A tool to enhance a collaborative collision avoidance decision-making method for intelligent ships in intersecting waters driven by a large language model.

Abstract

This invention provides a tool-enhanced, large language model-driven collaborative collision avoidance decision-making method for intelligent ships in intersecting waters. The method includes: establishing a formal model of the multi-agent ship collaborative collision avoidance problem in intersecting waters based on a partially observable Markov decision process; constructing an experimental framework for ship encounter scenarios by optimizing the observation space parameter set, action space constraints, and state transition probability function; designing a center-distributed two-layer large language model decision architecture for intelligent ships to generate collision avoidance strategies by integrating environmental perception data, coordination commands, navigation experience, and navigation rules; and constructing a tool enhancement system that integrates a mathematical engine and a navigation knowledge base, dynamically storing navigation experience, navigation rules, and water information through a semantic similarity retrieval mechanism. This invention, through theoretical modeling, architectural innovation, and tool enhancement, promotes the engineering application of large language models in the field of maritime decision-making, providing a new paradigm for solving the complex decision-making problem of multi-agent ship collaborative collision avoidance in intersecting waters.

Description

Technical Field

[0001] This invention relates to the fields of large models and intelligent ship technology, and more particularly to a tool-enhanced large language model-driven collaborative collision avoidance decision-making method for intelligent ships in intersecting waters. Background Technology

[0002] In the field of maritime navigation, confluence waters are high-risk areas for ship collisions. Confluence waters refer to areas where two or more shipping lanes intersect, resulting in a complex network of channels, frequent encounters between ships, and intricate navigation regulations. Ships navigating in confluence waters must adhere to strict navigation and collision avoidance rules to ensure navigational safety. Traditional ship collision avoidance decision-making relies primarily on the experience and judgment of human navigators; however, with the development of ship automation and intelligence, collaborative collision avoidance decision-making by intelligent ships has gradually become a research hotspot.

[0003] Currently, collaborative decision-making methods for solving complex collision avoidance problems in intersecting waters mainly include algorithms based on optimization theory and algorithms based on machine learning. Algorithms based on optimization theory optimize collision avoidance paths and strategies by establishing mathematical models, but their modeling process is complex, lacks flexibility, and struggles to adapt to complex dynamic environments. Machine learning-based algorithms learn collision avoidance strategies through data-driven approaches, but they have poor generalization capabilities, lack interactive capabilities, and struggle to cope with unknown navigation scenarios. In recent years, with the rise of generative pre-trained large language models (LLMs), they have demonstrated outstanding capabilities in understanding, reasoning, and human-computer interaction, providing new ideas for solving collision avoidance problems in intersecting waters. However, LLMs have shortcomings in continuous learning and struggle to adapt to constantly changing navigation environments and rules.

[0004] While existing collaborative decision-making methods can address collision avoidance issues in intersecting waters to some extent, they suffer from several drawbacks. Algorithms based on optimization theory involve complex modeling processes and struggle to adapt to dynamically changing navigation environments; machine learning-based algorithms exhibit poor generalization, lack interactive capabilities, and are ill-suited for complex navigation scenarios. Furthermore, while existing large language models excel in understanding, reasoning, and human-computer interaction, their parameters become fixed after pre-training, lacking continuous learning capabilities and making it difficult to adapt to constantly changing navigation environments and rules. Therefore, guiding large language models to make safe and reliable collaborative avoidance decisions through appropriate prompt word engineering, while simultaneously compensating for their shortcomings in continuous learning, has become a pressing technical challenge. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a tool to enhance a large language model-driven collaborative collision avoidance decision-making method for intelligent ships in intersecting waters. Through reasonable cue word engineering and a two-layer decision architecture, this invention guides the large language model to make safe and reliable collaborative avoidance decisions, thereby improving the reliability of intelligent ships' collision avoidance decisions in intersecting waters.

