Multi-ship dynamic joint planning method and system based on trajectory

Abstract

The invention relates to a trajectory-based multi-ship dynamic joint planning method and system, belongs to the technical field of ship route planning, and solves the problem that a model cannot effectively predict multi-ship dynamic disordered behaviors in the prior art. The method comprises the following steps: predicting a future motion track of a first ship according to historical data of the first ship by using an established ship track prediction model; determining a preset action and an initial condition of a second ship according to a future motion track by utilizing an established ship path planning model, and simulating and executing the preset action to plan a motion path of the second ship; and obtaining a dynamic interaction behavior between the two ships according to the future movement track of the first ship and the movement path of the second ship by using an established ship dynamic interaction model, and adjusting the planning of the movement path of the second ship based on the interaction behavior. According to the technical scheme, simulation and prediction of the dynamic interaction behaviors of the multiple ships are effectively achieved.

Description

Technical Field

[0001] This invention relates to the field of ship route planning technology, and in particular to a planning method and system for dynamic joint planning of multiple ships based on trajectory. Background Technology

[0002] Ship trajectory prediction, which predicts a ship's future course based on its current trajectory, has a wide range of applications. Conventional ship trajectory prediction is generally used for itinerary planning and route dynamic management of ocean-going vessels, typically providing a simple long-term prediction based on common waterways that does not require frequent adjustments.

[0003] However, long-term planning is only suitable for single-vehicle management and safe waters. In complex sea conditions, congested waters, or emergency situations, it is difficult to achieve high-precision and dynamic prediction, planning, and adjustment. Currently, there is a considerable amount of new research on ship trajectory prediction. With the assistance of GPS (Global Positioning System) and AIS (Automatic Identification System) information, existing prediction models can make relatively accurate and timely predictions of single-vehicle trajectories. Single-vehicle trajectory prediction information (which can predict the trajectories of multiple vessels separately, but each prediction only considers the data of a single vessel) can be used for ship scheduling, effectively avoiding collisions with static or low-speed obstacles (such as other vessels about to enter or leave port), ensuring navigational safety, and improving maritime navigation efficiency. It can also be used to identify abnormal or dangerous states of vessels, monitor abnormal vessels, and provide timely warnings or rescue.

[0004] However, existing single-ship trajectory prediction models are still only idealized predictions based on safe environments and reasonable behavior. With the increase in maritime traffic, the complexity of ship behavior also increases. Existing prediction models can only perform orderly planning under the condition that all ships comply with rules and regulations. In the event of emergencies, such as loss of ship control, crew violations, or other urgent situations requiring intervention, it is necessary to predict the joint behavior of multiple ships, and to anticipate and adjust the behavior of other ships in real time based on the target ship's ever-changing state. Existing prediction models cannot support such dynamic, multi-ship interactive joint prediction and planning (where at least one ship may exhibit disorderly behavior). Summary of the Invention

[0005] Based on the above analysis, the embodiments of the present invention aim to provide a planning method and system for dynamic joint operation of multiple ships based on trajectories, in order to solve the problem that ship trajectory prediction models in the prior art cannot effectively predict the dynamic disordered behavior of multiple ships.

[0006] On one hand, embodiments of the present invention provide a trajectory-based multi-ship dynamic joint planning method, the method comprising the following steps:

[0007] Using an established ship trajectory prediction model, predict the future trajectory of at least one first ship based on historical data of at least one first ship;

[0008] Using an established ship path planning model, the preset actions and initial conditions of at least one second ship are determined based on the future motion trajectory of the first ship, and the preset actions are simulated to plan the motion path of the at least one second ship.

[0009] Using the established ship dynamic interaction model, the dynamic interaction behavior between the first ship and the second ship is obtained based on the future motion trajectory of the first ship and the motion path of the second ship, and the planning of the motion path of the second ship is adjusted based on the dynamic interaction behavior.

[0010] Based on a further improvement of the above method, the ship dynamic interaction model uses a dual-channel sequence network to obtain the dynamic interaction behavior between the first ship and the second ship.

[0011] Based on a further improvement of the above method, the dual-channel sequence network is a dual LSTM network, wherein the first LSTM network model in the dual LSTM network processes the data of the first ship; and the second LSTM network model processes the data of the second ship.

[0012] Based on further improvements to the above method, the ship dynamic interaction model includes an input layer, an encoder, an interaction layer, a decoder, and an output layer, wherein...

