Method and system for trajectory prediction of maritime moving targets with intent guidance and environmental constraints

Abstract

The application provides a maritime moving target trajectory prediction method and system based on intention guidance and environmental constraints, and belongs to the technical field of maritime traffic control and trajectory prediction. The method comprises the following steps: extracting a maritime traffic channel based on AIS trajectory data; statistically analyzing the environmental characteristics of channel nodes and edges; representing historical trajectories as a set of channel nodes and edges; constructing a behavior imitation learning model, and obtaining a global reward function through a neural network; training an intention discrimination model to output the behavior probability of short-term channel nodes and long-term ports or anchorages; based on the observed trajectory and the intention discrimination result, the velocity sequence is decomposed using wavelet transform under the guidance of intention, the velocity change is predicted, and the future trajectory coordinates are generated by interpolation. The application constructs a channel network in a data-driven manner, avoids subjective bias of artificial rules, introduces an intention guidance mechanism to suppress long-term prediction error accumulation, and combines with environmental feature statistics to enhance the adaptability to complex marine scenarios.

Description

Technical Field

[0001] This invention relates to the field of maritime traffic control and trajectory prediction technology, and in particular to a method and system for predicting the trajectory of a moving maritime target under intent guidance and environmental constraints. Background Technology

[0002] Predicting the trajectory of moving targets at sea is one of the core technologies for maritime situational awareness and intelligent control decision-making, and it is widely used in fields such as maritime surveillance, navigation safety, and port management. Accurately predicting the future trajectory of maritime targets is of great significance for ensuring navigation safety, optimizing route planning, and preventing maritime risks. However, the movement of moving targets at sea is influenced by multiple factors, including sea state changes, shipping lane habits, port distribution, and weather conditions. Lacking fixed, structured network constraints, their behavior exhibits significant randomness and uncertainty, especially in long-term trajectory prediction, which faces the severe challenge of error accumulation.

[0003] Existing methods for predicting the trajectory of moving targets at sea can be mainly divided into two categories:

[0004] One category is prediction methods based on imitation and generative models. These methods assume that the motion of moving targets follows certain underlying rules or constraints. They simulate future trajectories by mining behavioral patterns from expert experience or historical trajectory data and constructing rule-based or generative models. Researchers typically abstract the interactions between moving targets as attraction or repulsion, or use models such as cellular automata and probabilistic neural networks to simulate target behavioral decisions. With the development of deep learning, variational autoencoders, generative adversarial networks, and attention mechanisms have been introduced into trajectory generation tasks. These methods extract spatiotemporal features from trajectories through representation learning, generating interpretable future trajectories. However, the abstraction of behavioral rules in these methods is highly subjective, making it difficult to fully characterize the real decision-making process in complex marine environments. The generated trajectories generally suffer from realism issues and have limited generalization ability.

[0005] Another category is data-based learning-based prediction methods. These methods view the future trajectory of a moving target as a mapping of the interaction between its historical behavior and environmental features. By mining motion patterns, temporal dependencies, and spatial correlations within historical trajectories, they construct end-to-end prediction models. Commonly used techniques include recurrent neural networks, long short-term memory networks, encoder-decoder structures, and Gaussian process regression. Some studies have introduced techniques such as bundle search and Monte Carlo sampling to address the need for multimodal prediction. These methods can model target behavior and decisions in high-dimensional space and are highly adaptable. However, their predictive performance is highly dependent on the quality and coverage of the training data, and the problem of error accumulation becomes increasingly prominent as the prediction time increases, making it difficult to effectively cope with the uncertainties of future environmental and target decision-making.

[0006] In summary, existing trajectory prediction methods suffer from the following common problems in complex maritime scenarios: (1) Imitation learning-based methods are highly subjective, lacking in trajectory realism and generalization ability; (2) Data learning-based methods have high requirements for data quality, resulting in severe accumulation of long-term prediction errors and weak interpretability; (3) Maritime moving targets lack clear ground constraints and exhibit strong randomness in their behavior, making it difficult for existing methods to accurately capture their long-term movement patterns and intention changes. Therefore, a high-precision trajectory prediction method that can integrate environmental constraints and behavioral intentions and adapt to random maritime scenarios is urgently needed. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for predicting the trajectory of a moving maritime target under intent guidance and environmental constraints. This method overcomes the problems of weak realism, poor generalization ability, insufficient interpretability, and long-term prediction error accumulation in current trajectory prediction methods. It effectively solves the problem of long-term trajectory prediction error accumulation caused by the randomness of future scenarios and the uncertainty of behavioral decisions of moving maritime targets.

[0008] To achieve the above objectives, this invention provides a method for predicting the trajectory of a moving maritime target under intent guidance and environmental constraints, comprising the following steps:

[0009] Step S1: Based on the AIS trajectory dataset after unifying spatiotemporal coordinates, combined with the island data exclusion layer, and considering the density of trajectory sampling points of historical visits by moving targets at sea, construct a maritime traffic channel network.

[0010] Step S2: Based on the maritime traffic channel network, statistically analyze the speed distribution, heading distribution, and sea depth distribution characteristics of key nodes and adjacent areas of maritime traffic channels.

[0011] Step S3: Represent each historical trajectory with a destination of a port or anchorage as a set consisting of key nodes and edges of each maritime traffic channel, and divide it into a training set and a validation set;

[0012] Step S4: Construct a behavior imitation learning model, solve the objective function through a neural network model, output the reward values ​​of each directed node and edge of the maritime traffic channel, and obtain the global reward function by weighted fusion.

[0013] Step S5: Train the intent discrimination model. Input the trajectory dataset, the set of ports or anchorages, the statistical features of maritime traffic channels, and the global reward function. Output the short-term key nodes of maritime traffic channels and the long-term port or anchorage behavior probability of the moving target at sea.

[0014] Step S6: Based on the observed trajectory of the moving target, the set of ports or anchorages, the statistical characteristics of maritime traffic channels, the global reward function, and the intent discrimination model, determine the short-term and long-term behavioral intent of the moving target. Under the guidance of the intent, use wavelet transform to decompose the velocity sequence, predict the velocity change of the moving target, and interpolate to generate future trajectory coordinates.