[0006] The technical means employed in this invention are as follows:

[0007] A tool to enhance a large language model-driven collaborative collision avoidance decision-making method for intelligent ships in intersecting waters includes:

[0008] S1. Based on partially observable Markov decision processes, a formal model is established for the multi-agent ship cooperative collision avoidance problem in intersecting waters.

[0009] S2. By optimizing the observation space parameter set, action space constraints, and state transition probability function, an experimental framework for ship encounter scenarios is constructed.

[0010] S3, Design Center - Distributed Two-Layer Large Language Model Decision Architecture, including a distributed layer and a central coordination layer, is used by intelligent ships to generate collision avoidance strategies by integrating environmental perception data, coordination instructions, navigation experience and navigation rules.

[0011] S4. The system enhances the construction tools, including a scenario description module, a conflict description module, a decision-making module, and a simulation environment module. It integrates a mathematical engine and a navigation knowledge base, and dynamically stores navigation experience, navigation rules, and waterway information through a semantic similarity retrieval mechanism.

[0012] Further, step S1 specifically includes:

[0013] S11. The confluence area is defined as the waterway formed by the intersection and convergence of two or more waterways, and is further classified into two-way confluence area, three-way confluence area, four-way confluence area, and multi-way confluence area according to the number of branch waterway intersections.

[0014] S12. Consider navigation restrictions caused by the characteristics of the waterway and the vessel itself, including water depth and waterway restrictions, vessel inertia, and vessel speed requirements.

[0015] S13. The multi-agent ship cooperative collision avoidance problem in intersecting waters is constructed as a formal model based on a partially observable Markov Decision Process (POMDP), represented by a six-tuple as follows:

[0016] G = {V,S,A,P,[O]} i ],[R i ]}

[0017] Where V is the multi-agent ship ensemble, representing the collaborative decision-making entity; S is the joint state space, encompassing ship kinematic parameters and environmental characteristics; A is the distributed action space, defining the set of maneuvering commands that each agent can execute; P is the state transition function, constructed based on the ship kinematic model; Oi R represents the local observation space of agent i, reflecting the limitations of sensor information acquisition; i It is an individualized reward function that integrates safety and energy efficiency optimization objectives.

[0018] Further, step S2 specifically includes:

[0019] S21. Define the set of interactive ships within the sensing radius L of the intelligent ship i as N. i Observation matrix Representing local situational awareness information, where |N i | represents the number of observable ships, |F| = 6| represents the dimension of ship state features; the single ship state feature vector is defined as:

[0020]

[0021] Where, x i ,y i Let v be the longitudinal and lateral coordinates of ship i in the geographic coordinate system; x,i ,v y,i The velocity component in the corresponding direction; cosφ i With sinφ i Commonly representing the ship's heading angle φ i The trigonometric function form avoids the problem of periodic discontinuity of angles;

[0022] S22. Construct the system's global observation space, which is composed of the local observations of all agents, as follows:

[0023]

[0024] in, This indicates the observation tensor splicing operation; O1 represents the local observation space of smart ship 1, O2 represents the local observation space of smart ship 2, and O |V| This represents the local observation space of the intelligent ship V;

[0025] S23. Employing a speed-control-based discretization decision-making strategy, the action space of the intelligent ship is defined as discrete semantic commands A. i ={α dec ,α hold ,α acc}, which correspond to the commands for deceleration cruise, constant speed maintenance, and acceleration respectively;

[0026] S24. Once the high-level decision-making command is generated, the low-level motion controller will calculate and generate the corresponding target speed, driving the intelligent ship to complete the cooperative collision avoidance maneuver. This design follows the maritime practice principle of "using the vehicle to yield more and the rudder to yield less" in converging waters, achieving safe collision avoidance through longitudinal motion control.