[0013] The input layer processes the data from the first ship and the second ship to generate time-series data;

[0014] The encoder extracts key features from the time-series data of the first and second ships through the encoding module of the LSTM network;

[0015] The interaction layer captures the dynamic interaction relationship between the first vessel and the second vessel, and obtains the interaction characteristics of the first vessel and the second vessel.

[0016] The decoder predicts the future trajectories of the first and second vessels based on the key features and the interaction features.

[0017] The output layer outputs at least one prediction result for the first and second ships after the current moment, item by item, according to the trajectory point sequence.

[0018] Based on a further improvement to the above method, at least one model in the method uses a dynamically adjusted neighborhood orientation grid for prediction, the dynamic radius of which is:

[0019] R = R0 + k1·v - k2·p - k3·k + E;

[0020] Where R0 is the initial grid radius, v is the current velocity, k1 is the velocity adjustment coefficient; p is the trajectory density, k2 is the trajectory density adjustment coefficient; k is the average curvature value, k3 is the curvature adjustment coefficient; and E is the environmental attribute correction value.

[0021] On the other hand, embodiments of the present invention provide a trajectory-based multi-ship dynamic joint planning system, the system comprising:

[0022] A ship trajectory prediction model is used to predict the future trajectory of at least one first ship based on historical data of at least one first ship;

[0023] A ship path planning model is used to determine the preset actions and initial conditions of at least one second ship based on the future motion trajectory of the first ship, and to simulate the execution of the preset actions to plan the motion path of the at least one second ship.

[0024] The ship dynamic interaction model is used to obtain the dynamic interaction behavior between the first ship and the second ship based on the future motion trajectory of the first ship and the motion path of the second ship, and to adjust the planning of the motion path of the second ship based on the dynamic interaction behavior.

[0025] Based on further improvements to the above system, the ship dynamic interaction model uses a dual-channel sequence network to capture the dynamic interaction behavior between the first ship and the second ship.

[0026] Based on further improvements to the above system, the dual-channel sequence network is a dual LSTM network, wherein the first LSTM network model in the dual LSTM network processes the data of the first ship; and the second LSTM network model processes the data of the second ship.

[0027] Based on further improvements to the above system, the ship dynamic interaction model includes an input layer, an encoder, an interaction layer, a decoder, and an output layer, wherein...

[0028] The input layer processes the data from the first ship and the second ship to generate time-series data;

[0029] The encoder extracts key features from the time-series data of the first and second ships through the encoding module of the LSTM network;

[0030] The interaction layer captures the dynamic interaction relationship between the first vessel and the second vessel, and obtains the interaction characteristics of the first vessel and the second vessel.

[0031] The decoder predicts the future trajectories of the first and second vessels based on the key features and the interaction features.

[0032] The output layer outputs at least one prediction result for the first and second ships after the current moment, item by item, according to the trajectory point sequence.

[0033] Based on further improvements to the above system, the system also includes:

[0034] A dynamic mesh module is used for prediction using a dynamically adjusted neighborhood orientation mesh, the dynamic radius of which is:

[0035] R = R0 + k1·v - k2·p - k3·k + E;

[0036] Where R0 is the initial grid radius, v is the current velocity, k1 is the velocity adjustment coefficient; p is the trajectory density, k2 is the trajectory density adjustment coefficient; k is the average curvature value, k3 is the curvature adjustment coefficient; and E is the environmental attribute correction value.

[0037] This invention, through simulation and prediction of the dynamic interactive behavior of multiple vessels, can effectively identify, prevent, and intervene in emergencies in complex environments. It achieves real-time monitoring and assessment of vessel navigation status, enhances the adaptability and effectiveness of emergency plans, reduces the risk of maritime accidents, and provides scientific and technical support for multi-vessel management, especially joint management of multiple vessels in complex near-shore environments. At the same time, it strengthens the emergency response capability to emergencies and comprehensively improves maritime traffic safety and control efficiency.

[0038] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0039] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0040] Figure 1 This is a flowchart of a trajectory-based multi-ship dynamic joint planning method in one embodiment of the present invention;

[0041] Figure 2This is a schematic diagram of the ship dynamic interaction model architecture in a preferred embodiment of the present invention. Detailed Implementation

[0042] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0043] A specific embodiment of the present invention discloses a trajectory-based multi-ship dynamic joint planning method, such as... Figure 1 As shown, the method includes the following steps:

[0044] S11, using the established ship trajectory prediction model, predict the future motion trajectory of the at least one first ship based on the historical data of the at least one first ship;

[0045] S12, using the established ship path planning model, determine the preset actions and initial conditions of at least one second ship based on the future motion trajectory of the first ship, and simulate the execution of the preset actions to plan the motion path of the at least one second ship;

[0046] S13, using the established ship dynamic interaction model, obtain the dynamic interaction behavior between the first ship and the second ship based on the future motion trajectory of the first ship and the motion path of the second ship, and adjust the planning of the motion path of the second ship based on the dynamic interaction behavior.

