A ship trajectory prediction method for a navigable bridge area based on a graph neural network

Abstract

The application relates to a ship trajectory prediction method based on a graph neural network in a navigation bridge area, which comprises the following steps: S1: setting a ship monitoring field in the navigation bridge area, acquiring ship information and scene information; the monitoring field comprises an image information acquisition module, a ship basic information acquisition module and an environment information acquisition module; S2: establishing a ship prediction model in the navigation bridge area based on a graph neural network; historical characteristic data is divided into two parts: single object characteristics and traffic flow characteristics; based on the single object characteristics, a single ship trajectory prediction model is constructed through the graph neural network; random ship flow simulation results are generated and are incorporated into the graph neural network as new characteristics, are combined with the single ship trajectory prediction model, and a ship cluster trajectory prediction model is formed; and S22: model prediction and training. Compared with the prior art, the application has the advantages of multi-source data fusion, improved prediction accuracy and reliability, targeted model design, strong adaptability and the like.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of ship trajectory prediction, in particular to a ship trajectory prediction method for a navigation bridge area based on a graph neural network. BACKGROUND

[0002] In recent years, the global shipping industry has developed rapidly, and the increase in ship-bridge collision accidents highlights the urgent need for a powerful ship trajectory prediction method in accident prevention. However, there are few ship trajectory prediction methods for navigation bridge areas, and traditional ship trajectory prediction methods do not fully consider the environmental characteristics and bridge structure characteristics of the bridge area. SUMMARY

[0003] The purpose of the present application is to overcome the above-mentioned defects of the prior art and provide a ship trajectory prediction method for a navigation bridge area based on a graph neural network, which realizes ship trajectory prediction considering the global information of the navigation bridge area and has strong portability and important reference significance for active ship collision warning of the bridge.

[0004] The application provides a ship trajectory prediction method for a navigation bridge area based on a graph neural network, comprising the following steps:

[0005] S1: setting a ship monitoring field for a navigation bridge area, acquiring ship information and scene information; the monitoring field comprises an image information acquisition module, a ship basic information acquisition module and an environmental information acquisition module;

[0006] The image information acquisition module comprises a bridge deck main camera and a shoreline auxiliary camera, and the image information acquisition module acquires information in a target instance segmentation and target tracking manner; the bridge deck main camera is mainly used for acquiring ship appearance and trajectory features, and has high requirements for the field of view, and a camera with good performance can be considered; considering the cost, the shoreline auxiliary camera is an auxiliary camera and is mainly used to supplement the depth information of the image; the entire image information acquisition module is required to realize full coverage of the selected navigation bridge area.

[0007] The ship basic information acquisition module comprises an AIS data receiver and a radar; the AIS is used to acquire static information, dynamic information, voyage information and safety information of the ship; and the radar is used to compensate for the low recognition accuracy of the image information acquisition module under low visibility conditions.

[0008] The environmental information acquisition module comprises a water level gauge and a wind speed sensor. The water level gauge is used to acquire real-time water level and flow rate, and the wind speed sensor is used to acquire real-time wind speed.

[0009] S2: establishing a ship prediction model for a navigation bridge area based on a graph neural network:

[0010] S21: Data integration and model establishment: Using the entire monitoring system, various types of feature information of the navigation bridge area can be obtained, including environmental features, bridge site features, and ship features. These multi-source data are integrated to form historical feature data. The historical feature data is divided into two parts:

[0011] Single object features: Covering the static and dynamic features of a single ship, as well as the span of the navigation bridge, which does not consider the coupling relationship between similar objects. Based on these single object features, a single ship trajectory prediction model is constructed through a graph neural network.

[0012] Traffic flow features: Used to consider the coupling relationship between ships. Due to the complexity of ship traffic flow features, a multi-level Monte Carlo sampling method is used to generate random ship flow simulation results, which are incorporated into the graph neural network as new features, combined with the single ship trajectory prediction model to form a ship cluster trajectory prediction model.

[0013] S22: Model prediction and training: The data collected by the monitoring field at time t is input into the model as features, and through continuous loop operation, the subsequent trajectory of the ship can be predicted. At the same time, the predicted data can be used as new input for model training to achieve data-driven effect.

[0014] Further, in S1, MMSI, as the internationally unified unique ship code, can be defined as the label of the subsequent target detection model. This module identifies the basic information of the ship as the label in the target instance segmentation and target tracking process, to facilitate data integration later.

