Collaborative collision avoidance decision-making method for intelligent ships in intersection water area driven by tool-enhanced large language model

Abstract

The invention provides an intersection water area intelligent ship collaborative collision avoidance decision method driven by a tool enhanced large language model, which comprises the following steps: establishing a formalized model of an intersection water area multi-agent ship collaborative collision avoidance problem based on a partial observable Markov decision process; by optimizing an observation space parameter set, an action space constraint condition and a state transition probability function, an experimental framework oriented to a ship encounter scene is constructed; a center-distributed double-layer large language model decision-making framework is designed, and is used for generating a collision avoidance strategy by integrating environment perception data, coordination instructions, navigation experience and navigation rules of the intelligent ship; a tool enhancement system is constructed, a mathematical engine and a navigation knowledge base are integrated, and navigation experience, navigation rules and water area information are dynamically stored through a semantic similarity retrieval mechanism. Through theoretical modeling, architecture innovation and tool enhancement, engineering application of a large language model in the field of navigation decision making is promoted, and a new normal form is provided for solving the complex decision making problem of collaborative collision avoidance of multiple intelligent ships in an intersection water area.

Description

Technical Field

[0001] The present invention relates to the technical fields of large models and intelligent ships. Specifically, it particularly relates to a method for collaborative collision avoidance decision-making of intelligent ships in intersection waters driven by a tool-enhanced large language model. Background Art

[0002] In the field of ship navigation, intersection waters are areas with a high incidence of ship collision accidents. Intersection waters refer to the waters formed by the intersection of two or more shipping lanes, where the waterways crisscross, ships meet frequently, and navigation regulations are complex. Ships navigating in intersection waters need to follow strict navigation rules and collision avoidance rules to ensure navigation safety. Traditional ship collision avoidance decisions mainly rely on the experience and judgment of human drivers. However, with the development of ship automation and intelligence, collaborative collision avoidance decision-making for intelligent ships has gradually become a research hotspot.

[0003] Currently, the collaborative decision-making methods for solving complex collision avoidance problems in intersection waters mainly include algorithms based on optimization theory and algorithms based on machine learning. Algorithms based on optimization theory establish mathematical models to optimize collision avoidance paths and strategies. However, their modeling process is complex, with insufficient flexibility and difficulty in adapting to complex dynamic environments. Algorithms based on machine learning learn collision avoidance strategies in a data-driven manner. However, they have poor generalization ability and lack interaction capabilities, making it difficult to handle unknown navigation scenarios. In recent years, with the rise of generative pre-trained large language models (LLMs), they have demonstrated excellent capabilities in understanding, reasoning, and human-computer interaction, providing new ideas for solving collision avoidance problems in intersection waters. However, LLMs have deficiencies in continuous learning and are difficult to adapt to changing navigation environments and rules.

[0004] Although existing collaborative decision-making methods can solve the collision avoidance problem in intersection waters to a certain extent, they have many drawbacks. Algorithms based on optimization theory have a complex modeling process and are difficult to adapt to dynamically changing navigation environments. Algorithms based on machine learning have poor generalization ability and lack interaction capabilities, making it difficult to handle complex navigation scenarios. In addition, although existing large language models perform well in understanding, reasoning, and human-computer interaction, their model parameters are fixed after pre-training, lacking continuous learning ability and being difficult to adapt to changing navigation environments and rules. Therefore, how to guide large language models to make safe and reliable collaborative avoidance decisions through reasonable prompt engineering while making up for their deficiencies in continuous learning has become an urgent technical problem to be solved. Summary of the Invention

[0005] In response to the above-mentioned technical problems, a method for collaborative collision avoidance decision-making of intelligent ships in intersection waters driven by a tool-enhanced large language model is provided. The present invention guides large language models to make safe and reliable collaborative avoidance decisions through reasonable prompt engineering and a two-layer decision-making architecture, improving the reliability of collision avoidance decisions of intelligent ships in intersection waters.

