Ship communication behavior prediction method, simulation method and prediction system based on AIS trajectory semantics

Abstract

The application relates to a ship communication behavior prediction method based on AIS trajectory semantics, a simulation method and a prediction system, and belongs to the information technology field, and solves the problem that in the prior art, a business parameter configuration is disengaged from actual communication behavior of a ship in a ship dynamic modeling simulation process. The method comprises the following steps: acquiring ship AIS data of a target ship, performing pretreatment, and acquiring trajectory feature vectors at multiple time steps; performing semantic labeling on the trajectory feature vectors by using a semantic classification model to obtain a scene label time sequence; performing semantic time sequence prediction based on the scene label time sequence to obtain a predicted scene label; and determining a corresponding predicted communication behavior of the ship by using a semantic-communication model based on the scene label time sequence and the predicted scene label. The application realizes AIS trajectory semantic mining, communication behavior prediction and demand dynamic adaptation, and can be applied to intelligent and dynamic configuration of simulation business parameters.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method, simulation method and prediction system for predicting ship communication behavior based on AIS trajectory semantics. Background Technology

[0002] In the field of maritime satellite communication simulation, Automatic Identification System (AIS) data has become a core data source for ship dynamic modeling because it can provide dynamic information such as ship latitude, longitude, speed, and heading. However, the application of AIS data in current technologies is limited to the extraction of basic dynamic parameters, without delving into the intrinsic relationship between the ship's navigation intentions and communication needs hidden behind the trajectory data. This leads to a disconnect between the configuration of business parameters and the actual communication behavior of the ship during the simulation process.

[0003] Current ship communication simulation technology has the following key shortcomings in its application of AIS data: 1) Superficial application of AIS data, lacking semantic mining of trajectory data. Current technologies only use AIS data as a tool to obtain basic parameters such as ship position and speed, without performing semantic analysis on the trajectory data. For example, there are significant differences in the AIS trajectory characteristics of a ship that has been anchored in a port for a long time and is sailing at a constant speed in the open ocean. In real-world scenarios, the former requires high-frequency transmission of data such as cargo scheduling and berth applications, while the latter mainly relies on low-bandwidth equipment monitoring data. However, current simulations have not identified these semantic differences.

[0004] 2) Static communication requirements configuration leads to a disconnect from actual navigation behavior. In existing simulations, parameters such as ship communication service types, bandwidth requirements, and latency thresholds are mostly manually preset or fixed, and cannot be dynamically adjusted according to the ship's real-time navigation status. When the ship's navigation status changes (such as switching from ocean voyage to port operations), communication requirements will change accordingly, but the existing simulation system cannot adapt in a timely manner, resulting in a large deviation between QoS simulation results and actual scenarios.

[0005] 3) Insufficient ability to predict ship communication behavior. Existing technologies have not established a correlation between AIS trajectory characteristics and ship communication behavior, making it impossible to predict future changes in ship communication needs based on historical AIS trajectory data. For example, analyzing historical AIS trajectories can reveal that ships enter port operation areas at fixed times each month, but existing simulations cannot adjust communication service parameters in advance based on this pattern, resulting in insufficient predictive ability and accuracy in simulations.

[0006] Therefore, there is an urgent need for a ship communication behavior prediction method, simulation method, and prediction system based on AIS trajectory semantics to solve the problem of the disconnect between the configuration of service parameters and the actual communication behavior of ships during the simulation process. This would enable AIS trajectory semantic mining, communication behavior prediction, and dynamic adaptation to requirements. It can be applied to the intelligent and dynamic configuration of simulation service parameters, enhance the conformity between ship communication simulation scenarios and actual navigation communication needs, and improve the guiding value of simulation results for the design and optimization of ship satellite communication systems. Summary of the Invention

[0007] Based on the above analysis, the embodiments of the present invention aim to provide a ship communication behavior prediction method, simulation method, and prediction system based on AIS trajectory semantics. Through semantic annotation, time-series prediction, and semantic-communication model, it solves the problems of missing AIS trajectory semantic mining, disconnect between business parameter configuration and actual ship communication behavior, and insufficient prediction capability in the ship dynamic modeling and simulation process of the prior art. It realizes in-depth utilization of AIS data, accurate prediction of communication behavior, and dynamic parameter adaptation.

[0008] On one hand, embodiments of the present invention provide a method for predicting ship communication behavior based on AIS trajectory semantics, including: Obtain the AIS dataset of the target vessel, preprocess the AIS dataset to obtain trajectory feature vectors at multiple time steps; For all the trajectory feature vectors, semantic annotation is performed using a semantic classification model to obtain a temporal sequence of scene labels; Based on the time series sequence of the scene labels, perform time series prediction to obtain the predicted scene labels; Based on the time sequence of the scene labels and the predicted scene labels, the predicted communication behavior corresponding to the target ship is determined using a semantic-communication model.

[0009] Furthermore, the ship AIS dataset is preprocessed to obtain trajectory feature vectors at multiple time steps, including: The ship AIS dataset is subjected to trajectory windowing to obtain trajectory sub-segments at multiple time steps; For each time step, feature extraction is performed on the trajectory sub-segment to obtain the corresponding static and dynamic features; Based on the static and dynamic features of each time step, a trajectory feature vector is constructed for each time step.

[0010] Furthermore, semantic annotation is performed on all the aforementioned trajectory feature vectors using a semantic classification model to obtain a temporal sequence of scene labels, including: The trajectory feature vector at each time step is input into the semantic classification model to obtain the corresponding trajectory scene label; wherein, the semantic classification model is obtained through training; The trajectory scene labels of all the time steps are integrated in time to form the scene label time sequence.

[0011] Further, based on the time-series sequence of the scene labels, time-series prediction is performed to obtain predicted scene labels, including: Obtain the predicted duration, and extract historical tag time series data from the scene tag time series based on the predicted duration; Based on the historical tag time series data, the initial prediction probability of each navigation scenario tag is obtained using an LSTM neural network model; Based on the predicted initial probabilities of all the navigation scene labels, the predicted scene label corresponding to the predicted duration is determined.

[0012] Further, based on the initial probability of prediction for each of the navigation scene labels, determining the predicted scene label corresponding to the predicted duration includes: Based on the aforementioned ship AIS dataset, obtain the corresponding target ship type and target season label; Based on the target vessel type, the target type label weight for each navigation scenario label is determined according to the vessel type-semantic label weight mapping relationship; Based on the target seasonal label, the target seasonal label coefficient for each of the navigation scenario labels is determined according to the seasonal-semantic label coefficient mapping relationship; Based on the target type label weight and the target season label coefficient, the initial prediction probability of each navigation scenario label is corrected to obtain the corresponding corrected prediction probability; Based on the predicted correction probability, the predicted scene label corresponding to the predicted duration is determined.

[0013] Furthermore, the semantic-communication model includes a semantic-communication mapping table; based on the scene label time series and the predicted scene label, the semantic-communication model is used to determine the predicted communication behavior corresponding to the ship, including: Perform time series analysis on the scene label time series and the predicted scene label to determine the currently effective label, and determine whether the predicted scene label is consistent with the currently effective label; If they match, the current communication behavior corresponding to the currently active tag is obtained, and the current communication behavior is used as the predicted communication behavior; wherein, the predicted communication behavior includes the communication service type and the communication parameter range; If there is a discrepancy, the predicted communication behavior is determined using the semantic-communication mapping table based on the predicted scene label.

[0014] Furthermore, the semantic-communication model also includes a feature-semantic mapping table; using the current communication behavior as the predicted communication behavior includes: The communication service type corresponding to the current communication behavior is used as the communication service type corresponding to the predicted communication behavior; Obtain the current discrete feature combination corresponding to the currently effective label, and obtain the predicted discrete feature combination corresponding to the predicted scene label; Based on the currently active label and the feature-semantic mapping table, it is determined whether there is a fluctuation between the current discrete feature combination and the predicted discrete feature combination; If there is no fluctuation, the range of communication parameters corresponding to the current communication behavior shall be used as the range of communication parameters corresponding to the predicted communication behavior.

[0015] Furthermore, before performing time series analysis on the scene label time series and the predicted scene label, the method further includes: Obtain the time sequence of the scene labels and the discrete feature combination corresponding to the predicted scene labels; Based on all the discrete feature combinations, determine the accuracy of the scene label time sequence and the first label of the predicted scene label; The ship communication data of the target ship is acquired, the ship communication data and the ship AIS data are synchronized in time, and the ship communication behavior sequence corresponding to the scene tag time sequence is extracted. Based on the scene tag time sequence and the ship communication behavior sequence, the accuracy of the second tag of the scene tag time sequence is determined according to the semantic-communication mapping table; If the accuracy of the first label and / or the accuracy of the second label fail, the time sequence of the scene label and / or the predicted scene label are reacquired.

[0016] On the other hand, embodiments of the present invention provide a method for simulating ship communication behavior based on AIS trajectory semantics, including: Based on any one of the first aspects, obtain the predicted scene label and the predicted communication behavior; The predicted scene labels and the predicted communication behaviors are standardized to obtain standardized simulation results. The standardized simulation results are subjected to compliance verification, and the standardized simulation results that pass the compliance verification are taken as the final simulation results.

[0017] In another aspect, embodiments of the present invention provide a ship communication behavior prediction system based on AIS trajectory semantics, comprising: The data acquisition module is used to acquire the ship AIS dataset of the target ship, preprocess the ship AIS dataset, and obtain trajectory feature vectors at multiple time steps. The semantic annotation module is used to perform semantic annotation on all the trajectory feature vectors using a semantic classification model to obtain a temporal sequence of scene labels. The semantic prediction module is used to perform semantic temporal prediction based on the temporal sequence of the scene labels to obtain the predicted scene labels; The communication prediction module is used to determine the predicted communication behavior of the ship based on the time series sequence of the scene labels and the predicted scene labels, using a semantic-communication model.

[0018] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. By processing ship AIS data to obtain trajectory feature vectors and performing semantic annotation, the trajectory semantics of AIS data can be mined to obtain scene label time series and further obtain predicted scene labels. Through the established semantic-communication model, predicted communication behavior can be determined, enabling the prediction of future changes in ship communication behavior. This can be used for intelligent and dynamic configuration of service parameters in ship dynamic modeling and simulation, enabling communication service parameters to be dynamically adjusted in real time with AIS data during simulation, improving the consistency between simulation configuration and actual ship communication behavior, solving the problem of disconnect between simulation service parameter configuration and actual ship communication behavior, and significantly improving the realism and reliability of QoS simulation.

[0019] 2. By extracting features from AIS data and obtaining trajectory feature vectors, a semantic classification model is used to obtain trajectory scene labels. The ship's navigation intention is extracted from AIS data, extending the value of AIS data from location description to behavioral understanding. The trajectory scene labels (such as anchorage waiting, port operations, and ocean cruise) are accurately identified, realizing the semantic parsing and in-depth utilization of AIS data. This solves the problems of shallow application and semantic missingness of AIS trajectory data in existing technologies.

[0020] 3. Establish an end-to-end predictive data link from AIS data to the navigation scenario (track scenario label) and then to the predicted communication behavior. This enables the prediction of changes in the future communication needs of ships and can provide dynamic and forward-looking parameter configuration references for subsequent simulation systems, thereby improving the accuracy of ship communication simulation and the guiding value of AIS data.

[0021] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0022] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart illustrating a method for predicting ship communication behavior based on AIS trajectory semantics in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the process of obtaining trajectory feature vectors at multiple time steps in an embodiment of the present invention. Figure 3 This is a schematic diagram of the process of obtaining trajectory scene labels using a semantic classification model in an embodiment of the present invention; Figure 4 This is a schematic diagram of the training process of the semantic classification model in an embodiment of the present invention; Figure 5 This is a schematic diagram of the process for obtaining predicted scene labels based on scene label time series in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the process of performing time-series analysis on scene tag time-series sequences to determine predicted communication behavior in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the process of verifying the accuracy of the scene label time sequence and the predicted scene label in an embodiment of the present invention. Figure 8 This is a schematic diagram of the main modules of a ship communication behavior prediction system based on AIS trajectory semantics in an embodiment of the present invention. Detailed Implementation

[0023] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0024] A specific embodiment of the present invention discloses a method for predicting ship communication behavior based on AIS trajectory semantics, such as... Figure 1 As shown, it includes: Step S1: Obtain the AIS dataset of the target vessel, preprocess the AIS dataset to obtain trajectory feature vectors for multiple time steps.

[0025] Obtain the target vessel's AIS dataset, preprocess the AIS dataset to obtain trajectory feature vectors for multiple time steps, specifically including steps S11-S12.

[0026] Step S11: Obtain the AIS dataset of the target vessel.

[0027] The AIS dataset of the target vessel is obtained. This dataset includes AIS data from multiple sampling time points for the target vessel, which refers to any vessel whose communication behavior needs to be predicted. In this embodiment, the AIS data includes the vessel identifier, vessel type, and multiple navigation parameters. The vessel identifier uses the vessel's MMSI code. The navigation parameters include latitude and longitude, speed, heading, rate of turn, turning angle, positioning time, dwell time, and anchorage / port dwell records (duration or number of times). This data can be acquired in real-time or by inputting a preset or historically collected vessel AIS dataset through a visual interface, serving as the basis for communication behavior prediction.

[0028] In this embodiment, the ship AIS dataset can be the AIS data of the target ship in the last few hours (e.g., the AIS data corresponding to the target ship's port operations in the last 4 hours), or the AIS data of the last few days (e.g., the AIS data collected during the target ship's 72-hour ocean cruise), or the AIS data of multiple weeks or even longer periods, without any limitation.

[0029] Step S12: Preprocess the ship AIS dataset to obtain trajectory feature vectors for multiple time steps.

[0030] The ship AIS dataset is preprocessed to obtain trajectory feature vectors at multiple time steps for subsequent semantic annotation at those time steps. This preprocessing includes trajectory windowing and feature extraction, such as... Figure 2 As shown, step S12 specifically includes steps S121-S123.

[0031] Step S121: Perform trajectory windowing on the ship AIS dataset to obtain trajectory sub-segments at multiple time steps.

[0032] The ship AIS dataset is segmented into tracks to obtain track segments across multiple time steps. Based on the timestamps of the positioning time, the ship AIS dataset is segmented into non-overlapping and non-omission segments according to preset time steps, such as 1 hour and 0.5 hours, following a chronological order. All AIS data corresponding to each time step forms the track segment for that time step. During the segmentation process, the timestamps of the positioning time in the navigation parameters are retained to ensure the continuity and non-overlapping and non-omission of time intervals. This ensures that the time boundaries of each track segment are clear and reproducible, facilitating subsequent prediction and time alignment with communication data. For example, if the ship AIS dataset includes a 72-hour continuous navigation track, and the preset time step is 1 hour, the 72-hour continuous track is divided into 72 track segments with 72 time steps. It should be noted that if the time length of the last window does not reach the preset time step, the actual remaining time length is retained as the last track segment to ensure track integrity and consistency.

[0033] It should be noted that for the target vessel, the position and status points of the vessel at each positioning time can be determined based on the positioning time, latitude and longitude, speed and heading information. These points are used as the trajectory points of the target vessel. The curve or scattered points formed by connecting all the trajectory points in chronological order of the positioning time constitute the navigation trajectory of the target vessel.

[0034] Furthermore, to improve the accuracy of AIS trajectory semantic understanding, preprocessing also includes data cleaning and / or trajectory smoothing, using the cleaned and smoothed ship AIS dataset for trajectory windowing and feature extraction.

[0035] In this embodiment, data cleaning includes outlier removal, determination of the validity of missing data, and completion of missing values.

