A meteorological perception-based ship navigation behavior anomaly detection method

Abstract

This invention discloses a method for detecting anomalies in ship navigation behavior based on meteorological perception, belonging to the field of ship management technology. The main steps include: First, spatiotemporal alignment and standardization of ship AIS data and marine meteorological data are performed. Then, static attributes, dynamic trajectories, and meteorological environmental information are feature-encoded, and the dynamic and meteorological features are dynamically fused through a learnable meteorological gating unit. The fused feature sequence is input into a Transformer model to capture long-term temporal dependencies and predict the ship's state at the next moment. Finally, by calculating the prediction residuals and comparing them with an adaptive threshold dynamically generated based on historical error statistical distribution, accurate determination of ship behavior anomalies is achieved. This invention realizes the modeling of the dynamic correlation between ship behavior and the marine meteorological environment, thereby effectively distinguishing between reasonable risk avoidance behavior caused by severe weather and genuine abnormal behavior, significantly reducing the false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of ship management technology, specifically to a method for detecting abnormal ship navigation behavior based on meteorological perception. Background Technology

[0002] Currently, maritime regulatory authorities generally rely on dynamic and static navigation data collected by Automatic Identification Systems (AIS) to monitor and detect anomalies in ship behavior using data-driven methods. Existing technologies are mostly based on the geometric characteristics of ship trajectories, such as changes in position, speed, and heading, employing statistical methods or recurrent neural network models (such as LSTM) for pattern learning and anomaly detection. However, these methods often overlook the substantial impact of marine meteorological environments—such as wind, waves, and currents—on ship navigation decisions. In actual navigation, ships often take proactive measures such as slowing down and changing course when encountering severe sea conditions; these actions are essentially reasonable responses that comply with safety regulations. However, because existing detection models lack the ability to perceive meteorological conditions, they are prone to misjudging such reasonable behaviors as "abnormalities," leading to a large number of false alarms. This not only affects regulatory efficiency but also increases the burden of manual review. Furthermore, existing methods have significant limitations at the data level: AIS data is high-frequency and discrete, while meteorological data is typically provided in hourly, gridded formats. The two are difficult to directly align in terms of time and space, hindering the effective fusion of multi-source data. At the model level, while traditional time-series networks such as LSTM can process sequential data, their ability to capture long-distance dependencies is limited, and most anomaly detection still relies on manually set fixed thresholds, failing to adapt to dynamic changes in different sea areas and weather conditions. Consequently, their generalization ability is insufficient in real-world complex sea conditions. In summary, existing ship behavior anomaly detection technologies suffer from problems such as lack of environmental awareness, difficulties in data fusion, weak model adaptability, and rigid thresholds, making it difficult to achieve accurate and reliable anomaly identification in highly dynamic and changeable marine environments. Summary of the Invention

[0003] To achieve accurate and reliable identification of abnormal ship navigation behavior in highly dynamic and variable marine environments, this invention proposes a method for detecting abnormal ship navigation behavior based on meteorological perception, comprising the following steps: S1: Acquire the AIS data of the target vessel and the corresponding marine meteorological data within the spatiotemporal range, and perform spatiotemporal alignment and standardization preprocessing on the AIS data and marine meteorological data; S2: Input the static attribute information, dynamic trajectory sequence information and environmental information in the AIS data and the marine meteorological data into the corresponding encoders for feature encoding to obtain the corresponding high-dimensional feature vectors; S3: Through a learnable meteorological gating fusion unit, the influence factor of the meteorological feature vector is dynamically calculated, and the encoded dynamic feature vector and the meteorological feature vector are weighted and fused based on the influence factor. S4: Input the fused feature vector sequence into the sequence prediction model based on the Transformer architecture, and use the multi-head attention mechanism to capture the spatiotemporal dependency of ship behavior in order to output the predicted value of the ship's state at the next moment. S5: Calculate the residual between the actual observed value and the predicted value of the ship's status at the current moment. Based on the statistical distribution of the historical predicted residuals, dynamically generate an adaptive judgment threshold. By comparing the residual with the adaptive judgment threshold, determine whether the ship's behavior is abnormal.