[0006] The technical means adopted by the present invention are as follows:

[0007] A method for enhancing the collision avoidance decision-making of intelligent ships in converging waters driven by a large language model, including:

[0008] S1. Based on the partially observable Markov decision process, establish a formal model for the cooperative collision avoidance problem of multi-agent ships in converging waters;

[0009] S2. By optimizing the observation space parameter set, action space constraint conditions, and state transition probability function, construct an experimental framework for ship encounter scenarios;

[0010] S3. Design a central-distributed two-layer large language model decision-making architecture, including a distributed layer and a central coordination layer, for intelligent ships to generate collision avoidance strategies based on comprehensive environmental perception data, coordination instructions, navigation experience, and navigation rules;

[0011] S4. Construct a tool-enhanced system, including a scenario description module, a conflict description module, a decision-making module, and a simulation environment module, integrate a mathematical engine and a navigation knowledge base, and dynamically store navigation experience, navigation rules, and water area information through a semantic similarity retrieval mechanism.

[0012] Further, step S1 specifically includes:

[0013] S11. Define the converging waters as the waters formed by the intersection and convergence of two or more waterways, and divide them into two-way converging waters, three-way converging waters, four-way converging waters, and multi-way converging waters according to the number of branch waterway intersections

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

[0015] S13. Construct the cooperative collision avoidance problem of multi-agent ships in converging waters into a formal model based on the partially observable Markov decision process (POMDP), and represent it using a six-tuple as follows:

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

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

[0018] Furthermore, step S2 specifically includes:

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

[0020]

[0021] Among them, x i ,y i are the longitudinal and transverse position coordinates of ship i in the geographic coordinate system; v x,i ,v y,i is the velocity component in the corresponding direction; cosφ i and sinφ i Commonly represent the ship heading angle φ i The trigonometric form of , avoids the problem of angular periodic discontinuity;

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

[0023]

[0024] in, Represents 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| represents the local observation space of the smart ship V;

[0025] S23, adopting a discretization decision strategy based on speed control, the action space of the smart ship is defined as discrete semantic instructions A i ={α dec ,α hold ,α acc}, corresponding to deceleration cruise, constant speed maintenance and acceleration propulsion instructions respectively;

[0026] S24. Once the high-level decision instructions are generated, the low-level motion controller calculates the corresponding target speed, driving the intelligent ship to complete the coordinated collision avoidance maneuver. This design adheres to the navigational principle of "giving way to the vehicle more often and to the rudder less often" in converging waters, achieving safe collision avoidance through longitudinal motion control.

[0027] S25. Discretize the longitudinal propulsion control quantity of the ship into five standardized telegraph orders, and establish the mapping relationship between ship speed and telegraph order as follows:

[0028] Set the target speed as v target ∈{0, 3, 6, 9, 12} kn. When an acceleration / deceleration instruction is received, the target ship speed is adjusted step by step according to the gear;

[0029] S26. Construct a PID controller according to the deviation between the target speed and the current speed as follows:

[0030]

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

[0032] Furthermore, in step S3, in the designed center-distributed double-layer large language model decision-making architecture:

[0033] The distributed layer autonomously completes environmental perception and decision-making reasoning by the intelligent ship driven by the large language model;

[0034] The central coordination layer configures a large language model coordinator, and determines the passing sequence by using the conflict severity quantification model and the International Regulations for Preventing Collisions at Sea, 1972.

[0035] Furthermore, step S3 specifically includes:

[0036] S31. Define ship traffic conflicts (a ship in navigation is forced by another ship in space and time, resulting in a collision risk between the ship and the other ship, and the ship must take actions to avoid the collision), and divide ship traffic conflicts into crossing conflicts, head-on conflicts and following conflicts according to the conflict angle;

[0037] S32. Quantify the severity of ship traffic conflicts, including quantifying the severity of following conflicts and quantifying the severity of crossing conflicts;

[0038] S33. Utilize the outstanding characteristics of the large language model in text understanding, reasoning and human-computer interaction, establish prompt engineering, construct a collaborative collision avoidance thinking chain of state perception, intention sharing, conflict mediation and decision-making, and guide the large language model to make reliable decisions.

