A combined navigation optimization method integrating residual fitting and significant wave height constraint USV

Abstract

This invention belongs to the field of marine GNSS / INS integrated navigation technology and discloses an integrated navigation optimization method that integrates residual fitting and significant wave height constraint USV. First, a GRU neural network model is built and trained while the GNSS is operating normally. The trained model is then used to predict the GNSS position increment to obtain the predicted GNSS position. Next, the residual between the predicted and actual GNSS positions is calculated, and a residual fitting function is pre-constructed. When GNSS lock-up occurs, the pre-constructed residual fitting function dynamically corrects the measurement noise covariance matrix, solving the problem of inaccurate measurement noise covariance matrix settings. Simultaneously, the vertical position information of the positioning result is calculated using the sea surface significant wave height constraint integrated navigation system. This invention improves positioning accuracy from both horizontal and vertical directions, enhancing the positioning accuracy and stability of the unmanned vessel integrated navigation system in GNSS interruption scenarios.

Description

Technical Field

[0001] This invention belongs to the field of marine GNSS / INS integrated navigation technology, and particularly relates to an integrated navigation optimization method that combines residual fitting and significant wave height constraint USV. Background Technology

[0002] Unmanned surface vessels (USVs) can perform a variety of tasks such as maritime search and rescue, river patrol, and marine mapping, and navigation systems are key to their operations.

[0003] Global Navigation Satellite System (GNSS) offers high positioning accuracy and is available 24 / 7, but its signal is easily interrupted under complex operating conditions. Inertial Navigation System (INS) is resistant to external environmental interference and has strong autonomy, but errors accumulate over time, making it difficult for a single system to achieve high-precision positioning.

[0004] Currently, the mainstream approach uses GNSS / INS integrated navigation, which combines their advantages to achieve continuous and accurate positioning. However, after GNSS signal lock-on is lost, the system positioning error still diverges rapidly. Three solutions are currently available to address this issue.

[0005] One approach is to utilize external auxiliary sensors of the GNSS / INS integrated navigation system, such as magnetometers, odometers, and barometers. Information from these external sensors can improve the overall robustness and positioning accuracy of the GNSS / INS integrated navigation system. However, adding auxiliary sensors also increases the overall cost and complexity of the integrated navigation system.

[0006] Second, motion constraint algorithms are used. Motion constraint algorithms can improve the positioning accuracy and reliability of integrated navigation systems without using external sensors. However, existing motion constraint algorithms often use fixed constraint values ​​and cannot adaptively adjust the constraint values ​​according to the actual motion state of the USV.

[0007] Third, we used Artificial Neural Networks (ANN) algorithms to study the robustness of GNSS / INS integrated navigation systems after GNSS lock-up. While neural network algorithms offer greater flexibility, they present challenges such as the inability to define the measurement noise covariance matrix, and the real-time performance of the neural network algorithm's solutions is also a concern.

[0008] In summary, the three solutions to address the divergence of INS positioning errors after GNSS lock-up have their own advantages and disadvantages. Adding additional auxiliary sensors can effectively improve positioning accuracy, but it also increases the complexity and cost of the system. Improvements at the algorithm level can enhance the robustness of the integrated navigation system without adding additional auxiliary sensors.

[0009] Therefore, for low-cost maritime USV GNSS / INS integrated navigation systems, the focus is on solving the problem of positioning error divergence caused by insufficient number of visible satellites or multipath effects in maritime GNSS denial environments (such as satellite obstruction, strong multipath effects on the sea surface, and maritime electromagnetic interference).

