intelligent system, improving the adaptability, interpretability, and overall reliability of ship collision avoidance decisions in real and complex environments.

Description

Technical Field

[0001] This invention relates to the field of intelligent shipping technology, specifically to a method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning. Background Technology

[0002] With the rapid development of intelligent shipping technology, transforming international maritime collision avoidance rules into computer-processable knowledge models has become a key research direction for achieving autonomous and safe navigation of ships. Currently, utilizing knowledge graph technology for structured representation and logical reasoning of collision avoidance rules is the mainstream technical approach. The common practice is to extract entities and relationships from the rule text through manual or natural language processing, construct entity models, and store them in a graph database. Then, scenario matching and behavioral reasoning are performed based on predefined logical rules. This method has initially achieved the formalization and visualization of collision avoidance knowledge, providing a basic support for decision support. However, the knowledge graphs constructed by such systems are essentially static; once established, the knowledge is fixed and difficult to evolve autonomously with the accumulation of navigation practice, maritime cases, and updates to external regulations. They must rely on tedious manual revisions by experts, leading to delayed knowledge updates and an inability to respond promptly to emerging scenarios and risks. Meanwhile, existing systems only use real-time sensor data as factual input for reasoning, lacking the ability to proactively mine and summarize unwritten empirical collision avoidance patterns and risk knowledge from massive amounts of historical navigation data and accident cases. This results in a single basis for decision-making and insufficient adaptability in complex or regional navigation scenarios. Furthermore, the system's reasoning mechanism is usually limited to mechanical rule matching, unable to perform differentiated weighting and personalized decisions based on different ship characteristics and navigation environments, nor can it evaluate and optimize the credibility of internal knowledge by comparing decision recommendations with feedback from actual operational effects, lacking a self-improving closed loop. These static, rigid limitations, lacking learning and evolutionary capabilities, make it difficult for existing knowledge graph systems to meet the high demands of real-time, adaptive, and continuously evolving intelligent collision avoidance systems in the dynamic and complex environment of the real sea. Summary of the Invention

[0003] To meet the high demands of real-time, adaptive, and continuously evolving intelligent collision avoidance systems in the dynamic and complex environment of the real sea, this invention proposes a self-learning-based method for constructing a knowledge graph of maritime collision avoidance rules, including the following steps: S1: Real-time acquisition and fusion of multi-source navigation data from ship sensors, environmental perception devices and historical databases to form a standardized spatiotemporal-entity-event feature sequence; S2: Construct an initial knowledge graph based on international maritime collision avoidance rules, and attach dynamic attributes including confidence and support to the knowledge elements in the graph; S3: Through a self-learning engine, the standardized spatiotemporal-entity-event feature sequence and case data from historical databases are continuously analyzed to update the entities and relationships in the knowledge graph in an incremental manner, and empirical collision avoidance rules and risk warning knowledge are extracted from them and injected into the knowledge graph in the form of dynamic attributes. S4: For the current navigation scenario, retrieve relevant rule knowledge, experience knowledge and risk knowledge from the updated knowledge graph, and perform comprehensive reasoning based on their dynamic attributes to generate collision avoidance decision suggestions with multi-source explanations; S5: Based on actual operational feedback and safety assessment results, dynamically adjust the confidence and weight of relevant knowledge in the knowledge graph.

[0004] This invention establishes a self-learning closed loop of "data fusion - dynamic knowledge update - reasoning decision - feedback optimization," transforming a static collision avoidance rule knowledge graph into a dynamic intelligent system that can continuously absorb navigation experience, adaptively adjust, and constantly optimize itself. This significantly improves the adaptability, interpretability, and overall reliability of ship collision avoidance decisions in real and complex environments.

[0005] Furthermore, in step S1, the multi-source navigation data includes real-time sensing data and historical data, wherein: the real-time sensing data includes at least ship dynamic information from AIS, radar, and GNSS, as well as meteorological and oceanographic data from environmental sensors, and the historical data includes at least historical AIS tracks and maritime accident reports.

[0006] Furthermore, in step S3, mining empirical collision avoidance rules includes: performing sequence pattern mining on safe encounter cases in historical AIS trajectories, extracting and abstracting successful collision avoidance operation patterns as empirical collision avoidance rules.