[0047] This invention, through simulation and prediction of the dynamic interactive behavior of multiple vessels, can effectively identify, prevent, and intervene in emergencies in complex environments. It achieves real-time monitoring and assessment of vessel navigation status, enhances the adaptability and effectiveness of emergency plans, reduces the risk of maritime accidents, and provides scientific and technical support for multi-vessel management, especially joint management of multiple vessels in complex near-shore environments. At the same time, it strengthens the emergency response capability to emergencies and comprehensively improves maritime traffic safety and control efficiency.

[0048] In this embodiment of the invention, existing and relatively mature ship trajectory prediction models are first used to predict the motion trajectories of individual ships in order to identify abnormal or potentially dangerous target ships. The ship trajectory prediction model mainly collects and analyzes historical and real-time trajectory data of ships, and uses machine learning algorithms to establish a ship behavior prediction model to predict the future navigation trajectory of ships.

[0049] Specifically, the process of establishing the ship trajectory prediction model includes the following steps:

[0050] Collect historical (including real-time data collected at the current moment) trajectory data of all vessels in the current sea area (hereinafter referred to as the first vessel), including but not limited to vessel type, size, speed, course, position (latitude and longitude), timestamp, etc. Data sources may include GPS, AIS data, meteorological data, etc.

[0051] The trajectory data is preprocessed, including removing obvious erroneous data and stationary ship data, filling in missing data, and denoising the trajectory points. Preferably, the preprocessing also includes normalizing the numerical features in the trajectory data so that each feature is within the same scale range, avoiding imbalance in model training caused by data of different scales.

[0052] Key features are extracted from historical trajectory data, and each historical trajectory point is labeled with a trajectory point direction classification tag. These key features include spatial, temporal, and environmental features. Spatial features include, but are not limited to, the ship's current position, speed, heading, acceleration, acceleration direction, and information about neighboring points in the grid. Temporal features include, but are not limited to, the current time, time interval, and seasonal features (such as season and time period). Environmental features include, but are not limited to, wind speed, ocean current direction and intensity, and weather conditions. The trajectory point direction classification tag divides the possible target directions of the current trajectory point into a grid of multiple neighboring directions (generally set as a 3x3 grid with eight neighboring direction grids: North, Northeast, East, Southeast, South, Southwest, West, and Northwest; alternatively, directions and grids can be set according to hexagonal or other shapes). Based on the historical trajectory data, the grid where each trajectory point is located at the next moment is used as the label for the target direction at the current moment. The trajectory point orientation classification label is generally obtained directly from historical trajectory data. This directly obtained label is the key information of the sample data and is mainly used for model training and implementation. However, it is obviously impossible to directly obtain the target orientation at future moments from historical trajectory data. Therefore, the trajectory point orientation classification label at future moments is marked by the model after processing. This is the main prediction output of the model.

[0053] Based on the trajectory prediction requirements, at least one machine learning model from Random Forest, Support Vector Machine, and KNN is selected. Processed historical trajectory data (including but not limited to preprocessing, feature extraction, and labeling) is used as the training set to train the model to predict the target grid direction for each trajectory point. That is, the model input features are the key features of the current moment (i.e., the spatial, temporal, and environmental features of the ship mentioned earlier); the output is the ship's direction of motion at the next moment, represented by the trajectory point's direction classification label (the probability that the ship will enter a certain neighborhood grid at the next moment, which actually includes both direction and position prediction). The model can also continuously output prediction results for multiple subsequent moments through iterative prediction. In this way, the model processes and outputs multiple features such as position, velocity, direction, and time, which are difficult to express simply, using a single grid classification task, effectively reducing model complexity, improving prediction efficiency, and enhancing prediction accuracy.

