An evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation

Abstract

The application discloses an evolution simulation method based on multi-agent cognitive modeling and graph network propagation, constructs a unified state space by acquiring port multi-source data, constructs an image for a navigation subject based on the space and generates a cognitive agent; the agent is connected to a large language model to generate an optimal manipulation intention, the intention is mapped into a continuous control quantity, and the continuous control quantity is subjected to the action of a physical response model and a residual correction term and is projected into a safe feasible region to generate a predicted motion state of a ship; a time-varying navigation interaction graph is constructed, and multi-ship interaction risks integrated with multiple factors are calculated, according to the interaction graph and the risks, a risk ship is determined through a selection function, a directed risk message is generated through a message function, and the risk state of an affected ship is updated through a state update function; a closed-loop verification loss is constructed according to the updated risk state and historical data, the model is iteratively corrected and updated, a risk propagation chain and a regional risk hot area are output, and the authenticity and deducibility of navigation risk evolution are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of computer simulation technology, specifically to an evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation. Background Technology

[0002] In simulation analysis of navigation safety in complex port areas, existing technologies typically simplify ships as trajectory points, using AIS data and indicators such as CPA / TCPA to determine collision risks, and relying on rule-based kinematic models or finite state machines to extrapolate the navigation process. However, navigation risks in port areas are not generated independently by a single ship, but rather propagate and amplify dynamically in a multi-ship interaction network under the coupled effects of multiple factors such as "people-ship-environment-rules-relationships". Existing methods are limited by their model architecture. On the one hand, they lack the ability to model the cognitive differences and decision-making preferences of navigation subjects such as captains and pilots, making it difficult to generate interpretable maneuvering behaviors that conform to real navigation experience, resulting in insufficient credibility of simulation scenarios. On the other hand, their risk assessments mostly focus on local, instantaneous encounter situations between two ships, lacking an effective mechanism to describe how local disturbances spread along the multi-ship relationship network to a regional scale and form systemic risks, thus making it difficult to clearly characterize and extrapolate the risk propagation chain and the scope of impact. Summary of the Invention

[0003] The purpose of this invention is to provide an evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation to solve the problems mentioned in the background art.

[0004] An evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation includes: Acquire basic port area data, vessel navigation status data, navigation rules data, environmental data, and historical trajectory data to construct a complex port area navigation scenario with a unified state space representation; Based on a unified state space, construct a main profile vector for the captain, pilot and VTS supervisor, and generate a navigation subject cognitive intelligent agent; By integrating the general aviation cognitive agent into a large language model, a set of candidate manipulation intentions is generated under structured prompt constraints, and the optimal manipulation intention is obtained. The optimal maneuvering intention is converted into a continuous control quantity through a mapping function from intention to control quantity. This control quantity is then processed by the physical motion response model and the historical trajectory residual correction term, and projected onto the safe and feasible region consisting of the channel boundary and risk threshold to generate the predicted motion state of the ship at the next moment. Using ship intelligent agents as nodes and encounter relationships, relative motion relationships, traffic organization relationships, and risk impact relationships as edges, a time-varying navigation interaction graph is constructed. Based on the time-varying navigation interaction graph, the multi-ship interaction risk that integrates encounter types, environmental disturbances, and differences in behavioral intentions is calculated. Based on the time-varying navigation interaction diagram and the multi-ship interaction risk, the risky vessel and its affected vessel set are determined by the selection function. Based on the risky vessel and its affected vessel set, a directed risk message from the risky vessel to the affected vessel is generated by the message function. Based on the directed risk message, the risk status of the affected vessel is updated by the state update function. Based on the updated risk status of affected vessels, historical AIS trajectories, VTS handling records, and expert review opinions, a closed-loop verification loss is constructed. The predicted motion status, motion model parameters, and main body profile of vessels are iteratively corrected based on the closed-loop verification loss, and the next round of simulation is updated to output the risk propagation chain and regional risk hotspots.

[0005] Furthermore, by acquiring basic port area data, vessel navigation status data, navigation rule data, environmental data, and historical trajectory data, a complex port area navigation scenario with a unified state space representation is constructed, including: Acquire basic port area data, including channel boundary data, restricted navigation area data, anchorage data, berth data, and construction area data, and generate a port area spatial constraint field based on the basic port area data; The system acquires ship navigation status data, which includes ship planar position, speed, heading, draft, length, beam, ship type, maneuverability, and hazardous materials attributes. Based on the ship navigation status data, it generates a ship motion and attribute state vector. Acquire navigation rule data, which includes traffic organization status, navigation control rules, pilotage plans and tugboat resource information, and generate operational constraint vectors based on the navigation rule data; Acquire environmental data, including wind field data, wave field data, flow velocity and direction field data, and visibility data, and generate environmental field vectors based on the environmental data; Acquire historical trajectory data, including AIS historical trajectory records and VTS handling records, and generate a historical behavior memory bank based on the historical trajectory data; By combining the ship motion and attribute state vectors, environmental field vectors, port area spatial constraint field, operational constraint vectors, historical behavior memory, and observation input vectors consisting of the relative positional relationships, encounter relationships, and rule constraint relationships between each ship agent and surrounding ship agents, a complex port area navigation scenario with a unified state space expression is constructed.

[0006] Furthermore, based on a unified state space, a cognitive intelligent agent containing subject profile vectors is constructed for the captain, pilot, and VTS supervisor, generating a navigation subject cognitive intelligent agent, including: Obtain a historical behavior memory bank in a unified state space, extract historical navigation experience data and typical encounter handling pattern data of each navigation entity based on the historical behavior memory bank, and generate experience pattern features. Obtain the operational constraint vector in the unified state space, extract the understanding level data of each navigation entity on navigation rules and port area organization constraints based on the operational constraint vector, and generate rule understanding ability characteristics; Obtain the historical behavior memory bank and ship motion and attribute state vector in a unified state space. Extract the historical maneuvering habit data and response time delay feature data of each navigation subject based on the historical behavior memory bank and ship motion and attribute state vector, and generate maneuvering habit features. Based on the characteristics of experience patterns, rule understanding ability, and operational habits, and combined with the risk preference parameters and organizational role parameters of each general aviation entity, a subject profile vector is constructed that includes experience patterns, rule understanding ability, risk preference, organizational role, and operational habits. Based on the subject profile vector and the observation input vector, a general aviation subject cognitive agent is generated. The general aviation subject cognitive agent contains a cognitive decision function, which is used to output a set of candidate manipulation intentions based on the subject profile vector and the observation input data.

[0007] Furthermore, the general aviation cognitive agent is integrated into a large language model to generate a set of candidate manipulation intentions under structured prompt constraints, thereby obtaining the optimal manipulation intention, including: Obtain the subject profile vector and observation input data of the general aviation subject cognitive intelligent agent, and construct a structured prompt template based on the subject profile vector and observation input data. The structured prompt template includes general aviation subject role information, current situation description information, general aviation rule constraint information, and historical behavior memory data in the historical behavior memory bank. The structured prompt template is input into the large language model. Based on the large language model, a set of candidate maneuvering intentions is generated under the constraints of the structured prompt template. The set of candidate maneuvering intentions includes intentions to maintain course, intentions to fine-tune course, intentions to avoid obstacles in advance, intentions to slow down and wait, intentions to keep right, intentions to request VTS coordination, intentions to postpone arrival, and intentions to adjust berthing and departure order. Obtain the rule compliance score, maneuverability score, and conflict suppression score for each candidate maneuvering intention in the candidate maneuvering intention set. The rule compliance score is determined based on the degree of compliance between the candidate maneuvering intention and the navigation rules and organizational instructions. The maneuverability score is determined based on the degree to which the candidate maneuvering intention meets the ship's maneuverability and safety boundaries. The conflict suppression score is determined based on the intensity of multi-ship conflicts that the candidate maneuvering intention may cause. Based on the rule compliance score, manipulation feasibility score, and conflict suppression score, and combined with the confidence probability of each candidate manipulation intention by the large language model, a weighted comprehensive score is given for each candidate manipulation intention. The candidate manipulation intent with the highest score is selected from the set of candidate manipulation intents based on the weighted comprehensive score, and then selected as the optimal manipulation intent.

[0008] Furthermore, the optimal maneuvering intention is converted into a continuous control quantity through a mapping function from intention to control quantity. This control quantity is then processed by the physical motion response model and the historical trajectory residual correction term, and projected onto the safe and feasible region comprised of the channel boundary and risk threshold to generate the predicted motion state of the ship at the next moment, including: Obtain the optimal maneuvering intention and the subject profile vector of the general aviation subject cognitive intelligent agent. Based on the optimal maneuvering intention and the subject profile vector, convert the optimal maneuvering intention into a continuous control quantity containing longitudinal acceleration and steering angular velocity through the mapping function from intention to control quantity. Obtain the continuous control quantity and the environmental field vector in the unified state space. Based on the continuous control quantity and the environmental field vector, generate the physical motion prediction state that satisfies the ship size, inertia, rudder effect and acceleration / deceleration boundaries through the physical motion response model. Obtain the physical motion prediction state and the historical behavior memory bank in the unified state space. Based on the physical motion prediction state and the historical behavior memory bank, generate the residual correction prediction state, which includes corrections for port channel geometry and local environmental disturbances, through the historical trajectory residual correction term. Obtain the residual correction prediction state and the port area spatial constraint field and operational constraint vector in the unified state space. Based on the residual correction prediction state, the port area spatial constraint field and operational constraint vector, project the residual correction prediction state onto the safe and feasible domain composed of the channel boundary, restricted navigation area, berth front edge and risk threshold through the risk constraint projection function to generate the ship's predicted motion state at the next moment.