[0004] This invention introduces a meteorological gating fusion unit and a Transformer-based sequence prediction model to model the dynamic relationship between ship behavior and the marine meteorological environment. This effectively distinguishes between reasonable risk avoidance behavior caused by severe weather and real abnormal behavior, significantly reducing the false alarm rate.

[0005] Furthermore, in step S1, the spatiotemporal alignment specifically includes: A cubic spline interpolation algorithm was used to perform interpolation alignment of marine meteorological data in the time dimension. A bilinear interpolation algorithm is used to interpolate gridded marine meteorological data to the real-time latitude and longitude position of the ship.

[0006] Further, in step S1, the standardization preprocessing specifically includes: The Z-score normalization method was used to process continuous variables in AIS data and marine meteorological data.

[0007] Furthermore, in step S2, the static attribute information includes the target vessel's length, beam, and draft; the dynamic trajectory sequence information includes the target vessel's latitude and longitude position, speed over land, and heading over land; and the environmental information includes wind speed, wind direction, wave height, and current speed.

[0008] Furthermore, in step S3, the meteorological gating fusion unit calculates the fused feature vector using the following formula: In the formula, The fused feature vector For dynamic feature vectors, For static feature vectors, For meteorological feature vectors, This is element-wise multiplication. It is the Sigmoid activation function. This is the weight matrix.

[0009] Furthermore, in step S4, the sequence prediction model based on the Transformer architecture includes an encoder structure composed of multiple stacked Transformer encoders.

[0010] Furthermore, in step S5, the adaptive judgment threshold is dynamically generated based on the statistical distribution of historical prediction residuals, calculated using the following formula: In the formula, To adaptively determine the threshold, The mean of the historical forecast residuals. The standard deviation of the historical forecast residuals. This is an adjustable sensitivity coefficient.

[0011] Furthermore, in step S5, the step of determining whether the ship's behavior is abnormal by comparing the residual with the adaptive judgment threshold specifically involves: If the component of the residual in any dimension exceeds the adaptive judgment threshold of the corresponding dimension, the ship's behavior is judged to be abnormal.

[0012] Furthermore, in step S2, the feature encoding process includes injecting Gaussian noise into the meteorological feature vector.

[0013] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The present invention proposes a method for detecting abnormal ship navigation behavior based on meteorological perception. By introducing a meteorological gating fusion unit and a sequence prediction model based on Transformer, it realizes the modeling of the dynamic relationship between ship behavior and marine meteorological environment, thereby effectively distinguishing reasonable risk avoidance behavior caused by severe weather from real abnormal behavior and significantly reducing the false alarm rate. (2) By using cubic spline and bilinear interpolation techniques, the problem of spatiotemporal alignment between high-frequency AIS data and low-frequency grid meteorological data was solved, ensuring the continuity and authenticity of the input data; (3) The meteorological gating mechanism adaptively adjusts the weight of meteorological characteristics based on real-time sea conditions, simulates the nonlinear changes in ship decision-making under different environmental conditions, and enhances the system's adaptability to complex meteorological environments. (4) Adaptive thresholds based on the statistical distribution of historical prediction residuals are used instead of traditional fixed thresholds, so that the system can dynamically adjust the judgment criteria according to the actual sea conditions, thereby further improving the robustness and generalization ability of the detection. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of a method for detecting abnormal ship navigation behavior based on meteorological perception. Detailed Implementation