[0010] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0011] The purpose of this invention is to propose a combined navigation optimization method for USVs that integrates residual fitting and significant wave height constraints. When GNSS lock is lost, pseudo-GNSS observations are obtained by using the GNSS position increment of the USV predicted by the GRU neural network. Based on the residual fitting function pre-constructed during the GNSS uninterrupted period, the noise covariance matrix of the GNSS position observation is adaptively corrected. At the same time, the calculated significant sea surface wave height is introduced as a vertical positioning constraint. Through the synergistic cooperation of noise correction and wave height constraints, the positioning accuracy and reliability of the navigation system in three dimensions of the USV at sea after GNSS lock is lost are improved.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: The combined navigation optimization method integrating residual fitting and significant wave height constraint USV includes the following steps: Step 1. Build a GRU neural network model to predict GNSS position increments; the input of the GRU neural network is the raw observations of the IMU within a window period and the heading angle of the inertial navigation system INS; Step 2. When the GNSS is working normally, the inertial navigation system and the satellite navigation system perform the combined extended Kalman filtering normally, continuously outputting reliable positioning results; and simultaneously train the GRU neural network model. The GNSS position increment is predicted based on the GRU neural network model, thereby obtaining the predicted GNSS position; The difference between the predicted GNSS position and the actual GNSS position is used to pre-construct a residual fitting function; Simultaneously, by utilizing the wave surface vertical displacement calculated through integrated navigation, the current significant wave height is calculated, and the reasonable fluctuation range for the dominant wave vertical elevation is determined. This provides prior information for vertical positioning constraints; in This represents the average elevation when the GNSS is operating normally. Significant wave height; Step 3. When the GNSS loses lock, define the time of GNSS loss as time k; First, the GNSS position increment at time k is predicted using the trained GRU neural network model, thus obtaining the predicted GNSS position at time k. The predicted GNSS position is then substituted into the Extended Kalman Filter (EKF) for filtering and updating. Subsequently, the measurement noise covariance matrix R of EKF is corrected using a pre-built residual fitting function, and the GNSS position is gradually predicted at subsequent time points by updating the sliding window, thus simultaneously completing the iterative optimization of R. Simultaneously calculate the elevation value. , and when When fluctuations exceed the reasonable range, limit them to the range. This is done by applying vertical prior constraints to suppress vertical positioning divergence until the GNSS returns to normal operation.

[0013] The present invention has the following advantages: As mentioned above, to address the shortcomings of existing neural network-based GNSS position increment prediction methods, which directly incorporate predicted information into the KF model solution without considering the reasonable setting of the measurement noise covariance matrix, this invention proposes a combined navigation optimization method that integrates residual fitting based on neural network prediction with a significant wave height constraint USV. This method first constructs a GRU neural network model, trains it while the GNSS is operating normally, and outputs the GNSS position increment. Based on the GNSS position increment, the predicted GNSS position is calculated. Then, based on the residual between the predicted and actual GNSS positions, a residual fitting function is constructed. When GNSS lock-up occurs, the measurement noise covariance matrix is ​​dynamically corrected based on this residual fitting function, thus solving the problem of inaccurate setting of the measurement noise covariance matrix. Furthermore, this invention utilizes the vertical position information from the positioning results calculated by the sea surface significant wave height constraint combined navigation system. This invention improves positioning accuracy from both horizontal and vertical perspectives, effectively enhancing the positioning accuracy and stability of the unmanned vessel combined navigation system in GNSS interruption scenarios. Attached Figure Description

[0014] Figure 1 This is a flowchart of the combined navigation optimization method that integrates residual fitting and significant wave height constraint USV in an embodiment of the present invention. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 This invention proposes a robust integrated navigation method that integrates residual fitting noise correction based on neural network position increment prediction with significant sea surface wave height constraints. It utilizes a neural network to predict the GNSS position increment of a USV during GNSS interruption, establishes a residual model between the predicted and actual GNSS positions using a fitted residual fitting function, and adaptively corrects the GNSS position observation noise covariance matrix based on this residual fitting function. Furthermore, due to the influence of the sea surface environment, relying solely on the residual fitting method is insufficient to effectively improve vertical positioning accuracy. Therefore, this invention introduces the calculated significant sea surface wave height as a vertical positioning constraint, integrating environmental observation information into the integrated navigation solution process. Through the synergistic combination of noise correction and wave height constraints, this invention effectively improves the positioning accuracy and reliability of the three-dimensional navigation system for USVs at sea after GNSS lock-up.

[0016] like Figure 1 As shown, the combined navigation optimization method integrating residual fitting and significant wave height constraint USV includes the following steps: Step 1. Build a GRU neural network model to predict GNSS position increments; the input of the GRU neural network is the raw observations of the IMU within a window period and the heading angle of the inertial navigation system INS.

