A Real-Time Processing Method for GNSS-INS Data with Comprehensive Data Quality

Abstract

This invention discloses a real-time GNSS-INS data processing method with comprehensive data quality, belonging to the field of navigation and positioning. It includes: deploying a GNSS accuracy prediction model on a navigation terminal; acquiring GNSS data quality indicators in real time and inferring the predicted GNSS accuracy value for the current epoch; calculating the ratio of this predicted value to the cumulative INS error, and comparing it with an adaptive threshold in segments: if the value is higher than a first threshold, gross errors are removed based on INS; if it is lower than a second threshold, INS is corrected using GNSS; if it is between the two, a fusion solution is performed by adjusting the transition weight associated with the adjustment factor; simultaneously, a two-dimensional reliability judgment space is constructed using the predicted GNSS accuracy value and the cumulative INS error to comprehensively evaluate the reliability of the positioning results. This invention effectively improves the continuous positioning accuracy, stability, and reliability of results in complex scenarios such as urban canyons and tree-lined roads.

Description

Technical Field

[0001] This invention relates to the field of navigation and positioning technology, and specifically to a real-time GNSS-INS data processing method with comprehensive data quality. Background Technology

[0002] With the rapid development of applications such as autonomous driving, UAV mapping, and mobile surveying, the integrated navigation technology of Global Navigation Satellite System (GNSS) and Inertial Navigation System (INS) has become the mainstream solution for providing continuous, reliable, and high-precision positioning services. GNSS has the advantage of high long-term absolute positioning accuracy, while INS has the characteristics of good short-term relative measurement continuity and strong anti-interference capability; the fusion of the two can form a complementary relationship. However, the fusion positioning performance of GNSS / INS integrated navigation is highly dependent on the weight allocation of the two types of observation information in Kalman filtering, and the rationality of the weight allocation depends on a comprehensive evaluation of GNSS data quality and INS error status.

[0003] In complex environments such as urban canyons, tree-lined roads, and tunnel transition sections, GNSS signals are susceptible to multipath effects, obstruction, and electromagnetic interference, leading to a significant decrease in positioning accuracy. If fixed weights or nominal accuracy are still used for fusion in such situations, abnormal GNSS observations will contaminate the INS mechanical arrangement results, causing a deterioration in combined positioning accuracy. Conversely, INS errors accumulate and diverge over time; if the INS drift state is not identified in time after GNSS accuracy recovers, the fusion result will deviate from the true value for a long period. Therefore, real-time and accurate comprehensive evaluation of the current epoch GNSS data quality and INS error level, and dynamic adjustment of the weights between the two types of observation information accordingly, to achieve conditional two-way auxiliary processing, is crucial for improving the continuity and reliability of integrated navigation in complex environments.

[0004] In existing technologies, GNSS / INS data quality assessment and weight allocation strategies mainly include the following categories: The first category consists of simple dynamic models with fixed weights or based on nominal accuracy. These methods do not adequately consider the dynamic impact of the actual observation environment on GNSS signal quality, and in complex scenarios, they cannot reflect the true error level of GNSS positioning results, easily leading to inaccurate weight allocation.

[0005] The second category is variance component estimation algorithms. These methods estimate the variance of observation noise through post-hoc residual statistics and then adjust the weights. However, they are computationally inefficient, have slow convergence speed, and exhibit significant lag, making it difficult to meet the real-time requirements of navigation terminals.

[0006] The third category consists of threshold methods and statistical tests based on single or combined indicators. Threshold methods determine quality by setting fixed thresholds such as signal-to-noise ratio, carrier-to-noise ratio, or cycle slips. While simple to implement, they lack adaptability to dynamic environmental changes. Statistical tests require extensive historical data statistics and hypothesis testing, resulting in heavy computational burdens, poor real-time performance, and difficulty in handling complex dynamic environments.

[0007] The fourth category is traditional machine learning methods. Algorithms such as Support Vector Machines and Random Forests are used for GNSS data quality classification or prediction. However, GNSS data quality indicators typically include categorical features such as satellite system identifiers (GPS / GLONASS / BDS / Galileo) and signal frequency identifiers (L1 / L2 / L5). Traditional machine learning models need to convert these into numerical features through one-hot encoding or label encoding, which can easily lead to feature dimension explosion, increase the risk of model overfitting and computational burden, and is not conducive to real-time deployment on embedded navigation terminals.

[0008] In summary, current technologies have not yet solved the technical problem of how to comprehensively and accurately evaluate GNSS data quality and predict its solution accuracy in real time at navigation terminals, thereby enabling adaptive and conditional bidirectional assistance and weight adjustment between GNSS and INS. Therefore, there is an urgent need for a novel method that can integrate data quality information, natively process mixed-type features, has strong anti-overfitting capabilities, and high inference efficiency to support real-time data quality control and high-precision fusion solution for GNSS / INS integrated navigation in complex environments. Summary of the Invention

[0009] Based on the above analysis, the present invention aims to disclose a real-time GNSS-INS data processing method with comprehensive data quality; and to solve the technical problems of inaccurate weight allocation, lack of two-way assistance, and insufficient reliability assessment in GNSS / INS integrated navigation under complex environments.