[0015] Further, in S1, the information collected by the environmental information collection module also includes bridge and bridge site information, including bridge navigation hole information and shoreline engineering.

[0016] Further, in S21, when integrating multi-source data, data fusion technology is used, including Kalman filtering and Bayesian fusion, to fuse and process data from different sources to improve data accuracy and reliability.

[0017] Further, in S21, feature selection and dimensionality reduction are performed on historical feature data to remove redundant features and retain features that have a significant impact on trajectory prediction. Principal component analysis (PCA) and linear discriminant analysis (LDA) methods are used for feature selection and dimensionality reduction.

[0018] Further, in S21, an attention mechanism is introduced into the graph neural network, allowing the model to pay more attention to important features and nodes for trajectory prediction. The attention mechanism can adaptively assign weights to different features and nodes, improving the prediction accuracy of the model.

[0019] Further, in S21, the model is optimized using a reinforcement learning algorithm. Reinforcement learning can adjust the parameters of the model through interaction with the environment to achieve optimal trajectory prediction. The ship's navigation trajectory is taken as the state, and the prediction accuracy is taken as the reward. The model is continuously optimized through the reinforcement learning algorithm.

[0020] Further, in S22, in addition to using the predicted data for model training, the predicted results can also be compared and analyzed with the actual navigation trajectory to evaluate the prediction accuracy of the model. The mean square error (MSE) and mean absolute error (MAE) indicators are used to evaluate the performance of the model.

[0021] Further, in S21, the data integration of the collected information is achieved through the following methods:

[0022] S211: Determine the data source and type;

[0023] S212: Correlate with the ship as the core. Take the MMSI code of the ship as the key identifier. Correlate the basic information of the ship obtained by the AIS data receiver with the appearance and trajectory information of the ship obtained by the target instance segmentation and target tracking in the image information collection module. Different types of data of the same ship can be matched. For radar data, the same can be matched and associated with the ship in AIS and image information according to the position and other information of the ship. Especially in low visibility, radar data can be used to supplement and correct the deficiencies of ship position information in other data.

[0024] S213: Fuse environmental information and ship information. Fuse the water level, flow rate, wind speed, and bridge and bridge site information obtained by the environmental information collection module with the relevant information of the ship. Combine real-time water level and flow rate information with ship navigation dynamic information to analyze the influence of water flow on ship navigation trajectory. Associate wind speed information with ship stability and other characteristics to consider the effect of wind on ship navigation.

[0025] S214: Construct historical feature data;

[0026] S215: Data preprocessing and normalization.

[0027] Further, in S22, the specific process of constructing the model is as follows:

[0028] S221: Determine the nodes and edges of the graph;

[0029] S222: Select a graph neural network architecture: select GCN graph convolutional neural network or GAT graph attention network or GraphSAGE;

[0030] S223: Design the layer structure of the model:

[0031] Input layer: The initial feature vector of the node is input into the model, various attribute features of the ship, environmental parameters, bridge and bridge site information, these features are preliminarily linearly transformed or embedded mapped through the input layer, and are converted into low-dimensional vector representation suitable for graph neural network processing;

[0032] Hidden layer: Stack multiple graph neural network layers, GCN layer, GAT layer or GraphSAGE layer as hidden layer; in each layer, the features of the neighbor nodes are aggregated and the features of the current node are updated, so as to gradually extract high-order relationships and complex patterns in the graph data; the number of hidden layers and the number of neurons need to be adjusted according to the specific problem and data characteristics, and the optimal structure is generally determined through experiments;

[0033] Output layer: The output layer is designed according to the target of the prediction task; if the trajectory of the ship is predicted, the output layer can be a fully connected layer, which outputs the position coordinates, speed and heading information of the ship at a certain time step in the future; if the ship collision or other events are predicted, the output layer can be a binary classification layer, which outputs the probability value to represent the possibility of collision.

[0034] S224: Define loss function and optimizer;

[0035] S225: Model training and adjustment.

[0036] Compared with the prior art, the present application has the following advantages:

[0037] (1) Multi-source data fusion, improve the prediction accuracy and reliability: This method comprehensively utilizes multi-source data such as image information, ship basic information and environmental information, which can comprehensively describe the ship navigation situation in the navigation bridge area from multiple angles, provides rich information for trajectory prediction, and improves the prediction accuracy and reliability.

[0038] (2) Targeted model design, strong adaptability: Different trajectory prediction models are designed for single ship and ship cluster respectively, considering the coupling relationship between ships and the complexity of traffic flow characteristics, so that the model can better adapt to different scenes and needs.