[0007] Furthermore, in step S3, mining risk warning knowledge includes: performing natural language processing and causal analysis on maritime accident reports, extracting key factors that lead to the accident, and associating them with relevant scenarios and behavioral nodes in the knowledge graph.

[0008] Furthermore, in step S3, the self-learning engine also includes a context adaptation module, which is used to analyze historical data of a specific ship or route to form a personalized knowledge subgraph and activate it when a relevant context is identified.

[0009] Furthermore, in step S2, the knowledge elements include entities, relationships, and attributes, and the dynamic attributes also include at least one of timestamps, traceability identifiers, and contextual conditions.

[0010] Furthermore, in step S3, updating the knowledge graph incrementally includes: automatically discovering and creating new navigation entities based on real-time data streams, and identifying and updating the dynamic spatiotemporal relationships between entities.

[0011] Furthermore, in step S4, the comprehensive reasoning includes: retrieving international maritime collision avoidance rules, rules of experience, risk warnings, and contextual preference knowledge related to the current scenario, and performing weighted calculations based on their confidence, support, and matching degree with the current scenario to generate multiple collision avoidance action plans ranked by recommendation level.

[0012] Furthermore, in step S4, the accompanying multi-source interpretation includes: explicitly indicating the collision avoidance rule clauses on which the generated collision avoidance decision recommendation is based, the statistical data supporting the decision, and relevant historical similar case references.

[0013] Furthermore, in step S5, the dynamic adjustment includes: comparing the collision avoidance decision suggestions recommended by the system with the actual actions taken by the driver and the safety assessment results, generating feedback signals, thereby increasing the confidence of the adopted and verified safety knowledge, or decreasing the confidence of the rejected knowledge or knowledge that leads to adverse consequences.

[0014] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The present invention proposes a method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning. By integrating multi-source navigation data in real time and giving each knowledge element in the knowledge graph dynamic attributes such as confidence and support, the system can continuously reflect the latest status and experience in navigation practice. On this basis, the self-learning engine automatically extracts empirical collision avoidance rules and risk warnings that are not explicitly stipulated from historical data through incremental learning and case mining, and dynamically injects them into the graph, enriching the sources of decision-making basis. (2) Based on context adaptation, personalized knowledge subgraphs are formed for specific ships and routes, thereby providing more adaptive decision support in a changing environment; (3) By comprehensively utilizing rule knowledge, experience knowledge and risk knowledge, and by weighting and integrating them according to their dynamic attributes, collision avoidance suggestions that conform to international rules and are close to actual scenarios are generated. Each suggestion is accompanied by multi-source explanations to enhance the interpretability of the system and the driver's trust. Attached Figure Description

[0015] Figure 1 A step diagram illustrating a self-learning-based method for constructing a knowledge graph of maritime collision avoidance rules; Figure 2 This is a module diagram of a self-learning-based knowledge graph construction system for maritime collision avoidance rules. Detailed Implementation

[0016] This invention relates to a method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning, particularly applicable to collision avoidance decision support in intelligent ships and autonomous navigation systems. The technical solution of this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0017] To clearly illustrate the specific implementation process of this invention and enable those skilled in the art to fully understand and implement this solution, the five core steps of this invention will be described below in conjunction with the system architecture and workflow. Figure 1 As shown, it includes the following steps: S1: Real-time acquisition and fusion of multi-source navigation data from ship sensors, environmental perception devices and historical databases to form a standardized spatiotemporal-entity-event feature sequence; S2: Construct an initial knowledge graph based on international maritime collision avoidance rules, and attach dynamic attributes including confidence and support to the knowledge elements in the graph; S3: Through a self-learning engine, the standardized spatiotemporal-entity-event feature sequence and case data from historical databases are continuously analyzed to update the entities and relationships in the knowledge graph in an incremental manner, and empirical collision avoidance rules and risk warning knowledge are extracted from them and injected into the knowledge graph in the form of dynamic attributes. S4: For the current navigation scenario, retrieve relevant rule knowledge, experience knowledge and risk knowledge from the updated knowledge graph, and perform comprehensive reasoning based on their dynamic attributes to generate collision avoidance decision suggestions with multi-source explanations; S5: Based on actual operational feedback and safety assessment results, dynamically adjust the confidence and weight of relevant knowledge in the knowledge graph.