[0010] This invention discloses a real-time GNSS-INS data processing method with comprehensive data quality, comprising the following steps: S1. Deploy the GNSS resolution accuracy prediction model to the navigation terminal, calculate the GNSS data quality index in real time, and infer the GNSS resolution accuracy prediction value for the current epoch; wherein, the GNSS resolution accuracy prediction model is a CatBoost regression model trained based on the positioning reference true value obtained by offline precise resolution and a single GNSS resolution accuracy label. S2. The ratio of the predicted GNSS solution accuracy to the cumulative INS error of the current epoch is compared in segments with an adaptive threshold dynamically determined based on the integrated error of the GNSS-INS dual system: when the ratio is higher than the first threshold, the GNSS observation noise covariance is adjusted to make INS dominate the filtering, and a double-difference residual statistic is constructed based on the coordinates derived from INS to eliminate gross errors; when the ratio is lower than the second threshold, the GNSS positioning result is used to correct the INS; when the ratio is between the two thresholds, a transition weight is constructed based on the ratio and an adjustment factor that takes into account the accuracy of both systems, and the GNSS and INS observations participate in the Kalman filter fusion solution together according to the transition weight to obtain the fused positioning result; S3. Construct a two-dimensional reliability judgment space using the predicted value of GNSS solution accuracy and the cumulative error of INS, and comprehensively judge the reliability of the positioning result.

[0011] Furthermore, the GNSS data quality indicators include multipath error, data integrity rate, carrier-to-noise ratio, satellite geometric accuracy factor, satellite number variation, current epoch cycle slip count, and cycle slip ratio.

[0012] Furthermore, the training process of the CatBoost regression model includes: Collect raw GNSS observation data and synchronous IMU data under different environments, and use high-precision integrated navigation equipment to perform offline precision calculations to obtain the reference true value; Using only raw GNSS observation data, a single GNSS RTK solution is performed, and the three-dimensional spatial distance from the reference true value is calculated as the accuracy label; After removing outlier samples, the CatBoost model is trained using GNSS data quality indicators as input features and accuracy labels as regression target values.

[0013] Furthermore, S2 includes the following sub-steps: S2-1. Calculate the cumulative INS error for the current epoch, and obtain the ratio of the predicted GNSS solution accuracy to the cumulative INS error; S2-2. Dynamically determine the adaptive threshold based on the integrated error of the dual system. The adaptive threshold includes a first threshold determined according to the integrated error relationship between the predicted value of GNSS solution accuracy and the cumulative error of INS, and a second threshold set to 1. S2-3. Compare the ratio with the adaptive threshold in segments: When the ratio is higher than the first threshold, adjust the GNSS observation noise covariance to the maximum value so that the filtering result depends almost entirely on the INS output. Use the three-dimensional coordinates derived from INS as the approximate true value to construct the double-difference pseudorange residual test statistic and the double-difference carrier residual test statistic respectively to eliminate the gross errors in the original GNSS observation. S2-4. When the ratio is lower than the second threshold, keep the GNSS observation noise covariance matrix as the nominal value, use the GNSS positioning results to perform feedback correction on the INS, estimate and compensate the position error, velocity error, attitude error and inertial device zero bias error of the INS system, and restart the inertial mechanical orchestration integration from the corrected state. S2-5. When the ratio is between the second threshold and the first threshold, a transition weight is constructed based on the ratio and an adjustment factor that takes into account the accuracy of both systems. GNSS and INS observations participate in the Kalman filter fusion solution together according to the transition weight. S2-6. Substitute the clean GNSS data obtained after gross error removal and INS feedback correction, along with the corrected INS information, into a GNSS / INS combined Kalman filter using an adaptive covariance matrix stochastic model. Calculate the fused positioning result through time updates and measurement updates.

[0014] Furthermore, in S2-1, the calculated cumulative INS error for the current epoch is: ; in, To achieve zero bias in the accelerometer, To achieve zero bias in the gyroscope, It is the acceleration due to gravity. For time intervals.

[0015] Furthermore, in S2-2, the first threshold is determined based on the comprehensive error relationship between the GNSS solution accuracy prediction value and the INS cumulative error. for: Second threshold .

[0016] Furthermore, in S2-3, test statistics for double-difference pseudorange residuals and double-difference carrier residuals are constructed for each non-reference satellite; When any test statistic exceeds the corresponding threshold, the observation value of the corresponding frequency point of the satellite is determined to have gross errors and is removed, thereby obtaining clean GNSS data after removing gross errors.

[0017] Further, in S2-5, when the ratio is between the second threshold and the first threshold, the transition weight is: ; in, As a regulating factor, The first threshold, This is the second threshold.

