A ship standard height wind speed correction model construction method, medium and system

Abstract

The present application provides a kind of ship standard height wind speed correction model construction method, medium and system, belong to ship standard height wind speed correction model construction technical field, the present application is by multi-antenna beidou array acquisition carrier phase difference observation data, space baseline network is modeled as weighted undirected graph, introduce wave age parameter dynamic adjustment zanok coefficient and impose gikhonov regularization constraint to obtain sea roughness, fusion obtains the altitude of wind sensor;Original relative wind speed is input into neural-kalman hybrid filter network, covariance self-adaptive estimation is estimated using artificial intelligence, and inner loop is refined iteration to complete nonlinear folding of logarithmic wind profile, and finally corrected wind speed is output after adaptive kalman filtering, solve the technical problem that wind speed height folding error is difficult to real-time accurate compensation caused by dynamic change of ship wind sensor height.

Description

Technical Field

[0001] This invention belongs to the technical field of ship standard altitude wind speed correction model construction, specifically, it relates to a method, medium and system for constructing a ship standard altitude wind speed correction model. Background Technology

[0002] In the field of shipborne meteorological observation, the standardization of wind speed conversion requires uniformly converting the measured wind speed from sensors to a 10-meter reference height to eliminate systematic biases caused by differences in installation height. Traditional methods rely on a fixed installation height and a logarithmic wind profile model, combined with ship speed vector decomposition to obtain the true wind speed. This method can yield relatively reliable conversion results under static conditions such as calm seas or near-shore anchoring.

[0003] However, under dynamic sea conditions such as swells and strong winds, the ship continuously experiences rolling, pitching, and heave movements. The wind sensors exhibit significant instantaneous altitude deviations as the ship's attitude changes, rendering the traditional fixed altitude assumption invalid. Simultaneously, sea surface roughness dynamically changes with wave age, and a strong nonlinear coupling exists between the Chanok coefficient and frictional wind speed. Traditional static parameterization methods are prone to iterative divergence during rapid sea state transitions. Furthermore, the attitude calculation accuracy of the shipborne global navigation satellite system decreases under multipath interference environments, further amplifying altitude conversion errors.

[0004] In existing technologies, due to the superposition of multiple factors such as dynamic changes in ship attitude, nonlinear coupling of sea surface roughness, and multipath interference, the instantaneous height of wind sensors cannot be accurately estimated in real time. Furthermore, there are continuous deviations in the height input during the logarithmic wind profile conversion process, making it difficult to meet the engineering requirements of real-time correction accuracy for wind speed at a standard 10-meter height under dynamic sea conditions. In other words, existing technologies suffer from the technical problem of difficulty in accurately compensating for wind speed-height conversion errors caused by dynamic changes in ship wind sensor height in real time. Summary of the Invention

[0005] In view of this, the present invention provides a method, medium and system for constructing a ship standard height wind speed correction model, which can solve the technical problem in the prior art that the wind speed height conversion error caused by the dynamic change of the ship's wind sensor height is difficult to compensate for in real time and accurately.

[0006] The present invention is implemented as follows: The first aspect of the present invention provides a method for constructing a ship standard altitude wind speed correction model, comprising the following steps: The system utilizes a multi-antenna BeiDou array to synchronously acquire absolute elevation data of the main antenna and phase differential observation data of multiple baseline carriers, and simultaneously records the original relative wind speed. The spatial baseline network of the multi-antenna BeiDou array is modeled as a weighted undirected graph. The graph Laplace matrix is ​​constructed, the generalized eigenvalue problem is solved, the ship's roll and pitch angles are recovered, and the rotation matrix is ​​constructed. Establish the spatial geometric vectors of the antenna array and the wind sensor in the carrier coordinate system, and calculate the instantaneous vertical deviation of the wind sensor center by combining the rotation matrix; The minimum description length candidate model is evaluated for the vertical motion within the current time window, and the vertical displacement prediction value is output. The absolute elevation of the main antenna is fused with the instantaneous vertical deviation and the vertical displacement prediction value to obtain the altitude of the wind sensor. The wave age parameter is introduced to dynamically adjust the Chanok coefficient, and Tikhonov regularization constraint is applied to the sea surface roughness iterative equation to obtain the sea surface roughness. The raw relative wind speed at the altitude of the wind sensor is input into the neural-Kalman hybrid filter network, and the smoothed standard 10-meter height wind speed is output. Using the geometric accuracy factor of the BeiDou satellite as the observation weight, an adaptive Kalman filter is applied to the standard 10-meter height wind speed sequence to output the final corrected wind speed.

[0007] The multi-antenna BeiDou array consists of no less than three BeiDou receiving antennas. The antenna spatial layout is configured according to the three-axis direction of the carrier coordinate system. The absolute elevation of the main antenna is obtained by precise single-point positioning calculation of the BeiDou main antenna. The original relative wind speed is directly output by the shipborne wind sensor.

[0008] The weighted undirected graph uses each BeiDou receiving antenna as a vertex and the baseline between any two BeiDou receiving antennas as an edge. The edge weight is determined by the signal-to-noise ratio and geometric strength comprehensive score of the carrier phase difference observations of the corresponding baseline. The comprehensive score is obtained by weighted least squares fitting to obtain the signal-to-noise ratio weight coefficient and the geometric strength weight coefficient.

[0009] Wherein, the Graph Laplacian matrix satisfy ,in For degree matrix, Let be the edge weight matrix; the generalized eigenvalue problem is... Take the three eigenvectors corresponding to the second to fourth smallest eigenvalues, and form a low-dimensional embedding space for the attitude parameter manifold. The roll and pitch angles are recovered from the eigenvector coordinate system through least squares mapping.

[0010] The steps for solving the generalized eigenvalue problem also include adopting a partial ambiguity fixing strategy, prioritizing the fixing of the baseline subset with the highest comprehensive score, fixing integer ambiguities by combining an improved least squares ambiguity decorrelation adjustment search tree pruning algorithm, and using an inter-epoch ambiguity inheritance mechanism to directly inherit the fixed baseline solution that is not affected by the cycle slip after the cycle slip.

[0011] The steps for solving the generalized eigenvalue problem include establishing a multipath hemispherical map, statistically analyzing the average carrier phase residuals of each BeiDou receiving antenna according to the two-dimensional grid of satellite elevation and azimuth angles to form a multipath prior template, subtracting the corresponding multipath prior template value from the table during real-time calculation, and using adaptive spectral subtraction to suppress the remaining residuals in the frequency domain.

[0012] The selection of candidate models with minimum description length is performed within a sliding time window. The candidate model set includes three types: sinusoidal superposition model, autoregressive model, and piecewise linear model. The minimum description length criterion value is composed of the sum of the model parameter encoding length and the data condition encoding length. The optimal vertical motion model is used to extrapolate and predict the vertical displacement prediction value of the next epoch.

[0013] The wave age parameter is defined as the ratio of peak wave velocity to frictional wind speed. The piecewise functional relationship between the Channock coefficient and the wave age parameter is obtained by fitting peak wave velocity and frictional wind speed data under various typical sea conditions. The peak wave velocity is provided in real time by a shipboard wave measuring radar or buoy.

[0014] The neural-Kalman hybrid filter network consists of four parts: a long short-term memory module, a linear projection layer, a differentiable Kalman filter layer, and an attention gating module. The hidden state of the long short-term memory module is mapped to the process noise covariance matrix and the observation noise covariance matrix through two independent linear projection layers, respectively. The output of the projection layer is guaranteed to be positive definite by the Softplus activation function.

[0015] The attention gating module monitors the sea state change characteristics identified by the long short-term memory module. When the change confidence exceeds the dynamic change threshold, the gating signal triggers the covariance matrix fast reset submodule to perform a weighted reset on the process noise covariance matrix and the observation noise covariance matrix.

[0016] The dynamic mutation threshold is determined by a resource dynamic adjustment function. The resource dynamic adjustment function calculates a comprehensive state index based on three inputs: the geometric accuracy factor of the Beidou satellite, the standard deviation of the carrier phase residual, and the current sea state level. The comprehensive state index is normalized to the interval of 0 to 1, and the covariance matrix is ​​configured to reset the weights for the three dynamic mutation threshold intervals respectively.