[0017] The design process of the GRU neural network model is as follows: When GNSS is available, the input layer of the neural network is the raw measurement value of the IMU within a window period and the heading angle and heading angular velocity vector calculated by the INS.

[0018] Traditional integrated navigation calculations do not involve neural networks. In order to solve the problem of short-term GNSS interruptions, this invention trains neural networks during normal GNSS periods, thereby predicting GNSS position increments when GNSS is interrupted.

[0019] The raw measurements from the Inertial Measurement Unit (IMU) and the heading angle from the Inertial Navigation System (INS) are the most significant factors affecting GNSS position increments.

[0020] The raw measurements from the IMU include three-dimensional angular velocities from the gyroscope. and three-dimensional force from accelerometer .

[0021] Input vector of GRU neural network model The formula is expressed as follows: .

[0022] in Three-dimensional angular velocity exist Transpose of components in three directions. Three-dimensional specific force exist Transpose of components in three directions; Indicates the heading angle.

[0023] GNSS position increment output by the GRU neural network model It is expressed as follows: .

[0024] in , , These represent the position increments in the north, east, and vertical directions, respectively.

[0025] After obtaining the GNSS position increment at time k, the predicted GNSS position at time k is further obtained; the calculation process for the predicted GNSS position at time k is as follows: When calculating the predicted GNSS position, the actual GNSS position at time k-1 is used as the baseline, and the GNSS position increment at time k output by the GRU neural network model is superimposed to obtain the predicted GNSS position at time k, as shown in the following formula: .

[0026] In the formula, Represents the actual GNSS position at time k-1. This represents the GNSS position increment at time k relative to time k-1, predicted by the GRU neural network model. Let k be the predicted GNSS position at time k.

[0027] This invention improves the model's ability to predict GNSS increments under prolonged GNSS interruption conditions (under conditions of GNSS signal anomalies or interruptions) by pre-constructing a GRU (Gross Root Regulator) and using IMU observation data and heading angles as inputs, extracting dynamic evolution features from the time series data. Furthermore, to mitigate the accumulation of prediction errors during GNSS interruptions, a residual fitting function is used to model the errors, and the measurement noise covariance matrix R of the predicted GNSS position is corrected to adapt to changes in input information and system state, thereby effectively reducing the accumulation of prediction errors during GNSS interruptions.

[0028] Step 2. When the GNSS is working normally, the inertial navigation system and the satellite navigation system perform the extended Kalman filter combination normally and continuously output reliable positioning results; and the GRU neural network model is trained simultaneously.

[0029] The GNSS position increment is predicted based on the GRU neural network model, thereby obtaining the predicted GNSS position.

[0030] The predicted GNSS position, calculated by the GRU neural network model, is subtracted from the actual GNSS position. A residual fitting function is pre-constructed to achieve dynamic correction of the measurement noise covariance matrix in step 3.

[0031] Simultaneously, by utilizing the wave surface vertical displacement calculated through integrated navigation, the current significant wave height is calculated, and the reasonable fluctuation range for the dominant wave vertical elevation is determined. This provides prior information for vertical positioning constraints.

[0032] in This represents the average elevation when the GNSS is operating normally. Significant wave height.

[0033] In step 2, the process of pre-constructing the residual fitting function is as follows: When GNSS observations are missing at time k, a time series is constructed by selecting residual data from m consecutive times prior to time k. A function model of the residuals changing over time, i.e., the residual fitting function, is established through a fitting method to characterize the temporal pattern of the position prediction bias of the GRU neural network model. The specific process is as follows: First, calculate the GNSS position residuals for m consecutive times from k-m+1 to k. The formula for calculating the residual at time k is: ; In the formula, Let k be the GNSS position residual at time k; Let k be the actual GNSS position at time k.

[0034] Subsequently based on The residual sequence within a time period consisting of 1 time points , , Construct a Bayesian polynomial residual fitting function for the corresponding dimension.

[0035] Three of the residual sequences were obtained by subtracting the predicted GNSS position from the actual GNSS position during the normal GNSS phase, resulting in residual sequences for latitude, longitude, and elevation.

