Ship trajectory analysis and anomaly detection method and system based on multi-source data fusion

Abstract

The invention relates to a ship trajectory analysis and anomaly detection method and system based on multi-source data fusion, and relates to the technical field of maritime affair intelligent supervision, and the method comprises the steps: extracting spatial-temporal features, environmental risk features and ship interaction features through real-time fusion of AIS data, environmental data and static archive data; self-adaptive weighted fusion is carried out by using a multi-head self-attention mechanism to form a unified feature vector; then inputting the data into a space-time diagram neural network model, taking a ship as a node, constructing an edge by a spatial distance, and aggregating space-time neighborhood information to output an abnormal probability score; and finally, a reference threshold value is set based on historical normal data, dynamic adjustment is performed according to the real-time environment risk index, and self-adaptive early warning is realized. According to the method, the defects of single sensing dimension, shallow space-time modeling and rigid early warning mechanism in the prior art are effectively overcome, and the accuracy, environmental adaptability and real-time performance of anomaly detection are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of maritime intelligent supervision technology, specifically to a method and system for ship trajectory analysis and anomaly detection based on multi-source data fusion. Background Technology

[0002] Detection of abnormal ship navigation behavior is a core component of intelligent maritime supervision, crucial for ensuring maritime traffic safety and improving waterway operational efficiency. The widespread adoption of Automatic Identification Systems (AIS) has provided a data foundation for trajectory-based anomaly detection research, prompting the field to evolve from the traditional "human monitoring" model towards automation and intelligence.

[0003] In existing technologies, several ship anomaly detection schemes based on machine learning and deep learning have emerged. For example, the prior art document "Method, System and Storage Medium for Detecting Abnormal Navigation Behavior of Inland Waterway Vessels" (Publication No.: CN116010829A) proposes a scheme that integrates the ship's own trajectory motion characteristics with the interaction characteristics of the ship with buoys in the same waterway through an attention mechanism autoencoder model, and uses cluster analysis to obtain different ship motion patterns, thereby constructing a local autoencoder model for real-time anomaly identification. This scheme has achieved detection of microscopic abnormal behavior to a certain extent and represents an advanced level in this field.

[0004] However, after in-depth analysis, the inventors discovered that this type of existing technology still has obvious limitations, mainly in the following three aspects: First, the perception dimension is singular, lacking a collaborative perception of environmental risks and dynamic interactions between ships. Existing solutions (such as CN116010829A) mainly focus on the interaction between ships and static channel facilities (such as buoys), failing to effectively integrate real-time meteorological and sea state data (such as visibility and wave height). This results in the system's inability to distinguish the inherent differences in ship behavior under calm conditions versus adverse environments such as poor visibility and large waves. Furthermore, its interaction feature modeling fails to fully consider dynamic, spatially varying collision avoidance relationships between ships, and cannot effectively capture key abnormal behaviors such as dangerous approach and violations of collision avoidance rules, leading to an incomplete detection context.

[0005] Second, the spatiotemporal modeling is superficial and fails to achieve deep spatiotemporal correlation mining. Existing methods mostly use a combination of autoencoders and clustering, but their extraction and fusion of spatiotemporal features remain superficial. In particular, it lacks explicit modeling of the spatiotemporal graph structure of ship groups, and cannot simultaneously capture the temporal dynamics of nodes (ships), the complex influence of ships in their spatial neighborhood, and the changing patterns of these influences over time through graph neural networks, thus limiting the model's ability to represent complex anomaly patterns.

[0006] Third, the early warning mechanism is rigid and unable to adapt to the external environment. Existing technologies rely heavily on fixed thresholds or thresholds based solely on historical trajectory statistics for anomaly detection. This rigid mechanism faces a dilemma in the complex and ever-changing real-world navigation environment: if the threshold is set too high, it may miss real threats in high-risk environments such as severe weather; if the threshold is set too low, it will generate a large number of invalid alarms in normal environments, interfering with regulatory decisions and lacking the intelligent adjustment capability to adapt to environmental risks.