[0039] Further, in step S32, quantifying the severity of following collision conflicts and quantifying the severity of crossing collision conflicts specifically include:

[0040] Based on the impact of Fujii's ship domain on the relationship between ships, quantify the severity of following collision conflicts between two ships in the same waterway as follows:

[0041]

[0042] Where d represents the distance between two ships in the same waterway; t represents the time of collision between the two ships; L represents the ship length;

[0043] Crossing collisions generally occur near the conflict points in the convergence area. According to the influence of the minimum distance of approach and the minimum time of approach on the collision risk of ships, use DTCP (distance from the ship to the ship traffic conflict point) and TTCP (time for the ship to reach the conflict point) to quantify the severity of crossing collision conflicts as follows:

[0044]

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

[0046]

[0047] Where TTCP i represents and TTCP j represents the times for ships i and j to reach the ship traffic conflict point; v i and v j represent the ship speeds of intelligent ships i and j.

[0048] Further, step S4 specifically includes:

[0049] S41. Retrieve and enhance static information:

[0050] Collect information on navigation rules and water area conditions in the convergence area, clean it, slice the cleaned text data, and use a text embedding model to convert it into semantic vectors and store them in a vector database;

[0051] Vectorize the input prompt, retrieve the k fragment vectors with the largest dot product similarity in the vector database, fuse the prompt vector and the fragment vectors to enhance the input, and send the enhanced context to the large language model for decision-making;

[0052] S42. Retrieve and enhance dynamic navigation experience information:

[0053] A decision result evaluation module is established. For the decision results during the navigation process, semantic evaluations are generated by calculating the conflict risk level. For example, "This ship operation intensifies the conflict and should be avoided."

[0054] The scenario description, conflict description, decision results, and decision evaluations are stored in JSON format to establish links among scenarios, actions, and evaluations.

[0055] They are converted into semantic vectors using a text embedding model and stored in a vector database. During decision-making, the k scenarios with the highest similarity to the scenario in the database (cosine similarity) are retrieved, and the original scenario prompt input, the k scenarios, and their evaluation results are jointly output to the large language model.

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

[0057]

[0058] where A i , B i are metric vectors, and n represents the vector dimension.

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

[0060] 1. The method for intelligent ship collaborative collision avoidance decision-making in the confluence waters driven by a large language model enhanced by a tool provided by the present invention guides the large language model to make safe and reliable collaborative avoidance decisions through reasonable prompt engineering and a two-layer decision-making architecture, improving the reliability of collision avoidance decisions for intelligent ships in the confluence waters.

[0061] 2. The method for intelligent ship collaborative collision avoidance decision-making in the confluence waters driven by a large language model enhanced by a tool provided by the present invention constructs a tool-enhanced system that dynamically stores navigation experience, navigation rules, and water area information, making up for the deficiencies of the LLM in continuous learning and enhancing the model's adaptability to the changing navigation environment and rules.

[0062] 3. Through theoretical modeling, architecture innovation, and tool enhancement, the present invention promotes the engineering application of large language models in the field of navigation decision-making, providing a new paradigm for solving the complex decision-making problem of multi-intelligent ship collaborative collision avoidance in the confluence waters.

[0063] For the above reasons, the present invention can be widely promoted in the fields of large models, intelligent ships, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

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

[0066] Figure 2 This is a schematic diagram of the types of intersection waters of the present invention.

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

[0068] Figure 4 This is a schematic diagram of traffic conflict points in a four-way intersection water area.

[0069] Figure 5 This is a collaborative decision-making thinking chain diagram for intersection waters.

[0070] Figure 6 This is a retrieval-augmented generation framework diagram. Detailed implementation manners

[0071] To enable those skilled in the art of the present technology to better understand the solution of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

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

[0073] As Figure 1 shown, the present invention provides a tool-enhanced large language model-driven intelligent ship collaborative collision avoidance decision-making method for intersection waters, including:

[0074] S1. Based on the partially observable Markov decision process, establish a formal model for the collaborative collision avoidance problem of multi-agent ships in intersection waters;

[0075] S2. Construct an experimental framework for ship encounter scenarios by optimizing the observation space parameter set, action space constraint conditions, and state transition probability function;

[0076] S3. Design a central-distributed two-layer large language model decision-making architecture, including a distributed layer and a central coordination layer, for intelligent ships to generate collision avoidance strategies based on comprehensive environment perception data, coordination instructions, navigation experience, and navigation rules;

[0077] S4. Construct a tool enhancement system, including a scenario description module, a conflict description module, a decision-making module, and a simulation environment module, integrate a mathematical engine and a navigation knowledge base, and dynamically store navigation experience, navigation rules, and water area information through a semantic similarity retrieval mechanism.

