Ocean-facing GNSS buoy environment perception measurement covariance dynamics modeling method

Abstract

The application discloses a marine GNSS buoy environment perception measurement covariance dynamics modeling method, and relates to the technical field of marine observation and satellite navigation. The method comprises the following steps: collecting buoy multi-source sensor data to extract a standardized feature vector; hierarchically encoding wave and tide and other multi-scale marine environment features; combining a wave state parameter driven attention mechanism to quantify the sea surface multipath effect; using a multilayer perception machine and a decomposition algorithm to deduce a process noise matrix; modeling the measurement noise covariance evolution as a differential equation driven by the matrix, and applying a sea state adaptive spectrum constraint; triggering a hierarchical safety fallback strategy according to the monitoring index of the covariance matrix; and injecting the safety covariance matrix into a fusion filter to perform parameter extraction and alarm. The application ensures the smoothness and stability of the covariance evolution from the mathematical structure, takes into account the environmental adaptive capability, and guarantees the reliable positioning and operation of the buoy in extreme sea conditions.

Description

Technical Field

[0001] This invention relates to the field of marine observation and satellite navigation technology, specifically to a method for modeling the covariance dynamics of marine GNSS buoy environmental perception measurement. Background Technology

[0002] GNSS buoys, as core infrastructure in modern ocean observation systems, primarily function as buoy platforms deployed on the sea surface, equipped with high-precision Global Navigation Satellite System (GNSS) receivers. These platforms enable monitoring of sea level changes over vast ocean areas, inversion of sea wave parameters, and early warning of extreme marine disasters such as tsunamis. In actual operation, the buoy platform continuously receives radio frequency signals transmitted by multiple navigation satellites and utilizes precise point positioning (PPP) or real-time dynamic differential positioning (RTK) techniques to acquire the three-dimensional spatial coordinate sequence of the buoy antenna in real time, thereby extracting data parameters reflecting real ocean dynamic processes.

[0003] In the multi-sensor fusion positioning process of GNSS buoys, the Extended Kalman Filter (EKF) is currently the most widely used data fusion framework. When performing the measurement update step, this filter heavily relies on the measurement noise covariance matrix to quantify the error level of each GNSS observation. When the value of the measurement noise covariance matrix is ​​lower than the actual noise level, the filter assigns excessively high confidence weights to noisy observations, directly leading to spurious jumps in the positioning results that do not match actual physical motion. Conversely, if the value of the matrix is ​​excessively higher than the actual noise level, the filter cannot fully utilize the high-precision observation information, thus lowering the overall positioning accuracy. Therefore, the ability to estimate the GNSS measurement noise covariance in real time and accurately in a dynamic ocean environment is a key factor determining the buoy positioning accuracy and the quality of subsequent ocean parameter inversion.

[0004] However, existing GNSS buoy positioning methods generally suffer from the technical limitation of failing to simultaneously consider the dynamic adaptability of the marine environment and the temporal smoothness of the covariance matrix evolution when estimating the measurement noise covariance matrix. On the one hand, while constant covariance estimation methods can guarantee temporal smoothness, they cannot reflect the true fluctuations in signal quality when the sea state (such as from calm sea to rough seas) changes drastically. On the other hand, while pure data-driven methods based on deep neural networks possess strong environmental perception and adaptive response capabilities, these methods often treat covariance estimation as a static regression problem at discrete moments, failing to establish reasonable dynamic constraints between covariance estimates at adjacent moments. This makes them prone to drastic temporal jumps during sudden environmental changes. Once these discontinuous covariance jumps are injected into the downstream positioning filter, they will directly cause numerical instability in the filter state update or even lead to divergence failure. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a dynamic modeling method for environmental perception measurement covariance of marine GNSS buoys. This method solves the problem that the measurement noise covariance matrix on which multi-sensor fusion positioning calculation of marine GNSS buoys relies in dynamic and long-term observation environments is difficult to simultaneously take into account environmental adaptability and temporal smoothness.

[0006] To achieve the above objectives, the present invention provides a method for modeling the covariance dynamics of marine GNSS buoy environmental perception measurement, comprising the following steps:

[0007] Collect multi-source sensor data from an ocean GNSS buoy platform, extract quality index vectors, estimate the buoy's motion state, and stitch together and output standardized feature vectors;

[0008] The short-term fluctuation component of signal quality caused by waves, the mid-time feature vector of diurnal variation of tides and ionosphere, and the long-term trend component of slow change of sea state and atmospheric conditions are extracted and spliced ​​to obtain a multi-scale environmental feature vector.

[0009] The query vector is obtained by summarizing the elevation angle distribution histogram of visible satellites. The sea surface roughness factor is calculated using the significant wave height and wind speed, and a key matrix is ​​generated. The attention weight is obtained by calculating the inner product similarity between the query vector and the key matrix, and the multipath sensing spatial embedding vector is output.

[0010] The standardized feature vector, multi-scale environmental feature vector, and multipath sensing space embedding vector are concatenated and input into a multilayer perceptron for encoding to generate a high-dimensional embedding vector. The process noise matrix is ​​then constructed based on the parameterization of the high-dimensional embedding vector using a decomposition algorithm.

[0011] The evolution of the GNSS measurement noise covariance matrix is ​​modeled as a differential equation driven by the process noise matrix. Spectral constraints are imposed on the system matrix, and the lower bound of the system matrix eigenvalues ​​is dynamically adjusted using the sea state index. Discretized recursive equations are then expanded to perform parallel computation of the covariance matrix elements.

[0012] Calculate the monitoring indicators of the covariance matrix, trigger graded safety rollback based on the degree of anomaly of the monitoring indicators, and output the safety covariance matrix after safety rollback;

[0013] The safety covariance matrix is ​​injected into the measurement update stage of the multi-sensor fusion filter to extract ocean observation parameters, and the abnormal water level alarm and the lower bound of the eigenvalue are automatically switched based on the rate of change of the residual sea level height.

[0014] Furthermore, the specific process of acquiring multi-source sensor data from the marine GNSS buoy platform, extracting quality index vectors, estimating the buoy's motion state, and splicing to output a standardized feature vector includes: extracting a quality index vector containing various precision attenuation factors, the number of visible satellites, and the average carrier-to-noise ratio of each satellite from multi-constellation GNSS receiver data; performing logarithmic transformation on various precision attenuation factors to compress the dynamic range, and normalizing the number of visible satellites by the maximum value in the sequence; estimating the buoy's motion state using inertial measurement unit data and the buoy motion model, calculating the radial distance and radial velocity in the local reference coordinate system and performing normalization; and splicing the normalized radial distance, normalized radial velocity, logarithmically transformed precision attenuation factors, normalized number of visible satellites, and average carrier-to-noise ratio to output a standardized feature vector.