[0017] The differentiable Kalman filter layer is internally configured with an inner loop for refinement iteration. Each iteration uses the previous posterior state to update the Jacobian matrix to adapt to the nonlinear curvature of the logarithmic wind profile mapping. The geometric accuracy factor of the BeiDou satellite directly participates in the real-time scaling calculation of the observation noise covariance matrix.

[0018] The training dataset of the neural-Kalman hybrid filter network collects raw observation data of multi-antenna Beidou arrays, wave height data of shipborne wave measuring radar and reference anemometer data under various typical sea conditions. The wind speed measured by the reference anemometer at a height of 10 meters is used as a label and divided into training samples by sliding window.

[0019] The neural-Kalman hybrid filter network is trained using the mean squared error loss function, the optimizer uses the adaptive moment estimation algorithm, and the early stopping strategy terminates training when the validation set loss has not decreased for several consecutive rounds.

[0020] Among them, the baseline length between antennas of the multi-antenna BeiDou array ranges from 1 to 3 meters, the absolute elevation accuracy of the main antenna is better than 0.2 meters, the original relative wind speed sampling frequency ranges from 1 to 10 Hz, the sliding time window length is 60 seconds, and the regularization coefficient range of the Tikhonov regularization constraint is... ~ The number of hidden units in the long short-term memory module is 128, the number of inner loop refinement iterations is 2 to 3, and the weights of the covariance matrix reset range from 0.1 to 0.9.

[0021] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for constructing a ship standard height wind speed correction model.

[0022] A third aspect of the present invention provides a system for constructing a ship standard height wind speed correction model, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0023] This invention constructs a multi-antenna BeiDou array map attitude calculation mechanism to recover the ship's roll and pitch angles in real time. It also combines the spatial geometric vector calculation of the carrier coordinate system to calculate the instantaneous vertical deviation of the wind sensor. Then, it uses the minimum description length criterion to adaptively select the vertical motion prediction model. After the wind sensor's altitude is accurately estimated in real time, it is input into a neural-Kalman hybrid filter network to complete the dynamic nonlinear conversion of the logarithmic wind profile.

[0024] The minimum cut attitude calculation method of this invention automatically assigns multipath contaminated baselines to low-weight subgraphs without explicit outlier removal, ensuring stable convergence of attitude angles under multipath interference and thus guaranteeing the robustness of wind sensor height estimation. The minimum description length candidate model selection mechanism automatically switches the prediction model type at the moment of sea state change, avoiding prediction biases caused by fixed models during state transitions and ensuring that the vertical displacement prediction value consistently closely approximates the actual motion. In the neural-Kalman hybrid filter network, the long short-term memory module drives real-time adaptive estimation of the covariance matrix, the attention gating module quickly resets the covariance matrix during sudden sea state changes, and the inner loop refines and iteratively tracks the nonlinear curvature of the logarithmic wind profile, jointly ensuring that the accuracy of wind speed calculation does not degrade when the height deviation is large.

[0025] In summary, the present invention solves the technical problem mentioned in the background art of the difficulty in real-time and accurate compensation for wind speed and height conversion errors caused by the dynamic changes in the height of ship wind sensors. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 A time-series comparison of roll and pitch angles for minimum cut attitude calculation under strong swell conditions.

[0028] Figure 3 The time series diagram of standard 10-meter height wind speed output by the neural-Kalman hybrid filter network before and after the sea state change is shown.

[0029] Figure 4 The graph shows the distribution of the root mean square error of the final corrected wind speed over time under three sea states. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0031] like Figure 1 The diagram shown is a flowchart of a method for constructing a ship standard altitude wind speed correction model according to the first aspect of the present invention. This method includes the following steps: S10. Use a multi-antenna BeiDou array to synchronously acquire the absolute elevation of the main antenna and the phase differential observation data of multiple baseline carriers, and synchronously record the original relative wind speed. S20. Model the spatial baseline network of the multi-antenna BeiDou array as a weighted undirected graph, construct the graph Laplace matrix, solve the generalized eigenvalue problem, recover the ship's roll and pitch angles, and construct the rotation matrix. S30. Establish the spatial geometric vectors of the antenna array and the wind sensor in the carrier coordinate system, and calculate the instantaneous vertical deviation of the wind sensor center by combining the rotation matrix. S40. Perform minimum description length candidate model evaluation on the vertical motion within the current time window and output the vertical displacement prediction value; fuse the absolute elevation of the main antenna with the instantaneous vertical deviation and the vertical displacement prediction value to obtain the wind sensor altitude; introduce the wave age parameter to dynamically adjust the Chanok coefficient and apply Tikhonov regularization constraint to the sea surface roughness iterative equation to obtain the sea surface roughness. S50: Input the raw relative wind speed at the altitude of the wind sensor into the neural-Kalman hybrid filter network, and output the smoothed standard 10-meter height wind speed; S60 uses the geometric accuracy factor of BeiDou satellites as the observation weight to perform adaptive Kalman filtering on the standard 10-meter height wind speed sequence and outputs the final corrected wind speed.

[0032] In step S10, the multi-antenna BeiDou array consists of no fewer than three BeiDou receiving antennas, with the baseline length between antennas ranging from 1 to 3 meters. The antenna spatial layout is configured according to the three-axis direction of the carrier coordinate system to ensure the observability of the two attitude angles, roll and pitch. The absolute elevation of the main antenna is obtained by precise single-point positioning calculation of the BeiDou main antenna, with an accuracy better than 0.2 meters. The accuracy threshold is determined by repeated experiments and statistical root mean square error in no fewer than 10 port static reference scenarios. The original relative wind speed is directly output by the shipborne wind sensor, with a sampling frequency range of 1 to 10 Hz.

[0033] In step S20, the weighted undirected graph uses each BeiDou receiving antenna as a vertex and the baseline between any two BeiDou receiving antennas as an edge. The edge weight is determined by the combined score of the signal-to-noise ratio (SNR) and geometric strength of the carrier phase differential observations of the corresponding baseline. The combined score is obtained by statistically analyzing the correlation between the SNR of each baseline and the corresponding attitude calculation residual in no less than 20 sets of measured data under different sea states, and by using weighted least squares fitting to obtain the SNR weight coefficient and the geometric strength weight coefficient. The sum of the two coefficients after normalization is 1. The graph Laplacian matrix... satisfy ,in For degree matrix, This is the edge weight matrix. Take the number of Beidou receiving antennas The generalized eigenvalue problem is The algorithm extracts three eigenvectors corresponding to the second to fourth smallest eigenvalues, spans them into a low-dimensional embedding space of the attitude parameter manifold, and recovers the roll and pitch angles from the eigenvector coordinate system using least-squares mapping. The computational cost of the algorithm is... , 3 is selected; when a baseline is contaminated by multipath error, the weight of the corresponding edge is automatically reduced, and the baseline is assigned to a low-weight subgraph in the graph segmentation, realizing adaptive robust estimation based on topology structure, without the need to explicitly set the outlier removal threshold; the graph minimum cut attitude solution algorithm transforms multipath contamination identification and isolation into graph topology analysis, the weight of contaminated baselines is automatically decayed and grouped with high-quality baselines in the spectrum segmentation, so that the attitude solution maintains stable convergence under multipath interference, and has natural robustness to scenarios where multiple baselines are contaminated at the same time.

[0034] Step S20 further includes: employing a partial ambiguity fixing strategy, prioritizing the fixing of the baseline subset with the highest comprehensive score; and combining an improved least-squares ambiguity decorrelation adjustment search tree pruning algorithm to control the integer ambiguity fixing time to 1–3 minutes. The upper bound of the fixed time for integer ambiguity is determined by statistically analyzing the fixed success rate and time consumption distribution in no less than 50 cycle slip simulation experiments, and taking the upper bound of time consumption corresponding to a fixed success rate of no less than 95%. The epoch ambiguity inheritance mechanism directly inherits the baseline fixed solution that is not affected by the cycle slip after the cycle slip, reducing the number of ambiguity parameters that need to be searched again. It also includes: establishing a multipath hemispherical map, collecting carrier phase residual data of each Beidou receiving antenna for no less than 7 days under anchoring or low-speed uniform operation conditions, dividing it into 5°×5° grids according to the satellite elevation angle 0~90° and azimuth angle 0~360°, and statistically analyzing the average carrier phase residual value in each grid as the multipath prior template; in real-time calculation, subtracting the corresponding multipath prior template value from the table according to the current satellite elevation angle and azimuth angle; using adaptive spectral subtraction to suppress the frequency domain of the remaining residual, with a frequency domain resolution range of 0.01~0.1Hz, determined by the power spectrum analysis of the remaining residual.