[0015] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0016] Current methods for detecting anomalies in ship navigation primarily rely on data from Automatic Identification Systems (AIS) and employ statistical, machine learning, or deep learning-based approaches to analyze the geometric characteristics of ship trajectories. These methods determine whether ship behavior deviates from normal by identifying abnormal fluctuations in parameters such as position, speed, and heading. However, these methods have a significant limitation: their analysis process is often isolated from the actual marine environment in which the ship operates, failing to incorporate real-time meteorological and hydrological factors such as wind, waves, and currents into the decision-making system. In actual navigation, ship operations are heavily constrained by the environment. For example, when encountering strong winds and waves, captains typically take proactive measures such as slowing down or changing course to mitigate risks. Due to the lack of perception and integration capabilities of current technologies regarding meteorological environments, the system is highly susceptible to misclassifying these adaptive behaviors that comply with safety regulations as "abnormal speed" or "abnormal heading," leading to frequent false alarms on the monitoring platform. This not only increases the workload of maritime regulators but also reduces their efficiency in responding to genuine risk signals. Therefore, enabling anomaly detection systems to possess "environmental perception" capabilities and accurately distinguish between "environmentally driven reasonable behavior" and "genuine abnormal behavior" has become a key challenge in improving the intelligence level of maritime traffic safety supervision. To address the aforementioned problems, this invention proposes a novel method for detecting ship anomalies by deeply integrating meteorological information, such as... Figure 1 As shown, the method mainly includes the following steps: S1: Acquire the AIS data of the target vessel and the corresponding marine meteorological data within the spatiotemporal range, and perform spatiotemporal alignment and standardization preprocessing on the AIS data and marine meteorological data; S2: Input the static attribute information, dynamic trajectory sequence information and environmental information in the AIS data and the marine meteorological data into the corresponding encoders for feature encoding to obtain the corresponding high-dimensional feature vectors; S3: Through a learnable meteorological gating fusion unit, the influence factor of the meteorological feature vector is dynamically calculated, and the encoded dynamic feature vector and the meteorological feature vector are weighted and fused based on the influence factor. S4: Input the fused feature vector sequence into the sequence prediction model based on the Transformer architecture, and use the multi-head attention mechanism to capture the spatiotemporal dependency of ship behavior in order to output the predicted value of the ship's state at the next moment. S5: Calculate the residual between the actual observed value and the predicted value of the ship's status at the current moment. Based on the statistical distribution of the historical predicted residuals, dynamically generate an adaptive judgment threshold. By comparing the residual with the adaptive judgment threshold, determine whether the ship's behavior is abnormal.

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the complete process and technical details of this invention will be described in detail below in conjunction with specific implementation scenarios. This embodiment uses the real-time navigation monitoring of a large container ship by a maritime traffic management system in a certain sea area as a background, systematically demonstrating the entire chain of technical implementation processes from multi-source data acquisition to final anomaly detection and output. Those skilled in the art will understand that the specific details described herein are only for explaining this invention and are not intended to limit it. Any modifications and substitutions made without departing from the principles of this invention should be covered within the protection scope of this invention.

[0018] In practical operation, the maritime traffic management system employing the technical solution of this invention receives two core data streams in parallel: one is a data stream from the Automatic Identification System (AIS) of ships, originating from shore-based AIS base stations or satellites. This data stream is continuously updated at a high frequency (e.g., once per second to every few seconds), recording the ship's identification and real-time movement status. The other is a gridded marine meteorological data stream from the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis database. This data provides globally covered environmental elements such as sea surface wind fields, waves, and ocean currents at a lower frequency (typically once per hour). Taking a target container ship as an example, the system acquires its recent AIS messages, parsing static attribute information such as a length of 200 meters, a beam of 30 meters, and a draft of 10 meters, as well as a dynamic trajectory sequence changing over time, including timestamps, latitude and longitude coordinates, speed over land, and heading over land. Simultaneously, based on the approximate sea area and time range reported by the ship, the system extracts wind speed, wind direction, significant wave height, and current velocity components from the corresponding spatiotemporal grid within the ERA5 database. These two types of data exhibit significant differences in spatiotemporal scale in their original state: AIS data is a high-frequency discrete sequence attached to the locations of moving ships, while ERA5 data is a low-frequency snapshot fixed on regular spatial grid points. To achieve subsequent deep fusion analysis, this scale mismatch must first be addressed.