[0015] Preferably, step S1 includes the following steps:

[0016] Step S11: For the trajectory sampling points after unifying coordinates, removing outliers, and resampling, calculate the local density of each point:

[0017] Step S12: Calculate the minimum distance from each trajectory sampling point to any higher density point;

[0018] Step S13: Calculate the decision value for each trajectory sampling point, sort them by decision value, determine the number of key nodes through inflection point analysis, and select the top... The trajectory sampling points were selected as candidates for key nodes in maritime traffic channels.

[0019] Step S14: Remove candidate nodes located within the polygons of the island data exclusion layer to obtain a set of key nodes for maritime traffic channels;

[0020] Step S15: Construct Voronoi elements for each key node of the maritime traffic channel and calculate the element boundary using the perpendicular bisector.

[0021] Step S16: Based on the Voronoi diagram, restore the topological relationships between nodes to form a maritime traffic channel network.

[0022] Preferably, step S2 includes the following steps:

[0023] Step S21: For each key node of a maritime traffic channel, statistically analyze the velocity distribution characteristics of all trajectory sampling points within its neighborhood, including the maximum value, minimum value, median, and average value.

[0024] Step S22: Statistically analyze the distribution characteristics of ground headings, including the maximum value, minimum value, median, and average value;

[0025] Step S23: Statistically analyze the distribution characteristics of the ship's bow heading, including the maximum value, minimum value, median, and average value;

[0026] Step S24: Statistically analyze the ocean depth distribution characteristics, including the maximum value, minimum value, median, and average value;

[0027] Step S25: Combine the spatial location of each key node in the maritime traffic channel with its corresponding speed, heading to the ground, heading of the ship, and sea depth distribution characteristics to form the environmental feature vector of that node.

[0028] Preferably, step S3 includes the following steps:

[0029] Step S31: Based on the proximity principle, match the preprocessed trajectory sampling points to key nodes of the maritime traffic channel, and describe the trajectory as a process from the starting point through the node sequence to the port or anchorage.

[0030] Step S32: Use a physical dynamics model to interpolate and complete the AIS trajectory data, and use differential equations to describe the relationship between the position of the moving target and time.

[0031] Step S33: In the physical dynamics model, further consider the rate of change of velocity and direction;

[0032] Step S34: Discretize the differential equation and update the velocity and direction according to the time interval;

[0033] Step S35: Update the position coordinates using the updated velocity and direction, combined with the time interval;

[0034] Step S36: Represent the processed trajectory as a structured data sequence containing timestamp, position coordinates, heading to ground, bow heading, instantaneous speed, and water depth information;

[0035] Step S37: Select trajectories whose endpoints are located within a predefined port or anchorage area as expert trajectories, and represent them as a structured form of node access sequence and destination label.

[0036] Preferably, step S4 includes the following steps:

[0037] Step S41: Model the behavior of the moving target as a Markov decision process and construct a reward function parameterized by a neural network to evaluate the immediate reward obtained under a specific state-action pair.

[0038] Step S42: Design a multilayer perceptron as a reward function approximator, inputting the state-action joint feature vector and outputting a scalar reward value;

[0039] Step S43: For any candidate path, its cumulative reward is defined as the sum of the rewards of each edge in the path;

[0040] Step S44: Employ the maximum entropy inverse reinforcement learning framework, with the objective function being to maximize the log-likelihood of the expert trajectory, thereby maximizing the probability of the expert trajectory in all feasible paths.

[0041] Step S45: Train the neural network using an optimization algorithm until the log-likelihood of the validation set no longer increases, and obtain the final reward function;

[0042] Step S46: For each directed edge in the maritime traffic channel network, calculate its expected reward value and obtain the global reward function by weighted fusion according to the historical trajectory access frequency.

[0043] Preferably, step S5 includes the following steps:

[0044] Step S51: Extract temporal features representing motion continuity and environmental interaction features representing the interaction between the target and the environment from the processed structured trajectory data;

[0045] Step S52: Based on the global reward function, calculate the maximum cumulative reward from the current key node of the maritime traffic channel to each candidate node as the reward feature;

[0046] Step S53: Add long and short time intent labels to the input trajectory segment, using the actual next node after the last observation node as the short time label and the final port or anchorage as the long time label.

[0047] Step S54: Use oversampling technology to balance the sample imbalance problem of long and short time intent labels and generate synthetic samples;

[0048] Step S55: Train the short-term and long-term intent discrimination model and output the short-term intent and long-term intent of the moving target. The short-term intent is the probability of selecting the key node of the next maritime traffic channel, and the long-term intent is the probability distribution of the destination port / anchorage.

[0049] Preferably, step S6 includes the following steps:

[0050] Step S61: Input the current observed trajectory segment of the moving target, obtain the long and short time intent probability distribution through the trained intent discrimination model, and select the short time channel node and long time port / anchorage with the highest probability as the intent target;

[0051] Step S62: On the maritime traffic route network that integrates reward values, with the criterion of maximizing cumulative reward, plan the optimal path from the current position to the short-term intention target, and then from the short-term intention target to the long-term intention target, and connect them to obtain the complete planned path;

[0052] Step S63: Extract the velocity sequence from the historical observation sequence, perform multi-level discrete wavelet transform on it, and decompose it into approximate coefficients that characterize the low-frequency trend and detail coefficients that characterize fluctuations at different frequencies.

[0053] Step S64: Construct a path-constrained prediction network. The input is a feature vector formed by concatenating the wavelet coefficients after standardization and flattening with the current speed and path reference point. The output is the speed increment and heading angle for the next multiple steps.

[0054] Step S65: Based on the predicted velocity increment and heading angle, reconstruct the predicted coordinates for future moments using the kinematic model, and output a multimodal predicted trajectory that conforms to the navigation preferences of the moving target.