[0018] Furthermore, S3 includes the following sub-steps: S3-1. Set the allowable error threshold for the positioning result, and use the allowable error threshold as the boundary to divide the two-dimensional reliability judgment space into four regions: high-precision consistency region, GNSS single-system trust region, INS single-system trust region and dual-system failure region. S3-2. When the predicted value of the GNSS solution accuracy and the cumulative error of the INS are both less than the allowable error threshold, the fused positioning result is in the high-precision consistency zone. Perform single INS and single GNSS consistency check: independently obtain the single INS positioning result and the single GNSS positioning result, convert the two to the same coordinate system, calculate the three-dimensional position space deviation, and output the three-level reliability level according to the ratio of the three-dimensional position space deviation to the allowable error threshold. S3-3. When only one of the GNSS solution accuracy prediction value and the INS cumulative error is less than the allowable error threshold, the fusion positioning result is in the single system confidence zone and is determined to be relatively reliable. S3-4. When the predicted value of the GNSS solution accuracy and the cumulative error of the INS are both greater than or equal to the allowable error threshold, the fused positioning result is in the dual system failure zone, is determined to be unreliable, and triggers a system alarm.

[0019] Furthermore, in S3-2, the three reliability levels include: If the three-dimensional positional spatial deviation is less than one-third of the allowable error threshold, the fusion positioning result is determined to be highly reliable. If the three-dimensional position spatial deviation is greater than or equal to one-third of the allowable error threshold and less than the allowable error threshold, the fusion positioning result is determined to be reliable. If the three-dimensional positional spatial deviation is greater than or equal to the allowable error threshold, it is downgraded to unreliable.

[0020] Compared with traditional methods, the present invention has the following technical advantages: By using a CatBoost regression model to predict GNSS solution accuracy in real time and natively handle categorical features, model complexity is reduced, facilitating lightweight deployment of navigation terminals. Segmented comparisons are performed using the ratio of GNSS prediction accuracy to INS cumulative error and an adaptive threshold, enabling adaptive dynamic adjustment and conditional bidirectional assistance of GNSS and INS observation weights. When GNSS accuracy deteriorates, INS-led filtering is used, and a double-difference residual statistic is constructed based on INS-derived coordinates to eliminate gross errors. When INS drift intensifies, GNSS positioning results are used to correct INS cumulative errors. In the transition interval, an adjustment factor that considers the accuracy of both systems is constructed to ensure fusion continuity, thereby improving the continuous positioning accuracy and long-term stability of integrated navigation in complex environments. Simultaneously, a two-dimensional reliability judgment space is constructed using GNSS solution accuracy predictions and INS cumulative errors, transforming abstract error predictions into semantic reliability levels, providing downstream navigation applications with confidence labels that can be directly used for functional safety decisions. Attached Figure Description

[0021] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of the real-time GNSS-INS data processing method in an embodiment of the present invention; Figure 2 This is a two-dimensional reliability judgment space diagram of the GNSS solution accuracy prediction value and the INS cumulative error in an embodiment of the present invention. Figure 3 This is a comparison chart of the localization results of the fixed weight strategy and the method of the present invention in this embodiment; wherein, Figure 3 (a) in the figure represents the fixed-weight localization result; Figure 3 (b) in the figure represents the positioning result of the method of the present invention. Detailed Implementation

[0022] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and, together with the embodiments of the present invention, serve to illustrate the principles of the present invention.

[0023] One embodiment of the present invention discloses a real-time GNSS-INS data processing method with comprehensive data quality, such as... Figure 1 As shown, it includes the following steps: S1. Deploy the GNSS resolution accuracy prediction model to the navigation terminal, calculate the GNSS data quality index in real time, and infer the GNSS resolution accuracy prediction value for the current epoch; wherein, the GNSS resolution accuracy prediction model is a CatBoost regression model trained based on the positioning reference true value obtained by offline precise resolution and a single GNSS resolution accuracy label. S2. The ratio of the predicted GNSS solution accuracy to the cumulative INS error of the current epoch is compared in segments with an adaptive threshold dynamically determined based on the integrated error of the GNSS-INS dual system: when the ratio is higher than the first threshold, the GNSS observation noise covariance is adjusted to make INS dominate the filtering, and a double-difference residual statistic is constructed based on the coordinates derived from INS to eliminate gross errors; when the ratio is lower than the second threshold, the GNSS positioning result is used to correct the INS; when the ratio is between the two thresholds, a transition weight is constructed based on the ratio and an adjustment factor that takes into account the accuracy of both systems, and the GNSS and INS observations participate in the Kalman filter fusion solution together according to the transition weight to obtain the fused positioning result; S3. Construct a two-dimensional reliability judgment space using the predicted value of GNSS solution accuracy and the cumulative error of INS, and comprehensively judge the reliability of the positioning result.