[0009] Furthermore, using ship intelligent agents as nodes and encounter relationships, relative motion relationships, traffic organization relationships, and risk impact relationships as edges, a time-varying navigation interaction graph is constructed. Based on this graph, multi-ship interaction risks that integrate encounter types, environmental disturbances, and differences in behavioral intentions are calculated, including: Obtain the predicted motion state of the ship, and calculate the relative position and relative velocity between any two ship agents based on the predicted motion state and the ship motion and attribute state vector. Calculate the nearest meeting time and the nearest meeting distance based on relative position and relative speed; Obtain the optimal maneuvering intention and the environmental field vector in the unified state space. Based on the nearest encounter time, the nearest encounter distance, the optimal maneuvering intention, and the environmental field vector, and combined with the encounter type correction coefficient, calculate the multi-ship interaction risk of the fusion distance risk item, the time urgency risk item, the encounter type risk item, the environmental risk factor, and the degree of conflict of behavioral intentions. Based on the risk of multi-ship interaction, and combining spatial and temporal scale parameters, the interaction edge weight between any two ship agents is calculated. Using ship intelligent agents as nodes and interaction edge weights as edges, a time-varying navigation interaction graph containing an adjacency matrix and an interaction strength matrix is ​​constructed.

[0010] Furthermore, based on the time-varying navigation interaction diagram and multi-ship interaction risks, a selection function is used to determine the at-risk vessels and their affected vessel sets, including: Obtain the interaction edge weights and multi-ship interaction risks in the time-varying navigation interaction graph. Based on the interaction edge weights and multi-ship interaction risks, and combined with the initial risk state of each ship agent based on the current interaction relationship and the navigation resource occupation risk, calculate the risk ship discrimination score of each ship agent. Obtain the optimal maneuvering intention and the operational constraint vector in the unified state space. Based on the optimal maneuvering intention and the operational constraint vector, and combined with the degree of conflict between the intentions of surrounding vessels, update the risk vessel identification score. Based on the risk vessel identification score and the preset risk vessel identification threshold, ship agents with scores exceeding the threshold are selected as the risk vessel set; Based on the risk vessel set and the interaction edge weights in the time-varying navigation interaction graph, ship agents that have interaction edge weights with each risk vessel exceeding a preset impact threshold are selected as the affected vessel set corresponding to each risk vessel.

[0011] Furthermore, based on the at-risk vessel and its set of affected vessels, a directed risk message is generated from the at-risk vessel to the affected vessels using a message function. Based on the directed risk message, the risk status of the affected vessels is updated using a state update function, including: The system obtains the multi-ship interaction risk, optimal maneuvering intention, and environmental field vector in the unified state space for each risky ship in the risky ship set. Based on the multi-ship interaction risk, optimal maneuvering intention, and environmental field vector, and combined with the interaction edge weights between the risky ship and the affected ship, a directed risk message from the risky ship to the affected ship is generated through a message function. The directed risk message contains the interaction risk information between the two ships, the candidate maneuvering intention conflict information, and the environmental disturbance amplification coefficient. Obtain the set of directed risk messages corresponding to each affected vessel. Based on the set of directed risk messages, aggregate each directed risk message in the set of directed risk messages using a message aggregation function to generate an aggregated risk message. Obtain the individual risk status and aggregated risk messages of the affected vessels. Based on the individual risk status and aggregated risk messages, and combined with the main risk response coefficient and message coupling amplification coefficient, update the individual risk status of the affected vessels through the status update function. Based on the updated individual risk status of each affected vessel and the interaction intensity matrix in the time-varying navigation interaction diagram, the regional navigation risk is calculated using the regional risk aggregation function.

[0012] Furthermore, based on the updated risk status of the affected vessels, historical AIS tracks, VTS handling records, and expert review opinions, a closed-loop verification of losses is constructed, including: Obtain the predicted motion state of the ship and its historical AIS trajectory. Calculate the weighted state error norm based on the deviation between the predicted motion state of the ship and the actual ship state at the corresponding moment in the historical AIS trajectory. Obtain the updated risk status and VTS handling records of the affected vessels, and calculate the risk trend error based on the deviation between the updated risk status of the affected vessels and the observed risk at the corresponding time in the VTS handling records. Obtain the optimal manipulation intention and expert review opinions, and calculate the behavior label deviation loss based on the deviation between the optimal manipulation intention and the reasonable behavior labels recognized in the expert review opinions; Obtain time-varying navigation interaction maps and historical AIS tracks; calculate the relationship network deviation based on the deviation between the time-varying navigation interaction maps and the real multi-ship relationship map constructed based on historical AIS tracks. Based on the weighted state error norm, risk trend error, behavioral label bias loss, and relationship network bias, a closed-loop verification loss is constructed.

[0013] Furthermore, based on the closed-loop verification loss, the predicted motion state, motion model parameters, and main body profile of the ship are iteratively corrected, and the next round of simulation is updated, outputting the risk propagation chain and regional risk hotspots, including: Obtain the closed-loop verification loss and the predicted motion state of the ship. Based on the closed-loop verification loss, perform state-level correction on the predicted motion state of the ship and generate the state-corrected predicted motion state of the ship. The ship's predicted motion state and closed-loop verification loss after state correction are obtained. Based on the ship's predicted motion state and closed-loop verification loss after state correction, the parameters of the physical motion response model and the parameters of the historical trajectory residual correction term are corrected at the parameter level to generate updated motion model parameters. The updated motion model parameters and closed-loop verification loss are obtained. Based on the updated motion model parameters and closed-loop verification loss, the empirical pattern features, rule understanding ability features, risk preference parameters and manipulation habit features in the subject profile vector are corrected at the parameter level to generate the updated subject profile. Obtain the state-corrected predicted motion state of the ship, the updated motion model parameters, and the updated subject profile, and write these data back into the historical behavior memory bank for the next round of simulation. Based on the updated risk status of affected vessels and the time-varying navigation interaction diagram in the next round of simulation, the directed risk message transmission path between each risk vessel in the risk vessel set and the affected vessel set is extracted to generate a risk propagation chain. Based on the regional navigation risks in the next round of simulations, regional risk hotspots are generated and output.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention provides an evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation. It constructs a unified state space by acquiring multi-source data from port areas, and then uses this space to create profiles of navigation entities and generate cognitive agents. The agents are integrated into a large language model to generate optimal maneuvering intentions, which are then mapped into continuous control variables. These variables are processed by a physical response model and residual correction terms, projected onto the safe and feasible region, and generate predicted ship motion states. A time-varying navigation interaction graph is constructed, and the multi-ship interaction risks incorporating multiple factors are calculated. Based on the interaction graph and risks, risky ships are identified through a selection function, directed risk messages are generated through a message function, and the risk states of affected ships are updated through a state update function. A closed-loop verification loss is constructed based on the updated risk states and historical data. The model is iteratively corrected and updated, outputting the risk propagation chain and regional risk hotspots, effectively improving the realism and predictability of navigation risk evolution. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the principle of the evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation, as described in a specific implementation. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0017] Please see Figure 1 This invention provides a technical solution. The technical solution of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0018] An evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation includes: Acquire basic port area data, vessel navigation status data, navigation rules data, environmental data, and historical trajectory data to construct a complex port area navigation scenario with a unified state space representation; Based on a unified state space, construct a main profile vector for the captain, pilot and VTS supervisor, and generate a navigation subject cognitive intelligent agent; By integrating the general aviation cognitive agent into a large language model, a set of candidate manipulation intentions is generated under structured prompt constraints, and the optimal manipulation intention is obtained. The optimal maneuvering intention is converted into a continuous control quantity through a mapping function from intention to control quantity. This control quantity is then processed by the physical motion response model and the historical trajectory residual correction term, and projected onto the safe and feasible region consisting of the channel boundary and risk threshold to generate the predicted motion state of the ship at the next moment. Using ship intelligent agents as nodes and encounter relationships, relative motion relationships, traffic organization relationships, and risk impact relationships as edges, a time-varying navigation interaction graph is constructed. Based on the time-varying navigation interaction graph, the multi-ship interaction risk that integrates encounter types, environmental disturbances, and differences in behavioral intentions is calculated. Based on the time-varying navigation interaction diagram and the multi-ship interaction risk, the risky vessel and its affected vessel set are determined by the selection function. Based on the risky vessel and its affected vessel set, a directed risk message from the risky vessel to the affected vessel is generated by the message function. Based on the directed risk message, the risk status of the affected vessel is updated by the state update function. Based on the updated risk status of affected vessels, historical AIS trajectories, VTS handling records, and expert review opinions, a closed-loop verification loss is constructed. The predicted motion status, motion model parameters, and main body profile of vessels are iteratively corrected based on the closed-loop verification loss, and the next round of simulation is updated to output the risk propagation chain and regional risk hotspots.