[0078] When specifically implemented, as a preferred implementation manner of the present invention, step S1 specifically includes:

[0079] S11. Define the confluence waters as waters formed by the intersection and convergence of two or more shipping lanes. Ships in the confluence waters usually sail according to the shipping lanes or channels stipulated at sea. The number of confluence shipping lanes basically determines the type of ship traffic flow, as Figure 2 shown, and it is 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 shipping lane intersections

[0080] S12. Consider the navigation restrictions caused by the characteristics of the waterway and the ship itself, including water depth and waterway restrictions, ship inertia, and ship speed requirements; in this embodiment, specifically: 1) In the confluence waters, the navigable waters are restricted due to water depth and waterway restrictions. Large ships are difficult to use large-angle steering to avoid oncoming ships in the confluence waters. The priority of giving way by speed is greater than that of giving way by rudder, that is, strictly control the ship speed to avoid; 2) The ship inertia is extremely large and the action is sluggish. A safety distance of more than four times the ship's length needs to be maintained between ships to avoid collision accidents; 3) The ship needs a certain ship speed to maintain normal maneuverability. Moreover, when a large ship turns from one waterway to another, a ship speed greater than 3.5 knots is required to ensure the rudder effect.

[0081] S13. Construct the multi-agent ship cooperative collision avoidance problem in the confluence waters as a formal model based on the Partially Observable Markov Decision Process (POMDP), which is represented by a six-tuple as follows:

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

[0083] Among them, \(V\) is a multi-agent ship set, representing the collaborative decision-making entity; \(S\) is the joint state space, covering ship kinematic parameters and environmental characteristics; \(A\) is the distributed action space, defining the set of maneuvering instructions that each agent can execute; \(P\) is the state transition function, constructed based on the ship kinematic model; \(O\) i represents the local observation space of agent \(i\), reflecting the limitations of sensor information acquisition; \(R\) i is the individual reward function, integrating the goals of safety and energy efficiency optimization. In this embodiment, at decision-making moment \(t\), agent \(i\) obtains the local observation \(o\) i , \(t\in O\) i , and generates a maneuvering instruction \(a_{i,t}\in A\) through the policy \(\pi\). After multi-agent action synthesis, it triggers the state transition \(S_{t + 1}\sim P(\cdot|S_t,a_t)\). This modeling method can describe the temporal dependence and observation uncertainty of interactions between ships, enhancing the mathematical logic of centralized-distributed collaborative decision-making.

[0084] Specifically in implementation, as a preferred implementation manner of the present invention, step S2 specifically includes:

[0085] S21. Define the set of interactive ships within the sensing radius \(L\) of intelligent ship \(i\) as \(N\) i , and the observation matrix represents the local situation awareness information, where \(|N|\) i is the number of observable ships, and \(|F| = 6\) is the dimension of ship state characteristics; the single-ship state characteristic vector is defined as:

[0086]

[0087] where \(x\) i , \(y\) i are the longitudinal and lateral position coordinates of ship \(i\) in the geographic coordinate system; \(v\) x,i , \(v\) y,i are the velocity components in the corresponding directions; \(\cos\varphi\) i and \(\sin\varphi\) i together represent the trigonometric function form of the ship's heading angle \(\varphi\) i to avoid the problem of discontinuous angle periodicity;

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

[0089]

[0090] Among them, represents the observation tensor splicing operation; \(O_1\) represents the local observation space of intelligent ship 1, \(O_2\) represents the local observation space of intelligent ship 2, and \(O\) |V| represents the local observation space of intelligent ship \(V\);