[0035] In step S30, the spatial geometric vector describes the three-dimensional positional relationship between the phase centers of each Beidou receiving antenna and the center of the wind sensor in the carrier coordinate system, and is obtained by measurement from the shipborne structural drawings; the rotation matrix is ​​constructed from the roll and pitch angles calculated in step S20; the instantaneous vertical deviation is obtained by extracting the vertical component through matrix multiplication of the spatial geometric vector and the rotation matrix; the carrier phase differential technology eliminates satellite clock error and receiver clock error through the carrier phase double difference observation values ​​between the main antenna and the slave antenna, and the calculation accuracy of the vertical displacement of the wind sensor center reaches the decimeter level; the lower limit of the decimeter-level accuracy is determined by statistically analyzing the root mean square error in no less than 30 repeated experiments in a static reference scenario, requiring the root mean square error to not exceed 0.12m.

[0036] In step S40, the minimum description length candidate model is selected within a length of 60. The process is executed within a sliding time window. The candidate model set includes three types: sinusoidal superposition model, autoregressive model, and piecewise linear model. The parameters of the sinusoidal superposition model are the frequency, amplitude, and initial phase of each component; the order of the autoregressive model is... The range is 2–8, pre-determined from offline experimental data by the Akaike information criterion; the number of inflection points in the piecewise linear model ranges from 1–3, and the inflection point positions and slopes are obtained through piecewise least squares estimation; the minimum description length criterion value is composed of the sum of the model parameter encoding length and the data condition encoding length; the optimal vertical motion model is used to extrapolate and predict the vertical displacement prediction value of the wind sensor in the next epoch, serving as the input to the process equation of the adaptive Kalman filter; the algorithm automatically switches the model type at the instant of sea state change, eliminating the prediction of the fixed model during the state transition period. The bias is low, requiring only parallel evaluation of the likelihood functions of 3 to 5 candidate models, resulting in low computational overhead. The wave age parameter is defined as the ratio of peak wave velocity to frictional wind speed. The piecewise functional relationship between the Chanock coefficient and the wave age parameter is obtained by collecting peak wave velocity data and corresponding measured frictional wind speed data under at least 5 typical sea states, calculating the wave age parameter and inverted Chanock coefficient under each sea state, and fitting the inflection point position and slope of each segment of the piecewise function. The peak wave velocity is provided in real time by a shipboard wave measuring radar or buoy. The regularization coefficient range of the Tikhonov regularization constraint is... ~ The altitude of the wind sensor was determined experimentally by minimizing the iterative residuals of sea surface roughness on data from different sea state transition periods; ,in Take 10m, The absolute elevation of the main antenna. For instantaneous vertical deviation, This is the predicted value for vertical displacement.

[0037] In step S50, the specific structure of the neural-Kalman hybrid filter network is as follows: the network consists of four parts: a long short-term memory module, a linear projection layer, a differentiable Kalman filter layer, and an attention gating module; the long short-term memory module has 128 hidden layer units, and the input sequence length corresponds to 60. The historical observation sequence has an input feature dimension of 3, corresponding to the carrier phase residual, the BeiDou satellite geometric accuracy factor, and the original relative wind speed, respectively. The hidden state of the Long Short-Term Memory (LSTM) module is mapped to the process noise covariance matrix and the observation noise covariance matrix through two independent learnable linear projection layers. The output of the projection layer is activated by the Softplus activation function to ensure positive definiteness. The differentiable Kalman filter layer uses differentiable matrix operations to implement the prediction step and update step. The gradient propagates back from the filtered posterior state, penetrates the Kalman filter mathematical structure, and flows back to the LSM module parameters. The attention gating module monitors the sea state change features identified by the LSM module. When the change confidence exceeds the dynamic change threshold, the gating signal triggers the covariance matrix fast reset submodule to perform a weighted reset on the process noise covariance matrix and the observation noise covariance matrix. The reset weight ranges from 0.1 to 0.9, determined by minimizing the number of convergence steps after reset in no less than 10 simulated sea state change experiments. The differentiable Kalman filter layer has 2 to 3 internal loop refinement iterations. Each iteration uses the previous posterior state to update the Jacobian. The matrix is ​​adapted to the nonlinear curvature of the logarithmic wind profile mapping. The network memory allocation strategy is as follows: the weights of the long short-term memory module reside in the memory, the process noise covariance matrix and the observation noise covariance matrix are allocated to the independent first CUDA stream, and are executed in parallel with the matrix operation of the differentiable Kalman filter layer. The attention gating module is allocated to the second CUDA stream to achieve dual-stream parallelism and reduce single-epoch inference latency. The Beidou satellite geometric precision factor is used as the observation weight and directly participates in the real-time scaling calculation of the observation noise covariance matrix in the first CUDA stream, forming an explicit functional relationship between the Beidou satellite geometric precision factor and the observation noise covariance matrix. The neural-Kalman hybrid filter network enables the long short-term memory module to adaptively estimate the covariance parameter by learning historical time-series features. The attention gating module quickly resets the covariance matrix when the sea state changes abruptly, so that the filter maintains convergence during the state transition period. The inner loop refinement structure of the differentiable Kalman filter layer targets the nonlinear amplification effect of the logarithmic wind profile and tracks the curvature change of the mapping function through multiple Jacobian matrix updates, so that the wind speed conversion accuracy remains stable when the height deviation is large.

[0038] The steps for establishing the training dataset for the neural-Kalman hybrid filter network specifically include: collecting raw observation data from a multi-antenna BeiDou array, shipboard wave height data from a wave measuring radar, and reference anemometer data under no less than six typical sea states, with a time span of no less than 30 days; using the wind speed measured by the reference anemometer at a height of 10m as a label, and using the carrier phase residual sequence, the BeiDou satellite geometric precision factor sequence, and the raw relative wind speed sequence as input features, according to 60... The sliding window is divided into training samples; the training set and validation set are divided in an 8:2 ratio, and the test set is taken from sea state data that does not overlap with the training set.

[0039] The specific steps for training the neural-Kalman hybrid filter network include: supervising training on the standard 10-meter height wind speed output using a mean squared error loss function; selecting an adaptive moment estimation algorithm as the optimizer, with an initial learning rate range of [value missing]. ~ The batch size is determined by convergence curve experiments on the validation set; the batch size ranges from 32 to 128, determined by the memory capacity constraint; the training epochs are no less than 100 epochs, and an early stopping strategy is adopted to terminate training when the validation set loss has not decreased for 20 consecutive epochs; the positive definiteness constraints of the process noise covariance matrix and the observation noise covariance matrix are automatically satisfied during forward propagation by the Softplus activation function.

[0040] The dynamic mutation threshold is determined by a resource dynamic adjustment function; the resource dynamic adjustment function calculates a comprehensive state index based on three inputs: the BeiDou satellite geometric accuracy factor, the standard deviation of the carrier phase residual, and the current sea state level. The weighted sum of the three inputs is normalized to the 0-1 range. The weighting coefficients are obtained by minimizing the false triggering rate of the attention gating module in at least 15 sets of sea state change experiments. When a higher dynamic mutation threshold is used, the covariance matrix reset weights are set to 0.1–0.3; when When a moderate dynamic mutation threshold is used, the covariance matrix reset weights are set to 0.3–0.7; when At that time, a lower dynamic mutation threshold was adopted, and the covariance matrix reset weight was set to 0.7 to 0.9. The dynamic mutation thresholds for the three intervals were obtained by minimizing the number of reset filter convergence steps on the measured data within the corresponding comprehensive state index interval.

[0041] The graph minimum cut attitude solution algorithm refers to: representing the spatial baseline network of a multi-antenna array as a weighted undirected graph, obtaining a low-dimensional embedded representation of attitude parameters by solving the generalized eigenvalue problem of the graph Laplacian matrix, and using the graph structure to automatically identify and isolate the multi-path contaminated baseline for attitude calculation.

[0042] The minimum description length candidate model selection refers to an algorithm that, based on the minimum description length criterion, evaluates multiple candidate models in parallel for vertical motion data within a sliding time window, and automatically selects the optimal vertical motion model and extrapolates the predicted vertical displacement value by taking the minimum sum of the model parameter encoding length and the data condition encoding length as the criterion.