[0024] Specifically, in S1, using only the raw GNSS observation data from the integrated navigation data (without IMU information) and combining it with the satellite broadcast ephemeris output by the receiver, the following GNSS data quality indicators are calculated: Multipath error (MP): Calculated by differencing pseudorange observations with smoothed carrier phase observations; Data Completeness (DI): The ratio of the number of satellites actually observed to the theoretically visible number at the current epoch; Carrier-to-noise ratio C / N0: The statistical value of the signal carrier-to-noise ratio for each satellite at each frequency; Satellite Geometric Precision Factor (GDOP): A geometric precision factor calculated based on the current spatial distribution of satellites; Satellite number change ΔN: The change in the number of visible satellites between adjacent epochs; Current epoch cycle slip count (CS): The number of cycle slips occurring in the current epoch is detected by polynomial fitting or MW combination of carrier phase observations; Cycle slip ratio (CR): The ratio of the number of satellites that experienced cycle slips to the number of observed satellites at two different time points.

[0025] The above-mentioned indicators are calculated using existing GNSS data processing technology, utilizing only the raw observation data and broadcast ephemeris output by the GNSS receiver, without involving INS auxiliary information.

[0026] Specifically, in S1, the training process of the CatBoost regression model includes: 1) Collect raw GNSS observation data and synchronous IMU data under different environments, and use high-precision integrated navigation equipment to perform offline precision calculations to obtain the reference true value; Specifically, GNSS observation data and synchronous high-precision IMU data are collected in different environments (such as open environments, urban canyons, tree-lined roads, tunnel transition sections, etc.); raw data are collected using high-precision integrated navigation equipment (such as tactical or navigation-grade IMUs), and offline precise calculations are performed using post-processing software such as Inertial Explorer to obtain high-precision positioning results at the centimeter or even millimeter level, which serve as the positioning reference true values ​​for subsequent model training.

[0027] 2) Perform single GNSS RTK calculations using only raw GNSS observation data, and calculate the three-dimensional spatial distance from the reference true value as the accuracy label; Specifically, using only raw GNSS observation data, the standard Kalman filter algorithm is employed to perform single GNSS RTK positioning calculations, yielding single GNSS positioning results. Compare this result with the high-precision reference value. The difference is calculated, and the three-dimensional spatial distance is taken as the single GNSS solution accuracy label for that epoch: ; in, For high-precision reference three-dimensional coordinates, The coordinates are obtained from the Kalman filter solution of a single GNSS array.

[0028] 3) After removing outlier samples, train the CatBoost model using GNSS data quality indicators as input features and accuracy labels as regression target values; include: (1) Perform statistical analysis on the calculated solution accuracy label sequence and remove outliers; The 3σ criterion or the box plot method (IQR method) can be used to calculate the mean of the precision label sequence. and standard deviation Eliminate those that meet the requirements Outlier samples; or calculate quartiles. , and interquartile range Remove those located in Samples outside the specified interval. After removing outliers, the remaining samples constitute the model training samples; (2) Construction of CatBoost regression prediction model; The training sample set is constructed by using GNSS data quality indicators as input features and single GNSS solution accuracy after removing outliers as the regression target value. This paper utilizes the CatBoost algorithm to construct a nonlinear mapping relationship between GNSS data quality indicators and GNSS solution accuracy. The CatBoost algorithm natively supports categorical features, directly processing categorical data (such as different satellite system identifiers: GPS / GLONASS / BDS / Galileo, different signal frequency identifiers: L1 / L2 / L5, etc.) without one-hot encoding or label encoding, and without requiring normalization preprocessing for numerical features. During model training, the categorical feature index is defined directly through the `cat_features` parameter, and the algorithm automatically calculates the optimal split points for categorical features, thus avoiding the feature dimension explosion problem caused by one-hot encoding, reducing the risk of model overfitting, and improving training and inference efficiency.

[0029] By adjusting CatBoost's key hyperparameters (such as number of iterations, learning rate, tree depth, L2 regularization coefficient, etc.) through grid search or Bayesian optimization, and using root mean square error (RMSE) or mean absolute error (MAE) as the loss function, a regression prediction model for GNSS resolution accuracy is trained. The model takes a GNSS data quality index vector for the current epoch as input and outputs the predicted GNSS solution accuracy. .

[0030] Specifically, in S1, model deployment and real-time inference.

[0031] The trained GNSS resolution accuracy prediction model Deployed to navigation terminals (such as vehicle-mounted integrated positioning devices, UAV flight control systems, or handheld mapping terminals). The model is embedded in the terminal processor in a lightweight form (such as ONNX format or CatBoost native C++ library), and real-time inference is achieved through a predictive inference engine based on GNSS solution accuracy. In real-time navigation scenarios, based on the raw observation data (pseudorange, carrier phase, Doppler, etc.) received by the GNSS receiver and satellite broadcast ephemeris, GNSS data quality indicators such as multipath error (MP), data integrity rate (DI), carrier-to-noise ratio (C / N0), satellite geometrical accuracy factor (GDOP), satellite number change (ΔN), current epoch cycle slip count (CS), and cycle slip ratio (CR) are calculated in real time and used as models. The input features are then used as input to the deployed prediction model. The inference engine outputs the GNSS solution accuracy prediction value for the current epoch. .

