GNSS multipath error suppression method based on spatial-temporal characteristic decoupling

Abstract

The invention discloses a GNSS multipath error suppression method based on spatial-temporal feature decoupling, and relates to the technical field of global navigation satellite system high-precision positioning, and the method comprises the steps: obtaining GNSS observation data, carrying out the preprocessing, and extracting a time series data set; constructing a satellite geometric feature, training the satellite geometric feature by using the first-stage extreme gradient lifting model, and predicting to obtain a geometric component predicted value; constructing a time sequence feature set based on the prediction residual of the geometric component prediction value, and training the time sequence feature set by using a second-stage extreme gradient lifting model to obtain a time-varying component prediction value; and obtaining a multipath correction value according to the geometric component predicted value and the time-varying component predicted value, correcting the original GNSS observation value, and performing performance verification. The method is stable for different satellites in different observation stations, does not need strict orbit repeatability assumption, is wider in application range, and can provide a high-performance and easy-to-deploy technical scheme for practical engineering application.

Description

Technical Field

[0001] This invention relates to the field of high-precision positioning technology for global navigation satellite systems, and in particular to a GNSS multipath error suppression method based on spatiotemporal feature decoupling. Background Technology

[0002] In GNSS high-precision positioning applications, multipath effect is the main source of error restricting positioning accuracy. This error arises from the interference and superposition of satellite signals reflected by surrounding reflectors such as buildings and the ground at the receiving antenna with the direct signals, leading to systematic deviations in the observations. Existing multipath suppression methods mainly include two categories: sidereal filtering and multipath hemispherical charts. Sidereal filtering utilizes the periodic repeatability of satellite orbits to correct current observations using the residuals from the same time period of the previous day. However, it is highly dependent on the strict repeatability of the orbit. When there is a slight drift in the satellite orbit or a change in the reflection environment, the correction effect will rapidly decay or even introduce negative correction. In practical applications, the performance of this method deteriorates significantly after 10 days.

[0003] Multipath hemispherical images construct spatial gridded correction models using satellite elevation and azimuth angles as indices. However, the use of grid mean as a correction factor leads to the irreversible loss of high-frequency multipath information during spatial discretization, making it difficult to capture short-term dynamic fluctuations. In particular, it is insufficient in suppressing transient interference caused by moving reflectors.

[0004] While existing research has applied machine learning algorithms to multipath modeling, many employ simple feature fusion strategies, directly mixing satellite geometric features such as elevation and azimuth angles with temporal features such as historical observations as input into the model. This ignores the physical mechanisms of multipath errors, which include long-period geometric components determined by the static reflection environment and short-term time-varying components caused by dynamic environmental changes. These two types of components differ fundamentally in their physical origins and evolutionary patterns. Direct fusion leads to mutual interference between features. Specifically, highly correlated temporally lagging features dominate model learning, suppressing the modeling ability of satellite geometric features, resulting in performance inferior to using spatial features alone. This invention aims to address the technical problems of poor environmental adaptability, high-frequency information loss, and feature fusion interference in existing GNSS multipath suppression methods in practical engineering applications. Summary of the Invention

[0005] In view of the problems existing in the GNSS multipath error suppression method based on spatiotemporal feature decoupling, this invention is proposed. Therefore, the problem to be solved by this invention is how to provide a GNSS multipath error suppression method based on spatiotemporal feature decoupling.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a GNSS multipath error suppression method based on spatiotemporal feature decoupling, which includes: acquiring GNSS observation data and preprocessing it, and extracting a time series dataset containing timestamps, satellite elevation angles, satellite azimuth angles and multipath residuals;

[0008] Satellite geometric features are constructed based on satellite elevation and azimuth angles. The satellite geometric features are trained using the first-stage extreme gradient boosting model. The trained first-stage extreme gradient boosting model is then used for prediction to obtain the predicted values ​​of the geometric components.

[0009] Based on the prediction residuals of the geometric component predictions, a time-series feature set containing lag features and time-varying residual features is constructed. The time-series feature set is trained using the second-stage extreme gradient boosting model, and the trained second-stage extreme gradient boosting model is used for prediction to obtain the time-varying component predictions.

[0010] Multipath correction values ​​are obtained based on the predicted values ​​of geometric components and time-varying components. The original GNSS observations are then corrected using the multipath correction values, and the performance is verified.