[0091] S23. Aiming at the special needs of ship cooperative collision avoidance in converging waters, a discrete decision-making strategy based on speed control is adopted. Given that LLM has significant advantages in the field of knowledge-driven semantic reasoning, but its numerical analysis ability is relatively limited, the action space of the intelligent ship is defined as discrete semantic instructions A. i ={α dec ,α hold ,α acc}, corresponding to deceleration cruise, constant speed maintenance and acceleration propulsion instructions respectively;

[0092] S24. Once the high-level decision instructions are generated, the low-level motion controller calculates the corresponding target speed, driving the intelligent ship to complete the coordinated collision avoidance maneuver. This design adheres to the navigational principle of "giving way to the vehicle more often and to the rudder less often" in converging waters, achieving safe collision avoidance through longitudinal motion control.

[0093] S25. Discretize the longitudinal propulsion control quantity of the ship into five-speed standardized engine clock instructions, and establish a mapping relationship between ship speed and engine clock instructions as follows:

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

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

[0096]

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

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

[0099] The distributed layer is driven by a large language model, and the intelligent ship autonomously completes environmental perception and decision reasoning;

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

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

[0102] S31. Define ship traffic conflicts (when a navigating ship is approached by another ship in space and time, resulting in a collision risk between the ship and the other ship, and the ship must take actions to avoid the collision), as Figure 3 shown, and classify ship traffic conflicts into crossing conflicts, head-on conflicts, and following conflicts according to the conflict angle;

[0103] S32. Quantify the severity of ship traffic conflicts, including quantifying the severity of following conflicts and quantifying the severity of crossing conflicts; specifically including:

[0104] Based on the influence of Fujii's ship domain on the relationship between ships, quantify the severity of the following conflict between two ships in the same waterway as follows:

[0105]

[0106] where d represents the distance between two ships in the same waterway; t represents the time of collision between the two ships; L represents the ship length;

[0107] As Figure 4 shown, crossing conflicts generally occur near the conflict point in the convergence area. According to the influence of the minimum distance of approach and the minimum time of approach on the ship collision risk, use DTCP (distance from the ship to the ship traffic conflict point) and TTCP (time for the ship to reach the conflict point) to quantify the severity of crossing conflicts as follows:

[0108]

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

[0110]

[0111] where TTCP i represents and TTCP j represents the times for ships i and j to reach the ship traffic conflict point; v i and v j represent the ship speeds of intelligent ships i and j.

[0112] S33. Utilize the outstanding characteristics of large language models in text understanding, reasoning, and human-computer interaction to establish prompt engineering, construct a collaborative collision avoidance thinking chain of state perception, intention sharing, conflict mediation, and decision-making, and guide the large language model to make reliable decisions. As Figure 5As shown in the figure. Conduct a dual safety risk assessment before making a decision, check the set of alternative actions before the decision result, and eliminate the actions that may cause a rear-end collision or exacerbate the severity of the conflict due to insufficient safety distance between the selected action and the ship due to excessive speed.

[0113] When specifically implemented, as a preferred implementation manner of the present invention, step S4 specifically includes:

[0114] S41. Retrieve and enhance static information:

[0115] Collect information on navigation rules and water area conditions in the intersection water area, clean it, slice the cleaned text data, and convert it into semantic vectors using a text embedding model and store them in a vector database;

[0116] Vectorize the input prompt, retrieve the k segment vectors with the largest dot product similarity in the vector database, fuse the prompt vector and the segment vector to enhance the input, and send the enhanced context to the large language model for decision-making; the calculation formula of the dot product similarity is as follows:

[0117]

[0118] Among them, A i , B i is a metric vector, n represents the vector latitude, and the retrieval enhancement process is as Figure 6 shown.

[0119] S42. Retrieve and enhance dynamic navigation experience information:

[0120] Set up a decision result evaluation module, and generate a semantic evaluation by calculating the conflict risk for the decision result during the navigation process; for example: "This ship operation action exacerbates the conflict and should be avoided."