[0036] 1) Outlier Removal. In this embodiment, outliers refer to values ​​of navigation parameters that exceed the corresponding parameter thresholds. For example, the parameter thresholds for speed are the target ship's maximum design speed, latitude and longitude are 0-90 degrees, and heading is 0-360 degrees. If the speed value at a certain point in time exceeds the target ship's maximum design speed, or the latitude and longitude value is negative or greater than 90 degrees, then the value is an outlier and is removed.

[0037] 2) Determine the validity of missing data, remove invalid missing data, and complete the valid missing data in the next step. In this embodiment, valid missing data refers to scattered and isolated small-scale missing data, which is an occasional packet loss during the collection or transmission process, rather than continuous data loss caused by equipment failure or signal interruption. For example, missing data that meets the following three judgment conditions at the same time is identified as valid missing data: (1) In the continuous collection time sequence of the target ship, the sampling time point of the single sampling time point missing value or the sampling time point of the continuous missing value does not exceed the validity threshold. The sampling time point is the time of each collection of the ship trajectory point, which can be the positioning time (time stamp) as the sampling time point. For example, the validity threshold includes the segment validity threshold (e.g., 2 sampling time points) and the global validity threshold (e.g., 5% sampling point missing rate); (2) The navigation parameter values ​​corresponding to the adjacent sampling time points of the missing value are valid values ​​and there are no abnormal values; (3) Based on the navigation parameter values ​​of the adjacent sampling time points, the missing value has no sudden change. For example, the relative difference of the speed of the adjacent sampling points is less than the speed stability threshold (e.g., 1%), and the ship trajectory satisfies physical continuity after interpolation. Other missing data are invalid missing data. Examples include data with consecutive missing values ​​(e.g., no corresponding value for 5 consecutive sampling time points), missing data with outliers in consecutive data positions, data with missing core identifiers (core identifiers include the ship's MMSI code and positioning time), and missing data when the target ship is actively disconnected.

[0038] 3) Use nearest neighbor interpolation to complete valid missing data. For example, for two sampling time points with consecutive missing values, calculate the average of the navigation parameter values ​​corresponding to the adjacent sampling time points, and use this average as the navigation parameter value corresponding to the two sampling time points to ensure the continuity and accuracy of time information and navigation parameters.

[0039] Furthermore, the cleaning quality of the ship AIS dataset after data cleaning is judged. For example, the effective data retention rate is greater than 95%, and / or the proportion of effective missing data completion does not exceed 3%, and / or the deviation of the completed value from the data at adjacent sampling time points does not exceed 10%. The ship AIS dataset that passes the cleaning quality judgment is then smoothed.

[0040] In this embodiment, trajectory smoothing employs a Kalman filter algorithm to eliminate trajectory fluctuations caused by positioning noise, generating a continuous and smooth ship AIS dataset. Specifically, the navigation parameter data for all sampling time points are arranged according to the chronological order of positioning time (i.e., sampling time points). The Kalman filter algorithm is then used to sequentially filter and optimize the navigation parameters such as speed, heading, latitude, and longitude at each sampling time point, ensuring that the smoothed ship AIS dataset is completely consistent with the time series of the original ship AIS dataset. The core objective of trajectory smoothing is to eliminate temporal trajectory fluctuations caused by positioning noise (such as sudden changes in the position or speed of the same ship at consecutive time points), making the ship's navigation trajectory smooth. The trajectory points retain the original timestamp information (i.e., the timestamp corresponding to the sampling time point) and are continuously connected in the time series without time breaks or disordered order, thus completely restoring the ship's continuous navigation process in the time dimension. It should be noted that trajectory points refer to the ship's position and status points determined by the sampling time point and the corresponding latitude, longitude, speed, and heading information. The navigation trajectory refers to the curve or scattered points formed by connecting all trajectory points in the chronological order of the sampling time points.

[0041] In other embodiments, trajectory smoothing employs a moving average filter. Specifically, a fixed-length filtering sliding window is selected (the window size can be adjusted according to the sampling frequency of the ship AIS dataset, for example, set to 5-10 sampling time points). The navigation parameters (such as speed, heading, latitude and longitude) within the filtering sliding window are calculated by arithmetic mean. The average value is used to replace the original value of the center sampling time point of the filtering sliding window. The smoothing process of the navigation trajectory is completed by sliding the window sequentially.

[0042] Furthermore, the smoothing quality of the ship AIS dataset after trajectory smoothing is judged. For example, the smoothed speed and position coordinates have a good fit with the original valid data of ≥98%, there is no trajectory distortion caused by over-smoothing, and the true navigation trend of the ship is preserved.

[0043] Step S122: Extract features from the trajectory sub-segment at each time step to obtain the corresponding static and dynamic features.

[0044] Feature extraction is performed on the trajectory sub-segment at each time step to obtain the corresponding static and dynamic features, specifically including steps S1221-S1222.

[0045] Step S1221: Extract the core feature dataset corresponding to the core features from the trajectory sub-segment.

[0046] In this embodiment, based on preset core features, each core feature at all sampling time points is extracted from the trajectory sub-segment to form a core feature dataset. The core features can be set empirically, such as speed and heading. For example, for the trajectory sub-segment corresponding to 9:00-10:00, the speed and heading at each sampling time point are extracted to form the core feature dataset corresponding to that trajectory sub-segment.

[0047] Furthermore, to ensure that the determined core features have the ability to accurately express trajectory semantics, this embodiment determines the core features through steps A-D: Step A: Obtain the first ship navigation dataset; wherein, the first ship navigation dataset includes multiple first sample AIS data and corresponding actual scene labels.

[0048] The first ship navigation dataset is obtained, including multiple first sample AIS data and corresponding actual scenario labels. For example, it can be historical navigation AIS datasets of multiple ships collected historically or obtained from public channels. The AIS data for each ship at each sampling time point is the first sample AIS data, and the corresponding navigation scenario is the corresponding actual scenario label. In this embodiment, a navigation scenario label candidate set is constructed based on common scenarios in the ship navigation industry, including five categories of navigation scenario labels: anchorage waiting, port operations, near-shore fishing, ocean cruising, and emergency avoidance, as well as the corresponding basic navigation scenario connotations (e.g., "port operations" corresponds to the loading, unloading, and scheduling-related navigation behaviors carried out by ships in port areas). In other embodiments, this can be adjusted according to actual needs. It should be noted that the actual scenario label corresponding to each first sample AIS data can be selected by experts from the navigation scenario label candidate set based on the first sample AIS data and experience, or it can be determined by combining ship communication behavior with the navigation scenario label candidate set. For example, the first ship navigation dataset also includes ship communication behavior data. The corresponding actual scenario label is determined based on the communication behavior data of each ship at each time point. For example, if the ship communication behavior data includes port scheduling communications such as berth applications and cargo scheduling, then the corresponding actual scenario label is determined to be "port operations."

[0049] Step B: Extract multiple trajectory basic features from the first ship navigation dataset, and calculate the correlation between each trajectory basic feature and the corresponding actual scene label.

[0050] All quantitative indicators in the first ship navigation dataset are extracted as corresponding trajectory basic features, such as latitude and longitude, speed, heading, and positioning time. The Pearson correlation coefficient is used to calculate the correlation between each trajectory basic feature and its corresponding actual scene label. In this embodiment, to improve data quality, based on the ship's MMSI code, all first sample AIS data in the first ship navigation dataset are divided into single-ship navigation AIS datasets for each ship. For each single-ship navigation AIS dataset, outlier removal, effective missing data interpolation completion, and trajectory smoothing are performed according to step S121 before trajectory basic feature extraction. It should be noted that in this embodiment, persistent data loss in the first ship navigation dataset is permissible.

[0051] First, a unique continuous numerical mapping is performed on the navigation scenario labels to obtain the corresponding coded values, satisfying the requirement that the objects of Pearson correlation coefficient analysis must be two continuous variables. For example, the coded value for anchorage waiting is 1, for port operations it is 2, for near-shore fishing it is 3, for ocean cruising it is 4, and for emergency avoidance it is 5. Based on the coded value corresponding to each navigation scenario label, the coded value of the actual scenario label corresponding to each trajectory basic feature is determined.

[0052] Secondly, for each trajectory feature, the Pearson correlation coefficient with the mapped actual scene label is calculated. For example, In the formula, Indicates the first The Pearson correlation coefficient between the basic features of the trajectory and the actual scene label, with a value range of [value range missing]. The closer the absolute value is to 1, the stronger the correlation. Indicates the first In the first sample of AIS data, the first The numerical values ​​of the basic features of a trajectory. Indicates all the first The sample mean of the basic features of each trajectory (e.g. ), Indicates the first The encoded value of the actual scene label corresponding to the first sample AIS data. This represents the sample mean of all actual scene label encoded values ​​(e.g., ), This indicates the number of AIS data points in the first sample.

[0053] Furthermore, to reduce computational load, before calculating the Pearson correlation coefficient, the basic trajectory features can be preliminarily screened based on data stability. The Pearson correlation coefficient is then calculated on the selected basic trajectory features. For example, basic trajectory features that cannot be stably extracted (e.g., those with a data missing rate > 10%) or those clearly lacking industry relevance are removed. Only quantifiable and reproducible basic trajectory features are retained for Pearson correlation coefficient calculation.

[0054] Step C: Determine strongly correlated basic features based on the correlation of each trajectory's basic features.

[0055] Step B quantifies the linear correlation between individual trajectory features and actual scene labels, i.e., correlation. Based on a correlation threshold, trajectory features weakly correlated with actual scene labels are eliminated, and strongly correlated features are identified, further narrowing the scope of core feature selection. For example, trajectory features with a correlation threshold (e.g., 0.6) are retained as strongly correlated features.

[0056] Step D: Perform semantic label discrimination verification on each strongly related basic feature, and determine the strongly related basic features that pass the verification as core features.

[0057] Semantic label discriminability verification is performed on each strongly related basic feature to test whether each strongly related basic feature can effectively distinguish different real-world scene labels, that is, whether there is a significant statistical difference in the value of a certain strongly related basic feature under different real-world scene labels. Strongly related basic features that pass the verification, that is, those strongly related basic features that can effectively distinguish different real-world scene labels, are identified as core features.

[0058] In this embodiment, statistical difference testing methods (such as ANOVA and Kruskal-Wallis test) are used to perform cross-label numerical distribution tests on each strongly correlated basic feature. For example, for each strongly correlated basic feature, the p-value of the difference in values ​​of that feature under different real-world scenario labels is calculated. If there is a significant statistical difference (e.g., p-value < 0.05), it indicates that the strongly correlated basic feature can effectively distinguish different navigation scenarios and is retained; if there is no significant statistical difference (e.g., p-value ≥ 0.05), it indicates that the strongly correlated basic feature cannot support the identification of different real-world scenario labels and is removed.

[0059] Through the correlation analysis in steps B-C and the discrimination verification in step D, the obtained core features can effectively express and distinguish different actual scenario labels, that is, the obtained core features can effectively express and distinguish different navigation scenario labels. In this embodiment, the top 5 strongly correlated basic features with the highest differences are selected as core features to provide a basis for the extraction of subsequent static and dynamic features. For example, the 5 core features are speed, dwell time, heading, turning angle, and anchorage / port dwell records (duration or number).

[0060] Through steps A-D, adhering to the core principle of "aligning with the physical laws of ship navigation and adapting to the needs of trajectory semantic mining," the determination of core features is completed in three steps: extracting basic trajectory features, performing correlation analysis, and verifying discriminability. The entire process is an engineered statistical analysis process without the need for model training, ensuring that the final output core features are strongly correlated with the ship navigation scenario, can effectively distinguish different semantic labels, and can be stably extracted from AIS data. This provides a basis for subsequent static and dynamic feature extraction, as well as the construction of trajectory feature vectors.

[0061] Step S1222: Perform time-series statistics on the core feature dataset to obtain static and dynamic features.

[0062] Temporal statistics are performed on each core feature in the core feature dataset to obtain static and dynamic features. In this embodiment, the temporal statistical results representing the global statistical features of the trajectory sub-segment are used as static features, including average speed, total dwell time, and number of turns; the temporal statistical results representing the temporal changes within the trajectory sub-segment are used as dynamic features, reflecting the real-time trend of navigation status changes, including speed change rate, turning angle distribution, and port / anchorage dwell frequency.

[0063] The average speed is the arithmetic mean of the speeds at all sampling time points of the trajectory sub-segment.

[0064] The total dwell time is the sum of the durations covered by the sampling time points when the speed does not reach the travel threshold. For example, if the travel threshold is 2 knots, and there are two sampling time points where the speed is 1 knot, then the total dwell time is the duration of the two sampling intervals.

[0065] The number of turns is the number of times the heading change angle reaches the turning threshold. For example, the heading change angle... ,like This constitutes one turn, where, , The first The, the The heading corresponding to each sampling time point For example, the turning threshold. .

[0066] The rate of change in airspeed is the maximum ratio of the change in airspeed at adjacent sampling time points to the time interval. For example, In the formula, Indicates the rate of change in air speed. , The first The, the The speed corresponding to each sampling time point , The first The, the The timestamp corresponding to each sampling time point.

[0067] The steering angle distribution includes the mean and variance of the steering angles at all sampling time points of the trajectory sub-segment.

[0068] Port / anchorage dwell frequency is the ratio of the number of times a vessel enters the port / anchorage area to the total duration of a trajectory segment. For example, In the formula, Number of port / anchorage entries This represents the total duration of the trajectory sub-segment.

[0069] Step S123: Construct the trajectory feature vector for each time step based on the static features and dynamic features of each time step.

[0070] After extracting features from the trajectory sub-segments at each time step, the corresponding static and dynamic features are obtained, and the trajectory feature vector at each time step is constructed.

[0071] In this embodiment, all static and dynamic features are first subjected to Z-score standardization to eliminate the dimensional differences between different features. For example... In the formula, Indicates the first The first time step A standardized static feature or a standardized dynamic feature, with a value range. Indicates the first The first time step A static or dynamic feature , Representing the first time step of each time step The mean and standard deviation of each static or dynamic feature are calculated; then, for each time step, all static and dynamic features are arranged in order of static features first, followed by dynamic features, forming the trajectory feature vector for that time step. For example, the first... The trajectory feature vector at each time step is ,in, , , , , , , The numbers represent the order of the numbers. Standardized average speed, standardized total dwell time, standardized number of turns, standardized speed change rate, standardized mean of turn angle distribution, standardized variance of turn angle distribution, and standardized port / anchorage dwell frequency for each time step.

[0072] Based on the physical laws of ship navigation and the needs of trajectory semantic mining, the core features of the trajectory are selected and determined. Then, static features and dynamic features are extracted according to the temporal change features of the features. Finally, the features are standardized and vectorized to form a trajectory feature vector with a unified dimension, which provides standardized input for the subsequent semantic classification model.

[0073] Step S2: For all the trajectory feature vectors, perform semantic annotation using a semantic classification model to obtain the scene label time sequence.

[0074] For all trajectory feature vectors, semantic annotation is performed using a semantic classification model to obtain a temporal sequence of scene labels, specifically including steps S21-S22.

[0075] Step S21: Input the trajectory feature vector of each time step into the semantic classification model to obtain the corresponding trajectory scene label; wherein, the semantic classification model is obtained through training.

[0076] Using the trained semantic classification model, the trajectory feature vector at each time step is semantically labeled to obtain the trajectory scene label corresponding to each time step.

[0077] In this embodiment, the semantic classification model adopts the Transformer model, including a position encoding layer, a Transformer Encoder layer, and an output layer. It adapts to a lightweight conventional structure for trajectory feature vectors, simplifying the number of layers in the conventional position encoding layer and Transformer Encoder layer, thus reducing computational cost. The trajectory feature vector at each time step is input into the trained semantic classification model to obtain the corresponding trajectory scene label, such as... Figure 3 As shown, it includes steps S211-S214.