[0027] S25. Discretize the ship's longitudinal propulsion control quantities into five standardized engine telegraph commands, and establish a speed-engine command mapping relationship as follows:

[0028] Set the target speed as v target ∈{0,3,6,9,12}kn, when an acceleration / deceleration command is received, the target speed is adjusted in step increments according to the gear;

[0029] S26. Based on the deviation between the target speed and the current speed, construct a PID controller as follows:

[0030]

[0031] Among them, K p Let e(t) represent the proportional gain factor, and e(t) represent the speed deviation, e(t) = v target -v actual v target V represents the target velocity. actual Indicates the current speed; K i K represents the integral cumulative adjustment factor, e(τ) represents the real-time speed deviation, and K represents the integral cumulative adjustment factor. d Represents the differential dynamic suppression coefficient;

[0032] Furthermore, in step S3, within the designed central-distributed two-layer large language model decision architecture:

[0033] The distributed layer is driven by a large language model, enabling intelligent ships to autonomously complete environmental perception and decision-making reasoning.

[0034] The central coordination layer is equipped with a large language model coordinator, which uses a conflict severity quantification model and the 1972 International Regulations for Preventing Collisions at Sea to determine the passage sequence.

[0035] Further, step S3 specifically includes:

[0036] S31. Define a ship traffic conflict (when a ship is approaching another ship in space and time, causing a collision risk between the ship and the other ship, the ship must take action to avoid the collision), and classify ship traffic conflicts into cross-traffic conflicts, head-on conflicts and tail-end conflicts according to the perspective of the conflict.

[0037] S32. Quantify the severity of ship traffic conflicts, including quantifying the severity of rear-end collisions and the severity of cross-traffic conflicts;

[0038] S33. Leveraging the prominent characteristics of large language models in text understanding, reasoning, and human-computer interaction, establish a prompt word engineering system and construct a collaborative collision avoidance thinking chain based on state awareness, intent sharing, conflict resolution, and decision-making to guide large language models in making reliable decisions.

[0039] Furthermore, in step S32, quantifying the severity of rear-end collisions and quantifying the severity of cross-collisions specifically includes:

[0040] Based on the influence of Fujii Shipbuilding's field on inter-ship relationships, the severity of rear-end collisions between two ships in the same waterway is quantified as follows:

[0041]

[0042] Where d represents the distance between two ships in the same channel; t represents the time when the two ships collide; and L represents the length of the ship.

[0043] Cross-collision incidents typically occur near the point of conflict in the intersection area. Based on the impact of minimum encounter distance and minimum encounter time on the risk of ship collision, the severity of cross-collision incidents is quantified using DTCP (distance from ship to the point of conflict) and TTCP (time to the point of conflict), as follows:

[0044]

[0045] Among them, DTCP i and DTCP j Let i and j represent the distances from intelligent ships i and j to the point of ship traffic conflict, respectively; ΔTTCP represents the time difference of arrival at the point of ship traffic conflict, and its calculation formula is as follows:

[0046]

[0047] Among them, TTCP i Indicates TTCP j Indicates the time when ships i and j arrive at the point of conflict in ship traffic; v i and v j This represents the speed of intelligent ships i and j.

[0048] Further, step S4 specifically includes:

[0049] S41. Enhance retrieval for static information:

[0050] Information on navigation rules and water conditions in the confluence waters is collected, cleaned, and then sliced. The cleaned text data is converted into semantic vectors using a text embedding model and stored in a vector database.

[0051] The input prompts are vectorized, the k fragment vectors with the highest dot product similarity in the vector database are retrieved, the prompt vectors and fragment vectors are fused to enhance the input, and the enhanced context is sent to the large language model for decision-making.

[0052] S42. Enhanced retrieval of dynamic navigation experience information:

[0053] A decision-making outcome evaluation module is established to generate semantic evaluations based on the decision-making outcomes during the navigation process by calculating the conflict risk; for example: "This maneuvering action exacerbated the conflict and should be avoided."

[0054] Store the scene description, conflict description, decision result, and decision evaluation in JSON format, and establish links between the scene, action, and evaluation.

[0055] The text embedding model is used to convert the text into semantic vectors and store them in a vector database. When making a decision, the k scenes with the highest similarity to the scenes in the database (cosine similarity) are retrieved. The original scene prompts are input together with the k scenes and their evaluation results and output to the large language model.