[0054] However, existing technologies typically use fixed-size grids for neighborhood meshing and prediction. While this approach can handle most low-speed, safe situations, it is clearly unsuitable for the complex, highly uncertain, and dynamically interactive scenarios addressed in this invention. Therefore, in a preferred embodiment of this invention, a dynamic adjustment scheme is provided for the grid size used in the model. The goal of the model's grid selection is to provide a concise, effective, and accurate representation of the ship's motion direction at the next moment. A fixed-size grid covers a limited area, while the actual ship speed directly affects its future motion range; higher speeds result in a wider motion range, potentially exceeding the grid's coverage. Conversely, directly using a large-size grid would be insufficient to meet the high accuracy requirements of the task. This invention preferably dynamically adjusts the neighborhood grid size (which can be the radius of the grid's inscribed circle or the side length of the grid polygon) based on historical data (speed, trajectory density, curvature, etc.) of multiple (or all) ships in the current sea area and environmental factors to adapt to the actual conditions of the current sea area and / or a specific target ship. Specifically, the grid radius can be dynamically adjusted using the following formula:

[0055] R = R0 + k1·v - k2·p - k3·k + E;

[0056] Wherein, R0 is the initial grid radius, which is generally the default value under traditional low-speed or no-movement conditions;

[0057] v represents the current speed. In this invention, a larger grid size is used for scenarios with higher speeds to effectively capture the potential target direction of high-speed ships, while the grid size can be appropriately reduced for lower speeds to achieve higher accuracy and prediction of more targets together; k1 is the speed adjustment coefficient, which determines the sensitivity of the grid radius to speed changes.

[0058] p represents the trajectory density, which is obtained by calculating the spatial distribution density of historical trajectory points of multiple ships (or all ships) in the current sea area (the number of trajectory points per unit area, i.e., number of points / unit area). The higher the trajectory point density, the more complex and diverse the trajectories in the area (such as near a port), requiring a smaller grid radius for finer prediction. Conversely, a larger grid can be used for simpler trajectories (such as in the main channel area). k2 is the trajectory density adjustment coefficient.

[0059] k is the average curvature value, representing the degree of turning. High curvature indicates a more varied route and requires a finer grid, while low curvature indicates a straight route and can use a larger grid. k3 is the curvature adjustment coefficient. The average curvature value is usually calculated based on the recommended route for the current sea area. In sea areas without a recommended route, the average value can be calculated based on the historical trajectories of multiple (or all) ships that have sailed in the current sea area.

[0060] E is an environmental attribute correction value used to impose regional constraints on the grid radius. Depending on different marine environmental conditions, the vessel's mobile range is subject to certain limiting factors. Smaller grid radii can be used in narrow waters, while larger grid radii can be used in wide sea areas with a greater operational range. This correction value is closely related to the current marine environment, but for a specific sea area, the environmental conditions are relatively stable, and the correction value is usually a fixed value. Therefore, the sea area where the vessel is located can be determined based on its current position, and the E value corresponding to the current sea area can be directly selected from the pre-stored values ​​in the system.

[0061] Furthermore, during model training, the model can be validated and tested as needed, and optimized as appropriate. Specifically, the trained model can be validated using a test set, and its performance can be evaluated using cross-validation and loss functions (such as cross-entropy). For example, the error between the predicted result and the actual trajectory can be calculated, and metrics such as accuracy and recall can be used for evaluation. Based on the validation results, the model's hyperparameters can be adjusted according to the model performance objectives, and the model structure can be optimized. Given that there are already many ship trajectory prediction models in the existing technology, the specific process of model selection, training, validation, and optimization will not be elaborated here.

[0062] After completing a series of operations such as model training, verification, and optimization, an established ship trajectory prediction model can be obtained. This established ship trajectory prediction model acquires the historical data of the first ship in real time (including but not limited to historical trajectory and current position, speed, heading, etc.) and predicts the future motion trajectory of the ship based on the trained model; the prediction result can be the direction and position of motion in the future.

[0063] The ultimate goal of this invention is to support dynamic joint situational awareness in multi-ship interactions. After obtaining a prediction of the future trajectory of the first ship, it can determine whether there is any abnormal behavior based on historical trajectories, predicted trajectories, and actual movement (e.g., judging collision risk based on the predicted trajectory, judging whether it is an illegal vessel based on the path from the departure point to the current position, judging whether an accident or emergency has occurred based on abnormal changes in speed and heading in the historical trajectory, judging whether there is any suspicious behavior based on deviations from the predicted trajectory). Then, it coordinates at least one surrounding vessel (hereinafter referred to as the second vessel) in the current sea area near the abnormal target vessel to respond to the target vessel exhibiting abnormal behavior. In some simple scenarios, vessels around the abnormal target vessel can simply give way; however, in some complex scenarios, such as rescue, firefighting, coast guard and coastal defense inspections, pirate attacks, or even military operations, in addition to unrelated vessels giving way, some working vessels in the vicinity need to quickly approach and control the abnormal target vessel. During the approach, the state of the target vessel may change at any time, and contingency plans must be prepared for these possible changes.