[0078] Step S211: Based on the trajectory feature vector, obtain the trajectory encoding vector using the position encoding layer.

[0079] In this embodiment, the Positional Encoding Layer uses a fixed positional encoding vector and a trajectory feature vector to perform vector fusion to generate a trajectory encoding vector. This preserves the global correlation of features and simplifies the temporal positional encoding logic of a conventional Transformer. For example, the fixed positional encoding vector... (in, , , , , , , (representing the 1st to 7th position codes in sequence) and trajectory feature vector Using element-wise addition (e.g.) Vector fusion is performed to obtain a trajectory encoding vector with location information, which retains the feature correlation while supplementing the location identifier.

[0080] Furthermore, the semantic classification model also includes an input layer and a feature mapping layer. The input layer receives the trajectory feature vector at each time step; the feature mapping layer maps the trajectory feature vector (7-dimensional) to a high-dimensional feature vector (e.g., 64-dimensional), denoted as the first mapped feature vector. This is implemented using a fully connected layer (weight initialization uses Xavier normal initialization to avoid gradient vanishing), increasing the feature dimensionality and enhancing feature representation, providing sufficient feature information for subsequent self-attention calculations. For example, the 7-dimensional trajectory feature vector... After linear transformation by a fully connected layer and Xavier weight initialization, it is converted into a 64-dimensional first-mapping feature vector. (in, For the mapped first (1 vector element), to complete feature dimension enhancement. The trajectory feature vector (i.e., the first mapped feature vector) after passing through the feature mapping layer is input into the encoding layer to obtain the corresponding trajectory encoding vector. At this point, the dimension of the fixed-position encoding vector is the same as the dimension of the first mapped feature vector. For example, the fixed-position encoding vector... (in, For the first (vector elements) and the 64-dimensional first mapping feature vector By adding element by element, a trajectory encoding vector with location information is obtained. .

[0081] Step S212: Input the trajectory encoding vector into the Transformer Encoder layer and perform self-attention fusion calculation to obtain the fused feature vector; wherein, the Transformer Encoder layer adopts a stack of lightweight Encoders.

[0082] In this embodiment, the Transformer Encoder layer uses two lightweight encoder layers stacked together. Each encoder layer includes a multi-head self-attention layer and a feed-forward network (FFN) layer. The multi-head self-attention layer includes four attention heads, each with a dimension determined by the dimension of the trajectory encoding vector. For example, if a 64-dimensional trajectory encoding vector is obtained after passing through the feature mapping layer and the position encoding layer, then each attention head is set to 16 dimensions. Through attention weight calculation, the correlation between trajectory feature vectors (such as the correlation between speed and number of turns) is captured, and the attention weighted feature vector is output. The feedforward neural network sub-layer adopts a conventional structure of "linear transformation + ReLU activation + linear transformation". The first linear transformation maps the attention feature vector (e.g., 64-dimensional) to a higher-dimensional feature vector (e.g., 256-dimensional), denoted as the second mapped feature vector. The ReLU activation function enhances the nonlinear expressive power. The second linear transformation maps the second mapped feature vector (e.g., 256-dimensional) back to the dimension of the attention feature vector (e.g., 64-dimensional), and outputs the feature vector after nonlinear optimization, i.e., the fused feature vector.

[0083] Furthermore, residual connection layers and layer normalization layers are added to each encoder's multi-head self-attention layer and feedforward neural network layer to avoid gradient vanishing during model training and accelerate model convergence. For example, (1) a 64-dimensional trajectory encoding vector with location information is input into the multi-head self-attention layer of the first layer encoder. The attention weights between features are calculated by the four attention heads respectively. After weighted fusion, a 64-dimensional attention feature vector is output. After passing through the residual connection layer (added with the feature vector input to this layer) and the normalization layer, a standardized 64-dimensional feature vector is obtained; (2) the normalized feature vector is input into the feedforward neural network layer of the first layer encoder. After “64-dimensional → 256-dimensional linear transformation → ReLU activation → 256-dimensional → 64-dimensional linear transformation”, an optimized 64-dimensional feature vector is output. After passing through the residual connection layer and the normalization layer again, the feature optimization of the first layer encoder is completed; (3) the 64-dimensional feature vector output by the first layer encoder is repeated in steps (1)-(2) and input into the second layer encoder. Finally, a 64-dimensional core feature vector after two attention fusions and nonlinear optimizations is output as the fused feature vector.

[0084] Step S213: Input the fused feature vector into the output layer to obtain the corresponding label probability vector.

[0085] In this embodiment, the output layer uses a combination of a fully connected layer and a softmax activation function to map the fused feature vector output from the Transformer Encoder layer into a label probability vector with the same dimension as the number of categories of the navigation scene labels. For example, in this embodiment, the navigation scene labels include 5 categories, so a 5-dimensional label probability vector is output. For instance, a 64-dimensional fused feature vector is input to the fully connected layer and linearly transformed into a 5-dimensional probability vector. After normalization by the Softmax activation function, the final probability vector is output as the label probability vector, thus completing the transformation of data from feature vector to label probability.

[0086] Step S214: Determine the trajectory scene label corresponding to the trajectory feature vector based on the label probability vector.

[0087] The trajectory scene label corresponding to the trajectory feature vector is determined based on the label probability vector. In this embodiment, for the label probability vector of each time step, the navigation scene label corresponding to the dimension with the largest value is selected as the trajectory scene label for that time step, achieving a one-to-one match between each time step and the trajectory scene label. For example, the five navigation scene labels are, in order, anchorage waiting, port operations, near-shore fishing, ocean patrol, and emergency avoidance. The 5-dimensional label probability vector for a certain time step... The third dimension has the largest value, which determines that the trajectory scene label corresponding to this time step is nearshore fishing.

[0088] In other embodiments, the semantic classification model adopts a CNN-LSTM hybrid model. First, the CNN (convolutional neural network) extracts local correlation features (such as the short-term correlation between speed and turning angle) from the trajectory feature vector. Then, the LSTM (long short-term memory network) captures the temporal dependency of trajectory features (such as the change pattern of speed mutation rate in different time periods). Finally, the fully connected layer + Softmax activation function outputs the trajectory scene label corresponding to the trajectory segment.

[0089] By integrating the static and dynamic features of ship navigation trajectories, and based on a semantic classification model, accurate identification of trajectory scene labels is achieved. The classification accuracy of ship navigation scenes can reach 92%, which is a significant improvement over traditional learning algorithms (such as SVM and random forest).

[0090] Step S22: Perform time integration on the trajectory scene labels of all the time steps to form the scene label time sequence.

[0091] The trajectory scene labels of all time steps are integrated in time, and all time steps and their corresponding trajectory scene labels are arranged in chronological order to form a scene label time sequence in the format of "time step: trajectory scene label".

[0092] Furthermore, based on the continuity and consistency of trajectory scene labels, time merging is performed on all time steps and their corresponding trajectory scene labels. Adjacent time steps with the same trajectory scene label are merged into time intervals, and the same trajectory scene label is used as the trajectory scene label for that time interval. All time intervals or time steps and their corresponding trajectory scene labels are sorted in chronological order to form a scene label time sequence. For example, if there are "08:00-09:00: Port Operations", "09:00-10:00: Port Operations", "10:00-11:00: Port Operations", and "11:00-12:00: Nearshore Fishing", then the first three time steps are merged into the time interval 08:00-11:00, forming the scene label time sequence "08:00-11:00: Port Operations; 11:00-12:00: Nearshore Fishing".

[0093] Merging consecutive time steps with consistent trajectory scene labels can simplify semantic annotation results, improve annotation efficiency, reduce the amount of trajectory scene label annotations, reduce data redundancy, facilitate subsequent data storage and manual verification, and better highlight the main navigation scenarios of the target ship. For example, it can clearly present the document navigation status of the target ship within a continuous time period, avoid being fragmented by scattered time steps, better fit the continuity of the target ship, and improve the readability of the scene label time sequence.

[0094] Based on the positioning time included in AIS data, trajectory windows are formed according to a preset time step. The preset time step can be adjusted based on the historical navigation scenarios of different ships. The trajectory sub-segments in the window are semantically labeled by the semantic classification model step by step, ensuring the accuracy and rationality of the scene label time sequence.

[0095] Furthermore, in this embodiment, the semantic classification model is trained through steps E-I. Figure 4 This is a flowchart of the corresponding process.

[0096] Step E: Obtain the second and third ship navigation datasets; wherein, the second ship navigation dataset includes multiple second sample AIS data and corresponding actual scene labels; the third ship navigation dataset includes multiple third sample AIS data.

[0097] In this embodiment, the second vessel navigation dataset includes multiple second sample AIS data sets and corresponding navigation scenario labels. It can be collected from publicly available datasets in the vessel navigation industry or AIS datasets filed with maritime authorities (containing multiple vessel types and navigation scenarios). The second vessel navigation dataset is divided into trajectories using the vessel's MMSI code, with a total data volume of no less than 100,000 navigation trajectories or trajectory segments. The AIS data for each vessel at each sampling time point is used as the second sample AIS data. The actual scenario label corresponding to the second sample AIS data refers to the navigation scenario label corresponding to the second sample AIS data. This label can be selected by experts based on the second sample AIS data and experience from a candidate set of navigation scenario labels, or it can be determined through vessel communication behavior.

[0098] In this embodiment, the third ship navigation dataset includes multiple third sample AIS data, which can be the same as the first ship navigation dataset and / or the second ship navigation dataset, or other AIS datasets in the industry (including multiple ship types and multiple navigation scenarios) can be collected separately, with the AIS data of each ship at each sampling time point as the third sample AIS data.

[0099] It should be noted that the second and third ship navigation datasets in this embodiment can allow for continuous data loss. For example, if a ship is missing data for one day, the navigation trajectory before and after that day will be treated as two trajectory segments of the ship and processed separately.

[0100] Step F: Preprocess the second and third ship navigation datasets to obtain the corresponding trajectory feature vectors.

[0101] Referring to step S12, the AIS data corresponding to each navigation trajectory or trajectory segment in the second and third ship navigation datasets are preprocessed to obtain the trajectory feature vectors of multiple time steps corresponding to each navigation trajectory or trajectory segment.

[0102] Step G: Cluster all trajectory feature vectors corresponding to the second ship navigation dataset to obtain multiple feature vector clusters. Statistically analyze the cluster features of each feature vector cluster to determine the corresponding navigation scene label.

[0103] Clustering is performed on all trajectory feature vectors corresponding to the second ship navigation dataset. In this embodiment, the K-means clustering algorithm is used for unsupervised clustering. Cluster feature statistics are performed on the trajectory feature vectors in each feature vector cluster after clustering, including mean, standard deviation, and extreme value range.

[0104] Based on the actual scene label and cluster features of each second sample AIS data corresponding to each feature vector cluster, and combined with the common scenarios in the shipping industry, the navigation scene label corresponding to each feature vector cluster is determined from the navigation scene label candidate set. For example, if the cluster features of a certain feature vector cluster are characterized by an average speed lower than the travel threshold, an average port / anchorage dwelling frequency higher than the temporary dwelling threshold (e.g., 0.8), and the proportion of port operations among all actual scene labels corresponding to the feature vector cluster is higher than the category threshold (e.g., 0.7), then port operations are determined to be the navigation scene label corresponding to that feature vector cluster.

[0105] Step H: Based on each feature vector cluster and its corresponding navigation scene label, determine the feature range corresponding to each navigation scene label.

[0106] For each feature vector cluster, the numerical distribution of each dimension of all trajectory feature vectors within the cluster is statistically analyzed, i.e., the numerical distribution of each dynamic or static feature. Based on a first confidence threshold (e.g., 95%), a reasonable range for each dynamic or static feature is determined, constituting the feature range corresponding to that feature vector cluster. This forms a navigation scene label-feature range mapping table, completing the automated calibration and normalization of the feature range corresponding to each feature vector cluster, thus obtaining the feature range corresponding to each navigation scene label. It should be noted that the trajectory feature vectors contain standardized dynamic or static features, and the feature range corresponding to the original dimensions of the dynamic or static features can be determined through inverse operations.

[0107] For example, the feature range of dynamic and static characteristics corresponding to port operations: standardized average speed (The corresponding average speed in its original dimensions) Standardized number of turns (i.e., number of turns) Standardized port / anchorage dwelling frequency (i.e., port / anchorage dwell frequency) ); the characteristic range corresponding to ocean cruise: (Right now Standardized speed mutation rate (i.e., speed change rate) Standardized turning angle distribution variance (i.e., uniform distribution of turning angles); the corresponding characteristic range of the anchorage waiting area: (Right now Standardized total stay time (i.e., total length of stay) ); the characteristic range corresponding to nearshore fishing: (Right now Standardized steering angle (i.e., moderate average steering angle); characteristic range corresponding to emergency avoidance: (Right now ), (i.e., steering angle) ).

[0108] Furthermore, to improve the effectiveness of feature range, box plot analysis is used to remove outlier clusters from all feature vector clusters, and the feature range of the remaining feature vector clusters is confirmed. In this embodiment, the number of retained feature vector clusters is consistent with the number of navigation scene labels in the navigation scene label candidate set (e.g., 5 categories). In other embodiments, the proportion of outlier clusters removed is controlled to not exceed 3% of all feature vector clusters. If the number of retained feature vector clusters exceeds the number of navigation scene labels, cluster merging is performed based on the navigation scene label corresponding to each feature vector cluster.

[0109] Furthermore, when the second ship navigation dataset adds new AIS data (e.g., an AIS dataset for a new ship type), or when the navigation scenario label candidate set is updated according to routine scenarios in the ship navigation industry, steps E-I are re-executed to update the scenario label-feature range mapping table. This includes updating the feature ranges corresponding to existing navigation scenario labels and / or adding new navigation scenario labels, updating the feature ranges corresponding to all navigation scenario labels, and simultaneously updating the navigation scenario label candidate set, thereby improving the adaptability and universality of navigation scenario labels and feature ranges. In other embodiments, a periodic update method can also be used.

[0110] By constructing a candidate set of navigation scene labels, the feature range corresponding to each navigation scene label is automatically, standardized, and normalized based on industry data (the second ship navigation dataset), and updated according to the settings. This avoids the subjectivity and large errors of purely manual definition, and solves the problem of the disconnect between purely automated label definition and actual industry scenarios, ensuring the adaptability and universality of navigation scene labels and feature ranges.

[0111] Step 1: Based on the feature range, construct the first training set using the trajectory feature vectors corresponding to all third-sample AIS data, and train the semantic classification model.

[0112] In this embodiment, based on the feature range corresponding to each navigation scene label, a corresponding navigation scene label is added to the trajectory feature vector corresponding to each third sample AIS data in the third ship navigation dataset. All trajectory feature vectors with added navigation scene labels corresponding to the third ship navigation dataset are used as the first training set to train the semantic classification model, thereby enabling the semantic classification model to learn the mapping relationship between trajectory feature vectors and navigation scene labels.

[0113] In other embodiments, the trajectory feature vectors of all third sample AIS data corresponding to the third ship navigation dataset can be divided into multiple training vector subsets according to the feature range corresponding to each navigation scenario label. The trajectory feature vectors of each training vector subset satisfy the feature range of the same type of navigation scenario label. Positive sample pairs are constructed using the trajectory feature vectors of the same training vector subset, and negative sample pairs are constructed using the trajectory feature vectors of different training vector subsets. The semantic classification model is trained using the positive sample pairs and negative sample pairs.

[0114] Step S3: Perform semantic temporal prediction based on the temporal sequence of the scene labels to obtain the predicted scene labels.