[0019] In this embodiment, port area basic data is used to define spatial constraints such as channels, anchorages, and berths, and can be obtained through the port area geographic information system. Vessel navigation status data characterizes the vessel's position, speed, heading, and dimensions, and is provided by the AIS system. Navigation rule data includes operational constraints such as traffic organization, control instructions, and pilotage plans, and originates from the VTS management system. Environmental data describes external disturbance conditions such as wind, waves, currents, and visibility, and can be obtained from meteorological and hydrological services. Historical trajectory data records past vessel movement patterns and VTS handling cases, and is stored in a historical log database. A unified state space integrates the above multi-source heterogeneous data into a structured description, providing a consistent data foundation for subsequent simulations and deductions.

[0020] The subject profile vector represents the personalized cognitive characteristics of a navigation subject in terms of experience patterns, risk preferences, operational habits, and role positioning. Its parameters can be initially set based on historical behavior records and expert knowledge. The navigation subject cognitive intelligent agent refers to a computational model that integrates the subject profile vector with the current observation input. It can simulate the situational awareness and decision-making process of a real captain, pilot, or VTS supervisor, transforming the simulated object from an indistinguishable trajectory point into a decision-making subject with anthropomorphic judgment capabilities.

[0021] Structured cue constraints are input instructions that organize current scenario information, subject role definitions, navigation rules, and decision objectives into a preset format. These instructions guide the large language model in reasoning within the port navigation professional framework. Based on this, the large language model generates a set of candidate maneuvering intentions containing multiple actionable actions. The optimal maneuvering intention is selected from this set through a built-in evaluation mechanism that comprehensively considers the candidate intention's compliance with navigation rules, its feasibility to the ship's maneuvering boundaries, and the potential risk of inducing new conflicts.

[0022] The intention-to-control quantity mapping function is used to parse semantic maneuvering intentions into continuous control quantities including acceleration and steering angular velocity. In specific implementations, this mapping function can be constructed in various ways, such as establishing rule-based lookup table mappings or parametric mappings, or it can be derived physically, calculating the required control quantities using kinematic formulas based on the desired speed change and steering angle, combined with the ship's current motion state. The physical motion response model can be implemented using a simplified kinematic model or an MMG ship maneuvering model, simulating the ship's dynamic response based on the control quantities. The historical trajectory residual correction term can be constructed by learning the ship's historical motion deviations under similar scenarios, used to locally correct the output of the physical model. The safe and feasible region is jointly defined by the channel boundary and a preset risk threshold, ensuring through constrained projection that the final generated predicted ship motion state does not exceed the acceptable range of physical space and safety risks.

[0023] The multi-ship interaction risk calculated based on the time-varying navigation interaction diagram is based on the traditional two-ship encounter index, and further integrates the encounter situation type, real-time environmental disturbance intensity and the degree of difference in maneuvering intentions between adjacent ships, so as to obtain a more comprehensive quantitative value of multi-ship coupling risk.

[0024] The selection function can be constructed based on risk thresholds or sorting methods to identify vessels whose risk status or influence exceeds a preset threshold as risk vessels, and then filter out a set of closely related affected vessels based on the edge weights in the interaction graph. The message function can generate directed risk messages using vector concatenation or weighted summation. This message encapsulates the risk attributes of the risk source, environmental disturbance factors, and the correlation strength between the two vessels, and specifies the transmission direction from the risk source to the affected vessels. The state update function can be implemented using gated loop units or graph message passing networks, enabling affected vessels to update their own risk status according to a fusion function after receiving messages from one or more risk sources, thus achieving dynamic simulation of the propagation and superposition of risk along the relationship network.

[0025] The closed-loop verification loss is constructed by comprehensively comparing the deviations between the predicted state and risk state output by the simulation and the historical AIS trajectory, VTS handling records and expert review opinions. It is used to adjust the motion model parameters and the subject profile vector in reverse to correct the system error of the simulator.

[0026] This embodiment integrates multi-source data through a unified state space to support the construction of the basic scenario. Based on this space, a subject profile and cognitive agent are constructed, enabling the simulated object to possess anthropomorphic decision-making capabilities. The agent connects to a large language model and generates optimal intent under structured prompts and constraints, achieving a combination of rules and semantics. This intent undergoes mapping transformation, physical response, residual correction, and safety projection to generate an executable predicted motion state. Furthermore, a time-varying interaction graph constructed with ships as nodes and various relationships as edges quantifies the coupling impact of multiple ships and, through selection, messaging, and state update functions, realizes the dynamic transmission of risk from the source ship to the affected ship. Finally, a closed-loop verification loss constructed using historical data and expert knowledge is used to iteratively correct the model, endowing the simulator with self-correction capabilities. This embodiment combines cognitive modeling, graph network propagation, and closed-loop correction to improve the realism and long-term operational stability of the navigation risk evolution simulation.

[0027] In some embodiments, basic port area data, vessel navigation status data, navigation rule data, environmental data, and historical trajectory data are acquired to construct a complex port area navigation scenario with a unified state space representation, including: Acquire basic port area data, including channel boundary data, restricted navigation area data, anchorage data, berth data, and construction area data, and generate a port area spatial constraint field based on the basic port area data; The system acquires ship navigation status data, which includes ship planar position, speed, heading, draft, length, beam, ship type, maneuverability, and hazardous materials attributes. Based on the ship navigation status data, it generates a ship motion and attribute state vector. Acquire navigation rule data, which includes traffic organization status, navigation control rules, pilotage plans and tugboat resource information, and generate operational constraint vectors based on the navigation rule data; Acquire environmental data, including wind field data, wave field data, flow velocity and direction field data, and visibility data, and generate environmental field vectors based on the environmental data; Acquire historical trajectory data, including AIS historical trajectory records and VTS handling records, and generate a historical behavior memory bank based on the historical trajectory data; By combining the ship motion and attribute state vectors, environmental field vectors, port area spatial constraint field, operational constraint vectors, historical behavior memory, and observation input vectors consisting of the relative positional relationships, encounter relationships, and rule constraint relationships between each ship agent and surrounding ship agents, a complex port area navigation scenario with a unified state space expression is constructed.

[0028] In this embodiment, the port area spatial constraint field can be implemented by establishing a geofence or coordinate boundary set using the acquired data on channel boundaries, restricted navigation areas, anchorages, berths, and construction areas. This is used to subsequently determine whether a vessel's position is within a permitted or prohibited area. The vessel motion and attribute state vector is organized according to a preset dimensional order based on the acquired vessel planar position, speed, heading, draft, length, beam, hull type, maneuverability, and hazardous materials attributes. Each dimension corresponds to a specific data field. The operational constraint vector is encoded based on the timeliness and priority rules in traffic organization status, navigation control rules, pilotage plans, and tugboat resource information, and is used to characterize the logical constraints at the port area operational level.

[0029] The environmental field vector is obtained by interpolating wind field, wave field, current velocity and direction field, and visibility data onto a spatial grid, and is used to describe the distribution of external disturbances at various locations in the port area. The historical behavior memory bank extracts key event nodes and typical handling patterns from the acquired historical trajectory data, indexes them by time series and ship identifiers, forming a set of experience records that can be retrieved later. The observation input vector consists of the real-time calculated values ​​of the relative positional relationships, encounter relationships, and rule constraint relationships between each ship agent and surrounding ship agents. The unified state space stores the above vectors and fields according to a preset data structure, enabling subsequent steps to quickly access the corresponding ship motion and attribute state vectors, environmental field vectors, port area spatial constraint fields, operational constraint vectors, historical behavior memory bank, and observation input vectors based on ship identifiers, timestamps, or spatial locations.

[0030] This embodiment generates multiple independent vectors and fields representing physical space, ship attributes, operating rules, environmental conditions, historical memory, and real-time perception by structurally decomposing and parametrically transforming various types of port navigation data. These heterogeneous data are then organically combined into a unified state space, solving the problem of difficulty in merging multi-source data with different formats and semantics. This provides a standardized and operable input for agent modeling and computation based on this space in subsequent steps.

[0031] In some embodiments, based on a unified state space, a cognitive agent containing subject profile vectors is constructed for the captain, pilot, and VTS supervisor, generating a navigation subject cognitive agent, including: Obtain a historical behavior memory bank in a unified state space, extract historical navigation experience data and typical encounter handling pattern data of each navigation entity based on the historical behavior memory bank, and generate experience pattern features. Obtain the operational constraint vector in the unified state space, extract the understanding level data of each navigation entity on navigation rules and port area organization constraints based on the operational constraint vector, and generate rule understanding ability characteristics; Obtain the historical behavior memory bank and ship motion and attribute state vector in a unified state space. Extract the historical maneuvering habit data and response time delay feature data of each navigation subject based on the historical behavior memory bank and ship motion and attribute state vector, and generate maneuvering habit features. Based on the characteristics of experience patterns, rule understanding ability, and operational habits, and combined with the risk preference parameters and organizational role parameters of each general aviation entity, a subject profile vector is constructed that includes experience patterns, rule understanding ability, risk preference, organizational role, and operational habits. Based on the subject profile vector and the observation input vector, a general aviation subject cognitive agent is generated. The general aviation subject cognitive agent contains a cognitive decision function, which is used to output a set of candidate manipulation intentions based on the subject profile vector and the observation input data.