[0043] The neural-Kalman hybrid filter network refers to a hybrid filter network that deeply couples a long short-term memory module with a differentiable Kalman filter layer, uses the hidden state of the long short-term memory module to drive the process noise covariance matrix and the observation noise covariance matrix in real time, uses the attention gating module to respond to sudden changes in sea state to trigger a rapid reset of the covariance matrix, and has an internal loop for refinement iteration to adapt to the nonlinear curvature of the logarithmic wind profile.

[0044] The Tikhonov regularization constraint refers to a constraint method that introduces a regularization term into the sea surface roughness iterative equation, in exchange for the norm of the solution being penalized, to ensure the stable convergence of the inverse problem under any sea state, thus avoiding the strong coupling and iterative divergence between the Chanok coefficient and the frictional wind speed.

[0045] The multipath hemispherical map refers to the average carrier phase residual of each BeiDou receiving antenna calculated according to the two-dimensional grid of satellite elevation and azimuth angles, forming a multipath prior template, which is used in real-time calculation to subtract the corresponding multipath error from the spatial statistical lookup table.

[0046] The specific implementation of step S10 is as follows: The multi-antenna BeiDou array consists of no fewer than three BeiDou receiving antennas, with a baseline length between antennas ranging from 1 to 3 meters. The antenna spatial layout is configured according to the three axes of the carrier coordinate system to ensure the observability of the roll and pitch angles. The absolute elevation of the main antenna is calculated by the BeiDou main antenna precise single-point positioning algorithm, with an accuracy better than 0.2 meters. This accuracy threshold is determined by repeated experiments and statistical root mean square error calculation in no fewer than 10 port static reference scenarios. The raw relative wind speed is directly output by the shipborne wind sensor, with a sampling frequency range of 1 to 10 Hz. It is synchronized with the BeiDou observation data timestamp and stored synchronously as the raw input for subsequent processing.

[0047] The specific implementation of step S20 is as follows: The spatial baseline network of the multi-antenna BeiDou array is modeled as a weighted undirected graph, with each BeiDou receiving antenna as a vertex and the baseline between any two BeiDou receiving antennas as an edge. The edge weight is determined by the comprehensive score of the signal-to-noise ratio (SNR) and geometric strength of the carrier phase differential observations of the corresponding baseline. The comprehensive score is obtained by statistically analyzing the correlation between the SNR of each baseline and the corresponding attitude calculation residual in no less than 20 sets of measured data under different sea states, and by using weighted least squares fitting to obtain the SNR weight coefficient and the geometric strength weight coefficient. The sum of the two coefficients after normalization is 1. (Graph Laplacian matrix) satisfy ,in For degree matrix, Let be the edge weight matrix. The generalized eigenvalue problem is... The three eigenvectors corresponding to the second to fourth smallest eigenvalues ​​are taken and spanned into a low-dimensional embedding space of the attitude parameter manifold. The roll and pitch angles are recovered from the eigenvector coordinate system through least-squares mapping to construct the rotation matrix. When a baseline is contaminated by multipath error, the weight of the corresponding edge is automatically reduced, and the baseline is assigned to a low-weight subgraph in the graph segmentation, achieving adaptive robust estimation. In addition, a partial ambiguity fixing strategy is adopted, prioritizing the fixing of the subset of baselines with the highest comprehensive score. Combined with an improved least-squares ambiguity decorrelation adjustment search tree pruning algorithm, the integer ambiguity fixing time is controlled within 1 to 3 seconds. The epoch ambiguity inheritance mechanism directly inherits the fixed solution of the baseline that is not affected by the cycle slip after the cycle slip. Under anchored or low-speed uniform operating conditions, carrier phase residual data of each BeiDou receiving antenna is collected for no less than 7 days. The data is divided into 5°×5° grids according to the satellite elevation angle of 0-90° and the azimuth angle of 0-360°. The average carrier phase residual value in each grid is used as the multipath prior template. In real-time calculation, the corresponding multipath prior template value is subtracted from the table based on the current satellite elevation angle and azimuth angle. The remaining residual is suppressed in the frequency domain by adaptive spectral subtraction. The frequency domain resolution range is 0.01-0.1Hz.

[0048] The specific implementation of step S30 is as follows: Spatial geometric vectors describe the three-dimensional positional relationship between the phase centers of each BeiDou receiving antenna and the center of the wind sensor in the carrier coordinate system, obtained by measurement from the shipborne structural drawings. The rotation matrix is ​​constructed from the roll and pitch angles calculated in step S20. The instantaneous vertical deviation is obtained by extracting the vertical component through matrix multiplication of the spatial geometric vector and the rotation matrix, achieving a calculation accuracy at the decimeter level. The lower limit of the decimeter-level accuracy is determined by statistically analyzing the root mean square error (RMSE) in at least 30 repeated experiments under a static reference scenario, requiring the RMSE to not exceed 0.12m. Carrier phase differential technology eliminates satellite clock errors and receiver clock errors through the double-difference observations of the carrier phase between the main antenna and the slave antenna, ensuring high-precision extraction of the vertical component.

[0049] The specific implementation of step S40 is as follows: The selection of the minimum description length candidate model is performed within a sliding time window of 60 seconds. The candidate model set includes three types: sinusoidal superposition model, autoregressive model, and piecewise linear model. The parameters of the sinusoidal superposition model are the frequency, amplitude, and initial phase of each component; the order of the autoregressive model ranges from 2 to 8, and is predetermined in offline experimental data by the Akaike information criterion; the number of inflection points in the piecewise linear model ranges from 1 to 3, and the inflection point positions and slopes are obtained through piecewise least squares estimation. The minimum description length criterion value is composed of the sum of the model parameter encoding length and the data condition encoding length. The model with the smallest criterion value is selected as the optimal vertical motion model, and the predicted vertical displacement value for the next epoch is extrapolated. The altitude of the wind sensor meets the requirements. ,in The absolute elevation of the main antenna. For instantaneous vertical deviation, This represents the predicted vertical displacement. The wave age parameter is defined as the ratio of peak wave velocity to frictional wind speed. The piecewise functional relationship between the Chanok coefficient and the wave age parameter is obtained through fitting under at least five typical sea states. The regularization coefficient range of the Tikhonov regularization constraint is... ~ The method is determined by experiments that minimize the iterative residual of sea surface roughness on data from different sea state transition periods, ensuring that the sea surface roughness iteration converges stably under any sea state.

[0050] The specific implementation of step S50 is as follows: The neural-Kalman hybrid filter network consists of four parts: a long short-term memory module, a linear projection layer, a differentiable Kalman filter layer, and an attention gating module. The long short-term memory module has 128 hidden layer units, the input sequence length corresponds to a 60-second historical observation sequence, and the input feature dimension is 3, corresponding to the carrier phase residual, the BeiDou satellite geometric accuracy factor, and the original relative wind speed, respectively. The hidden state of the long short-term memory module is mapped to the process noise covariance matrix and the observation noise covariance matrix through two independent linear projection layers, respectively. The output of the projection layer is guaranteed to be positive definite by the Softplus activation function. The differentiable Kalman filter layer uses differentiable matrix operations to implement the prediction step and the update step. The gradient propagates backward from the filtered posterior state, penetrates the Kalman filter mathematical structure, and flows back to the parameters of the long short-term memory module. It internally sets 2 to 3 inner loop refinement iterations. Each iteration uses the previous posterior state to update the Jacobian matrix to adapt to the nonlinear curvature of the logarithmic wind profile mapping. The attention gating module monitors sea state abrupt change characteristics. When the confidence level of the abrupt change exceeds the dynamic abrupt change threshold, it triggers a rapid reset of the covariance matrix, with reset weights ranging from 0.1 to 0.9. The dynamic abrupt change threshold is determined by a resource dynamic adjustment function. A comprehensive state index is calculated based on three inputs: the geometric accuracy factor of the BeiDou satellite, the standard deviation of the carrier phase residual, and the current sea state level. After normalizing the comprehensive state index to the 0-1 range, reset weights are configured for each of the three threshold ranges. The training dataset collects raw observation data from a multi-antenna BeiDou array under at least six typical sea states, spanning at least 30 days. The wind speed measured at a height of 10 meters using a reference anemometer is used as the label. The dataset is divided into training samples using a 60-second sliding window, with the training and validation sets divided in an 8:2 ratio. Training employs a mean squared error loss function, and the optimizer uses an adaptive moment estimation algorithm. An early stopping strategy is used to terminate training when the validation set loss does not decrease for 20 consecutive rounds.