[0036] Represents the residual sequence along the latitudinal direction; Residual sequence representing longitude direction; Represents the residual sequence in the elevation direction.

[0037] The formula for the Bayesian polynomial residual fitting function is as follows: ; In the formula These are residual fitting functions for latitude, longitude, and altitude, respectively, which can characterize the variation of residuals over time and output a 95% confidence interval to quantify the uncertainty of the fitting results.

[0038] For time series indexing, , respectively, are the polynomial orders of the latitude, longitude, and altitude residual fitting functions. , , .

[0039] , , , respectively, are the posterior mean of the polynomial coefficients obtained by Bayesian fitting.

[0040] The measurement noise covariance matrix calibrated before GNSS interruption, i.e., when GNSS is operating normally. Using the residual fitting function as a benchmark, a correction coefficient matrix is ​​constructed, and then... Dynamic corrections are performed to obtain the measurement noise covariance matrix that adapts to the prediction bias characteristics during GNSS outages. The specific formula is as follows: .

[0041] in The diagonal correction coefficient matrix, constructed based on the residual fitting function, is expressed as: .

[0042] in The values ​​of the residual fitting function for latitude, longitude, and altitude at time k are taken, and the absolute values ​​are used to ensure that the correction coefficients are non-negative. The variance form of the reference measurement noise covariance matrix calibrated before GNSS interruption is as follows: .

[0043] in This represents the variance of the latitude positioning error. The variance representing the longitude positioning error. The variance of the elevation positioning error, along with the other two factors, constitutes the reference measurement noise before GNSS interruption.

[0044] During USV maritime operations supported by GNSS / INS integrated navigation technology, the velocity and position data calculated by the integrated navigation can not only be used for the vehicle's own positioning and attitude control, but the U-direction velocity information also implicitly contains characteristic parameters of sea surface wave motion. As one of the elements of the dynamic marine environment, the wave height, period, and spectral characteristics of ocean waves directly affect the USV's navigation stability and the accuracy of the navigation system. Therefore, performing wave spectrum analysis on the GNSS / INS integrated navigation results to extract significant wave height information further improves the positioning accuracy and reliability of USVs at sea.

[0045] In step 2, the process of solving for the vertical displacement of the wavefront is as follows: Before performing wave spectrum analysis on GNSS / INS integrated navigation results, it is necessary to extract the vertical displacement of the wave surface.

[0046] First, the U-direction velocity of the GNSS / INS loose combined navigation results is integrated to obtain a sequence of relative changes in sea surface height including tides, waves, and various environmental disturbances. Then, the low-frequency tidal trend in the sequence is separated by filtering using the moving average method.

[0047] This invention employs the moving average method for filtering. The difference between the sea surface height sequences before and after filtering is calculated, and the tidal component is removed to obtain the pure vertical wave surface displacement. The moving average method is simple in principle, provides stable filtering results, and is suitable for filtering long-term time-series data.

[0048] The principle behind this method is as follows: A time series of a certain odd length is selected and averaged. This average is used as the trend value at the midpoint of the selected sequence. By continuously shifting and selecting a certain length of sequence, the average value of each period is obtained, forming a new sampling sequence. Because filtering eliminates the influence of random factors, it better reflects the basic development trend of the random process. The process of applying a moving average to a time series is essentially a filtering process.

[0049] In step 2, the significant wave height extraction process is as follows: For the filtered wavefront vertical displacement information, an autoregressive (AR) model is used for spectral analysis. The AR parameter model method establishes a mathematical model based on the sampled data, extrapolates the data and autocorrelation function, and makes its length exceed the original data sequence, avoiding window truncation, improving the resolution of the estimated spectrum, and making the spectrum estimation results more reflective of the global nature of the random signal.

[0050] In the AR model, a stationary random process is represented. Let the variance be... White noise sequence The sequence output after passing through a certain linear excitation system is... Or its autocorrelation function estimate The parameters, then by The parameters of the original stationary random process are used to inversely deduce the original stationary random process. The power spectrum was calculated to determine the significant wave height.

[0051] The specific process is as follows: The functional relationship before and after the input of the noise sequence is represented by the difference equation of the P-order AR model.