[0018] Meanwhile, the overall architecture of the technical solution of this invention can be referred to Figure 2 The system module connection diagram shown mainly consists of four layers: multi-source perception and fusion layer, dynamic knowledge graph layer, self-learning engine layer, and intelligent reasoning and application layer. Each layer is closely connected through data flow and control flow, forming a complete closed loop from environmental perception to knowledge learning, then to reasoning and decision-making, and finally to self-optimization through feedback.

[0019] In practical implementation, the first step is to have various sensors and data interfaces in the multi-source sensing module collect real-time navigation data of the ship itself and its surrounding environment. This includes dynamic information from AIS (Automatic Identification System), radar, and GNSS (Global Navigation Satellite System), such as position, heading, speed, and rate of turn, as well as data from meteorological sensors and marine environmental monitoring equipment, such as wind speed, wind direction, visibility, wave height, and ocean currents. Furthermore, it requires access to electronic nautical chart databases and static and dynamic geographic information provided by port VTS (Vessel Traffic Service) systems, and the retrieval of historical AIS trajectory data and maritime accident investigation reports stored locally or in the cloud, forming a multi-source, heterogeneous navigation data set. This data is processed by the data fusion module, which synchronizes the time and coordinate systems of data from different sources. Then, through target association and tracking algorithms, various types of information about the same ship are fused to form a complete dynamic profile of the ship. Based on this, further preprocessing operations such as data cleaning, outlier filtering, and missing value imputation are performed, and representative feature vectors are extracted from the raw data. Finally, a standardized, structured "spatiotemporal entity event feature" data sequence is output. This sequence not only contains the state information of ships at various times, but also identifies encounter events between ships, environmental change events, etc., providing high-quality input for the subsequent construction and updating of the knowledge graph.

[0020] After obtaining the standardized data sequence, a dynamic knowledge graph layer is used for its construction and initialization phase. The core task of this phase is to build a structured knowledge framework based on the International Maritime Collision Avoidance Regulations (COLREGs) and endow it with dynamic evolution capabilities. Specifically, firstly, ontology modeling methods are used to abstract and define core concepts in the collision avoidance domain, forming a series of entity types such as ships, behaviors (e.g., straight-ahead navigation, yielding, turning, deceleration), encounter scenarios (e.g., head-on encounters, cross encounters, overtaking), and environmental conditions (e.g., poor visibility, narrow waterways, dense traffic areas). Simultaneously, the relationships between these entities are defined, such as a "yielding relationship" between "ship A" and "ship B," "a certain behavior" occurring in a "certain scenario," and "a certain environment" affecting the "application of a certain rule." Unlike traditional static knowledge graphs, in the graph constructed in this invention, each entity, relationship, and attribute is attached with a series of dynamic attributes. These attributes constitute the metadata of the knowledge elements, enabling them to reflect the real-time state and reliability of the knowledge. Key dynamic attributes include confidence level, which quantifies the credibility of the knowledge in the current system; support level, which records the frequency with which the knowledge is supported by historical data or success cases; timestamp, which marks the creation time or last update time of the knowledge; source identifier, which clarifies the source of the knowledge, such as whether it originates from international rule texts, the mining of a historical case, or the operational habits of a specific ship; and contextual conditions, which define the scope of the knowledge's effective application, such as whether it only applies to a certain ship type, a specific geographical area, or certain visibility conditions. A knowledge graph initialized in this way is no longer a fixed knowledge base, but an intelligent knowledge carrier with its own data framework, capable of dynamic filling and adjustment.

[0021] A key technology of this invention lies in the self-learning engine layer, which is crucial for driving the knowledge graph from static to dynamic and enabling continuous knowledge evolution. The self-learning engine consists of multiple collaborative functional modules that work together to achieve in-depth analysis and knowledge extraction from navigation data.

[0022] The incremental learning module is responsible for processing the real-time standardized data stream from the data fusion layer. This module analyzes the data stream in real time, automatically discovering and creating new navigational entities (such as newly appearing target ships), and identifying and updating the dynamic spatiotemporal relationships between entities (such as changes in distance or relative bearing between two ships). This ensures that the entity and relationship network in the knowledge graph remains synchronized with the real physical world. For example, when the system detects a new ship entering its range via AIS data, the incremental learning module creates a corresponding "ship" entity node in the knowledge graph and continuously updates its position, speed, and other attributes over time. Simultaneously, it establishes or updates the "spatial relationship" edges between this ship and other ships.