[0011] As a preferred embodiment of the GNSS multipath error suppression method based on spatiotemporal feature decoupling described in this invention, the step of acquiring GNSS observation data and performing preprocessing includes:

[0012] A base station and a rover are set up in the monitoring area to obtain dual-frequency uncombined pseudorange and carrier phase observations.

[0013] The satellite clock error, receiver clock error, tropospheric delay and ionospheric delay are eliminated by inter-station double-difference processing to obtain the double-difference residual;

[0014] The double-difference residuals are restored to the single-difference residuals of each satellite, and low-pass filtering is performed to suppress observation noise, resulting in the filtered single-difference multipath residual sequence.

[0015] Data is organized separately according to satellite and observation type to form a dataset containing timestamps, satellite elevation angles, satellite azimuth angles, and multipath residuals.

[0016] As a preferred embodiment of the GNSS multipath error suppression method based on spatiotemporal feature decoupling described in this invention, the step of training the satellite geometric features using the first-stage extreme gradient boosting model includes:

[0017] The first stage of input features is constructed by extracting satellite elevation angle and satellite azimuth angle from the acquired dataset. The elevation angle is trigonometrically transformed to calculate the sine and cosine values, and the azimuth angle is trigonometrically transformed to calculate the sine and cosine values, thus forming an input vector containing six features.

[0018] A sliding window strategy is adopted, using data from multiple days prior to the target date as the training set, and outlier removal and Z-score standardization are performed.

[0019] Train the first-stage extreme gradient boosting model and set the model hyperparameters, including the number of decision trees, the maximum depth of the trees, the learning rate, the sample sampling ratio, the feature sampling ratio, and the L1 and L2 regularization coefficients.

[0020] The trained first-stage extreme gradient boosting model is used to predict each epoch of the target day, and the geometric component prediction is obtained by inputting the satellite geometric features.

[0021] As a preferred embodiment of the GNSS multipath error suppression method based on spatiotemporal feature decoupling described in this invention, the construction of a time-series feature set including hysteresis features and residual time-varying features includes:

[0022] For each sample in the training set, the first-stage residual is obtained by subtracting the geometric component prediction value from the observed multipath residual value; the lag feature includes the residual values ​​of multiple epochs prior to the current epoch.

[0023] The time-varying features of the residuals include: velocity features and acceleration features calculated based on lag residuals, sliding window standard deviation of residuals from multiple epochs, and short-term trend slope obtained by linear fitting of recent residual sequences;

[0024] A time series feature set containing lag characteristics and residual time-varying characteristics is obtained.

[0025] As a preferred embodiment of the GNSS multipath error suppression method based on spatiotemporal feature decoupling described in this invention, the step of training the temporal feature set using a second-stage extreme gradient boosting model includes:

[0026] The training data in the second stage are subjected to feature construction and standardization, and the prediction residuals of the extreme gradient boosting model in the first stage are used as the training target.

[0027] Set the model hyperparameters, including the number of decision trees, the maximum depth of the trees, the learning rate, the sample sampling ratio, the feature sampling ratio, and the L1 and L2 regularization coefficients;

[0028] The trained second-stage extreme gradient boosting model is used to predict each epoch of the target day. The lag characteristics and time-varying characteristics of the residuals are calculated based on the historical residuals in the historical residual buffer and input into the second-stage extreme gradient boosting model to obtain the predicted values ​​of the time-varying components.

[0029] As a preferred embodiment of the GNSS multipath error suppression method based on spatiotemporal feature decoupling described in this invention, the multipath correction value obtained from the geometric component prediction value and the time-varying component prediction value includes:

[0030] The sum of the predicted values ​​of the geometric components and the predicted values ​​of the time-varying components is used as the multipath correction value. The multipath correction value is subtracted from the original pseudorange observation value to obtain the corrected pseudorange.

[0031] The corrected carrier phase is obtained by subtracting the multipath correction value from the original carrier phase observation value;

[0032] The double-difference positioning solution is re-performed using the corrected observations to obtain the corrected coordinate sequence.

[0033] As a preferred embodiment of the GNSS multipath error suppression method based on spatiotemporal feature decoupling described in this invention, the performance verification includes:

[0034] A sidereal solar filtering method model and a multipath hemispherical diagram method model were constructed as a comparison baseline. The residual reduction rate of each method model on pseudorange and carrier phase observations was calculated to evaluate the multipath suppression performance.