[0064] Therefore, in an embodiment of the present invention, a ship path planning model is further provided to plan the movement of surrounding second ships according to preset actions based on the future movement trajectory of the first ship. This ship path planning model realizes the initial joint interaction of multiple ships.

[0065] Specifically, in this embodiment of the invention, the ship path planning model selects appropriate preset actions from the execution plan library based on the judgment result of the abnormal behavior of the first ship (such as normal cruising when no abnormality occurs, avoidance when there is a risk of collision, rescue / parallel observation when an accident / emergency occurs, approach control or forced interception when there is illegal vessel or suspicious behavior, etc.). It determines the initial conditions of the actions (such as relative distance, relative speed, deviation direction, timing of action start, and future destination, etc.) based on the future trajectory of the first ship and the real-time data of the second ship (current position, speed, heading, etc.). The model then simulates the execution of the preset actions to obtain the planned path for the second ship. Furthermore, the judgment of abnormal behavior of the first ship and the selection of corresponding actions for the second ship can be improved by maintaining a continuously updated strategy library. This strategy library is expanded and improved by automatically collecting confirmed abnormal behaviors and judgment criteria, while simultaneously recording the actual executed response actions, to achieve the identification of abnormal behaviors and the selection of preset actions.

[0066] When the second vessel intervenes in response to the abnormal behavior of the first vessel, the situation on the ground becomes more complex. For example, the target first vessel may detect the intervention of the second vessel at any time and may adopt different behaviors depending on its own circumstances, such as active interaction (e.g., actively communicating and signaling, stopping for inspection, or actively approaching), passive interaction (e.g., reducing speed but not stopping, refusing to communicate, or ignoring communication and continuing to sail), or malicious interaction (e.g., accelerating / turning to escape, defensive behavior, or aggressive behavior). In order to effectively deal with various emergencies, this embodiment of the invention further predicts the possible behavioral changes of the first vessel through a vessel dynamic interaction model and adjusts the planning of the second vessel's movement path accordingly.

[0067] Specifically, in this embodiment of the invention, considering the mutual influence between the behaviors of the target vessel (first vessel) and the vessel performing the action (second vessel), sequence modeling techniques are used to capture the dynamic interaction patterns between them. Preferably, in this embodiment, the sequence modeling technique employs a multi-agent joint sequence modeling framework, based on a dual-channel sequence network constructed using dual LSTM (Long Short-Term Memory) networks, combined with an attention mechanism to capture the dynamic interaction patterns between the target vessel and the law enforcement vessel. Specifically, the first LSTM network model in the dual LSTM network is primarily used to process the data from the first vessel, while the second LSTM network model is used to process the data from the second vessel.

[0068] In embodiments of the present invention, the ship trajectory prediction model, the ship path planning model, and the ship dynamic interaction model can be three relatively independent models, each with different input data and performing corresponding functions, achieving functional continuity only through the cascading of partial input and output data. In another preferred embodiment, a model can be implemented by sub-functions of other models. Specifically, in the dual LSTM network of the ship dynamic interaction model, the first LSTM network model can be used to implement the basic functions of the ship trajectory prediction model and dynamically updated in the interaction mode to meet the needs of the ship dynamic interaction model; the second LSTM network model can be used to implement the basic functions of the ship path planning model and dynamically updated in the interaction mode to meet the needs of the ship dynamic interaction model.

[0069] Further as Figure 2As shown, the architecture of the ship dynamic interaction model using the above preferred embodiment consists of five layers: an input layer, an encoder, an interaction layer, a decoder, and an output layer. The dual LSTM network mainly performs dynamic updates in the interaction layer based on the iteration of interaction behavior. Specifically, the input layer processes the data of the first ship (including historical and real-time data) and the second ship (including historical and real-time data), normalizing features such as position and velocity, and combining the time step and the number of steps to generate time-related serialized data, resulting in a time-discrete sequence of motion states (e.g., including position, velocity, direction, and acceleration).

[0070] The encoder extracts key features from the time-series data of the first and second ships using the encoding module of an LSTM network. The first LSTM network model is used to extract features from the time-discrete motion state sequence of the first ship, yielding key features of its historical trajectory data. Similarly, the second LSTM network model is used to extract features from the time-discrete motion state sequence of the second ship, yielding key features of its historical trajectory data.

