Offshore ship tracking method and device based on multi-source data fusion and storage medium

Abstract

The application discloses a near-shore ship tracking method and device based on multi-source data fusion and a storage medium. The method comprises the following steps: acquiring first radar point cloud data of a wind farm millimeter wave radar, breakwater vibration sensor data, AIS data and second radar point cloud data of a multi-radar station, and processing the data; performing spatial alignment and time alignment on the processed first radar point cloud data and second radar point cloud data, and performing time alignment on the processed AIS data; fusing the aligned data to obtain target state information; predicting multiple states of a ship at a future time based on the target state information, obtaining corresponding multiple predicted trajectories, correcting the predicted trajectories based on a wind farm wake model, and screening an optimal trajectory from the corrected predicted trajectories as a ship track. Thus, the multi-dimensional risk resistance capability can be improved, the safety guarantee is higher, and the supervision efficiency in a complex environment such as a new near-shore wind farm can be improved.

Description

Technical Field

[0001] This application relates to the field of intelligent maritime monitoring technology, and in particular to a nearshore vessel tracking method, device and storage medium based on multi-source data fusion. Background Technology

[0002] In the field of nearshore vessel traffic monitoring, traditional technologies mainly rely on the collaborative operation of a single radar station and an onboard Automatic Identification System (AIS). Due to the linear propagation characteristics of radar waves, the construction of breakwaters and other wind farms along the coast creates persistent monitoring blind spots due to dense wind turbine towers and fixed obstacles, resulting in an exponential increase in target loss rate with the density of obstructions. Simultaneously, AIS systems are susceptible to data distortion caused by human tampering and equipment failure, and their coverage on small vessels such as fishing boats is less than 60%. While existing solutions integrate radar data, AIS data, and video surveillance, the complex nearshore environment makes the monitoring effectiveness of vessel traffic service (VTS) systems susceptible to impacts in these new and complex nearshore environments, leading to safety hazards. Summary of the Invention

[0003] This application provides a nearshore vessel tracking method, device, and storage medium based on multi-source data fusion, which can improve perception accuracy, enhance multi-dimensional risk resistance, and provide higher security. It can effectively solve the problem of continuous vessel tracking in complex nearshore environments and improve regulatory efficiency in complex environments such as new nearshore wind farms.

[0004] In a first aspect, embodiments of this application provide a near-shore vessel tracking method based on multi-source data fusion, including:

[0005] Acquire first-type radar point cloud data from the millimeter-wave radar of the wind farm, vibration sensor data from the breakwater, AIS data, and second-type radar point cloud data from multiple radar stations, and process the first-type radar point cloud data, the vibration sensor data, the AIS data, and the second-type radar point cloud data.

[0006] The processed first-type and second-type radar point cloud data are spatially and temporally aligned, and the processed AIS data is temporally aligned to obtain aligned data; wherein, the spatial alignment is based on geomagnetic fingerprint matching; the temporal alignment is triggered based on the processed vibration sensor data meeting preset conditions;

[0007] The aligned data are then fused to obtain the target state information;

[0008] Based on the target state information, multiple states of the ship at future times are predicted, resulting in multiple predicted trajectories. The predicted trajectories are then corrected based on the wind farm wake model, and the optimal trajectory is selected from the corrected predicted trajectories as the ship's track.

[0009] Secondly, embodiments of this application provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the method provided in embodiments of this application.

[0010] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method provided in embodiments of this application.

[0011] The technical solution provided in this embodiment acquires first-type radar point cloud data from a wind farm millimeter-wave radar, vibration sensor data from a levee slope, AIS data, and second-type radar point cloud data from multiple radar stations. The acquired data is processed by spatially and temporally aligning the processed first-type and second-type radar point cloud data, and temporally aligning the processed AIS data. Spatial alignment is based on geomagnetic fingerprint matching, and temporal alignment is triggered when vibration sensor data meets prediction conditions. The aligned data is then fused to obtain target state information. Based on this target state information, multiple future states of the ship are predicted, resulting in multiple predicted trajectories. These predicted trajectories are then corrected based on a wind farm wake model, and a selection is made from the corrected predicted trajectories. The optimal trajectory is determined as the ship's track. This involves processing, aligning, and fusing multi-source data to improve perception accuracy, enhance multi-dimensional risk resistance, and provide higher safety assurance. Spatial alignment of data based on geomagnetic fingerprint matching enables precise spatial alignment in wind farm environments. The predicted trajectory is then corrected using a wind farm wake model to select the optimal trajectory as the ship's track, taking into account the impact of wind farms on ship tracks. In summary, by processing and fusing multi-source data and using a wind farm wake model to correct the predicted trajectory, perception accuracy can be improved, multi-dimensional risk resistance can be enhanced, and higher safety assurance can be provided. This effectively solves the problem of continuous ship tracking in complex near-shore environments and improves regulatory efficiency in complex environments such as new near-shore wind farms. Attached Figure Description

[0012] Figure 1 A flowchart of a nearshore vessel tracking method based on multi-source data fusion provided for the implementation of this application;

[0013] Figure 2 For the collaborative calibration flowchart;

[0014] Figure 3 To output a ship trajectory flowchart;

[0015] Figure 4 This is a schematic diagram of the system architecture;

[0016] Figure 5 This is a simulation rendering;

[0017] Figure 6 A structural block diagram of a nearshore ship tracking device based on multi-source data fusion is provided for embodiments of this application;

[0018] Figure 7 This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0019] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] Figure 1 This is a flowchart of a nearshore vessel tracking method based on multi-source data fusion provided in an embodiment of this application. The method can be executed by a nearshore vessel tracking device based on multi-source data fusion. The device is located in a VTS system, which can be configured with electronic devices such as computers. The method can be applied to complex nearshore scenarios such as new offshore wind power corridors.

[0021] like Figure 1 As shown, the method provided in this application embodiment may include:

[0022] S110: Acquire the first type of radar point cloud data of the millimeter-wave radar of the wind farm, the vibration sensor data of the breakwater, the AIS data, and the second type of radar point cloud data of multiple radar stations, and process the first type of radar point cloud data, the vibration sensor data, the AIS data, and the second type of radar point cloud data.

[0023] In this embodiment, in complex environments such as new nearshore wind farms, millimeter-wave radar arrays of wind farms, vibration sensors of breakwaters, multiple radar stations, AIS, etc., can be deployed to collect raw data of the target water area. The raw data includes first-type radar point cloud data of the wind farm millimeter-wave radar, vibration sensor data of the breakwater, AIS data, and second-type radar point cloud data of multiple radar stations.

[0024] Among them, millimeter-wave radars are installed in key locations within the wind farm in a regular spatial layout to ensure comprehensive coverage of the entire wind farm; they can collect raw electromagnetic echo signals from the waters surrounding the wind turbines and output first-type radar point cloud data in polar coordinate form. ,in Slope distance It is the azimuth angle. Radial velocity, This refers to the signal-to-noise ratio.