[0032] In summary, step S1 uses the CatBoost regression model to natively process categorical features (satellite systems, signal frequencies, etc.) in GNSS data. This eliminates the need for one-hot encoding or label encoding, avoids feature dimension explosion, reduces the risk of overfitting, and provides high training and inference efficiency. It is also easy to deploy in a lightweight manner to embedded navigation terminals such as vehicles and drones.

[0033] Specifically, S2 includes: S2-1. Calculate the cumulative INS error for the current epoch, and obtain the ratio of the predicted GNSS solution accuracy to the cumulative INS error; S2-2. Dynamically determine the adaptive threshold based on the integrated error of the dual system. The adaptive threshold includes a first threshold determined according to the integrated error relationship between the predicted value of GNSS solution accuracy and the cumulative error of INS, and a second threshold set to 1. S2-3. Compare the ratio with the adaptive threshold in segments: When the ratio is higher than the first threshold, adjust the GNSS observation noise covariance to the maximum value so that the filtering result depends almost entirely on the INS output. Use the three-dimensional coordinates derived from INS as the approximate true value to construct the double-difference pseudorange residual test statistic and the double-difference carrier residual test statistic respectively to eliminate the gross errors in the original GNSS observation. S2-4. When the ratio is lower than the second threshold, keep the GNSS observation noise covariance matrix as the nominal value, use the GNSS positioning results to perform feedback correction on the INS, estimate and compensate the position error, velocity error, attitude error and inertial device zero bias error of the INS system, and restart the inertial mechanical orchestration integration from the corrected state. S2-5. When the ratio is between the second threshold and the first threshold, a transition weight is constructed based on the ratio and an adjustment factor that takes into account the accuracy of both systems. GNSS and INS observations participate in the Kalman filter fusion solution together according to the transition weight. S2-6. Substitute the clean GNSS data obtained after gross error removal and INS feedback correction, along with the corrected INS information, into a GNSS / INS combined Kalman filter using an adaptive covariance matrix stochastic model. Calculate the fused positioning result through time updates and measurement updates.

[0034] Specifically, in S2-1, the calculated cumulative INS error for the current epoch is: ; in, To achieve zero bias in the accelerometer, To achieve zero bias in the gyroscope, It is the acceleration due to gravity. For time intervals.

[0035] Specifically, in S2-2, the first threshold for: Second threshold The value is 1.

[0036] The adaptive transition weights, which are formed by segmented comparisons based on the dynamically determined adaptive threshold of the GNSS-INS dual-system integrated error, are as follows: ; in, As an adjustment factor, the accuracy of both the GNSS and INS systems is taken into account.

[0037] Specifically, in S2-3, when hour, This is the maximum value, which can be set as needed, for example... At this point, the GNSS observation noise covariance matrix is ​​adjusted to... ,in The initial covariance matrix is ​​constructed based on the receiver's nominal pseudorange accuracy and the current observation geometry; due to As the Kalman gain approaches its maximum, the filtering result becomes almost entirely dependent on the INS mechanical orchestration output. Meanwhile, high-precision coordinates derived using INS are utilized. As an approximate true value, gross errors are eliminated from the raw GNSS observations.

[0038] Specifically, test statistics for double-difference pseudorange residuals and double-difference carrier residuals are constructed for each non-reference satellite:

[0039] in, To obtain satellite coordinates using satellite broadcast ephemeris, The rover coordinates derived from the INS system, These are GNSS double-difference pseudorange observations. These are double-difference carrier observations. This is the double-difference ionospheric delay error. This is the double-difference tropospheric delay error. For the corresponding carrier wavelength, For double-difference ambiguity, For pseudorange observation accuracy, For carrier observation accuracy.

[0040] When any of the above test statistics exceeds the corresponding threshold, it is determined that the observation value of the corresponding frequency point of the satellite has gross errors and is removed, thereby obtaining clean GNSS data after removing gross errors.

[0041] Specifically, in S2-4, when hour, At this time, the GNSS observation noise covariance matrix retains its nominal value. GNSS is used for filtering with nominal weights. The clean GNSS positioning results after removing gross errors are used as external high-precision observations to perform feedback correction on the INS.

[0042] Specifically, the difference between the 3D position calculated by GNSS and the 3D position obtained by INS mechanical arrangement is used as the measurement information for the GNSS / INS combined Kalman filter. Through the measurement update process of Kalman filtering, the position error, velocity error, attitude error, gyroscope bias error, and accelerometer bias error of the INS system are estimated in real time. The estimated error state is fed back to the INS navigation solution unit to compensate and correct the position, velocity, attitude, and inertial device bias of the INS. The inertial mechanical arrangement integration is restarted from the corrected state, thereby suppressing the divergence of INS error over time.

[0043] Specifically, in S2-5, when hour, Take intermediate transition value GNSS and INS observations participate in the Kalman filter fusion solution together with adjusted weights. In this stage, the filter itself completes the optimal fusion based on the stochastic model.

[0044] The clean GNSS data, after removing gross errors and processed by the above two-way auxiliary method, along with the corrected INS information, are then substituted into the GNSS / INS combined Kalman filter; wherein, the stochastic model for GNSS observations is adopted... Adjusted observation noise covariance matrix The final fusion localization result is calculated by combining time and measurement updates using Kalman filtering.

