GNSS deformation monitoring coordinate reliability quantitative evaluation and optimization method and system

Abstract

The application provides a GNSS deformation monitoring coordinate reliability quantitative evaluation and optimization method and system, and belongs to the field of deformation monitoring. The method comprises the following steps: based on the data of the monitoring station and the reference station, GNSS multi-dimensional derived features and coordinates are solved, and the overall absolute deviation of the monitoring station is determined; the SGM solution sample is divided into normal and abnormal samples, a data set is constructed, and robust normalization and hierarchical sample equalization are performed to obtain a target data set; a physical prior constraint adaptive weighted fusion loss function is used to train an MLP reliability quantitative model, GNSS multi-dimensional derived features are input into the trained complete model to obtain a reliability quantitative score; based on the score, the Kalman filter observation noise is adaptively adjusted, and the state of the solution coordinates is updated to obtain the optimized station coordinates. The application realizes reliable and universal quantitative evaluation of the solution coordinates, and realizes noise suppression in the smooth stage and rapid response in the deformation stage through adaptive filtering, thereby effectively reducing the false alarm and missed alarm risks.

Description

Technical Field

[0001] This invention relates to the field of GNSS high-precision deformation monitoring data processing, and specifically to a method and system for quantitative evaluation and optimization of GNSS deformation monitoring coordinate reliability. Background Technology

[0002] Since its inception, GNSS high-precision positioning technology has revolutionized the field of surface deformation monitoring, expanding its applications from early experimental explorations to various engineering fields such as geological disaster prevention, structural health monitoring, and mine subsidence detection. Real-time dynamic positioning (RTK), precise point positioning (PPP), and static GNSS measurement (SGM) algorithms are widely used in engineering practice. Among these, after long-term practical demonstration and engineering verification, SGM, with its advantages of millimeter-level accuracy and high stability, remains the preferred solution for engineering scenarios such as geological disaster and structural health monitoring, and the need for verification of its solution reliability is even more urgent. SGM is essentially a simple short baseline solution, which theoretically can converge positioning accuracy to a few millimeters.

[0003] Although GNSS algorithms, such as SGM, possess theoretical accuracy at the millimeter or centimeter level, in real deformation monitoring environments, GNSS signals are significantly affected by complex environmental factors surrounding the antenna. When complex reflection or interference conditions, severe electromagnetic interference, or gross errors occur at a given time, the current calculation results will exhibit significant deviations, including unmodeled residual errors such as multipath errors. Consequently, the actual accuracy is often lower than the nominal value. In deformation monitoring, when the monitored target exhibits abrupt changes (such as landslides or sudden-type mudslides), it is difficult to distinguish in real time between abnormal deviations in GNSS results and actual displacement phenomena. Currently, a certain upper limit for displacement thresholds is typically used to determine whether a hazard alarm is triggered. Therefore, when the two cannot be effectively distinguished, frequent false alarms are highly likely. To address these issues, existing technologies have made various attempts. At the algorithm optimization level, some scholars have introduced auxiliary sensors (such as accelerometers) to supplement observations when GNSS signals are blocked, in order to maintain millimeter-level monitoring capabilities in complex environments. Others have proposed using data-driven techniques to invert unmodeled errors, reducing the impact of multipath errors and non-line-of-sight signals on positioning accuracy. Still others have adopted robust estimation-based outlier automatic detection methods to identify and process outliers of varying sizes, or explored fusing GNSS and inertial navigation system data to improve stability in complex noisy environments. Furthermore, some solutions, from a data quality assessment perspective, qualitatively describe the reliability of the solution results using indicators such as pseudorange multipath error, signal-to-noise ratio, and cycle slip ratio. However, the above methods still have obvious shortcomings: In terms of algorithm optimization, existing methods focus too much on making the coordinate sequence more in line with the accuracy requirements, but neglect the compatibility with the real displacement. That is, in the process of real-time detection and correction, they cannot effectively distinguish between coordinate outliers and real displacement changes, and are very likely to misjudge real displacement features as outliers and remove them, thus causing the problem of missed alarms; In terms of multi-index evaluation, existing schemes can only qualitatively describe the relationship between each index and the solution result, and cannot capture the nonlinear correlation between the two. Moreover, there is no absolute correspondence between the value of a single index and the solution accuracy, and there are special scenarios where the index score is poor but the solution result is occasionally reliable.

[0004] In summary, the core problem faced by existing technologies is that they fail to capture the complex nonlinear relationship between multi-dimensional indicators and solution reliability, and cannot achieve accurate quantitative differentiation between abnormal deviations and true displacements, thus the core contradiction between false alarms and missed alarms cannot be fundamentally resolved. Summary of the Invention

[0005] In view of this, it is necessary to provide a GNSS deformation monitoring coordinate reliability quantification assessment and optimization method and system to solve the technical problems in the existing technology that fail to capture the complex nonlinear relationship between multi-dimensional indicators and solution reliability, and cannot achieve accurate quantitative differentiation between abnormal deviations and true displacements, resulting in false alarms and missed alarms.