[0007] Therefore, existing technologies generally face three major challenges in practical applications: incomplete detection dimensions, shallow spatiotemporal modeling, and rigid early warning mechanisms. There is an urgent need for an anomaly detection method and system that can deeply integrate multi-source information, accurately model complex spatiotemporal relationships, and have intelligent environmental perception capabilities. Summary of the Invention

[0008] To address the problems of existing technologies, this invention provides a method and system for ship trajectory analysis and anomaly detection based on multi-source data fusion.

[0009] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: Firstly, a method for ship trajectory analysis and anomaly detection based on multi-source data fusion, comprising the following steps: S1: Multi-source data fusion steps: Real-time acquisition of Automatic Identification System (AIS) data, environmental data, and static vessel profile data; spatiotemporal alignment of the data using the Maritime Mobile Service Identifier (MMSI) and Coordinated Universal Time (UTC) as the association keys, and Z-score standardization of all numerical features to form a fused data sequence; the Z-score standardization formula is: ,in These are the original eigenvalues. and The mean and standard deviation of this feature are calculated from the historical training dataset; the environmental data includes at least wind speed, wind direction, wave height, visibility, and surface current speed. S2: Multidimensional feature extraction and fusion steps: extract spatiotemporal features, environmental risk features, and ship interaction features from the fused data sequence; input the three types of features into the fusion model based on the multi-head self-attention mechanism (MHSAM) for adaptive weighting and concatenation, and output a unified feature vector; S3: Spatiotemporal Graph Neural Network Analysis Steps: Input the unified feature vector into the pre-trained spatiotemporal graph neural network (ST-GNN) model; the ST-GNN model uses each ship as a graph node, and the spatial Euclidean distance between ships is less than a preset neighborhood radius. To construct the graph edges, where The value ranges from 3 to 10 nautical miles; the ST-GNN model aggregates the features of a node itself and its neighboring nodes through alternating stacked spatial graph convolutional layers and temporal convolutional layers, and outputs a probability score representing the degree of abnormality in ship behavior. ; S4: Dynamic threshold early warning step: Based on historical normal flight trajectory data, calculate the anomaly probability score output by the ST-GNN model. The distribution is used, and the 99.5th percentile of this distribution is selected as the benchmark threshold. Based on the real-time calculated total environmental risk index The baseline threshold is linearly adjusted, resulting in a dynamic threshold. The calculation formula is: ,in This is a preset proportionality coefficient, with a value ranging from 0.1 to 0.5; if If so, an early warning message will be generated and issued.

[0010] In one specific implementation of the first aspect, in the multidimensional feature extraction and fusion step: the spatiotemporal features include the average speed calculated based on a time sliding window. Standard deviation of speed Average rate of change of heading Average longitude of trajectory points and average latitude ; The environmental risk characteristics are obtained through the following quantification rules: when the visibility value Visibility risk index below 5000 meters When the wave height value When the wave height exceeds 2 meters, the wave height risk index is high. The overall environmental risk index The weighted sum of the risk indices for each component is calculated using the following formula: ,in and The preset weighting coefficients, and + =1; The ship interaction characteristics include the Euclidean distance to the nearest ship. And the predicted meeting distance based on the current velocity vector within a preset 10-minute time window. Number of ships less than 0.5 nautical miles .

[0011] In one specific implementation of the first aspect, the ST-GNN model in the spatiotemporal graph neural network analysis step is trained in the following manner: The spatial graph convolutional layer is implemented using a graph attention network (GAT), and the temporal convolutional layer is implemented using a one-dimensional convolutional neural network (TCN). The model minimizes the loss function. Training is performed, and the loss function is... Cross-entropy loss and Regularization loss The sum of ,in This is the preset regularization strength coefficient, and a value of 0.01 is recommended. The training data consists of historical multi-source fusion data and corresponding verified anomaly labels, which are derived from publicly available accident reports and manually labeled anomaly events by maritime authorities.