[0056] Further, in step S41, the formula for calculating the dot product similarity is as follows:

[0057]

[0058] Among them, A i B i For the metric vector, n represents the vector dimension.

[0059] Compared with the prior art, the present invention has the following advantages:

[0060] 1. The present invention provides a tool-enhanced large language model-driven collaborative collision avoidance decision-making method for intelligent ships in intersecting waters. Through reasonable cue word engineering and a two-layer decision architecture, it guides the large language model to make safe and reliable collaborative avoidance decisions, thereby improving the reliability of intelligent ships' collision avoidance decisions in intersecting waters.

[0061] 2. The present invention provides a tool-enhanced large language model-driven collaborative collision avoidance decision-making method for intelligent ships in intersecting waters. The constructed tool-enhanced system compensates for the shortcomings of LLM in continuous learning by dynamically storing navigation experience, navigation rules and water information, and enhances the model's adaptability to the ever-changing navigation environment and rules.

[0062] 3. This invention promotes the engineering application of large language models in the field of maritime decision-making through theoretical modeling, architectural innovation and tool enhancement, and provides a new paradigm for solving the complex decision-making problem of multi-intelligent ship cooperative collision avoidance in intersecting waters.

[0063] Based on the above reasons, this invention can be widely applied in fields such as large-scale models and intelligent ships. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a flowchart of the method of the present invention.

[0066] Figure 2 This is a schematic diagram of the types of confluence waterways in this invention.

[0067] Figure 3 This is a schematic diagram illustrating the conflict types of the present invention.

[0068] Figure 4 This is a schematic diagram of a traffic conflict point at the intersection of four roads in a waterway.

[0069] Figure 5 A thought process diagram for collaborative decision-making in converging waters.

[0070] Figure 6 Generate a framework diagram to enhance retrieval. Detailed Implementation

[0071] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0072] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0073] like Figure 1 As shown, this invention provides a tool-enhanced large language model-driven collaborative collision avoidance decision-making method for intelligent ships in intersecting waters, comprising:

[0074] S1. Based on partially observable Markov decision processes, a formal model is established for the multi-agent ship cooperative collision avoidance problem in intersecting waters.

[0075] S2. By optimizing the observation space parameter set, action space constraints, and state transition probability function, an experimental framework for ship encounter scenarios is constructed.

[0076] S3, Design Center - Distributed Two-Layer Large Language Model Decision Architecture, including a distributed layer and a central coordination layer, is used by intelligent ships to generate collision avoidance strategies by integrating environmental perception data, coordination instructions, navigation experience and navigation rules.

[0077] S4. The system enhances the construction tools, including a scenario description module, a conflict description module, a decision-making module, and a simulation environment module. It integrates a mathematical engine and a navigation knowledge base, and dynamically stores navigation experience, navigation rules, and waterway information through a semantic similarity retrieval mechanism.

[0078] In a specific implementation, as a preferred embodiment of the present invention, step S1 specifically includes:

[0079] S11. A confluence area is defined as a waterway formed by the intersection and convergence of two or more shipping lanes. Vessels in a confluence area typically navigate according to prescribed sea routes or channels. The number of confluence routes largely determines the type of vessel traffic flow, such as... Figure 2 As shown, the waterways are divided into two-way confluence waters, three-way confluence waters, four-way confluence waters, and multi-way confluence waters based on the number of branch waterway intersections.

[0080] S12. Considering the navigation restrictions caused by the characteristics of the waterway and the vessel itself, including water depth and waterway restrictions, vessel inertia, and vessel speed requirements; in this embodiment, specifically: 1) Navigable waters are limited by water depth and waterway restrictions in the confluence area, making it difficult for large vessels to make large-angle turns to avoid oncoming vessels in the confluence area. The priority of yielding to the vehicle is greater than that of yielding to the rudder, that is, the speed must be strictly controlled to avoid the collision; 2) Vessels have extremely high inertia and slow movement. Vessels need to maintain a safe distance of more than 4 times their length to avoid collision accidents; 3) Vessels need a certain speed to maintain normal maneuverability. Moreover, when large vessels turn from one waterway to another, they need a speed of more than 3.5 knots to ensure rudder effectiveness.