[0023] In addition to processing real-time data, the system also possesses the ability to mine deep knowledge from historical data, primarily accomplished by the case-driven mining module. This module focuses on extracting valuable information from two types of historical data: learning empirical rules from successful cases and learning risk warning knowledge from failed cases. For mining empirical rules, the module utilizes massive amounts of historical AIS trajectory data, employing algorithms such as sequence pattern mining and cluster analysis to identify safe and efficient ship encounter and avoidance maneuver sequences. For example, at a busy port entrance, data analysis might reveal a widely followed cooperative avoidance pattern (such as specific turning timing and magnitude) between large cargo ships and ferries, though not explicitly stated in the rules. The module abstracts this recurring and proven safe maneuver pattern, formalizing it as an empirical collision avoidance rule, assigning it initial confidence and support, and then injecting it as a new knowledge node into the knowledge graph, establishing connections with relevant scenarios and behavioral nodes. For mining risk warning knowledge, the module performs natural language processing and causal analysis on stored maritime accident investigation reports. Through text parsing and event chain extraction techniques, the module can extract key factors leading to accidents, such as "failure to activate radar in time under poor visibility conditions" and "erroneous acceleration behavior in a cross-encounter situation." These risk factors are then formalized into "risk warning" knowledge nodes and strongly correlated with corresponding scenario nodes (such as "cross-encounter in poor visibility") and behavior nodes (such as "acceleration") in the knowledge graph, playing a warning role in subsequent reasoning.

[0024] Furthermore, to provide more personalized decision support, the self-learning engine also includes a context-adaptive module. This module can perform specialized analysis on historical navigation data for specific vessels (e.g., based on their type, size, and maneuvering characteristics) or specific routes (e.g., a fixed trade route). By analyzing the vessel's or route's long-term AIS tracks and operational records, the module can learn its unique operational patterns or regional navigation habits, such as the specific route or speed a vessel typically uses when navigating a narrow waterway. These personalized patterns are summarized and constructed into a lightweight personalized knowledge subgraph, attached to the main knowledge graph. When the system identifies a relevant context for the current navigation through real-time data (e.g., recognizing that the vessel is indeed that specific vessel and is located in that narrow waterway), it automatically activates the corresponding personalized subgraph, allowing its knowledge to participate in reasoning, thereby providing decision suggestions that are more aligned with the vessel's actual operational preferences.

[0025] After the knowledge graph is continuously enriched and updated by the self-learning engine, the intelligent reasoning and application layer is responsible for using this dynamic knowledge base to generate collision avoidance decisions in real time.

[0026] When a ship is underway, this layer first activates the scene recognition module. This module uses the real-time perceived status information (position, heading, speed, etc.) of the ship and surrounding vessels, as well as environmental information (visibility, wind and waves, etc.), to perform matching calculations with various predefined standardized encounter scenario models in the knowledge graph. By calculating parameters such as the relative bearing, distance, approach speed, DCPA (closest encounter distance), and TCPA (time to nearest encounter point) between the two ships, the module can quickly and accurately determine which scenario or combination of these scenarios exists, such as a head-on encounter, a cross encounter (the ship is giving way or traveling in the same direction), overtaking, or being overtaken.