[0015] Furthermore, the specific process of extracting the short-term fluctuation component of signal quality caused by waves, the mid-time feature vectors of the diurnal variation of tides and ionosphere, and the long-term trend component of the slowly changing sea state and atmospheric conditions, and concatenating them to obtain the multi-scale environmental feature vector includes: extracting the short-term fluctuation component of signal quality caused by waves through a high-pass filter; generating the mid-time feature vectors of the diurnal variation of tides and ionosphere using sine-cosine coding; extracting the long-term trend component of the slowly changing sea state and atmospheric conditions through a low-pass filter; and concatenating the short-term fluctuation component, the mid-time feature vector, and the long-term trend component to obtain the multi-scale environmental feature vector.

[0016] Further, the process of summarizing the elevation angle distribution histogram of visible satellites to obtain the query vector, calculating the sea surface roughness factor using significant wave height and wind speed to generate a key matrix, calculating the inner product similarity between the query vector and the key matrix to obtain the attention weight, and outputting the multipath sensing spatial embedding vector includes: summarizing the elevation angle distribution histogram of visible satellites at the current time and performing L2 normalization to obtain the query vector; calculating the sea surface roughness factor using significant wave height and wind speed to generate a key matrix that modulates the typical sea surface multipath pattern; calculating the inner product similarity between the query vector and the key matrix, obtaining the attention weight through low-temperature Softmax, and outputting the multipath sensing spatial embedding vector through weighted summation and linear projection.

[0017] Furthermore, the specific process of constructing the process noise matrix based on the parameterization of high-dimensional embedding vectors using the decomposition algorithm includes: constructing a symmetric positive definite process noise matrix based on the parameterization of high-dimensional embedding vectors using the LDL decomposition algorithm; ensuring that the diagonal elements of the diagonal positive matrix have strict numerical positivity through the Softplus function; and generating the off-diagonal elements of the unit lower triangular matrix through linear mapping.

[0018] Furthermore, the specific process of modeling the evolution of the GNSS measurement noise covariance matrix as a differential equation driven by the process noise matrix, imposing spectral constraints on the system matrix, dynamically adjusting the lower bound of the system matrix eigenvalues ​​using the sea state index, and performing parallel computation of the covariance matrix elements by expanding the discretized recursive equation includes: modeling the evolution of the GNSS measurement noise covariance matrix as a Lyapunov differential equation and imposing spectral constraints on the system matrix; calculating the sea state index based on normalized effective wave height and wind speed, and dynamically adjusting the lower bound of the system matrix eigenvalues ​​using the sea state index; expanding the discretized recursive equation using the diagonalized structure of the system matrix to obtain a scalar convolution structure, and performing parallel computation of the covariance matrix elements through zero-padding fast Fourier transform.

[0019] Further, the specific process of calculating the monitoring indicators of the covariance matrix, triggering graded safety backoff based on the degree of anomaly of the monitoring indicators, and outputting the safety covariance matrix after safety backoff includes: calculating the condition number indicator, trace change rate indicator, and minimum eigenvalue indicator of the covariance matrix; triggering graded safety backoff based on the degree of anomaly of the three monitoring indicators; calculating the weighted mean of the sliding time window to replace the current estimate under the first-level backoff, switching to conservative constant covariance replacement under the second-level backoff, marking the current location solution as unavailable and issuing an alarm under the third-level backoff; and outputting the safety covariance matrix after safety backoff and the corresponding quality flag.

[0020] Furthermore, the specific process of injecting the safety covariance matrix into the measurement update stage of the multi-sensor fusion filter to extract ocean observation parameters includes: injecting the safety covariance matrix into the measurement update stage of the multi-sensor fusion filter using an extended Kalman filter framework, using the safety covariance matrix to control the weight of GNSS observations in the state update; and extracting residual sea level height, significant wave height, mean wave period, and main wave direction from the high-precision three-dimensional position time series output by the multi-sensor fusion filter.

[0021] Furthermore, the specific process of automatically switching between abnormal water level alarms and eigenvalue lower bounds based on the rate of change of residual sea level height includes: when the rate of change of residual sea level height exceeds a preset threshold and the duration meets the judgment window, an abnormal water level alarm is issued, and the lower bound of the system matrix eigenvalues ​​is automatically switched to the lower bound of tsunami eigenvalues, allowing the covariance matrix to shrink rapidly.

[0022] This invention provides a method for modeling the covariance dynamics of marine GNSS buoy environmental perception measurements. It offers the following advantages:

[0023] 1. This invention models the evolution of the GNSS measurement noise covariance matrix at discrete time as a continuous-time Lyapunov differential equation, applies spectral constraints to the system matrix, and dynamically adjusts the lower bound of the system matrix eigenvalues ​​using a sea state index constructed based on significant wave height and wind speed. This limits the shrinkage rate of the logarithmic determinant of the covariance matrix, achieving a balance between the temporal smoothness of the measurement noise covariance evolution and environmental adaptability in the marine environment, and avoiding the divergence of downstream multi-sensor fusion filters caused by non-physical jumps in covariance.

[0024] 2. This invention obtains a query vector by summarizing the elevation angle distribution histogram of visible satellites and performing L2 normalization. It calculates the sea surface roughness factor using the effective wave height and wind speed, and generates a key matrix that modulates the typical sea surface multipath pattern modulated by the sea surface roughness factor. Then, it calculates the inner product similarity between the query vector and the key matrix to extract attention weights. This realizes a quantitative expression of the mapping relationship between wave state and low elevation angle satellite multipath intensity, and overcomes the technical defect that multipath error is difficult to quantify and model in the dynamic observation environment of ocean buoys.

[0025] 3. This invention calculates the condition number index, trace change rate index, and minimum eigenvalue index of the covariance matrix in real time, and triggers a hierarchical safety backoff strategy based on the degree of anomaly of the above three monitoring indicators, including weighted mean substitution, conservative constant covariance substitution, and marking the unavailability of the localization solution, to output a safe covariance matrix. This achieves the interception of abnormal covariance output states in long-term observation tasks, ensuring the continuous availability of data and the reliability of system operation of unattended marine GNSS buoy platforms under harsh sea conditions. Attached Figure Description

[0026] Figure 1 This is a graph showing the experimental results of covariance prediction in this invention;

[0027] Figure 2 The figure shows the experimental results of exponential convergence verification under different initial covariances of the present invention;

[0028] Figure 3 This is a diagram showing the experimental results of the sea state adaptive spectrum constraint effect of the present invention;

[0029] Figure 4 This is a diagram showing the experimental results of the EKF positioning root mean square error comparison of the present invention;

[0030] Figure 5 This is a flowchart illustrating the technical route of the present invention. Detailed Implementation

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

[0032] Please see the appendix Figure 5 This invention provides a method for dynamic modeling of covariance in marine GNSS buoy environmental perception measurements. By constructing an exponentially stable dynamic system driven by the marine environment, the method models the GNSS measurement noise covariance as a physical process that evolves continuously over time, thus balancing the rapid adaptability of the marine environment with the temporal smoothness of the covariance evolution. The specific implementation process includes:

[0033] Step S1: Multi-source buoy observation data acquisition and preprocessing. Multi-source sensor data is acquired from the marine GNSS buoy platform, including data from multi-constellation GNSS receivers, inertial measurement units (IMUs), and marine meteorological sensors. Accuracy attenuation factor, visible satellite count, and carrier-to-noise ratio are extracted from the multi-constellation GNSS receiver data. Logarithmic transformation is used to compress the dynamic range of the accuracy attenuation factor, and the visible satellite count is normalized to its maximum value. Simultaneously, the six-degree-of-freedom motion state of the buoy is estimated using IMU data combined with the buoy motion model. The radial distance and radial velocity of the buoy relative to the mooring center are calculated and normalized. Finally, a standardized feature vector is output.