[0052] ; in This represents the wavefront displacement sequence, where P represents the model order. These are the parameters for the AR model.

[0053] This represents the historical wavefront displacement value j steps backward from the nth sampling point, where n represents the nth sampling point. It is a white noise sequence. The summation index is used to traverse historical samples from order 1 to order P in the AR model.

[0054] Where the model transfer function for: .

[0055] Power spectrum The formula is expressed as follows: .

[0056] in Indicates the sampling time interval. It is a white noise sequence The variance.

[0057] In AR models, AR model parameters and autocorrelation function The following relationship exists: .

[0058] The steps for estimating the wave spectrum are as follows: I. Wavefront displacement sequence from observation data Estimate the autocorrelation function .

[0059] II. Autocorrelation function obtained from step I Using the Levinsion-Durbin algorithm as input, the parameters of the AR model are solved. and white noise variance .

[0060] III. AR model parameters and white noise variance Substitute the equation into the power spectrum equation, solve for the wave power spectrum, and calculate the significant wave height.

[0061] This invention utilizes USV motion response or external observation data to invert significant wave height parameters and constructs a mapping relationship between "wave height and elevation". By introducing significant wave height as an observation constraint into the integrated navigation system, it solves the problem of easy divergence in vertical positioning of GNSS / INS integrated navigation systems under GNSS interruption.

[0062] Step 3. When the GNSS loses lock, define the time of GNSS loss as time k.

[0063] First, the GNSS position increment at time k is predicted using the trained GRU neural network model, thus obtaining the predicted GNSS position at time k. The predicted GNSS position is then substituted into the EKF filter for updating.

[0064] Subsequently, the measurement noise covariance matrix R of the extended Kalman filter is corrected using a pre-constructed residual fitting function. The GNSS position is then progressively predicted at subsequent time points by updating the sliding window, and the iterative optimization of R is completed simultaneously.

[0065] Simultaneously calculate the elevation value. , and when When fluctuations exceed the reasonable range, limit them to the range. This is done by applying vertical prior constraints to suppress vertical positioning divergence until the GNSS returns to normal operation.

[0066] Considering that the USV undergoes wave-dominated vertical motion around mean sea level over a certain time scale, its elevation variation amplitude is statistically significantly limited by wave height. The process of applying vertical prior constraints is as follows: Let the average elevation during the normal working time before the GNSS lock-up be . Significant wave height is The elevation of the solution result of the integrated navigation system (GNSS / INS loose combination solution) The reasonable fluctuation range is expressed as: ; When the elevation exceeds the above-mentioned reasonable fluctuation range, it is corrected to meet the prior constraints, as shown in the following formula: .

[0067] This invention addresses the moment when GNSS signal lock-off occurs by constructing a combined navigation method based on pseudo-GNSS position incremental observations and significant wave height constraints, and by introducing multi-source information to jointly constrain the system state. On the one hand, pseudo GNSS observations predicted based on the GRU model are used as horizontal position observation information, while on the other hand, the pseudo observation noise covariance is adaptively corrected by combining the residual fitting function. On the other hand, the introduction of significant sea surface wave height information to construct vertical constraints enables continuous correction of the navigation state under GNSS rejection conditions, thereby effectively suppressing the divergence of inertial navigation errors and improving the positioning accuracy and robustness of the system.

[0068] Compared with the prior art, the present invention has at least the following advantages: (1) By constructing a neural network model based on GRU, adaptive correction of inertial navigation error is achieved in the event of GNSS signal interruption or abnormality, reducing error accumulation and improving positioning continuity and stability.

[0069] (2) By establishing the relationship between wave-induced vertical motion and elevation error, and by introducing significant wave height into the navigation calculation process, this invention enhances the system's anti-interference capability and robustness under complex sea conditions.

[0070] (3) The present invention does not require additional hardware equipment. The present invention does not require any additional sensors. Wave height calculation can be achieved by using the output data of the GNSS / INS system on the unmanned surface vessel (USV).

[0071] Example 2 This embodiment 2 describes a computer device that includes a memory and one or more processors.

[0072] Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the combined navigation optimization method that integrates residual fitting and significant wave height constraint USV in Embodiment 1 above.