[0032] In this embodiment, the experience pattern feature is generated by extracting the relative motion parameters (such as relative distance change curves, relative azimuth rate of change, and turning trigger timing) of each navigation entity in typical encounter events from the historical behavior memory bank into feature vectors, and then matching them with a preset typical encounter pattern template library. The rule understanding ability feature is generated by statistically analyzing the frequency and magnitude of events that deviate from navigation rules or port area organization constraints in the historical behavior of each navigation entity based on the operational constraint vector, and then normalizing the results. The maneuvering habit feature is generated by extracting statistical quantities such as the mean and variance of the turning angular velocity and the delay time of the speed change response of each navigation entity in similar scenarios from the historical behavior memory bank and the ship motion and attribute state vector, and then parameterizing these parameters.

[0033] Risk preference parameters are determined based on the quantiles of the nearest encounter distance distribution when each navigation entity took evasive action in historical encounter events; organizational role parameters are directly set by the functional codes of the navigation entity in the port area operation system, with captains, pilots, and VTS supervisors each corresponding to different code values. The entity profile vector is composed of experience pattern characteristics, rule understanding ability characteristics, operation habit characteristics, risk preference parameters, and organizational role parameters, assembled in a preset dimensional order.

[0034] The cognitive decision function preferably adopts a multi-input, single-output mapping structure. The input is a concatenation of the subject profile vector and the observation input vector, and the output is a priority score vector for each candidate manipulation intention. This function can be implemented using a rule weight matrix or a lightweight inference model trained with domain data. The general aviation subject cognitive agent is composed of the subject profile vector and the cognitive decision function, and is used to output the score results of each candidate manipulation intention based on the current situation in subsequent steps.

[0035] This embodiment generates experiential pattern features by parameterizing encounter events in the historical behavior memory bank, generates rule understanding ability features by statistically quantifying rule deviations in the operational constraint vector, and generates manipulation habit features by parameterizing the statistics of historical manipulation behavior. Combined with risk preference parameters and organizational role parameters, a subject profile vector that reflects the personalized cognitive characteristics of the general aviation subject is constructed. This generates a general aviation subject cognitive agent containing cognitive decision functions, transforming abstract maritime experience and rule understanding into computable feature parameters and decision models. This provides a foundation for subsequent steps to generate manipulation decisions based on the agent that conform to the actual behavior patterns of general aviation subjects.

[0036] In some embodiments, the general aviation subject cognitive agent is connected to a large language model to generate a set of candidate manipulation intentions under structured prompt constraints, thereby obtaining the optimal manipulation intention, including: Obtain the subject profile vector and observation input data of the general aviation subject cognitive intelligent agent, and construct a structured prompt template based on the subject profile vector and observation input data. The structured prompt template includes general aviation subject role information, current situation description information, general aviation rule constraint information, and historical behavior memory data in the historical behavior memory bank. The structured prompt template is input into the large language model. Based on the large language model, a set of candidate maneuvering intentions is generated under the constraints of the structured prompt template. The set of candidate maneuvering intentions includes intentions to maintain course, intentions to fine-tune course, intentions to avoid obstacles in advance, intentions to slow down and wait, intentions to keep right, intentions to request VTS coordination, intentions to postpone arrival, and intentions to adjust berthing and departure order. Obtain the rule compliance score, maneuverability score, and conflict suppression score for each candidate maneuvering intention in the candidate maneuvering intention set. The rule compliance score is determined based on the degree of compliance between the candidate maneuvering intention and the navigation rules and organizational instructions. The maneuverability score is determined based on the degree to which the candidate maneuvering intention meets the ship's maneuverability and safety boundaries. The conflict suppression score is determined based on the intensity of multi-ship conflicts that the candidate maneuvering intention may cause. Based on the rule compliance score, manipulation feasibility score, and conflict suppression score, and combined with the confidence probability of each candidate manipulation intention by the large language model, a weighted comprehensive score is given for each candidate manipulation intention. The candidate manipulation intent with the highest score is selected from the set of candidate manipulation intents based on the weighted comprehensive score, and then selected as the optimal manipulation intent.

[0037] In this embodiment, the structured prompt template is constructed by filling the navigation subject role information, current situation description information, navigation rule constraint information and historical behavior memory data into the template framework in a preset field order. Each information fragment occupies a fixed position in the template, forming a standardized input format that can be parsed by a large language model.

[0038] The rule compliance score is calculated by comparing each candidate maneuvering intention with each rule entry in the navigation rule base, counting the number of compliant and conflicting rule entries, and then weighting the results with preset weights for each rule entry. The maneuverability score checks whether the maneuvering parameters required for each candidate maneuvering intention, such as the steering angle and speed range, are within the safe maneuvering boundaries of the ship's current state, and is calculated in reverse based on the degree to which these boundaries are exceeded; the greater the deviation, the lower the score. The conflict mitigation score is calculated by simulating the encounter situation with surrounding vessels after the execution of each candidate maneuvering intention, assessing the number or severity of newly generated multi-ship conflicts.

[0039] The confidence probability is the probability distribution value output by the large language model when generating each candidate manipulation intention, representing the model's inherent assessment of the rationality of each intention under the current situation. The weighted comprehensive score is obtained by multiplying the rule compliance score, manipulation feasibility score, conflict suppression score, and confidence probability by preset weight coefficients and then summing them. Each weight coefficient is preset based on the port area security strategy or expert experience.

[0040] This embodiment guides the large language model to generate a set of candidate manipulation intentions by constructing a structured prompt template that includes roles, situations, rules, and historical memories. Then, the optimal manipulation intention is selected from the set by a weighted comprehensive score of rule compliance, manipulation feasibility, conflict suppression, and confidence probability. This constrains the semantic generation capability of the large language model within the framework of general aviation safety, avoiding the model from directly generating unexecutable or rule-violating actions, and ensuring that the final output manipulation intention achieves optimal results in terms of semantic rationality, rule compliance, manipulation feasibility, and conflict suppression.

[0041] In some embodiments, the optimal maneuvering intention is converted into a continuous control quantity through a mapping function from intention to control quantity. This control quantity is then processed by a physical motion response model and a historical trajectory residual correction term, and projected onto a safe and feasible region comprised of channel boundaries and risk thresholds to generate the predicted ship motion state for the next moment, including: Obtain the optimal maneuvering intention and the subject profile vector of the general aviation subject cognitive intelligent agent. Based on the optimal maneuvering intention and the subject profile vector, convert the optimal maneuvering intention into a continuous control quantity containing longitudinal acceleration and steering angular velocity through the mapping function from intention to control quantity. Obtain the continuous control quantity and the environmental field vector in the unified state space. Based on the continuous control quantity and the environmental field vector, generate the physical motion prediction state that satisfies the ship size, inertia, rudder effect and acceleration / deceleration boundaries through the physical motion response model. Obtain the physical motion prediction state and the historical behavior memory bank in the unified state space. Based on the physical motion prediction state and the historical behavior memory bank, generate the residual correction prediction state, which includes corrections for port channel geometry and local environmental disturbances, through the historical trajectory residual correction term. Obtain the residual correction prediction state and the port area spatial constraint field and operational constraint vector in the unified state space. Based on the residual correction prediction state, the port area spatial constraint field and operational constraint vector, project the residual correction prediction state onto the safe and feasible domain composed of the channel boundary, restricted navigation area, berth front edge and risk threshold through the risk constraint projection function to generate the ship's predicted motion state at the next moment.

[0042] In this embodiment, the mapping function from intention to control quantity matches the semantic category of the optimal maneuvering intention with a preset maneuvering parameter lookup table, and then personalizes the matched parameters by combining the maneuvering habit features in the subject profile vector, ultimately outputting a continuous control quantity including longitudinal acceleration and steering angular velocity. The preset maneuvering parameter lookup table is pre-established based on ship type, size, and maneuvering performance data, establishing a mapping relationship between each semantic intention and the corresponding acceleration and steering angular velocity ranges. The mapping values ​​can be obtained through ship maneuverability tests or simulation calibration.

[0043] The physical motion response model, based on the continuous control variables and the wind, wave and current disturbances in the environmental field vector, takes the current ship motion state as the starting point and uses dynamic integration to recursively calculate the ship's position, speed and heading at the next moment, generating a physical motion prediction state that satisfies the ship's dimensions, inertia, rudder effectiveness and acceleration / deceleration boundaries.

[0044] The historical trajectory residual correction term retrieves past navigation segments from the historical behavior memory bank that are similar to the scene in which the physical motion prediction state is located, extracts the deviation pattern between the actual trajectory and the physical model prediction trajectory in the segment, and superimposes the deviation pattern onto the physical motion prediction state to generate a residual correction prediction state that includes corrections for port channel geometry and local environmental disturbances.