[0115] In this embodiment, based on the scene label time sequence, an LSTM neural network model is used to perform semantic time sequence prediction to obtain the predicted scene label, such as... Figure 5 As shown, step S3 specifically includes steps S31-S33.

[0116] Step S31: Obtain the predicted duration, and extract historical tag time series data from the scene tag time series based on the predicted duration.

[0117] In this embodiment, based on the prediction duration requirement, historical tag time series data with a time length that matches the scene tag time series are extracted and used as input to the LSTM neural network model for semantic time series prediction.

[0118] First, the prediction duration is obtained, and based on this, the length of the historical label time series data to be extracted is determined. The prediction duration can be set as needed. The LSTM neural network model needs a sufficiently long length of historical label time series data to learn the switching patterns of trajectory scene labels at different times of the ship, as well as temporal correlation features (such as the average duration of "anchorage waiting" and the dependency relationship between adjacent trajectory scene labels). The length of the historical label time series data directly determines the prediction accuracy. If the historical label time series data is too short, the model cannot capture the complete label switching patterns, easily leading to prediction bias; if it is too long, it may introduce redundant noise, increase computational costs, and reduce short-term prediction accuracy. Therefore, there is a strong correlation between the length of the historical data input to the model and the prediction duration, and the two must be matched appropriately. Furthermore, the integrity of periodic coverage is verified for a time length determined according to a preset ratio, adapting to actual ship navigation scenarios. This is combined with the typical duration of labels for different trajectory scenarios (e.g., short duration for "emergency avoidance" and long duration for "ocean cruising") to ensure that historical data contains samples consistent with the label types and switching patterns that may appear within the predicted duration. This achieves effective transfer from "historical patterns to future predictions," guaranteeing the accuracy and practicality of the prediction results. For example, the prediction duration is... According to the preset ratio (For example, 2-3), calculate the time length of historical tag time series data. The switching integrity is verified for all trajectory scene labels corresponding to the time length. For example, it is determined whether each type of navigation scene label appears at least once. If so, it is considered to meet the switching integrity. Otherwise, the preset ratio is increased and the time length is recalculated until the requirement that each type of navigation scene label appears at least once is met. The final time length is then determined as the time length of the historical label time series data to be extracted.

[0119] Secondly, based on the final determined time length, historical tag time-series data of the corresponding time length is extracted from the scene tag time-series sequence. In this embodiment, the last time step of the scene tag time-series sequence is kept consistent with that of the historical tag time-series data before the historical tag time-series data is extracted. For example, if the scene tag time-series sequence includes trajectory scene tags from 0:00 on January 1st to 24:00 on March 1st, and the required time length is 24 hours, then the trajectory scene tags from 0:00 on March 1st to 24:00 on March 1st, based on all extracted trajectory scene tags and their corresponding time steps, constitute the historical tag time-series data.

[0120] It should be noted that when extracting historical label time-series data from the scene label time-series sequence, the extraction is performed according to time steps and input into the LSTM neural network model. For example, the scene label time-series sequence includes time steps and merged time intervals, such as "08:00-11:00: Port operations; 11:00-12:00: Offshore fishing". The extracted historical label time-series data is then formed into time-series data according to time steps, such as "08:00-09:00: Port operations; 09:00-10:00: Port operations; 10:00-11:00: Port operations; 11:00-12:00: Offshore fishing". This allows the LSTM neural network model to identify the time-series changes of trajectory scene labels, such as changes in time steps while the trajectory scene labels remain unchanged, or changes in trajectory scene labels when time steps change. This accurately captures the time boundaries of trajectory scene label changes, enabling real-time monitoring and prediction, and avoiding a decrease in time-series accuracy due to time step merging.

[0121] Step S32: Based on the historical tag time series data, use the LSTM neural network model to obtain the initial predicted probability of each navigation scenario tag.

[0122] Historical tag time-series data is input into an LSTM neural network model to obtain the initial prediction probability for each navigation scenario tag. In this embodiment, the LSTM neural network model predicts the probability of each navigation scenario tag appearing in the predicted time period based on the input historical tag time-series data, i.e., the initial prediction probability. The initial prediction probabilities of all navigation scenario tags can be output as a prediction vector, and the dimension of the prediction vector is consistent with the number of navigation scenario tag types.

[0123] Furthermore, if the prediction duration includes multiple time steps, the LSTM neural network model is used to output the initial prediction probability of each navigation scenario label corresponding to each time step within the prediction duration based on historical label time series data. In this embodiment, the initial prediction probability of multiple prediction time steps is obtained directly based on historical label time series data. In other embodiments, the LSTM neural network model can also be used to predict time steps one by one, and the previously obtained prediction results and historical label time series data are used together as the input of the LSTM neural network model to predict the next prediction time step.

[0124] Step S33: Based on the initial probability of prediction for each of the navigation scene labels, determine the predicted scene label corresponding to the predicted duration.

[0125] Based on the initial predicted probability of each navigation scenario label, the predicted scenario label corresponding to the predicted duration is determined. In this embodiment, the navigation scenario label with the highest initial predicted probability is selected as the predicted scenario label corresponding to the predicted duration.

[0126] If the prediction duration includes multiple time steps, then for each time step within the prediction duration, the initial probability of the prediction is determined for each of the multiple navigation scene labels corresponding to each time step. The navigation scene label with the highest initial probability is selected as the prediction scene label for that time step, thereby determining the prediction scene labels for all time steps within the prediction duration.

[0127] Furthermore, based on the time length determined in step S31, historical vector time series data of the corresponding time length can be extracted from all trajectory feature vectors of the target vessel. The historical vector time series data and historical label time series data are input into the LSTM neural network model to obtain the predicted feature vector and the initial predicted probability of each navigation scenario label. The corresponding predicted scenario label is determined with reference to step S33, that is, the predicted trajectory feature vector (predicted feature vector) and the predicted navigation scenario label (predicted scenario label) are obtained simultaneously.

[0128] In this embodiment, the LSTM neural network model is trained. For example, based on the first ship navigation dataset and / or the second ship navigation dataset, the trajectory is divided by the ship's MMSI code. The navigation scene labels corresponding to all sample AIS data of each ship are arranged in chronological order based on temporal continuity to form one or more sample label sequences for that ship. All sample label sequences of all ships constitute the LSTM training set, which is used to train the LSTM neural network model. This urges the LSTM neural network model to predict the initial probability of each future navigation scene label based on the input label sequence, thereby determining the future navigation scene label. Furthermore, referring to step S1, trajectory windowing can be performed on all sample AIS data for each ship to determine the sample feature vector at each time step. Based on all navigation scene labels at each time step, the sample scene label for that time step is determined according to a voting mechanism. Data pairs are formed according to time step-sample feature vector-sample scene label, arranged chronologically based on temporal continuity, constituting one or more sample feature-label sequences for that ship. All sample feature-label sequences for all ships constitute the LSTM training set, used to train the LSTM neural network model. This prompts the LSTM neural network model to predict future trajectory feature vectors (i.e., predicted feature vectors) and future navigation scene labels (i.e., predicted scene labels) based on the input sample feature vector sequence and sample scene label sequence. The training method can adopt existing model training methods, and no restrictions are imposed here.

[0129] Furthermore, to improve prediction accuracy, when determining the prediction scenario label corresponding to the prediction duration based on the initial prediction probability of each navigation scenario label, the initial prediction probability is also corrected by combining ship type (such as cargo ship, fishing boat, cruise ship) and seasonal factors (such as peak port season, fishing season). The prediction scenario label corresponding to the prediction duration is determined based on the corrected prediction probability, specifically including steps S331-S335.

[0130] Step S331: Based on the ship AIS dataset, obtain the corresponding target ship type and target season label.

[0131] First, based on the target vessel's AIS dataset, obtain the target vessel's type, i.e., the target vessel type (e.g., cargo ship, fishing boat, cruise ship, passenger ship).

[0132] Secondly, the current month is determined based on the positioning time, and the corresponding seasonal tag is then used as the target seasonal tag for the target vessel. In this embodiment, based on the typical scenarios in the shipping industry, specific time periods corresponding to different months are used as seasonal tags, such as peak port season (July-August, November-December), fishing season (April-May, September-October), and normal season (January-March, June). The target seasonal tag is determined based on the specific time period corresponding to the current month. In other embodiments, the target seasonal tag can also be determined based on the four seasons, such as spring, summer, autumn, and winter, according to the current month.

[0133] Step S332: Based on the target ship type, determine the target type label weight for each navigation scenario label according to the ship type-semantic label weight mapping relationship.

[0134] Navigation scenarios for different ship types exhibit significant differences, and the occurrence probability and switching patterns of each navigation scenario label show a fixed preference. This embodiment, based on the target ship type, determines the type label weight of each navigation scenario label within the prediction duration of the target ship according to the ship type-semantic label weight mapping relationship; this weight is the target type label weight. The ship type-semantic label weight mapping relationship is pre-constructed and includes the correspondence between ship types and the type label weight of each navigation scenario label.

[0135] This embodiment constructs a ship type-semantic label weight mapping relationship based on a first ship navigation dataset and / or a second ship navigation dataset and / or a third ship navigation dataset. First, the first ship navigation dataset and / or the second ship navigation dataset and / or the third ship navigation dataset are classified according to ship type (cargo ship, fishing boat, cruise ship, passenger ship, etc.), forming multiple single-class ship data subsets. Second, for each ship type's corresponding single-class ship data subset, the duration percentage of each navigation scene label is calculated, and this duration percentage is used as the type label weight for that navigation scene label. In this embodiment, the total type label weight is 1; a higher weight indicates a higher probability that the ship type will appear under that navigation scene label. Finally, a ship type-semantic label weight mapping relationship is formed according to the ship type-navigation scene label-type label weight. For example, cargo ships: port operations (0.4), ocean patrol (0.35), anchorage waiting (0.2), emergency avoidance (0.03), near-shore fishing (0.02); 2. fishing vessels: near-shore fishing (0.55), anchorage waiting (0.25), port operations (0.15), emergency avoidance (0.03), ocean patrol (0.02); 3. cruise ships: port operations (0.35), ocean patrol (0.4), anchorage waiting (0.2), emergency avoidance (0.03), near-shore fishing (0.02).

[0136] Step S333: Based on the target seasonal label, determine the target seasonal label coefficient for each of the navigation scenario labels according to the seasonal-semantic label weight coefficient relationship.

[0137] Seasonal changes directly affect the navigation needs of ships, thereby altering the probability of occurrence of each navigation scenario label. This embodiment, based on the target seasonal label of the target ship, determines the seasonal label coefficient of each navigation scenario label within the prediction timeframe of the target ship according to the seasonal-semantic label coefficient mapping relationship; this coefficient is the target seasonal label coefficient. The seasonal-semantic label coefficient mapping relationship is pre-constructed and includes the correspondence between seasonal labels and the seasonal label coefficients of each navigation scenario label.

[0138] This embodiment constructs a seasonal-semantic label coefficient mapping relationship based on a first ship navigation dataset and / or a second ship navigation dataset and / or a third ship navigation dataset. First, the first and / or second and / or third ship navigation datasets are classified according to seasonal labels, forming multiple single-season ship data subsets. Second, for each seasonal label's corresponding single-season ship data subset, the duration percentage of each navigation scenario label is calculated, and this duration percentage is used as the seasonal label coefficient for that navigation scenario label. A higher coefficient indicates a higher probability of that navigation scenario label appearing during that seasonal label. Finally, a seasonal-semantic label coefficient mapping relationship is formed according to the seasonal label - navigation scenario label - seasonal label coefficient. In other embodiments, the seasonal label coefficient is the ratio of the duration of each navigation scenario label in the normal period to the duration of the peak season / fishing season to the corresponding normal period. For example, in the peak season: port operations (1.2), anchorage waiting (1.1), ocean cruise (0.9), near-shore fishing (0.85), emergency avoidance (1.0); in the fishing season: near-shore fishing (1.25), port operations (0.9), anchorage waiting (1.05), ocean cruise (0.85), emergency avoidance (1.0); in the normal period: all are 1.0 (no adjustment).

[0139] Step S334: Based on the target type label weight and the target season label coefficient, the predicted initial probability of each navigation scenario label is corrected to obtain the corresponding predicted corrected probability.

[0140] Based on the target type label weight and / or the target season label coefficient, the initial prediction probability of each navigation scenario label within the target ship's prediction duration is corrected to obtain the corresponding corrected prediction probability. In this embodiment, following the logic of "ship type priority, supplemented by seasonal factors," the initial prediction probability is first corrected based on the target type label weight. For example... In the formula, For the first The initial predicted probability after the first correction for each navigation scene label For the first The target type label weight of each navigation scenario label For the first The initial predicted probability of each navigation scenario label before correction is calculated; then, based on the target season label coefficient, the initial predicted probability after the first correction is corrected a second time. For example, In the formula, For the first The second revised initial probability prediction for each navigation scenario label For the first The target seasonal label coefficients for each navigation scenario label are obtained through two corrections, resulting in a more accurate prediction of the corrected probability for the actual navigation scenario.

[0141] Furthermore, a reasonableness judgment is made on all predicted correction probabilities to avoid prediction deviations where the predicted correction probabilities do not match the ship type. For example, for fishing vessels, if the probability of "ocean-going cruise" is the highest among all predicted correction probabilities, which does not match the characteristics of fishing vessels fishing in coastal waters and operating in ports, then return to step S1 or step S2 to obtain the predicted correction probabilities again.

[0142] Furthermore, if step S331 fails to obtain the ship type, the target ship type is set to empty. In step S332, based on the empty target ship type, the target type label weight for each navigation scenario label is determined to be the default type weight (e.g., 0.2). The initial prediction probability of the navigation scenario label is corrected based on the default type weight and the target seasonal label coefficient to obtain the corresponding prediction correction probability. If step S331 results in unmatched seasonal information (e.g., ambiguous definition of special time periods) or no target seasonal label is obtained, the initial prediction probability of the navigation scenario label is corrected only based on the target type label weight to obtain the corresponding prediction correction probability.

[0143] Understandably, when determining the trajectory scene label corresponding to the trajectory feature vector based on the label probability vector in step S214, steps S331-S335 can be referenced to obtain the target type label weight and target season label coefficient for each navigation scene label at each time step. The label probability vector is then modified element-by-element to obtain the corresponding modified label probability vector, thereby determining the corresponding trajectory scene label and improving the accuracy and generalization ability of semantic annotation. Furthermore, a separate semantic classification model and semantic-communication model can be trained for each ship type to improve the model's adaptability to ship types.

[0144] Step S335: Determine the predicted scene label corresponding to the predicted duration based on the predicted correction probability.

[0145] Based on the predicted correction probability of each navigation scenario label, the predicted scenario label corresponding to the predicted duration is determined. In this embodiment, the navigation scenario label with the highest predicted correction probability is selected as the predicted scenario label corresponding to the predicted duration.

[0146] In other embodiments, if the predicted correction probabilities of multiple navigation scene labels are close (e.g., the difference is less than 0.05), multiple navigation scene labels are retained as predicted scene labels. In step S40, accuracy verification is performed, and the predicted correction probability is selected from several predicted scene labels that have passed the accuracy verification as the final predicted scene label, thereby improving the effectiveness of the predicted scene labels.

[0147] Furthermore, if the prediction duration includes multiple time steps, the initial prediction probability is corrected for each time step. For the multiple navigation scene labels corresponding to each time step, the navigation scene label with the highest prediction correction probability is selected as the prediction scene label for that time step, thereby determining the prediction scene labels for all time steps within the prediction duration.

[0148] Step S4: Based on the time sequence of the scene label and the predicted scene label, use the semantic-communication model to determine the predicted communication behavior corresponding to the target ship.