[0081] S13. The multi-agent ship cooperative collision avoidance problem in intersecting waters is constructed as a formal model based on a partially observable Markov Decision Process (POMDP), represented by a six-tuple as follows:

[0082] G = {V,S,A,P,[O]} i ],[R i ]}

[0083] Where V is the multi-agent ship ensemble, representing the collaborative decision-making entity; S is the joint state space, encompassing ship kinematic parameters and environmental characteristics; A is the distributed action space, defining the set of maneuvering commands that each agent can execute; P is the state transition function, constructed based on the ship kinematic model; O i R represents the local observation space of agent i, reflecting the limitations of sensor information acquisition; i The individualized reward function integrates safety and energy efficiency optimization objectives. In this embodiment, at decision time t, agent i obtains local observation o based on the current state St. i ,t∈O i The manipulation instructions ai,t∈A are generated through policy π, and after being synthesized by multi-agent actions, a state transition St+1~P(·|St,at) is triggered. This modeling method can characterize the temporal dependence and observation uncertainty of interactions between ships, and enhance the mathematical logic of center-distributed collaborative decision-making.

[0084] In a specific implementation, as a preferred embodiment of the present invention, step S2 specifically includes:

[0085] S21. Define the set of interactive ships within the sensing radius L of the intelligent ship i as N. i Observation matrix Representing local situational awareness information, where |N i | represents the number of observable ships, |F| = 6| represents the dimension of ship state features; the single ship state feature vector is defined as:

[0086]

[0087] Where, x i ,y i Let v be the longitudinal and lateral coordinates of ship i in the geographic coordinate system; x,i ,v y,i The velocity component in the corresponding direction; cosφ i With sinφ i Commonly representing the ship's heading angle φ i The trigonometric function form avoids the problem of periodic discontinuity of angles;

[0088] S22. Construct the system's global observation space, which is composed of the local observations of all agents, as follows:

[0089]

[0090] in, This indicates the observation tensor splicing operation; O1 represents the local observation space of smart ship 1, O2 represents the local observation space of smart ship 2, and O |V| This represents the local observation space of the intelligent ship V;

[0091] S23. To address the specific needs of collaborative collision avoidance among vessels in intersecting waters, a discretized decision-making strategy based on speed control is adopted. Given the significant advantages of LLM in knowledge-driven semantic reasoning, while its numerical analytical capabilities are relatively limited, the action space of the intelligent ship is defined as discrete semantic instructions A. i ={α dec ,α hold ,α acc}, which correspond to the commands for deceleration cruise, constant speed maintenance, and acceleration respectively;

[0092] S24. Once the high-level decision-making command is generated, the low-level motion controller will calculate and generate the corresponding target speed, driving the intelligent ship to complete the cooperative collision avoidance maneuver. This design follows the maritime practice principle of "using the vehicle to yield more and the rudder to yield less" in converging waters, achieving safe collision avoidance through longitudinal motion control.

[0093] S25. Discretize the ship's longitudinal propulsion control quantities into five standardized engine telegraph commands, and establish a speed-engine command mapping relationship as follows:

[0094] Set the target speed as v target ∈{0,3,6,9,12}kn, when an acceleration / deceleration command is received, the target speed is adjusted in step increments according to the gear;

[0095] S26. Based on the deviation between the target speed and the current speed, construct a PID controller as follows:

[0096]

[0097] Among them, K p Let e(t) represent the proportional gain factor, and e(t) represent the speed deviation, e(t) = v target -v actual v target V represents the target velocity. actual Indicates the current speed; K i K represents the integral cumulative adjustment factor, e(τ) represents the real-time speed deviation, and K represents the integral cumulative adjustment factor. d Represents the differential dynamic suppression coefficient;

[0098] In a specific implementation, as a preferred embodiment of the present invention, in step S3, the designed central-distributed two-layer large language model decision architecture includes:

[0099] The distributed layer is driven by a large language model, enabling intelligent ships to autonomously complete environmental perception and decision-making reasoning.