[0045] The risk-constrained projection function checks whether the spatial location of the residual-corrected predicted state exceeds the channel boundary or restricted navigation area defined by the port area spatial constraint field, and whether the risk index exceeds the risk threshold set by the operational constraint vector. For the portion exceeding the constraints, it performs a backtracking correction within the constraint boundary, generating the predicted motion state of the ship at the next moment. The risk threshold set by the operational constraint vector is determined according to the port area safety management system, channel grade, and safety standards in navigation rules, and different threshold values ​​can be set for different waterways and different ship types.

[0046] This embodiment transforms semantic manipulation intent into continuous control quantities through a mapping function from intent to control quantities. It simulates the dynamic response of the ship under control quantities and environmental disturbances through a physical motion response model. It introduces data-driven deviation correction through historical trajectory residual correction terms to compensate for modeling errors in the physical model. Finally, it ensures that the final state does not violate spatial boundaries and safety thresholds through a risk constraint projection function. This transforms the semantic intent of the cognitive decision-making layer into an executable ship motion state that conforms to physical laws, has been corrected by historical data, and meets safety constraints.

[0047] In some embodiments, a time-varying navigation interaction graph is constructed using ship intelligent agents as nodes and encounter relationships, relative motion relationships, traffic organization relationships, and risk impact relationships as edges. Based on this time-varying navigation interaction graph, multi-ship interaction risks that integrate encounter types, environmental disturbances, and differences in behavioral intentions are calculated, including: Obtain the predicted motion state of the ship, and calculate the relative position and relative velocity between any two ship agents based on the predicted motion state and the ship motion and attribute state vector. Calculate the nearest meeting time and the nearest meeting distance based on relative position and relative speed; Obtain the optimal maneuvering intention and the environmental field vector in the unified state space. Based on the nearest encounter time, the nearest encounter distance, the optimal maneuvering intention, and the environmental field vector, and combined with the encounter type correction coefficient, calculate the multi-ship interaction risk of the fusion distance risk item, the time urgency risk item, the encounter type risk item, the environmental risk factor, and the degree of conflict of behavioral intentions. Based on the risk of multi-ship interaction, and combining spatial and temporal scale parameters, the interaction edge weight between any two ship agents is calculated. Using ship intelligent agents as nodes and interaction edge weights as edges, a time-varying navigation interaction graph containing an adjacency matrix and an interaction strength matrix is ​​constructed.

[0048] In this embodiment, the relative position between any two ship agents is calculated by the vector difference between their position coordinates, and the relative speed is calculated by the vector difference between their velocity vectors. The closest encounter time and closest encounter distance are derived from the relative positions, relative speeds, and headings of the two ships using geometric kinematics methods, thus determining the location of the closest approach point in space and the time required to reach that point.

[0049] The risk of multi-ship interaction is calculated by integrating distance risk, time urgency risk, encounter type risk, environmental risk factors, and degree of conflict of behavioral intentions. It is expressed by formula (1) as follows: ; In formula (1), For the first The ship and the first A ship in Risks associated with multi-ship interactions at any given time; This is a normalization function used to map the overall risk value to the [0,1] interval; This is the weighting coefficient for the distance risk term, used to adjust the contribution of distance risk to the total risk; Based on the nearest meeting distance Calculated distance risk term; This is the weighting coefficient for the time urgency risk item, used to adjust the contribution of time urgency risk to the total risk; Based on the most recent meeting time The time urgency and risk factors for calculation; This is the weighting coefficient for the encounter type risk item, used to adjust the contribution of encounter type risk to the total risk; For the first The ship and the first A ship in The risk item for the type of encounter at any given moment indicates the risks brought about by different encounter scenarios such as intersection, meeting, and overtaking. These are the weighting coefficients for environmental risk factors, used to adjust the contribution of environmental risk to the total risk. for Environmental risk factors at any given time represent the amplification or reduction effect of environmental conditions such as wind, waves, currents, and visibility on risks. This is a weighting coefficient for the degree of conflict in behavioral intentions, used to adjust the contribution of conflict in intentions to the total risk; For the first The ship and the first A ship in The degree of conflict in behavioral intentions at any given moment indicates the difference between the current or anticipated maneuvering intentions of the two ships.

[0050] Furthermore, the weights of the interactive edges are expressed by formula (2) as follows: ; In formula (2), For the first The ship and the first A ship in The weight of the interaction edge at any given moment; This is a spatial scale parameter used to control the decay rate of the distance to the edge weights; This is a time scale parameter used to control the decay rate of edge weights based on time urgency. This is the encounter type correction coefficient, used to correct the edge weights according to different encounter types (such as intersection and pair).

[0051] The encounter type correction factor is determined based on the relative bearing and heading angle between the two ships. Different types of encounters correspond to different correction factor values. The spatial scale parameter is set based on the width of the port channel and the ship density, while the temporal scale parameter is set based on the ship maneuvering response time and the encounter warning advance time.

[0052] The time-varying navigation interaction graph is constructed with ship intelligent agents as nodes and interaction edge weights as edges. The adjacency matrix generates binary connection relationships by judging whether the interaction edge weights exceed a preset connection threshold, and the interaction strength matrix directly stores the weight values ​​of each edge.

[0053] This embodiment calculates the relative position, relative speed, nearest encounter time, and nearest encounter distance between two ships using geometric kinematics. By integrating multi-dimensional risk calculation formulas that consider distance risk, time urgency, encounter type, environmental disturbances, and intentional conflict, it obtains the multi-ship interaction risk. By introducing exponential decay functions of spatial and temporal scale parameters combined with encounter type correction coefficients, it calculates the interaction edge weights. Finally, it constructs a time-varying navigation interaction graph containing an adjacency matrix and an interaction intensity matrix, quantifying the dynamic interaction relationships between ships into a computable graph structure. This provides a structured relationship model for subsequent risk propagation analysis based on this graph.

[0054] In some embodiments, based on time-varying navigation interaction diagrams and multi-ship interaction risks, a selection function is used to determine at-risk vessels and their affected vessel sets, including: Obtain the interaction edge weights and multi-ship interaction risks in the time-varying navigation interaction graph. Based on the interaction edge weights and multi-ship interaction risks, and combined with the initial risk state of each ship agent based on the current interaction relationship and the navigation resource occupation risk, calculate the risk ship discrimination score of each ship agent. Obtain the optimal maneuvering intention and the operational constraint vector in the unified state space. Based on the optimal maneuvering intention and the operational constraint vector, and combined with the degree of conflict between the intentions of surrounding vessels, update the risk vessel identification score. Based on the risk vessel identification score and the preset risk vessel identification threshold, ship agents with scores exceeding the threshold are selected as the risk vessel set; Based on the risk vessel set and the interaction edge weights in the time-varying navigation interaction graph, ship agents that have interaction edge weights with each risk vessel exceeding a preset impact threshold are selected as the affected vessel set corresponding to each risk vessel.

[0055] In this embodiment, the initial risk state of each ship agent based on the current interaction relationship is derived from the aggregation of multi-ship interaction risks in the current time-varying navigation interaction graph, reflecting the basic risk level of the ship in the current interaction network. The risk of navigation resource occupation is quantified by judging whether the ship's current position occupies resources such as key sections of the waterway, anchorage, or berth front, and the degree of impact of the occupation behavior on the passage of other ships.

[0056] The risk vessel identification score is expressed by formula (3) as follows: ; In formula (3), For the first A ship in Risk assessment score for vessels at any given time; This represents the weighting coefficient for the risk status of a single ship. For the first A ship in The risk status of a single ship at any given moment; The weighting coefficient for the maximum ship-to-ship interaction risk; For the first A ship in Risks associated with ship-to-ship interactions at any given moment; The weighting coefficients for the weighted ship pair interaction risk; For the first A ship in The weight of the interaction edge at any given moment; The weighting coefficient for the rate of change of risk; This is for simulating step size; Weighting coefficients for the risk of navigation resource occupation; For the first A ship in Risk of occupying air traffic resources at specific times; This is a weighting coefficient representing the degree of conflict of intent. For the first A ship in The degree of conflict with surrounding vessels at any given time. The weighting coefficients in formula (3) can be determined based on expert experience or by fitting and calibrating historical risk event data.

[0057] The degree of conflict between the intentions of surrounding vessels is determined by comparing the differences between the optimal maneuvering intentions of the vessel and those of surrounding vessels, and then weighting and aggregating the results based on the interaction edge weights. The risk vessel identification threshold is set according to the statistical distribution of vessel identification scores in historical risk events.

[0058] The affected vessel set is expressed by formula (4) as follows: ; In formula (4), For the first Risky vessels The affected ship group at any given time; for A collection of ship intelligent agents in real time; For the first Risky vessel and the first The affected vessels The weight of the interaction edge at any given moment; The preset impact threshold; For the first Risky vessel and the first The affected vessels The most recent meeting time; This is the time window for the impact of the risk; For the first Risky vessel and the first The affected vessels The closest moment will meet distance; This refers to the spatial impact distance threshold. The preset impact threshold is set based on the port area's traffic density and safety control requirements; the risk impact time window is set based on the vessel maneuvering response time and the risk warning advance time; and the spatial impact distance threshold is set based on the port area's channel width and vessel safety distance requirements.