[0149] Based on the target ship's AIS dataset, semantic deep understanding of AIS data is performed to obtain the semantic temporal pattern of historical AIS trajectory (i.e., the temporal sequence of scene labels), thereby obtaining the predicted scene labels. Then, the semantic-communication model is used to determine the predicted communication behavior corresponding to the target ship.

[0150] The semantic-communication model can employ a reinforcement learning model, using the temporal changes of ship semantic labels (temporal sequence of scene labels and predicted scene labels) as environmental state input. By utilizing the semantic-communication model, the predicted communication behavior of the target ship can be determined, and the ship's communication needs can be predicted in advance.

[0151] The semantic-communication model (reinforcement learning type) is obtained through training. The first ship navigation dataset and / or the second ship navigation dataset and / or the third ship navigation dataset also include ship communication data. Taking the first ship navigation dataset as an example, a sample label time sequence is constructed using the navigation scene labels of all first sample AIS data, and a sample communication behavior sequence (including communication service type and communication parameter range) is constructed using the communication behavior of ship communication data. The sample label time sequence and the sample communication behavior sequence are time-aligned to form a reinforcement learning training set. The sample label time sequence is used as the environmental state input, and the sample communication behavior sequence is used as the label. The reward function is "communication adaptation accuracy + transmission efficiency" (the higher the adaptation accuracy and the better the transmission efficiency, the higher the reward value). Through multiple rounds of iterative training, the reinforcement learning model continuously learns the adaptation rules of "navigation scene label - communication behavior" to achieve automatic prediction of communication behavior.

[0152] By combining semantic annotation, temporal prediction, and semantic-communication model, the future communication behavior of ships can be predicted in advance (i.e., predictive communication behavior), realizing end-to-end processing from semantic prediction to communication demand prediction. This can be applied to satellite communication resource scheduling simulation, providing forward-looking data support for simulation and significantly improving the guiding value of simulation results for system design. For example, the simulation prediction latency is ≤100ms, and the accuracy of predicting communication demand in the next 24 hours is ≥88%.

[0153] Furthermore, to improve the interpretability of predicted communication behavior, the semantic-communication model can also adopt a dynamic mapping model of "trajectory semantic-communication behavior" based on association rule mining, which does not require complex deep learning training. This model includes a semantic-communication mapping table, based on the time sequence of scene labels and predicted scene labels, and uses the semantic-communication model to determine the predicted communication behavior corresponding to the target ship, specifically including steps S41-S43.

[0154] Step S41: Perform time series analysis on the scene label time series and the predicted scene label to determine the currently effective label and determine whether the predicted scene label is consistent with the currently effective label.

[0155] A time series analysis is performed on the scene label time series sequence and the predicted scene label to determine the currently effective label. In this embodiment, if the prediction time is a single time step, the changes in the trajectory scene label as time progresses are obtained based on the scene label time series sequence. The trajectory scene label corresponding to the last time step in the scene label time series sequence is extracted as the currently effective label, and it is determined whether the predicted scene label and the currently effective label are consistent.

[0156] Furthermore, if the prediction duration includes multiple time steps, the judgment is performed step by step. For the predicted scene label of the first time step within the prediction duration, the trajectory scene label corresponding to the last time step in the scene label time sequence is used as the currently effective label. The predicted scene label of the first time step within the prediction duration is compared with the currently effective label to determine the subsequent prediction communication behavior. For the predicted scene labels of other time steps within the prediction duration, the predicted scene label of the previous time step within the prediction duration is used as the currently effective label for judgment.

[0157] Step S42: If they match, obtain the current communication behavior corresponding to the currently effective tag, and use the current communication behavior as the predicted communication behavior; wherein, the predicted communication behavior includes the communication service type and the communication parameter range.

[0158] If the predicted scenario label matches the currently active label, then the current communication behavior corresponding to the currently active label is obtained, and this current communication behavior is used as the predicted communication behavior. The predicted communication behavior includes the communication service type and the communication parameter range. In this embodiment, the current communication behavior can be obtained by acquiring the target vessel's communication behavior data, including the current communication service type and the communication parameter range. The current communication service type and communication parameter range can then be used as the communication service type and communication parameter range corresponding to the predicted communication behavior.

[0159] Furthermore, to improve the accuracy of predicting communication behavior, the semantic-communication model also includes a feature-semantic mapping table, which uses the current communication behavior as the predicted communication behavior, including steps S421-S424.

[0160] Step S421: Use the communication service type corresponding to the current communication behavior as the communication service type corresponding to the predicted communication behavior.

[0161] The communication service type corresponding to the current communication behavior is used as the communication service type corresponding to the predicted communication behavior, that is, as the communication service type corresponding to the predicted duration.

[0162] Step S422: Obtain the current discrete feature combination corresponding to the currently effective label, and obtain the predicted discrete feature combination corresponding to the predicted scene label.

[0163] Obtain the current discrete feature combination corresponding to the currently effective label, and obtain the predicted discrete feature combination corresponding to the predicted scene label. In this embodiment, based on all trajectory feature vectors, the current discrete feature combination is determined using an equal-width discretization method. For example, based on all trajectory feature vectors, the numerical range of each dimension is determined. Using an equal-width discretization method, the continuous numerical range of each dimension is divided into M intervals of equal width. According to the trajectory feature vector corresponding to the currently effective label, the interval in which the value of each dimension is located is determined. The value of each dimension is converted into the corresponding interval (i.e., discrete feature), forming a discrete feature combination, i.e., the current discrete feature combination. In other embodiments, the numerical range of each static feature and dynamic feature can also be determined based on the first ship navigation dataset and / or the second ship navigation dataset and / or the third ship navigation dataset. The interval is determined according to equal-frequency discretization or cluster discretization, thereby determining the interval position of the trajectory feature vector corresponding to the currently effective label, thus forming the current discrete feature combination. For example, the average speed ranges from 0 to 20 knots, divided into three intervals: 0-6 knots is the first interval (low), 7-13 knots is the second interval (medium), and 14-20 knots is the third interval (high). If the average speed in the trajectory feature vector corresponding to the currently active label is 8, then the corresponding discrete feature is medium.

[0164] To predict discrete feature combinations, firstly, the predicted feature vector corresponding to the predicted scene label is obtained through the LSTM neural network model. Then, referring to the current method for determining discrete feature combinations, the discrete feature combination corresponding to the predicted feature vector is determined, which is the predicted discrete feature combination.

[0165] Step S423: Based on the feature-semantic mapping table, determine whether there is a fluctuation between the current discrete feature combination and the predicted discrete feature combination.

[0166] Based on the feature-semantic mapping table, it is determined whether the volatility of the current discrete feature combination and the predicted discrete feature combination exceeds the volatility threshold. In this embodiment, the feature-semantic mapping table is pre-constructed and includes a rule mapping antecedent and a corresponding rule mapping consequent, wherein the rule mapping antecedent is a discrete feature combination and the rule mapping consequent is a navigation scene label.

[0167] First, based on the currently effective label, find the rule mapping consequent that is the same as the currently effective label from the feature-semantic mapping table, and extract the corresponding rule mapping antecedent. The rule mapping antecedent includes core feature items and non-core feature items. Based on the core feature items and non-core feature items of the rule mapping antecedent, divide the multiple discrete features in the current discrete feature combination and the predicted discrete feature combination into current core feature items, current non-core feature items, current redundant feature items, predicted core feature items, predicted non-core feature items, and predicted redundant feature items.

[0168] For example, the currently effective label is "port operations," and the antecedent of the mapping rule corresponding to "port operations" in the feature-semantic mapping table is { Low, High}, among which, The discrete average speed is the core feature term. This represents the discretized port / anchorage dwelling frequency, which is a non-core feature term. The current discrete feature combination is: Predicting discrete feature combinations ,but (Current discrete average speed) (Predicted discretized average speed) represents the current core feature term and the predicted core feature term, respectively. (Current frequency of discrete port / anchorage stays) (Predicted discretized port / anchorage dwell frequency) represents the current non-core feature term and the predicted non-core feature term, respectively. (Total stay duration) (Current number of turns) (Current speed change rate) (Current steering angle distribution mean) (Current steering angle distribution variance) is the current redundant feature term. (Predicted total stay duration) (Predicted number of turns) (Predicted speed change rate) (Predicted steering angle distribution mean) (Predicted steering angle distribution variance) is a predicted redundancy feature term.

[0169] Secondly, determine whether the current core feature item, the predicted core feature item, and the core feature item in the rule mapping antecedent are consistent. If there is a discrepancy, return to step S1 or step S3 to re-obtain the scene label time sequence or the predicted scene label. If they are all consistent, then determine whether the current non-core feature item and the predicted non-core feature item are consistent. If the current non-core feature item and the predicted non-core feature item are consistent, it is considered that there is no fluctuation between the current discrete feature combination and the predicted discrete feature combination. If the current non-core feature item and the predicted non-core feature item are inconsistent, it is considered that there is a fluctuation between the current discrete feature combination and the predicted discrete feature combination. For example, if , If both are low, then the current core feature term and the predicted core feature term are considered consistent with the core feature term in the rule mapping antecedent. Next, the consistency between the current non-core feature term and the predicted non-core feature term is checked. If... , If both are high, then it is considered that there is no fluctuation between the current discrete feature combination and the predicted discrete feature combination. For high, If the value is in the middle, then fluctuations are considered to exist.

[0170] Step S424: If there is no fluctuation, then the communication parameter range corresponding to the current communication behavior is taken as the communication parameter range corresponding to the predicted communication behavior.

[0171] If there is no fluctuation, the range of communication parameters corresponding to the current communication behavior will be used as the range of communication parameters corresponding to the predicted communication behavior.

[0172] Furthermore, step S42 also includes: Step S425: If fluctuations exist, the range of communication parameters corresponding to the current communication behavior is adjusted and used as the range of communication parameters corresponding to the predicted communication behavior.

[0173] If fluctuations occur, the communication service type will not be changed; only the range of communication parameters will be adjusted. For example, the bandwidth of port scheduling communication will be slightly adjusted from 1.5Mbps to 1.2Mbps to ensure that the communication parameters are adapted to slight changes in the navigation trajectory.

[0174] Furthermore, based on the feature range corresponding to each navigation scenario label determined in step H, the magnitude of volatility is judged. If the standardized feature values ​​corresponding to both the current non-core feature item and the predicted non-core feature item are within the feature range corresponding to the currently effective label, then the volatility is considered to meet the volatility requirements. The communication parameter range corresponding to the current communication behavior is then adjusted and used as the communication parameter range corresponding to the predicted communication behavior. For example, For high, For the middle, Corresponding predictive standardized features If the standardized average speed of port operations is within the characteristic range of [0.6, 1.2], then the volatility is considered to meet the volatility requirement; if the standardized feature value corresponding to the predicted non-core feature item is not within the feature range corresponding to the currently effective label, then the volatility requirement is not met, and the process returns to step S3 to reacquire the prediction scenario label, or it is determined that the prediction scenario label is inconsistent with the currently effective label, and the prediction communication behavior is determined based on step S43.

[0175] In other embodiments, if in step S423 , If both are high, then according to , The current standardized feature and the predicted standardized feature are used to determine whether there is fluctuation. If the difference between the current standardized feature and the predicted standardized feature does not exceed the fluctuation threshold (e.g., 5%), then there is no fluctuation. The communication parameter range corresponding to the predicted communication behavior is determined according to step S424. If the difference is greater than the fluctuation threshold, then there is fluctuation. The communication parameter range corresponding to the predicted communication behavior is determined according to step S425. For high, If the result is negative, return to step S3 to reacquire the predicted scene label, or determine that the predicted scene label is inconsistent with the currently effective label, and determine the predicted communication behavior based on step S43.

[0176] Step S43: If there is no consistency, the predicted communication behavior is determined based on the predicted scene label using the semantic-communication mapping table; wherein the predicted communication behavior includes the communication service type and the communication parameter range.

[0177] If the predicted scenario label is inconsistent with the currently effective label, the predicted communication behavior is determined based on the predicted scenario label using a semantic-communication mapping table. The semantic-communication mapping table includes the correspondence between navigation scenario labels and ship communication behaviors. Based on the predicted scenario label, the semantic-communication mapping table is used to find the navigation scenario label that matches the predicted scenario label, and the corresponding ship communication behavior is extracted; this is the predicted communication behavior. For example, if the predicted scenario label is "ocean cruising," and the ship communication behavior corresponding to "ocean cruising" in the semantic-communication mapping table is "Communication service type: ocean navigation communication; communication parameter range {bandwidth: 0.5-1Mbps, latency: ≤100ms, transmission frequency: 1 time / 1min}", then the predicted communication behavior (i.e., the new communication service type and parameter range) is determined to be "Communication service type: ocean navigation communication; communication parameter range {bandwidth: 0.5-1Mbps, latency: ≤100ms, transmission frequency: 1 time / 1min}", and the switch is completed quickly with a response time of ≤1s, ensuring that the communication service and navigation scenario are synchronously adapted.

[0178] Furthermore, after determining the predicted communication behavior in step S42 or step S43, the method further includes: Step S44: Obtain the expected effective duration of the predicted scene label. If the expected effective duration reaches the label switching threshold, then perform communication adaptation based on the predicted communication behavior.

[0179] The expected duration of the predicted scene labels is obtained. If the expected duration reaches the label switching threshold, communication adaptation is performed based on the predicted communication behavior. By using the label switching threshold, communication adaptation is performed with stable and effective scene labels to avoid frequent communication switching caused by instantaneous label fluctuations. The label switching threshold can be set empirically, for example, 1 minute.

[0180] If the predicted scene label is consistent with the currently effective label, then the effective duration of the currently effective label is obtained, which is the cumulative duration of adjacent time steps that are the same as the currently effective label in the scene label time sequence. The expected effective duration of the predicted scene label is the sum of the duration of a single time step and the effective duration. If the predicted scene label is inconsistent with the currently effective label, then the expected effective duration is the duration of a single time step.

[0181] Furthermore, if the prediction duration includes multiple time steps, the expected effective duration of the current predicted scene label also includes the duration to be effective, which is the cumulative duration of adjacent time steps where the predicted scene labels of all time steps corresponding to the prediction duration are the same as the current predicted scene label.

[0182] By using a semantic-communication model (dynamic mapping type) to achieve time-series awareness and dynamic adaptation, it can automatically identify temporal changes in semantic tags (such as time interval switching and tag type changes) and simultaneously complete the matching and adjustment of communication service types and parameter ranges without manual intervention. It can also output the process of judging consistency and fluctuation along with the predicted communication behavior, improving the traceability and interpretability of the prediction results.

[0183] Furthermore, if any abnormal situation occurs in steps S1-S4, such as the inability to identify trajectory scene labels, a mismatch between the current discrete feature combination and the mapping rule antecedents, or frequent changes in the predicted scene labels (e.g., the expected effective duration is always a single time step), then a redundancy adaptation mechanism is activated. In this embodiment, the redundancy adaptation mechanism includes retrieving general ship communication services (e.g., bandwidth 0.3-0.5Mbps, latency ≤200ms), while simultaneously feeding back abnormal information. It may also include reminding manual troubleshooting to ensure uninterrupted communication services and no adaptation failure.

[0184] By using a semantic-communication model to determine the predictive communication behavior of the target vessel, the system achieves a simple and fast response, adapting to the real-time requirements of the temporal changes in vessel trajectory scene labels. Through temporal change capture and dynamic fine-tuning of communication parameters, it ensures both the accuracy of communication service matching and the flexibility of adaptation.

[0185] Furthermore, to improve the reliability of predicting communication behavior, before performing time-series analysis on the scene tag time-series sequence, step S4 further includes: Step S40: Verify the accuracy of the scene label time series and the predicted scene label.