[0034] Step S2: Multi-scale marine environmental feature encoding. Marine environmental factors affecting GNSS measurement quality are divided into three layers according to time scale for feature encoding. A high-pass filter is used to extract short-time fluctuation components of signal quality caused by waves; sine-cosine coding is used to extract mid-time quasi-periodic features of tidal and ionospheric diurnal variations; and a low-pass filter is used to extract long-time trend components of slowly changing sea state and atmospheric conditions. The short-time fluctuation components, mid-time quasi-periodic features, and long-time trend components are concatenated to obtain a multi-scale environmental feature vector.

[0035] Step S3: Sea Surface Multipath Attention Embedding Based on Wave State Parameters. Construct a query vector containing the elevation angle distribution of currently visible satellites and perform L2 normalization. Calculate the sea surface roughness factor using significant wave height and wind speed, and generate a key vector encoding the multipath pattern through sea surface roughness factor modulation. Obtain the attention weight by calculating the similarity of the inner product between the query vector and the key vector. Perform weighted summation and linear projection on the value matrix to output the multipath sensing spatial embedding vector.

[0036] Step S4: Positive Definite Process Noise Matrix Inference. The multi-scale environmental feature vector, multipath sensing space embedding vector, and quality index vector are concatenated and input into a multilayer perceptron for encoding into a high-dimensional embedding vector. Based on the high-dimensional embedding vector, a strictly symmetric positive definite process noise matrix is ​​constructed and output using the LDL decomposition algorithm.

[0037] Step S5: Lyapunov Covariance Dynamics Evolution and Sea State Adaptive Spectral Constraints. The evolution of the GNSS measurement noise covariance is modeled as a Lyapunov differential equation. Spectral constraints are imposed on the system matrix to ensure the exponential stability of the system and the shrinkage rate of the logarithmic determinant of the covariance. A sea state adaptive mechanism is introduced, defining a sea state index using significant wave height and wind speed. The lower bound of the eigenvalues ​​of the system matrix is ​​dynamically adjusted based on the sea state index. Discretized recursive equations are expanded using the diagonalized structure of the system matrix, and parallel computation is performed using a fast Fourier transform.

[0038] Step S6: Covariance Anomaly Detection and Graded Safety Rollback. The condition number, trace change rate, and minimum eigenvalue indices are calculated in real-time for the covariance matrix. Graded safety rollback is triggered based on the degree of anomaly detection for these three indices: for minor anomalies, a weighted mean of the covariance matrix over a sliding time window is used; for moderate anomalies, a conservative constant covariance is used; for severe anomalies, the localized solution is marked as unavailable and an alarm is issued. After rollback processing, a safety-verified covariance matrix is ​​output.

[0039] Step S7: Navigation Filter Integration and Ocean Parameter Output. The safety-verified covariance matrix is ​​injected into the measurement update stage of the multi-sensor fusion filter to control the weight of GNSS observations in state updates. Residual sea level height, significant wave height, mean wave period, and dominant wave direction are extracted from the buoy's 3D position time series output by the filter. A tsunami alarm is triggered when the rate of change and duration of the residual sea level height meet the criteria, and the spectral constraint in Step S5 is automatically switched to a fast response mode that allows for rapid contraction of the covariance matrix.

[0040] Step S1: Multi-source buoy observation data acquisition and preprocessing. This step aims to acquire multi-source sensor data from the buoy platform and extract standardized feature vectors to characterize signal quality and motion state. The specific implementation process includes:

[0041] In step S101, multi-source sensor data is acquired on the marine GNSS buoy platform. The data includes raw observations and navigation solutions output from multi-constellation GNSS receivers, acceleration and angular velocity output from the inertial measurement unit, and environmental data output from marine meteorological sensors.

[0042] In step S102, a quality index vector is extracted from the GNSS receiver data. , This represents the dimension of the quality metric vector. The quality metric vector includes various accuracy attenuation factors and the number of visible satellites. and the average carrier-to-noise ratio of each satellite Various precision attenuation factors include geometric precision attenuation factor, horizontal precision attenuation factor, and vertical precision attenuation factor.

[0043] In step S103, a logarithmic transformation is performed on various precision attenuation factors to compress the dynamic range:

[0044]

[0045] in, Indicates the first Precision attenuation factor in The value at time, It is the natural logarithm function.

[0046] Normalize the number of visible satellites by the maximum value in the sequence:

[0047] in, This represents the maximum number of visible satellites in a time series. This is a time-series index.

[0048] In step S104, the six-degree-of-freedom motion state of the buoy is estimated using inertial measurement unit data and the buoy motion model. The buoy position is then determined. and speed Project the coordinates onto the local reference coordinate system of the mooring area where the buoy is located. Calculate the radial distance of the buoy relative to the center of the mooring area in the local reference coordinate system. and radial velocity and radial distance and radial velocity Normalization to Interval.

[0049] In step S105, the standardized feature vector is output. , This represents the dimension of the standardized feature vector. The standardized feature vector is obtained by concatenating the following components:

[0050]

[0051] Step S2: Marine Multi-Scale Environmental Feature Coding. This step aims to encode the marine environmental factors affecting GNSS measurement quality into three layers based on time scales such as waves, tides, and sea state. The specific implementation process includes:

[0052] In step S201, the short-time fluctuation component of the signal quality caused by waves is extracted. This is measured by the buoy attitude sensor. Antenna normal vector offset angle at time 1 , Vertical acceleration at time as well as Carrier-to-noise ratio at time As input, the signal quality fluctuation component in the wave frequency band is extracted using a high-pass filter. In the formula, This indicates the current moment. The time period range corresponding to the wave frequency band is 0.5 seconds to 25 seconds. The formula for calculating the short-time fluctuation component is as follows:

[0053]

[0054] in, express The short-term fluctuation component at time, Indicates the cutoff frequency as High-pass filtering operation, Determined based on the wave response transfer function of the buoy. The value ranges from 0.04 Hz to 2.0 Hz.

[0055] In step S202, the mid-time quasi-periodic characteristics of tidal and ionospheric diurnal variations are extracted. The current time... Mapped to tidal phase angle and solar hour angle The mapping formula is as follows:

[0056]

[0057]

[0058] in, The tidal period of the semi-diurnal tide M2 ​​is 12.42 hours. This represents the solar day cycle, with a value of 24 hours. This indicates that the decimal part of the input value is taken.