[0025] Optionally, processing the first type of radar point cloud data includes: adding points whose distance to the target point from the wind farm is less than the search radius to the target point's neighborhood, and generating clusters based on the target point's neighborhood; filtering out abnormal size clusters from the clusters, outputting target clusters, and filtering the target clusters; wherein points whose distance to the wind turbines in the wind farm is less than a preset distance do not participate in clustering. Specifically, clustering can be performed using an improved density-based spatial clustering of applications with noise (DBSCAN) clustering algorithm, introducing wind turbine location constraints, i.e., adjusting distance weights based on the wind farm structure, for radar point cloud data clustering to identify valid targets. Specifically, it may include the following steps: 1) Initializing parameters, including setting the neighborhood search radius. ( Dynamically adjusted based on millimeter-wave radar resolution (e.g., based on radar scan period T and ship's maximum speed); minimum number of clustered points. Load wind turbine coordinate database 2) Neighborhood search, for each point Calculate its relationship with all points wind farm distance ;like ,Will join in 3) Cluster expansion, starting from the core point (the number of points in the neighborhood ≥ MinPts), recursively merge points that satisfy the following conditions: Boundary points; marking distances from wind turbine units Points smaller than 3σ are considered occluded false targets and do not participate in clustering; 4) Output results, filter out anomalous size clusters to obtain target clusters, including the centroid coordinates, velocity vectors, and corresponding size estimation information of the target clusters. Optionally, Kalman filtering or extended Kalman filtering can be applied to the target clusters to suppress noise based on the real-time signal-to-noise ratio (e.g., when environmental interference causes a decrease in the signal-to-noise ratio, observation noise can be suppressed by dynamically adjusting the covariance).

[0026] In this embodiment, optionally, the distance to the wind farm is calculated based on the following formula:

[0027] ;

[0028] in, These are the i-th and j-th points in the first type of radar point cloud data, representing the spatial location information detected by the millimeter-wave radar. Each point corresponds to the spatial coordinates of a reflection source (which can be an effective target reflection point, an interference point, or a noise point). These are the balance coefficients, default =0.7, =0.3, balancing the spatial proximity of radar point clouds with the impact of wind turbine obstruction (i.e., controlling the weight of the influence of "Euclidean distance" and "wind turbine position constraint"); Let be the coordinates of the kth wind turbine, and σ be the spatial attenuation factor, which is taken as 1 / 4 of the millimeter-wave radar wavelength; This refers to the number of wind turbine units (which can be understood as the number of wind turbine units within the field of view of millimeter-wave radar). for and The Euclidean distance between them; for and The distance between wind farms.

[0029] in, For the position constraint term of the wind turbine, its essence is a point. to the wind farm Location of each wind turbine The improved clustering algorithm uses a "Gaussian distance weighted sum" to address the unique interference in wind farms. For example, the rotation of wind turbine blades generates numerous radar reflection points (creating noise or false targets). When ships or other valid targets approach wind turbines, traditional distance measurement methods may misclassify them as noise. The improved algorithm incorporates a "wind turbine location constraint" into the clustering process by introducing a wind farm distance calculation method. When a point is close to a wind turbine (a reflection point from the turbine blades), the value of the "wind turbine location constraint" is large, increasing the distance between that point and other normal target points, thus identifying it as noise (and excluding it). When a point is a valid target such as a ship or personnel, if it is far from the wind turbine, the value of the "wind turbine location constraint" is small, and the classification is mainly determined by Euclidean distance, ensuring normal clustering. In other words, the algorithm "understands" that the wind turbine location is an "interference source area," automatically reducing the clustering priority of points in that area and improving the recognition rate of valid targets (such as ships).

[0030] Therefore, by using an improved DBSCAN algorithm for clustering and incorporating a wind farm distance calculation method, interference from wind turbines can be suppressed. Specifically, the traditional DBSCAN clustering algorithm misclassifies the massive number of reflection points generated by the rotation of wind turbine blades as "dense clusters" (due to their short Euclidean distance). After the improvement, these points will be identified as such because they are close to the wind turbine location. The "wind turbine location constraint" is increased by adding distance, which is then judged as noise to prevent clustering results from being overwhelmed by wind turbines. Clustering using the method provided in this application embodiment can also enhance the identification of effective targets. Specifically, for point clouds of targets such as ships and inspection personnel, if they are far from wind turbines, the distance in the wind farm distance calculation formula degenerates to an approximate Euclidean distance, resulting in normal clustering; if they are close to wind turbines (e.g., ships entering the wind farm warning zone), the distance can be adjusted based on the wind turbine location, allowing for "priority identification" of targets abnormally close to wind turbines. Clustering using the method provided in this application embodiment can adapt to various wind farm layouts. Specifically, different wind farms have different numbers and locations of wind turbines; by adjusting... , , It can be flexibly adapted to various wind farm structures, making it more versatile.

[0031] In this embodiment, the processing of vibration sensor data for the breakwater includes filtering and normalization to eliminate noise interference and highlight vibration characteristics. Specifically, a wavelet packet threshold denoising algorithm is used to obtain the denoised vibration signal, and the amplitude and dominant characteristic frequency of the vibration sensor signal at the event trigger time are extracted for subsequent time alignment triggering.

[0032] In this embodiment, the processing of AIS data may include parsing and verification, and extracting the Maritime Mobile Service Identifier (MMSI) and draft of the vessel. Speed And heading information, etc.

[0033] In this embodiment, the processing of the second type of radar point cloud data from multiple radar stations may include coordinate system transformation and time synchronization. Polar coordinate point cloud data of multiple radar stations is obtained through a maritime private network, and the polar coordinates are converted into the WGS84 coordinate system or the UTM unified geographic coordinate system. The time reference error of multiple radar stations is controlled to ≤1ms through the GPS / BeiDou satellite timing system to eliminate the timing misalignment caused by asynchronous sampling.

[0034] S120: Spatially and temporally align the processed first-type radar point cloud data and the second-type radar point cloud data, and temporally align the AIS data to obtain aligned data; wherein, the spatial alignment is based on geomagnetic fingerprint matching; the temporal alignment is triggered based on the processed vibration sensor data meeting preset conditions.

[0035] In this embodiment, spatial and temporal alignment can be performed using a spatiotemporal alignment module. This module can integrate a dual-modal spatiotemporal alignment algorithm, specifically employing a dual-modal spatiotemporal alignment algorithm for both spatial and temporal alignment. Spatial alignment can be based on geomagnetic fingerprint matching; temporal alignment is triggered when the sum of the amplitudes of vibration signals from all vibration sensors exceeds a preset threshold, or when the amplitude of continuously sampled data exceeds a certain threshold. Optionally, spatial and temporal alignment can be performed on first-type and second-type radar point cloud data, and temporal alignment can be performed on AIS data. Spatial alignment can also be understood as spatial calibration, and temporal alignment as temporal calibration. Since radar point cloud data and AIS reference data use different coordinate systems—radar point cloud data using radar polar coordinates and AIS data using a global geographic coordinate system—spatial alignment is necessary to unify data from different coordinate systems into a common reference coordinate system (global geographic coordinate system) for subsequent processing. Millimeter-wave radar, multiple radar stations, and AIS systems are prone to clock drift issues, leading to clock inconsistencies; therefore, temporal alignment of the corresponding data is also required.