[0186] Accuracy verification is performed on the scene label time series and the predicted scene labels. Then, time series analysis is conducted on the scene label time series and the predicted scene labels that pass the accuracy verification to determine the predicted communication behavior. For example, such as... Figure 6 The diagram illustrates a process that first verifies the accuracy of the time-series sequence of scene labels and the predicted scene labels, then performs time-series analysis on the verified time-series sequence of scene labels and the predicted scene labels, and finally determines the predicted communication behavior. In this embodiment, the accuracy verification includes a first label accuracy verification and / or a second label accuracy verification, such as... Figure 7 As shown, step S40 includes steps S401-S405.

[0187] Step S401: Obtain the time sequence of the scene label and the discrete feature combination corresponding to the predicted scene label.

[0188] Referring to the method for obtaining the current discrete feature combination in step S422, the trajectory feature vectors of other time steps in the scene label time series are discretized to obtain the discrete feature combination corresponding to each time step. The discrete feature combination predicted in step S422 is the discrete feature combination corresponding to the predicted scene label. If the prediction duration is multiple time steps, the corresponding discrete feature combination is obtained step by step, referring to the method for obtaining the predicted discrete feature combination in step S422.

[0189] Step S402: Based on all the discrete feature combinations, determine the accuracy of the scene label time sequence and the first label of the predicted scene label.

[0190] Based on the discrete feature combinations of all time steps, the accuracy of the first label of the scene label time sequence and the predicted scene label is judged. In this embodiment, the feature-semantic mapping table is used to judge the accuracy of the first label of any trajectory scene label and the predicted scene label one by one. Any discrete feature combination is used as the feature combination to be judged, and the corresponding trajectory scene label or predicted scene label is used as the label to be judged for specific explanation.

[0191] From the feature-semantic mapping table, find the rule mapping consequent that is the same as the label to be judged, extract the corresponding rule mapping antecedent (including core feature items and non-core feature items), and match the feature combination to be judged with the rule mapping antecedent item by item. If the core feature item of the feature combination to be judged is the same as the core feature item of the rule mapping antecedent, then the accuracy judgment of the first label of the label to be judged is considered to have passed. For example, the rule mapping antecedent is { Low, High}, among which, The discrete average speed is the core feature term. This represents the discrete port / anchorage dwelling frequency, which is a non-core feature item. The feature combination to be judged is { Low, Low, Low, Low, Low, Low, High}, where the discretized average speed (Core feature term) is low, and it corresponds to the core feature term in the antecedent of the rule mapping ( If the accuracy of the first label is the same as that of the second label, then the accuracy judgment of the first label to be judged is considered to have passed. When determining the predicted communication behavior, the accuracy judgment of the first label can be output together to enhance the interpretability of trajectory scene labels and predicted scene labels. In other embodiments, non-core feature items (discrete port / anchorage dwelling frequency) and redundant discrete features ( Low, Low, Low, Low, The low value does not affect the accuracy of the first label judgment result.

[0192] Further, if the core feature term of the feature combination to be judged is different from the core feature term of the rule mapping antecedent, then return to step S1 or step S3 to re-obtain the scene label time series sequence or predicted scene label. In other embodiments, if the core feature term of the feature combination to be judged is different from the core feature term of the rule mapping antecedent, then obtain all rule mapping antecedents and corresponding mapping confidences involved in the feature combination to be judged from the feature-semantic mapping table. If the largest mapping confidence is greater than the second confidence threshold (e.g., 80%), then correct the trajectory scene label or predicted scene label corresponding to the feature combination to be judged to the rule mapping consequent corresponding to that confidence, that is, correct the label to be judged. The corrected trajectory scene label or predicted scene label is subjected to the second label accuracy test, or is used as the trajectory scene label or predicted scene label that has passed the accuracy test for time series analysis. Through double verification matching (feature combination verification + mapping confidence verification), the accuracy of the trajectory scene label or predicted scene label is ensured; if there is no mapping confidence greater than the second confidence threshold, then the second label accuracy test is performed.

[0193] Step S403: Obtain the ship communication data of the target ship, synchronize the ship communication data and the ship AIS data in time, and extract the ship communication behavior sequence corresponding to the scene tag time sequence.

[0194] The ship communication data of the target ship is acquired, and the ship communication data and ship AIS data are synchronized in time. According to the same preset time step as the trajectory window, the ship communication behavior corresponding to each time step is determined as the trajectory communication behavior of that time step, and the ship communication behavior sequence corresponding to the scene label time sequence is obtained.

[0195] Step S404: Based on the scene tag time sequence and the ship communication behavior sequence, determine the accuracy of the second tag of the scene tag time sequence according to the semantic-communication mapping table.

[0196] Based on the scene label time sequence and the ship communication behavior sequence, the accuracy of the second label of the scene label time sequence is judged according to the semantic-communication mapping table. In this embodiment, the accuracy of the second label includes basic matching and continuity test.

[0197] First, based on the scene label time sequence and the ship communication behavior sequence, a basic matching is performed step by step according to the semantic-communication mapping table. For each time step, the corresponding trajectory scene label and trajectory communication behavior are extracted from the scene label time sequence and the ship communication behavior sequence. Based on the trajectory scene label of that time step, the navigation scene label that is the same as the trajectory scene label of that time step is searched in the semantic-communication mapping table, and the corresponding ship communication behavior is extracted. It is determined whether the trajectory communication behavior and the ship communication behavior are the same. If they are the same, the basic matching is considered to be successful. For example, if the trajectory scene label to be judged is "ocean cruise", the ship communication type in the corresponding ship communication behavior in the semantic-communication mapping table is ocean navigation communication, the communication parameter range is {bandwidth: 0.5-1Mbps, latency: ≤100ms, transmission frequency: 1 time / 1min}, and the current actual communication service (i.e., trajectory communication behavior) is ocean cruise, and the bandwidth and latency are also within the preset parameter range, then the basic matching is considered to be successful.

[0198] Secondly, the continuity of the scene label time sequence and ship communication behavior sequence that have passed the basic matching is checked. Combined with the common scenarios in the ship navigation industry, the trajectory scene label and trajectory communication behavior at each time step are judged to see whether they are continuous and reasonable with the trajectory scene label and trajectory communication behavior at the adjacent time steps. This ensures that the scene label time sequence and ship communication behavior sequence fit the continuity of ship navigation and avoids logical defects in the scene label time sequence. For example, if the trajectory scenario label for the time step to be judged is "port operation" and the trajectory communication behavior is "port dispatch communication", and the trajectory scenario label for adjacent time steps is "port operation" and the corresponding trajectory communication behavior is "port dispatch communication" (consistent with the current trajectory communication behavior to be judged), then the time step to be judged is considered to have "good communication service adaptation" and the continuity check passes. If the trajectory scenario label for adjacent time steps is "port operation" (corresponding to port dispatch communication) and "anchorage waiting" (corresponding to anchorage watch communication), combined with the ship navigation logic in the conventional scenario of the shipping industry (anchorage waiting may occur after port operation, and it is reasonable for the communication service to switch from dispatch to watch), then the time step to be judged is also considered to have "good communication service adaptation" and the continuity check passes. If the trajectory scenario label for adjacent time steps is "ocean cruise" (corresponding to ocean navigation communication), and there is a short-term alternation of ocean cruise-port operation-ocean cruise in one time step without ship navigation logic continuity, then the time step to be judged is considered to have "poor communication service adaptation" and the continuity check fails.

[0199] If "basic matching fails" or "communication service compatibility is poor", return to step S1 to reacquire the scene label time sequence, or combine more adjacent time step data for further judgment to avoid misjudgment.

[0200] The accuracy of the second label is improved by using the semantic-communication mapping table to eliminate misjudgments that contradict the trajectory scene label and communication service, thereby improving the accuracy of the judgment and conforming to the actual scenario logic of matching ship navigation and communication, thus further improving the accuracy of trajectory scene label.

[0201] It should be noted that in this embodiment, the accuracy of the first tag and the accuracy of the second tag are judged step by step to accurately capture time-series changes and adapt to communication prediction needs. This provides more accurate time-series data with consistent time granularity for subsequent communication behavior prediction, ensuring that the tag status of each time step is monitored in real time and that the corresponding communication service is matched for each time step.

[0202] Step S405: If the accuracy of the first label and / or the accuracy of the second label fail, then reacquire the scene label time sequence and / or the predicted scene label.

[0203] If the accuracy of the first label and / or the accuracy of the second label fail, return to step S1 or step S3 to reacquire the scene label time series or predict the scene label.

[0204] By judging the accuracy of the first label and / or the second label of the predicted scene label time sequence, the accuracy of AIS data interpretation and the accuracy of the predicted scene label are improved, thereby improving the reliability of the predicted communication behavior.

[0205] In other embodiments, before step 31, the scene label time sequence can be judged for first label accuracy and / or second label accuracy, and semantic time sequence prediction can be performed using the scene label time sequence (or the corrected scene label time sequence) judged by first label accuracy and second label accuracy to obtain the predicted scene label.

[0206] In this embodiment, a feature-semantic mapping table is constructed through steps J-N to clarify the navigation scene labels corresponding to different AIS data features (i.e., combinations of discrete navigation features).

[0207] Step J: Obtain the fourth ship navigation dataset; wherein, the fourth ship navigation dataset includes multiple fourth sample AIS data and corresponding actual scene labels.

[0208] In this embodiment, the fourth ship navigation dataset includes multiple fourth sample AIS data and corresponding actual scene labels. The AIS data for each ship at each sampling time point is considered the fourth sample AIS data. The actual scene label corresponding to the fourth sample AIS data refers to the navigation scene label corresponding to the fourth sample AIS data. This label can be selected by experts based on the second sample AIS data and experience from a candidate set of navigation scene labels, or it can be determined through ship communication behavior. The fourth ship navigation dataset can differ from the first, second, and third ship navigation datasets. It can also be constructed based on the feature vector clusters after clustering the first and / or second ship navigation datasets and their corresponding navigation scene labels, or the first training set corresponding to the third ship navigation dataset. The fourth ship navigation dataset is divided into trajectories using the ship's MMSI code, forming single-ship AIS datasets of different ship navigation trajectories or trajectory segments.

[0209] Step K: Preprocess the fourth ship navigation dataset to obtain the corresponding trajectory feature vector.

[0210] Referring to step S12, the AIS dataset corresponding to each navigation trajectory or trajectory segment of the fourth ship navigation dataset is preprocessed to obtain the corresponding trajectory feature vector.

[0211] Step L: Time-align the trajectory feature vectors and actual scene labels corresponding to the fourth ship navigation dataset to form a feature-semantic association dataset.

[0212] For the trajectory feature vectors and actual scene labels corresponding to the fourth ship navigation dataset, time alignment is performed based on time steps to form a feature-semantic association dataset. For example, the... The data format for each time step is as follows: ,like In this embodiment, if the same time step includes multiple different actual scene labels, the preset time step length is adjusted, or a voting mechanism is adopted to use the actual scene label with the highest number of votes as the actual scene label for that time step.

[0213] Step M: Perform frequent itemset mining on the feature-semantic association dataset to obtain high-order frequent itemsets.

[0214] This embodiment employs the Apriori algorithm, a classic association rule mining algorithm, to mine frequent itemsets in a feature-semantic association dataset, obtaining high-order frequent itemsets. The Apriori algorithm is primarily used to discover frequent itemsets and their association rules in a dataset. It uses a layer-by-layer search approach, utilizing support and confidence to filter out meaningful rules. Core concepts include support (the frequency of an itemset in the dataset), confidence (the probability of another itemset occurring given the occurrence of a given itemset), and frequent itemsets are defined as itemsets with support no less than the minimum support (itemsets are subsets of "discrete feature combinations + actual scene labels").

[0215] First, the feature-semantic association dataset is discretized to form a discretized association dataset.

[0216] Since the Apriori algorithm is suitable for discrete data, the trajectory feature vectors need to be discretized first. Using the same discretization method as in step S422, the continuous values ​​of each dimension of the trajectory feature vector are divided into M discrete intervals. Each dimension of the trajectory feature vector at each time step in the feature-semantic association dataset is converted into a corresponding discrete feature. The discrete features of all dimensions constitute a discrete feature combination, which, together with the corresponding real-scene label, forms a discretized association dataset. After discretization, the trajectory feature vector at each time step is converted into a discrete feature combination, while the real-scene label remains unchanged. For example, each dimension of the trajectory feature vector is divided into 3 discrete intervals, corresponding to three levels: "low," "medium," and "high." The trajectory feature vector and the real-scene label at a certain time step in the feature-semantic association dataset are... The converted discrete feature combination and the actual scene label are .

[0217] Secondly, based on the discretized association dataset, two core parameters of the Apriori algorithm are set: minimum support (min_support) and minimum confidence (min_confidence). For example, minimum support is set to 5%, meaning that the frequency of a particular itemset in the entire dataset should not be less than 5%, used to remove invalid itemsets with excessively low frequency. Minimum confidence is set to 70%, meaning that the confidence level of a particular itemset should not be less than 70%, ensuring the effectiveness and practicality of the rules and removing itemsets with low confidence. In other embodiments, these parameters can be flexibly adjusted based on feedback from the actual dataset.

[0218] Next, 1-itemset mining is performed on the discretized associated dataset to form 1-frequent itemsets.

[0219] Iterate through all discrete features and real-scene labels in the discretized association dataset, calculate the frequency (support) of each single item combination of discrete feature and real-scene label in the discretized association dataset, and select the single item combinations with the minimum support to form 1-frequent itemsets. For example... , .

[0220] Finally, high-order frequent itemsets are mined based on all 1-frequent itemsets to obtain high-order frequent itemsets.

[0221] Based on all 1-frequent itemsets, a "self-join + pruning" strategy is used to progressively mine 2-itemsets, 3-itemsets...k-itemsets until no higher-order frequent itemsets can be mined. Specifically, 1-frequent itemsets are paired to generate candidate 2-itemsets, and a self-join is performed. The support of each candidate 2-itemset is calculated, and the candidate itemset with the minimum support is selected to obtain the 2-frequent itemset, completing pruning. The above process is repeated, self-joining 2-frequent itemsets with 1-frequent itemsets to generate candidate 3-itemsets, which are then pruned to obtain the 3-frequent itemsets, until all candidate itemsets of a certain order have been pruned, at which point mining stops. The final higher-order frequent itemsets are obtained. For example, , .

[0222] Step N: Extract mapping rules from high-order frequent itemsets to obtain standardized semantic mapping rules and construct a feature-semantic mapping table.

[0223] Based on all the high-order frequent itemsets obtained from mining, an initial semantic mapping rule is constructed in the form of "discrete combination of features → navigation scene label" (i.e., the antecedent is a discrete combination of features and the consequent is a navigation scene label). All the extracted initial semantic mapping rules are filtered and standardized by mapping confidence to obtain standardized semantic mapping rules, which constitute a feature-semantic mapping table.

[0224] First, based on all high-order frequent itemsets, multiple initial semantic mapping rules are constructed; where the antecedent of the initial semantic mapping rule is a combination of discrete features and the consequent is the corresponding navigation scene label.

[0225] Based on all the high-order frequent itemsets obtained from mining, multiple initial semantic mapping rules are constructed. Specifically, based on each high-order frequent itemset, an initial semantic mapping rule is constructed. The discrete feature combination formed by all discrete features in the high-order frequent itemset is used as the antecedent of the initial semantic mapping rule, and the actual scene label (i.e., the navigation scene label) in it is used as the consequent of the initial semantic mapping rule, thus forming an initial semantic mapping rule.

[0226] Secondly, the mapping confidence of each initial semantic mapping rule is calculated, and the initial semantic mapping rules whose mapping confidence reaches the third confidence threshold are selected as candidate semantic mapping rules.