[0059] The mid-time feature vector is generated using sine-cosine coding. The formula for calculating the mid-time feature vector is as follows:

[0060]

[0061] in, express The mid-time eigenvector at time t. It is a sine function. It is a cosine function.

[0062] In step S203, the long-term trend components of slowly changing sea state and atmospheric conditions are extracted. Effective wave height at time , 10-meter wind speed at sea at any given time and The swell cycle of time As input, a low-pass filter is used to extract the trend component. The formula for calculating the trend component is as follows:

[0063]

[0064] in, express The long-term trend component of time. Indicates the cutoff frequency as The low-pass filtering operation, The range of values ​​is to , Represents hours.

[0065] In step S204, the three-layer features are concatenated to obtain a multi-scale environmental feature vector. The short-time fluctuation components are then... , mid-time eigenvectors and long-term trend components By concatenating the vectors, we obtain a multi-scale environmental feature vector:

[0066]

[0067] in, Represents multi-scale environmental feature vectors. The dimension is The real space, is the dimension of the multi-scale environmental feature vector.

[0068] Step S3: Sea Surface Multipath Attention Embedding Based on Wave State Parameters. This step aims to construct an attention mechanism using wave state parameters such as significant wave height and wind speed to quantify the dynamic interference intensity of sea surface multipath effects on satellite signals at different elevation angles. The specific implementation process includes:

[0069] In step S301, a query vector for the satellite elevation angle distribution is defined. The current time... visible satellite elevation angle This is summarized into an elevation angle distribution histogram. The satellite elevation angle range from 0 degrees to 90 degrees is divided into... Each elevation angle interval covers [number] elevation angle intervals. The range of elevation angles. The number of satellites within each elevation angle range is counted as the elevation angle distribution vector. For the elevation angle distribution vector L2 normalization is performed to obtain the query vector. The calculation formula is as follows:

[0070]

[0071] in, Represents the elevation angle distribution vector The L2 norm.

[0072] In step S302, a wave-state-based bond vector is constructed. This is done using the current time step. effective wave height and wind speed Calculate the sea surface roughness factor :

[0073]

[0074] in, and For learnable parameters, This is a Softplus function used to guarantee the sea surface roughness factor. .generate The key vectors constitute the key matrix. Each key vector encodes a typical sea surface multipath pattern, ranging from high reflectivity on calm sea surfaces to scattered reflectivity on rough sea surfaces. The key values ​​are determined by the sea surface roughness factor. Modulation, the first row key vector The calculation formula is as follows:

[0075]

[0076] in, and The first learnable parameter matrix OK.

[0077] In step S303, attention weights and output are calculated. Query vector With the key matrix composed of key vectors Similarity is calculated using the inner product, and attention weights are obtained through low-temperature softmax. :

[0078]

[0079] in, Temperature parameters The value ranges from 0.01 to 0.5. Temperature parameter. The lower the value, the more focused the attention is on the multipath pattern that best matches the current elevation angle distribution. The attention output is a value matrix. The weighted summation, followed by linear projection, yields the multipath sensing spatial embedding vector. :

[0080]

[0081] in, To output the projection matrix, For the transpose of the value matrix, The dimension of the embedding vector in the multipath sensing space.

[0082] Step S4: Positive Definite Process Noise Matrix Inference. This step aims to input the fused multidimensional features into a multilayer perceptron and parameterize the output process noise matrix, which strictly satisfies the symmetric positive definite property, using the LDL decomposition algorithm. The specific implementation process includes:

[0083] In step S401, the multi-scale environmental feature vector is... Multipath sensing spatial embedding vector and quality index vector The concatenated vectors are input into a multilayer perceptron and encoded as high-dimensional embedding vectors. High-dimensional embedding vectors The calculation formula is as follows:

[0084]

[0085] in, This represents a multilayer perceptron. The dimension is The high-dimensional embedding vectors. The multilayer perceptron contains 2 to 4 hidden layers, each with a dimension of 64 to 256, and uses ReLU as the activation function.

[0086] In step S402, the symmetric positive definite process noise matrix is ​​output through LDL decomposition parameterization. :

[0087]

[0088] in, It is a unit lower triangular matrix. It is a diagonal positive matrix. Represents a unit lower triangular matrix The transpose of . Let be the dimension of the covariance matrix. For 3D GNSS positioning, take . .

[0089] In step S403, the diagonal positive matrix The diagonal elements are guaranteed to be strictly positive using the Softplus function. The formula for calculating the diagonal elements is as follows:

[0090]

[0091] in, Represents a diagonal positive matrix The Middle Line number The diagonal elements of the column, For the Softplus function, For learnable parameter vectors, Represents the learnable parameter vector The transpose of . Unit lower triangular matrix The off-diagonal elements are obtained by linear mapping:

[0092]

[0093] in, Represents a unit lower triangular matrix The Middle Line number off-diagonal elements of a column, condition Indicates row index Strictly greater than column index , For learnable parameter vectors, Represents the learnable parameter vector The transpose of .

[0094] The construction method of LDL decomposition guarantees the process noise matrix For any high-dimensional embedding vector All are symmetric positive definite matrices.

[0095] Step S5: Lyapunov Covariance Dynamics Evolution and Sea State Adaptive Spectral Constraints. This step aims to model the covariance evolution as a continuous differential equation and introduce a sea state adaptive mechanism to dynamically adjust the spectral constraints, ensuring the time smoothness and exponential stability of the covariance evolution system. The specific implementation process includes:

[0096] In step S501, the GNSS measurement noise covariance matrix is... The evolution model is represented by Lyapunov differential equations:

[0097]

[0098] in, Represents the GNSS measurement noise covariance matrix The derivative with respect to time, Represents the system matrix. Representing the system matrix transpose, This represents the noise matrix during the inference process. This represents the dimension of the covariance matrix.

[0099] In step S502, the system matrix is... Apply spectral constraints. (The system matrix is ​​then...) Parameterization into diagonalizable form ,in Represents an eigenvalue diagonal matrix. Let represent an invertible matrix. The real parts of all eigenvalues ​​are constrained to satisfy:

[0100]

[0101] in, Denotes the lower bound of the real part of the eigenvalues. Indicates the first The real parts of each eigenvalue.

[0102] The rate of contraction of the logarithmic determinant of the covariance is set to satisfy:

[0103]

[0104] in, Indicates the maximum contraction rate. Represents the GNSS measurement noise covariance matrix The determinant of the eigenvalues. When the lower bound of the eigenvalues ​​satisfies At that time, the smoothness constraint is guaranteed by the system structure.

[0105] In step S503, a sea state adaptive mechanism is introduced to dynamically adjust the lower bound of the eigenvalue. Define the sea state rating index. :

[0106]

[0107] in, Indicates the normalized effective wave height. Indicates the normalized wind speed. and The weighting coefficients and Sea State Rating Index The higher the value of the sea state index, the worse the sea conditions.