[0019] To address the aforementioned data heterogeneity, this invention designs a spatiotemporal alignment and standardization preprocessing workflow. The core of time alignment lies in compensating for the insufficient temporal resolution of meteorological data. The system processes each meteorological variable (such as significant wave height) independently, using observations at adjacent hourly times (such as UTC 00:00 and 01:00) as control nodes, and employs a cubic spline interpolation algorithm to construct a smooth time function curve. This algorithm not only ensures that the curve passes through all nodes but also ensures the continuity of the curve's first and second derivatives, thus realistically simulating the continuous evolution of meteorological parameters within an hour, rather than a simple linear transition. When it is necessary to obtain the meteorological value corresponding to a precise AIS timestamp (e.g., UTC 00:30:15), simply substitute that time into the curve function; this method effectively preserves the physical continuity of meteorological changes. Spatial alignment aims to accurately match the gridded meteorological data to the real-time location of the moving ship. Based on the specific latitude and longitude reported by the ship's AIS, the system locates the ERA5 grid cell where the ship is situated and selects the grid points at the four corners of that cell. Subsequently, a bilinear interpolation algorithm was used to perform weighted calculations based on the distances between the ship's position and these four corner points, ultimately obtaining the wind speed, wave height, and other environmental values ​​for the precise location of the ship. At this point, each AIS data point possessed a time-synchronized and spatially matched meteorological environmental label. To eliminate the impact of differences in feature dimensions and numerical ranges on model training, all continuous variables, including speed, heading, wind speed, and wave height, underwent Z-score standardization. This involves calculating the mean and standard deviation of each feature based on the historical training dataset, and then transforming the real-time data by (value - mean) / standard deviation to ensure all input features are within a similar distribution range.

[0020] After achieving precise data alignment and standardization, the system proceeds to the feature encoding and adaptive fusion step. The goal of this step is to transform different types of data into unified, semantically rich, high-dimensional feature representations and dynamically establish a correlation model between ship behavior and environmental factors. To this end, the system constructs three parallel feature encoding pathways. The static feature encoding pathway processes the ship's inherent physical attributes, mapping the three scalar values ​​of length, beam, and draft to a high-dimensional vector (e.g., 64-dimensional) through a fully connected layer. This vector implicitly encodes the ship's dimensions, tonnage, and inherent wind and wave resistance, providing prior knowledge for the model to understand the differences in environmental responses among different ship types. The dynamic feature encoding pathway processes the ship's historical trajectory sequence. The system extracts continuous AIS dynamic data using a sliding time window (e.g., containing the past 30 time steps, each step being 2 minutes). The latitude, longitude, speed, and heading parameters of each time step are concatenated and mapped to feature vectors of the same dimension through a linear projection layer (fully connected layer), thus forming a sequence representing the ship's recent motion pattern. The meteorological feature encoding pathway processes the aligned environmental data, encoding wind, wave, and current parameters for each time step into a high-dimensional meteorological feature sequence in a similar manner. Notably, to enhance the model's robustness to inherent uncertainties and measurement noise in the meteorological data and to prevent overfitting, a Gaussian noise injection technique is introduced at the output of the meteorological feature encoder. Specifically, during model training, a random perturbation with a mean of zero and a small standard deviation is added to the generated meteorological feature vectors. This allows the model to learn more generalizable patterns of environmental behavior associations, rather than simply memorizing specific combinations of meteorological values.

[0021] The feature vectors generated by the above three encoding paths are static feature vectors. Dynamic feature sequences and meteorological characteristic sequences —This data will be fed into one of the core components of this invention: the meteorological gating fusion unit. The design of this unit is inspired by the captain's decision-making and cognitive process during actual navigation, namely, dynamically adjusting attention to environmental threats based on real-time sea conditions. The gating unit uses a learnable weight matrix... , will the current moment , and The concatenated vectors along the feature dimensions undergo a linear transformation, followed by a Sigmoid activation function. Compression produces a scalar gating value g between 0 and 1. The physical meaning of this gating value can be interpreted as "the attention weight the model should assign to the ship's own historical dynamic characteristics under the current integrated state." When the system senses severe weather conditions (such as high waves and strong winds), the meteorological characteristics... The information it carries is crucial for predicting the ship's next move; the gating unit will be trained to output a smaller value. This means that the model will "adopt" more of the environmental constraints indicated by meteorological features when making decisions. Conversely, in calm weather, The value approaches 1, and the model primarily relies on the ship's own motion inertia and historical patterns for prediction. Finally, the feature vectors are fused. Through the formula: , The calculation shows that, among which This represents element-wise multiplication. This mechanism enables data-driven, soft, adaptive feature weighting fusion, giving the model intelligent capabilities for environmental perception and decision-making trade-offs.