[0055] This invention also provides an intention-guided and environment-constrained trajectory prediction system for moving maritime targets, used to implement the aforementioned intention-guided and environment-constrained trajectory prediction method for moving maritime targets, comprising:

[0056] The waterway extraction module is used to extract maritime traffic channels based on the AIS trajectory dataset after the spatiotemporal coordinates are unified, combined with the island data exclusion layer, and considering the density of trajectory sampling points of historical visits by moving targets at sea.

[0057] The environmental feature statistics module is used to statistically analyze the speed distribution, heading distribution, and sea depth distribution characteristics of key nodes and the adjacent areas of maritime traffic channels, based on maritime traffic routes.

[0058] The trajectory representation module is used to represent each historical trajectory with a destination of a port or anchorage as a set consisting of key nodes and edges of each maritime traffic channel, and to divide it into a training set and a validation set;

[0059] The behavior imitation learning module is used to build a behavior imitation learning model. It solves the objective function through a neural network model, outputs the reward values ​​of each directed node and edge of the maritime traffic channel, and then weights and fuses them to obtain the global reward function.

[0060] The intent discrimination module is used to train the intent discrimination model. It takes as input trajectory dataset, port or anchorage set, maritime traffic channel statistical features and global reward function, and outputs the short-term maritime traffic channel key node and long-term port or anchorage behavior probability of the moving target at sea.

[0061] The trajectory prediction module is used to determine the short-term and long-term behavioral intentions of a moving target based on the observed trajectory of the moving target, the set of ports or anchorages, the statistical characteristics of maritime traffic channels, the global reward function, and the intention discrimination model. Under the guidance of the intention, the module uses wavelet transform to decompose the velocity sequence, predicts the velocity changes of the moving target, and interpolates to generate future trajectory coordinates.

[0062] Therefore, the present invention employs the above-described method and system for predicting the trajectory of a moving maritime target under intent guidance and environmental constraints, and the beneficial technical effects are as follows:

[0063] (1) To address the problem that traditional imitation learning methods rely on manually abstracted behavioral rules, resulting in strong subjectivity and insufficient trajectory realism and generalization ability, this invention extracts maritime moving target traffic channels through a data-driven approach and constructs a channel network based on the density of actual trajectory sampling points, thus avoiding the subjective bias of human rules. Furthermore, it employs neural inverse reinforcement learning to objectively learn the global reward function from historical trajectories, accurately reflecting the target's decision preferences in different environments. By combining the statistics of channel node environmental characteristics (speed, heading, water depth distribution), the model can adapt to different sea areas and target types, improving the realism of predicted trajectories and the model's generalization ability.

[0064] (2) To address the shortcomings of data-based learning methods in long-term time-series prediction where errors accumulate step by step, this invention introduces an intent-guided mechanism. By training a long- and short-term intent discrimination model, it accurately identifies the short-term channel node access intent and the long-term port or anchorage destination intent of a moving target. Under intent guidance, the optimal path is planned based on a global reward function, combining macro-level behavioral intent with micro-level speed prediction. This effectively constrains the divergence trend of the predicted trajectory, suppresses error accumulation from a mechanistic perspective, and thus improves the accuracy of trajectory prediction over long time spans.

[0065] (3) To address the problems of mobile targets at sea lacking fixed ground constraints, exhibiting strong randomness in behavior, and the poor adaptability and interpretability of existing methods, this invention maps discrete trajectories into structured graphs by constructing a traffic and waterway network, and statistically analyzes the environmental characteristics of the node neighborhoods, enabling the model to perceive changes in dynamic environmental factors such as sea conditions and waterway habits. Simultaneously, the intention discriminant model's output of waterway node access probabilities and destination selection probabilities provides an interpretable intermediate process for the prediction results, not only improving the model's adaptability to complex marine scenarios but also providing a reliable decision-making basis for applications such as maritime traffic monitoring and navigation safety early warning. Attached Figure Description

[0066] Figure 1 This is a flowchart of the method for predicting the trajectory of a moving target at sea under the guidance and environmental constraints of the present invention.

[0067] Figure 2 Map for marine area research;

[0068] Figure 3 Comparison of predicted trajectories of moving targets at sea (Trajectory 1);

[0069] Figure 4 Comparison of predicted trajectories of moving targets at sea (trajectory 2);

[0070] Figure 5 Comparison of predicted trajectories of moving targets at sea (trajectory 3);

[0071] Figure 6Comparison of predicted trajectories of moving targets at sea (trajectory four);

[0072] Figure 7 Comparison of predicted trajectories of moving targets at sea (Trajectory 5). Detailed Implementation

[0073] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0074] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0075] Example 1

[0076] like Figure 1 As shown, the method for predicting the trajectory of a moving maritime target under intent guidance and environmental constraints includes the following steps:

[0077] Step S1: Based on the AIS trajectory dataset after unifying spatiotemporal coordinates, and combined with the island data exclusion layer, the density of trajectory sampling points of historical visits by moving targets at sea is considered to extract maritime traffic channels.

[0078] The original AIS trajectory dataset is represented as Each trajectory It consists of a series of sampling points. Indicates the first The first of the trajectories Each sampling point at the timestamp latitude and longitude Instantaneous velocity is Ground heading The ship's bow course The current water depth is To simplify notation, superscripts will be omitted below without causing confusion. Direct writing:

[0079] ;

[0080] Step S11: For the trajectory sampling points after unifying coordinates, removing outliers, and resampling, calculate the local density of each point:

[0081] ;

[0082] in, For the first Local density of each trajectory sampling point For trajectory sampling points and trajectory sampling points The Euclidean distance between them is represented by the maximum distance traveled over 3 minutes at the instantaneous speed of each trajectory sampling point in the maritime AIS trajectory dataset; This represents the cutoff distance, taking the percentile of the distances between all points. It is a step function. .

[0083] Step S12: Sample each trajectory point Calculate the minimum Euclidean distance from it to all points with higher local density. :

[0084] ;

[0085] in, The first and Local density of each trajectory sampling point; Indicates that all satisfying local density Greater than point Find the match The distance value with the smallest distance; Indicates when When it is the point with the maximum global density, then... Set the distance to the farthest point to ensure it is identified as the cluster center.