[0227] In this embodiment, the mapping confidence of each initial semantic mapping rule is expressed by the formula. Calculate, where, Represents the initial semantic mapping rules The mapping confidence, Indicates the antecedent (discrete combination of features). Indicates the consequent (navigation scenario label). express and Simultaneous support (i.e., the frequency of the initial semantic mapping rule appearing in all initial semantic mapping rules). express The support of a single occurrence (i.e., the frequency of the antecedent in all initial semantic mapping rules) is used to select initial semantic mapping rules whose mapping confidence reaches the third confidence threshold (e.g., 70%) as candidate semantic mapping rules, and to eliminate initial semantic mapping rules with low mapping confidence or insufficient credibility.

[0228] Furthermore, to improve the simplicity of the feature-semantic mapping table, redundant rules are removed from all candidate semantic mapping rules. Specifically, all candidate semantic mapping rules are compared pairwise. If two candidate semantic mapping rules have completely identical antecedents but different consequents, the candidate semantic mapping rule with lower mapping confidence is removed, and the candidate semantic mapping rule with higher mapping confidence is retained. If two candidate semantic mapping rules have the same consequent, and the antecedent of one candidate semantic mapping rule is a subset of the antecedent of the other candidate semantic mapping rule, and the difference in mapping confidence between the two candidate semantic mapping rules does not exceed the redundancy threshold (e.g., 5%), then the candidate semantic mapping rule with the simpler antecedent is retained, and the other is removed to avoid rule redundancy. For example, there are two candidate semantic mapping rules. The mapping confidence level is 85.5%. If the mapping confidence level is 87.5%, then retain. .

[0229] Finally, the candidate semantic mapping rules are standardized to obtain standardized semantic mapping rules, and all standardized semantic mapping rules constitute a feature-semantic mapping table.

[0230] Candidate semantic mapping rules are standardized by defining the antecedent (discrete feature combination), consequent (semantic label), support, and mapping confidence for each rule, thus forming standardized semantic mapping rules. All standardized semantic mapping rules constitute a feature-semantic mapping table. Example: Rule 1: Support 6.8%, mapping confidence 85.5%; Rule 2: Support 5.7%, mapping confidence 75.1%; Rule 3: The support is 5.2%, and the mapping confidence is 82.5%. Furthermore, based on the support of low-order frequent itemsets, the antecedent (discrete feature combination) can be divided into core feature items and non-core feature items. For example, for rule 1, two 1-frequent itemsets... The support rate is 5. If the support is 1, then it is determined. For core feature items, These are non-core feature terms.

[0231] The Apriori algorithm is used to extract mapping rules from associated datasets, uncovering potential associations between trajectory feature vectors and navigation scene labels, forming a feature-semantic mapping table. This table can be applied to assist in optimizing semantic classification models. When the model outputs labels that are inconsistent with the labels derived from the mapping rules, the mapping rule with a mapping confidence of ≥80% can be selected to correct the model labeling error. It can also be used as a basis for the interpretability of trajectory scene labels and predicted scene labels. When outputting trajectory scene labels and predicted scene labels, the corresponding mapping rules are output simultaneously to explain the rationality of the trajectory scene labels and predicted scene labels.

[0232] Furthermore, in this embodiment, a semantic-communication mapping table is constructed through steps O-P to clarify the communication requirements under different semantic scenarios.

[0233] Step O: Construct a semantic-communication association dataset.

[0234] Obtain the fifth ship navigation dataset, which includes multiple fifth-sample AIS data sets, as well as the corresponding real-world scenario labels and ship communication behavior data for each fifth-sample AIS data set. The ship communication behavior data includes communication service type and communication parameter range. Based on the fifth ship navigation dataset, establish a semantic-communication association dataset that includes real-world scenario labels, communication service types, and communication parameter ranges (communication mapping rules).

[0235] Step P: Extract the communication mapping rules between navigation scenario labels and communication service types and communication parameter ranges from the semantic-communication association dataset to form a semantic-communication mapping table.

[0236] Referring to steps M-N, the Apriori algorithm is used to extract the communication mapping rules between navigation scene labels and communication service types and communication parameter ranges from the semantic-communication association dataset, forming a semantic-communication mapping table. For example, "port operation" corresponds to the communication service type "cargo scheduling service" and the communication parameter range "bandwidth requirement ≥ 50Mbps, latency threshold ≤ 300ms"; "ocean cruise" corresponds to "equipment monitoring service" and "bandwidth requirement ≤ 5Mbps, latency threshold ≤ 5s".

[0237] In other embodiments, the semantic-communication mapping table can also be constructed and updated via step R and / or step S.

[0238] Step R: Based on common knowledge in the ship communication industry, construct communication mapping rules for navigation scenario tags and communication service behaviors to form a semantic-communication mapping table.

[0239] By collecting publicly known information from the ship communication industry, extracting rules, and standardizing formats, a communication mapping rule is constructed that links navigation scenario labels to communication service behaviors (including communication service types and communication parameter ranges). The antecedent is the navigation scenario label, and the consequent is the communication service behavior (including communication service types and communication parameter ranges), forming a semantic-communication mapping table. The communication service types required for different ship navigation scenarios (port operations, ocean cruises, etc. corresponding to navigation scenario labels) are established standards in the fields of maritime communication and ship satellite communication. For example, the communication service types for port operations are necessarily port scheduling communications such as cargo scheduling and berth application, while the communication service types for ocean cruises are necessarily ocean navigation communications such as ship equipment monitoring and navigation positioning. The parameter ranges such as bandwidth, latency, and transmission frequency corresponding to different communication service types all follow the technical specifications of the ship satellite communication industry, maritime communication standards (such as relevant ITU-R maritime communication recommendations), and the performance thresholds of equipment hardware. For example, the bandwidth requirement for high-bandwidth cargo scheduling data transmission is ≥50Mbps, and the latency threshold for low-latency emergency communication is ≤300ms.

[0240] Step S: Update the semantic-communication mapping table based on known updates and / or practical data from the ship communication industry.

[0241] Based on publicly known updates in the ship communication industry, such as revisions to ship communication industry standards and maritime communication standards, or changes in actual ship communication operational needs, the communication mapping rules for navigation scenario tags and communication service behaviors are updated to achieve semantic-communication mapping table updates. For example, when organizations such as the International Maritime Organization (IMO) and the International Telecommunication Union (ITU) revise ship communication technical standards (e.g., adjusting the bandwidth minimum for port communication or latency requirements for ocean surveillance), or when domestic maritime authorities issue new ship communication specifications, the corresponding specification parameters are automatically extracted, replacing the original parameter ranges. The matching relationship between communication service behaviors and navigation scenario tags is updated as the scenario-based service requirements in the specifications are adjusted.

[0242] Based on practical data, new mapping rules for navigation scenario labels and communication service behaviors are created to update the semantic-communication mapping table. For example, if a certain type of vessel (such as a newly added large cruise ship) has new communication service needs in a specific navigation scenario (such as passenger communication during a cruise ship's ocean voyage), based on feedback from actual practical data, new navigation scenario label-communication service behavior (including communication service type-communication parameter range) communication mapping rules are created in the semantic-communication mapping table, or the existing parameter range is expanded to adapt to practical needs.

[0243] Furthermore, in this embodiment, a semantic-communication model is constructed based on the feature-semantic mapping table and the semantic-communication mapping table to determine the predicted communication behavior based on the temporal sequence of scene labels and the predicted scene labels. Specifically, the semantic-communication model is constructed through the following steps.

[0244] First, the core architecture of the semantic-communication model is defined. In this embodiment, the semantic-communication model adopts a four-layer architecture of "rule storage layer + temporal parsing layer + matching decision layer + output adaptation layer", which meets the real-time requirements of ship AIS trajectory semantic annotation.

[0245] The rule storage layer stores feature-semantic mapping tables and semantic-communication mapping tables, using a key-value pair storage model. Navigation scenario tags serve as the core index, associating the antecedent (discrete combination of features) of semantic mapping rules with their corresponding support and mapping confidence, and the consequent (communication business behavior) of communication mapping rules. All stored data supports real-time updates; when the feature-semantic mapping tables and semantic-communication mapping tables are updated or adjusted, the rule storage layer updates synchronously. For example, {port operations: {semantic mapping rules: [{ Low, High}→Port operations, support 6.8%, mapping confidence 78.3%], communication mapping rule: port scheduling communication, parameter range: {bandwidth: 1-2Mbps, latency: ≤50ms, transmission frequency: 1 time / 30s}}}.

[0246] The temporal parsing layer receives the temporal sequence of scene labels and performs temporal analysis. For example, it extracts all time steps and their corresponding trajectory scene labels from the temporal sequence of scene labels, identifies whether the trajectory scene labels of adjacent time steps are consistent (to determine if there is a label switch), marks the time node of the label switch (e.g., 12:00-12:01 is the label switch period), and extracts the trajectory scene label (i.e., the effective semantic label) and the corresponding current discrete feature combination for the current time, providing temporal basis and feature support for subsequent matching decisions.

[0247] The matching decision layer, based on the mapping rules of the rule storage layer and the parsing results of the temporal parsing layer, achieves accurate matching of "feature-tag-communication service". It employs a "double-check matching" logic to ensure matching accuracy. For example, firstly, based on the currently active tags extracted by the temporal parsing layer, it retrieves the corresponding standardized semantic mapping rules from the rule storage layer to verify whether the current discrete combination of features meets the preconditions of the standardized semantic mapping rules (e.g., the discrete combination of features corresponding to port operations is...). Low, First, verify the mapping confidence of the standardized semantic mapping rule. If the mapping confidence is ≥80%, directly confirm the matching of the corresponding communication service type and parameter range. If the mapping confidence is between 70% and 80%, combine the semantic tags of adjacent time intervals and the adaptation of communication services to make a comprehensive judgment (e.g., if adjacent intervals are all port operations and are adapted to port scheduling communication, then confirm the match). If the mapping confidence is <70%, retrieve the default communication service and parameters corresponding to the semantic tag in the rule storage layer to avoid matching failure.

[0248] The output adaptation layer converts the communication service type and communication parameter range determined by the matching decision layer into a format directly recognizable by the ship's communication module (such as JSON format). Simultaneously, it associates the timing information (time interval, effective duration) of the current semantic tag and outputs the adaptation result. The output result supports real-time feedback and synchronization with the ship's communication control module, enabling automatic switching of communication services and parameter adjustment. Example output format: {Time interval: 08:00-12:00, Semantic tag: Port operation, Communication service: Port scheduling communication, Parameter range: Bandwidth 1-2Mbps, Latency ≤50ms, Transmission frequency 1 time / 30s}.

[0249] Secondly, dynamic parameters of the model are set based on the communication requirements of the target vessel. In this embodiment, these dynamic parameters include a label switching threshold and a mapping confidence priority. The label switching threshold controls the frequency of communication service matching and switching. For example, it is set to 1 minute, meaning that communication adaptation (i.e., communication service switching) is only performed when the expected effective duration is ≥1 minute, avoiding frequent switching caused by instantaneous label fluctuations. The mapping confidence priority controls the accuracy of the verification. For example, semantic mapping rules with a mapping confidence of ≥80% are set as the highest priority, 70%-80% as medium priority, and communication services are set as the lowest priority by default. Matching is performed according to priority order to ensure matching accuracy. Both parameters can be flexibly adjusted according to the actual navigation scenario (e.g., ocean-going, port), supporting automated calibration. In this embodiment, communication requirements can be set manually or determined by the trajectory scene label of the current time step. For example, if the communication requirements for ocean cruise require high accuracy, the label switching threshold is set to 1 minute, the mapping confidence priority is set to 90%, the standardized semantic mapping rule is the highest priority, 70%-90% is the medium priority, and the default communication service is the lowest priority.

[0250] Finally, a joint test was conducted on the semantic classification model, the LSTM neural network model, and the semantic-communication model to ensure data fluency, accuracy, and effectiveness. The scene label time sequence output by the semantic classification model and the predicted scene label output by the LSTM neural network model can be transmitted to the time sequence parsing layer in real time, ensuring response speed (≤1s) and accuracy, and ensuring no matching errors, switching delays, or other issues.

[0251] By establishing a semantic-communication mapping table based on industry-standard knowledge, and using the semantic mapping rules mined by the Apriori algorithm to establish a feature-semantic mapping table, a semantic-communication model is constructed. This model is then combined with a semantic classification model and an LSTM neural network model to predict communication behavior based on the temporal changes of AIS data. This model is then applied to automatic switching of communication services and fine-tuning of communication parameters, enabling the prediction of future communication behavior changes of ships through semantic mining of AIS trajectories. It supports dynamic adjustment of mapping rules based on ship type and navigation environment, with a parameter matching error of ≤8%. It can be used for intelligent and dynamic configuration of business parameters in ship dynamic modeling and simulation, solving the problem of the disconnect between simulation business parameter configuration and actual ship communication behavior.

[0252] Furthermore, the comprehensive accuracy of semantic classification models, LSTM neural network models, and semantic-communication models can be verified through matching accuracy, parameter matching precision, and switch precision. Among these, matching accuracy... In the formula represents the number of true positive samples, represents the number of true negative samples, represents the number of false positive samples, represents the number of false negative samples; the parameter matching accuracy , where in the formula represents the th communication parameter number of samples that meet the parameter matching condition, represents the total number of samples, represents the total number of communication parameters, where parameter matching means that the communication parameter range corresponding to the predicted communication behavior output by the model intersects with the sample communication parameter range, and the intersection proportion is greater than a proportion threshold (e.g., 80%). For example, if the bandwidth in the sample communication parameters is 1-2Mbps, the model outputs 1.2-1.8Mbps, the intersection is 1.2-1.8Mbps with a proportion of 80%, which is determined as matching; the time sequence switching accuracy is used to measure the accuracy and timeliness of communication service / parameter switching when the model performs semantic label time sequence switching, , represents the number of samples with timely and accurate communication adaptation switching, represents the total number of samples with label switching. The criterion for determining timely communication adaptation switching is that after label switching, the response time of model service / parameter switching is ≤ 1s; the criterion for determining accurate communication adaptation switching is that the switched service / parameter is consistent with the standard service / parameter corresponding to the new label.

[0253] The comprehensive model accuracy is obtained by weighted summation of the matching accuracy, the parameter matching accuracy and the time sequence switching accuracy. After both the individual verification and the comprehensive verification are satisfied, the model is considered to be trained or constructed. For example, the individual verification includes matching accuracy ≥ 92% (ensuring accurate matching of core guaranteed communication service types), communication parameter matching accuracy ≥ 88% (ensuring that the communication parameter range adapts to simulation and actual communication requirements), and time sequence switching accuracy ≥ 85% (ensuring no interruption of communication services and timely adaptation during label switching); if any individual index fails to meet the standard, even if the final accuracy is ≥ 90%, it is still determined as not up to the standard, and the corresponding link needs to be optimized in a targeted manner (for example, if the parameter matching accuracy fails to meet the standard, the parameter range mapping logic is optimized); the comprehensive verification includes that the final model accuracy is ≥ 90%, which measures the overall matching performance of the model, ensures that the services and parameters of the vast majority of samples are accurately matched, and conforms to actual navigation and communication requirements; if the final accuracy is < 90%, it is determined as not up to the standard, and backtracking optimization is required (such as semantic-communication mapping table, supplementing training associated data sets, retraining the semantic classification model, etc.).

[0254] An embodiment of the present invention provides a ship communication behavior simulation method based on AIS trajectory semantics, comprising: Step 1: A ship communication behavior prediction method based on AIS trajectory semantics to obtain prediction scene labels and prediction communication behavior.