[0108] Adaptive eigenvalue lower bound The calculation formula is:

[0109]

[0110] in, The characteristic value representing the lower bound of calm sea state has a large absolute value, allows for rapid contraction, and its range is [value missing]. to ; The characteristic value representing the lower bound of severe sea state is small in absolute value, conservative and smooth, and its range is [value range missing]. to .

[0111] In step S504, during the GNSS sampling period The discretized Lyapunov recurrence equation is as follows:

[0112]

[0113] in, and They represent the first The discrete time and the first GNSS measurement noise covariance matrix at discrete time points Represents the discretized system matrix. This represents the process noise matrix in the discretization process.

[0114] By utilizing the diagonalized structure of the system matrix, the discretized recurrence equations are expanded to obtain:

[0115]

[0116] in, , , and The eigenvalues ​​are denoted as . The expanded convolutional structure eliminates temporal dependencies and is computed in parallel via zero-padding Fast Fourier Transform.

[0117] In step S505, the training loss function is calculated. The training loss function is a weighted sum of the negative log-likelihood and the smoothing penalty:

[0118]

[0119] in, Represents the negative log-likelihood term. This represents the smoothing penalty term. This represents the smoothing penalty weight.

[0120] The negative log-likelihood term is defined as:

[0121]

[0122] in, Indicates the length of the time series. Represents the covariance matrix The determinant, Indicates GNSS measurement residuals. Indicates the actual location. Indicates the GNSS measurement location.

[0123] The smoothing penalty term uses the squared hinge loss:

[0124]

[0125] The smoothing penalty term was used only for comparison in ablation experiments.

[0126] Step S6: Covariance Anomaly Detection and Graded Safety Rollback. This step aims to monitor the health status of the covariance matrix in real time and trigger different levels of safety rollback strategies based on the degree of anomaly to ensure the reliability of the long-term observation system. The specific implementation process includes:

[0127] In step S601, at each time... Calculate the following three monitoring indicators. Calculate the condition number indicator. :

[0128]

[0129] in, Indicates time covariance matrix The largest eigenvalue, Indicates time covariance matrix The minimum eigenvalue, condition number index It reflects the numerical stability of the covariance matrix.

[0130] Calculate the trace value change rate index :

[0131]

[0132] in, Indicates time The trace of the covariance matrix Indicates time The trace of the covariance matrix This indicates the absolute value operation, specifically the trace value change rate index. It reflects the relative rate of change of covariance.

[0133] Calculate the minimum eigenvalue index :

[0134]

[0135] Minimum eigenvalue index It reflects the positive definiteness margin of the covariance matrix.

[0136] In step S602, a graded safety rollback is triggered based on the degree of abnormality of the three monitoring indicators. A preset threshold for the number of mild abnormality conditions is set. Threshold for rate of change of mildly abnormal trace values Threshold for the number of moderately abnormal conditions Minimum eigenvalue threshold for moderate anomalies and the threshold for the number of severe abnormal conditions .

[0137] When the condition number index or trace value change rate index At this time, a level 1 rollback is triggered. A level 1 rollback uses the most recent... The weighted mean of the covariance matrix at each time step replaces the current estimate, outputting the safe covariance matrix. :

[0138]

[0139] in, The time window length, Indicates time The covariance matrix, For exponentially decaying weights, the exponentially decaying weights satisfy... , This is the attenuation factor.

[0140] When the condition number index or minimum eigenvalue index When this occurs, a second-level backoff is triggered. The second-level backoff switches to a conservative constant covariance, outputting a safe covariance matrix. :

[0141]

[0142] in, This is the upper bound of the mean of the trace values ​​of the recent covariance matrix. for 3D identity matrix Let be the dimension of the covariance matrix.

[0143] When the positive definiteness test of the covariance matrix fails, i.e., the minimum eigenvalue index fails... or conditional index When this happens, a level 3 rollback is triggered. The level 3 rollback marks the current location solution as unavailable and issues a system alarm signal.

[0144] In step S603, the safety covariance matrix is ​​output after the safety rollback. and corresponding quality marks Safety covariance matrix Corresponding discrete time The obtained safety covariance matrix Quality Mark These correspond to normal, level 1 rollback, level 2 rollback, and level 3 alarm, respectively.

[0145] Step S7: Navigation Filter Integration and Ocean Parameter Output. This step aims to inject the covariance matrix after safe rollback into the positioning filter, thereby extracting high-precision ocean observation parameters and executing abnormal water level alarms based on the residual sea level height. The specific implementation process includes:

[0146] In step S701, the security covariance matrix that has been verified for security is... The measurement update stage is injected into the multi-sensor fusion filter. The multi-sensor fusion filter employs an extended Kalman filter framework, and the state vector includes the buoy position. ,speed and posture In the GNSS measurement update step, the Kalman gain is:

[0147]

[0148] in, For state prediction covariance, For the measurement matrix, To measure the transpose of the matrix, Discrete time The corresponding security covariance matrix, Invert a matrix. Safety covariance matrix. Directly control the weight of GNSS observations in state updates.

[0149] In step S702, ocean observation parameters are extracted from the high-precision buoy three-dimensional position time series output by the extended Kalman filter. For sea level height, the low-frequency component of the buoy's vertical coordinate is taken, where the cutoff period of the low-frequency component is greater than the maximum wave period. After subtracting the tidal model prediction value, the residual sea level height is obtained. For wave parameters, spectral analysis is performed on the wave frequency band components of the vertical coordinate to extract the significant wave height. Mean wave period And the main wave direction. Significant wave height. The mean wave period is the square root of the zeroth moment of the spectrum. It is the square root of the ratio of the zeroth moment to the second moment of the spectrum.

[0150] In step S703, when the rate of change of residual sea level exceeds a preset threshold and the duration exceeds the judgment window, an abnormal water level alarm is triggered. In tsunami warning mode, the spectral constraint in step S5 automatically switches to fast response mode, temporarily lowering the eigenvalue bound. Set as . The absolute value is relatively large, ranging from -5.0 to -2.0. This allows for rapid shrinkage of the covariance matrix to avoid smoothing constraints masking real sea surface anomaly signals.