[0051] The specific implementation of step S60 is as follows: using the geometric accuracy factor of the BeiDou satellite as the observation weight, an adaptive Kalman filter is performed on the standard 10-meter height wind speed sequence. When the geometric accuracy factor of the BeiDou satellite is small, the observation weight is high, and vice versa. In this way, the confidence of the observation information is enhanced when the satellite geometric conditions are good, and the process prediction is relied upon when the geometric conditions are poor to output the final corrected wind speed.

[0052] It should be noted that the key technologies of this invention include: the minimum cut attitude solution technology transforms multipath contaminated baseline identification into topological analysis of a weighted undirected graph. The weight of the contaminated baseline is automatically decayed and grouped with high-quality baselines in the spectral segmentation, avoiding the problem of high-quality baselines being mistakenly removed due to improper threshold setting in traditional explicit outlier detection methods under strong multipath environments, thus ensuring the stability and observability of attitude angles in complex electromagnetic environments; the minimum description length candidate model selection technology automatically matches the physical laws of vertical motion with information theory criteria, and can switch the prediction model without manual intervention when the swell cycle changes or the sea state changes abruptly, eliminating the systematic prediction bias of the fixed model during the state transition period, and ensuring the continuous accuracy of wind sensor altitude estimation across the entire sea state range; the neural-Kalman hybrid filtering technology deeply couples deep temporal learning with physical filtering structure, with the long short-term memory module driving the covariance matrix to adapt in real time, the inner loop refining and iteratively tracking the nonlinear curvature of the logarithmic wind profile, and the attention gating module quickly restoring the convergence of the filter when the sea state changes abruptly. The three key technologies cover the three levels of measurement input reliability, height estimation adaptability, and conversion process robustness, forming a complete error compensation closed loop. Their synergistic effect makes the overall correction accuracy significantly better than the effect of any single technology applied independently when multiple interference factors are superimposed.

[0053] It should be noted that when ships undertake long-distance ocean voyages across multiple climate zones, the environmental sea state can change abruptly from low waves and high-frequency swells to high waves and low-frequency swells within hours. Accompanied by strong wind shear, the dominant frequency and amplitude of the vertical motion of the wind sensor simultaneously undergo large-scale jumps. Traditional vertical displacement prediction methods based on a single fixed-parameter model will produce systematic prediction biases lasting tens of seconds to several minutes during such state transitions. This leads to persistently underestimating or overestimating the altitude of the wind sensor, and the logarithmic wind profile conversion introduces non-healing cumulative errors. The reason for these technical problems is that the frequency response characteristics of the fixed-parameter model do not match the actual vertical motion patterns after the state transition. Model switching relies on manually set sea state discrimination thresholds. When sea states change continuously or multiple factors overlap, the threshold judgment lags behind the actual physical state changes, causing the erroneous model to continue participating in predictions until the next manual intervention. Common solutions to these technical problems include increasing the model update frequency or introducing more discriminative features to improve threshold sensitivity. However, increasing the update frequency can lead to frequent false switching under short-term noise disturbances, causing non-physical oscillations in the predicted values. Adding discriminative features requires pre-labeling a large number of sea state samples to train the discriminator, which is difficult to obtain sufficient labeled data covering all climate zones in actual ship deployments. Furthermore, the discriminator itself has limited generalization ability under transitional sea states not covered by the training set, making it difficult to accurately identify complex sea states with multiple overlapping factors. This invention effectively solves this technical problem. The minimum description length candidate model selection technique does not rely on any sea state discrimination threshold, but instead directly quantifies the descriptive efficiency of each candidate model for vertical motion data within the current time window using information theory criteria. The candidate model set includes three types: sinusoidal superposition models, autoregressive models, and piecewise linear models, corresponding to motion laws dominated by different physical mechanisms. At each epoch, the three types of models are evaluated in parallel for the minimum description length criterion value of the current window data. The model with the smallest criterion value is the optimal descriptive model for the current motion state, without the need for pre-determining the sea state type. When sea state changes, the efficiency of the original model that dominated the motion pattern decreases, and its criterion value naturally increases. The criterion value of the new matching model decreases accordingly. The system automatically completes the model switch at the information theory level, with a switching delay of only one sliding window update cycle, without any manual intervention or external sea state annotation data. This mechanism ensures that the wind sensor's altitude estimation is always driven by the model that best matches the actual motion pattern during navigation across all climate zones, fundamentally eliminating the systematic prediction bias of the fixed model during sea state transitions.

[0054] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for constructing a ship standard height wind speed correction model.

[0055] A third aspect of the present invention provides a system for constructing a ship standard height wind speed correction model, comprising the aforementioned computer-readable storage medium. The system is any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0056] Specifically, the principle of this invention is: The fundamental reason why this invention can solve the above-mentioned technical problems is that the core source of the wind speed-height conversion error is the uncertainty of the instantaneous height of the wind sensor, which in turn stems from three mutually coupled physical mechanisms: dynamic changes in ship attitude, nonlinear coupling of sea surface roughness, and multipath interference in the measurement system. This invention establishes corresponding decoupling and compensation paths for each of these three mechanisms, and fuses the outputs of each path into the final corrected wind speed through a unified filtering framework.

[0057] In the attitude calculation stage, this invention models the spatial baseline network of a multi-antenna BeiDou array as a weighted undirected graph, and maps the attitude parameters to a low-dimensional manifold space through generalized eigenvalue decomposition of the graph Laplacian matrix. When a baseline is contaminated by multiple paths, its corresponding edge weights automatically decay and are naturally assigned to a low-weight subgraph during spectral segmentation. The influence of contaminated baselines is isolated at the mathematical structure level, eliminating the need to rely on explicit outlier detection thresholds. This mechanism ensures the stable observability of roll and pitch angles under complex electromagnetic environments, providing a reliable foundation for subsequent vertical deviation calculation.

[0058] In the real-time height estimation stage of the wind sensor, this invention establishes a spatial geometric vector between the antenna array and the wind sensor in the carrier coordinate system, and extracts the vertical component by combining the rotation matrix to obtain the instantaneous vertical deviation. Simultaneously, the minimum description length criterion evaluates three candidate models—sinusoidal superposition, autoregressive, and piecewise linear—in parallel within a sliding time window, automatically selecting the optimal model based on minimizing the sum of the parameter encoding length and the data condition encoding length, and predicting the vertical displacement for the next epoch. This mechanism enables height estimation to adaptively match the optimal motion law under different operating conditions such as swell cycle changes and sudden sea state changes, avoiding the systematic deviation of fixed models during state transition periods.

[0059] In the parameterization of sea surface roughness, this invention introduces a wave age parameter to dynamically adjust the Chanok coefficient and applies Tikhonov regularization constraints to the iterative equation of sea surface roughness. By penalizing the norm of the solution, the stable convergence of the inverse problem under any sea state is obtained, thus cutting off the path of strong coupling and iterative divergence between the Chanok coefficient and frictional wind speed from the root.

[0060] In the wind speed calculation stage, the neural-Kalman hybrid filter network deeply couples the temporal feature learning capability of the long short-term memory module with the physical structural constraints of the differentiable Kalman filter layer. The long short-term memory module learns the temporal patterns of historical carrier phase residuals, BeiDou satellite geometric accuracy factors, and the original relative wind speed, driving the adaptive estimation of the process noise covariance matrix and the observation noise covariance matrix in real time. The attention gating module triggers a rapid reset of the covariance matrix when it detects abrupt changes in sea state, ensuring the filter maintains convergence during state transitions. The inner loop refinement iteration of the differentiable Kalman filter layer tracks the nonlinear curvature change of the logarithmic wind profile mapping function through multiple Jacobian matrix updates, ensuring that the calculation accuracy does not degrade when the wind sensor height deviation is large. These components form a complete technical closed loop, jointly guaranteeing the real-time and accurate correction of the standard 10-meter height wind speed under dynamic sea states.

[0061] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment is described in detail below.