[0255] A method for predicting ship communication behavior based on AIS trajectory semantics (steps S1-S4) is used to obtain prediction scene labels and predict communication behavior.

[0256] Step 2: Standardize the predicted scene labels and the predicted communication behaviors to obtain standardized simulation results.

[0257] The predicted scene labels and predicted communication behaviors are standardized and transformed into standardized simulation parameters. These standardized simulation results are then output to the ship satellite communication simulation system to achieve dynamic adaptation of the simulation parameters.

[0258] In this embodiment, the standardized simulation results include service type configuration, bandwidth requirement threshold, latency QoS requirements, peak communication periods, etc., supporting real-time call and updating of parameter configurations by the simulation system. It employs engineering mapping and standardized processing adapted to ship communication scenarios, based on predicted scenario labels (predicted future navigation scenarios) and predicted communication behaviors. These predicted behaviors include communication service types (e.g., port scheduling communication, ocean navigation communication) and communication parameter ranges (e.g., bandwidth 1-2Mbps, latency ≤50ms). Standardized parameters adapted to ship communication simulation are used, following industry-standard simulation formats (e.g., JSON, XML), with fixed parameter values ​​(no ambiguity range), standardized parameter naming, and standardized format conversion, allowing direct import into simulation tools (e.g., NS-3, OPNET).

[0259] Among them, the parameter values ​​are fixed, including determining fixed parameter values ​​from the predicted "communication parameter range" according to "simulation accuracy requirements + industry standard values", prioritizing the middle value of the range (e.g., bandwidth 1-2Mbps, fixed as 1.5Mbps); if the range is a critical value constraint (e.g., latency ≤50ms), the critical value (50ms) is taken; if there are special scenario requirements (e.g., low bandwidth scenario in ocean), the lower limit of the range (e.g., bandwidth 0.5Mbps) is taken to ensure that the values ​​fit the predicted scenario.

[0260] Standardized parameter naming includes associating the predicted "communication service type" and "predicted scenario label" with the simulation parameters, unifying parameter identifiers (e.g., naming the bandwidth parameter of port scheduling communication as "port_scheduling_bandwidth"), avoiding parameter confusion during simulation, and following the industry standards for ship communication simulation.

[0261] The format standardization conversion includes integrating the solidified parameters and associated prediction scenario labels / communication service types according to the format required by the simulation tool to generate a standardized simulation parameter package.

[0262] Here is an example in JSON format: { "predicted_scenario": "Port Operations", / / Predicted scenario label (related scenario) "communication_service": "port_scheduling", / / Predicted communication service type (standardized naming) "simulation_parameters": { "bandwidth": 1.5, / / Fixed bandwidth (unit: Mbps) "delay": 50, / / Curing delay (unit: ms) "transmission_frequency": 2 / / Fixed transmission frequency (unit: times / minute, converted from 1 time / 30s) } } Step 3: Perform compliance verification on the standardized simulation results, and take the standardized simulation results that pass the compliance verification as the final simulation results.

[0263] The standardized simulation results are subjected to compliance verification, and the standardized simulation results that pass the compliance verification are used as the final simulation results. In this embodiment, the compliance verification includes checking the rationality of parameter values ​​(such as the bandwidth not exceeding the maximum supported value of the ship communication module) and the correctness of the format (such as the parameter units being consistent and no missing items), ensuring that the converted parameters can be directly imported into the simulation tool. If the verification fails, it will automatically backtrack to the parameter solidification step and readjust the values.

[0264] Based on predicted scenario labels and predicted communication behavior, through standardized processing and compliance verification, the ambiguity of the prediction results is eliminated and the simulation format is unified, enabling seamless connection between "predicted communication behavior" and subsequent application "simulation tools", completing dynamic adaptation to requirements, improving the realism of simulation, and realizing dynamic adjustment of simulation business parameters with ship navigation scenarios. This solves the problem of static simulation parameters in traditional simulations and improves the matching degree between simulation scenarios and actual communication needs by more than 60%.

[0265] This invention provides a ship communication behavior prediction system based on AIS trajectory semantics, such as... Figure 8 As shown, it includes: The data acquisition module is used to acquire the AIS data of the target vessel, preprocess the AIS data, and obtain trajectory feature vectors for multiple time steps. The semantic annotation module is used to perform semantic annotation on all the trajectory feature vectors using a semantic classification model to obtain a temporal sequence of scene labels. The semantic prediction module is used to perform semantic temporal prediction based on the temporal sequence of the scene labels to obtain the predicted scene labels; The communication prediction module is used to determine the predicted communication behavior of the ship based on the time sequence of the scene label and the predicted scene label, using a semantic-communication model.

[0266] Furthermore, it also includes: The simulation output module is used to standardize the predicted scene labels and the predicted communication behavior, or to standardize the simulation results; to perform compliance verification on the standardized simulation results, and to take the standardized simulation results that pass the compliance verification as the final simulation results.

[0267] This invention uses satellite communication simulation of ocean-going cargo ships as an application scenario to describe in detail the implementation process and results of a ship communication behavior prediction system based on AIS trajectory semantics.

[0268] (1) Scene parameter settings: AIS dataset: Cargo ship MMSI code is 412345678, collection period is 10 seconds, collection duration is 72 hours, including latitude and longitude, speed, heading and other data; Historical communication data: Same-time vessel communication records, including cargo scheduling data (bandwidth requirement 50-80Mbps), equipment monitoring data (bandwidth requirement 2-5Mbps), and crew communication data (bandwidth requirement 1-2Mbps). Supporting geographic data: Locations of Chinese coastal ports, ocean shipping route planning data, and anchorage distribution.

[0269] (2) Implementation process: Step a: The data acquisition module receives the above data, preprocesses the AIS dataset, removes 3 abnormal speed data (>25 knots, the maximum design speed of the cargo ship is 22 knots), smooths the trajectory through Kalman filtering, extracts core features such as 12 turning times and 4 hours of port stay, and constructs a trajectory feature vector based on the core features; Step b: The semantic annotation module performs semantic annotation on the trajectory feature vector, completes the semantic recognition of the trajectory, and outputs the scene label time sequence, such as "Day 1 00:00-04:00: Anchorage waiting", "04:00-08:00: Port operation", "08:00-72:00: Ocean cruise"; Step c: The semantic prediction module predicts semantic changes in the next 24 hours and obtains predicted scene labels, such as "0-4 hours: ocean cruise", "4-8 hours: anchorage waiting", and "8-24 hours: port operations". Step d: The communication prediction module determines the predicted communication behavior, i.e. the communication demand prediction result, based on the time sequence of the scene tags and the predicted scene tags. For example, "ocean cruise" corresponds to equipment monitoring (bandwidth 3Mbps), "anchorage waiting" corresponds to crew communication (bandwidth 2Mbps), and "port operation" corresponds to cargo scheduling (bandwidth 60Mbps). Step e: The simulation output module converts the prediction results into simulation parameters and outputs them to the ship satellite communication simulation system. The system automatically adjusts the service bandwidth configuration for each time period.

[0270] (3) Implementation results: Semantic annotation accuracy: The accuracy rate of trajectory scene label recognition reached 93.2%, with the accuracy rates of "port operations" and "ocean cruise" labels being 95.1% and 94.5%, respectively. Prediction accuracy: The accuracy rate of predicting communication service types in the next 24 hours reaches 90.5%, and the bandwidth demand prediction error is ≤7%, which meets the accuracy requirements of simulation parameter configuration. Simulation adaptability: Based on the simulation results of the predicted parameters, compared with the actual QoS data of the cargo ship's communication, the throughput error is ≤5% and the delay error is ≤8%, which is more than 60% more accurate than the traditional fixed parameter simulation.

[0271] The above-described method and system embodiments are based on the same principles, and their related aspects can be referenced from each other to achieve the same technical effects. For specific implementation processes, please refer to the foregoing embodiments, which will not be repeated here.

[0272] In summary, the ship communication behavior prediction method, simulation method, and prediction system based on AIS trajectory semantics of the present invention have at least one of the following beneficial effects: 1. By processing ship AIS data to obtain trajectory feature vectors and performing semantic annotation, the trajectory semantics of AIS data can be mined to obtain scene label time series and further obtain predicted scene labels. Through the established semantic-communication model, predicted communication behavior can be determined, enabling the prediction of future changes in ship communication behavior. This can be used for intelligent and dynamic configuration of service parameters in ship dynamic modeling and simulation, enabling communication service parameters to be dynamically adjusted in real time with AIS data during simulation, improving the consistency between simulation configuration and actual ship communication behavior, solving the problem of disconnect between simulation service parameter configuration and actual ship communication behavior, and significantly improving the realism and reliability of QoS simulation.

[0273] 2. By extracting features from AIS data and obtaining trajectory feature vectors, a semantic classification model is used to obtain trajectory scene labels. The ship's navigation intention is extracted from AIS data, extending the value of AIS data from location description to behavioral understanding. The trajectory scene labels (such as anchorage waiting, port operations, and ocean cruise) are accurately identified, realizing the semantic parsing and in-depth utilization of AIS data. This solves the problems of shallow application and semantic missingness of AIS trajectory data in existing technologies.

[0274] 3. Establish an end-to-end predictive data link from AIS data to the navigation scenario (track scenario label) and then to the predicted communication behavior. This enables the prediction of changes in the future communication needs of ships and can provide dynamic and forward-looking parameter configuration references for subsequent simulation systems, thereby improving the accuracy of ship communication simulation and the guiding value of AIS data.

[0275] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0276] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting ship communication behavior based on AIS trajectory semantics, characterized in that, include: Obtain the AIS dataset of the target vessel, preprocess the AIS dataset to obtain trajectory feature vectors at multiple time steps; For all the trajectory feature vectors, semantic annotation is performed using a semantic classification model to obtain a temporal sequence of scene labels; Based on the time series sequence of the scene labels, perform time series prediction to obtain the predicted scene labels; Based on the time sequence of the scene labels and the predicted scene labels, the predicted communication behavior corresponding to the target vessel is determined using a semantic-communication model; wherein, the semantic-communication model includes a semantic-communication mapping table; Based on the time-series sequence of the scene labels and the predicted scene labels, a semantic-communication model is used to determine the predicted communication behavior of the ship, including: Perform time series analysis on the scene label time series and the predicted scene label to determine the currently effective label, and determine whether the predicted scene label is consistent with the currently effective label; If they match, the current communication behavior corresponding to the currently active tag is obtained, and the current communication behavior is used as the predicted communication behavior; wherein, the predicted communication behavior includes the communication service type and the communication parameter range; If there is a discrepancy, the predicted communication behavior is determined using the semantic-communication mapping table based on the predicted scene label.

2. The method according to claim 1, characterized in that, The ship AIS dataset is preprocessed to obtain trajectory feature vectors at multiple time steps, including: The ship AIS dataset is subjected to trajectory windowing to obtain trajectory sub-segments at multiple time steps; For each time step, feature extraction is performed on the trajectory sub-segment to obtain the corresponding static and dynamic features; Based on the static and dynamic features of each time step, a trajectory feature vector is constructed for each time step.

3. The method according to claim 1, characterized in that, For all the trajectory feature vectors, semantic annotation is performed using a semantic classification model to obtain a temporal sequence of scene labels, including: The trajectory feature vector at each time step is input into the semantic classification model to obtain the corresponding trajectory scene label; wherein, the semantic classification model is obtained through training; The trajectory scene labels of all the time steps are integrated in time to form the scene label time sequence.

4. The method according to claim 1, characterized in that, Based on the time series sequence of the scene labels, time series prediction is performed to obtain predicted scene labels, including: Obtain the predicted duration, and extract historical tag time series data from the scene tag time series based on the predicted duration; Based on the historical tag time series data, the initial prediction probability of each navigation scenario tag is obtained using an LSTM neural network model; Based on the predicted initial probabilities of all the navigation scene labels, the predicted scene label corresponding to the predicted duration is determined.

5. The method according to claim 4, characterized in that, Based on the initial predicted probability of each of the navigation scene labels, the predicted scene label corresponding to the predicted duration is determined, including: Based on the aforementioned ship AIS dataset, obtain the corresponding target ship type and target season label; Based on the target vessel type, the target type label weight for each navigation scenario label is determined according to the vessel type-semantic label weight mapping relationship; Based on the target seasonal label, the target seasonal label coefficient for each of the navigation scenario labels is determined according to the seasonal-semantic label coefficient mapping relationship; Based on the target type label weight and the target season label coefficient, the initial prediction probability of each navigation scenario label is corrected to obtain the corresponding corrected prediction probability; Based on the predicted correction probability, the predicted scene label corresponding to the predicted duration is determined.

6. The method according to claim 1, characterized in that, The semantic-communication model also includes a feature-semantic mapping table; Using the current communication behavior as the predicted communication behavior includes: The communication service type corresponding to the current communication behavior is used as the communication service type corresponding to the predicted communication behavior; Obtain the current discrete feature combination corresponding to the currently effective label, and obtain the predicted discrete feature combination corresponding to the predicted scene label; Based on the currently active label and the feature-semantic mapping table, it is determined whether there is a fluctuation between the current discrete feature combination and the predicted discrete feature combination; If there is no fluctuation, the range of communication parameters corresponding to the current communication behavior shall be used as the range of communication parameters corresponding to the predicted communication behavior.

7. The method according to claim 1, characterized in that, Before performing time series analysis on the scene label time series and the predicted scene label, the method further includes: Obtain the time sequence of the scene labels and the discrete feature combination corresponding to the predicted scene labels; Based on all the discrete feature combinations, determine the accuracy of the scene label time sequence and the first label of the predicted scene label; The ship communication data of the target ship is acquired, the ship communication data and the ship AIS data are synchronized in time, and the ship communication behavior sequence corresponding to the scene tag time sequence is extracted. Based on the scene tag time sequence and the ship communication behavior sequence, the accuracy of the second tag of the scene tag time sequence is determined according to the semantic-communication mapping table; If the accuracy of the first label and / or the accuracy of the second label fail, the time sequence of the scene label and / or the predicted scene label are reacquired.

8. A simulation method for ship communication behavior based on AIS trajectory semantics, characterized in that, include: Based on any one of claims 1-7, predictive scene labels and predictive communication behaviors are obtained; The predicted scene labels and the predicted communication behaviors are standardized to obtain standardized simulation results. The standardized simulation results are subjected to compliance verification, and the standardized simulation results that pass the compliance verification are taken as the final simulation results.

9. A ship communication behavior prediction system based on AIS trajectory semantics, characterized in that, include: The data acquisition module is used to acquire the ship AIS dataset of the target ship, preprocess the ship AIS dataset, and obtain trajectory feature vectors at multiple time steps. The semantic annotation module is used to perform semantic annotation on all the trajectory feature vectors using a semantic classification model to obtain a temporal sequence of scene labels. The semantic prediction module is used to perform semantic temporal prediction based on the temporal sequence of the scene labels to obtain the predicted scene labels; The communication prediction module is used to determine the predicted communication behavior of the ship based on the time series sequence of the scene labels and the predicted scene labels, using a semantic-communication model; wherein, the semantic-communication model includes a semantic-communication mapping table; Based on the time-series sequence of the scene labels and the predicted scene labels, a semantic-communication model is used to determine the predicted communication behavior of the ship, including: Perform time series analysis on the scene label time series and the predicted scene label to determine the currently effective label, and determine whether the predicted scene label is consistent with the currently effective label; If they match, the current communication behavior corresponding to the currently active tag is obtained, and the current communication behavior is used as the predicted communication behavior; wherein, the predicted communication behavior includes the communication service type and the communication parameter range; If there is a discrepancy, the predicted communication behavior is determined using the semantic-communication mapping table based on the predicted scene label.

Images