[0036] In this embodiment, optionally, the step of spatially aligning and temporally aligning the processed first-type radar point cloud data and the second-type radar point cloud data includes: determining an objective function with the goal of minimizing the comprehensive optimization value of the objective function; solving the objective function to obtain the optimal spatial location and the optimal temporal offset; wherein the objective function includes a cost term for geomagnetic fingerprint similarity and a cost term for spatiotemporal consistency; and performing spatial and temporal alignment on the processed first-type radar point cloud data and the second-type radar point cloud data based on the optimal spatial location and the optimal temporal offset; wherein the geomagnetic fingerprint similarity cost term is determined based on the geomagnetic fingerprint matching.

[0037] The objective function is:

[0038] ;

[0039] in, The spatial location to be solved; The time offset to be solved; and These are the optimization weights, used to balance the importance of the cost term for geomagnetic fingerprint similarity and the cost term for spatiotemporal consistency. The default values ​​are 0.6 and 0.4 respectively; The difference between the observed geomagnetic fingerprint vector at location O and the geomagnetic fingerprint in the geomagnetic fingerprint database (e.g., Mahalanobis distance) is the cost term for geomagnetic fingerprint similarity, which can be determined based on geomagnetic fingerprint matching. Details will be provided later. The smaller the value, the higher the geomagnetic fingerprint matching degree; This is the original timestamp; For the k-th sensor at timestamp... The observed values, for example if It is a millimeter-wave radar or multiple radar stations, which can represent the situation in... The target slant range and azimuth angle observed at all times; Let k be the installation location of the k-th sensor; In the observation model for the k-th sensor, assuming the target location is at point O, the theoretical observation value is given. In the case of multiple radar stations, the known radar station locations are... Then it can be based on the location and The geometric relationship between them is used to calculate the theoretical slant range and azimuth. If the sensor is a millimeter-wave radar or multiple radar stations, It is a geometric mapping function, a function relating the target location at point O. The specific function form can be determined based on the reference position of the k-th sensor (different algorithms apply to different types of sensors). The theoretical tilt moment and azimuth can be calculated using this function. K represents the number of sensors; This is the cost term for spatiotemporal consistency. The smaller this value, the more consistent the observations on multiple sensors are in spatiotemporal space. The sensors include the wind farm millimeter-wave radar or multiple radar stations. GeometricTransform is a geometric transformation mapping function, which refers to the mathematical process of transforming the position of the target in the global coordinate system to the measured values ​​(distance and azimuth) in the local coordinate system of the sensor. It is used to map the position of the target in the global coordinate system to the local position of a specific sensor.

[0040] In this embodiment, spatial alignment and temporal alignment are not completely independent, but rather a joint spatial-temporal optimization objective function is constructed through the implementation of collaborative calibration. To achieve deep fusion and find the optimal solution simultaneously, we can rely more on the spatiotemporal consistency cost of multiple sensors. When the data quality of the sensors is poor, we can rely more on geomagnetic fingerprint matching. By constructing a spatial-temporal joint optimization objective function, we can effectively solve the limitations of single spatial or single temporal alignment.

[0041] In this embodiment, it may further include acquiring the current geomagnetic fingerprint observation value, searching for candidate points within a preset radius in the geomagnetic fingerprint database, determining the similarity between the current geomagnetic fingerprint observation value and the geomagnetic fingerprint of the candidate point, performing particle filtering optimization on the candidate points whose similarity meets preset conditions, outputting the optimal position estimate, and using the optimal position estimate as the initial value of the geomagnetic fingerprint similarity cost term.

[0042] The similarity is determined based on the following formula:

[0043] ;

[0044] in, , This represents the current geomagnetic fingerprint observation value, which is also the current geomagnetic observation vector. These are the geomagnetic intensities along the x, y, and z axes, respectively. The geomagnetic intensity gradient; Candidate point location Geomagnetic fingerprints; This represents the similarity between the current geomagnetic fingerprint observation and the geomagnetic fingerprint of the i-th candidate point; , is the covariance matrix of the geomagnetic fingerprint, used to eliminate differences in dimensions and reflect the reliability of each feature.

[0045] Specifically, current geomagnetic fingerprint observations can be obtained through mobile geomagnetic probes deployed on ships or buoys, or through geomagnetic interpolation at locations initially estimated via data fusion. The geomagnetic fingerprint database can be constructed using grid points (e.g., 1m × 1m resolution) in the wind farm and surrounding waters. Above, geomagnetic fingerprints for each point are generated through measurement or high-precision geomagnetic models. , These can be locations in the geomagnetic fingerprint database. The geomagnetic intensity and geomagnetic intensity gradient along the x, y, and z axes; a search can be performed within a preset radius to obtain candidate points, and the similarity to each candidate point can be calculated. The preset radius can be 500 meters or other values. Filtering optimization can be performed on the top three candidate points ranked by similarity to output the optimal location estimate, which is then used as the initial value for the geomagnetic fingerprint similarity cost term in the objective function. Specifically, the filtering optimization can involve assigning weights to the top three candidate points and performing a weighted summation of their positions to obtain the optimal location estimate. The weights assigned to the candidate points can be calculated based on the following formula:

[0046] ;

[0047] in, This can represent the k-th filtering optimization, the Top-order... i Normalized weights for (i=1,2,3) candidate points; It is the minimum similarity between the current geomagnetic fingerprint and the geomagnetic fingerprint observation among all candidate points; The "standard deviation" of the similarity distribution; The normalization coefficients are for the Gaussian distribution; output the optimal position estimate: .

[0048] In this embodiment, the method provided by this application may further include: when the sum of the amplitudes of the vibration signals of all vibration sensors exceeds a preset threshold, recording the local timestamp of each sensor and obtaining the corresponding reference timestamp, and determining the original time offset of each sensor based on the local timestamp and the reference timestamp; obtaining the global time offset by fitting using the dynamic weighted least squares method, and using the global time offset as the initial value of the time offset to be solved;

[0049] The global time offset is determined based on the following formula:

[0050] ;

[0051] in, For the global time offset; This is the local timestamp of the k-th sensor; This is the reference timestamp for the k-th sensor; The weight corresponding to the k-th sensor;

[0052] ;

[0053] in, Let be the amplitude of the vibration signal from the k-th vibration sensor; Let be the dominant characteristic frequency of the vibration signal of the k-th vibration sensor; where the vibration signal of the vibration sensor is generated by the impact of the ship's wake.