[0035] The residual reduction rate is defined as the root mean square of the individual error residuals before correction minus the root mean square of the individual error residuals after correction, divided by the root mean square of the individual error residuals before correction, multiplied by 100%.

[0036] By using simulated dynamic relative positioning calculations, the root mean square errors of positioning in multiple directions of each method model are compared.

[0037] The performance stability of each method model in continuous observation and its adaptability in dynamic multipath simulation scenarios are analyzed.

[0038] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a GNSS multipath error suppression method based on spatiotemporal feature decoupling.

[0039] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements the steps of a GNSS multipath error suppression method based on spatiotemporal feature decoupling.

[0040] The beneficial effects of this invention are as follows: This invention achieves significant improvements in multipath suppression performance, positioning accuracy, long-term stability, and dynamic environment adaptability. Regarding multipath suppression performance, this method can reduce the root mean square of the residuals of pseudorange observation frequency and carrier phase observation frequency. In terms of positioning accuracy, the application of this method improves the accuracy of simulated dynamic positioning, meeting the requirements of high-precision applications. Regarding long-term stability, this method maintains stable performance during the testing period, reducing the model update frequency to 7 to 10 days and lowering computational costs. Regarding dynamic environment adaptability, this method significantly outperforms traditional methods in dynamic multipath simulation experiments simulating vehicle passage scenarios, and has a good suppression effect on transient interference caused by moving reflectors. Furthermore, this method achieves optimal performance across all observation types and is stable across different stations and satellites. It does not require strict orbit repeatability assumptions, making it more widely applicable and providing a high-performance and easily deployable technical solution for practical engineering applications. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of a GNSS multipath error suppression method based on spatiotemporal feature decoupling; Figure 2 This is a comparison chart of the residual reduction rate between this method and existing technologies; Figure 3 This is a comparison chart of the long-term stability of our method and existing technologies; Figure 4 This is a comparison chart of the positioning errors of our method and existing technologies in static multipath scenarios; Figure 5 This is a distribution diagram of the feature importance of our method under different modeling strategies; Figure 6 This is a comparison chart of the positioning errors of our method and existing technologies in dynamic multipath scenarios. Detailed Implementation

[0043] To make the above-mentioned objects, features, and advantages of the present invention more readily understood, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0044] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0045] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0046] Example 1

[0047] Reference Figure 1 This is the first embodiment of the present invention, which provides a GNSS multipath error suppression method based on spatiotemporal feature decoupling, comprising:

[0048] Step 1: Acquire and preprocess observation data. This involves acquiring observation data, including Global Navigation Satellite System (GNSS) observation data, satellite ephemeris, and coordinates of the base station and rover, to prepare data for multipath error modeling. The specific steps are as follows:

[0049] Step 1-1: Set up a base station and a rover station in the monitoring area. Both the base station and the rover station use the same type of geodesic antenna and receiver. The sampling rate is set to 1s. The base station is installed in a geologically stable area far away from reflectors. The rover station is installed on the structure to be monitored. The baseline length is controlled within 10 meters to form a short baseline configuration.

[0050] Steps 1-2: Obtain experimental data through the Curtin University Global Navigation Satellite System Data Center.

[0051] Steps 1-3: Use the Global Navigation Satellite System data processing software RTKLIB to perform parameter estimation of dual-frequency uncombined pseudorange and carrier phase observations. The processing parameters are set as follows: cutoff elevation angle 10 degrees, static mode to estimate rover position, and eliminate common errors such as satellite clock bias receiver clock bias tropospheric delay and ionospheric delay through inter-station differential to obtain a posterior residual sequence containing multipath error and observation noise.

[0052] Steps 1-4: Extract single-difference residuals from double-difference residuals as observations for multipath error. Select a reference satellite and restore the double-difference residuals of other satellites and the reference satellite to single-difference residuals of each satellite. Perform low-pass filtering on the single-difference residuals, using a Butterworth filter with a cutoff frequency of 0.1 Hz to remove high-frequency observation noise. Extract the filtered single-difference multipath residual sequences from the pseudorange observations of P1 and P2 and the carrier phase observations of L1 and L2, respectively. At the same time, record the satellite elevation angle and azimuth angle information corresponding to each epoch.

[0053] Steps 1-5: Store the above data according to satellite number and observation type. Each observation type of each satellite constitutes an independent time series dataset. The dataset contains four fields: timestamp, satellite elevation angle, satellite azimuth angle, and multipath residual.