[0062] The specific implementation of step S10 is as follows: The multi-antenna BeiDou array consists of no fewer than three BeiDou receiving antennas, with baseline lengths between antennas ranging from 1 to 3 meters. The antenna spatial layout is configured according to the three axes of the carrier coordinate system. The absolute elevation of the main antenna is... The accuracy, better than 0.2m, was obtained through precise single-point positioning calculations using the BeiDou main antenna. This accuracy threshold was determined by repeatedly conducting experiments and statistically analyzing the root mean square error in no fewer than 10 static reference scenarios at ports. (Original relative wind speed) The data is directly output from the shipborne wind sensor, with a sampling frequency range of 1–10 Hz, and is recorded synchronously with the observation data from the BeiDou array.

[0063] The specific implementation of step S20 is as follows: the spatial baseline network of the multi-antenna BeiDou array is modeled as a weighted undirected graph. ,in For the antenna vertex set, For the baseline edge set, This is the edge weight matrix. Edge weights Signal-to-noise ratio of carrier phase differential observations corresponding to the baseline With geometric strength The overall score is determined. Baseline Signal-to-noise ratio of carrier phase differential observations Baseline The geometric intensity score is calculated from the reciprocal of the geometric dilution factor of the satellite elevation angle distribution corresponding to that baseline. The comprehensive scoring formula is expressed as follows: ; In the formula, This is a reference value for weight normalization, with an empirical value of 1. Normalized reference value for signal-to-noise ratio Geometric strength normalized reference value and The signal-to-noise ratio weighting coefficient and the geometric intensity weighting coefficient satisfy the following conditions: The two coefficients were obtained by statistically analyzing the correlation between the baseline signal-to-noise ratio and the attitude calculation residuals from no fewer than 20 sets of measured data under different sea states, and then using weighted least squares fitting. (Graph Laplacian matrix) satisfy: ; In the formula, For degree matrix, Take the number of Beidou receiving antennas The generalized eigenvalue problem is: ; In the formula, For feature vectors, For the corresponding eigenvalues, the three eigenvectors corresponding to the 2nd to 4th smallest eigenvalues ​​are used to span the low-dimensional embedding space of the pose parameter manifold, where the roll angle is... With pitch angle The coordinate system is recovered from the eigenvector coordinate system through least squares mapping, and the mapping relationship is expressed as follows: ; In the formula, This is a reference value for angle normalization, with an empirical value of 1 rad. , , The eigenvectors corresponding to the 2nd to 4th smallest eigenvalues The reference value for normalizing the feature vector is 1, which is an empirical value. The least squares mapping matrix was obtained by fitting offline calibration data. Normalized reference value and Same. Rotation matrix From the roll angle With pitch angle Build: ; The computational cost of the algorithm is , Take 3. When a baseline is contaminated by multipath error, the corresponding edge weight is... Automatic reduction: the baseline is assigned to a low-weight subgraph in the map segmentation, achieving adaptive robust estimation based on topology. Step S20 also includes: adopting a partial ambiguity fixing strategy, prioritizing the fixing of the subset of baselines with the highest comprehensive score, and combining an improved least-squares ambiguity decorrelation adjustment search tree pruning algorithm to control the integer ambiguity fixing time to 1-3 seconds. The upper bound of this time is determined by statistically analyzing the fixing success rate and time consumption distribution in no less than 50 cycle slip simulation experiments, taking the upper bound of time consumption corresponding to a fixing success rate of no less than 95%. The epochal ambiguity inheritance mechanism directly inherits the baseline fixing solution that is not affected by the cycle slip after the cycle slip. The multipath hemispherical map is divided into 5°×5° grids according to the satellite elevation angle of 0-90° and the azimuth angle of 0-360°, and the mean value of the carrier phase residual in each grid is statistically analyzed. As a multi-path prior template, in For satellite elevation index, This is the azimuth index, subtracted from the current carrier phase residual during real-time calculation. The remaining residual is suppressed in the frequency domain by adaptive spectral subtraction, with a frequency domain resolution range of 0.01 to 0.1 Hz.

[0064] The specific implementation of step S30 is as follows: In the carrier coordinate system, the spatial geometric vectors of the antenna array and the wind sensor Measurements obtained from shipboard structural drawings describe the three-dimensional positional relationship between the phase centers of each BeiDou receiving antenna and the center of the wind sensor. Instantaneous vertical deviation. The vertical component is obtained by matrix multiplication of spatial geometric vectors and rotation matrices, as expressed by the following formula: ; In the formula, It is a vertical unit vector, dimensionless. Take 10m The rotation matrix constructed in step S20 The unit is meters. The unit is m, and both sides are divided by Dimensionless. The carrier phase double-difference observation eliminates satellite clock error and receiver clock error, achieving decimeter-level accuracy in calculating the vertical displacement of the wind sensor center. The lower limit of accuracy is determined by statistically analyzing the root mean square error in no less than 30 repeated experiments under a static reference scenario, requiring the root mean square error to not exceed 0.12m.

[0065] The specific implementation of step S40 is as follows: the selection of candidate models with the minimum description length is performed within a sliding time window of 60 seconds. The candidate model set includes three categories: sinusoidal superposition models, autoregressive models, and piecewise linear models. The formula for the sinusoidal superposition model is as follows: ; In the formula, For the sinusoidal superposition model in The predicted vertical displacement at time t, in meters. The total number of sinusoidal components For the first Each sinusoidal component amplitude, in meters. For the first The frequency of each sinusoidal component is expressed in Hz. Time, in seconds This is a time normalization reference value, with an empirical value of 1 second. For the first Each sinusoidal component has an initial phase, measured in rad; each term is divided by... The value is dimensionless. The autoregressive model formula is expressed as follows: ; In the formula, For autoregressive models in The predicted vertical displacement at time t, in meters. For the first The order of the autoregressive coefficients is dimensionless; the order is... The range is 2 to 8, which is predetermined by the Akaike information criterion in offline experimental data. for The vertical displacement observation at time t, in meters. The residual term is in meters; each term is divided by... The dimensionless dimension is given by the piecewise linear model as follows: ; In the formula, For piecewise linear models in The predicted vertical displacement at time t, in meters. For the first Slope, in units For the first Each inflection point time is measured in seconds. For the first Segment intercept, in meters; number of inflection points ranges from 1 to 3, and the location and slope of the inflection points are obtained through piecewise least squares estimation. and Both are dimensionless, and the dimensions on both sides are uniform. Minimum descriptive length criterion value. Length encoded by model parameters With data conditional encoding length The sum is composed of: ; In the formula, , , All units are bits, and the dimensions on both sides are consistent. The optimal vertical motion model is used to extrapolate and predict the vertical displacement of the wind sensor in the next epoch. Wind sensor altitude satisfy: ; In the formula, , , , All units are in meters (m), and each term is divided by... Dimensionless after wave age parameter. Defined as peak wave velocity Frictional wind speed The ratio: ; In the formula, Peak wave velocity, in units of Real-time information is provided by marine wave-measuring radar or buoys. Frictional wind speed, unit: Obtained by wind speed observation inversion A dimensionless wave age parameter. Chanock coefficient. With wave age parameter The piecewise functional relationship between the components was obtained by collecting peak wave velocity data and corresponding measured frictional wind speed data under at least five typical sea states, calculating wave age parameters and inverted Chanok coefficients for each sea state, and fitting the inflection point positions and slopes of each segment of the piecewise function. Tikhonov regularization constraints were applied to the sea surface roughness iterative equation, and the objective function is expressed as follows: ; In the formula, Sea surface roughness, in meters (m). This is a normalized reference value for sea surface roughness, with the empirical value being... m The coefficient matrix is ​​obtained by linearizing the strongly coupled iterative equation of the Channock coefficient and frictional wind speed, and is dimensionless. This corresponds to the linearized observation vector, in units of m. The regularization coefficient is dimensionless and has a range of [value missing]. ~ The method was determined experimentally by minimizing the iterative residuals of sea surface roughness on data from different sea state transition periods. and All units are , divided by The following is dimensionless.