[0006] To address the aforementioned technical problems, in a first aspect, the present invention provides a method for quantitative evaluation and optimization of GNSS deformation monitoring coordinate reliability, comprising: SGM calculation is performed based on GNSS monitoring station data and GNSS reference station data to obtain GNSS multidimensional derived features and calculated coordinates of monitoring stations; the total absolute deviation of monitoring stations is determined based on the calculated coordinates. Based on the quartile threshold method, the samples corresponding to each SGM solution are divided into normal samples and abnormal samples according to the total absolute deviation. The original dataset is constructed based on the normal samples, the abnormal samples and the GNSS multidimensional derived features. The original dataset is then subjected to site-level robust normalization and hierarchical site-level sample balancing to obtain the target dataset. An adaptive weighted fusion loss function with physical prior constraints is used to train an MLP reliability metric model based on the target dataset, resulting in a fully trained MLP reliability metric model. The GNSS multidimensional derived features corresponding to the baseline to be tested are input into the fully trained MLP reliability metric model to obtain the reliability metric score of the corresponding SGM solution. Based on the reliability metric score, the baseline observation noise of the Kalman filter in the SGM solution is adaptively adjusted to obtain the optimized observation noise; based on the optimized observation noise, the solution coordinates are updated using Kalman filtering to obtain the optimized monitoring station coordinates.

[0007] In one possible implementation, the GNSS multidimensional derived features include multidimensional raw data quality features and multidimensional solution process features; The multidimensional raw data quality features include signal-to-noise ratio correlation features, cycle slip ratio features, data integrity rate features, pseudorange multipath features, and noise features; The multidimensional solution process features include single-point positioning accuracy features, unfixed solution ratio features, variance anomaly features, ambiguity variance ratio statistical features, unit weight mean square error statistical features, accuracy factor features, pre-verification residual features, and post-verification residual features.

[0008] In one possible implementation, determining the overall absolute deviation of the monitoring station based on the calculated coordinates includes: Obtain the calculated coordinates of the monitoring stations for the SGM solution; Determine the solution coordinates in , , The deviation values ​​relative to the coordinates of the reference station in three dimensions are used to determine the Euclidean distance based on the deviation values, thus obtaining the overall absolute deviation of the monitoring station.

[0009] In one possible implementation, the step of performing site-level robust normalization and hierarchical site-level sample equalization on the original dataset to obtain the target dataset includes: All the GNSS multidimensional derived features are processed using site-level robust normalization on the original dataset to generate a transitional dataset; The transition dataset is subjected to hierarchical site-level synthetic minority class oversampling to achieve sample balancing and obtain the target dataset.

[0010] In one possible implementation, the step of applying site-level robust normalization to all the GNSS multidimensional derived features of the original dataset to generate a transition dataset includes: The original dataset is divided into independent site subsets based on the monitoring site identifiers; For each GNSS derived feature in each station subset, the median and interquartile range of the corresponding feature are determined respectively; if the interquartile range of a certain feature of a certain station is equal to 0, the normalization value of that feature is set to 0; if the interquartile range of a certain feature of a certain station is greater than 0, the feature is standardized based on the median and interquartile range of the corresponding station. After standardizing the GNSS-derived features of all stations, a normalized transition dataset is obtained.

[0011] In one possible implementation, the hierarchical site-level synthetic minority class oversampling of the transition dataset to achieve sample equalization and obtain the target dataset includes: A site feature statistical vector is constructed using a single site as a unit, and the distribution similarity between sites is determined by cosine similarity. Sites with similarity greater than or equal to a preset similarity threshold are identified as similar sites. For stations with no fewer than 5 anomalous samples, synthetic samples are generated by interpolation based on the anomalous samples of the station. For stations with fewer than 5 anomalous samples, anomalous samples from all similar stations of the station are merged, and a preset number of neighboring samples are selected and combined with Gaussian noise to generate synthetic samples that do not overlap with the original samples. After generating synthetic samples for each site, the total number of its abnormal samples is limited to a preset proportion of the number of normal samples at that site to obtain the target dataset.

[0012] In one possible implementation, the input to the MLP reliable metric model is the GNSS multidimensional derived features after site-level robust normalization, and the output is the corresponding baseline SGM solution coordinates. Interval reliability metric score; the MLP reliability metric model includes a feature attention layer set between the input layer and the classifier module, and a classifier module and an output layer connected in series with the feature attention layer; The feature attention layer is an attention weighting module added to the front end of the network and built into a two-layer perceptron sub-network. It is used to adaptively generate attention weights for features of each dimension and obtain weighted features by weighting the input GNSS multidimensional derived features through Hadamard product. The classifier module is used to perform nonlinear mapping and deep feature extraction on the weighted features to obtain deep features; The output layer is used to receive the deep features, activate the deep features through an activation function, and output a reliable metric score corresponding to the baseline coordinate calculation.

[0013] In one possible implementation, the construction of the adaptive weighted fusion loss function for the physical prior constraints includes: The physical constraint loss is constructed by including two types of physical penalty terms: normal matrix condition number constraint and carrier kurtosis anomaly constraint, as shown in the following equation:

[0014] in, For the physical constraint loss, To assess the reliability of model predictions, For high reliability threshold, , For physical penalty weights, The condition number of the normal matrix. For carrier residual kurtosis; The adaptive weight loss is constructed as shown in the following equation:

[0015] in, For a reliable and true metric score of the sample, For the reliability metric score predicted by a fully trained MLP reliability metric model, For adaptive weighting coefficients

[0016] The physical constraint loss With the adaptive weight loss The weighted summation yields the adaptive weighted fusion loss function for the physical prior constraints, as shown in the following equation:

[0017] in, The adaptive weighted fusion loss function is the physical prior constraints. This is the physical constraint balance coefficient.

[0018] In one possible implementation, the state noise of the Kalman filter in the SGM solution is adaptively adjusted based on the reliability metric score to obtain the optimized solution coordinates, as shown in the following equation:

[0019] in, For the optimized state noise, As the reference state noise, and This is used as an adjustment coefficient to control the sensitivity of observation noise covariance scaling. For the aforementioned reliability metric score, The current solution coordinates, These are the predicted values ​​of the prior state coordinates. To predict residual bias, This is the average value of the predicted residual time series vector.