[0151] Based on the same inventive concept as the aforementioned method embodiments, this invention provides a modeling device for environmental perception measurement covariance dynamics of marine GNSS buoys. Each functional module in the modeling device corresponds one-to-one with each step in the aforementioned method embodiments, and the entire processing logic of the environmental perception measurement covariance dynamics modeling method is implemented by running a computer program. Specifically, the modeling device includes the following processing modules:

[0152] The multi-source buoy observation data acquisition and preprocessing module is used to acquire multi-source sensor data from the marine GNSS buoy platform. This multi-source sensor data includes raw observations and navigation solutions output from multi-constellation GNSS receivers, acceleration and angular velocity output from the inertial measurement unit (IMU), and environmental data output from marine meteorological sensors. The module extracts a quality index vector from the multi-constellation GNSS receiver data, containing various accuracy attenuation factors, the number of visible satellites, and the average carrier-to-noise ratio (CNR) of each satellite. It performs logarithmic transformation on the accuracy attenuation factors and normalizes the number of visible satellites by the maximum value in the sequence. The module also estimates the buoy's motion state using IMU data and the buoy motion model. In a local reference coordinate system, it calculates and normalizes the radial distance and radial velocity of the buoy relative to the mooring center. The normalized radial distance, normalized radial velocity, logarithmically transformed accuracy attenuation factors, normalized number of visible satellites, and average CNR are then concatenated into a standardized feature vector.

[0153] The marine multi-scale environmental feature coding module extracts short-term fluctuation components of signal quality caused by waves using a high-pass filter, generates mid-time feature vectors of diurnal tidal and ionospheric variations using sine-cosine coding, and extracts long-term trend components of slowly changing sea state and atmospheric conditions using a low-pass filter. The module then concatenates the short-term fluctuation components, mid-time feature vectors, and long-term trend components to obtain the multi-scale environmental feature vector.

[0154] The wave-state parameter-based sea surface multipath attention embedding module is used to summarize the elevation angles of visible satellites at the current moment into an elevation angle distribution histogram and perform L2 normalization to obtain a query vector. It calculates the sea surface roughness factor using significant wave height and wind speed, and generates a key matrix consisting of key vectors modulated by the sea surface roughness factor that encode typical sea surface multipath patterns. The module then calculates the similarity between the inner product of the query vector and the key matrix, obtains the attention weights through low-temperature Softmax, performs weighted summation and linear projection on the value matrix, and outputs a multipath sensing spatial embedding vector.

[0155] The positive definite process noise matrix inference module is used to concatenate multi-scale environmental feature vectors, multipath sensing space embedding vectors, and quality index vectors, and then input them into a multilayer perceptron for encoding into a high-dimensional embedding vector. The module utilizes the LDL decomposition algorithm to parameterize and output a symmetric positive definite process noise matrix based on the high-dimensional embedding vector. In the LDL decomposition, the diagonal elements of the diagonal positive matrix are guaranteed to have strict numerical positivity using the Softplus function, and the off-diagonal elements of the unit lower triangular matrix are obtained through linear mapping.

[0156] The Lyapunov Covariance Dynamics Evolution and Sea State Adaptive Spectral Constraint module models the evolution of the GNSS measurement noise covariance matrix as a Lyapunov differential equation, imposes spectral constraints on the system matrix, and defines a sea state index based on normalized significant wave height and wind speed. This index is then used to dynamically adjust the lower bound of the system matrix eigenvalues. The module utilizes the diagonalized structure of the system matrix to expand the discretized recursive equations, obtaining a scalar convolution structure that eliminates time-series dependencies. It then computes the covariance matrix elements in parallel using a zero-padding fast Fourier transform.

[0157] The covariance anomaly detection and graded safety backoff module calculates the condition number, trace change rate, and minimum eigenvalue indices for the covariance matrix, triggering graded safety backoffs based on the degree of anomaly in these three indicators. When triggering a first-level backoff, the module replaces the current estimate with the weighted mean of the covariance matrix over a sliding time window; when triggering a second-level backoff, it switches to a conservative constant covariance; and when triggering a third-level backoff, it marks the current solution as unavailable and issues an alarm. The module outputs the safety covariance matrix after safety backoff and its corresponding quality flags.

[0158] The navigation filtering integration and ocean parameter output module injects a safety covariance matrix into the measurement update stage of a multi-sensor fusion filter employing an extended Kalman filter framework. This safety covariance matrix controls the weight of GNSS observations in the state update process. The module extracts residual sea level height, significant wave height, mean wave period, and dominant wave direction from the high-precision 3D position time series output by the filter. When the rate of change of residual sea level height exceeds a preset threshold and the duration meets a judgment window, the module triggers an abnormal water level alarm and automatically switches the lower bound of the system matrix eigenvalues ​​to the lower bound of tsunami eigenvalues ​​with larger absolute values, allowing for rapid contraction of the covariance matrix.

[0159] Furthermore, the present invention provides a computer-readable storage medium. A computer program is stored on the computer-readable storage medium. When the computer program is executed by a processor, it implements all the steps in the aforementioned method for modeling the covariance dynamics of environmental perception measurement for marine GNSS buoys.

[0160] When a computer program is executed by a processor, the following processing logic is implemented:

[0161] Multi-source sensor data is collected from an ocean GNSS buoy platform, including data from multiple constellation GNSS receivers, inertial measurement units (IMUs), and marine meteorological sensors. A quality index vector containing various accuracy attenuation factors, the number of visible satellites, and the average carrier-to-noise ratio (CNR) of each satellite is extracted from the multi-constellation GNSS receiver data. Logarithmic transformation is applied to various accuracy attenuation factors to compress the dynamic range, and the number of visible satellites is normalized to the maximum value in the sequence. The buoy's motion state is estimated using IMU data and a buoy motion model, and the radial distance and radial velocity in the local reference coordinate system are calculated and normalized. The normalized radial distance, normalized radial velocity, logarithmically transformed accuracy attenuation factors, normalized number of visible satellites, and average CNR are concatenated to output a standardized feature vector.

[0162] Short-time fluctuation components of signal quality caused by waves are extracted using a high-pass filter. Sine-cosine coding is used to generate mid-time feature vectors for diurnal tidal and ionospheric variations. Long-time trend components of slowly changing sea state and atmospheric conditions are extracted using a low-pass filter. The short-time fluctuation components, mid-time feature vectors, and long-time trend components are concatenated to obtain a multi-scale environmental feature vector.

[0163] The query vector is obtained by summarizing the elevation angle distribution histogram of visible satellites at the current time and performing L2 normalization. The sea surface roughness factor is calculated using significant wave height and wind speed. A key matrix encoding typical sea surface multipath patterns, modulated by the sea surface roughness factor, is generated. The similarity between the inner product of the query vector and the key matrix is ​​calculated, and attention weights are obtained through low-temperature softmax. Finally, a multipath sensing spatial embedding vector is output through weighted summation and linear projection.

[0164] The standardized feature vector, multi-scale environmental feature vector, and multipath sensing space embedding vector are concatenated and input into a multilayer perceptron to generate a high-dimensional embedding vector. A symmetric positive definite process noise matrix is ​​constructed based on the parameterization of the high-dimensional embedding vector using the LDL decomposition algorithm. The Softplus function ensures that the diagonal elements of the diagonal positive matrix have strict numerical positivity, and the off-diagonal elements of the unit lower triangular matrix are generated through linear mapping.