[0027] Once the current scenario is identified, the multi-strategy reasoning module can retrieve all knowledge entries related to the identified scenario from the dynamic knowledge graph. These knowledge entries come from diverse sources, including not only rule knowledge directly derived from international maritime collision avoidance regulations, but also empirical rules and risk warnings mined from historical data, as well as personalized preference knowledge generated through context adaptation. Due to the diversity of knowledge sources, for the same scenario, knowledge from different sources may provide the same, complementary, or even seemingly conflicting action recommendations. For example, a rule might stipulate that in a certain intersection situation, the yielding vessel should turn to starboard, but an empirical rule mined from local vessel operating practices might suggest that, due to the influence of currents in that area, turning slightly to port is safer and more efficient. To address the issues of fusion and conflict among multi-source knowledge, the reasoning module employs a weighted comprehensive evaluation mechanism based on dynamic attributes. For each retrieved relevant knowledge, the system comprehensively considers its confidence (representing the reliability of the knowledge itself), support (representing the number of times the knowledge has been verified in practice), and its matching degree with the parameters of the current specific scenario (representing the applicability of the knowledge in the current context). The system calculates the final recommended score for each knowledge-based action plan using a preset or adaptive weighting function. Finally, the system generates one or more collision avoidance action plans (such as "turn 20 degrees to the right", "decelerate to 10 knots and maintain course", "sound five short blasts and remain alert"), and sorts them according to their comprehensive recommended scores.

[0028] More importantly, to enhance driver trust in the system and achieve effective human-machine collaboration, the system provides a detailed, traceable, multi-source explanation for each generated decision suggestion. This explanation clearly identifies one or more core collision avoidance rules upon which the suggestion is based (e.g., "based on Rule 15 of the COLREGs Rules regarding cross-traffic collisions"), displays relevant statistical data supporting the decision (e.g., "in the past 100 similar scenarios, the encounter safety rate of this approach is 99.5%), and provides relevant historical similar cases for reference (e.g., "refer to the case where, on a certain day in 2023, in the same body of water, vessel A and vessel B successfully avoided a collision using similar maneuvers"). This comprehensive explanation greatly enhances the transparency and acceptability of the system's decisions.

[0029] To enable continuous self-optimization of the technical solution of this invention, a key feedback optimization closed loop is constructed, primarily reflected in the feedback optimization module of the self-learning engine layer. This module establishes a connection between system decisions and actual navigation operations and results. After outputting collision avoidance decision suggestions, it continuously monitors the actual actions taken by the pilot. Simultaneously, the system also accesses or generates a safety assessment of the collision avoidance operation's outcome, based on indicators such as whether subsequent ship trajectories are safe, whether a safe distance is maintained between the two vessels, and whether the collision hazard has been averted. The core function of the feedback optimization module is to compare the system-recommended strategy with the pilot's actual actions and, combined with the safety assessment results, generate a feedback signal. This signal is used to dynamically adjust the confidence and weight of relevant knowledge in the knowledge graph.

[0030] Specific adjustment strategies include: if a rule of thumb is recommended by the system and adopted by the pilot, and subsequent safety assessments confirm its safety and effectiveness, then the confidence and support for this rule of thumb will increase, and its influence in similar scenarios in the future will grow. Conversely, if a rule or piece of knowledge recommended by the system (whether an international rule or mined experience) is repeatedly rejected by the pilot (this may mean that the knowledge is inconsistent with actual operating habits or the local environment), or more seriously, if adopting the knowledge leads to adverse consequences or near-accident situations, then the confidence of that knowledge will be lowered. Through this mechanism of testing truth in practice, the knowledge in the knowledge graph is no longer static but can be continuously optimized and self-corrected based on feedback from the real world. Invalid or inapplicable knowledge will be gradually marginalized, while knowledge that has been repeatedly verified to be effective will be strengthened. This process enables the autonomous evolution of the knowledge graph, allowing the decision-making ability of the entire system to continuously improve with the accumulation of operating time, becoming more aligned with the complex and ever-changing maritime environment.

[0031] In summary, this invention fundamentally overcomes the inherent shortcomings of traditional static knowledge graph systems, such as lagging knowledge updates, insufficient adaptability, and lack of self-optimization, by constructing a self-learning-based knowledge graph for building and reasoning about maritime collision avoidance rules. This method integrates multi-source navigation data in real time and assigns dynamic attributes such as confidence and support to each knowledge element in the knowledge graph, enabling the system to continuously reflect the latest status and experience in navigation practice. Based on this, the self-learning engine automatically extracts unwritten empirical collision avoidance rules and risk warnings from historical data through incremental learning and case mining, dynamically injecting them into the graph to enrich the sources of decision-making basis. Simultaneously, the system also possesses context-adaptive capabilities, enabling it to form personalized knowledge subgraphs for specific vessels and routes, thereby providing more adaptive decision support in changing environments. During the reasoning phase, the system comprehensively utilizes rule knowledge, experience knowledge, and risk knowledge, and performs weighted fusion based on their dynamic attributes to generate collision avoidance suggestions that conform to international rules and closely reflect real-world scenarios. Each suggestion is accompanied by multi-source explanations, enhancing the system's interpretability and driver trust. Through a feedback optimization mechanism, the system uses actual operational results and safety assessments as feedback signals to dynamically adjust the confidence and weight of knowledge in the graph, thereby achieving a complete closed loop of "perception-learning-reasoning-decision-feedback" and continuous self-evolution.