[0054] Specifically, the vibration signal is continuously monitored within a window period (e.g., 5 seconds). The sum of the amplitudes of the vibration signals... Exceeding the preset threshold A valid global synchronization event is determined when the amplitude of the vibration signal is sampled continuously for five consecutive times and the amplitude exceeds a certain threshold; the local timestamp of the sensor (radar, AIS) at the time of the synchronization event is recorded. Get the reference timestamp when the event occurred. (Typically provided by the BeiDou / GPS timing module); calculate the raw time offset of the data stream from each sensor relative to the reference time. The global time bias is obtained by using dynamic weighted least squares fitting. The reason is that not all sensor observations are reliable, data is subject to noise interference, and reliability varies. Therefore, the algorithm needs to best reflect the time deviation caused by real target (ship) events while minimizing random noise and interference. By incorporating ship physical characteristics (amplitude, frequency), the algorithm can intelligently select the most reliable time synchronization information, thus deriving a robust and interference-resistant "global time offset." The weights corresponding to each sensor are as follows: The amplitude of its vibration signal The degree of matching with the typical ship wake frequency (e.g., 2.5Hz) is jointly determined. The calculated global time offset is then used... Correct subsequent data streams, either by using linear interpolation or dynamic adjustment, to correct the timestamps of the data in the data stream. ;in, This is the corrected timestamp; it is recalibrated periodically or when a global synchronization event occurs again to address clock drift.

[0055] In this embodiment, the corresponding time alignment of the AIS data includes: using the global time offset as the optimal time offset to align the AIS data. Specifically, the sensor is identified as AIS, the global time offset is determined using the above method, and the above formula is used... Time alignment is performed on the AIS data. Therefore, by achieving time alignment using the above method, the timing discrepancies caused by the drift of the built-in clocks of various sensors can be resolved.

[0056] In this embodiment, the process of solving the objective function can be as follows, that is, the collaborative calibration process can be referred to Figure 2 Specifically, this may include: S1: Initialization: Load the geomagnetic fingerprint database and sensor network topology, set the clock synchronization period (default 300 seconds), and use the optimal location estimate given by geomagnetic fingerprint matching and the global time offset given by event-driven timestamp correction as the initial values ​​for joint optimization. Set weights. S2. Iterative Optimization: Employing nonlinear least squares optimization algorithms such as Levenberg-Marquardt (LM), the objective function is minimized. In each iteration, a. Spatial perturbation: Based on the optimal position estimate obtained from geomagnetic matching, a theoretical time-biased model is generated, and the model is then used to estimate the current optimal position. a. Minor positional perturbations in the vicinity; b. Temporal perturbations: Correct the sensor data time axis using global time offset, within the current global time offset. a. Generate small temporal perturbations in the vicinity; c. For each set of candidate values ​​for spatial location and global temporal offset, calculate the geomagnetic fingerprint similarity cost term and apply the global temporal offset to the data from all sensors. For each sensor Calculate the corrected observations With at that location Theoretical observations The difference is obtained by weighted summation of the two cost terms. Based on the change in the cost function, determine the next search direction, eventually converging to make... Minimum solution , These are the optimal spatial location and the optimal time offset, respectively; d. Output the optimal spatial location and the optimal time offset. ,Will Global time synchronization for all sensors. S3. Anomaly handling: When When the spatial deviation of a location at a certain moment is greater than 10 meters compared to the previous moment, an anomaly is triggered, causing the geomagnetic fingerprint database to be reloaded; three consecutive instances of time deviation exceeding the limit occur. Hardware clock diagnostics are initiated at the specified time. Spatial alignment provides positional constraints for temporal alignment, while temporal alignment provides timing consistency for spatial matching. The two share gradient information during the optimization process, forming a closed-loop feedback.

[0057] S130: The aligned data is fused to obtain the target state information.

[0058] In this embodiment, the aligned data can be fused using a Dynamic Weighted Fusion Algorithm (DW-STAN) engine. Specifically, the aligned data is processed through an improved spatiotemporal attention network to obtain a dynamic weight matrix, and data fusion is performed based on this dynamic weight matrix to obtain target state information. The input data for the improved spatiotemporal attention network can be found in Table 1, and the fused target state information can be... .in, These are the x-axis position, y-axis position, x-axis velocity, and y-axis velocity of the fused target. This is confidence level information.

[0059] Table 1

[0060]

[0061] As shown in Table 1, The elements in the table are the target's x-axis position, y-axis position, x-axis velocity, y-axis velocity after spatial and temporal alignment, and an initial confidence level based on quality assessment. The elements in the table represent longitude, latitude, speed over ground, heading, draft, and confidence level, which is usually considered a high value (AIS data typically has a confidence level of 0.95, which is usually considered a high fixed value).

[0062] In this embodiment, optionally, fusing the aligned data to obtain target state information includes: processing the aligned data through a spatiotemporal convolutional layer and an attention layer in an improved spatiotemporal attention network to obtain fused spatiotemporal correlation features; processing the spatiotemporal correlation features through an LSTM layer in the improved spatiotemporal attention network to obtain hidden state information; converting the hidden state information into a dynamic weight matrix through a fully connected layer in the improved spatiotemporal attention network; and determining the target state information based on the dynamic weight matrix and the spatiotemporal correlation features; wherein, the target state information includes position information, velocity information, and confidence information;

[0063] The attention coefficients in the attention layer are:

[0064] ;

[0065] in, The attention coefficient for the i-th spatiotemporal location corresponding to the j-th spatiotemporal location; Let i be the query vector for the i-th spatiotemporal location; Let j be the key vector at the j-th spatiotemporal location; Let be the key vector at the k-th spatiotemporal location; It is the "similarity calculation term" in the classic attention mechanism, which uses the query vector based on its spatiotemporal location. and key vector Calculate feature similarity ; where, through linear transformation The corresponding query vector and key vector can be obtained; For vector dimensions; This is a transpose operation. Let be the static visibility probability from the i-th spatial location to the k-th spatial location; The static visibility probability from the i-th spatial location to the j-th spatial location can be interpreted as the probability that there is no obstruction and direct line of sight between a sensor (such as radar) at spatial grid point i and a target at spatial grid point j. It is generally calculated in advance based on a digital map of a wind farm using a ray tracing algorithm, forming a visibility probability matrix. This can be done offline in advance, generating a visibility probability matrix for online querying. The calculation process may include the following steps:

[0066] Loading a 3D digital model of a wind farm This model contains precise geometric information and geographical locations of all wind turbine towers, blades, and other structures. It defines the set of possible spatial locations I for the sensors to be calculated and the set of possible spatial locations J for the targets (typically corresponding to uniform spatial grid points). For each pair of locations to be calculated... Innovate a ray from spatial network point i to spatial network point j Simulate the propagation path of a radar beam or line of sight, and calculate the ray. With wind farm 3D model Check if any of the triangular facets of obstacles intersect. If any intersection exists, the line of sight is considered obstructed, and the view is set to... If there is no intersection, then the line of sight is considered unobstructed, and the position is set... To address factors such as positioning errors, map accuracy, and beamwidth in practical engineering, the Monte Carlo ray tracing method can be used: within a specified range near spatial grid points i and j, multiple ray tracing operations (e.g., ...) can be performed. (Number of random samples) generate multiple slightly jittery rays, and statistical analysis is performed. Number of times the line of sight was unobstructed in the second sampling Calculate the static visibility probability This method yields It is a continuous value between [0, 1], which better reflects the uncertainty of the real world. Here, 1 represents complete unobstructed access, with no possibility of being blocked by wind turbines; 0 represents complete obstruction of the line of sight, with the target (spatial grid point j) absolutely located within the blind zone of the sensor (spatial grid point i); values ​​between 0 and 1 indicate partial obstruction or uncertainty. This is introduced into the calculation of the attention coefficient. This is equivalent to injecting prior physical spatial structure knowledge when calculating the correlation between two locations. It proactively reduces the attention weight from the sensor (spatial grid point i) to the occluded target (spatial grid point j), thus preventing the model from learning correlations that are inherently impossible. This significantly improves the accuracy and reliability of the algorithm in complex occlusion environments. Yes The calculation results are stored as a pre-computed visibility probability lookup table or matrix. During the fusion phase, the visibility probability between any pair of grid points can be obtained by querying this table or matrix. The value of is used as a bias term in the calculation of the attention coefficient to explicitly encode the shading effect of wind turbines. This can actively reduce the sensor weight in the shaded area and effectively solve the problem of inaccurate sensor data in the shaded area.

[0067] In this embodiment, fusing the aligned data includes the following steps: S1: Input preprocessing, fusing the aligned data in a unified spatiotemporal grid ( Encoding is performed within a time window × spatial grid, and normalization is applied accordingly; spatiotemporal feature matrices are extracted. ,in S2: The attention coefficients in the attention layer are calculated using the formula above. These coefficients quantify the importance of different spatiotemporal locations. The query vector for each spatial grid is calculated as follows: (The text then repeats the formula, so the translation stops here.) and key vector Superimposed visibility probability logarithmic terms S3. Dynamic weight matrix generation: The spatiotemporal correlation features output from the attention layer are input into the LSTM layer. The input dimensions of the LSTM layer are: (Signal strength 3D + environmental factors 4D + hidden state 256D); Signal strength characteristics (such as the average signal-to-noise ratio from millimeter-wave radar) (AIS signal strength) The environmental disturbance factor (a scalar value obtained by weighted fusion of meteorological data such as wind speed, wave height, and rainfall intensity); The hidden state of the LSTM (carrying historical weight information) is then used to output a dynamic weight matrix through a fully connected layer, with the dynamic weights allocated as follows: S4. Fusion and Correction: Perform weighted fusion and add a Kalman correction term to compensate for system errors, outputting the fused target state matrix. (Location, velocity, confidence level). Specifically, the Hungarian algorithm, etc., can be used to combine observations from different data sources ( The data is associated with the tracked target to confirm that it originates from the same target; multiple observations associated with the same target are fused, where continuous state variables such as position and velocity are calculated using a weighted average; and classification information (such as ship type) is determined using a weighted voting method, with the overall confidence level... The confidence level is determined by the weights of different data sources and their initial confidence levels, and is calculated based on the following formula: ;

[0068] in, This represents the initial confidence value of each sensor's observation in this instance. For millimeter-wave radar, the quality of the target clusters formed after clustering can be assessed to calculate the initial confidence value. For multiple radar stations, the initial confidence level can be calculated based on the quality assessment of their own sensor observations. For AIS, the confidence level is generally considered high, with a fixed value of 0.95. Based on signal strength characteristics by the LSTM layer and environmental interference The dynamic weight matrix is ​​calculated. The corresponding weight elements. The overall confidence score in the output target state information. The weights and confidence scores determine the final target state matrix. This provides all the necessary information for subsequent tracking algorithms.

[0069] S140: Based on the target state information, predict multiple states of the ship at future times to obtain multiple predicted trajectories. Correct the predicted trajectories based on the wind farm wake model. Select the optimal predicted trajectory from the corrected predicted trajectories as the ship's track.

[0070] In this embodiment, to address uncertainties such as obstruction and clutter, an improved multiple hypothesis tracking (MHT) algorithm can be used to predict multiple predicted trajectories of the vessel based on target state information. Specifically, for each predicted trajectory... Interactive multi-model (IMM) filters can be used based on target state information. Position and velocity in the data predict the state at the next moment; state data Each state from the target state to the predicted next moment can form a corresponding predicted trajectory.

[0071] In this embodiment, optionally, the predicted trajectory is corrected based on the wind farm wake model. Correcting the predicted trajectory based on the wind farm wake model includes: calculating the corresponding predicted heading angle based on the state data in the predicted trajectory; correcting the predicted heading angle based on the wind farm wake model; and updating the state data in the predicted trajectory using the corrected predicted heading angle.

[0072] Specifically, the predicted heading angle can be calculated based on the velocity vector in the state data: ;in, These are the predicted x-coordinate, y-coordinate, x-direction velocity, and y-direction velocity for the next moment. It is the arctangent function in the four quadrants; To predict the heading angle. For each predicted trajectory Calculate the predicted location Downwind wind turbines distance Calculate heading compensation using a wind farm wake model ;in, ; The characteristic length of the wake is often taken as 3 times the tower height of the wind turbine. This is an empirical coefficient, with a value range of 0.05-0.15.

[0073] In this embodiment, the predicted heading angle is corrected: Using the corrected heading angle Update the velocity vector to obtain the correct updated velocity vector. This is for future updates.

[0074] In this embodiment, for the corrected predicted trajectory, prior knowledge (channel depth) can be used to eliminate physically impossible or illegal predicted trajectories, reducing computational complexity and improving tracking accuracy. Specifically, for each predicted trajectory... Get its current position Query the water depth of the electronic nautical chart at this location. Calculate the compliance score The compliance constraints are as follows:

[0075] ;

[0076] Among them, all (hard constraint pruning) were removed. A predicted trajectory, for example, if a ship's draft is greater than the usable depth, implying a risk of grounding, is physically impossible; among which, The draft can be extracted from AIS data; The real-time tide height (from the tidal model).

[0077] In this embodiment, after removing non-compliant predicted trajectories, compliant trajectories are obtained. The most probable trajectories are then selected as the final output, representing the ship's path. Specifically, for each compliant trajectory... Calculate the Bayesian posterior probability ,in It is the observational likelihood; The compliance score is either 0 or 1. It is a dynamic model likelihood used to evaluate the smoothness of a trajectory (e.g.: Select the compliant trajectory with the highest posterior probability. As the final tracking result; can be The status is converted into a standardized track data stream and output to the VTS system. The output data may display: timestamp (UTC); position: {longitude, latitude} (WGS84); sog, cog: ground speed and heading; confidence: confidence information; navigational_status: compliance flag. For details on the process of outputting the ship's track, please refer to [link / reference needed]. Figure 3 .