[0071] The interaction layer is used to capture the dynamic interaction relationship between the first and second vessels. It employs an attention mechanism, calculating interaction weights based on key features of the historical trajectory data of both vessels—that is, the dependence of the first vessel on the second vessel's time step. Then, it uses these weights to fuse the key features of the second vessel, generating enhanced interaction features for the first vessel. The same method is used to generate enhanced interaction features for the second vessel, ultimately outputting the interaction features of both vessels. The enhanced interaction features are augmented features obtained by fusing information from another vessel with weights on top of the original features, still focusing on individual vessels. The interaction features, on the other hand, are feature vectors specifically describing the interaction relationship between the two vessels, focusing on the interaction relationship and no longer distinguishing between individual vessels.

[0072] Specifically, the interaction layer first extracts key features from the historical trajectory data of the first and second vessels. Preferably, information such as position, velocity, direction, and acceleration are extracted from the historical trajectory data of the first and second vessels, and arranged into a time-series format according to time step order. The key feature sequence of the first vessel can be: F1 = [f 1_1 f 1_2 , ..., f 1_T The key feature sequence of the second ship: F2 = [f 2_1 f 2_2 , ..., f 2_T ]; where f 1_t f 2_t The feature vector consists of information such as position, velocity, direction, and acceleration at time t, where T is the total number of steps after dividing the historical time according to the time step.

[0073] Define the first ship query vector Q1, the second ship key vector K2, the first ship value vector V1, and the second ship value vector V2 as follows:

[0074] Q1 = F1·W q ,

[0075] K2 = F2·W k ,

[0076] V1 = F1·W v ,

[0077] V2 = F2·W v ,

[0078] Among them, W q W k W v This is the weight matrix to be learned.

[0079] The attention score is defined as: Where d k The dimension of the key vector is used to control scaling.

[0080] The interaction dependency weights are then obtained using the softmax function, i.e., interaction dependency weights α = softmax(S); the specific softmax function used can be:

[0081] Where α ij S represents the dependency weight between the state of the first ship at time step i and the state of the second ship at time step j. ij and S ik This represents the raw attention score for the corresponding time step.

[0082] Then, interactive enhancement features for the first and second ships are generated respectively. The interactive enhancement feature for the first ship is F. 1e =F1 + α·V2; that is, for each time step t, the interactive enhancement feature component of the first ship is Where f 1_t Let α be the original feature vector (d-dimensional vector) of the first ship at time step t. tj Let v be the interaction weight (a scalar) between the first vessel at time step t and the second vessel at time step j. 2j Let F be the value vector (d-dimensional vector) of the second vessel at time step j. The interaction enhancement feature of the second vessel is F. 2e =F2+α'·V1; where α' is the transpose of matrix α.

[0083] Based on the interaction enhancement features of each vessel, the interaction features between the two vessels are obtained (i.e., the interaction enhancement feature components are used to update the feature vectors in the key feature sequences of the two vessels).

[0084] The decoder predicts the future trajectories of the first and second vessels based on key features extracted by the encoder and interaction features output by the interaction layer. It can further receive current state information (such as current position, speed, and direction) of the second and / or first vessels for iterative processing. The first LSTM network model decoder can be used to predict the trajectory of the first vessel step-by-step (e.g., updating the cell states of the sequence based on the forward dependencies of the time-series data), outputting a sequence of future trajectory points for the first vessel (position, speed, direction, time, etc.); the second LSTM network model decoder can be used to predict the trajectory of the second vessel step-by-step (e.g., updating the cell states of the sequence based on the forward dependencies of the time-series data), outputting a sequence of future trajectory points for the second vessel (position, speed, direction, time, etc.); during the output process, the interaction dependency weights can also be dynamically updated based on real-time data.

[0085] The output layer outputs at least one prediction result for the first and second ships after the current time step, sequentially according to the trajectory point sequence. Preferably, the prediction results include a prediction of the trajectory of the first ship and a path planning for the second ship. For example, based on a multilayer perceptron (MLP) or a fully connected layer, the adjacent grid and probability of each ship at the next time step are predicted step by step. Simultaneously, the policy library judges the probability of the result (e.g., collision risk) based on the prediction results of the two ships and selects the corresponding interaction strategy (mainly selecting the response action of the second ship). The prediction of the ship trajectory may include multiple results with different probabilities. Therefore, when outputting the predicted trajectory, these different results can be output according to their probability, and the corresponding probability information can be attached to each result. Since the preferred embodiment of the present invention can be iteratively updated, the output of the output layer can also be dynamically updated accordingly.