[0022] After obtaining the fused feature sequence, the system uses a sequence prediction model based on the Transformer architecture to accurately predict the future state of the ship. In this embodiment, the model consists of a stacked Transformer encoder with six layers. Each encoder layer contains two core sub-layers: a multi-head self-attention mechanism and a feedforward neural network, supplemented by residual connections and layer normalization to ensure the stability and depth of the training process. The multi-head self-attention mechanism allows the model to simultaneously analyze the complex dependencies between time steps within the input sequence from multiple different representation subspaces (i.e., multiple "attention heads"). For ship trajectory prediction tasks, this means that the model can capture multiple spatiotemporal patterns in parallel: for example, one attention head may focus on identifying the stability pattern of the ship maintaining its course during long-distance voyages, another head may specifically detect the periodic deceleration behavior of the ship when approaching the port area, and still others may associate the ship's operational responses under similar weather conditions several hours earlier. Through this mechanism, the Transformer model can effectively learn and memorize the ship's long-term navigation habits and behavioral paradigms in specific environments, overcoming the memory decay problem that traditional recurrent neural networks are prone to in long sequence modeling. After deep processing by multiple Transformer encoders, the contextual information of the sequence is fully extracted. The system then aggregates the information of the entire sequence through an attention pooling layer, or directly uses the feature representation of the last time step, and then decodes it through a prediction head (fully connected layer) to output the prediction for the next time step. Predicted values ​​of critical ship conditions at any given time It typically includes latitude and longitude location, speed over the ground, and heading over the ground.

[0023] After the prediction step is completed, the system enters the final anomaly detection and judgment stage. When Actual AIS observation data at time Upon receipt, the system immediately calculates its value against the predicted value. The residual vector between The residuals intuitively quantify the degree to which the ship's actual behavior deviates from the learned reasonable behavior pattern under the current weather conditions. To scientifically evaluate the residuals, this invention employs an adaptive threshold generation mechanism based on historical performance. The system maintains a sliding window, continuously recording the residuals of each dimension generated over a recent historical period (e.g., the last 1000 predictions). Based on this window data, the mean of the residuals for each state dimension (e.g., longitude, latitude, speed, and heading) is dynamically calculated. and standard deviation Adaptive decision threshold Then according to the formula Dynamically generated. Among them, It is a configurable sensitivity coefficient, typically set between 0 and 1. Maritime regulators can flexibly adjust it according to the risk level, traffic density, or specific regulatory needs of different sea areas. Value: In open, low-risk waters, the value can be appropriately increased. The value can be adjusted to reduce false alarms; in high-risk and sensitive areas such as port entrances and narrow waterways, the value can be lowered. This threshold value is used to improve detection sensitivity and ensure that potential risks are not overlooked. The core advantage of this threshold mechanism lies in its adaptability: when the ship is in severe sea conditions, the uncertainty of the model prediction itself increases, leading to fluctuations in the historical residuals. The natural expansion of the residual vector dynamically raises the anomaly detection threshold. This allows the system to tolerate reasonable but significant deviations in risk avoidance behavior caused by weather, effectively distinguishing between abnormal behavior and reasonable responses. The judgment logic is clear and explicit: the system will use the current residual vector... Each component is compared with the dynamic threshold calculated for the corresponding dimension. If the absolute value of the residual in any dimension exceeds its threshold, the system immediately determines that the ship's behavior is abnormal at that moment and triggers a graded warning (such as highlighting the target ship on the VTS electronic chart, issuing an audible and visual alarm, and recording the anomaly type, time, and location); if the residuals of all dimensions do not exceed the threshold, the ship is determined to be sailing normally.

[0024] In summary, the present invention proposes a method for detecting abnormal ship navigation behavior based on meteorological perception. By introducing a meteorological gating fusion unit and a Transformer-based sequence prediction model, it achieves modeling of the dynamic correlation between ship behavior and the marine meteorological environment. This effectively distinguishes between reasonable risk avoidance behavior caused by severe weather and real abnormal behavior, significantly reducing the false alarm rate.