[0020] Secondly, the present invention also provides a GNSS deformation monitoring coordinate reliability measurement, evaluation, and optimization system, comprising: GNSS reference station, used to collect GNSS data from the reference station; GNSS monitoring station, used to collect GNSS data from the monitoring station; The memory is used to store the program, the GNSS data of the base station, and the GNSS data of the monitoring station; A processor, coupled to the memory, is used to execute a program stored in the memory to implement the steps in the GNSS deformation monitoring coordinate reliability measurement evaluation and optimization method described in any of the above implementations. The display, coupled to the processor, is used to output the reliability quantification score corresponding to the SGM solution result and the optimized monitoring station coordinates.

[0021] The beneficial effects of this invention are as follows: The GNSS deformation monitoring coordinate reliability quantification assessment and optimization method provided by this invention firstly extracts multi-dimensional derived features from the monitoring station and base station data through SGM calculation, and after determining the overall absolute deviation based on the calculated coordinates, labels each calculated sample as normal or abnormal using the quartile threshold method, transforming the qualitative and ambiguous reliability judgment into a quantifiable label, laying a data foundation for subsequently capturing the complex nonlinear correlation between multi-dimensional indicators and calculation reliability; secondly, it adopts station-level robust normalization to eliminate the interference of cross-site distribution offset on the feature space, which can adapt to the environmental heterogeneity under different monitoring scenarios, and at the same time adopts a hierarchical station-level synthetic minority class oversampling strategy to improve the model bias towards the majority class caused by sample imbalance; finally, it uses an adaptive weighted fusion loss function with physical prior constraints to train the MLP reliability based on the target dataset. A quantified model was developed to obtain a fully trained MLP reliability quantification model. The model training bias was suppressed through a loss function, addressing both differences in station size and the disparity in the number of positive and negative samples. Based on this, the multidimensional derived features of GNSS to be evaluated were input into the fully trained model, outputting a score that quantitatively characterizes the reliability of the SGM solution. This ensures that abnormal coordinate deviations and actual displacements exhibit a distinguishable distribution in the score space, achieving a stable and universally applicable quantitative assessment of the reliability of the GNSS solution. Furthermore, the baseline observation noise of the Kalman filter was adaptively adjusted based on the reliability quantification score. The optimized observation noise was used to update the Kalman filter state of the solution coordinates, resulting in optimized monitoring station coordinates. This achieved smooth filtering of coordinate fluctuations during stable phases and accurate response to deformation characteristics during actual deformation, further reducing the risk of false alarms and missed alarms in deformation monitoring. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in 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.

[0023] Figure 1 A schematic flowchart of an embodiment of the GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method provided by the present invention; Figure 2 This is a schematic diagram of an embodiment of the present invention for determining the total absolute deviation of a monitoring station based on calculated coordinates; Figure 3 This is a schematic flowchart of an embodiment of the present invention that divides samples corresponding to a baseline into normal samples and abnormal samples; Figure 4This is a schematic diagram of an embodiment of the process for obtaining a target dataset provided by the present invention; Figure 5 Provided by the present invention Figure 4 A schematic diagram of an embodiment of S401; Figure 6 Provided by the present invention Figure 4 A schematic diagram of an embodiment of S402; Figure 7 A schematic diagram of an embodiment of the MLP reliable metric model provided by the present invention; Figure 8 Provided by the present invention Figure 7 A schematic diagram of an embodiment of the classifier module; Figure 9 This is a schematic flowchart of another embodiment of the GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method provided by the present invention; Figure 10 This is a schematic diagram of an embodiment of the GNSS deformation monitoring coordinate reliability quantification evaluation and optimization system provided by the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0026] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] Before demonstrating the embodiments, the following terms will be explained.

[0029] GNSS: Global Navigation Satellite System, refers to a radio navigation and positioning system that uses artificial satellites to achieve global coverage, including the US GPS, China's BeiDou (BDS), Russia's GLONASS, and the EU's Galileo.

[0030] Baseline: In GNSS surveying, the baseline refers to the spatial line connecting two GNSS receiver stations, which includes two elements: length and direction, and is usually represented by the difference in three-dimensional coordinates.

[0031] Reference station: In GNSS deformation monitoring, the reference station is a GNSS receiving station deployed in a geologically stable and long-term unchanging area. Its coordinates serve as a known fixed reference point for differential calculation with the monitoring station.

[0032] Monitoring station: In GNSS deformation monitoring, the monitoring station is a GNSS receiving station deployed on the target to be measured (such as landslides, dams, and bridges). Its coordinates may change over time and it is the direct observation object of deformation monitoring.

[0033] Deformation monitoring refers to the process of repeatedly observing the spatial changes of the ground surface or engineering structure over time and obtaining deformation information by analyzing the differences in the observation results.

[0034] SMOTE: Synthetic Minority Over-sampling Technique, is a classic method for dealing with the problem of imbalanced data classes. It generates new synthetic samples by interpolating between existing minority class samples to balance the class distribution in the dataset.

[0035] This invention provides a method and system for quantitative evaluation and optimization of coordinate reliability for GNSS deformation monitoring, which will be described below.