[0059] This embodiment constructs a multi-dimensional risk vessel discrimination score calculation formula that integrates single-vessel risk status, ship-to-ship interaction risk, risk change rate, navigation resource occupation risk, and degree of intent conflict. It then uses a risk vessel discrimination threshold to filter out a set of risk vessels. Finally, it uses multiple conditions such as interaction edge weight, nearest encounter time, and nearest encounter distance to determine the set of affected vessels corresponding to each risk vessel. This allows for the accurate identification of key vessels with risk source characteristics and their potential impact range from a complex multi-vessel interaction network, providing clear risk sources and propagation targets for the subsequent directed dissemination of risk information.

[0060] In some embodiments, based on the risky vessel and its set of affected vessels, a directed risk message is generated from the risky vessel to the affected vessels via a message function. Based on the directed risk message, the risk status of the affected vessels is updated via a state update function, including: The system obtains the multi-ship interaction risk, optimal maneuvering intention, and environmental field vector in the unified state space for each risky ship in the risky ship set. Based on the multi-ship interaction risk, optimal maneuvering intention, and environmental field vector, and combined with the interaction edge weights between the risky ship and the affected ship, a directed risk message from the risky ship to the affected ship is generated through a message function. The directed risk message contains the interaction risk information between the two ships, the candidate maneuvering intention conflict information, and the environmental disturbance amplification coefficient. Obtain the set of directed risk messages corresponding to each affected vessel. Based on the set of directed risk messages, aggregate each directed risk message in the set of directed risk messages using a message aggregation function to generate an aggregated risk message. Obtain the individual risk status and aggregated risk messages of the affected vessels. Based on the individual risk status and aggregated risk messages, and combined with the main risk response coefficient and message coupling amplification coefficient, update the individual risk status of the affected vessels through the status update function. Based on the updated individual risk status of each affected vessel and the interaction intensity matrix in the time-varying navigation interaction diagram, the regional navigation risk is calculated using the regional risk aggregation function.

[0061] In this embodiment, the candidate maneuvering intent conflict information is obtained by comparing the difference between the optimal maneuvering intents of the at-risk vessel and the affected vessel, and then judging the degree of conflict based on the encounter situation of the two vessels. The environmental disturbance amplification factor is calculated based on the degree of disturbance to the relative motion trajectory of the two vessels by environmental conditions such as wind, waves, current, and visibility; the more severe the environmental conditions, the larger the amplification factor value.

[0062] Directed risk information is expressed by formula (5) as follows: ; In formula (5), for From the moment The risky vessel points to the first Targeted risk information for the affected vessels; For the first Risky vessels Risk assessment score for vessels at any given time; The weighting coefficients for conflicting intent information; For the first Risky vessel and the first The affected vessels Conflicting information regarding candidate manipulation intentions at any given moment; The weighting coefficient for the environmental disturbance amplification factor; for Environmental disturbances at time 1 affect the first Risky vessel and the first The amplification factor of the relative motion of the affected ships. The weighting coefficients in Formula (5) can be obtained by fitting and calibrating based on expert experience or historical risk event data.

[0063] The message aggregation function is implemented using weighted summation or product aggregation. It is used to merge each directed risk message in the set of directed risk messages corresponding to each affected vessel according to the interaction edge weight or the risk vessel discrimination score to generate an aggregated risk message.

[0064] The subject risk response coefficient is set based on the risk preference parameter in the subject profile vector, and the message coupling amplification coefficient is set based on the ship's sensitivity to multi-source risk messages.

[0065] The updated individual vessel risk status of the affected vessel is expressed by formula (6) as follows: ; In formula (6), For the first A ship in The constantly updated risk status of a single vessel; For the first A ship in The risk status of a single ship at any given moment; For the first The main risk response coefficient of a vessel; for Time for the first A group of nearby vessels that pose a risk to a single vessel; For the first The message coupling amplification factor for a single vessel; for Time of the first Aggregated risk information for a single vessel; This is a pruning function used to restrict the updated risk state to a range.

[0066] The regional risk aggregation function calculates regional navigation risk by weighting the updated individual risk status of each affected vessel according to the weights in the interaction intensity matrix and then correcting it by combining the vessel density within the region.

[0067] This embodiment generates directed risk messages by fusing interaction edge weights, risk vessel identification scores, candidate maneuvering intent conflict information, and environmental disturbance amplification coefficients through message functions. Multi-source risk messages are then fused into aggregated risk messages through message aggregation functions. The individual risk status of affected vessels is updated by combining the subject risk response coefficient and message coupling amplification coefficient through state update functions. Finally, regional navigation risks are calculated through regional risk aggregation functions. This constructs a complete risk propagation chain from risk source identification, risk message generation and transmission, risk status update to regional risk aggregation, realizing dynamic quantitative simulation of the diffusion process from local disturbances to regional areas.

[0068] In some embodiments, a closed-loop verification loss is constructed based on the updated risk status of the affected vessel, historical AIS tracks, VTS handling records, and expert review opinions, including: Obtain the predicted motion state of the ship and its historical AIS trajectory. Calculate the weighted state error norm based on the deviation between the predicted motion state of the ship and the actual ship state at the corresponding moment in the historical AIS trajectory. Obtain the updated risk status and VTS handling records of the affected vessels, and calculate the risk trend error based on the deviation between the updated risk status of the affected vessels and the observed risk at the corresponding time in the VTS handling records. Obtain the optimal manipulation intention and expert review opinions, and calculate the behavior label deviation loss based on the deviation between the optimal manipulation intention and the reasonable behavior labels recognized in the expert review opinions; Obtain time-varying navigation interaction maps and historical AIS tracks; calculate the relationship network deviation based on the deviation between the time-varying navigation interaction maps and the real multi-ship relationship map constructed based on historical AIS tracks. Based on the weighted state error norm, risk trend error, behavioral label bias loss, and relationship network bias, a closed-loop verification loss is constructed.

[0069] In this embodiment, the weight matrix in the weighted state error norm is set according to the importance of each dimension of the ship's state or the measurement confidence level. Different weight values ​​can be assigned to dimensions such as position, speed, and heading to adjust the contribution of each dimension deviation to the total state error.

[0070] The weighting coefficient of risk trend error is set according to the importance of the risk status in the overall verification loss, and the focus can be placed on the accuracy of risk prediction according to the port area safety management requirements.

[0071] The behavioral label bias loss function can take the form of cross-entropy loss or hinge loss, and is used to measure the degree of difference between the optimal manipulation intention and the reasonable behavioral label recognized in the expert review opinion.

[0072] The weighting coefficients for relational network bias are set according to the importance of the relational network structure in risk propagation analysis. In high-density traffic scenarios, higher weights can be assigned to strengthen the constraint on the accuracy of the relational structure.

[0073] The real multi-ship relationship graph is generated based on the encounter relationships, relative motion relationships, and traffic organization relationships between ships in historical AIS trajectories, using the same edge definition rules as those used to construct the time-varying navigation interaction graph.

[0074] set up The first in the historical AIS trajectory A ship in The actual ship status at any given moment To map the simulated state to the observation matrix of observables, The first one is obtained by back-calculation from the VTS processing record. A ship in The risk of constant observation The first in the expert review opinion A ship in Given the appropriate behavior label at each moment, the errors are expressed by formula (7) as follows: ; In formula (7), For the first A ship in The state error at any given moment; For the first A ship in Predicted motion state of the ship at any given moment; For the first A ship in Risk trend error at any given moment; For the first A ship in The risk status is updated in real time; For the first A ship in Moment-time behavior label bias loss; For the first A ship in The optimal manipulation intention at any given moment.

[0075] Each error term can be normalized before being substituted into the closed-loop verification loss to eliminate the influence of different dimensions on the loss value.

[0076] Based on the above error term, the closed-loop verification loss is expressed by formula (8) as follows: ; In formula (8), for Closed-loop verification loss at any given moment; for A fleet of ships at risk at any given moment; Let the weighted state error norm be . This is the weight matrix; This represents the weighting coefficient for risk trend error; The weighting coefficients for behavioral label bias loss; The behavior label bias loss function; The weighting coefficients for the bias in the relational network; A real multi-ship relationship diagram constructed based on historical AIS trajectories; for A time-varying navigation interaction diagram; It is the square of the Frobenius norm.

[0077] This embodiment calculates the deviations between the simulated output and the actual observations in four dimensions: ship motion state, risk trend, maneuvering behavior, and relationship network. It obtains the weighted state error norm, risk trend error, behavior label deviation loss, and relationship network deviation, respectively. The four errors are then weighted and fused into a closed-loop verification loss, constructing a comprehensive verification index covering state, risk, behavior, and relationship structure. This provides a quantitative error feedback basis for simulator calibration based on this loss in subsequent steps.