[0012] In one specific implementation of the first aspect, the ST-GNN model is co-trained based on a federated learning framework, and the specific training steps include: Multiple local clients use local de-identified data to train their respective ST-GNN models; The global aggregation server periodically collects model weight gradients from each local client and performs a federated averaging algorithm. Aggregate to generate global model weights, where The number of clients participating in the aggregation; Distribute the updated global model weights to each local client; After each round of global aggregation, the model weight hash value of this round of aggregation, the MMSI prefix identifier of the participating client, and the timestamp are written into an immutable blockchain distributed ledger.

[0013] In the first aspect, the specific data flow and parameter determination rules for the ship trajectory analysis and anomaly detection method based on multi-source data fusion are as follows: 1. Data Fusion Process Details: The multi-source data fusion process forms the foundation for subsequent analysis. AIS data streams, environmental data streams, and static archive data are correlated in real-time in memory using MMSI and UTC timestamps, forming a fused data frame with a spatiotemporal label. Subsequently, pre-stored standardized parameters (mean) are invoked. and standard deviation The Z-score normalization calculation is performed on the numerical fields (speed, longitude, latitude, wave height) in the data frame, and the formula is as follows: .in, and The data is derived from offline statistics of historical training datasets (AIS and meteorological data of port waters over the past year), ensuring the stability of the data distribution.

[0014] 2. Method for determining feature weights and risk index: The total environmental risk index The calculation of visibility risk index is a crucial step in quantifying environmental context. The 5000-meter threshold is set based on the general definition of "poor visibility" in the International Maritime Organization's (IMO) International Regulations for Preventing Collisions at Sea and maritime regulatory practices; the wave height risk index formula... The 2-meter threshold is based on industry consensus on the safety of navigation operations for small and medium-sized vessels and port navigation guidelines. The weighting coefficients α and β are determined by analyzing the correlation between historical accident data and environmental factors. For example, in one embodiment, a threshold of 2 meters is set. =0.6, =0.4, to emphasize the dominant influence of visibility on navigation safety in specific waters.

[0015] 3. Model parameters and decision rule basis: The neighborhood radius in the ST-GNN model The nautical distance setting is based on the ship domain model and the conventional requirements for safe encounter distances in VTS (Vessel Traffic Service), ensuring that effective inter-ship interactions can be captured. The dynamic threshold formula... The proportionality coefficient in The value was determined empirically by performing a grid search on the validation set with a fixed step size (0.05) to find the optimal balance between anomaly detection rate and false positive rate. This rule allows the system to operate under harsh environments ( (Raising) The threshold is automatically lowered to improve detection sensitivity, which is consistent with the logic of safety warning.

[0016] Secondly, a ship trajectory analysis and anomaly detection system based on multi-source data fusion includes: The data fusion module is configured to perform the multi-source data fusion step, including storing feature means. and standard deviation Memory; The feature fusion module is configured to perform the multidimensional feature extraction and fusion steps, including a graphics processing unit (GPU) for running the MHSAM fusion model; The behavior analysis module, configured to perform the spatiotemporal graph neural network analysis steps, includes a GPU or tensor processor (TPU) for storing and running the ST-GNN model. The early warning decision module is configured to execute the dynamic threshold early warning step, including calculating the dynamic threshold. And the central processing unit (CPU) that generates early warning information.

[0017] In one specific implementation of the second aspect, the data fusion module pre-stores the mean values ​​of each feature used for Z-score standardization. and standard deviation The lookup table.

[0018] In one specific implementation of the second aspect, the behavior analysis module has a preset neighborhood radius. nautical miles. In one specific implementation of the second aspect, a preset proportional coefficient is used in the early warning decision module. .

[0019] In one specific implementation of the second aspect, a model training and update subsystem is further included, the subsystem comprising: Multiple local training clients and a global aggregation server based on a federated learning architecture; A blockchain unit that communicates with the global aggregation server, the blockchain unit being composed of multiple distributed nodes.