[0054] Step 2: First-stage geometric trend modeling, based on the satellite's spatial geometric position characteristics, modeling long-period systematic multipath errors caused by fixed reflectors.

[0055] The specific steps are as follows:

[0056] Step 2-1: Construct the first-stage input features. Extract the satellite elevation angle and satellite azimuth angle from the acquired dataset. Perform trigonometric transformation on the elevation angle to calculate the sine and cosine values. Perform trigonometric transformation on the azimuth angle to calculate the sine and cosine values. Finally, construct an input vector containing six features.

[0057] Step 2-2: Prepare training data using a sliding window strategy. For a target date, such as the year-end date 249, collect the cumulative data from the five days prior to the target date, i.e., the year-end dates 244 to 248, as the training set. After merging the five days of data, perform outlier cleanup and remove sample points whose absolute residual value exceeds three times the standard deviation. Perform Z-score standardization on all input features so that the mean of each feature is 0 and the standard deviation is 1.

[0058] Steps 2-3: Train the first-stage extreme gradient boosting model (XGBoost) with the following hyperparameters set as follows: number of decision trees equals 40, maximum tree depth equals 20, learning rate equals 0.3, sample sampling ratio equals 0.8, feature sampling ratio equals 0.9, L1 regularization coefficient equals 0.1, and L2 regularization coefficient equals 1.0.

[0059] Steps 2-4: Use the trained first-stage extreme gradient boosting model to predict each epoch of the target day, and input satellite geometric features to obtain the predicted values ​​of geometric components.

[0060] Step 3: Second stage residual time-varying modeling, modeling the short-time fluctuation components caused by dynamic environmental changes in the residual space after eliminating geometric trends.

[0061] The specific steps are as follows: Step 3-1: Calculate the first-stage residual. For each sample in the training set, subtract the obtained geometric component prediction value from the observed multipath residual value. That is, the residual is equal to the observed value minus the geometric component prediction value.

[0062] Step 3-2: Construct the second-stage input features, which include two categories: lag features and residual time-varying features.

[0063] The lag features are the residual values ​​of the previous 1 to 10 epochs, denoted as Lag-1 to Lag-10. That is, for the current epoch, the Lag-1 feature is R(t-1), the Lag-2 feature is R(t-2), and so on up to Lag-10, which is R(t-10).

[0064] The time-varying features of the residuals include: velocity feature V(t) equals R(t-1) minus R(t-2); acceleration feature Acc(t) equals V(t) minus V(t-1); the sliding window standard deviation σ(t) is the standard deviation of the residuals from R(t-10) to R(t-1); and the short-term trend is obtained by linearly fitting the residual values ​​of the most recent 5 epochs to obtain the slope of the fitted line. All time-series features are strictly constructed based on historical data prior to the current time to ensure causality.

[0065] Step 3-3: Perform feature construction and standardization on the training data of the second stage. Since the lagging features require the data of the first 10 epochs, the first 10 epochs of each satellite transit arc cannot be used to construct complete features and need to be removed. Perform Z-score standardization on the constructed feature matrix.

[0066] Steps 3-4: Train the second-stage extreme gradient boosting model (XGBoost) with the following hyperparameters set as follows: number of decision trees equals 100, maximum tree depth equals 14, learning rate equals 0.05, sample sampling ratio equals 0.7, feature sampling ratio equals 0.7, L1 regularization coefficient equals 1.0, and L2 regularization coefficient equals 5.0.

[0067] Steps 3-5: Use the trained second-stage extreme gradient boosting model to predict each epoch of the target day, maintain a historical residual buffer with a length of 10 epochs, calculate the lag features and time-varying features of the residuals based on the historical residuals in the buffer, and input them into the second-stage extreme gradient boosting model to obtain the predicted values ​​of the time-varying components.

[0068] Step 4: Calculation and application of multipath correction values. The two-stage predicted values ​​are combined to form the final multipath correction value and applied to the observation value correction.

[0069] The specific steps are as follows: Step 4-1, calculate the multipath correction value. The final multipath correction value is the sum of the predicted value of the geometric component and the predicted value of the time-varying component.

[0070] Step 4-2: Apply multipath correction to the original observations. The corrected pseudorange is equal to the original pseudorange minus the multipath correction value, and the corrected carrier phase is equal to the original carrier phase minus the multipath correction value.

[0071] Step 4-3: Use the corrected observations to recalculate the double-difference positioning to obtain the improved coordinate sequence.