[0086] It should be noted that the ship dynamic interaction model in this embodiment of the invention supports dynamic feedback updates. During the dynamic interaction of multiple ships, the states of the first and second ships will change in real time, which will inevitably affect the subsequent trajectory prediction. This embodiment of the invention can combine the previous output results of the model with dynamic real-time data and re-input the model to iteratively calculate the interaction features and iteratively simulate the output, thereby updating the trajectory prediction of both ships.

[0087] In a preferred embodiment of the present invention, the dual LSTM networks of the ship dynamic interaction model can each produce preliminary outputs even when some data support is lacking (especially when data from the other party is lacking), obtaining their respective preliminary prediction results. Subsequently, after obtaining data from both parties, the interaction layer is processed and the encoder is iterated, thereby updating the predictions of both parties in real time. In this way, when the first LSTM network model outputs its initial output, the initial prediction result of the future trajectory of the first ship can be obtained, realizing the function of the ship trajectory prediction model; when the second LSTM network model outputs its first iteration (using the aforementioned initial prediction result for iteration), the initial planning of the motion path of the second ship can be obtained (the predicted motion trajectory is the planned motion path), realizing the function of the ship path planning model; subsequently, in the process of more iterations of the dual LSTM network model, the changing motion trajectory of the first ship is predicted one by one, and the planning of the motion path of the second ship is adjusted accordingly, realizing the function of the ship dynamic interaction model.

[0088] Furthermore, the ship dynamic interaction model can also incorporate environmental factors (such as wind, waves, ocean currents, etc.) or motion data from other vessels. These factors can be used as additional inputs to the model to comprehensively simulate and predict the motion trajectories of the first and second vessels, thereby further improving the realism of the simulation.

[0089] On the other hand, embodiments of the present invention provide a trajectory-based multi-ship dynamic joint planning system, the system comprising:

[0090] A ship trajectory prediction model is used to predict the future trajectory of at least one first ship based on historical data of at least one first ship;

[0091] A ship path planning model is used to identify the behavior pattern of at least one first ship based on the future motion trajectory, determine the preset actions and initial conditions of at least one second ship according to the behavior pattern, and simulate the execution of the preset actions to plan the motion path of at least one second ship.

[0092] A ship dynamic interaction model is used to predict changes in the behavior of the first ship based on the future trajectory of the first ship and the motion path of the second ship, and to adjust the planning of the motion path of the second ship based on the changes in behavior.

[0093] Based on further improvements to the above system, the ship dynamic interaction model uses a dual-channel sequence network to capture the dynamic interaction behavior between the first ship and the second ship.

[0094] Based on further improvements to the above system, the dual-channel sequence network is a dual LSTM network, wherein the first LSTM network model in the dual LSTM network processes the data of the first ship to realize the function of the ship trajectory prediction model; the second LSTM network model processes the data of the second ship to realize the function of the ship path planning model.

[0095] Based on further improvements to the above system, the ship dynamic interaction model includes an input layer, an encoder, an interaction layer, a decoder, and an output layer, wherein...

[0096] The input layer processes the data from the first ship and the second ship to generate time-series data;

[0097] The encoder extracts key features from the time-series data of the first and second ships through the encoding module of the LSTM network;

[0098] The interaction layer captures the dynamic interaction relationship between the first vessel and the second vessel, and obtains the interaction characteristics of the first vessel and the second vessel.

[0099] The decoder predicts the future trajectories of the first and second vessels based on the key features and the interaction features.

[0100] The output layer outputs at least one prediction result for the first and second ships after the current moment, item by item, according to the trajectory point sequence.

[0101] Based on further improvements to the above system, the system also includes:

[0102] A dynamic mesh module is used for prediction using a dynamically adjusted neighborhood orientation mesh, the dynamic radius of which is:

[0103] R = R0 + k1·v - k2·p - k3·k + E;

[0104] Where R0 is the initial grid radius, v is the current velocity, k1 is the velocity adjustment coefficient; p is the trajectory density, k2 is the trajectory density adjustment coefficient; k is the average curvature value, k3 is the curvature adjustment coefficient; and E is the environmental attribute correction value.

[0105] In summary, by simulating and predicting the dynamic interactive behavior of multiple vessels, the embodiments of the present invention can effectively identify, prevent, and intervene in emergencies in complex environments. This enables real-time monitoring and assessment of vessel navigation status, enhances the adaptability and effectiveness of emergency plans, reduces the risk of maritime accidents, and provides scientific and technical support for multi-vessel management, especially joint management of multiple vessels in complex near-shore environments. At the same time, it strengthens the emergency response capability to emergencies and comprehensively improves maritime traffic safety and control efficiency.