[0078] In some embodiments, the predicted motion state, motion model parameters, and main body profile of the ship are iteratively corrected based on the closed-loop verification loss, and the next round of simulation is updated to output the risk propagation chain and regional risk hotspots, including: Obtain the closed-loop verification loss and the predicted motion state of the ship. Based on the closed-loop verification loss, perform state-level correction on the predicted motion state of the ship and generate the state-corrected predicted motion state of the ship. The ship's predicted motion state and closed-loop verification loss after state correction are obtained. Based on the ship's predicted motion state and closed-loop verification loss after state correction, the parameters of the physical motion response model and the parameters of the historical trajectory residual correction term are corrected at the parameter level to generate updated motion model parameters. The updated motion model parameters and closed-loop verification loss are obtained. Based on the updated motion model parameters and closed-loop verification loss, the empirical pattern features, rule understanding ability features, risk preference parameters and manipulation habit features in the subject profile vector are corrected at the parameter level to generate the updated subject profile. Obtain the state-corrected predicted motion state of the ship, the updated motion model parameters, and the updated subject profile, and write these data back into the historical behavior memory bank for the next round of simulation. Based on the updated risk status of affected vessels and the time-varying navigation interaction diagram in the next round of simulation, the directed risk message transmission path between each risk vessel in the risk vessel set and the affected vessel set is extracted to generate a risk propagation chain. Based on the regional navigation risks in the next round of simulations, regional risk hotspots are generated and output.

[0079] In this embodiment, state-level correction is achieved by multiplying the state error portion of the closed-loop verification loss by the state correction gain and then incrementally adding it to the predicted motion state of the ship.

[0080] Parameter-level correction generates updated motion model parameters by calculating the gradient of the closed-loop verification loss on the parameters of the physical motion response model and the parameters of the historical trajectory residual correction term, and then fine-tuning and updating the parameters along the gradient descent direction.

[0081] The correction of each feature in the subject profile vector is achieved by calculating the gradient of the closed-loop verification loss on the empirical pattern feature, rule understanding ability feature, risk preference parameter and manipulation habit feature. The feature parameters are then fine-tuned and updated along the gradient descent direction, while the updated parameters are restricted to a preset reasonable range of values, thus generating an updated subject profile.

[0082] After indexing the corrected ship motion state, updated motion model parameters, and updated subject profile according to timestamps and ship identifiers, the data is stored in the corresponding time point in the historical behavior memory bank for use in the next round of simulation.

[0083] The risk propagation chain is generated by connecting the risk source vessels and each affected vessel in chronological order according to the direction of directed risk message transmission between each risk vessel in the risk vessel set and the affected vessel set in the next round of simulation, forming a propagation path sequence from the risk source to the affected endpoint.

[0084] The regional risk hotspots are generated by using spatial interpolation or contour plotting to create a visual hotspot map that reflects the distribution of risk levels on the port area spatial grid in the next round of simulation.

[0085] This embodiment corrects the instantaneous deviation of the ship's predicted motion state through state-level correction, improves model accuracy by adjusting the parameters of the physical motion response model and the parameters of the historical trajectory residual correction term through parameter-level correction, optimizes the cognitive feature parameters of the agent through subject profile correction, and writes the corrected state, model parameters, and subject profile back to the historical behavior memory bank to form a closed loop. Thus, in the next round of simulation, the risk propagation chain and regional risk hot zone are output based on the corrected simulator, realizing the continuous suppression of errors and the visualization of output results by the simulator in continuous simulation.

[0086] By adopting the above technical solutions, this invention differs from existing technologies and has the following beneficial effects: It constructs a unified state space by acquiring multi-source data, generating navigation subject cognitive intelligent agents for captains, pilots, and VTS supervisors, enabling simulated objects to possess human-like situational awareness and decision-making capabilities; it connects the intelligent agents to a large language model and generates optimal maneuvering intentions under structured prompt constraints, converting these intentions into continuous control quantities via a mapping function, which are then combined with the physical motion response model and historical trajectory residual correction terms and projected onto the safe and feasible domain, generating a predicted ship motion state that conforms to physical laws and safety constraints; it constructs a time-varying navigation interaction graph with ship intelligent agents as nodes and various relationships as edges, calculating and integrating multi-ship interaction risks based on encounter types, environmental disturbances, and differences in behavioral intentions, and dynamically transmitting risks from at-risk vessels to affected vessels through selection functions, message functions, and state update functions; it constructs a closed-loop verification loss based on the updated risk state, historical AIS trajectory, VTS handling records, and expert review opinions, iteratively correcting the predicted ship motion state, motion model parameters, and subject profile, and outputting the risk propagation chain and regional risk hotspots. The above technical solution combines cognitive modeling, graph network propagation, and closed-loop correction, enabling navigation risk evolution simulation to have anthropomorphic decision-making capabilities, multi-ship coupled risk transmission and deduction capabilities, and self-correction capabilities in long-term operation, effectively improving the authenticity and credibility of the simulation results.

Claims

1. An evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation, characterized in that, include: Acquire basic port area data, vessel navigation status data, navigation rules data, environmental data, and historical trajectory data to construct a complex port area navigation scenario with a unified state space representation; Based on the unified state space, a main profile vector is constructed for the captain, pilot, and VTS supervisor to generate a navigation subject cognitive intelligent agent. The general aviation subject cognitive agent is connected to a large language model, and a set of candidate manipulation intentions is generated under structured prompt constraints to obtain the optimal manipulation intention; The optimal maneuvering intention is converted into a continuous control quantity through a mapping function from intention to control quantity. This control quantity is then processed by the physical motion response model and the historical trajectory residual correction term, and projected onto the safe and feasible region consisting of the channel boundary and risk threshold to generate the predicted motion state of the ship at the next moment. Using ship intelligent agents as nodes and encounter relationships, relative motion relationships, traffic organization relationships, and risk impact relationships as edges, a time-varying navigation interaction graph is constructed, and multi-ship interaction risks that integrate encounter types, environmental disturbances, and differences in behavioral intentions are calculated based on the time-varying navigation interaction graph. Based on the time-varying navigation interaction diagram and the multi-ship interaction risk, a risky vessel and its set of affected vessels are determined by a selection function. Based on the risky vessel and its set of affected vessels, a directed risk message from the risky vessel to the affected vessel is generated by a message function. Based on the directed risk message, the risk status of the affected vessel is updated by a state update function. Based on the updated risk status of the affected vessels, historical AIS trajectories, VTS handling records, and expert review opinions, a closed-loop verification loss is constructed. The predicted motion status, motion model parameters, and main body profile of the vessels are iteratively corrected based on the closed-loop verification loss, and the next round of simulation is updated to output the risk propagation chain and regional risk hotspots.

2. The evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation according to claim 1, characterized in that, Acquire basic port area data, vessel navigation status data, navigation rules data, environmental data, and historical trajectory data to construct a complex port area navigation scenario with a unified state space representation, including: Acquire basic port area data, which includes channel boundary data, restricted navigation area data, anchorage data, berth data, and construction area data; generate a port area spatial constraint field based on the basic port area data. The ship navigation status data is acquired, including the ship's planar position, speed, heading, draft, length, beam, ship type, maneuverability, and hazardous materials attributes. A ship motion and attribute state vector is generated based on the ship navigation status data. Acquire navigation rule data, which includes traffic organization status, navigation control rules, pilotage plans and tugboat resource information, and generate an operational constraint vector based on the navigation rule data; Acquire environmental data, including wind field data, wave field data, flow velocity and direction field data, and visibility data, and generate an environmental field vector based on the environmental data; Acquire historical trajectory data, including AIS historical trajectory records and VTS handling records, and generate a historical behavior memory bank based on the historical trajectory data; The complex port navigation scenario with a unified state space expression is constructed by combining the ship motion and attribute state vector, the environmental field vector, the port area spatial constraint field, the operation constraint vector, the historical behavior memory, and the observation input vector consisting of the relative positional relationship, encounter relationship, and rule constraint relationship between each ship agent and the surrounding ship agents.

3. The evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation according to claim 2, characterized in that, Based on the unified state space, a main profile vector is constructed for the captain, pilot, and VTS supervisor, generating a navigation subject cognitive intelligent agent, including: Obtain the historical behavior memory bank in the unified state space, extract the historical navigation experience data and typical encounter handling pattern data of each navigation entity based on the historical behavior memory bank, and generate experience pattern features; Obtain the operational constraint vector in the unified state space, and extract the understanding level data of each navigation entity on navigation rules and port area organization constraints based on the operational constraint vector to generate rule understanding ability features; Obtain the historical behavior memory bank and ship motion and attribute state vector in the unified state space, extract the historical maneuvering habit data and response time delay feature data of each navigation subject based on the historical behavior memory bank and ship motion and attribute state vector, and generate maneuvering habit features; Based on the aforementioned experience pattern characteristics, rule understanding ability characteristics, and operational habit characteristics, and combined with the risk preference parameters and organizational role parameters of each general aviation entity, a subject profile vector containing experience patterns, rule understanding ability, risk preference, organizational role, and operational habits is constructed. Based on the subject profile vector and the observation input vector, a general aviation subject cognitive agent is generated. The general aviation subject cognitive agent includes a cognitive decision function, which is used to output a set of candidate manipulation intentions based on the subject profile vector and the observation input data.