[0073] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0074] Example 3 This embodiment 3 describes a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the combined navigation optimization method of fusing residual fitting and significant wave height constraint USV in embodiment 1.

[0075] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0076] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A combined navigation optimization method integrating residual fitting and significant wave height constraint USV, characterized in that, Includes the following steps: Step 1. Build a GRU neural network model to predict GNSS position increments; the input of the GRU neural network is the raw observations of the IMU within a window period and the heading angle of the inertial navigation system INS; Step 2. When the GNSS is working normally, the inertial navigation system and the satellite navigation system perform the extended Kalman filter combination normally, continuously output reliable positioning results, and simultaneously train the GRU neural network model; The GNSS position increment is predicted based on the GRU neural network model, thereby obtaining the predicted GNSS position; the difference between the predicted GNSS position and the actual GNSS position is calculated, and a residual fitting function is pre-constructed; The process of pre-constructing the residual fitting function is as follows: when When GNSS observations are missing at any given time, select from... The continuous sequence starting from the beginning of time A time series is constructed from the residual data at each time point. A functional model of the residuals changing over time, i.e., the residual fitting function, is established through a fitting method. The process is as follows: First, calculate from k-m+1 to... Continuous time The GNSS position residuals at each time step; where The formula for calculating the time residual is: In the formula Let k be the GNSS position residual at time k; The actual GNSS position at time k; for Predicted GNSS position at any given time; Subsequently based on The residual sequence within a time period consisting of 1 time points , , Construct a Bayesian polynomial residual fitting function for the corresponding dimension; in Represents the residual sequence along the latitudinal direction; Residual sequence representing longitude direction; Residual sequence representing the elevation direction; The above three residual sequences are obtained by subtracting the predicted GNSS position from the actual GNSS position during the normal GNSS phase, resulting in residual sequences in the latitude, longitude, and elevation directions. The formula for the Bayesian polynomial residual fitting function is as follows: ； In the formula These are the residual fitting functions for latitude, longitude, and altitude, respectively. For time series indexing; These are the polynomial orders of the latitude, longitude, and altitude residual fitting functions, respectively. , , These are the coefficients of the polynomials in the corresponding directions; The coefficients are obtained by Bayesian estimation based on fitting the residual sequence; Measurement noise covariance matrix calibrated under normal GNSS operation Using the residual fitting function as a benchmark, a correction coefficient matrix is ​​constructed for each pair. Dynamic correction yields the measurement noise covariance matrix during GNSS outages. The formula is as follows: ； in The diagonal correction coefficient matrix, constructed based on the residual fitting function, is expressed as: ； in , where is the absolute value of the residual fitting function values ​​for latitude, longitude, and altitude at time k; The reference measurement noise covariance matrix calibrated during normal GNSS operation has the following variance form: ； in This represents the variance of the latitude positioning error. The variance representing the longitude positioning error. The variance of the elevation positioning error, together with the other two factors, constitutes the reference measurement noise when the GNSS is working normally. Simultaneously, by utilizing the wave surface vertical displacement calculated through integrated navigation, the current significant wave height is calculated, and the reasonable fluctuation range for the dominant wave vertical elevation is determined. This provides prior information for vertical positioning constraints; in This represents the average elevation when the GNSS is operating normally. Significant wave height; Step 3. When the GNSS loses lock, define the time of GNSS loss as time k; First, the GNSS position increment at time k is predicted using the trained GRU neural network model, thus obtaining the predicted GNSS position at time k. The predicted GNSS position is then substituted into the Extended Kalman Filter (EKF) for filtering and updating. Subsequently, the measurement noise covariance matrix R of EKF is corrected using a pre-built residual fitting function, and the GNSS position is gradually predicted at subsequent time points by updating the sliding window, thus simultaneously completing the iterative optimization of R. Simultaneously calculate the elevation value. , and when When fluctuations exceed the reasonable range, limit them to the range. This applies a vertical prior constraint until the GNSS returns to normal operation.