[0036] Figure 1 This is a schematic flowchart of an embodiment of the GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method provided by the present invention, as shown below. Figure 1As shown, the quantitative evaluation and optimization methods for the reliability measurement of GNSS deformation monitoring coordinates include: S101. Based on GNSS monitoring station data and GNSS reference station data, perform SGM calculation to obtain GNSS multidimensional derived features and the calculated coordinates of the monitoring stations; determine the total absolute deviation of the monitoring stations based on the calculated coordinates; S102. Based on the quartile threshold method, the samples corresponding to each SGM solution are divided into normal samples and abnormal samples according to the total absolute deviation. The original dataset is constructed based on the normal samples, abnormal samples and GNSS multidimensional derived features. The original dataset is then subjected to site-level robust normalization and hierarchical site-level sample balancing to obtain the target dataset. S103. Using an adaptive weighted fusion loss function with physical prior constraints, train the MLP reliability metric model based on the target dataset to obtain a fully trained MLP reliability metric model; input the GNSS multidimensional derived features corresponding to the baseline to be tested into the fully trained MLP reliability metric model to obtain the reliability metric score of the corresponding SGM solution result. S104. Based on the reliability metric scoring, the baseline observation noise of the Kalman filter in the SGM solution is adaptively adjusted to obtain the optimized observation noise; based on the optimized observation noise, the solution coordinates are updated by Kalman filtering to obtain the optimized monitoring station coordinates.

[0037] It should be noted that "each SGM solution" refers to a complete baseline solution and parameter estimation process performed on the GNSS data collected by the monitoring station and the base station within a solution time window. The output of each solution time window corresponds to an independent sample. A "sample" refers to a data record unit corresponding to each SGM solution, which specifically includes two parts: first, the GNSS multidimensional derived features output by this solution, and second, the label corresponding to this solution (i.e., normal or abnormal based on the total absolute deviation using the quartile threshold method).

[0038] The GNSS deformation monitoring coordinate reliability quantification assessment and optimization method provided in this invention first extracts multi-dimensional derived features from monitoring station and base station data through SGM calculation. Based on the calculated coordinates, after determining the overall absolute deviation, the quartile threshold method is used to label each calculated sample as normal or abnormal, transforming the qualitatively ambiguous reliability judgment into a quantifiable label. This lays the data foundation for subsequently capturing the complex nonlinear correlation between multi-dimensional indicators and calculation reliability. Second, station-level robust normalization is used to eliminate the interference of cross-site distribution offset on the feature space, enabling adaptation to environmental heterogeneity under different monitoring scenarios. Simultaneously, a hierarchical station-level synthetic minority class oversampling strategy is employed to improve the model bias towards the majority class caused by sample imbalance. Finally, an adaptive weighted fusion loss function with physical prior constraints is used to train an MLP reliability quantification model based on the target dataset. A fully trained MLP reliability metric model was obtained. The model training bias was suppressed from two dimensions: the difference in station size and the disparity in the number of positive and negative samples, through the loss function. On this basis, the multidimensional derived features of GNSS to be evaluated were input into the fully trained model, and a score that can quantitatively characterize the reliability of SGM solution results was output. This made the abnormal coordinate deviation and the true displacement present a distinguishable distribution in the score space, realizing a stable and universal quantitative evaluation of the reliability of GNSS solution results. Furthermore, the baseline observation noise of the Kalman filter was adaptively adjusted based on the reliability metric score. The optimized observation noise was used to update the Kalman filter state of the solution coordinates, resulting in optimized monitoring station coordinates. This achieved smooth filtering of coordinate fluctuations in the stable phase and accurate response to deformation characteristics when real deformation occurs, further reducing the risk of false alarms and missed alarms in deformation monitoring.

[0039] In some embodiments of the present invention, the algorithm parameter configuration of the SGM solution module is shown in Table 1.

[0040] Table 1: SGM Solver Parameter Configuration

[0041] The following is the meaning of each configuration item in Table 1: GNSS frequency: Specifies the satellite system and signal frequency used in the calculation. Dual-frequency observations can construct an ionospherically independent combination, eliminating first-order ionospheric delay errors.

[0042] Solution mode: Specifies the data processing method. Static mode requires the receiver to remain stationary during the observation period, obtaining the highest accuracy positioning results through joint processing of long-term continuous observation data.

[0043] Data processing filtering method: Specifies the filtering strategy for parameter estimation. Forward Kalman Filter recursively estimates the state using only current and historical epoch observation data, and is suitable for real-time or near-real-time processing scenarios.

[0044] Cutoff elevation angle: An angle threshold set to shield against obstructions and multipath effects; satellite signals below this angle are not processed. 15° can effectively eliminate low-elevation-angle satellite observations that are severely affected by atmospheric delay and noise.

[0045] Ionospheric error correction model: used to calculate and eliminate delay errors caused by signals passing through the ionosphere. The Klobuchar model estimates ionospheric delay using coefficients broadcast ephemeris data, offering high computational efficiency.

[0046] Tropospheric error correction model: Used to calculate and eliminate delay errors caused by signals passing through the troposphere. The Saastamoinen model estimates zenith tropospheric delay using meteorological parameters from the station and is a classic model in high-precision GNSS data processing.

[0047] Satellite coordinates: Specifies the ephemeris type used to calculate the satellite's position. Broadcast ephemeris refers to the predicted orbital parameters broadcast in real time by the satellite via navigation messages, which can be obtained in real time.

[0048] Integer ambiguity resolution: A fixed strategy for specifying the integer ambiguity of the carrier phase. Fix-and-hold uses the fixed integer ambiguity from the previous epoch as a virtual observation in the current filtering, maintaining the continuity of fixed ambiguity.

[0049] Solution time window: Specifies the length of observation data used in a single solution. A 2-hour observation window can accumulate enough satellite observation data to improve the success rate of ambiguity fixation and positioning accuracy.