[0066] The specific implementation of step S50 is as follows: The neural-Kalman hybrid filter network consists of four parts: a long short-term memory module, a linear projection layer, a differentiable Kalman filter layer, and an attention gating module. The long short-term memory module has 128 hidden layer units, the input sequence length corresponds to a 60-second historical observation sequence, and the input feature dimension is 3, corresponding to the carrier phase residuals. Geometric precision factor of BeiDou satellites relative wind speed from the original The carrier phase residual is obtained by subtracting the theoretically calculated value from the real-time carrier phase observation value output by the BeiDou receiver, and is expressed in meters (m). Long Short-Term Memory (LSTM) module hidden state. The process noise covariance matrix is ​​mapped through two independent, learnable linear projection layers. Covariance matrix of observation noise The projection layer output is guaranteed to be positive definite by the Softplus activation function, which is defined as follows: ; In the formula, Scalar elements output by the projection layer, dimensionless The function output is always positive, ensuring the positive definiteness of the covariance matrix. The mapping formula between the process noise covariance matrix and the observation noise covariance matrix is ​​expressed as follows: ; ; In the formula, , For learnable weight matrix , For the corresponding bias vector This means constructing a diagonal matrix from the vectors. The geometric precision factor for BeiDou satellites is dimensionless and directly participates as an observation weight. Real-time scaling, The smaller the value, the higher the reliability of the observation. The smaller the value. To distinguish it from the observation matrix symbol in the Kalman filter formula, the observation noise covariance matrix is ​​uniformly denoted as... The formulas for the prediction step and update step of the differentiable Kalman filter layer are expressed as follows: ; ; ; ; ; In the formula, for Predict the state vector at each time step for Post-hoc state vector for Post-hoc state vector To predict the covariance matrix for Posterior covariance matrix at time step For the posterior covariance matrix The state transition matrix is ​​obtained by linearizing the process equations of the optimal vertical motion model in step S40. The observation matrix is ​​obtained by linearizing the logarithmic wind profile function at the current operating point. Kalman gain matrix for Wind speed observation at any time The identity matrix is ​​used. The differentiable Kalman filter layer contains 2 to 3 inner loop refinement iterations. Each iteration updates the Jacobian matrix using the previous posterior state to adapt to the nonlinear curvature of the logarithmic wind profile mapping. The iteration format is as follows: ; In the formula, superscript This serves as the index for the number of iterations in the inner loop. The range is 0 to 2 For the first In the next iteration The Jacobian matrix obtained by taking the partial derivative of the working point with respect to the logarithmic wind profile mapping function For the corresponding number The Kalman gain matrix of the next iteration is derived from... The result is obtained by substituting into the Kalman gain formula. The attention gating module monitors the characteristics of sudden changes in sea state and integrates the state index. Geometric precision factor of BeiDou satellites Standard deviation of carrier phase residual With the current sea state level The three inputs are weighted, summed, and normalized to the interval 0-1. ; In the formula, This represents a normalization operation that linearly maps the weighted summation result within the parentheses to the interval 0 to 1. Specifically, it involves subtracting the minimum sample value from the summation and then dividing by the sample range. The standard deviation of the carrier phase residual is expressed in meters. The current sea state level is output in real time by the shipboard wave-measuring radar. This is a normalized reference value for the geometric precision factor, with an empirical value of 10. The normalized reference value for the standard deviation of the carrier phase residual is 0.05m, with an empirical value of 0.05m. This is a normalized reference value for sea state levels, with an empirical value of 9. , , Let be the weighting coefficient, satisfying The values ​​were obtained by minimizing the false trigger rate of the attention gating module in no fewer than 15 sets of sea state change experiments; all terms in parentheses are dimensionless. Dimensionless. When When a lower dynamic mutation threshold is used, the covariance matrix reset weights are set to 0.7–0.9; when A moderate dynamic mutation threshold is used, and the reset weight is set to 0.3–0.7; when A relatively high dynamic mutation threshold was used, and the weights were reset to 0.1–0.3. The network was trained under supervision using the mean squared error loss function, and the optimizer employed was the adaptive moment estimation algorithm. The initial learning rate ranged from [value missing]. ~ The batch size ranges from 32 to 128, and the training rounds are no less than 100. An early stopping strategy is adopted, which terminates training when the validation set loss does not decrease for 20 consecutive rounds.

[0067] The specific implementation of step S60 is as follows: using the geometric precision factor of the BeiDou satellite. As the observation weights, an adaptive Kalman filter is applied to the standard 10-meter height wind speed sequence to output the final corrected wind speed. The smaller the geometric precision factor, the higher the observation reliability. The final corrected wind speed is the output value of the standard 10-meter height wind speed sequence after adaptive Kalman filtering smoothing.

[0068] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effectiveness of the invention, technicians built a test environment and deployed a multi-antenna Beidou array and a shipborne wind sensor system on a near-shore comprehensive test platform to collect 72 hours of continuous observation data, covering three typical sea states: stable, medium waves, and strong swells. The entire process of the ship standard height wind speed correction model construction method of the invention was tested and verified.

[0069] In this embodiment, the corrected wind speed is the true wind speed, which takes into account the ship's own speed for correction, and the altitude of the wind sensor should be the height of the wind sensor relative to the instantaneous sea surface.

[0070] The BeiDou array consists of four BeiDou receiving antennas with a baseline length of approximately 2 meters between them, arranged according to the three axes of the carrier coordinate system. The absolute elevation of the main antenna is calculated using a precise single-point positioning algorithm with an accuracy better than 0.2 meters. The raw relative wind speed is sampled and output by a shipborne ultrasonic wind sensor at a frequency of 5 Hz. The multipath hemispherical map is established by collecting 7 days of carrier phase residual data during the anchoring phase before testing, with a grid resolution of 5°×5°. During testing, the multipath prior template corrects the carrier phase residual in real time, and the residual error is further suppressed in the frequency domain of 0.01–0.1 Hz using adaptive spectral subtraction.

[0071] In the minimum cut attitude calculation process of the graph, the spatial baseline network is constructed as a 4-vertex weighted undirected graph. The edge weights are determined by a weighted comprehensive score of the signal-to-noise ratio weight coefficient and the geometric strength weight coefficient. These two coefficients are obtained by weighted least-squares fitting of 20 sets of historical sea state data. The generalized eigenvalue decomposition of the graph Laplacian matrix takes the three eigenvectors corresponding to the 2nd to 4th smallest eigenvalues. The roll and pitch angles are recovered through least-squares mapping. Figure 2 As shown, under strong swell conditions, the roll and pitch angle time-series curves output by the graph minimum cut attitude solution method are in high agreement with the results from the reference attitude instrument. Even during periods of strong multipath interference, the attitude angle solution results do not show significant jumps because the weight of the contaminated baseline is automatically decayed and assigned to a low-weight subgraph, demonstrating the natural isolation capability of the graph topology for multipath errors. The integer ambiguity fixation time is stably controlled within 3 seconds after a cycle slip, and the epoch ambiguity inheritance mechanism effectively reduces the number of parameters that need to be re-searched.

[0072] In the carrier coordinate system, the spatial geometric vector between the antenna array and the wind sensor is determined by measurement from the test platform structural drawings. The vertical installation deviation of the wind sensor relative to the main antenna is 3.2m. The vertical component is extracted using the rotation matrix to obtain the instantaneous vertical deviation sequence, with a root mean square error not exceeding 0.12m. The minimum description length candidate model selection involves parallel evaluation of three candidate models—sinusoidal superposition model, autoregressive model, and piecewise linear model—within a 60s sliding window. The order of the autoregressive model is pre-determined to be 5 by the Akaike information criterion, and the maximum number of inflection points for the piecewise linear model is set to 3. Under stable sea states, the sinusoidal superposition model is frequently selected as the optimal model; under strong swell conditions, the autoregressive model has the smallest criterion value and becomes the dominant prediction model; during periods of rapid sea state transition, the model type automatically switches within a single window update cycle without any manual intervention. The wave age parameter is calculated by the ratio of the peak wave velocity provided in real-time by the shipboard wave measuring radar to the current frictional wind speed. The Chanok coefficient corresponding to the piecewise function relationship is determined by fitting historical data from five typical sea states, and the Tikhonov regularization coefficient is taken as... The sea surface roughness iteration remained stable and converged throughout the entire test period.

[0073] After fusing the absolute elevation of the main antenna, the instantaneous vertical deviation, and the predicted vertical displacement, the real-time altitude estimation results of the wind sensor are shown in Table 1: Table 1. Statistical table of wind sensor altitude under three sea states

[0074] As shown in Table 1, under strong swell sea conditions, the instantaneous standard deviation of the wind sensor's altitude reaches 0.89m. If a fixed nominal altitude is used for logarithmic wind profile conversion, a non-negligible systematic error will be introduced in epochs with large altitude deviations. This invention estimates the wind sensor's altitude in real time and inputs it into a neural-Kalman hybrid filter network, ensuring that the conversion process always uses the true altitude, thus eliminating the source of error caused by the fixed altitude assumption at the physical mechanism level.