[0165] The evolution of the GNSS measurement noise covariance matrix is ​​modeled as a Lyapunov differential equation. Spectral constraints are imposed on the system matrix. Sea state indexes based on normalized significant wave height and wind speed are calculated, and the lower bounds of the system matrix eigenvalues ​​are dynamically adjusted using these sea state indexes. The discretized recursive equations are expanded using the diagonalized structure of the system matrix to obtain a scalar convolution structure, and parallel computation of the covariance matrix elements is performed using a zero-padding fast Fourier transform.

[0166] Calculate the condition number, trace rate of change, and minimum eigenvalue index of the covariance matrix. Trigger tiered safety backoff based on the degree of anomaly of these three monitoring indicators. In Level 1 backoff, calculate the weighted mean of the sliding time window to replace the current estimate; in Level 2 backoff, switch to conservative constant covariance replacement; and in Level 3 backoff, mark the current location solution as unavailable and issue an alarm. Output the safety covariance matrix after safety backoff and the corresponding quality flags.

[0167] A safety covariance matrix is ​​injected into the measurement update stage of a multi-sensor fusion filter employing an extended Kalman filter framework, using the safety covariance matrix to control the weight of GNSS observations in state updates. Residual sea level height, significant wave height, mean wave period, and dominant wave direction are extracted from the high-precision 3D position time series output by the multi-sensor fusion filter. When the rate of change of residual sea level height exceeds a preset threshold and the duration meets a judgment window, an abnormal water level alarm is issued, and the lower bound of the system matrix eigenvalues ​​is automatically switched to the lower bound of tsunami eigenvalues ​​with larger absolute values, allowing for rapid contraction of the covariance matrix.

[0168] Computer-readable storage media include portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, optical disc read-only memory, magneto-optical disc, magnetic tape, or any physical combination of the aforementioned media. A computer program contains instructions that are compiled or interpreted into a machine-readable sequence of instructions that can be executed directly or indirectly by a processor.

[0169] To verify the effectiveness of the method of the present invention in actual marine environments, in conjunction with the appendix to the specification... Figure 1 To be continued Figure 4 Table 1 verifies the experimental data and results.

[0170] The experimental data originated from observations by GNSS buoys deployed in a nearshore area of ​​my country. The buoys were equipped with a multi-constellation GNSS receiver, a MEMS inertial measurement unit, and a marine meteorological sensor. The multi-constellation GNSS receiver supported joint positioning using BeiDou-3, GPS, GLONASS, and Galileo systems. The GNSS receiver had a sampling rate of 1Hz, the MEMS inertial measurement unit had a sampling rate of 100Hz, and the meteorological data had a sampling rate of 0.1Hz. The experimental data covered various sea state conditions, including significant wave heights from 0.3 m to 4.5 m and wind speeds from 1 m / s to 18 m / s. The post-processing solution from precise single-point positioning was used as the reference true value.

[0171] Three sets of comparative models were set up: a constant covariance model, an adaptive model based on an accuracy attenuation factor, and a multilayer perceptron direct prediction model. The measurement noise covariance matrix of the constant covariance model satisfies... , The scalar is determined by minimizing the global negative log-likelihood. The identity matrix is ​​used. The measurement noise covariance matrix of the adaptive model based on the accuracy attenuation factor is proportional to the horizontal and vertical accuracy attenuation factors. The multilayer perceptron direct prediction model adopts the same structure as the core network of this invention, but the multilayer perceptron direct prediction model has no Lyapunov dynamic constraints.

[0172] The overall performance comparison of each model is shown in Table 1:

[0173] Table 1: Comparison of the overall performance of different covariance models

[0174]

[0175] The accuracy of covariance prediction is evaluated based on Table 1. The evaluation metric used is the mean negative log-likelihood, calculated as follows:

[0176]

[0177] In the formula, Indicates the total number of sampling points. Indicates time The predicted measurement noise covariance matrix, Indicates time The GNSS measurement residuals. The mean negative log-likelihood of the method of this invention is 0.4501, which is lower than 1.3883 for the constant covariance model, 0.8743 for the adaptive model based on the accuracy attenuation factor, and 1.0398 for the multilayer perceptron direct prediction model. See attached. Figure 1 As shown, the measurement noise covariance matrix predicted by the method of the present invention has a higher degree of agreement with the actual measurement uncertainty fluctuation than the comparative model.

[0178] The smoothness of covariance is evaluated using Table 1. The evaluation metric is the root mean square value of the rate of change of the covariance trace, calculated as follows:

[0179]

[0180] In the formula, This represents the trace operation on the matrix. The root mean square value of the trace change rate of the time series covariance matrix output by the method of this invention is 0.0779, which is lower than 1.4492 for the multilayer perceptron direct prediction model.

[0181] Verify the exponential stability of the dynamic framework. (See attached document.) Figure 2 As shown, for different initial measurement noise covariance matrix settings, the time series of measurement noise covariance matrices output by the method of this invention can all converge to the same evolution path within 60 seconds.

[0182] Verify the effectiveness of the sea state adaptive spectrum constraint. By comparing the sea state adaptive spectrum constraint with the fixed spectrum constraint, as shown in the attached figure... Figure 3 As shown, under the condition of small effective wave height, the absolute value of the lower bound of the adaptive eigenvalue is large; under the condition of large effective wave height, the absolute value of the lower bound of the adaptive eigenvalue decreases.

[0183] Combine Table 1 and Appendix Figure 4 The navigation and positioning accuracy was evaluated. The average root mean square error (RMSE) of position was used as the evaluation metric. After injecting the safety covariance matrix generated by the method of this invention into the extended Kalman filter, the average RMSE of the method of this invention was 0.3202 meters, which is lower than the 0.3636 meters of the multilayer perceptron direct prediction model. In severe sea conditions, the positioning RMSE of the method of this invention remained within 0.32 meters.

[0184] The basic mathematical derivations of the multi-constellation GNSS positioning principle, extended Kalman filter framework, multilayer perceptron structure, high-pass and low-pass filter design, fast Fourier transform algorithm implementation, and spectrum analysis involved can be directly implemented by those skilled in the art using existing technical means according to actual needs. The aforementioned processes are well-known technologies in this field and will not be elaborated here.