[0072] Step 4-4: Store the first-stage residual of the current epoch, i.e., the observation multipath residual minus the predicted value of the geometric component, into the historical residual buffer for use in the next epoch.

[0073] Step 5: Performance verification and comparative analysis. This method is verified by comparing it with traditional methods.

[0074] The specific steps are as follows: Step 5-1, construct the comparison method model, including the sidereal solar filtering method model and the multipath hemispherical diagram method model.

[0075] Step 5-2: Statistically analyze the multipath suppression performance and calculate the residual reduction rate, which is defined as the root mean square of the single-difference residual before correction minus the root mean square of the single-difference residual after correction, divided by the root mean square of the single-difference residual before correction multiplied by 100%.

[0076] Step 5-3: Statistically analyze the improvement in positioning accuracy, analyze long-term stability, verify dynamic multipath adaptability, and verify the effectiveness of the two-stage decoupling strategy through SHAP analysis.

[0077] This embodiment also provides a computer device applicable to a GNSS multipath error suppression method based on spatiotemporal feature decoupling, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement all or part of the steps of the method described in the above embodiments of the present invention.

[0078] This embodiment also provides a storage medium storing a computer program thereon. When the computer program is executed by a processor, it performs the method in any optional implementation of the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0079] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0080] Example 2

[0081] Reference Figure 2 - Figure 6 This is the second embodiment of the present invention. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0082] The observation data used in this embodiment of the invention are 30 days from day 244 to day 273 of 2021, including 5 short baselines. The base station is CUT0, and the rover stations are CUAA, CUTA, CUTB, CUCC, and CUTC. The length of each baseline is less than 10 meters.

[0083] A comparative model was constructed. The orbital repetition period of the sidereal day filtering method was calculated to be 23 hours and 56 minutes based on the broadcast ephemeris. The multipath hemisphere diagram method discretized the celestial sphere into 32,400 grid units using a 1°×1° grid division.

[0084] like Figure 2 The data are the average results of a 25-day testing period (DOY 249-273, 2021) for five baselines at Curtin University (base station CUT0, and rover stations CUAA, CUTA, CUTB, CUCC, and CUTC).

[0085] The horizontal axis represents observation types P1, P2, L1, and L2, and the vertical axis represents the percentage reduction in residuals. The three bar charts represent sidereal-day filtering (SF), multipath hemisphere plot (MHM), and our method (ST-XGB), respectively. The figure shows that our method significantly outperforms traditional methods across all observation types.

[0086] Using the average results from 25 test days (249 to 273) at 5 stations, the average residual reduction rates of this method for P1, P2, L1, and L2 were 26.09%, 38.90%, 20.63%, and 20.49%, respectively. The corresponding results for the sidereal filtering method were 17.05%, 26.61%, 18.60%, and 18.00%, respectively, and the corresponding results for the multipath hemispherical diagram method were 12.34%, 22.58%, 15.05%, and 16.35%, respectively. Figure 2 As shown, this method achieves optimal performance across all observation types.

[0087] like Figure 3 As shown, the model was trained using only the data from the first 5 days to model the subsequent 25 days. The data represents the daily residual reduction rate of the 5 stations during the 25-day test period.

[0088] The horizontal axis represents the annual DOY from 249 to 273, and the vertical axis represents the percentage reduction rate of the residuals. This method maintains stable performance within 25 days. The sidereal day filter shows a linear decay trend, and P1 becomes negative after the 10th day. The multipath hemispherical plot is stable, but its performance is consistently lower than that of this method.

[0089] The residual reduction rate for each day was calculated as the average of five stations. During the 25-day test period, the standard deviations of the stable performance P1 and P2 were only 2.07% and 3.17%, respectively. The sidereal filtering showed that P1 dropped to a negative value as low as -7.90% after the 10th day (258th day of the year), indicating the introduction of negative correction. This proves that the method can remain effective for 7 to 10 days without daily updates.

[0090] like Figure 4 As shown, the data represents the time series of positioning errors for the CUCC station's annual day 249 under a conventional static multipath environment (without dynamic multipath disturbances). The horizontal axis represents time, and the vertical axis represents positioning error. The four curves represent the positioning results of the original solution, sidereal filtering, multipath hemispherical plot, and the proposed method, respectively. This method achieves optimal positioning accuracy in all three directions. Specifically, the root mean square (RMS) of the positioning error in the east direction decreased from 2.0 mm in the original solution to 1.6 mm, in the north direction from 2.3 mm to 1.9 mm, and in the celestial direction from 5.6 mm to 4.5 mm; representing improvements of 20.0%, 17.4%, and 19.6% respectively compared to the original solution.