[0121] Store the scene description, conflict description, decision result, and decision evaluation in JSON format, and establish a link between the scene, action, and evaluation;

[0122] Convert it into semantic vectors using a text embedding model and store them in a vector database. During decision-making, retrieve the k scenes with the largest similarity to the scene in the database (cosine similarity), and output the original scene prompt input together with the k scenes and their evaluation results 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, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for enhancing the intelligent ship collaborative collision avoidance decision-making in confluence waters driven by a large language model with a tool, characterized in that, Including: S1. Based on the partially observable Markov decision process, establish a formal model for the collaborative collision avoidance problem of multi-agent ships in the confluence waters; S2. By optimizing the observation space parameter set, action space constraint conditions, and state transition probability function, construct an experimental framework for ship encounter scenarios; S3. Design a centralized-distributed two-layer large language model decision-making architecture, including a distributed layer and a central coordination layer, for intelligent ships to generate collision avoidance strategies based on comprehensive environmental perception data, coordination instructions, navigation experience, and navigation rules; S4. Construct a tool-enhanced system, including a scenario description module, a conflict description module, a decision-making module, and a simulation environment module, integrate a mathematical engine and a navigation knowledge base, and dynamically store navigation experience, navigation rules, and water area information through a semantic similarity retrieval mechanism.

2. The intelligent ship cooperative collision avoidance decision-making method for confluence waters driven by a tool-enhanced large language model according to claim 1, characterized in that, Step S1 specifically includes: S11. Define the confluence waters as the waters formed by the intersection and convergence of two or more shipping lanes, and divide them into two-way confluence waters, three-way confluence waters, four-way confluence waters, and multi-way confluence waters according to the number of branch shipping lane intersections; S12. Consider the navigation restrictions caused by the characteristics of the waterway and the ship itself, including water depth and waterway restrictions, ship inertia, and ship speed requirements; S13. Construct the collaborative collision avoidance problem of multi-agent ships in the confluence waters as a formal model based on the partially observable Markov decision process, and represent it using a six-tuple as follows: G = {V, S, A, P, [O i ,[R i} Among them, \(V\) is a multi-agent ship set, representing the collaborative decision-making entity; \(S\) is the joint state space, covering ship kinematic parameters and environmental characteristics; \(A\) is the distributed action space, defining the set of maneuvering instructions that each agent can execute; \(P\) is the state transition function, constructed based on the ship kinematic model; \(O\) i represents the local observation space of agent \(i\), reflecting the limitations of sensor information acquisition; \(R\) i is the individual reward function, integrating the safety and energy efficiency optimization objectives.

3. The intelligent ship cooperative collision avoidance decision-making method driven by a large language model with tool enhancement for intersecting waters according to claim 1, wherein, Step S2 specifically includes: S21. Define the set of interactive vessels within the perception radius L of the intelligent vessel i as N i , and the observation matrix represents the local situation awareness information, where |N i | is the number of observable vessels, and |F| = 6| is the dimension of the vessel state characteristics; the single-vessel state characteristic vector is defined as: where x i , y i are the longitudinal and lateral position coordinates of ship i in the geographical coordinate system; v x,i , v y,i are the velocity components in the corresponding directions; cosφ i and sinφ i together represent the trigonometric function form of the ship's heading angle φ i to avoid the problem of discontinuous angle periodicity; S22. Construct a system global observation space, which is jointly composed of the local observations of all agents, as follows: Among them, represents the operation of splicing observation tensors; O1 represents the local observation space of intelligent ship 1, O2 represents the local observation space of intelligent ship 2, and O |V| represents the local observation space of intelligent ship V; S23. Adopt a discretized decision-making strategy based on speed control, and define the action space of the intelligent ship as discrete semantic instructions A i ={α dec , α hold , α acc}, corresponding to decelerated cruising, constant-speed maintenance, and accelerated propulsion instructions respectively; S24. After the high-level decision-making instruction is generated, the low-level motion controller will calculate and generate the corresponding target speed to drive the intelligent ship to complete the collaborative collision avoidance action. S25. Discretize the longitudinal propulsion control quantity of the ship into five gears of standardized telegraph orders, and establish a speed-telegraph order mapping relationship, as follows: Set the target speed to v target ∈ {0, 3, 6, 9, 12} kn. When an acceleration / deceleration command is received, the target speed is adjusted stepwise according to the gear S26. According to the deviation between the target speed and the current speed, construct a PID controller, as follows: Among them, K p represents the proportional gain factor, e(t) represents the speed deviation, and e(t) = v target - v actual , where v target represents the target speed, and v actual represents the current speed; K i represents the integral cumulative adjustment factor, e(τ) represents the real-time speed deviation, and K d represents the differential dynamic suppression coefficient.