[0045] In summary, step S2 dynamically determines the adaptive threshold based on the integrated error of the GNSS-INS dual systems and introduces an adjustment factor that takes into account the accuracy of both systems to construct transition weights. This smoothly adjusts the observation noise covariance as accuracy changes, preventing normal errors from being misjudged as gross errors and ensuring the effective utilization of observation data and the continuity of fusion calculations in complex environments. When GNSS accuracy deteriorates significantly, INS-dominated filtering and a double-difference residual statistic derived from INS coordinates are used to eliminate gross errors and suppress abnormal observation contamination. When INS drift intensifies, clean GNSS observation feedback is used to correct the INS error state and suppress error divergence. The transition interval is fused according to the optimal weights, thereby improving the continuous positioning accuracy and long-term stability in scenarios such as urban canyons and tree-lined roads.

[0046] Specifically, S3 includes: S3-1. Set the allowable error threshold for the positioning result, and use the allowable error threshold as the boundary to divide the two-dimensional reliability judgment space into four regions: high-precision consistency region, GNSS single-system trust region, INS single-system trust region and dual-system failure region. Specifically, while obtaining the fused positioning results, a two-dimensional reliability judgment space is constructed based on the following two prior pieces of information: Horizontal axis: GNSS solution accuracy prediction value ; Vertical axis: INS cumulative error .

[0047] Set the allowable error threshold for the positioning results as follows: (Based on the application scenario, such as 0.3 m or 1.0 m for autonomous driving scenarios, etc.), with and To define the boundaries, the two-dimensional coordinate plane is divided into four regions (A, B, C, and D), each corresponding to a different reliability level. For example... Figure 2 As shown; S3-2. When the predicted value of the GNSS solution accuracy and the cumulative error of the INS are both less than the allowable error threshold, the fused positioning result is in the high-precision consistency zone. Perform single INS and single GNSS consistency check: independently obtain the single INS positioning result and the single GNSS positioning result, convert the two to the same coordinate system, calculate the three-dimensional position space deviation, and output the three-level reliability level according to the ratio of the three-dimensional position space deviation to the allowable error threshold. In a more specific description, The fused location result is in region A, i.e. and At this point, both the GNSS prediction accuracy and the INS cumulative error are within acceptable ranges, indicating that both independent subsystems have high reliability.

[0048] Perform a consistency check between a single INS and a single GNSS, whereby... Single INS positioning results: Based on the current IMU's accelerometer and gyroscope data, the carrier's position information is obtained through independent integration using zero bias correction, initial alignment (static or dynamic), and mechanical arrangement algorithms; Single GNSS positioning results: Using clean GNSS observation data after removing gross errors at the current epoch, the carrier's position information is independently calculated using standard RTK or relative positioning algorithms; Coordinate System 1: Convert the single INS and single GNSS positioning results to the same coordinate system (such as the Earth-centered Earth-fixed Coordinate System ECEF or the local Northeast-Northern Sky Coordinate System ENU), and calculate the three-dimensional spatial deviation between the two. .

[0049] Specifically, the three reliability levels include: If the three-dimensional positional spatial deviation is less than one-third of the allowable error threshold, the fusion positioning result is determined to be highly reliable. Right now This indicates that the positioning results of the two independent subsystems are highly consistent, and the fusion result is very reliable; If the three-dimensional position spatial deviation is greater than or equal to one-third of the allowable error threshold and less than the allowable error threshold, the fusion positioning result is determined to be reliable. Right now This indicates that there are some differences between the two subsystems, but they are within acceptable limits, and the fusion result is reliable. If the three-dimensional positional spatial deviation is greater than or equal to the allowable error threshold, it is downgraded to unreliable.

[0050] If Even if the predicted accuracy of each value is lower than the threshold, the actual consistency is poor, so it is downgraded to unreliable and triggers an alarm for reassessment.

[0051] Note: The selection criteria are based on the fact that when the consistency deviation between the two independent systems is less than one-third of the allowable error, the fusion result can be considered to have a high degree of confidence.

[0052] S3-3. When only one of the GNSS solution accuracy prediction value and the INS cumulative error is less than the allowable error threshold, the fusion positioning result is in the single system confidence zone and is determined to be relatively reliable. The fused localization result is in region B or region C when any of the following conditions are met: Area B: and (GNSS accuracy is reliable, but INS has clearly drifted). Area C: and (INS accuracy is reliable, GNSS accuracy deteriorates).

[0053] At this point, only one subsystem is in a high-precision state, and the fusion result is mainly dominated by this reliable subsystem, while the weight of the other subsystem has been significantly reduced. Due to the lack of consistency cross-validation between the two independent systems, the positioning result is determined to be relatively reliable.

[0054] S3-4. When the predicted value of the GNSS solution accuracy and the cumulative error of the INS are both greater than or equal to the allowable error threshold, the fused positioning result is in the dual system failure zone, is determined to be unreliable, and triggers a system alarm.