[0086] Step S13: Based on each trajectory sampling point Local density and minimum Euclidean distance Calculate the decision value for each point. Sort by decision value and select the top Several trajectory sampling points were selected as candidates for key nodes in maritime traffic channels, and the K value was determined through... The inflection point was determined through analysis.

[0087] Step S14: Remove candidate nodes located within the polygons of the island exclusion layer to obtain the set of key nodes for maritime traffic channels. ,in Indicates the first The spatial coordinates of key nodes are derived from the original trajectory sampling points. Mid-decision value The top K-ranked players.

[0088] Step S15: Target the set of key nodes Build Figure i, the first Key Nodes corresponding i unit Defined as:

[0089] ;

[0090] in, Indicates the first indivual The i-element represents all elements on the plane that are connected to the i-element. The distance is no greater than that to any other joint node. A set of points at a distance; Represents the coordinates of any point in the sea area; Euclidean norm ( Norm, which is the distance between two points; This represents two distinct elements in the set of key nodes.

[0091] Step S16, based on By analyzing the adjacency relationships of unit i, an undirected graph is constructed to reconstruct the topological relationships of key nodes in maritime transportation.

[0092] ;

[0093] in, This is a map of a maritime transportation network, describing the accessibility between key nodes; The set of key nodes in the graph; Describe the edge set of a graph, if two i unit and If a boundary segment is shared, then at the critical nodes... and Add an undirected edge between them, denoted as . .

[0094] Step S2: Based on the maritime traffic channels, statistically analyze the speed distribution, heading to ground, heading to the bow, and depth distribution characteristics of key nodes and adjacent areas of the maritime traffic channels. For the heading to ground (COG) and heading to the bow (HDG) data of moving targets at sea within each key node area, since their value range is... Due to its periodicity, this invention performs phase correction on the heading sequence before statistical analysis: First, it determines whether the difference between the maximum and minimum values ​​of all heading values ​​within the region exceeds 180°. If it does, the data is considered to have crossed the 0° / 360° boundary, and all heading values ​​greater than 180° are subtracted by 360°, ensuring the entire sequence is distributed within a continuous linear interval. Subsequently, the maximum, minimum, median, and average values ​​are calculated on the corrected sequence. Finally, the results are uniformly converted back to... This serves as the heading characteristic of the node.

[0095] Step S21: For each key node of the maritime traffic channel Statistically analyze the velocity distribution characteristics of all trajectory segments within its neighborhood, including the maximum value. Minimum value , median ,average value ;

[0096] ;

[0097] ;

[0098] in, Indicates falling into i unit The total number of valid sampling points, This represents the characteristic distribution of the speed of moving targets in the sea area corresponding to key nodes of maritime traffic routes. Indicates the first The speed of each trajectory sampling point.

[0099] Step S22: Statistically analyze the distribution characteristics of ground headings, including the maximum value. Minimum value , median ,average value ;

[0100] ;

[0101] ;

[0102] in, This indicates the characteristic distribution of the heading relative to the ground of moving targets in the sea area corresponding to key nodes of maritime traffic channels. Indicates the corrected number The ground heading of each trajectory sampling point.

[0103] Step S23: Analyze the distribution characteristics of the ship's bow heading, including the maximum value. Minimum value , median ,average value ;

[0104] ;

[0105] ;

[0106] in, This indicates the characteristic distribution of the heading of a moving target ship in the sea area corresponding to a key node in the maritime traffic channel. Indicates the corrected number The trajectory sampling points are the ship's bow heading.

[0107] Step S24: Statistically analyze the ocean depth distribution characteristics, including the maximum value. Minimum value , median ,average value ;

[0108] ;

[0109] ;

[0110] in, This indicates the characteristic distribution of water depth for moving targets in the sea area corresponding to key nodes of maritime traffic routes. Indicates the first The ocean depth at which each trajectory sampling point is located.

[0111] Step S25: Identify key nodes of each maritime traffic route. The statistical features are concatenated into a sixteen-dimensional environmental feature vector. .

[0112] Step S3: Represent each historical trajectory with a destination of a port or anchorage as a set consisting of key nodes and edges of each maritime traffic channel, and divide it into a training set and a validation set.

[0113] Step S31: Process the original AIS trajectory dataset All trajectories are spatiotemporally aligned and missing values ​​are interpolated; a uniform time interval is set. For each trajectory In time interval Linear interpolation is performed within the interval to generate equally spaced trajectory sequences. ;

[0114] ;

[0115] ;

[0116] in, After interpolation at equal time intervals, the first... The first of the trajectories One trajectory sampling point, Indicates the first The number of sampling points for each trajectory This is the interpolated timestamp. This represents the index of the sampling points after interpolating the original trajectory at equal time intervals. Indicates the first The total number of sampling points after trajectory interpolation They represent the interpolated first and second digits respectively. The latitude, longitude, heading to the ground, heading at the bow, and sea depth of each trajectory sampling point are recorded. Each state variable is obtained by linear interpolation of adjacent valid sampling points. If there are invalid observations (signal loss greater than 30 minutes) in a certain time period, the trajectory segment is deleted.

[0117] Step S32, based on The i-nearest neighbor principle matches the interpolated trajectory to key nodes of the maritime traffic channel, describing the trajectory as a process from the starting point through a sequence of nodes to the port or anchorage.

[0118] Constructed using step S15 i picture The interpolated trajectory sequence Each sampling point in spatial location Mapped to its parent Unit i determines its associated key node index. :

[0119] ;

[0120] Therefore, we obtain the first... The node access sequence of the trajectory is :

[0121] .

[0122] Step S33: Eliminate moving targets at sea in a single Repeated nodes caused by short-term lingering within the i-graph cell, for Perform continuous deduplication:

[0123] ;

[0124] in, And satisfy For all If true, it means that adjacent nodes in the deduplicated trajectory sequence cannot be repeated. Indicates the first The number of duplicate nodes after deduplication of a trajectory. Indicates the first The number of trajectory sampling points after interpolation processing. Indicates the first After deduplication of the trajectory, the first The key node number for each access. For the first The visited node in the deduplication sequence of the trajectory Key node indexes.