4. The evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation according to claim 3, characterized in that, The general aviation cognitive agent is integrated into a large language model to generate a set of candidate manipulation intentions under structured prompt constraints, thereby obtaining the optimal manipulation intention, including: Obtain the subject profile vector and observation input data of the general aviation subject cognitive intelligent agent, and construct a structured prompt template based on the subject profile vector and the observation input data. The structured prompt template includes general aviation subject role information, current situation description information, general aviation rule constraint information, and historical behavior memory data in the historical behavior memory bank. The structured prompt template is input into the large language model. Based on the large language model, a set of candidate maneuvering intentions is generated under the constraints of the structured prompt template. The set of candidate maneuvering intentions includes intentions to maintain course, intentions to fine-tune course, intentions to avoid obstacles in advance, intentions to slow down and wait, intentions to keep right, intentions to request VTS coordination, intentions to postpone arrival, and intentions to adjust berthing and departure order. The rule compliance score, maneuverability score, and conflict suppression score of each candidate maneuvering intention in the candidate maneuvering intention set are obtained. The rule compliance score is determined based on the degree of compliance between the candidate maneuvering intention and the navigation rules and organizational instructions. The maneuverability score is determined based on the degree of satisfaction of the candidate maneuvering intention with the ship's maneuverability and safety boundaries. The conflict suppression score is determined based on the intensity of multi-ship conflicts that the candidate maneuvering intention may cause. Based on the rule compliance score, the manipulation feasibility score, and the conflict suppression score, and combined with the confidence probability of each candidate manipulation intention by the large language model, a weighted comprehensive score is given for each candidate manipulation intention. Based on the weighted comprehensive score, the candidate manipulation intent with the highest score is selected from the set of candidate manipulation intents as the optimal manipulation intent.

5. The evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation according to claim 4, characterized in that, The optimal maneuvering intention is converted into a continuous control quantity through an intention-to-control quantity mapping function. This control quantity is then processed by the physical motion response model and the historical trajectory residual correction term, and projected onto the safe and feasible region comprised of the channel boundary and risk threshold to generate the predicted ship motion state for the next moment, including: Obtain the optimal maneuvering intention and the subject profile vector of the general aviation subject cognitive intelligent agent. Based on the optimal maneuvering intention and the subject profile vector, convert the optimal maneuvering intention into a continuous control quantity containing longitudinal acceleration and steering angular velocity through a mapping function from intention to control quantity. The continuous control quantity and the environmental field vector in the unified state space are obtained. Based on the continuous control quantity and the environmental field vector, a physical motion prediction state that satisfies the ship size, inertia, rudder effect and acceleration / deceleration boundaries is generated through the physical motion response model. The physical motion prediction state and the historical behavior memory bank in the unified state space are obtained. Based on the physical motion prediction state and the historical behavior memory bank, a residual correction prediction state containing port channel geometry and local environmental disturbance correction is generated through the historical trajectory residual correction term. The residual correction prediction state and the port area spatial constraint field and operation constraint vector in the unified state space are obtained. Based on the residual correction prediction state, the port area spatial constraint field and the operation constraint vector, the residual correction prediction state is projected onto the safe and feasible domain composed of the channel boundary, the restricted navigation area, the berth front edge and the risk threshold through the risk constraint projection function to generate the ship's predicted motion state at the next moment.

6. The evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation according to claim 5, characterized in that, Using ship-to-ship intelligent agents as nodes and encounter relationships, relative motion relationships, traffic organization relationships, and risk impact relationships as edges, a time-varying navigation interaction graph is constructed. Based on this time-varying navigation interaction graph, multi-ship interaction risks that integrate encounter types, environmental disturbances, and differences in behavioral intentions are calculated, including: The predicted motion state of the ship is obtained, and the relative position and relative speed between any two ship agents are calculated based on the predicted motion state of the ship and the motion and attribute state vector of the ship. Calculate the nearest meeting time and the nearest meeting distance based on the relative position and the relative speed; Obtain the optimal maneuvering intent and the environmental field vector in the unified state space. Based on the nearest encounter time, the nearest encounter distance, the optimal maneuvering intent, and the environmental field vector, and combined with the encounter type correction coefficient, calculate the multi-ship interaction risk of the fusion distance risk item, the time urgency risk item, the encounter type risk item, the environmental risk factor, and the degree of conflict of behavioral intent. Based on the aforementioned multi-ship interaction risk, and combining spatial and temporal scale parameters, the interaction edge weight between any two ship agents is calculated. Using ship intelligent agents as nodes and the interaction edge weights as edges, a time-varying navigation interaction graph containing an adjacency matrix and an interaction strength matrix is ​​constructed.

7. The evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation according to claim 1, characterized in that, Based on the time-varying navigation interaction diagram and the multi-ship interaction risk, a selection function is used to determine the at-risk vessels and their affected vessel sets, including: Obtain the interaction edge weights and multi-ship interaction risks in the time-varying navigation interaction graph. Based on the interaction edge weights and multi-ship interaction risks, and combined with the initial risk state and navigation resource occupation risk of each ship agent based on the current interaction relationship, calculate the risk ship discrimination score of each ship agent. Obtain the optimal maneuvering intention and the operational constraint vector in the unified state space. Based on the optimal maneuvering intention and the operational constraint vector, and combined with the degree of conflict between the intentions of surrounding vessels, update the risk vessel identification score. Based on the risk vessel identification score and the preset risk vessel identification threshold, ship agents with scores exceeding the threshold are selected as a risk vessel set; Based on the set of risky vessels and the interaction edge weights in the time-varying navigation interaction graph, ship agents that have interaction edge weights with each risky vessel exceeding a preset influence threshold are selected as the set of affected vessels corresponding to each risky vessel.

8. The evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation according to claim 7, characterized in that, Based on the set of at-risk vessels and their affected vessels, a directed risk message is generated from the at-risk vessel to the affected vessel using a message function. Based on the directed risk message, the risk status of the affected vessels is updated using a state update function, including: Obtain the multi-ship interaction risk, optimal maneuvering intention, and environmental field vector in the unified state space for each risky vessel in the risky vessel set. Based on the multi-ship interaction risk, the optimal maneuvering intention, and the environmental field vector, and combined with the interaction edge weights between the risky vessel and the affected vessel, generate a directed risk message from the risky vessel to the affected vessel through a message function. The directed risk message includes two-ship interaction risk information, candidate maneuvering intention conflict information, and environmental disturbance amplification coefficient. Obtain the set of directed risk messages corresponding to each affected vessel. Based on the set of directed risk messages, aggregate each directed risk message in the set of directed risk messages using a message aggregation function to generate an aggregated risk message. Obtain the individual risk status of the affected vessel and the aggregated risk message. Based on the individual risk status and the aggregated risk message, and combined with the subject risk response coefficient and the message coupling amplification coefficient, update the individual risk status of the affected vessel through a status update function. Based on the updated individual risk status of each affected vessel and the interaction intensity matrix in the time-varying navigation interaction diagram, the regional navigation risk is calculated using a regional risk aggregation function.

9. The evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation according to claim 8, characterized in that, Based on the updated risk status of the affected vessels, historical AIS tracks, VTS handling records, and expert review opinions, a closed-loop verification of the loss is constructed, including: The ship's predicted motion state and historical AIS trajectory are obtained, and the weighted state error norm is calculated based on the deviation between the ship's predicted motion state and the actual ship state at the corresponding time in the historical AIS trajectory. Obtain the updated risk status and VTS handling record of the affected vessel, and calculate the risk trend error based on the deviation between the updated risk status of the affected vessel and the observed risk at the corresponding time in the VTS handling record; Obtain the optimal manipulation intention and expert review opinions, and calculate the behavior label deviation loss based on the deviation between the optimal manipulation intention and the reasonable behavior labels recognized in the expert review opinions; Obtain the time-varying navigation interaction map and the historical AIS trajectory, and calculate the relationship network deviation based on the deviation between the time-varying navigation interaction map and the real multi-ship relationship map constructed based on the historical AIS trajectory. Based on the weighted state error norm, the risk trend error, the behavior label bias loss, and the relationship network bias, a closed-loop verification loss is constructed.

10. The evolutionary simulation method based on multi-agent cognitive modeling and graph network propagation according to claim 9, characterized in that, The predicted motion state, motion model parameters, and main body profile of the ship are iteratively corrected based on the closed-loop verification loss, and the next round of simulation is updated. The risk propagation chain and regional risk hotspots are output, including: The closed-loop verification loss and the ship's predicted motion state are obtained. The ship's predicted motion state is then corrected at the state level based on the closed-loop verification loss to generate the state-corrected ship's predicted motion state. The ship's predicted motion state after state correction and the closed-loop verification loss are obtained. Based on the ship's predicted motion state after state correction and the closed-loop verification loss, the parameters of the physical motion response model and the parameters of the historical trajectory residual correction term are corrected at the parameter level to generate updated motion model parameters. Obtain the updated motion model parameters and the closed-loop verification loss, and perform parameter-level correction on the experience pattern features, rule understanding ability features, risk preference parameters and manipulation habit features in the subject profile vector based on the updated motion model parameters and the closed-loop verification loss to generate an updated subject profile; Obtain the corrected ship motion state, the updated motion model parameters, and the updated subject profile, and write the corrected ship motion state, the updated motion model parameters, and the updated subject profile back to the historical behavior memory bank for the next round of simulation. Based on the updated risk status of the affected vessels in the next round of simulation and the time-varying navigation interaction diagram, the directed risk message transmission path between each risk vessel in the risk vessel set and the affected vessel set is extracted to generate a risk propagation chain. Based on the regional navigation risks in the next round of simulation, regional risk hotspots are generated and output.

Images