[0050] Solution step size: Specifies the time interval between two adjacent solutions. In this configuration, the time window and step size are both 2 hours, meaning that a set of solution results is output every 2 hours, and adjacent windows do not overlap.

[0051] In some embodiments of the present invention, GNSS multidimensional derived features include multidimensional raw data quality features and multidimensional solution process features; Multidimensional raw data quality characteristics include signal-to-noise ratio correlation characteristics, cycle slip ratio characteristics, data integrity rate characteristics, pseudorange multipath characteristics, and noise characteristics; The multidimensional solution process features include single-point positioning accuracy features, unfixed solution ratio features, variance anomaly features, ambiguity variance ratio statistical features, unit weight mean square error statistical features, accuracy factor features, pre-verification residual features, and post-verification residual features.

[0052] In some embodiments of the present invention, the present invention extracts 88-dimensional GNSS derived features from two dimensions: the quality of the original data and the robustness of the solution process. The specific feature composition is shown in Table 2.

[0053] Table 2: Multidimensional Derived Features of GNSS

[0054] This embodiment constructs a total of 88-dimensional GNSS multidimensional derived features, consisting of 16-dimensional original data quality features and 72-dimensional solution process features. It fully covers two key dimensions: the original GNSS observation quality and the entire solution process. It can comprehensively extract various influencing factors that affect the reliability of GNSS deformation monitoring coordinates, and provide complete multidimensional feature inputs for subsequent preprocessing processes and feature attention-enhanced MLP models, effectively improving the recognition accuracy of coordinate reliability quantification assessment.

[0055] In some embodiments of the present invention, the total absolute deviation of the monitoring station is determined based on the calculated coordinates, such as... Figure 2 As shown, it includes: S201. Obtain the solution coordinates of the monitoring stations for SGM solution; S202, Determine the solution coordinates. , , The deviation values ​​relative to the coordinates of the reference station in three dimensions are used to determine the Euclidean distance based on the deviation values, thus obtaining the overall absolute deviation of the monitoring station.

[0056] In some embodiments of the present invention, the coordinates of each baseline are calculated. , , The respective deviations in the three-dimensional directions ( , , The total deviation can then be calculated using the Euclidean distance formula. As shown in Formula 1: (1) This embodiment calculates the Euclidean distance using three-dimensional coordinate deviation to obtain the overall absolute deviation of the monitoring station. This can comprehensively characterize the degree of overall coordinate offset in three dimensions, avoiding the one-sidedness of single-dimensional deviation judgment. It provides a comprehensive and unified quantitative judgment index for classifying normal and abnormal samples using the quartile threshold method, ensuring the accuracy and reliability of sample labeling results.

[0057] In some embodiments of the present invention, the baseline-corresponding samples are divided into normal samples and abnormal samples based on the quartile threshold method according to the total absolute deviation, such as... Figure 3 As shown, it includes: S301. Determine the 25th quantile Q1, 75th quantile Q3, and interquartile range IQR = Q3 - Q1 for the total absolute deviation sequence, and set the outlier threshold interval as [Q1 - 1.5IQR, Q3 + 1.5IQR]; S302. Mark the samples corresponding to the SGM solution results where the total absolute deviation exceeds the abnormal threshold range as abnormal samples, otherwise mark them as normal samples.

[0058] In some embodiments of the present invention, the quartile method is used to analyze each GNSS baseline. Perform outlier detection and calculation The 25th percentile (Q1), 75th percentile (Q3), and interquartile range (IQR = Q3 - Q1) of the sequence are used to set the anomaly threshold range as [Q1 - 1.5IQR, Q3 + 1.5IQR]. If the BIAS exceeds this range, it is marked as an abnormal deviation sample (LABEL = 1); otherwise, it is a normal displacement sample (LABEL = 0). Finally, a complete dataset corresponding to "88-dimensional GNSS derived features-labels" is constructed to provide a foundation for subsequent model training.

[0059] This embodiment uses the quartile threshold method to divide samples. It adaptively generates anomaly thresholds based on the overall absolute deviation distribution of the samples themselves, eliminating the need to manually set fixed discrimination thresholds and removing subjective bias caused by manual thresholds. This method constructs threshold intervals based on quartile statistics that resist extreme value interference, which can accurately distinguish between normal and abnormal samples and ensure that the sample labeling results are objective and accurate.

[0060] It should be noted that, unlike conventional neural network modeling datasets, GNSS-derived feature datasets exhibit three key characteristics: First, significant heterogeneity across multiple sites, with differences in equipment configuration and observation environment across different monitoring scenarios leading to a marked shift in feature distribution; second, redundant correlations exist in high-dimensional features, with some similar indicators (such as the 28 features corresponding to the pre-existing residuals) showing strong correlations; and third, an imbalance in the number of positive and negative samples, with abnormal deviation samples typically being an order of magnitude fewer than normal samples, easily causing the model to favor majority class learning. To address these data characteristics, this invention designs a targeted data preprocessing workflow and a lightweight model architecture.

[0061] This paper addresses the characteristics of the aforementioned data samples and designs a data preprocessing approach with the goals of eliminating sample distribution bias, filtering effective data, and balancing sample distribution, providing standardized and highly reliable input for model training.

[0062] In some embodiments of the present invention, the target dataset is obtained by performing site-level robust normalization and hierarchical site-level sample equalization on the original dataset, such as... Figure 4 As shown, it includes: S401. Apply site-level robust normalization to all GNSS multidimensional derived features of the original dataset to generate a transitional dataset. S402. Perform hierarchical site-level synthetic minority class oversampling on the transition dataset to achieve sample balancing and obtain the target dataset.