[0125] Step S34: Combine port and anchorage geographic information (predefined termination area) Each For a polygonal region, the endpoints of each trajectory are labeled and filtered: if the endpoint... If the destination is a port or anchorage, then it is marked as such. Otherwise, it will be marked as "Destination Not Found". This indicates the number of port anchorages in the study area; Indicates the first A predefined port anchorage area.

[0126] Step S35: Filter expert trajectories with a clear destination. Only trajectories whose ends stably lie within a predefined port or anchorage area are included in the intent learning dataset, and the index of that area is used as the destination label.

[0127] Ultimately, each expert trajectory is represented in a structured manner as follows: :

[0128] .

[0129] Step S36: For each key node in the structured expert trajectory Associate the corresponding environmental feature vectors (calculated in step S25) to all structured expert trajectory datasets. Grouping targets by their MMSI (Unique Target Identifier) ​​at sea, 80% are randomly selected as the training set. The rest are validation sets. This ensures that the trajectory of the same moving target at sea does not appear simultaneously in the training and validation set.

[0130] Step S4: Construct a behavior imitation learning model. Using the maximum entropy inverse reinforcement learning framework, the local rewards of state-action pairs are modeled through neural networks and weighted and fused into a global reward function to quantify potential decision preferences in the maritime mobile target traffic environment.

[0131] Step S41: Model the moving target at sea as a Markov decision process. State space Each state Includes location, motion status, water depth, and index of the current critical node. Action space Actions in Indicates starting from the current node Select the next adjacent node ,satisfy (See step S16, Maritime Traffic Network Map) ); State transition Determined by the kinematic model of a moving maritime target, the transition in the expert trajectory is a deterministic transition, representing a state change. Execute action Then, transition to the new state. The probability; reward function The parameterization learned by the neural network represents the state. Execute action Instant rewards received This represents the learnable parameters of a neural network.

[0132] Step S42: Design the reward network structure. A multilayer perceptron is used as the reward function approximator. The input is a joint state-action feature vector (including 38-dimensional features such as the current node's environmental features (calculated in step S25 for the 16-dimensional features of key nodes), the next node's environmental features (calculated in step S25 for the 16-dimensional features of key nodes), inter-node geometric features (distance, azimuth), and the current motion state of the moving target at sea (speed, heading towards land, bow heading, and water depth). The output is a scalar reward value. Reward Network The system includes hidden layers (3 fully connected layers, 256 neurons per layer). The first two layers use ReLU activation functions, and the third layer uses Tanh to limit the reward range. These are the learnable weight matrix and bias terms.

[0133] Step S43: Construct the global reward function. For any candidate path index... Its cumulative reward is defined as the sum of the rewards of each side:

[0134] ;

[0135] in, For moving targets at sea at nodes The current state (expert trajectory state). For the node Transfer to node The action, .

[0136] Step S44: Define the imitation learning objective function. Given an expert trajectory dataset. (From step S36, the trajectory with the port anchorage as the destination), each trajectory corresponds to a state-action sequence. , For the first The number of time steps for each trajectory. The maximum entropy objective is to maximize the log-likelihood of the expert trajectory:

[0137] ;

[0138] in, This represents the imitation learning objective function (log-likelihood), which maximizes the probability of the expert trajectory across all possible paths. This represents the set of all feasible paths from the origin to the same destination. This indicates that when the parameter is Under the reward function, path selection The reward obtained. This objective encourages the model to assign higher rewards to expert paths while maintaining the exploratory nature of the policy (maximum entropy regularization).

[0139] Step S45: Select the Adam optimizer, learning rate Weight decay The batch size is 32, the training iterations are 100, and an early stopping strategy is used: training is terminated if the validation set loss does not decrease for 15 consecutive iterations. After training, the parameters are fixed until the log-likelihood of the validation set no longer increases. The final reward function is obtained. .

[0140] Step S46: For the maritime traffic network map Each directed edge in Calculate its typical reward value;

[0141] ;

[0142] in, Represents node index Index to adjacent nodes Expected reward value, For the node index in the training set All state samples, each node The overall value is obtained by weighted average of its side rewards.

[0143] Step S5: Train the intent discrimination model. Input the processed structured trajectory dataset, port or anchorage set, maritime traffic channel statistical features and global reward function. Output the short-term intent (probability of key node selection in maritime traffic channel) and long-term intent (probability distribution of destination port / anchorage) of the moving maritime target.

[0144] Step S51: Extract multi-dimensional features from the processed structured trajectory data. For the moving target at time... observation window Extracting temporal features that characterize the continuity of motion Interaction characteristics with the environment :

[0145] ;

[0146] ;

[0147] ;

[0148] in, Input trajectory segment (time window) The sequence of motion states (including latitude and longitude, speed, heading relative to the ground, and bow heading), etc. An LSTM encoder with 128 hidden units. These are the learnable parameters of the LSTM. Key nodes defined for step S25 Environmental statistical characteristics The slave node output in step S46 All possible actions for departure The reward value, This represents vector concatenation. This is a feature of fusion.

[0149] Step S52: Construct supervised labeled training samples. After processing the input, structure each trajectory in the trajectory dataset. Add long and short time labels before retrieving the trajectory. Each node simulates online observation, at the last observation node. Extracting features Assign short-term tags Long-term tags are .

[0150] Step S53: Perform category balancing on long and short time intent labels. For each minority class (i.e., the intent class with fewer samples) samples... Perform SMOTE oversampling to generate synthetic samples to balance the class distribution.

[0151] For each sample feature Calculate its five nearest neighbors in a set of similar samples:

[0152] ;

[0153] Randomly select one of the nearest neighbors Generate synthetic samples:

[0154] ;

[0155] Assign the same label to the synthetic samples, repeat the process until the difference between samples of all classes does not exceed 5%, and output the balanced training set. , Indicates a uniform distribution Scalars of random sampling These are the original minority class sample features. For the selected nearest neighbors, The newly generated synthetic feature vector has labels that match the original minority class sample features. The corresponding tags are the same.