[0020] In the second aspect, the rules for internal module collaboration and parameter fixing are as follows: 1. Data Flow and Processing Logic Between Modules: This system achieves efficient data processing through modular pipelines. During initialization, the data fusion module loads pre-calculated feature mean values ​​from the solid-state drive. and standard deviation The lookup table is moved into memory, enabling high-speed standardization of the input data stream. The feature fusion module utilizes the parallel computing power of its GPU to perform numerous matrix multiplications and Softmax attention weight calculations in the MHSAM model. The behavior analysis module relies on its TPU's dedicated tensor processing core to efficiently complete the iterative operations of graph convolution and temporal convolution in ST-GNN. The CPU of the early warning decision module is responsible for performing lightweight logical judgments, i.e., continuously comparing the data from the TPU. Compared with the calculation based on the above method .

[0021] 2. System Parameter Fixation and Source: To ensure system stability and consistency, key parameters are fixed during deployment. These parameters are pre-stored in the data fusion module. and The lookup table is generated offline in batch processing of the entire historical training dataset before the system goes live, and then burned into memory, avoiding the latency and fluctuations of online computation. The neighborhood radius is set in the behavior analysis module. The nautical mileage, as a verified optimal value, is directly encoded in the system configuration file, balancing computational complexity with the integrity of the interaction relationship. Similarly, the proportional coefficient used in the early warning decision module... The optimal solution derived from extensive historical data experiments is set as the system constant. The solidification of these parameters ensures that the system exhibits stable and reliable detection performance under different deployment environments.

[0022] Thirdly, a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of a method for ship trajectory analysis and anomaly detection based on multi-source data fusion.

[0023] The beneficial effects of this invention are as follows: 1. This invention constructs a unified feature representation that comprehensively describes ship behavior and environmental situation through multi-source data fusion and multi-dimensional feature extraction, overcoming the limitations of a single data source. Utilizing an attention-based feature fusion model and a spatiotemporal graph neural network, the system can adaptively capture the deep coupling patterns of individual ship behavior, complex interactions between ships, and the dynamic influence of their environment. This enables a more accurate and robust probabilistic assessment of abnormal navigation behavior, significantly improving detection accuracy and scene adaptability. 2. This invention introduces an environmental risk-adaptive dynamic early warning mechanism, enabling the anomaly detection threshold to intelligently adjust according to changes in external navigation conditions. This effectively balances detection sensitivity and false alarm rate under different environments, enhancing the system's practicality and reliability. Furthermore, the system employs a federated learning architecture for collaborative model training and leverages blockchain technology to ensure the trustworthiness and transparency of the training process. While fully utilizing multi-party data to improve model performance, it ensures that ship privacy data remains local, providing a solid technical foundation for large-scale deployment in a distributed regulatory environment. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0025] Figure 2 This is a schematic diagram of the overall system architecture of the present invention.

[0026] Figure 3 This is a schematic diagram of the multidimensional feature extraction and fusion of the present invention.

[0027] Figure 4 This is a schematic diagram of the ST-GNN model structure of the present invention.

[0028] Figure 5 This is a schematic diagram illustrating the dynamic threshold adjustment principle of the present invention.

[0029] Figure 6 This is a schematic diagram of the federated learning and blockchain training process of the present invention.

[0030] Figure 7 This is a schematic diagram of the data calculation timing in Embodiment 2 of the present invention. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0032] like Figures 1 to 7 The method and system for ship trajectory analysis and anomaly detection based on multi-source data fusion are shown.

[0033] Example 1: Step 1: Data input as follows: AIS data: MMSI:123456789 :2023-10-2708:00:00,Lon:122.5,Lat:30.2,SOG:10.2knots,COG:95.5°.

[0034] Environmental data: :2023-10-2708:00:00,V:3000m,H:3.5m.

[0035] Static file: MMSI:123456789, ShipType:Cargo, L:150m; Pre-stored parameters: According to the patent description, the system has pre-stored standardized parameters calculated from historical data: =8.5, =4.2.

[0036] Step Two: Step-by-Step Calculation and Judgment Process Now, I will perform the calculations strictly according to the steps described in the patent: Step 1: Multi-source data fusion and standardization Spatiotemporal alignment: I correlated the above AIS and environmental data using MMSI and UTC time.