[0078] The technical solution provided in this embodiment acquires first-type radar point cloud data from a wind farm millimeter-wave radar, vibration sensor data from a levee slope, AIS data, and second-type radar point cloud data from multiple radar stations. The acquired data is processed by spatially and temporally aligning the processed first-type and second-type radar point cloud data, and temporally aligning the processed AIS data. Spatial alignment is based on geomagnetic fingerprint matching, and temporal alignment is triggered when vibration sensor data meets prediction conditions. The aligned data is then fused to obtain target state information. Based on this target state information, multiple future states of the ship are predicted, resulting in multiple predicted trajectories. These predicted trajectories are then corrected based on a wind farm wake model, and a selection is made from the corrected predicted trajectories. The optimal trajectory is determined as the ship's track. This involves processing, aligning, and fusing multi-source data to improve perception accuracy, enhance multi-dimensional risk resistance, and provide higher safety assurance. Spatial alignment of data based on geomagnetic fingerprint matching enables precise spatial alignment in wind farm environments. The predicted trajectory is then corrected using a wind farm wake model to select the optimal trajectory as the ship's track, taking into account the impact of wind farms on ship tracks. In summary, by processing and fusing multi-source data and using a wind farm wake model to correct the predicted trajectory, perception accuracy can be improved, multi-dimensional risk resistance can be enhanced, and higher safety assurance can be provided. This effectively solves the problem of continuous ship tracking in complex near-shore environments and improves regulatory efficiency in complex environments such as new near-shore wind farms.

[0079] To more clearly illustrate the method provided in this application's embodiments, an exemplary case is a port surrounded by a wind farm with 12 wind turbines (towers 80m high, spaced 300-500m apart); a breakwater 2.3km long (equipped with 46 vibration sensors); and three radar stations (traditional shore-based radar stations) with a coverage radius of 15 nautical miles. The traffic density is an average daily vessel traffic volume of >200 vessels (including fishing boats, cargo ships, wind power maintenance vessels, etc.). The system architecture can be referenced from... Figure 4 The data collected from millimeter-wave radar, embankment vibration sensors, AIS, and multiple radar stations can be processed through distributed edge computing nodes. Spatial and temporal alignment can be performed by a spatiotemporal alignment module. The aligned data fusion can be performed by the DW-STAN engine. The ship track is obtained through a track conflict resolver and displayed through the VTS system. The specific deployment is as follows: S1 perception layer deployment: The deployment and configuration of various sensors are shown in Table 2.

[0080] Table 2

[0081] Sensor type Deployment parameters Data characteristics millimeter-wave radar Eight nodes are arranged in a ring around the perimeter of the wind farm, spaced 350m apart, with an installation height of +8.5m (5.5m above the highest tide level). Scanning cycle 2 seconds, resolution 0.5°, detection range 5 nautical miles. Vibration sensors for breakwaters One triaxial vibration sensor is installed every 50 meters (46 in total), with a sampling rate of 100Hz. Detection frequency band: 0.5-5Hz; sensitivity: 0.01g Multiple radar stations It connects to three S-band radar stations in the port area (wavelength 10cm) and outputs polar coordinate point clouds (update rate 12 revolutions / minute). Data latency <200ms, distance accuracy ±15m

[0082] S2: Processing Layer Configuration: 1) Edge computing nodes are deployed in the wind farm control center and port control center. 2) Key algorithm parameters are set as follows: the attention dimension of the improved spatiotemporal attention network is 256-dimensional, the LSTM layer has 128 units; the MHT algorithm retains 5 predicted trajectories, and the compliance threshold is... 3) The spatiotemporal alignment cycle is: joint optimization calibration is performed every 5 minutes.

[0083] S3: Ship tracking using a 100,000-ton bulk carrier as an example:

[0084] 1) Initial target acquisition (t=0s): A target was detected 1.2 nautical miles northeast of wind turbine T03. 253 radar point cloud data points (SNR=9.2dB) were generated. Four effective target clusters (sizes 28m, 185m, etc.) were generated using the improved DBSCAN clustering algorithm, and two wind turbine reflection false targets (size <10m) were filtered out. MMSI (413789000) was correlated with AIS data for matching to obtain the ship's draft. .

[0085] 2) Perform spatiotemporal alignment (t=2s) and geomagnetic fingerprint matching: the current geomagnetic fingerprint observation value is Matched to wind turbine T07, at this time the error Better than the threshold of 5.0; event-driven time calibration was performed: characteristic frequency was detected by sensor No. 23 on the breakwater. Amplitude (Exceeding the threshold) ); Calculate global time bias Correct the data timestamp.

[0086] S4: Perform data fusion (t=4s): The attention coefficient distribution is shown in Table 3 below.

[0087] Table 3

[0088] Data source Attention coefficient Based on characteristics millimeter-wave radar 0.62 High signal-to-noise ratio (9.2dB), short range (1.2 nautical miles) Vibration sensors for breakwaters 0.18 Medium amplitude (0.82), frequency matching (2.1Hz) Port Area Radar Station 0.2 Low resolution (±15m), long distance (8.5 nautical miles)

[0089] Fusion output =[Longitude 122.31°, Latitude 30.12°, Speed ​​12.3 knots, Heading 135°], Confidence level 0.87.

[0090] S5: Trajectory Management: 1) Wake Compensation: Wind Speed Ship speed Festival Distance from wind turbine T07 ; Calculate heading compensation 2) Compliance check: Current real-time tide height Electronic nautical chart water depth ;check Mark compliance;

[0091] S6. Track Output: Update the VTS system display and trigger the channel passage permission signal.

[0092] The comparison of results showed improvements in several aspects, including "target loss rate", "heading estimation error", "violation detection rate", and "system robustness". The evaluation results are shown in Table 4. The simulated results can be used as a reference. Figure 5 .

[0093] Table 4

[0094]

[0095] Figure 6 This is a structural block diagram of a nearshore vessel tracking method device based on multi-source data fusion provided in an embodiment of this application, as shown below. Figure 6 As shown, the device includes:

[0096] The acquisition module 610 is used to acquire first type radar point cloud data of the wind farm millimeter-wave radar, vibration sensor data of the breakwater, AIS data, and second type radar point cloud data of multiple radar stations, and to process the first type radar point cloud data, the vibration sensor data, the AIS data, and the second type radar point cloud data.