[0156] Step S54: Train the moving target length-time intention discrimination model. Input the balanced training set. Output the probability distribution of behavioral intentions for short and long time periods. , (Because the action space of short-term intent prediction is dynamic and low-dimensional, while long-term intent involves high-dimensional fixed categories and requires reliable probability output, gradient boosting trees and fully connected neural networks are used for modeling to balance efficiency, accuracy and probability calibration performance.)

[0157] ;

[0158] ;

[0159] in, Indicates an action in the action space Selected Based on features The original score output by the gradient boosting tree model. For learnable parameters, Transform the raw scores into a probability distribution. , , , These are the learnable parameters for the first and second layers of the neural network, respectively.

[0160] After the model is trained, the parameters are fixed for real-time intent discrimination. During the inference phase, the system receives a time-series observation of any moving target, extracts fused features, calls the trained model, and outputs the probability of the moving target's short-term intent.

[0161] Step S6: After obtaining the short-term and long-term behavioral intentions of the moving target, construct a hybrid prediction model that combines intention guidance and environmental constraints. Through wavelet transform and LSTM network, model the non-stationary dynamics of the moving target's motion at sea, and generate a future trajectory that conforms to the moving target's navigation preferences.

[0162] Step S61: Input the moving target at time... A sequence of observations is used to obtain the short- and long-term intent distribution of the moving target through the intent discrimination model trained in step S5. and Select the one with the highest long-term intent probability. Each port / anchorage is considered the most likely destination for each destination. ,from Departing from the observation position at any time, via the maritime transportation network (Step S16 output) to maximize cumulative rewards. (The output of step S46) is a heuristic function that plans the optimal path. .reserve The high-scoring paths are the candidate path set (In this invention's experiments, only the prediction result with the highest probability is selected as the final short-term and long-term intention, i.e., let...) ).

[0163] Step S62: For each candidate path The interpolation is compared with the prediction timestamp (the sampling interval of this invention is...). Predicted duration Aligned reference coordinate sequence:

[0164] ;

[0165] in, These are the latitude and longitude coordinates of the predicted trajectory points after interpolation.

[0166] Step S63: Extract velocity sequences from historical observation sequences. , Given the length of the historical observation sequence, a three-level Discrete Wavelet Transform (DWT) is applied, using the Daubechies-4 wavelet basis. This decomposes the sequence into three scale approximations representing the low-frequency trend term and detail coefficients representing fluctuations at different frequencies. All coefficients are standardized, flattened, and then concatenated with the current velocity and reference point coordinates to form the input feature vector.

[0167] Step S64: Train a path-constrained prediction network. The path-constrained prediction network uses LSTM as its core and employs two layers of LSTM (128 hidden units per layer) to output the future path of the moving target. Step speed increment and heading angle .

[0168] Step S65: Reconstruct the predicted coordinates and output the predicted trajectory of the moving target in multimodal mode.

[0169] ;

[0170] ;

[0171] in, , For the first Predicted latitude and longitude For the first Step speed estimate To predict the speed increment of the network output, For the first The heading angle of the step, the sampling interval of this invention is .

[0172] The final predicted trajectory is: .

[0173] The invention will be further illustrated below with specific examples.

[0174] The study area for moving maritime targets was located in a certain sea area (107-123°E, 5-23°N). AIS trajectory data collected in 2019 was used, containing 241 moving maritime target trajectories. The sampling interval was 3 minutes. Each trajectory included the target's ID, longitude, latitude, timestamp, speed, heading towards land, and bow heading. The maritime study area is as follows: Figure 2 As shown.

[0175] To more accurately measure the prediction performance of the moving target trajectory prediction model proposed in this invention, this invention constructs four indicators: average displacement error (ADE), hit rate (HR), final displacement error (FDE), and maximum absolute error (MaxAE) to evaluate the trajectory prediction effect of the model from different perspectives.

[0176] Wherein, the average displacement error ADE represents the average distance between the predicted trajectory sampling points and the actual trajectory sampling points at each time step within the prediction time domain:

[0177] ;

[0178] in, To predict the number of sampling points within the time period, For the first Predicted coordinates at each time point. For the first The actual coordinates at each moment.

[0179] Hit Rate (HR) represents the distance between the actual trajectory point and the predicted trajectory at each time step within the prediction time domain, within a given tolerance range, and the distance threshold. Ratio:

[0180] ;

[0181] in, This indicates that the condition within {} is evaluated; if the condition is met, the counter is incremented by 1. In the prediction of the trajectory of a moving target at sea, the threshold is taken as the fourth quartile of the historical behavior trajectory step of the moving target, which is approximately 1670m.

[0182] The final displacement error (FDE) is used to measure the distance between the predicted trajectory and the actual trajectory at the last time point.

[0183] ;

[0184] in, The predicted coordinates for the last time point. These are the actual coordinates of the last time point.

[0185] Maximum absolute error It is used to measure the maximum distance between the predicted trajectory and the actual trajectory at all time points.

[0186] .

[0187] Comparative verification:

[0188] All comparative models served as the baseline method in this study. To ensure fairness, all model frameworks were implemented using the PyTorch framework and ran in a unified hardware and software environment. The Adam optimizer was used, with an initial learning rate of 1×10⁻⁶. −3 The learning rate was adjusted using cosine annealing. The training batch size was uniformly set to 32, and an early stopping strategy was employed to prevent overfitting. Based on this framework, the hyperparameters of each model were first initialized according to the recommended configurations in the original literature, and then adaptively tuned on a unified validation set to ensure that each model achieves its potential performance on a given task.

[0189] ①LSTM: Introduces a gating mechanism, dynamically adjusting the retention or discard of information through forget gate, input gate, and output gate, learning the dependencies between trajectory points, inputting historical data of the moving target as input, and recursively generating future trajectory points.

[0190] ②GRU: Similar to LSTM, it merges the forget gate and input gate into a single update gate, simplifying the state update process while retaining the ability to learn the long-term dependencies of the target's historical movement behavior.