[0106] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0107] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A trajectory-based multi-ship dynamic joint planning method, characterized in that, The method includes the following steps: Using an established ship trajectory prediction model, predict the future trajectory of at least one first ship based on historical data of at least one first ship; Using an established ship path planning model, the preset actions and initial conditions of at least one second ship are determined based on the future motion trajectory of the first ship, and the preset actions are simulated to plan the motion path of the at least one second ship. Using the established ship dynamic interaction model, the dynamic interaction behavior between the first ship and the second ship is obtained based on the future trajectory of the first ship and the motion path of the second ship, and the motion path planning of the second ship is adjusted based on the dynamic interaction behavior.

2. The method according to claim 1, characterized in that, The ship dynamic interaction model uses a dual-channel sequence network to obtain the dynamic interaction behavior between the first ship and the second ship.

3. The method according to claim 2, characterized in that, The dual-channel sequence network is a dual LSTM network, wherein the first LSTM network model in the dual LSTM network processes the data of the first ship, and the second LSTM network model processes the data of the second ship.

4. The method according to claim 1, characterized in that, The ship dynamic interaction model includes an input layer, an encoder, an interaction layer, a decoder, and an output layer, wherein... The input layer processes the data from the first ship and the second ship to generate time-series data; The encoder extracts key features from the time-series data of the first and second ships through the encoding module of the LSTM network; The interaction layer captures the dynamic interaction relationship between the first vessel and the second vessel, and obtains the interaction characteristics of the first vessel and the second vessel. The decoder predicts the future trajectories of the first and second vessels based on the key features and the interaction features. The output layer outputs at least one prediction result for the first and second ships after the current moment, item by item, according to the trajectory point sequence.

5. The method according to claim 1, characterized in that, At least one model in the method uses a dynamically adjusted neighborhood orientation grid for prediction, the dynamic radius of which is: R = R0 + k1·v - k2·p - k3·k + E; Where R0 is the initial grid radius, v is the current velocity, k1 is the velocity adjustment coefficient; p is the trajectory density, k2 is the trajectory density adjustment coefficient; k is the average curvature value, k3 is the curvature adjustment coefficient; and E is the environmental attribute correction value.

6. A trajectory-based multi-ship dynamic joint planning system, characterized in that, The system includes: A ship trajectory prediction model is used to predict the future trajectory of at least one first ship based on historical data of at least one first ship; A ship path planning model is used to determine the preset actions and initial conditions of at least one second ship based on the future motion trajectory of the first ship, and to simulate the execution of the preset actions to plan the motion path of the at least one second ship. The ship dynamic interaction model is used to obtain the dynamic interaction behavior between the first ship and the second ship based on the future motion trajectory of the first ship and the motion path of the second ship, and to adjust the planning of the motion path of the second ship based on the dynamic interaction behavior.

7. The system according to claim 6, characterized in that, The ship dynamic interaction model uses a dual-channel sequence network to capture the dynamic interaction behavior between the first ship and the second ship.

8. The system according to claim 7, characterized in that, The dual-channel sequence network is a dual LSTM network, wherein the first LSTM network model in the dual LSTM network processes the data of the first ship, and the second LSTM network model processes the data of the second ship.

9. The system according to claim 6, characterized in that, The ship dynamic interaction model includes an input layer, an encoder, an interaction layer, a decoder, and an output layer, wherein... The input layer processes the data from the first ship and the second ship to generate time-series data; The encoder extracts key features from the time-series data of the first and second ships through the encoding module of the LSTM network; The interaction layer captures the dynamic interaction relationship between the first vessel and the second vessel, and obtains the interaction characteristics of the first vessel and the second vessel. The decoder predicts the future trajectories of the first and second vessels based on the key features and the interaction features. The output layer outputs at least one prediction result for the first and second ships after the current moment, item by item, according to the trajectory point sequence.

10. The system according to claim 6, characterized in that, The system also includes: A dynamic mesh module is used for prediction using a dynamically adjusted neighborhood orientation mesh, the dynamic radius of which is: R = R0 + k1·v - k2·p - k3·k + E; Where R0 is the initial grid radius, v is the current velocity, k1 is the velocity adjustment coefficient; p is the trajectory density, k2 is the trajectory density adjustment coefficient; k is the average curvature value, k3 is the curvature adjustment coefficient; and E is the environmental attribute correction value.

Images