[0063] This embodiment employs site-level robust normalization and hierarchical site-level oversampling sequentially for preprocessing. Site-level robust normalization standardizes each site independently, avoiding cross-site distribution shifts and resisting extreme value interference. Hierarchical site-level oversampling balances samples only within a single site, preventing cross-site mixing from disrupting the inherent distribution of the site. The two methods work together to solve the problems of feature shift and class imbalance in multi-site heterogeneous scenarios, providing the model with high-quality input data that is suitable for the model.

[0064] It should be noted that before performing site-level robust normalization and hierarchical site-level sample equalization on the original dataset, the original data is first cleaned. Dataset cleaning aims to perform dual verification of the completeness and validity of the original dataset.

[0065] Data cleaning includes: removing samples with invalid or null values ​​in each dimension of features to ensure that the feature vectors correspond one-to-one with the labels; verifying data consistency through timestamps and baseline identifiers to remove invalid records with time mismatches or failed baseline calculations, thus avoiding interference from invalid data on model training; and retaining all labeled abnormal deviation samples and normal displacement samples to ensure that the dataset fully covers the data distribution characteristics in real monitoring scenarios.

[0066] In some embodiments of the present invention, site-level robust normalization is applied to all GNSS multidimensional derived features of the original dataset to generate a transitional dataset, such as... Figure 5 As shown, it includes: S501. Divide the original dataset into independent site subsets according to the monitoring site identifier; S502. For each GNSS derived feature in each station subset, determine the median and interquartile range of the corresponding feature. If the interquartile range of a feature in a certain dimension of a station is equal to 0, then the normalization value of that feature is set to 0. If the interquartile range of a feature in a certain dimension of a station is greater than 0, then the feature is standardized based on the median and interquartile range of the corresponding station. S503. After standardizing the GNSS-derived features of all stations, a normalized transition dataset is obtained.

[0067] To address the feature distribution shift caused by multi-site heterogeneity, this paper abandons the traditional global normalization method and adopts a site-independent robust normalization strategy. The dataset is divided into independent subsets based on site identifiers, and the feature statistics for each dimension of each site are calculated separately to avoid cross-site distribution interference. The median and interquartile range (IQR) are selected as robust statistics to replace the mean and standard deviation, which are susceptible to extreme values, thus enhancing the preprocessing's robustness against interference. The first site 3D features ( Standardization is performed according to Formula 2 to ensure that the characteristic value ranges are uniform and the inherent distribution characteristics within the site are preserved.

[0068] (2) in, These are the standardized feature values ​​after site-level robust normalization. For the first The first site The original observed values ​​of the dimensional features, For the first The first result calculated from all samples within each site The median of the dimensional feature is a robust statistic that is not affected by extreme outliers. For the first The first site The upper quartile Q3 of the dimension feature is subtracted from the lower quartile Q1.

[0069] This embodiment divides the site into independent subsets and calculates the median and interquartile range for each site and feature to achieve standardization, which can eliminate the problem of feature distribution offset between multiple sites. It relies on robust statistics to weaken the interference of extreme values, and adds a fallback logic of setting the interquartile range to 0 to avoid calculation anomalies, unifying the feature value range and fully preserving the inherent distribution characteristics of each site.

[0070] In some embodiments of the present invention, hierarchical site-level synthetic minority class oversampling is performed on the transition dataset to achieve sample balancing and obtain the target dataset, such as... Figure 6 As shown, it includes: S601. Construct a site feature statistical vector with a single site as a unit, and determine the distribution similarity between sites through cosine similarity. Sites with similarity greater than or equal to a preset similarity threshold are identified as similar sites. S602. For stations with no less than 5 abnormal samples, generate synthetic samples based on the abnormal samples of this station through interpolation; for stations with less than 5 abnormal samples, merge the abnormal samples of all similar stations of this station, and select a preset number of neighboring samples to combine with Gaussian noise to generate synthetic samples that do not overlap with the original samples. S603. After generating synthetic samples for each site, limit the total number of abnormal samples to a preset proportion of the number of normal samples at that site to obtain the target dataset.

[0071] To address the imbalance issue of a low proportion of outlier samples, a hierarchical site-level sample augmentation strategy is designed to avoid cross-site distribution confusion inherent in traditional SMOTE. Using a single site as the basic processing unit, a site feature statistical vector is defined, as shown in Equation 3: (3) The cosine similarity measure is used to quantify the distribution similarity between stations, as shown in Formula 4: (4) in, , Sites and eigenstatistical vectors, Represents the vector dot product. This represents the L2 norm. When... When two sites are identified as similar sites, they are determined to be similar sites.

[0072] Number of abnormal samples The site is directly based on the site's internal abnormal sample set. Perform synthesis; for The website merges the abnormal sample sets of all its similar websites to build an expanded abnormal sample pool. ( (Similar sites). For target anomaly samples Randomly select its K nearest neighbor samples Generate a synthetic sample according to Formula 5: (5) In the formula, These are random weighting coefficients; For a single nearest neighbor sample selected randomly, Gaussian noise is used to avoid complete overlap between the synthesized sample and the original sample.

[0073] To balance model bias and overfitting risk, the total number of outlier samples after synthesis at each site is controlled to be within 30% of the number of normal samples at that site. ( For the site The number of synthetic anomalous samples, (This represents the number of normal samples at the site). The final balanced training set (target dataset) provides sufficient data support for the model to learn the features of abnormal samples.