4. A method for enhancing the intelligent ship collision avoidance decision-making in the confluence waters driven by a large language model for a tool, characterized in that, In step S3, in the designed centralized-distributed two-layer large language model decision-making architecture: The distributed layer is driven by a large language model to autonomously complete environmental perception and decision-making inference for intelligent ships; The central coordination layer is configured with a large language model coordinator, and uses a conflict severity quantification model and the "International Regulations for Preventing Collisions at Sea, 1972" to determine the passing sequence.

5. A method for enhancing the intelligent ship cooperative collision avoidance decision-making in the confluence waters driven by a large language model for a tool, characterized in that, Step S3 specifically includes: S31. Define ship traffic conflicts, and divide ship traffic conflicts into crossing conflicts, head-on conflicts, and following conflicts according to the conflict angle; S32. Quantify the severity of ship traffic conflicts, including quantifying the severity of following conflicts and quantifying the severity of crossing conflicts; S33. Utilize the outstanding characteristics of the large language model in text understanding, reasoning, and human-computer interaction, establish prompt engineering, construct a collaborative collision avoidance thinking chain of state perception, intention sharing, conflict mediation, and decision-making, and guide the large language model to implement reliable decisions.

6. The intelligent ship cooperative collision avoidance decision-making method driven by a large language model with tool enhancement for intersecting waters according to claim 1, wherein In step S32, quantifying the severity of following conflicts and quantifying the severity of crossing conflicts specifically includes: Based on the influence of Fujii's ship domain on the relationship between ships, quantify the severity of the following conflict between two ships in the same waterway, as follows: Where d represents the distance between two ships in the same waterway; t represents the time of ship collision; L represents the ship length; Cross conflicts generally occur near the conflict points in the convergence area. According to the influence of the minimum distance of approach and the minimum time of approach on the ship collision risk, DTCP (distance from ship to ship conflict point) and TTCP (time for ship to reach the conflict point) are used to quantify the severity of cross conflicts as follows: where DTCP i and DTCP j represent the distances from intelligent ships i and j to the ship traffic conflict point respectively; ΔTTCP represents the time difference to reach the ship traffic conflict point, and its calculation formula is as follows: Among them, TTCP i represents and TTCP j represents the times of ships i and j arriving at the ship traffic conflict point; v i and v j represent the ship speeds of intelligent ships i and j.

7. A collaborative collision avoidance decision-making method for intelligent ships in confluence waters driven by a tool-enhanced large language model according to claim 1, characterized in that, Step S4 specifically includes: S41. Perform retrieval enhancement on static information: Collect information on navigation rules and water area conditions in the convergence water area, clean it, slice the cleaned text data, and use a text embedding model to convert it into semantic vectors and store them in a vector database; Vectorize the input prompt, retrieve the k segment vectors with the largest dot product similarity in the vector database, fuse the prompt vector and the segment vectors to enhance the input, and send the enhanced context to the large language model for decision-making; S42. Perform retrieval enhancement on dynamic navigation experience information: Set up a decision result evaluation module to generate a semantic evaluation by calculating the conflict risk for the decision results during the navigation process; Store the scenario description, conflict description, decision result, and decision evaluation in JSON format, and establish a link between the scenario, action, and evaluation; Use a text embedding model to convert it into semantic vectors and store them in a vector database. During decision-making, retrieve the k scenarios with the highest similarity to the scenario in the database, and output the original scenario prompt input together with the k scenarios and their evaluation results to the large language model.

8. A method for enhancing the intelligent ship collaborative collision avoidance decision-making in the confluence waters driven by a large language model for tools, characterized in that, In step S41, the calculation formula for the dot product similarity is as follows: Among them, A i , B i are metric vectors, and n represents the vector dimension.

Images