[0055] The fused positioning result is in region D, i.e. and At this point, both the GNSS prediction accuracy and the INS cumulative error have exceeded the allowable thresholds, indicating that neither subsystem is reliable. Regardless of how the fusion weights are allocated, the positioning result is deemed unreliable, triggering a system alarm and prompting a switch to an alternative positioning method (such as map matching, V2X assistance, or manual takeover).

[0056] In summary, step S3 constructs a two-dimensional reliability judgment space for GNSS prediction accuracy and INS cumulative error, and combines single INS and single GNSS consistency verification to output a three-level reliability level of very reliable, relatively reliable, and unreliable. This transforms the abstract error prediction into a functional safety confidence label that can be directly used in downstream applications such as autonomous driving and surveying.

[0057] To verify the effectiveness of the algorithm, this embodiment processes observation data collected by a low-cost integrated positioning terminal, and performs positioning processing using a fixed GNSS / INS weight strategy and the strategy of this invention, respectively. Figure 3 As shown, where, Figure 3 (a) in the figure represents the fixed-weight localization result (left figure); Figure 3 (b) in the figure shows the positioning results of the present invention (right figure). The horizontal axis of both figures represents the observation time (unit: seconds), and the vertical axis represents the positioning deviation (unit: centimeters). In the legend, the blue dashed line represents the eastward (E) deviation, the red dashed line represents the northward (N) deviation, and the yellow dashed line represents the celestial (U) deviation.

[0058] The left figure (fixed-weight positioning results) shows that under the fixed-weight allocation strategy, the positioning deviations in each direction exhibit a clear systematic divergence trend as the observation time progresses. Specifically, the celestial (U) deviation gradually accumulates from approximately 100 seconds, reaching approximately 580 cm after 700 seconds; the eastward (E) deviation appears to increase from approximately 100 seconds later. The systemic offset of centimeters persists; although the north (N) offset is relatively small, a deviation of about 20 to 30 centimeters also occurs in the range of about 400 to 600 seconds. The above results indicate that the fixed weight strategy fails to effectively suppress the divergence error accumulated by the inertial navigation system over time, resulting in a serious deterioration in long-term positioning performance.

[0059] The right figure (positioning results of the invention) shows that after adopting the comprehensive data quality real-time processing method described in this invention, the positioning deviations in the east (E), north (N), and sky (U) directions were effectively controlled within ±1.5 cm throughout the entire observation period (0-1000 seconds), and fluctuated within ±0.5 cm for most of the time, without any systematic accumulation or divergence. Especially after a brief fluctuation in the initial stage (0-100 seconds), the deviations in each direction quickly converged and remained stable.

[0060] In summary, the comparison of the two figures shows that the solution in this embodiment significantly suppresses the cumulative error of inertial devices in complex environments by predicting GNSS solution accuracy in real time, adaptively adjusting GNSS / INS observation weights, and supplementing with conditional bidirectional error correction. This reduces the positioning deviation from the order of hundreds of centimeters to the order of centimeters or even sub-centimeters, effectively improving the long-term positioning accuracy and stability of the integrated navigation system.

[0061] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A real-time processing method for GNSS-INS data with comprehensive data quality, characterized in that, Includes the following steps: S1. Deploy the GNSS resolution accuracy prediction model to the navigation terminal, calculate the GNSS data quality index in real time, and infer the GNSS resolution accuracy prediction value for the current epoch; wherein, the GNSS resolution accuracy prediction model is a CatBoost regression model trained based on the positioning reference true value obtained by offline precise resolution and a single GNSS resolution accuracy label. S2. The ratio of the predicted GNSS solution accuracy to the cumulative INS error of the current epoch is compared in segments with an adaptive threshold dynamically determined based on the integrated error of the GNSS-INS dual system: when the ratio is higher than the first threshold, the GNSS observation noise covariance is adjusted to make INS dominate the filtering, and a double-difference residual statistic is constructed based on the coordinates derived from INS to eliminate gross errors; when the ratio is lower than the second threshold, the GNSS positioning result is used to correct the INS; when the ratio is between the two thresholds, a transition weight is constructed based on the ratio and an adjustment factor that takes into account the accuracy of both systems, and the GNSS and INS observations participate in the Kalman filter fusion solution together according to the transition weight to obtain the fused positioning result; S3. Construct a two-dimensional reliability judgment space using the GNSS solution accuracy prediction value and the INS cumulative error, and comprehensively judge the reliability of the positioning result; S3 includes the following sub-steps: S3-1. Set the allowable error threshold for the positioning result, and use the allowable error threshold as the boundary to divide the two-dimensional reliability judgment space into four regions: high-precision consistency region, GNSS single-system trust region, INS single-system trust region and dual-system failure region. S3-2. When the predicted value of the GNSS solution accuracy and the cumulative error of the INS are both less than the allowable error threshold, the fused positioning result is in the high-precision consistency zone. Perform single INS and single GNSS consistency check: independently obtain the single INS positioning result and the single GNSS positioning result, convert the two to the same coordinate system, calculate the three-dimensional position space deviation, and output the three-level reliability level according to the ratio of the three-dimensional position space deviation to the allowable error threshold. S3-3. When only one of the GNSS solution accuracy prediction value and the INS cumulative error is less than the allowable error threshold, the fusion positioning result is in the single system confidence zone and is determined to be relatively reliable. S3-4. When the predicted value of the GNSS solution accuracy and the cumulative error of the INS are both greater than or equal to the allowable error threshold, the fused positioning result is in the dual system failure zone, is determined to be unreliable, and triggers a system alarm.