[0074] This embodiment employs a hierarchical site-level oversampling strategy, relying on cosine similarity to perform hierarchical clustering of sites to classify similar sites. Synthetic samples are generated differently for two types of hierarchical scenarios: sufficient and scarce abnormal samples within a site. For sites with scarce samples, abnormal samples from similar sites in the same hierarchical layer can be retrieved to supplement the interpolation basis. At the same time, the maximum generation ratio of synthetic samples for each site is constrained. Hierarchical sampling can effectively balance the positive and negative samples of each site, avoid cross-hierarchical mixed sampling from destroying the original data distribution of the sites, suppress oversampling distortion, and adapt to heterogeneous GNSS datasets of multiple sites.

[0075] In some embodiments of this invention, to address the problems of high-dimensional feature redundancy and uneven contribution of key features, this invention designs a Feature Attention-Augmented MLP as a reliable metric model. This model strengthens the representation of key information and suppresses redundant interference through adaptive feature weight allocation. The overall model architecture is as follows: Figure 7 As shown.

[0076] The input to the MLP reliable metric model is the GNSS multidimensional derived features after site-level robust normalization, and the output is the SGM solution coordinates of the corresponding baseline. Interval reliability metric score; like Figure 7 As shown, the reliable metric model of MLP includes a feature attention layer set in the input layer and the classifier module, and a classifier module and an output layer connected in series in the feature attention layer; The feature attention layer is an attention weighting module added to the front end of the network and built into a two-layer perceptron sub-network. It is used to adaptively generate attention weights for features of each dimension. The weighted features are obtained by weighting the input GNSS multidimensional derived features through Hadamard product. The classifier module is used to perform nonlinear mapping and deep feature extraction on the weighted features to obtain deep features; The output layer receives deep features, activates these features using an activation function, and outputs a reliable metric score corresponding to the baseline coordinates.

[0077] It should be noted that the original MLP network consists of an input layer, a hidden layer, and an output layer connected in series, and the classifier module of this invention is the hidden layer. This invention introduces a feature attention layer containing a perceptron subnet after the input layer to strengthen key features that affect the reliability of baseline solution and filter redundant noise features.

[0078] This embodiment adds a feature attention layer with two built-in perceptron subnets between the input layer and the classifier module. This layer can adaptively assign attention weights to multi-dimensional GNSS-derived features, strengthen effective features, and compress feature dimensions to reduce computational overhead. The lightweight classifier module, consisting of only two hidden layers, completes the nonlinear deep extraction of weighted features. Finally, the output layer outputs an intuitive baseline solution reliability quantification score, which reduces the number of model parameters, enables real-time fast inference, filters data noise, and improves the accuracy of reliability assessment results in multi-site scenarios.

[0079] In some embodiments of the present invention, the feature attention layer serves as the core module of the model, aiming to adaptively mine key features that are strongly correlated with the reliability of the solution and suppress interference from redundant information. A lightweight two-layer perceptron is used to construct the attention network, mathematically expressed as shown in Equation 6: (6) In the formula, , The attention network weight matrix is ​​initialized using Xavier. , This represents the bias term, initialized to 0; Use the Sigmoid function to ensure the output attention weights. In the experiment, the Dropout probability was set to 0.2 to suppress overfitting of the attention network. Then, the original features were combined using the Hadamard product. With attention weight The fusion yields a weighted feature vector. This enables the enhancement of key features and the suppression of redundant features.

[0080] In some embodiments of the present invention, a schematic diagram of the classifier module structure is shown below. Figure 8 As shown, the classifier module is responsible for mapping weighted features to a reliable quantitative score for the SGM solution. It includes a feature compression layer, a nonlinear mapping intermediate layer, and a probability output layer. The feature compression layer uses a linear transformation structure to compress the attention-weighted 88-dimensional features to 32 dimensions, reducing feature dimensionality and simplifying model complexity while retaining key information and reducing redundant computation. The intermediate layer uses a nonlinear mapping design to further map the 32-dimensional compressed features to 16 dimensions, enabling in-depth mining of nonlinear correlation information between features and improving the model's ability to fit complex data distributions. The output layer uses a single-dimensional probability mapping design to map the 16-dimensional deep interaction features to a single probability value in the range of 0 to 1, achieving a quantitative score for the reliability of the SGM solution. The closer the score is to 1, the more reliable the solution; the closer it is to 0, the more likely the solution is an abnormal deviation, thus quantitatively distinguishing between abnormal deviations and true displacements.

[0081] In some embodiments of the present invention, the construction of the adaptive weighted fusion loss function of physical prior constraints includes: The physical constraint loss is constructed by including two types of physical penalty terms: normal matrix condition number constraint and carrier kurtosis anomaly constraint, as shown in the following equation:

[0082] in, For physical constraint loss, To assess the reliability of model predictions, For high reliability threshold, , For physical penalty weights, The condition number of the normal matrix. For carrier residual kurtosis; The adaptive weight loss is constructed as shown in Equation 7 below: (7) in, For a reliable and true metric score of the sample, For the reliability metric score predicted by a fully trained MLP reliability metric model, For adaptive weighting coefficients

[0083] Physical constraint loss With adaptive weight loss Weighted summation yields the adaptive weighted fusion loss function for physical prior constraints, as shown in Equation 8 below: (8) in, The adaptive weighted fusion loss function is defined by physical prior constraints. This is the physical constraint balance coefficient.