[0100] The central coordination layer is equipped with a large language model coordinator, which uses a conflict severity quantification model and the 1972 International Regulations for Preventing Collisions at Sea to determine the passage sequence.

[0101] In a specific implementation, as a preferred embodiment of the present invention, step S3 specifically includes:

[0102] S31. Define a maritime traffic conflict (where a vessel is in a situation where another vessel approaches the vessel in space and time, creating a risk of collision and necessitating action to avoid a collision). Figure 3 As shown, based on the angle of conflict, ship traffic conflicts are divided into cross-traffic conflicts, head-on conflicts, and rear-end collision conflicts;

[0103] S32. Quantify the severity of vessel traffic conflicts, including quantifying the severity of rear-end collisions and the severity of cross-traffic conflicts; specifically including:

[0104] Based on the influence of Fujii Shipbuilding's field on inter-ship relationships, the severity of rear-end collisions between two ships in the same waterway is quantified as follows:

[0105]

[0106] Where d represents the distance between two ships in the same channel; t represents the time when the two ships collide; and L represents the length of the ship.

[0107] like Figure 4 As shown, cross-collision generally occurs near the point of conflict in the intersection area. Based on the impact of minimum encounter distance and minimum encounter time on the risk of ship collision, the severity of cross-collision is quantified using DTCP (distance from ship to the point of conflict) and TTCP (time of arrival at the point of conflict), as follows:

[0108]

[0109] Among them, DTCP i and DTCP j Let i and j represent the distances from intelligent ships i and j to the point of ship traffic conflict, respectively; ΔTTCP represents the time difference of arrival at the point of ship traffic conflict, and its calculation formula is as follows:

[0110]

[0111] Among them, TTCP i Indicates TTCP j Indicates the time when ships i and j arrive at the point of conflict in ship traffic; v i and v j This represents the speed of intelligent ships i and j.

[0112] S33. Leveraging the prominent characteristics of large language models in text understanding, reasoning, and human-computer interaction, establish a prompt word engineering framework to construct a collaborative collision avoidance thought chain based on state awareness, intent sharing, conflict resolution, and decision-making, guiding the large language model to make reliable decisions. For example... Figure 5As shown. Before making a decision, a dual safety risk assessment is conducted, examining the set of alternative actions before the decision is made, and eliminating actions that, due to excessive speed, would result in insufficient safe distance from the vessel, leading to a rear-end collision or exacerbating the severity of the conflict.

[0113] In a specific implementation, as a preferred embodiment of the present invention, step S4 specifically includes:

[0114] S41. Enhance retrieval for static information:

[0115] Information on navigation rules and water conditions in the confluence waters is collected, cleaned, and then sliced. The cleaned text data is converted into semantic vectors using a text embedding model and stored in a vector database.

[0116] The input prompt is vectorized, and the k fragment vectors with the highest dot product similarity are retrieved from the vector database. The prompt vector and fragment vector are then fused to enhance the input, and the enhanced context is sent to the large language model for decision-making. The formula for calculating the dot product similarity is as follows:

[0117]

[0118] Among them, A i B i Let n be the metric vector, representing the vector dimension. The retrieval enhancement process is as follows: Figure 6 As shown.

[0119] S42. Enhanced retrieval of dynamic navigation experience information:

[0120] A decision-making outcome evaluation module is established to generate semantic evaluations based on the decision-making outcomes during the navigation process by calculating the conflict risk; for example: "This maneuvering action exacerbated the conflict and should be avoided."

[0121] Store the scene description, conflict description, decision result, and decision evaluation in JSON format, and establish links between the scene, action, and evaluation.