2. The combined navigation optimization method based on residual fitting and significant wave height constraint USV as described in claim 1, characterized in that, In step 1, the design process of the GRU neural network model is as follows: The input layer of the GRU neural network model consists of the raw measurements from the IMU within a window period and the heading angle calculated by the INS; the raw measurements from the IMU are the three-dimensional angular velocities from the gyroscope. and three-dimensional force from accelerometer ; Input vector of GRU neural network model The formula is expressed as follows: ； in Three-dimensional angular velocity exist Transpose of components in three directions. Three-dimensional specific force exist Transpose of components in three directions; Indicates the heading angle; GNSS position increment output by the GRU neural network model It is expressed as follows: ； in , , These represent the position increments in the north, east, and vertical directions, respectively.

3. The combined navigation optimization method based on residual fitting and significant wave height constraint USV as described in claim 2, characterized in that, In step 1, after obtaining After the GNSS position increment at time, obtain Predicted GNSS position at any given time; in The calculation process for the predicted GNSS position at a given time is as follows: Based on the actual GNSS position at time k-1, the output of the GRU neural network model is superimposed. The GNSS position increment at time k is used to obtain the predicted GNSS position at time k, as shown in the following formula: ; In the formula, represent The actual GNSS position at any given time. Predicted by the GRU neural network model Time relative to GNSS position increment at time t, for Predicted GNSS position at any given time.

4. The combined navigation optimization method based on residual fitting and significant wave height constraint USV as described in claim 1, characterized in that, In step 2, the process of solving for the vertical displacement of the wavefront is as follows: First, the U-direction velocity of the GNSS / INS loosely integrated navigation results is integrated to obtain a sequence of relative changes in sea surface height including tides, waves and various environmental disturbances. Then, the low-frequency tidal trend in the sequence is separated by filtering using the moving average method. The difference between the sea surface height sequences before and after filtering is calculated, and the tidal component is removed to obtain the pure vertical displacement of the wave surface.

5. The combined navigation optimization method based on residual fitting and significant wave height constraint USV as described in claim 4, characterized in that, In step 2, the significant wave height extraction process is as follows: For the filtered wavefront vertical displacement, an autoregressive AR model is used to perform spectral analysis, as follows: The functional relationship before and after the input of the noise sequence is represented by the difference equation of the P-order AR model; ； in This represents the wavefront displacement sequence, where P represents the model order. For AR model parameters; This represents the historical wavefront displacement value j steps backward from the nth sampling point, where n represents the nth sampling point. It is a white noise sequence; The summation index is used to traverse historical samples from order 1 to order P in the AR model; Where the model transfer function for: ； in, Indicates the sampling time interval. Indicates the frequency of ocean waves; Power spectrum The formula is expressed as follows: ； in Represents the sampling time interval, u(n) is a value with zero mean and variance of . A white noise sequence; In AR models, AR model parameters and autocorrelation function The following relationship exists: ； The steps for estimating the wave spectrum are as follows: I. Wavefront displacement sequence from observation data Estimate the autocorrelation function ; To determine the order of an AR model of order P, it is necessary to calculate the autocorrelation values ​​from order 0 to P. Equations were constructed to solve for the model parameters and noise variance, and then the wave power spectrum was calculated. II. Obtained from step I Using the Levinsion-Durbin algorithm as input, solve the problem. and ; III. and Substitute the equation into the power spectrum equation, solve for the wave power spectrum, and calculate the significant wave height.

6. The combined navigation optimization method of fusing residual fitting and significant wave height constraint USV as described in claim 1, characterized in that, In step 3, the process of applying the vertical prior constraint is as follows: Let the average elevation during the normal working time before the GNSS lock-up be . Significant wave height is The elevation of the solution result of the integrated navigation system The reasonable fluctuation range is expressed as: ； When the elevation exceeds the above-mentioned reasonable fluctuation range, it is corrected to meet the prior constraints. The correction formula is as follows: 。 7. A computer device, comprising a memory and one or more processors; characterized in that, The memory stores executable code, which, when executed by the processor, implements the steps of the combined navigation optimization method of fusing residual fitting and significant wave height constraint USV as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a program stored thereon; characterized in that, When executed by the processor, the program is used to implement the steps of the combined navigation optimization method of fusing residual fitting and significant wave height constraint USV as described in any one of claims 1 to 6.

Images