[0097] The spatiotemporal alignment module 620 is used to spatially and temporally align the processed first-type radar point cloud data and the second-type radar point cloud data, and to temporally align the processed AIS data to obtain aligned data; wherein, the spatial alignment is based on geomagnetic fingerprint matching; the temporal alignment is triggered based on the processed vibration sensor data meeting preset conditions;

[0098] The fusion module 630 is used to fuse the aligned data to obtain the target state information;

[0099] The trajectory determination module 640 is used to predict multiple states of the ship at future times based on the target state information, obtain multiple corresponding predicted trajectories, correct the predicted trajectories based on the wind farm wake model, and select the optimal trajectory from the corrected predicted trajectories as the ship's trajectory.

[0100] In an optional embodiment, processing the first type of radar point cloud data includes:

[0101] For the first type of radar point cloud data, points whose wind farm distance from the target point is less than the search radius are added to the neighborhood of the target point, and a cluster is generated based on the neighborhood of the target point;

[0102] Abnormally sized clusters are filtered from the clusters to output a target cluster, and the target cluster is then filtered; points whose distance from the wind turbine is less than a preset distance are not included in the clustering.

[0103] In an optional embodiment, the distance to the wind farm is calculated based on the following formula:

[0104] ;

[0105] in, These are the i-th and j-th points in the first type of radar point cloud data, respectively; These are the balance coefficients; Let be the coordinates of the kth wind turbine, and σ be the spatial attenuation factor; This refers to the number of wind turbine units; for and The Euclidean distance between them; for and The distance between wind farms.

[0106] In an optional embodiment, the step of spatially aligning and temporally aligning the processed first type of radar point cloud data and the second type of radar point cloud data includes:

[0107] A target function is determined, and the overall optimization value of the target function is minimized. The target function is solved to obtain the optimal spatial location and the optimal temporal offset. The target function includes a cost term for geomagnetic fingerprint similarity and a cost term for spatiotemporal consistency.

[0108] Based on the optimal spatial location and the optimal temporal offset, spatial and temporal alignment are performed on the processed first-type radar point cloud data and second-type radar point cloud data; wherein, the geomagnetic fingerprint similarity cost term is determined based on the geomagnetic fingerprint matching;

[0109] The objective function is:

[0110] ;

[0111] in, The spatial location to be solved; The time offset to be solved; and These are optimization weights; The difference between the geomagnetic fingerprint vector actually observed at location O and the geomagnetic fingerprint in the geomagnetic fingerprint database is the cost term for geomagnetic fingerprint similarity. This is the original timestamp; For the k-th sensor Corrected observations; Let k be the installation location of the k-th sensor; In the observation model of the k-th sensor, the theoretical observation value is assumed to be located at point O; K is the number of sensors. The cost of spatiotemporal consistency; wherein the sensor includes the wind farm millimeter-wave radar or multiple radar stations.

[0112] In an alternative embodiment, the initialization module is configured to:

[0113] Obtain the current geomagnetic fingerprint observation value and search for candidate points within a preset radius in the geomagnetic fingerprint database;

[0114] Determine the similarity between the current geomagnetic fingerprint observation and the geomagnetic fingerprint of the candidate point, perform particle filtering optimization on the candidate points whose similarity meets the preset conditions, output the optimal position estimate, and use the optimal position estimate as the initial value of the spatial position to be solved in the geomagnetic fingerprint similarity cost term;

[0115] The similarity is determined based on the following formula:

[0116] ;

[0117] in, , This represents the current geomagnetic fingerprint observation value, which is also the current geomagnetic observation vector. These are the geomagnetic intensities along the x, y, and z axes, respectively. The geomagnetic intensity gradient; Candidate point location Geomagnetic fingerprints; Let be the similarity between the current geomagnetic fingerprint observation and the geomagnetic fingerprint of the i-th candidate point.

[0118] In an optional embodiment, the time alignment is triggered based on processed vibration sensor data meeting preset conditions, including:

[0119] Trigger time alignment when the sum of the vibration signals from all vibration sensors exceeds a preset threshold;

[0120] The initialization module is also used for:

[0121] If the sum of the amplitudes of the vibration signals from all vibration sensors exceeds a preset threshold, record the local timestamp of each sensor, obtain the corresponding reference timestamp, and determine the original time offset of each sensor based on the local timestamp and the reference timestamp.

[0122] The global time offset is obtained by fitting using the dynamic weighted least squares method, and the global time offset is used as the initial value of the time offset to be solved.

[0123] Accordingly, time alignment is performed on the AIS data, including:

[0124] The global time offset is used as the optimal time offset to align the AIS data in time.

[0125] The global time offset is determined based on the following formula:

[0126] ;

[0127] in, For the global time offset; This is the local timestamp of the k-th sensor; This is the reference timestamp for the k-th sensor; The weight corresponding to the k-th sensor;

[0128] ;

[0129] in, Let be the amplitude of the vibration signal from the k-th vibration sensor; is the dominant characteristic frequency of the vibration signal of the k-th vibration sensor.

[0130] In an optional embodiment, fusing the aligned data to obtain target state information includes:

[0131] The aligned data is processed separately through the spatiotemporal convolutional layer and attention layer in the improved spatiotemporal attention network to obtain fused spatiotemporal correlation features;

[0132] The spatiotemporal correlation features are processed by the LSTM layer in the improved spatiotemporal attention network to obtain hidden state information;

[0133] The hidden state information is converted into a dynamic weight matrix through a fully connected layer in the improved spatiotemporal attention network, and the target state information is determined based on the dynamic weight matrix and the spatiotemporal correlation features; wherein, the target state information includes position information, velocity information, and confidence level;

[0134] The attention coefficients in the attention layer are:

[0135] ;

[0136] in, The attention coefficient for the i-th spatiotemporal location corresponding to the j-th spatiotemporal location; Let i be the query vector for the i-th spatiotemporal location; Let j be the key vector at the j-th spatiotemporal location; Let d be the visibility probability from the i-th spatial location to the j-th spatial location, and d be the vector dimension.

[0137] In an optional embodiment, the predicted trajectory is corrected based on a wind farm wake model, including:

[0138] Calculate the corresponding predicted heading angle based on the state data in the predicted trajectory;

[0139] The predicted heading angle is corrected based on the wind farm wake model, and the state data in the predicted trajectory is updated using the corrected predicted heading angle.

[0140] like Figure 7 As shown in the figure, this application provides an electronic device, including a processor 111, a communication interface 112, a memory 113, and a communication bus 114, wherein the processor 111, the communication interface 112, and the memory 113 communicate with each other through the communication bus 114.

[0141] Memory 113 is used to store computer programs;

[0142] In one embodiment of this application, when the processor 111 executes a program stored in the memory 113, it implements the method provided in any of the foregoing method embodiments, including:

[0143] Acquire first-type radar point cloud data from the millimeter-wave radar of the wind farm, vibration sensor data from the breakwater, AIS data, and second-type radar point cloud data from multiple radar stations, and process the first-type radar point cloud data, the vibration sensor data, the AIS data, and the second-type radar point cloud data.