2. The GNSS-INS data real-time processing method for comprehensive data quality according to claim 1, characterized in that, The GNSS data quality indicators include multipath error, data integrity rate, carrier-to-noise ratio, satellite geometric accuracy factor, satellite number variation, current epoch cycle slip count, and cycle slip ratio.

3. The GNSS-INS data real-time processing method for comprehensive data quality according to claim 1, characterized in that, The training process of the CatBoost regression model includes: Collect raw GNSS observation data and synchronous IMU data under different environments, and use high-precision integrated navigation equipment to perform offline precision calculations to obtain the reference true value; Using only raw GNSS observation data, a single GNSS RTK solution is performed, and the three-dimensional spatial distance from the reference true value is calculated as the accuracy label; After removing outlier samples, the CatBoost model is trained using GNSS data quality indicators as input features and accuracy labels as regression target values.

4. The GNSS-INS data real-time processing method for comprehensive data quality according to claim 1, characterized in that, S2 includes the following sub-steps: S2-1. Calculate the cumulative INS error for the current epoch, and obtain the ratio of the predicted GNSS solution accuracy to the cumulative INS error; S2-2. Dynamically determine the adaptive threshold based on the integrated error of the dual system. The adaptive threshold includes a first threshold determined according to the integrated error relationship between the predicted value of GNSS solution accuracy and the cumulative error of INS, and a second threshold set to 1. S2-3. Compare the ratio with the adaptive threshold in segments: When the ratio is higher than the first threshold, adjust the GNSS observation noise covariance to the maximum value so that the filtering result depends almost entirely on the INS output. Use the three-dimensional coordinates derived from INS as the approximate true value to construct the double-difference pseudorange residual test statistic and the double-difference carrier residual test statistic respectively to eliminate the gross errors in the original GNSS observation. S2-4. When the ratio is lower than the second threshold, keep the GNSS observation noise covariance matrix as the nominal value, use the GNSS positioning results to perform feedback correction on the INS, estimate and compensate the position error, velocity error, attitude error and inertial device zero bias error of the INS system, and restart the inertial mechanical orchestration integration from the corrected state. S2-5. When the ratio is between the second threshold and the first threshold, a transition weight is constructed based on the ratio and an adjustment factor that takes into account the accuracy of both systems. GNSS and INS observations participate in the Kalman filter fusion solution together according to the transition weight. S2-6. Substitute the clean GNSS data obtained after gross error removal and INS feedback correction, along with the corrected INS information, into a GNSS / INS combined Kalman filter using an adaptive covariance matrix stochastic model. Calculate the fused positioning result through time updates and measurement updates.

5. The GNSS-INS data real-time processing method for comprehensive data quality according to claim 4, characterized in that, In S2-1, the calculated cumulative INS error for the current epoch is: ； in, To achieve zero bias in the accelerometer, To achieve zero bias in the gyroscope, It is the acceleration due to gravity. For time intervals.

6. The GNSS-INS data real-time processing method for comprehensive data quality according to claim 4, characterized in that, In S2-2, the first threshold is determined based on the comprehensive error relationship between the predicted GNSS solution accuracy and the cumulative INS error. for: Second threshold .

7. The GNSS-INS data real-time processing method for comprehensive data quality according to claim 4, characterized in that, In S2-3, test statistics for double-difference pseudorange residuals and double-difference carrier residuals are constructed for each non-reference satellite. When any test statistic exceeds the corresponding threshold, the observation value of the corresponding frequency point of the satellite is determined to have gross errors and is removed, thereby obtaining clean GNSS data after removing gross errors.

8. The GNSS-INS data real-time processing method for comprehensive data quality according to claim 4, characterized in that, In S2-5, when the ratio is between the second threshold and the first threshold, the transition weight is: ； in, As a regulating factor, The first threshold, This is the second threshold.

9. The real-time GNSS-INS data processing method for comprehensive data quality according to claim 1, characterized in that, In S3-2, the three reliability levels include: If the three-dimensional positional spatial deviation is less than one-third of the allowable error threshold, the fusion positioning result is determined to be highly reliable. If the three-dimensional position spatial deviation is greater than or equal to one-third of the allowable error threshold and less than the allowable error threshold, the fusion positioning result is determined to be reliable. If the three-dimensional positional spatial deviation is greater than or equal to the allowable error threshold, it is downgraded to unreliable.

Images