[0185] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for modeling the covariance dynamics of marine GNSS buoy environmental perception measurement, characterized in that... Includes the following steps: Collect multi-source sensor data from an ocean GNSS buoy platform, extract quality index vectors, estimate the buoy's motion state, and stitch together and output standardized feature vectors; The short-term fluctuation component of signal quality caused by waves, the mid-time feature vector of diurnal variation of tides and ionosphere, and the long-term trend component of slow change of sea state and atmospheric conditions are extracted and spliced ​​to obtain a multi-scale environmental feature vector. The query vector is obtained by summarizing the elevation angle distribution histogram of visible satellites. The sea surface roughness factor is calculated using the significant wave height and wind speed, and a key matrix is ​​generated. The attention weight is obtained by calculating the inner product similarity between the query vector and the key matrix, and the multipath sensing spatial embedding vector is output. The standardized feature vector, the multi-scale environmental feature vector, and the multipath sensing spatial embedding vector are concatenated and input into a multilayer perceptron for encoding to generate a high-dimensional embedding vector. A process noise matrix is ​​then constructed based on the parameterization of the high-dimensional embedding vector using a decomposition algorithm. The evolution of the GNSS measurement noise covariance matrix is ​​modeled as a differential equation driven by the process noise matrix. Spectral constraints are imposed on the system matrix, and the lower bound of the system matrix eigenvalues ​​is dynamically adjusted using the sea state index. Discretized recursive equations are then expanded to perform parallel computation of the covariance matrix elements. Calculate the monitoring indicators of the covariance matrix, trigger graded safety rollback based on the degree of abnormality of the monitoring indicators, and output the safety covariance matrix after safety rollback; The security covariance matrix is ​​injected into the measurement update stage of the multi-sensor fusion filter to extract ocean observation parameters, and the abnormal water level alarm and the lower bound of the eigenvalue are automatically switched based on the rate of change of the residual sea level height.

2. The method for modeling the covariance dynamics of marine GNSS buoy environmental perception measurement according to claim 1, characterized in that, The process of collecting multi-source sensor data from an ocean GNSS buoy platform, extracting quality index vectors, estimating the buoy's motion state, and splicing together standardized feature vectors includes: Extract a quality index vector from multi-constellation GNSS receiver data, which includes various accuracy attenuation factors, the number of visible satellites, and the average carrier-to-noise ratio of each satellite. Logarithmic transformation is performed on various precision attenuation factors to compress the dynamic range, and the number of visible satellites is normalized according to the maximum value of the sequence. The buoy motion state is estimated using inertial measurement unit data and buoy motion model, and the radial distance and radial velocity in the local reference coordinate system are calculated and normalized. The normalized radial distance, normalized radial velocity, logarithmically transformed precision attenuation factor, normalized visible satellite count, and mean carrier-to-noise ratio are concatenated to output a standardized feature vector.

3. The method for modeling the covariance dynamics of marine GNSS buoy environmental perception measurement according to claim 1, characterized in that, The extraction of short-term fluctuation components of signal quality caused by waves, mid-time feature vectors of diurnal tidal and ionospheric variations, and long-term trend components of slowly changing sea state and atmospheric conditions are concatenated to obtain a multi-scale environmental feature vector, including: Short-time fluctuation components of signal quality caused by waves are extracted using a high-pass filter. Sine-cosine coding is used to generate mid-time feature vectors of diurnal tidal and ionospheric variations; Long-term trend components of slowly changing sea state and atmospheric conditions are extracted using a low-pass filter. By concatenating the short-term fluctuation component, the medium-term feature vector, and the long-term trend component, a multi-scale environmental feature vector is obtained.

4. The method for modeling the covariance dynamics of marine GNSS buoy environmental perception measurement according to claim 1, characterized in that, The query vector is obtained from the aggregated histogram of visible satellite elevation angles. The sea surface roughness factor is calculated using significant wave height and wind speed, and a key matrix is ​​generated. The attention weight is obtained by calculating the inner product similarity between the query vector and the key matrix. Finally, a multipath sensing spatial embedding vector is output, including: The query vector is obtained by summarizing the histogram of the elevation angle distribution of visible satellites at the current moment and performing L2 normalization. The sea surface roughness factor is calculated using the significant wave height and wind speed, and a key matrix encoding typical sea surface multipath patterns modulated by the sea surface roughness factor is generated. The similarity between the inner product of the query vector and the key matrix is ​​calculated, and the attention weight is obtained through low-temperature Softmax. The multipath perception space embedding vector is then output through weighted summation and linear projection.

5. The method for modeling the covariance dynamics of marine GNSS buoy environmental perception measurement according to claim 1, characterized in that, The parameterized construction process noise matrix based on the high-dimensional embedding vector using the decomposition algorithm includes: A symmetric positive definite process noise matrix is ​​constructed based on high-dimensional embedding vector parameterization using the LDL decomposition algorithm. The Softplus function ensures that the diagonal elements of the diagonal positive matrix have strict numerical positivity, and the off-diagonal elements of the unit lower triangular matrix are generated by linear mapping.

6. The method for modeling the covariance dynamics of marine GNSS buoy environmental perception measurement according to claim 1, characterized in that, The process of modeling the evolution of the GNSS measurement noise covariance matrix as a differential equation driven by the process noise matrix, applying spectral constraints to the system matrix, dynamically adjusting the lower bound of the system matrix eigenvalues ​​using the sea state index, and performing parallel computation of the covariance matrix elements by expanding discretized recursive equations includes: The evolution of the GNSS measurement noise covariance matrix is ​​modeled as a Lyapunov differential equation, and spectral constraints are imposed on the system matrix. Calculate the sea state index based on normalized significant wave height and wind speed, and use the sea state index to dynamically adjust the lower bound of the eigenvalues ​​of the system matrix; The discretized recursive equations are expanded using the diagonalized structure of the system matrix to obtain a scalar convolution structure, and the covariance matrix elements are computed in parallel using a zero-padding fast Fourier transform.

7. The method for modeling the covariance dynamics of marine GNSS buoy environmental perception measurement according to claim 1, characterized in that, The monitoring indicators for calculating the covariance matrix trigger a tiered safety rollback based on the degree of anomaly of the monitoring indicators, and output a safety covariance matrix after the safety rollback, including: Calculate the condition number index, trace rate of change index, and minimum eigenvalue index of the covariance matrix; A tiered safety rollback is triggered based on the degree of abnormality of three monitoring indicators; In the first-level backoff, the weighted mean of the sliding time window is calculated to replace the current estimate; in the second-level backoff, the conservative constant covariance is switched to replace it; in the third-level backoff, the solution at the current time is marked as unavailable and an alarm is issued. Output the safety covariance matrix and corresponding quality indicators after safe rollback.

8. The method for modeling the covariance dynamics of marine GNSS buoy environmental perception measurement according to claim 1, characterized in that, The step of injecting the security covariance matrix into the measurement update stage of the multi-sensor fusion filter to extract ocean observation parameters includes: The safety covariance matrix is ​​injected into the measurement update stage of the multi-sensor fusion filter using the extended Kalman filter framework, and the safety covariance matrix is ​​used to control the weight of GNSS observations in the state update. Residual sea level height, significant wave height, mean wave period, and main wave direction are extracted from the high-precision three-dimensional position time series output by the multi-sensor fusion filter.

9. The method for modeling the covariance dynamics of marine GNSS buoy environmental perception measurement according to claim 8, characterized in that, The automatic switching between abnormal water level alarms and characteristic value lower bounds based on the rate of change of residual sea level height includes: When the rate of change of the residual sea level exceeds a preset threshold and the duration meets the judgment window, an abnormal water level alarm is issued, and the lower bound of the system matrix eigenvalue is automatically switched to the lower bound of the tsunami eigenvalue, allowing the covariance matrix to shrink rapidly.

Images