[0032] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0033] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0034] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0035] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning, characterized in that, Including the following steps: S1: Real-time acquisition and fusion of multi-source navigation data from ship sensors, environmental perception devices and historical databases to form a standardized spatiotemporal-entity-event feature sequence; S2: Construct an initial knowledge graph based on international maritime collision avoidance rules, and attach dynamic attributes including confidence and support to the knowledge elements in the graph; S3: Through a self-learning engine, the standardized spatiotemporal-entity-event feature sequence and case data from historical databases are continuously analyzed to update the entities and relationships in the knowledge graph in an incremental manner, and empirical collision avoidance rules and risk warning knowledge are extracted from them and injected into the knowledge graph in the form of dynamic attributes. S4: For the current navigation scenario, retrieve relevant rule knowledge, experience knowledge and risk knowledge from the updated knowledge graph, and perform comprehensive reasoning based on their dynamic attributes to generate collision avoidance decision suggestions with multi-source explanations; S5: Based on actual operational feedback and safety assessment results, dynamically adjust the confidence and weight of relevant knowledge in the knowledge graph.

2. The method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning as described in claim 1, characterized in that, In step S1, the multi-source navigation data includes real-time sensing data and historical data. The real-time sensing data includes at least ship dynamic information from AIS, radar, and GNSS, as well as meteorological and oceanographic data from environmental sensors. The historical data includes at least historical AIS tracks and maritime accident reports.

3. The method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning as described in claim 2, characterized in that, In step S3, mining empirical collision avoidance rules includes: performing sequence pattern mining on safe encounter cases in historical AIS trajectories, extracting and abstracting successful collision avoidance operation patterns as empirical collision avoidance rules.

4. The method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning as described in claim 2, characterized in that, In step S3, mining risk warning knowledge includes: performing natural language processing and causal analysis on maritime accident reports, extracting key factors that lead to the accident, and associating them with relevant scenarios and behavioral nodes in the knowledge graph.

5. The method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning as described in claim 2, characterized in that, In step S3, the self-learning engine also includes a context adaptation module, which is used to analyze historical data of specific ships or routes to form personalized knowledge subgraphs and activate them when relevant contexts are identified.

6. The method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning as described in claim 1, characterized in that, In step S2, knowledge elements include entities, relationships, and attributes. Dynamic attributes also include at least one of timestamps, traceability identifiers, and contextual conditions.

7. The method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning as described in claim 1, characterized in that, In step S3, updating the knowledge graph incrementally includes: automatically discovering and creating new navigation entities based on real-time data streams, and identifying and updating the dynamic spatiotemporal relationships between entities.

8. The method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning as described in claim 1, characterized in that, In step S4, the comprehensive reasoning includes: retrieving international maritime collision avoidance rules, rules of experience, risk warnings and contextual preference knowledge related to the current scenario, and performing weighted calculations based on their confidence, support and matching degree with the current scenario to generate multiple collision avoidance action plans ranked by recommendation level.

9. The method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning as described in claim 1, characterized in that, In step S4, the accompanying multi-source interpretation includes: clearly indicating the collision avoidance rule clauses on which the generated collision avoidance decision recommendation is based, the statistical data supporting the decision, and relevant historical similar case references.

10. The method for constructing a knowledge graph of maritime collision avoidance rules based on self-learning as described in claim 1, characterized in that, In step S5, dynamic adjustment includes: comparing the collision avoidance decision suggestions recommended by the system with the actual actions taken by the driver and the safety assessment results, generating feedback signals, thereby increasing the confidence of the adopted and verified safety knowledge, or decreasing the confidence of the rejected knowledge or knowledge that leads to adverse consequences.

Images