[0037] Z-score standardization: I extracted the SOG (speed of air) of 10.2 knots from the AIS and performed standardization calculations using pre-stored parameters: Similarly, the same standardization operation is performed on other numerical features such as longitude, latitude, visibility, and wave height.

[0038] Step 2: Multidimensional Feature Extraction and Fusion Extracting spatiotemporal features: If only the current point exists within the time window, then the average speed is calculated. Other characteristics such as standard deviation are calculated similarly.

[0039] Calculate environmental risk characteristics: Visibility risk index: V=3000<5000, therefore High risk index: H=3.5>2, therefore ; Overall Environmental Risk Index: Utilizing example weights from patents. , Extracting ship interaction features: Calculating the distance to the nearest ship based on the positions and speeds of surrounding ships. Number of at-risk vessels in the sea .

[0040] MHSAM Feature Fusion: All the above features (standardized spatiotemporal features, risk index, and interaction features) are organized into a vector and input into the MHSAM model. The model performs a weighted sum of these features according to its pre-trained attention weights, and finally outputs a unified feature vector, such as [0.12, -0.45, 0.87, ..., 1.34].

[0041] Step 3: Spatiotemporal Graph Neural Network Analysis Construct a graph structure: Using the ship as a node, its attributes are the unified feature vector obtained in the previous step. Find the surrounding nodes with distances within... Ships within 1 nautical mile radius act as neighboring nodes, establishing connections.

[0042] ST-GNN inference: The constructed graph is input into a pre-trained ST-GNN model. The model performs calculations through its fixed GAT and TCN layers, and finally generates an anomaly probability score at the output layer using the Softmax function. Assume we obtain: Step 4: Dynamic Threshold Early Warning Determine the dynamic threshold: Benchmark threshold: Determined using historical normal data, according to the patent. .

[0043] Dynamic adjustment: using formulas and presets And the result calculated in the second step : Final judgment: Compare: .

[0044] Decision: Trigger an anomaly warning.

[0045] Example 2: 1. Data preparation and parameter setting Historical training dataset: This dataset uses AIS data, meteorological and marine environmental data, and static vessel records from a port area in the East China Sea throughout 2022, containing over 10 million trajectory point records. The mean values ​​of various features used for Z-score standardization were then calculated from these records. ) and standard deviation ( These are stored in a lookup table. Statistical values ​​for some key features are shown in the table below: feature Mean (μ) Standard deviation (σ) Source and Explanation Speed ​​(knots) 10.5 4.2 Based on historical data statistics, reflecting the distribution of normal navigation speeds in this waterway Longitude (degrees) 122.35 0.15 Based on historical data statistics, reflecting the central area of ​​ship activity Wave height (meters) 1.2 0.8 Based on historical data statistics, reflecting normal sea conditions Model and system parameters: The following validated parameters are used in this embodiment: Environmental risk weighting coefficient: =0.6, =0.4. This weight was determined based on logistic regression analysis of 500 labeled anomalous events in the historical dataset, with the regression coefficient of the visibility factor being approximately 1.5 times that of the wave height factor.

[0046] Neighborhood radius Nautical miles. This value is obtained by comparing on the validation set. Model performance at nautical miles ( (scores), discovery hour The highest score (0.873) was used to determine the score.

[0047] Dynamic threshold scaling factor This value is obtained by performing a grid search on the independent validation set with a step size of 0.05. Maximizing the score is the equilibrium point determined by the objective.

[0048] 2. Real-time data processing and calculation examples Suppose the system acquires the following multi-source data for the target vessel (MMSI: 123456789) at 08:00:00 UTC on October 27, 2023: AIS data: Longitude 122.4°, Latitude 30.5°, Speed ​​12.5 knots, Heading 95°.

[0049] Environmental data: Visibility =3000 meters, wave height =3.0 meters.

[0050] Static data: The vessel type is a cargo ship.

[0051] S1: Multi-source data fusion steps Spatiotemporal alignment: The above data is linked using MMSI and UTC time as keys.

[0052] Z-score standardization: Read parameters from a pre-stored lookup table and standardize numerical features.