[0122] The text embedding model is used to convert the text into semantic vectors and store them in a vector database. When making a decision, the k scenes with the highest similarity to the scenes in the database (cosine similarity) are retrieved. The original scene prompts are input together with the k scenes and their evaluation results and output to the large language model.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A tool-enhanced method for collaborative collision avoidance decision-making by intelligent ships in intersecting waters driven by a large language model, characterized in that, include: S1. Based on partially observable Markov decision processes, a formal model is established for the multi-agent ship cooperative collision avoidance problem in intersecting waters, including: S11. The confluence waters are defined as waters formed by the intersection and convergence of two or more waterways, and are further divided into two-way confluence waters, three-way confluence waters, four-way confluence waters, and multi-way confluence waters according to the number of branch waterways confluencing each other. S12. Consider navigation restrictions caused by the characteristics of the waterway and the vessel itself, including water depth and waterway restrictions, vessel inertia, and vessel speed requirements. S13. The multi-agent ship cooperative collision avoidance problem in intersecting waters is constructed as a formal model based on a partially observable Markov decision process, represented by a six-tuple as follows: G={V,S,A,P,[O i ],[R i ]} Where V is the multi-agent ship ensemble, representing the collaborative decision-making entity; S is the joint state space, encompassing ship kinematic parameters and environmental characteristics; A is the distributed action space, defining the set of maneuvering commands that each agent can execute; P is the state transition function, constructed based on the ship kinematic model; O i R represents the local observation space of agent i, reflecting the limitations of sensor information acquisition; i The individualized reward function integrates safety and energy efficiency optimization objectives; S2. By optimizing the observation space parameter set, action space constraints, and state transition probability function, an experimental framework for ship encounter scenarios is constructed, including: S21. Define the set of interactive ships within the sensing radius L of the intelligent ship i as N. i Observation matrix Representing local situational awareness information, where |N i | represents the number of observable ships, |F| = 6, and |F| = 6 represents the dimension of ship state features; the single ship state feature vector is defined as: Where, x i ,y i Let v be the longitudinal and lateral coordinates of ship i in the geographic coordinate system; x,i ,v y,i The velocity component in the corresponding direction; cosφ i With sinφ i Commonly representing the ship's heading angle φ i The trigonometric function form avoids the problem of periodic discontinuity of angles; S22. Construct the system's global observation space, which is composed of the local observations of all agents, as follows: in, This indicates the observation tensor splicing operation; O1 represents the local observation space of smart ship 1, O2 represents the local observation space of smart ship 2, and O |V| This represents the local observation space of the intelligent ship V; S23. Employing a speed-control-based discretization decision-making strategy, the action space of the intelligent ship is defined as discrete semantic commands A. i ={α dec ,α hold ,α acc }, which correspond to the commands for deceleration cruise, constant speed maintenance, and acceleration respectively; S24. After the high-level decision-making command is generated, the low-level motion controller will calculate and generate the corresponding target speed to drive the intelligent ship to complete the cooperative collision avoidance action. S25. Discretize the ship's longitudinal propulsion control quantities into five standardized engine telegraph commands, and establish a speed-engine command mapping relationship as follows: Set the target speed as v target ∈{0,3,6,9,12}kn, when an acceleration / deceleration command is received, the target speed is adjusted in step increments according to the gear; S26. Based on the deviation between the target speed and the current speed, construct a PID controller as follows: Among them, K p Let e(t) represent the proportional gain factor, and e(t) represent the speed deviation, e(t) = v target -v actual v target V represents the target velocity. actual Indicates the current speed; K i K represents the integral cumulative adjustment factor, e(τ) represents the real-time speed deviation, and K represents the integral cumulative adjustment factor. d Represents the differential dynamic suppression coefficient; S3, Design Center - Distributed Two-Layer Large Language Model Decision Architecture, including a distributed layer and a central coordination layer, is used by intelligent ships to generate collision avoidance strategies by integrating environmental perception data, coordination instructions, navigation experience and navigation rules. S4. The system enhances the construction tools, including a scenario description module, a conflict description module, a decision-making module, and a simulation environment module. It integrates a mathematical engine and a navigation knowledge base, and dynamically stores navigation experience, navigation rules, and waterway information through a semantic similarity retrieval mechanism.