[0039] (3) Data-driven learning mechanism: The predicted data is used as new input for model training, forming a data-driven learning mechanism, which can continuously optimize the performance of the model and improve the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 Flowchart of the ship trajectory prediction method based on graph neural network in the navigation bridge area.

[0041] Figure 2 Layout diagram of the monitoring field.

[0042] Figure 3 Figure 1 is a diagram of a graph neural network architecture at time t. DETAILED DESCRIPTION

[0043] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. In the technical solution, if the component model, material name, connection structure, control method, algorithm and other features are not explicitly stated, they are considered as common technical features disclosed in the prior art.

[0044] Embodiment 1

[0045] The present application provides a ship trajectory prediction method for a navigable bridge area based on a graph neural network, as shown in Figure 1 The method comprises the following steps:

[0046] S1: As shown in Figure 2 , a ship monitoring field for a navigable bridge area is set up, and ship information and scene information are obtained; the monitoring field comprises an image information acquisition module, a ship basic information acquisition module, and an environmental information acquisition module.

[0047] The image information acquisition module comprises a bridge deck main camera and a shore line auxiliary camera, and the image information acquisition module acquires information in a target instance segmentation and target tracking manner; the bridge deck main camera is mainly used for ship appearance and trajectory feature acquisition, and has high requirements for the field of view, and a camera with good performance can be considered; considering the cost, the shore line auxiliary camera is an auxiliary camera, and is mainly used to supplement the depth information of the image; the entire image information acquisition module is required to achieve full coverage of the selected navigable bridge area.

[0048] The ship basic information acquisition module comprises an AIS data receiver and a radar; the AIS is used to acquire static information, dynamic information, voyage information and safety information of the ship; the radar is used to compensate for the low recognition accuracy of the image information acquisition module under low visibility conditions.

[0049] The environmental information acquisition module comprises a water level gauge and a wind speed sensor. The water level gauge is used to acquire real-time water level and flow rate, and the wind speed sensor is used to acquire real-time wind speed.

[0050] S2: Establish a ship prediction model for a navigable bridge area based on a graph neural network:

[0051] S21: Data integration and model establishment: using the entire monitoring system, various feature information of the navigable bridge area can be acquired, including environmental features, bridge site features and ship features, and these multi-source data are integrated to form historical feature data; the historical feature data is divided into two parts:

[0052] Single object features: Covering the static and dynamic features of a single ship, as well as the span of the navigation bridge and other features that do not consider the coupling relationship between similar objects; based on these single object features, a single ship trajectory prediction model is constructed through a graph neural network.

[0053] Traffic flow features: used to consider the coupling relationship between ships; due to the complexity of ship traffic flow features, a multi-level Monte Carlo sampling method is used to generate random ship flow simulation results, which are incorporated into the graph neural network as new features, combined with the single ship trajectory prediction model, to form a ship cluster trajectory prediction model;

[0054] S22: Model prediction and training: as shown in Figure 3 the data collected by the monitoring field at time t is input into the model as features, and through continuous loop operation, the subsequent trajectory of the ship can be predicted; at the same time, the predicted data can be used as new input for model training to achieve data-driven effect.

[0055] In the specific embodiment, in S1, MMSI, as a unified international ship code, can be defined as the label of the subsequent target detection model. This module identifies the basic information of the ship as the label in the target instance segmentation and target tracking process, to facilitate data integration later.

[0056] In the specific embodiment, in S1, the information collected by the environment information collection module also includes bridge, bridge site information, including bridge navigation hole information, and shoreline engineering.

[0057] In the specific embodiment, in S21, when integrating multi-source data, data fusion technology is used, including Kalman filtering and Bayesian fusion, to fuse and process data from different sources to improve the accuracy and reliability of the data.

[0058] In the specific embodiment, in S21, the historical feature data is subjected to feature selection and dimensionality reduction processing to remove redundant features and retain features that have an important influence on trajectory prediction. Principal component analysis (PCA) and linear discriminant analysis (LDA) methods are used for feature selection and dimensionality reduction.

[0059] In the specific embodiment, in S21, the attention mechanism is introduced into the graph neural network, so that the model can pay more attention to important features and nodes for trajectory prediction. The attention mechanism can adaptively assign weights to different features and nodes, improving the prediction accuracy of the model.