[0053] Standardization of sailing speed: =(12.5-10.5) / 4.2≈0.476 Longitude standardization: =(122.4-122.35) / 0.15≈0.333 Wave height standardization: =(3.0-1.2) / 0.8=2.25 Visibility standardization: =(3000- ) / (Calculation process omitted, method is the same as before).

[0054] S2: Multidimensional Feature Extraction and Fusion Steps Spatiotemporal feature extraction (based on a 5-minute sliding window including the current point): Average speed =12.5 sections; Standard deviation of speed =0.8 sections; Average rate of change of heading =1.2° / minute; Quantification of environmental risk characteristics: Visibility Risk Index: =(5000-3000) / 1000=2.0 (because) <5000 meters); High risk index: =3.0-2=1.0 (because) >2 meters); Overall Environmental Risk Index: =0.6*2.0+0.4*1.0=1.6; Ship interaction feature extraction: Calculated distance to the nearest ship =1.2 nautical miles.

[0055] Predicting the next 10 minutes, Number of ships less than 0.5 nautical miles =1.

[0056] Feature fusion: Input all the above features into the MHSAM model and output a unified 128-dimensional feature vector. .

[0057] S3: Steps for Spatiotemporal Graph Neural Network Analysis Graph construction: Using the target ship as a node, search for all ships within 5 nautical miles of it (a total of 3 ships) as neighbor nodes to construct a spatiotemporal graph.

[0058] Anomaly scoring: unifying feature vectors Input the trained ST-GNN model. The model aggregates the spatiotemporal features of the target ship and its neighboring nodes, and finally outputs a probability score representing the degree of behavioral abnormality: =0.78.

[0059] S4: Dynamic Threshold Early Warning Steps Baseline threshold: based on historical normal data. 99.5th percentile =0.70.

[0060] Dynamic threshold calculation: = ×(1- · =0.70×(1-0.3×1.6)=0.70×(1-0.48)=0.70×0.52=0.364 Early warning decision-making: comparison (0.78)> (0.364), the condition is met. Therefore, the system immediately generates and issues an "abnormal behavior" warning for MMSI123456789, prompting maritime regulatory personnel to intervene.

[0061] 3. Summary of Implementation Examples This embodiment demonstrates the entire process from data fusion, feature extraction, model analysis to final early warning decision through a set of specific input data.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for ship trajectory analysis and anomaly detection based on multi-source data fusion, characterized in that, Includes the following steps: S1: Multi-source data fusion steps: Real-time acquisition of Automatic Identification System (AIS) data, environmental data, and static vessel profile data; spatiotemporal alignment of the data using the Maritime Mobile Service Identifier (MMSI) and Coordinated Universal Time (UTC) as the association keys, and Z-score standardization of all numerical features to form a fused data sequence; the Z-score standardization formula is: ,in These are the original eigenvalues. and The mean and standard deviation of this feature are calculated from the historical training dataset; the environmental data includes at least wind speed, wind direction, wave height, visibility, and surface current speed. S2: Multidimensional feature extraction and fusion steps: extract spatiotemporal features, environmental risk features, and ship interaction features from the fused data sequence; input the three types of features into the fusion model based on the multi-head self-attention mechanism (MHSAM) for adaptive weighting and concatenation, and output a unified feature vector; S3: Spatiotemporal Graph Neural Network Analysis Steps: Input the unified feature vector into the pre-trained spatiotemporal graph neural network (ST-GNN) model; the ST-GNN model uses each ship as a graph node, and the spatial Euclidean distance between ships is less than a preset neighborhood radius. To construct the graph edges, where The value ranges from 3 to 10 nautical miles; the ST-GNN model aggregates the features of a node itself and its neighboring nodes through alternating stacked spatial graph convolutional layers and temporal convolutional layers, and outputs a probability score representing the degree of abnormality in ship behavior. ; S4: Dynamic threshold early warning step: Based on historical normal flight trajectory data, calculate the anomaly probability score output by the ST-GNN model. The distribution is used, and the 99.5th percentile of this distribution is selected as the benchmark threshold. ; Based on the real-time calculated total environmental risk index The baseline threshold is linearly adjusted, resulting in a dynamic threshold. The calculation formula is: ,in This is a preset proportionality coefficient, with a value ranging from 0.1 to 0.5; if If so, an early warning message will be generated and issued.