2. The method for collaborative collision avoidance decision-making of intelligent ships in intersecting waters driven by a tool-enhanced large language model according to claim 1, characterized in that, In step S3, within the designed central-distributed two-layer large language model decision architecture: The distributed layer is driven by a large language model, enabling intelligent ships to autonomously complete environmental perception and decision-making reasoning. The central coordination layer is equipped with a large language model coordinator, which uses a conflict severity quantification model and the 1972 International Regulations for Preventing Collisions at Sea to determine the passage sequence.

3. The method for collaborative collision avoidance decision-making of intelligent ships in intersecting waters driven by a tool-enhanced large language model according to claim 1, characterized in that, Step S3 specifically includes: S31. Define ship traffic conflicts and classify them into cross-traffic conflicts, head-on conflicts, and tail-end collision conflicts based on the perspective of the conflict. S32. Quantify the severity of ship traffic conflicts, including quantifying the severity of rear-end collisions and the severity of cross-traffic conflicts; S33. Leveraging the prominent characteristics of large language models in text understanding, reasoning, and human-computer interaction, establish a prompt word engineering system and construct a collaborative collision avoidance thinking chain based on state awareness, intent sharing, conflict resolution, and decision-making to guide large language models in making reliable decisions.

4. The method for collaborative collision avoidance decision-making of intelligent ships in intersecting waters driven by a tool-enhanced large language model according to claim 1, characterized in that, In step S32, the severity of rear-end collisions and the severity of cross-traffic collisions are quantified, specifically including: Based on the influence of Fujii Shipbuilding's field on inter-ship relationships, the severity of rear-end collisions between two ships in the same waterway is quantified as follows: Where d represents the distance between two ships in the same channel; t represents the time when the two ships collide; and L represents the length of the ship. Cross-collision incidents typically occur near the point of conflict in the intersection area. Based on the impact of minimum encounter distance and minimum encounter time on the risk of ship collision, the severity of cross-collision incidents is quantified using the distance from ship to the point of conflict (DTCP) and the time to reach the point of conflict (TTCP), as follows: Among them, DTCP i and DTCP j Let i and j represent the distances from intelligent ships i and j to the point of ship traffic conflict, respectively; ΔTTCP represents the time difference of arrival at the point of ship traffic conflict, and its calculation formula is as follows: Among them, TTCP i Indicates TTCP j Indicates the time it takes for ships i and j to reach the point of conflict in ship traffic; v i and v j This represents the speed of intelligent ships i and j.

5. The tool-enhanced large language model-driven collaborative collision avoidance decision-making method for intelligent ships in intersecting waters according to claim 1, characterized in that, Step S4 specifically includes: S41. Enhance retrieval for static information: Information on navigation rules and water conditions in the confluence waters is collected, cleaned, and then sliced. The cleaned text data is converted into semantic vectors using a text embedding model and stored in a vector database. The input prompt is vectorized, the k fragment vectors with the highest dot product similarity in the vector database are retrieved, the prompt vector and fragment vector are fused to enhance the input, and the enhanced context is sent to the large language model for decision-making. S42. Enhanced retrieval of dynamic navigation experience information: A decision-making outcome evaluation module is established to generate a semantic evaluation based on the decision-making outcome during the navigation process by calculating the conflict risk. Store the scene description, conflict description, decision result, and decision evaluation in JSON format, and establish links between the scene, action, and evaluation. The text embedding model is used to convert the text into semantic vectors and store them in a vector database. When making a decision, the k scenes with the highest similarity to the scenes in the database are retrieved. The original scene prompts are input together with the k scenes and their evaluation results and output to the large language model.

6. The tool-enhanced large language model-driven collaborative collision avoidance decision-making method for intelligent ships in intersecting waters according to claim 5, characterized in that, In step S41, the formula for calculating the dot product similarity is as follows: Among them, A i B i For the metric vector, n represents the vector dimension.

Images