[0091] The comprehensive correction performance of each method was evaluated using a simulated dynamic relative positioning mode. The original solution had an accuracy of 2.0 mm in the east, 2.3 mm in the north, and 5.5 mm in the celestial direction. After correction by this method, the accuracy reached 1.5 mm in the east, 1.8 mm in the north, and 4.4 mm in the celestial direction, representing improvements of 25.0%, 21.7%, and 20.0% respectively compared to the original solution. After correction by the sidereal solar filter, the accuracy was 1.7 mm, 1.9 mm, and 4.6 mm. After correction by the multipath hemispherical diagram, the accuracy was 1.7 mm, 2.0 mm, and 4.7 mm. This method achieved the best accuracy in all three directions.

[0092] like Figure 5 As shown, the data is based on SHAP interpretation analysis of 25 test days at the CUCC station. The horizontal axis represents the modeling configuration M2 single-stage, M3 single-stage, and M5 two-stage first and second stages, while the vertical axis represents the percentage of feature importance.

[0093] The stacked bar chart shows the proportions of spatial features, hysteresis features, and residual time-varying features. The figure shows that in single-stage fusion, hysteresis features account for 85.3%, suppressing spatial features to 14.7%. After two-stage decoupling, the spatial features in the first stage recover to 100%, while the residual time-varying features in the second stage improve to 21.2%. This demonstrates the effectiveness of the two-stage strategy in solving the feature masking problem, indicating that this method is more practical than traditional methods in effectively improving multipath suppression performance and positioning accuracy.

[0094] like Figure 6 As shown, the data represents the time series of positioning errors for the CUCC station's annual day 249 in a dynamic multipath simulation experiment. The horizontal axis represents time, and the vertical axis represents positioning error. The four curves represent the original solution, the sidereal day filter, the multipath hemispherical map, and the simulated dynamic positioning results of the method of this invention, respectively.

[0095] In a scenario simulating vehicle passage through 300 dynamic multipath events injected over an annual cumulative day (249 days), the average residual reduction rate of this method remained at 26.82%, while that of the stellar-sun filtering was only 13.10%, and that of the multipath hemispherical map was 16.16%. In terms of positioning accuracy, this method achieved 1.7 mm in the east direction, 2.1 mm in the north direction, and 5.0 mm in the celestial direction, representing improvements of 26.1%, 19.2%, and 20.6% respectively compared to the original solution, demonstrating the robustness of this invention under strong dynamic interference environments.

[0096] In summary, this invention achieves significant improvements in multipath suppression performance, positioning accuracy, long-term stability, and dynamic environment adaptability. Furthermore, this method achieves optimal performance across all observation types and demonstrates stability across different stations and satellites. It does not require strict orbit repeatability assumptions, thus broadening its applicability and providing a high-performance and easily deployable technical solution for practical engineering applications.

[0097] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A GNSS multipath error suppression method based on spatiotemporal feature decoupling, characterized in that: include, Acquire GNSS observation data and preprocess it to extract a time series dataset containing timestamps, satellite elevation angles, satellite azimuth angles, and multipath residuals; Satellite geometric features are constructed based on satellite elevation and azimuth angles. The satellite geometric features are trained using the first-stage extreme gradient boosting model. The trained first-stage extreme gradient boosting model is then used for prediction to obtain the predicted values ​​of the geometric components. Based on the prediction residuals of the geometric component predictions, a time-series feature set containing lag features and time-varying residual features is constructed. The time-series feature set is trained using the second-stage extreme gradient boosting model, and the trained second-stage extreme gradient boosting model is used for prediction to obtain the time-varying component predictions. Multipath correction values ​​are obtained based on the predicted values ​​of geometric components and time-varying components. The original GNSS observations are then corrected using the multipath correction values, and the performance is verified.