[0191] ③EKF: Establish a dynamic model of the trajectory of a moving target, describe its motion law with a set of differential equations, describe the relationship between the measured value and the true state with the observation equation, iteratively update the state estimate in two steps of prediction and correction, and use the final state estimate to predict the future trajectory of the moving target.

[0192] ④Transformer: It captures the relationship between any position in the trajectory of a moving target through a self-attention mechanism to model global contextual information. It adopts an encoder-decoder architecture, where the encoder is responsible for extracting features from the input trajectory sequence and the decoder is responsible for generating future trajectory points.

[0193] To demonstrate the results of this invention, five trajectories were selected from the processed AIS trajectory dataset for visual comparison (e.g., Figures 3-7As shown in Table 1-6, the performance of the method of this invention in predicting the trajectory of moving targets at sea (within 60 minutes, trajectory sampling interval of 3 minutes) is analyzed. The results show that the method of this invention outperforms the comparative model in all four indicators, demonstrating its more accurate ability to capture the dynamic behavior of moving targets in long-term time-series predictions. Visualization results further confirm that the method maintains relatively stable prediction performance even in complex waterway environments, especially in judging the behavioral intentions of moving targets.

[0194] Table 1 Comparison of 3-minute prediction results for moving targets at sea

[0195]

[0196] Table 2 Comparison of 9-minute prediction results for moving targets at sea

[0197]

[0198] Table 3 Comparison of 15-minute prediction results for moving targets at sea

[0199]

[0200] Table 4 Comparison of 30-minute prediction results for moving targets at sea

[0201]

[0202] Table 5 Comparison of 45-minute prediction results for moving targets at sea

[0203]

[0204] Table 6 Comparison of 60-minute prediction results of moving targets at sea

[0205]

[0206] It is important to note that in the experimental setting with a prediction duration of 3 minutes (i.e., single-step prediction), since the prediction sequence contains only one time point, the average displacement error (ADE), final displacement error (FDE), and maximum absolute error (MaxAE) are mathematically equivalent, hence the values ​​of the three indicators in Table 1 are completely consistent. As the prediction duration increases (as shown in Tables 2-6, multi-step prediction), the three indicators exhibit differentiated results due to different calculation logics. The average displacement error (ADE) is significantly lower than the other two indicators; while the values ​​of the final displacement error (FDE) and maximum absolute error (MaxAE) remain highly consistent (or equal within the statistical precision range).

[0207] This phenomenon aligns with the kinematic characteristics of moving targets at sea. Due to the large inertia of ships, prediction errors primarily stem from uncertainties in the initial state and the accumulation of environmental disturbances, leading to a monotonically increasing trajectory deviation over time. This results in the error at the end of the trajectory typically being the maximum value in the entire prediction sequence. Thanks to the intent guidance and channel constraint mechanisms proposed in this invention, the generated ship trajectories are smooth and continuous, rarely exhibiting severe oscillations such as "significant deviations followed by regressions." This verifies the physical rationality and stability of the moving target trajectory prediction method for maritime objects in this invention.

[0208] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0209] Therefore, the present invention adopts the above-mentioned intention-guided and environmentally constrained method and system for predicting the trajectory of moving targets at sea. By constructing a traffic channel network through data-driven means and combining trajectory prediction guided by intention discrimination and reward function, it solves the problems of strong subjectivity of human rules, accumulation of long-term prediction errors, and poor adaptability to complex marine scenarios in existing methods, and achieves a significant improvement in the authenticity, accuracy and interpretability of the trajectory prediction of moving targets at sea.

[0210] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting the trajectory of a moving maritime target under intent guidance and environmental constraints, characterized in that, Includes the following steps: Step S1: Based on the AIS trajectory dataset after unifying spatiotemporal coordinates, combined with the island data exclusion layer, and considering the density of trajectory sampling points of historical visits by moving targets at sea, construct a maritime traffic channel network. Step S2: Based on the maritime traffic channel network, statistically analyze the speed distribution, heading distribution, and sea depth distribution characteristics of key nodes and adjacent areas of maritime traffic channels. Step S3: Represent each historical trajectory with a destination of a port or anchorage as a set consisting of key nodes and edges of each maritime traffic channel, and divide it into a training set and a validation set; Step S4: Construct a behavior imitation learning model, solve the objective function through a neural network model, output the reward values ​​of each directed node and edge of the maritime traffic channel, and obtain the global reward function by weighted fusion. Step S5: Train the intent discrimination model. Input the trajectory dataset, the set of ports or anchorages, the statistical features of maritime traffic channels, and the global reward function. Output the short-term key nodes of maritime traffic channels and the long-term port or anchorage behavior probability of the moving target at sea. Step S6: Based on the observed trajectory of the moving target, the set of ports or anchorages, the statistical characteristics of maritime traffic channels, the global reward function, and the intent discrimination model, determine the short-term and long-term behavioral intent of the moving target. Under the guidance of the intent, use wavelet transform to decompose the velocity sequence, predict the velocity change of the moving target, and interpolate to generate future trajectory coordinates. Step S4 includes the following steps: Step S41: Model the behavior of the moving target as a Markov decision process and construct a reward function parameterized by a neural network to evaluate the immediate reward obtained under a specific state-action pair. Step S42: Design a multilayer perceptron as a reward function approximator, inputting the state-action joint feature vector and outputting a scalar reward value; Step S43: For any candidate path, its cumulative reward is defined as the sum of the rewards of each edge in the path; Step S44: Employ the maximum entropy inverse reinforcement learning framework, with the objective function being to maximize the log-likelihood of the expert trajectory, thereby maximizing the probability of the expert trajectory in all feasible paths. Step S45: Train the neural network using an optimization algorithm until the log-likelihood of the validation set no longer increases, and obtain the final reward function; Step S46: For each directed edge in the maritime traffic channel network, calculate its expected reward value, and obtain the global reward function by weighted fusion according to the historical trajectory access frequency. Step S6 includes the following steps: Step S61: Input the current observed trajectory segment of the moving target, obtain the long and short time intent probability distribution through the trained intent discrimination model, and select the short time channel node and long time port / anchorage with the highest probability as the intent target; Step S62: On the maritime traffic route network that integrates reward values, with the criterion of maximizing cumulative reward, plan the optimal path from the current position to the short-term intention target, and then from the short-term intention target to the long-term intention target, and connect them to obtain the complete planned path; Step S63: Extract the velocity sequence from the historical observation sequence, perform multi-level discrete wavelet transform on it, and decompose it into approximate coefficients that characterize the low-frequency trend and detail coefficients that characterize fluctuations at different frequencies. Step S64: Construct a path-constrained prediction network. The input is a feature vector formed by concatenating the wavelet coefficients after standardization and flattening with the current speed and path reference point. The output is the speed increment and heading angle for the next several steps. Step S65: Based on the predicted velocity increment and heading angle, reconstruct the predicted coordinates for future moments using the kinematic model, and output a multimodal predicted trajectory that conforms to the navigation preferences of the moving target.