[0060] In the specific embodiment, in S21, the model is optimized by using a reinforcement learning algorithm. The reinforcement learning can adjust the parameters of the model through interaction with the environment to achieve optimal trajectory prediction; the sailing trajectory of the ship is taken as the state, and the prediction accuracy is taken as the reward, and the model is continuously optimized by the reinforcement learning algorithm.

[0061] In the specific embodiment, in S22, in addition to using the predicted data for model training, the predicted results can also be compared and analyzed with the actual sailing trajectory to evaluate the prediction accuracy of the model, and the mean square error MSE and the mean absolute error MAE indicators are used to evaluate the performance of the model.

[0062] In the specific embodiment, in S21, the data integration of the collected information is realized by the following methods:

[0063] S211: Determine the data source and type;

[0064] S212: Correlate around the ship, take the MMSI code of the ship as the key identifier, correlate the basic information of the ship obtained by the AIS data receiver with the appearance and trajectory information of the ship obtained by the target instance segmentation and target tracking in the image information collection module, so that different types of data of the same ship can be corresponded; for radar data, it can also be matched and correlated with the ships in AIS and image information according to the position information of the ship, especially in low visibility, radar data can be used to supplement and correct the deficiency of ship position information in other data;

[0065] S213: Fuse environmental information and ship information, fuse the water level, flow rate, wind speed and bridge information obtained by the environmental information collection module with the related information of the ship; combine the real-time water level and flow rate information with the sailing dynamic information of the ship to analyze the influence of water flow on the sailing trajectory of the ship; associate the wind speed information with the stability and other characteristics of the ship to consider the effect of wind on the sailing of the ship.

[0066] S214: Construct historical feature data;

[0067] S215: Data preprocessing and normalization.

[0068] In the specific embodiment, in S22, the specific process of constructing the model is as follows:

[0069] S221: Determine the nodes and edges of the graph;

[0070] Ship node: Each ship is regarded as a node in the graph. The node features can include static information of the ship, such as ship type, size, load capacity, etc.; dynamic information, such as position, speed, heading, etc.; and voyage information, safety information, etc. in AIS data.

[0071] environment nodes: considering environmental factors as nodes, for example, the location of water level meters and wind speed sensors can be regarded as environmental nodes. The features include real-time water level measured by water level meters, flow rate, real-time wind speed obtained by wind speed sensors, etc.

[0072] bridge and bridge site nodes: setting the locations of bridge navigation holes and shoreline engineering as nodes, and the node features include bridge span, navigation hole size, shoreline shape, etc.

[0073] edges between ships: the edges are determined according to the relative position and distance between ships. For example, when the distance between two ships is within a certain range, it is considered that there is an edge between them, and the edge features can be the distance between the two ships, relative speed, included angle, etc.

[0074] edges between ships and environment: if a ship is within the influence range of a water level meter or a wind speed sensor, an edge is established between the ship node and the corresponding environmental node. The edge features can be the distance between the ship and the environmental monitoring point, the influence degree of the environmental factor on the ship, etc.

[0075] edges between ships and bridges and bridge sites: when a ship approaches a bridge navigation hole or a shoreline engineering, an edge is established between the ship node and the corresponding bridge and bridge site node. The edge features can include the distance between the ship and the navigation hole or the shoreline, the included angle between the ship heading and the navigation hole axis, etc.

[0076] S222: Selecting a graph neural network architecture: selecting a GCN graph convolutional neural network or a GAT graph attention network or a GraphSAGE;

[0077] S223: Designing the layer structure of the model:

[0078] input layer: input the initial feature vector of the node into the model, including various attribute features of the ship, environmental parameters, and bridge and bridge site information. These features are preliminarily linearly transformed or embedded mapped through the input layer, and are converted into low-dimensional vector representations suitable for graph neural network processing;

[0079] hidden layer: stacking multiple graph neural network layers, such as GCN layers, GAT layers or GraphSAGE layers as hidden layers; in each layer, the features of the neighbor nodes are aggregated and the features of the current node are updated to gradually extract high-order relationships and complex patterns in the graph data; the number of hidden layers and the number of neurons need to be adjusted according to the specific problem and data characteristics, and the optimal structure is generally determined through experiments;

[0080] Output layer: design the output layer according to the target of the prediction task; if predicting the trajectory of a ship, the output layer can be a fully connected layer, outputting the position coordinates, speed, and heading information of the ship at a certain future time step; if predicting whether a ship will collide or not, the output layer can be a binary classification layer, outputting a probability value representing the likelihood of collision.