[0025] By employing cubic spline and bilinear interpolation techniques, the spatiotemporal alignment challenge between high-frequency AIS data and low-frequency gridded meteorological data was resolved, ensuring the continuity and authenticity of the input data. The meteorological gating mechanism adaptively adjusts the weights of meteorological features based on real-time sea conditions, simulating the nonlinear changes in ship decision-making under different environmental conditions and enhancing the system's adaptability to complex meteorological environments. An adaptive threshold based on the statistical distribution of historical prediction residuals is used instead of the traditional fixed threshold, enabling the system to dynamically adjust the judgment criteria according to actual sea conditions, further improving the robustness and scenario generalization ability of the detection.

[0026] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0027] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0028] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0029] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A method for detecting abnormal ship navigation behavior based on meteorological sensing, characterized in that, Including the following steps: S1: Acquire the AIS data of the target vessel and the corresponding marine meteorological data within the spatiotemporal range, and perform spatiotemporal alignment and standardization preprocessing on the AIS data and marine meteorological data; S2: Input the static attribute information, dynamic trajectory sequence information and environmental information in the AIS data and the marine meteorological data into the corresponding encoders for feature encoding to obtain the corresponding high-dimensional feature vectors; S3: Through a learnable meteorological gating fusion unit, the influence factor of the meteorological feature vector is dynamically calculated, and the encoded dynamic feature vector and the meteorological feature vector are weighted and fused based on the influence factor. S4: Input the fused feature vector sequence into the sequence prediction model based on the Transformer architecture, and use the multi-head attention mechanism to capture the spatiotemporal dependency of ship behavior in order to output the predicted value of the ship's state at the next moment. S5: Calculate the residual between the actual observed value and the predicted value of the ship's status at the current moment. Based on the statistical distribution of the historical predicted residuals, dynamically generate an adaptive judgment threshold. By comparing the residual with the adaptive judgment threshold, determine whether the ship's behavior is abnormal.

2. The method for detecting abnormal ship navigation behavior based on meteorological sensing as described in claim 1, characterized in that, In step S1, the spatiotemporal alignment specifically includes: A cubic spline interpolation algorithm was used to perform interpolation alignment of marine meteorological data in the time dimension. A bilinear interpolation algorithm is used to interpolate gridded marine meteorological data to the real-time latitude and longitude position of the ship.

3. The method for detecting abnormal ship navigation behavior based on meteorological sensing as described in claim 1, characterized in that, In step S1, the standardization preprocessing specifically includes: The Z-score standardization method was used to process continuous variables in AIS data and marine meteorological data.

4. The method for detecting abnormal ship navigation behavior based on meteorological sensing as described in claim 1, characterized in that, In step S2, the static attribute information includes the target vessel's length, beam, and draft; the dynamic trajectory sequence information includes the target vessel's latitude and longitude position, speed over land, and heading over land; and the environmental information includes wind speed, wind direction, wave height, and current speed.

5. The method for detecting abnormal ship navigation behavior based on meteorological sensing as described in claim 1, characterized in that, In step S3, the meteorological gating fusion unit calculates the fused feature vector using the following formula: In the formula, The fused feature vector For dynamic feature vectors, For static feature vectors, For meteorological feature vectors, This is element-wise multiplication. It is the Sigmoid activation function. This is the weight matrix.

6. The method for detecting abnormal ship navigation behavior based on meteorological sensing as described in claim 1, characterized in that, In step S4, the sequence prediction model based on the Transformer architecture includes an encoder structure composed of multiple stacked Transformer encoders.

7. The method for detecting abnormal ship navigation behavior based on meteorological sensing as described in claim 1, characterized in that, In step S5, the adaptive judgment threshold is dynamically generated based on the statistical distribution of historical prediction residuals and calculated using the following formula: In the formula, To adaptively determine the threshold, The mean of the historical forecast residuals. The standard deviation of the historical forecast residuals. This is an adjustable sensitivity coefficient.

8. The method for detecting abnormal ship navigation behavior based on meteorological sensing as described in claim 7, characterized in that, In step S5, determining whether ship behavior is abnormal by comparing the residual with an adaptive judgment threshold specifically involves: If the component of the residual in any dimension exceeds the adaptive judgment threshold of the corresponding dimension, the ship's behavior is judged to be abnormal.

9. The method for detecting abnormal ship navigation behavior based on meteorological sensing as described in claim 1, characterized in that, In step S2 The feature encoding process includes injecting Gaussian noise into the meteorological feature vector.

Images