2. The method for predicting the trajectory of a moving maritime target under intent guidance and environmental constraints according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: For the trajectory sampling points after unifying coordinates, removing outliers, and resampling, calculate the local density of each point: Step S12: Calculate the minimum distance from each trajectory sampling point to any higher density point; Step S13: Calculate the decision value for each trajectory sampling point, sort them by decision value, determine the number of key nodes through inflection point analysis, and select the top... The trajectory sampling points were selected as candidates for key nodes in maritime traffic channels. Step S14: Remove candidate nodes located within the polygons of the island data exclusion layer to obtain a set of key nodes for maritime traffic channels; Step S15: Construct Voronoi elements for each key node of the maritime traffic channel and calculate the element boundary using the perpendicular bisector. Step S16: Based on the Voronoi diagram, restore the topological relationships between nodes to form a maritime traffic channel network.

3. The method for predicting the trajectory of a moving maritime target under intent guidance and environmental constraints according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: For each key node of a maritime traffic channel, statistically analyze the velocity distribution characteristics of all trajectory sampling points within its neighborhood, including the maximum value, minimum value, median, and average value. Step S22: Statistically analyze the distribution characteristics of ground headings, including the maximum value, minimum value, median, and average value; Step S23: Statistically analyze the distribution characteristics of the ship's bow heading, including the maximum value, minimum value, median, and average value; Step S24: Statistically analyze the ocean depth distribution characteristics, including the maximum value, minimum value, median, and average value; Step S25: Combine the spatial location of each key node in the maritime traffic channel with its corresponding speed, heading to the ground, heading of the ship, and sea depth distribution characteristics to form the environmental feature vector of that node.

4. The method for predicting the trajectory of a moving maritime target under intent guidance and environmental constraints according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Based on the proximity principle, match the preprocessed trajectory sampling points to key nodes of the maritime traffic channel, and describe the trajectory as a process from the starting point through the node sequence to the port or anchorage. Step S32: Use a physical dynamics model to interpolate and complete the AIS trajectory data, and use differential equations to describe the relationship between the position of the moving target and time. Step S33: In the physical dynamics model, further consider the rate of change of velocity and direction; Step S34: Discretize the differential equation and update the velocity and direction according to the time interval; Step S35: Update the position coordinates using the updated velocity and direction, combined with the time interval; Step S36: Represent the processed trajectory as a structured data sequence containing timestamp, position coordinates, heading to ground, bow heading, instantaneous speed, and water depth information; Step S37: Select trajectories whose endpoints are located within a predefined port or anchorage area as expert trajectories, and represent them as a structured form of node access sequence and destination label.

5. The method for predicting the trajectory of a moving maritime target under intent guidance and environmental constraints according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Extract temporal features representing motion continuity and environmental interaction features representing the interaction between the target and the environment from the processed structured trajectory data; Step S52: Based on the global reward function, calculate the maximum cumulative reward from the current key node of the maritime traffic channel to each candidate node as the reward feature; Step S53: Add long and short time intent labels to the input trajectory segment, using the actual next node after the last observation node as the short time label and the final port or anchorage as the long time label. Step S54: Use oversampling technology to balance the sample imbalance problem of long and short time intent labels and generate synthetic samples; Step S55: Train the short-term and long-term intent discrimination model and output the short-term intent and long-term intent of the moving target. The short-term intent is the probability of selecting the key node of the next maritime traffic channel, and the long-term intent is the probability distribution of the destination port / anchorage.

6. A system for predicting the trajectory of a moving maritime target under intent guidance and environmental constraints, characterized in that, for implementing the method for predicting the trajectory of a moving maritime target under intent guidance and environmental constraints as described in any one of claims 1-5, it comprises: The waterway extraction module is used to extract maritime traffic channels based on the AIS trajectory dataset after the spatiotemporal coordinates are unified, combined with the island data exclusion layer, and considering the density of trajectory sampling points of historical visits by moving targets at sea. The environmental feature statistics module is used to statistically analyze the speed distribution, heading distribution, and sea depth distribution characteristics of key nodes and the adjacent areas of maritime traffic channels, based on maritime traffic routes. The trajectory representation module is used to represent each historical trajectory with a destination of a port or anchorage as a set consisting of key nodes and edges of each maritime traffic channel, and to divide it into a training set and a validation set; The behavior imitation learning module is used to build a behavior imitation learning model. It solves the objective function through a neural network model, outputs the reward values ​​of each directed node and edge of the maritime traffic channel, and then weights and fuses them to obtain the global reward function. The intent discrimination module is used to train the intent discrimination model. It takes as input trajectory dataset, port or anchorage set, maritime traffic channel statistical features and global reward function, and outputs the short-term maritime traffic channel key node and long-term port or anchorage behavior probability of the moving target at sea. The trajectory prediction module is used to determine the short-term and long-term behavioral intentions of a moving target based on the observed trajectory of the moving target, the set of ports or anchorages, the statistical characteristics of maritime traffic channels, the global reward function, and the intention discrimination model. Under the guidance of the intention, the module uses wavelet transform to decompose the velocity sequence, predicts the velocity changes of the moving target, and interpolates to generate future trajectory coordinates.

Images