[0081] S224: define the loss function and optimizer;

[0082] Loss function

[0083] Regression loss: if the model's output is continuous trajectory data, etc., the mean squared error (MSE) loss function can be used to measure the difference between the predicted value and the true value, such as calculating the mean squared error between the predicted ship position and the actual observed position.

[0084] Classification loss: for classification tasks, such as predicting whether a ship can safely pass through a navigation bridge area, the cross-entropy loss function can be used to measure the difference between the predicted probability distribution and the true label.

[0085] Optimizer: choose a suitable optimizer to update the model's parameters, such as stochastic gradient descent (SGD) and its variants Adagrad, Adadelta, RMSProp, Adam, etc. The Adam optimizer usually performs well in many scenarios, it can adaptively adjust the learning rate, update the parameters according to the gradient history information of the parameters, which helps the model converge faster.

[0086] S225: model training and adjustment.

[0087] Data division: divide the collected multi-source data into training set, validation set and test set. The training set is used to train the model, let the model learn the patterns and rules in the data; the validation set is used to adjust the model's hyperparameters during training, such as selecting the best number of hidden layers, learning rate, etc.; the test set is used to evaluate the model's performance on unseen data, ensuring the model has good generalization ability.

[0088] Model training: use the training set to train the constructed graph neural network model, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and update the parameters using the optimizer, iteratively train the model, make the model's loss function gradually decrease, and the model's performance continuously improve.

[0089] Hyperparameter adjustment: according to the performance on the validation set, adjust the model's hyperparameters, such as trying different graph neural network architectures, number of hidden layers, number of neurons, learning rate, number of attention heads, etc., through multiple experiments and comparisons, find the optimal combination of hyperparameters that make the model perform best.

[0090] Model evaluation: The trained model is evaluated on the test set using accuracy, recall, F1 score, mean squared error, and other evaluation metrics to measure the performance of the model, ensuring that the model can accurately predict the trajectory of the ship in the navigable bridge area and meet the needs of practical applications.

[0091] The components not described in detail in this embodiment are existing components that can be purchased in public channels.

[0092] The above description of the embodiments is to facilitate the understanding and use of the invention by ordinary skilled persons in the art. Those skilled in the art can easily make various modifications to these embodiments, and apply the general principles described herein to other embodiments without creative labor. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A method for predicting vessel trajectories in navigation bridge areas based on graph neural networks, characterized in that, Includes the following steps: S1: Set up a vessel monitoring field in the navigation bridge area to acquire vessel information and scene information; the monitoring field includes: image information acquisition module, basic vessel information acquisition module, and environmental information acquisition module; The image information acquisition module includes a main camera on the bridge deck and a secondary camera on the shoreline. The image information acquisition module uses target instance segmentation and target tracking to acquire information. The ship basic information acquisition module includes an AIS data receiver and radar; it acquires static information, dynamic information, voyage information and safety information of the ship through AIS; the radar is used to compensate for the low recognition accuracy of the image information acquisition module under low visibility conditions. The environmental information acquisition module includes a water level gauge and a wind speed sensor. The water level gauge acquires real-time water level and flow velocity, while the wind speed sensor acquires real-time wind speed. S2: Establishing a vessel prediction model for navigation bridge areas based on graph neural networks: S21: Data Integration and Model Building: Utilizing the entire monitoring system, various characteristic information of the navigation bridge area can be acquired, including environmental characteristics, bridge site characteristics, and vessel characteristics. This multi-source data is integrated to form historical characteristic data; the historical characteristic data is divided into two parts: Single object features: encompassing the static and dynamic features of a single vessel, as well as the span of a navigation bridge without considering the coupling relationship between similar objects; based on these single object features, a trajectory prediction model for a single vessel is constructed using a graph neural network; Traffic flow characteristics: used to consider the coupling relationships between ships; given the complexity of ship traffic flow characteristics, a multi-level Monte Carlo sampling method is used to generate random ship flow simulation results, which are then incorporated as new features into a graph neural network and combined with a single ship trajectory prediction model to form a ship cluster trajectory prediction model. S22: Model Prediction and Training: The data collected at time t by the monitoring field is used as a feature input into the model. Through continuous iterative operation, the subsequent trajectory of the ship can be predicted. At the same time, the predicted data can be used as a new input for model training to achieve a data-driven effect.

2. The method for predicting vessel trajectories in navigation bridge areas based on graph neural networks according to claim 1, characterized in that, In S1, MMSI, as an internationally unified unique ship code, can be defined as a label for the subsequent target detection model. This module identifies basic ship information and uses it as a label in the target instance segmentation and target tracking process to facilitate later data integration.