2. The GNSS multipath error suppression method based on spatiotemporal feature decoupling as described in claim 1, characterized in that: The acquisition and preprocessing of GNSS observation data includes: A base station and a rover are set up in the monitoring area to obtain dual-frequency uncombined pseudorange and carrier phase observations. The satellite clock error, receiver clock error, tropospheric delay and ionospheric delay are eliminated by inter-station double-difference processing to obtain the double-difference residual; The double-difference residuals are restored to the single-difference residuals of each satellite, and low-pass filtering is performed to suppress observation noise, resulting in the filtered single-difference multipath residual sequence. Data is organized separately according to satellite and observation type to form a dataset containing timestamps, satellite elevation angles, satellite azimuth angles, and multipath residuals.

3. The GNSS multipath error suppression method based on spatiotemporal feature decoupling as described in claim 1, characterized in that: The training of satellite geometric features using the first-stage extreme gradient boosting model includes: The first stage of input features is constructed by extracting satellite elevation angle and satellite azimuth angle from the acquired dataset. The elevation angle is trigonometrically transformed to calculate the sine and cosine values, and the azimuth angle is trigonometrically transformed to calculate the sine and cosine values, thus forming an input vector containing six features. A sliding window strategy is adopted, using data from multiple days prior to the target date as the training set, and outlier removal and Z-score standardization are performed. Train the first-stage extreme gradient boosting model and set the model hyperparameters, including the number of decision trees, the maximum depth of the trees, the learning rate, the sample sampling ratio, the feature sampling ratio, and the L1 and L2 regularization coefficients. The trained first-stage extreme gradient boosting model is used to predict each epoch of the target day, and the geometric component prediction is obtained by inputting the satellite geometric features.

4. The GNSS multipath error suppression method based on spatiotemporal feature decoupling as described in claim 1, characterized in that: The construction of the time-series feature set, which includes hysteresis features and residual time-varying features, includes: For each sample in the training set, the first-stage residual is obtained by subtracting the geometric component prediction value from the observed multipath residual value; the lag feature includes the residual values ​​of multiple epochs prior to the current epoch. The time-varying features of the residuals include: velocity features and acceleration features calculated based on lag residuals, sliding window standard deviation of residuals from multiple epochs, and short-term trend slope obtained by linear fitting of recent residual sequences; A time series feature set containing lag characteristics and residual time-varying characteristics is obtained.

5. The GNSS multipath error suppression method based on spatiotemporal feature decoupling as described in claim 4, characterized in that: The training of the time series feature set using the second-stage extreme gradient boosting model includes: The training data in the second stage are subjected to feature construction and standardization, and the prediction residuals of the extreme gradient boosting model in the first stage are used as the training target. Set the model hyperparameters, including the number of decision trees, the maximum depth of the trees, the learning rate, the sample sampling ratio, the feature sampling ratio, and the L1 and L2 regularization coefficients; The trained second-stage extreme gradient boosting model is used to predict each epoch of the target day. The lag characteristics and time-varying characteristics of the residuals are calculated based on the historical residuals in the historical residual buffer and input into the second-stage extreme gradient boosting model to obtain the predicted values ​​of the time-varying components.

6. The GNSS multipath error suppression method based on spatiotemporal feature decoupling as described in claim 1, characterized in that: The multipath correction value obtained based on the geometric component prediction value and the time-varying component prediction value includes: The sum of the predicted values ​​of the geometric components and the predicted values ​​of the time-varying components is used as the multipath correction value. The multipath correction value is subtracted from the original pseudorange observation value to obtain the corrected pseudorange. The corrected carrier phase is obtained by subtracting the multipath correction value from the original carrier phase observation value; The double-difference positioning solution is re-performed using the corrected observations to obtain the corrected coordinate sequence.

7. The GNSS multipath error suppression method based on spatiotemporal feature decoupling as described in claim 6, characterized in that: The performance verification includes: A sidereal solar filtering method model and a multipath hemispherical diagram method model were constructed as a comparison baseline. The residual reduction rate of each method model on pseudorange and carrier phase observations was calculated to evaluate the multipath suppression performance. The residual reduction rate is defined as the root mean square of the individual error residuals before correction minus the root mean square of the individual error residuals after correction, divided by the root mean square of the individual error residuals before correction, multiplied by 100%. By using simulated dynamic relative positioning calculations, the root mean square errors of positioning in multiple directions of each method model are compared. The performance stability of each method model in continuous observation and its adaptability in dynamic multipath simulation scenarios are analyzed.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the GNSS multipath error suppression method based on spatiotemporal feature decoupling as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the GNSS multipath error suppression method based on spatiotemporal feature decoupling as described in any one of claims 1 to 7.

Images