[0075] The neural-Kalman hybrid filter network was trained offline before testing. The training dataset covered six typical sea states over a 32-day period, with training samples split into 60-second sliding windows. The training and validation sets were divided in an 8:2 ratio. Test data was taken from sea state periods that did not overlap with the training set. During network inference, the Long Short-Term Memory (LSTM) module used the carrier phase residual, the BeiDou satellite geometric precision factor, and the original relative wind speed as inputs to drive the adaptive estimation of the process noise covariance matrix and the observation noise covariance matrix in real time. The BeiDou satellite geometric precision factor also directly participated in the real-time scaling calculation of the observation noise covariance matrix. In two significant sea state abrupt changes recorded during testing, the attention gating module detected that the abrupt change confidence exceeded the dynamic abrupt change threshold, triggering a rapid reset of the covariance matrix. The reset weights were set to 0.72 and 0.85 respectively based on the comprehensive state index. The filter recovered convergence within seconds after the abrupt change. Figure 3 As shown, no significant divergent oscillations were observed in the standard 10-meter height wind speed output before and after the two abrupt events, verifying the effectiveness of the attention gating module and the fast reset mechanism of the covariance matrix. The differentiable Kalman filter layer is configured with three inner-loop refinement iterations. Each iteration updates the Jacobian matrix to track the nonlinear curvature of the logarithmic wind profile mapping. In epochs where the wind sensor height deviation exceeds 0.5m, the cumulative correction amount of the inner-loop refinement iterations is significantly improved compared to the baseline without iterations, demonstrating the structure's targeted compensation capability for scenarios with large height deviations.

[0076] Finally, adaptive Kalman filtering is performed on the standard 10-meter height wind speed sequence using the geometric accuracy factor of the BeiDou satellite as the observation weight, and the final corrected wind speed is output. For example... Figure 4 As shown in Table 2, the final corrected wind speed error statistics for the three sea states are as follows: Table 2. Statistical Table of Final Corrected Wind Speed ​​Errors under Three Sea States

[0077] As shown in Table 2, the root mean square error of the final corrected wind speed is still controlled at a low level under strong swell sea conditions, which reflects the overall correction capability of the present invention under dynamic sea conditions.

[0078] Compared to traditional methods for calculating wind profiles using a fixed altitude and static logarithmic height, this invention offers the following advancements: Traditional methods treat the wind sensor height as a constant, failing to detect instantaneous altitude deviations caused by changes in ship attitude. This invention, however, incorporates instantaneous altitude deviations into the calculation input in real time through a cascade mechanism of image attitude calculation and minimum description length candidate model selection, compensating for the fundamental flaws of the fixed altitude assumption under dynamic sea states at the physical mechanism level. Traditional methods use fixed Chanok coefficients, failing to reflect the dynamic modulation of sea surface roughness by wave age. This invention introduces a piecewise function for wave age parameters and Tikhonov regularization constraints, ensuring stable convergence of sea surface roughness even during rapid sea state transitions. Traditional methods use fixed filtering parameters, unable to adapt to changes in signal statistical characteristics caused by abrupt sea state changes. This invention's neural-Kalman hybrid filtering network, driven by deep temporal learning, employs covariance adaptation and inner-loop refinement iteration, maintaining stable accuracy in the calculation process under both nonlinear mapping and state abrupt changes.

[0079] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0080] Table 3. Variable Explanation Table (Part 1)

[0081] Table 4. Variable Explanation Table (Part Two)

[0082] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a ship standard altitude wind speed correction model, characterized in that, Includes the following steps: The system utilizes a multi-antenna BeiDou array to synchronously acquire absolute elevation data of the main antenna and phase differential observation data of multiple baseline carriers, and simultaneously records the original relative wind speed. The spatial baseline network of the multi-antenna BeiDou array is modeled as a weighted undirected graph. The graph Laplace matrix is ​​constructed, the generalized eigenvalue problem is solved, the ship's roll and pitch angles are recovered, and the rotation matrix is ​​constructed. Establish the spatial geometric vectors of the antenna array and the wind sensor in the carrier coordinate system, and calculate the instantaneous vertical deviation of the wind sensor center by combining the rotation matrix; The minimum description length candidate model is evaluated for the vertical motion within the current time window, and the vertical displacement prediction value is output. The absolute elevation of the main antenna is fused with the instantaneous vertical deviation and the vertical displacement prediction value to obtain the altitude of the wind sensor. The wave age parameter is introduced to dynamically adjust the Chanok coefficient, and Tikhonov regularization constraint is applied to the sea surface roughness iterative equation to obtain the sea surface roughness. The raw relative wind speed at the altitude of the wind sensor is input into the neural-Kalman hybrid filter network, and the smoothed standard 10-meter height wind speed is output. Using the geometric accuracy factor of the BeiDou satellite as the observation weight, an adaptive Kalman filter is applied to the standard 10-meter height wind speed sequence to output the final corrected wind speed.

2. The method for constructing a ship standard altitude wind speed correction model according to claim 1, characterized in that, The multi-antenna BeiDou array consists of no less than three BeiDou receiving antennas. The antenna spatial layout is configured according to the three-axis direction of the carrier coordinate system. The absolute elevation of the main antenna is obtained by precise single-point positioning calculation of the BeiDou main antenna. The original relative wind speed is directly output by the shipborne wind sensor.

3. The method for constructing a ship standard altitude wind speed correction model according to claim 2, characterized in that, The weighted undirected graph uses each BeiDou receiving antenna as a vertex and the baseline between any two BeiDou receiving antennas as an edge. The edge weight is determined by the signal-to-noise ratio and geometric strength comprehensive score of the carrier phase difference observations of the corresponding baseline. The comprehensive score is obtained by weighted least squares fitting to obtain the signal-to-noise ratio weight coefficient and the geometric strength weight coefficient.

4. The method for constructing a ship standard altitude wind speed correction model according to claim 3, characterized in that, The neural-Kalman hybrid filter network consists of four parts: a long short-term memory module, a linear projection layer, a differentiable Kalman filter layer, and an attention gating module. The hidden state of the long short-term memory module is mapped to the process noise covariance matrix and the observation noise covariance matrix through two independent linear projection layers, respectively. The output of the projection layer is guaranteed to be positive definite by the Softplus activation function.

5. The method for constructing a ship standard altitude wind speed correction model according to claim 4, characterized in that, The steps for solving the generalized eigenvalue problem also include adopting a partial ambiguity fixing strategy, prioritizing the fixing of the baseline subset with the highest comprehensive score, fixing integer ambiguities by combining an improved least squares ambiguity decorrelation adjustment search tree pruning algorithm, and using an inter-epoch ambiguity inheritance mechanism to directly inherit the fixed baseline solution that is not affected by the cycle slip after the cycle slip.

6. The method for constructing a ship standard altitude wind speed correction model according to claim 5, characterized in that, The steps for solving the generalized eigenvalue problem also include establishing a multipath hemispherical map, statistically analyzing the average carrier phase residual of each BeiDou receiving antenna according to the two-dimensional grid of satellite elevation and azimuth angles to form a multipath prior template, subtracting the corresponding multipath prior template value from the table during real-time calculation, and using adaptive spectral subtraction to suppress the remaining residual in the frequency domain.

7. The method for constructing a ship standard altitude wind speed correction model according to claim 6, characterized in that, The selection of candidate models with the minimum description length is performed within a sliding time window. The candidate model set includes three categories: sinusoidal superposition model, autoregressive model, and piecewise linear model. The minimum description length criterion value is composed of the sum of the model parameter encoding length and the data condition encoding length. The optimal vertical motion model is used to extrapolate and predict the vertical displacement prediction value of the next epoch.

8. The method for constructing a ship standard altitude wind speed correction model according to claim 7, characterized in that, The wave age parameter is defined as the ratio of peak wave velocity to frictional wind speed. The piecewise functional relationship between the Channock coefficient and the wave age parameter is obtained by fitting peak wave velocity and frictional wind speed data under various typical sea conditions. The peak wave velocity is provided in real time by a shipboard wave measuring radar or buoy.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the method for constructing a ship standard height wind speed correction model according to any one of claims 1-8.

10. A system for constructing a ship standard altitude wind speed correction model, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images