2. The method for ship trajectory analysis and anomaly detection based on multi-source data fusion according to claim 1, characterized in that: In the multidimensional feature extraction and fusion step: the spatiotemporal features include the average speed calculated based on a time sliding window. Standard deviation of speed Average rate of change of heading Average longitude of trajectory points and average latitude ; The environmental risk characteristics are obtained through the following quantification rules: when the visibility value... Visibility risk index below 5000 meters When the wave height value When the wave height exceeds 2 meters, the wave height risk index is high. ; The overall environmental risk index The weighted sum of the risk indices for each component is calculated using the following formula: ,in and The preset weighting coefficients, and + =1; The ship interaction characteristics include the Euclidean distance to the nearest ship. And the predicted meeting distance based on the current velocity vector within a preset 10-minute time window. Number of ships less than 0.5 nautical miles .

3. The method for ship trajectory analysis and anomaly detection based on multi-source data fusion according to claim 1, characterized in that: The ST-GNN model in the spatiotemporal graph neural network analysis step is trained in the following way: The spatial graph convolutional layer is implemented using a graph attention network (GAT), and the temporal convolutional layer is implemented using a one-dimensional convolutional neural network (TCN). The model minimizes the loss function. Training is performed, and the loss function is... Cross-entropy loss and Regularization loss The sum of ,in This is the preset regularization strength coefficient, and a value of 0.01 is recommended. The training data consists of historical multi-source fusion data and corresponding verified anomaly labels, which are derived from publicly available accident reports and manually labeled anomaly events by maritime authorities.

4. The method for ship trajectory analysis and anomaly detection based on multi-source data fusion according to claim 1, characterized in that: The ST-GNN model is trained collaboratively based on a federated learning framework. The specific training steps include: Multiple local clients use local de-identified data to train their respective ST-GNN models; The global aggregation server periodically collects model weight gradients from each local client and performs a federated averaging algorithm. Aggregate to generate global model weights, where The number of clients participating in the aggregation; Distribute the updated global model weights to each local client; After each round of global aggregation, the model weight hash value of this round of aggregation, the MMSI prefix identifier of the participating client, and the timestamp are written into an immutable blockchain distributed ledger.

5. A ship trajectory analysis and anomaly detection system based on multi-source data fusion, characterized in that: include: The data fusion module is configured to perform the multi-source data fusion step, including storing feature means. and standard deviation Memory; The feature fusion module is configured to perform the multidimensional feature extraction and fusion steps, including a graphics processing unit (GPU) for running the MHSAM fusion model; The behavior analysis module, configured to perform the spatiotemporal graph neural network analysis steps, includes a GPU or tensor processor (TPU) for storing and running the ST-GNN model. The early warning decision module is configured to execute the dynamic threshold early warning step, including calculating the dynamic threshold. And the central processing unit (CPU) that generates early warning information.

6. The ship trajectory analysis and anomaly detection system based on multi-source data fusion according to claim 5, characterized in that: The data fusion module pre-stores the mean values ​​of each feature used for Z-score standardization. and standard deviation The lookup table.

7. The ship trajectory analysis and anomaly detection system based on multi-source data fusion according to claim 5, characterized in that: The preset neighborhood radius in the behavior analysis module In the sea.

8. The ship trajectory analysis and anomaly detection system based on multi-source data fusion according to claim 5, characterized in that: The pre-set proportional coefficient in the early warning decision module .

9. The ship trajectory analysis and anomaly detection system based on multi-source data fusion according to claim 5, characterized in that: It also includes a model training and update subsystem, which includes: Multiple local training clients and a global aggregation server based on a federated learning architecture; A blockchain unit that communicates with the global aggregation server, the blockchain unit being composed of multiple distributed nodes.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the ship trajectory analysis and anomaly detection method based on multi-source data fusion as described in any one of claims 1 to 4.

Images