3. The method for predicting vessel trajectories in navigation bridge areas based on graph neural networks according to claim 1, characterized in that, In S1, the environmental information acquisition module also collects information on bridges and bridge sites, including bridge navigation channel information and shoreline engineering.

4. The method for predicting vessel trajectories in navigation bridge areas based on graph neural networks according to claim 1, characterized in that, In S21, when integrating multi-source data, data fusion techniques, including Kalman filtering and Bayesian fusion, are used to fuse data from different sources in order to improve the accuracy and reliability of the data.

5. The method for predicting vessel trajectories in navigation bridge areas based on graph neural networks according to claim 1, characterized in that, In S21, feature selection and dimensionality reduction are performed on historical feature data to remove redundant features and retain features that have an important impact on trajectory prediction. Principal component analysis (PCA) and linear discriminant analysis (LDA) are used for feature selection and dimensionality reduction.

6. The method for predicting vessel trajectories in navigation bridge areas based on graph neural networks according to claim 1, characterized in that, In S21, an attention mechanism is introduced into the graph neural network, enabling the model to pay more attention to features and nodes that are important for trajectory prediction.

7. The method for predicting vessel trajectories in navigation bridge areas based on graph neural networks according to claim 1, characterized in that, In S21, reinforcement learning algorithms are used to optimize the model. Reinforcement learning can continuously adjust the model parameters through interaction with the environment to achieve optimal trajectory prediction. The ship's navigation trajectory is used as the state, and the accuracy of the prediction is used as the reward. The model is continuously optimized through reinforcement learning algorithms.

8. The method for predicting vessel trajectories in navigation bridge areas based on graph neural networks according to claim 1, characterized in that, In S22, in addition to using the predicted data for model training, the predicted results can also be compared and analyzed with the actual flight trajectory to evaluate the prediction accuracy of the model. The mean squared error (MSE) and mean absolute error (MAE) are used to evaluate the performance of the model.

9. A method for predicting vessel trajectories in navigation bridge areas based on graph neural networks according to claim 1, characterized in that, In S21, the data integration of the collected information is achieved through the following methods: S211: Specify the source and type of data; S212: Using ships as the core for association, and the ship's MMSI code as the key identifier, the basic ship information acquired by the AIS data receiver is associated with the ship's appearance and trajectory information acquired by the image information acquisition module through target instance segmentation and target tracking, so that different types of data of the same ship can be matched. For radar data, the ship's position information can also be matched and associated with ships in AIS and image information, especially when visibility is low, radar data can be used to supplement and correct the lack of ship position information in other data. S213: Integrate environmental and ship information by combining water level, flow velocity, wind speed, and bridge and bridge site information acquired by the environmental information acquisition module with relevant ship information; combine real-time water level and flow velocity information with ship navigation dynamics information to analyze the impact of water flow on ship navigation trajectory; and correlate wind speed information with ship stability characteristics to consider the effect of wind on ship navigation. S214: Construct historical feature data; S215: Data preprocessing and normalization.

10. A method for predicting vessel trajectories in navigation bridge areas based on graph neural networks according to claim 1, characterized in that, In S22, the specific process of building the model is as follows: S221: Determine the nodes and edges of the graph; S222: Select a graph neural network architecture: Choose GCN graph convolutional neural network, GAT graph attention network, or GraphSAGE; S223: Layer structure of the design model: Input layer: The initial feature vectors of the nodes are input into the model. Various attributes of the ship, environmental parameters, bridge and bridge site information are processed by the input layer through preliminary linear transformation or embedding mapping, converting them into low-dimensional vector representations suitable for graph neural network processing. Hidden layers: Stack multiple graph neural network layers, such as GCN, GAT, or GraphSAGE layers, as hidden layers; in each layer, higher-order relationships and complex patterns in the graph data are gradually extracted by aggregating the features of neighboring nodes and updating the features of the current node; the number of hidden layers and neurons needs to be adjusted according to the specific problem and data characteristics, and the optimal structure is generally determined through experiments; Output layer: The output layer is designed according to the objective of the prediction task. If the goal is to predict the trajectory of a ship, the output layer can be a fully connected layer that outputs the ship's position coordinates, speed, and heading information at a future time step. If the goal is to predict whether a collision event will occur, the output layer can be a binary classification layer that outputs probability values ​​representing the likelihood of a collision. S224: Define the loss function and optimizer; S225: Model training and tuning.

Images