[0144] The processed first-type and second-type radar point cloud data are spatially and temporally aligned, and the processed AIS data is temporally aligned to obtain aligned data; wherein, the spatial alignment is based on geomagnetic fingerprint matching; the temporal alignment is triggered based on the processed vibration sensor data meeting preset conditions;

[0145] The aligned data are then fused to obtain the target state information;

[0146] Based on the target state information, multiple states of the ship at future times are predicted, resulting in multiple predicted trajectories. The predicted trajectories are then corrected based on the wind farm wake model, and the optimal trajectory is selected from the corrected predicted trajectories as the ship's track.

[0147] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method provided in any of the foregoing method embodiments.

[0148] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0149] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0150] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this application.

Claims

1. A nearshore vessel tracking method based on multi-source data fusion, characterized in that, include: Acquire first-type radar point cloud data from the millimeter-wave radar of the wind farm, vibration sensor data from the breakwater, AIS data, and second-type radar point cloud data from multiple radar stations, and process the first-type radar point cloud data, the vibration sensor data, the AIS data, and the second-type radar point cloud data. The processed first-type and second-type radar point cloud data are spatially and temporally aligned, and the processed AIS data is temporally aligned to obtain aligned data; wherein, the spatial alignment is based on geomagnetic fingerprint matching; the temporal alignment is triggered based on the processed vibration sensor data meeting preset conditions; The aligned data are then fused to obtain the target state information; Based on the target state information, multiple states of the ship at future moments are predicted, resulting in multiple predicted trajectories. The predicted trajectories are then corrected based on the wind farm wake model, and the optimal trajectory is selected from the corrected predicted trajectories as the ship's track. The step of spatially aligning and temporally aligning the processed first-type radar point cloud data and the second-type radar point cloud data includes: A target function is determined, and the overall optimization value of the target function is minimized. The target function is solved to obtain the optimal spatial location and the optimal temporal offset. The target function includes a cost term for geomagnetic fingerprint similarity and a cost term for spatiotemporal consistency. Based on the optimal spatial location and the optimal temporal offset, spatial and temporal alignment are performed on the processed first-type radar point cloud data and second-type radar point cloud data; wherein, the geomagnetic fingerprint similarity cost term is determined based on the geomagnetic fingerprint matching; The objective function is: ； in, The spatial location to be solved; The time offset to be solved; and These are optimization weights; The difference between the geomagnetic fingerprint vector actually observed at location O and the geomagnetic fingerprint in the geomagnetic fingerprint database is the cost term for geomagnetic fingerprint similarity. This is the original timestamp; For the k-th sensor Corrected observations; Let k be the installation location of the k-th sensor; In the observation model of the k-th sensor, the theoretical observation value is assumed to be located at point O; K is the number of sensors. The cost of spatiotemporal consistency; wherein, the sensors include the wind farm millimeter-wave radar, or multiple radar stations; The time alignment is triggered based on the processed vibration sensor data meeting preset conditions, including: Trigger time alignment when the sum of the vibration signals from all vibration sensors exceeds a preset threshold; The method further includes: If the sum of the amplitudes of the vibration signals from all vibration sensors exceeds a preset threshold, record the local timestamp of each sensor, obtain the corresponding reference timestamp, and determine the original time offset of each sensor based on the local timestamp and the reference timestamp. The global time offset is obtained by fitting using the dynamic weighted least squares method, and the global time offset is used as the initial value of the time offset to be solved. Accordingly, time alignment is performed on the AIS data, including: The global time offset is used as the optimal time offset to align the AIS data in time. The global time offset is determined based on the following formula: ； in, For the global time offset; This is the local timestamp of the k-th sensor; This is the reference timestamp for the k-th sensor; The weight corresponding to the k-th sensor; ； in, Let be the amplitude of the vibration signal from the k-th vibration sensor; is the dominant characteristic frequency of the vibration signal of the k-th vibration sensor.

2. The method according to claim 1, characterized in that, Processing the first type of radar point cloud data includes: For the first type of radar point cloud data, points whose distance from the target point to the wind farm is less than the search radius are added to the neighborhood of the target point, and a cluster is generated based on the neighborhood of the target point. Filter out clusters of abnormal size from the clusters, output the target cluster, and filter the target cluster; wherein, points whose distance from the wind turbine is less than a preset distance do not participate in the clustering.

3. The method according to claim 2, characterized in that, The distance to the wind farm is calculated based on the following formula: ； in, These are the i-th and j-th points in the first type of radar point cloud data, respectively; These are the balance coefficients; Let be the coordinates of the k-th wind turbine; σ is the spatial attenuation factor; This refers to the number of wind turbine units; for and The Euclidean distance between them; for and The distance between wind farms.

4. The method according to claim 1, characterized in that, Also includes: Obtain the current geomagnetic fingerprint observation value and search for candidate points within a preset radius in the geomagnetic fingerprint database; The similarity between the current geomagnetic fingerprint observation and the geomagnetic fingerprint of the candidate point is determined. Particle filtering optimization is performed on the candidate points whose similarity meets the preset conditions to output the optimal position estimate. The optimal position estimate is used as the initial value of the spatial position to be solved in the geomagnetic fingerprint similarity cost term. The similarity is determined based on the following formula: ； in, , This represents the current geomagnetic fingerprint observation value, which is also the current geomagnetic observation vector. These are the geomagnetic intensities along the x, y, and z axes, respectively. The geomagnetic intensity gradient; Candidate point location Geomagnetic fingerprints; Let be the similarity between the current geomagnetic fingerprint observation and the geomagnetic fingerprint of the i-th candidate point.

5. The method according to claim 1, characterized in that, The process of fusing the aligned data to obtain target state information includes: The aligned data is processed separately through the spatiotemporal convolutional layer and attention layer in the improved spatiotemporal attention network to obtain fused spatiotemporal correlation features; The spatiotemporal correlation features are processed by the LSTM layer in the improved spatiotemporal attention network to obtain hidden state information; The hidden state information is converted into a dynamic weight matrix through a fully connected layer in the improved spatiotemporal attention network, and the target state information is determined based on the dynamic weight matrix and the spatiotemporal correlation features; wherein, the target state information includes position information, velocity information, and confidence level; The attention coefficients in the attention layer are: ； in, The attention coefficient for the i-th spatiotemporal location corresponding to the j-th spatiotemporal location; Let i be the query vector for the i-th spatiotemporal location; Let j be the key vector at the j-th spatiotemporal location; Let be the key vector at the k-th spatiotemporal location; Let be the static visibility probability from the i-th spatial location to the j-th spatial location; d represents the static visibility probability from the i-th spatial location to the k-th spatial location; d is the vector dimension.

6. The method according to claim 1, characterized in that, The predicted trajectory is corrected based on the wind farm wake model, including: Calculate the corresponding predicted heading angle based on the state data in the predicted trajectory; The predicted heading angle is corrected based on the wind farm wake model, and the state data in the predicted trajectory is updated using the corrected predicted heading angle.

7. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed in a computer, causes the computer to perform the method described in any one of claims 1-6.

Images