[0084] This embodiment achieves synergistic constraints between physical mechanisms and data-driven approaches by constructing an adaptive weighted fusion loss function based on physical prior constraints. When the model predicts an artificially high reliability score in ill-conditioned or observationally anomalous scenarios, the activation mechanisms in the two types of physical penalty terms are triggered, significantly increasing the physical constraint loss to correct model parameters in reverse, suppressing prediction biases that violate the laws of GNSS physical errors, and ensuring that the score closely matches the actual error variation patterns. Simultaneously, the adaptive weighted loss guides the model to focus on difficult samples at the reliability boundary, alleviating the sample imbalance problem. The weighted fusion of the two losses makes the reliability score both physically reasonable and scenario-adaptable.

[0085] In some embodiments of the present invention, the state noise of the Kalman filter in the SGM solution is adaptively adjusted based on the reliability metric score to obtain the optimized solution coordinates, as shown in Equation 9 below: (9) in, For the optimized state noise, As the reference state noise, and This is used as an adjustment coefficient to control the sensitivity of observation noise covariance scaling. For reliable quantification of scoring, The current solution coordinates, These are the predicted values ​​of the prior state coordinates. To predict residual bias, This is the average value of the predicted residual time series vector.

[0086] This embodiment constructs an adaptive adjustment mechanism for state noise by fusing reliability scores and prediction residual bias. The reliability score item... Responsible for dynamically adjusting the state noise baseline based on the current reliability of the solution and predicting residual correction terms. This mechanism is responsible for secondary correction of state noise based on the deviation between current observations and historical predictions, and the two work together to achieve adaptive optimization of state noise. This mechanism takes into account both data fitting accuracy and the rationality of physical mechanisms, effectively solving the problems of discrete predictions and missing physical mechanisms in pure data-driven models. It achieves adaptive filtering that ensures sufficient smoothness of coordinate sequences in the stationary phase and rapid response in the deformation phase, further reducing the risk of false alarms and missed alarms in deformation monitoring.

[0087] In some embodiments of the present invention Figure 9 This is a schematic diagram of another embodiment of the GNSS deformation monitoring coordinate reliability quantification assessment and optimization method provided by the present invention. The method is divided into three parts: data generation and annotation, data preprocessing, and model evaluation output. First, GNSS monitoring station data and GNSS reference station data are input into the SGM (Site-Generated Model) to obtain GNSS multidimensional derived data, thus constructing the original dataset. Simultaneously, the total absolute deviation is calculated, and the quartile threshold method is used to classify normal and abnormal samples. Then, data cleaning, station-level robust normalization, and hierarchical station-level SMOTE processing are sequentially performed on the original dataset to eliminate multi-site distribution offsets and balance the number of samples of each type. The processed data is then fed into a feature attention-enhanced MLP (Multi-Level Processing) model, passed through the input layer to the feature attention layer to strengthen effective features, and then the classifier module performs discrimination. The model is optimized by combining an adaptive weighted fusion loss function with physical prior constraints, and finally, the GNSS deformation monitoring coordinate reliability quantification assessment result is output.

[0088] Accordingly, embodiments of the present invention also provide a GNSS deformation monitoring coordinate reliability quantification evaluation and optimization system 1000, such as... Figure 10 As shown, it includes: GNSS reference station 1001 is used to collect GNSS data from the reference station. GNSS monitoring station 1002 is used to collect GNSS data from the monitoring station; Memory 1003 is used to store programs, base station GNSS data, and monitoring station GNSS data; The processor 1004, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method in the above embodiments. Display 1005, coupled to the processor, is used to output the reliability quantification score corresponding to the SGM solution results.

[0089] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.), and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0090] The above provides a detailed description of the GNSS deformation monitoring coordinate reliability quantitative evaluation and optimization method and system provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for quantitative evaluation and optimization of coordinate reliability measurement in GNSS deformation monitoring, characterized in that, include: SGM calculation is performed based on GNSS monitoring station data and GNSS reference station data to obtain GNSS multidimensional derived features and the calculated coordinates of the monitoring stations; The overall absolute deviation of the monitoring stations is determined based on the calculated coordinates; Based on the quartile threshold method, the samples corresponding to each SGM solution are divided into normal samples and abnormal samples according to the total absolute deviation. The original dataset is constructed based on the normal samples, the abnormal samples, and the GNSS multidimensional derived features. The target dataset is obtained by performing site-level robust normalization and hierarchical site-level sample equalization on the original dataset. An adaptive weighted fusion loss function with physical prior constraints is used to train an MLP reliability metric model based on the target dataset, resulting in a fully trained MLP reliability metric model. The GNSS multidimensional derived features corresponding to the baseline to be tested are input into the fully trained MLP reliability metric model to obtain the reliability metric score of the corresponding SGM solution. Based on the reliability metric score, the baseline observation noise of the Kalman filter in the SGM solution is adaptively adjusted to obtain the optimized observation noise; based on the optimized observation noise, the solution coordinates are updated using Kalman filtering to obtain the optimized monitoring station coordinates.

2. The GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method according to claim 1, characterized in that, The GNSS multidimensional derived features include multidimensional raw data quality features and multidimensional solution process features; The multidimensional raw data quality features include signal-to-noise ratio correlation features, cycle slip ratio features, data integrity rate features, pseudorange multipath features, and noise features; The multidimensional solution process features include single-point positioning accuracy features, unfixed solution ratio features, variance anomaly features, ambiguity variance ratio statistical features, unit weight mean square error statistical features, accuracy factor features, pre-verification residual features, and post-verification residual features.

3. The GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method according to claim 1, characterized in that, The determination of the overall absolute deviation of the monitoring stations based on the calculated coordinates includes: Obtain the calculated coordinates of the monitoring stations for the SGM solution; Determine the solution coordinates in , , The deviation values ​​relative to the coordinates of the reference station in three dimensions are used to determine the Euclidean distance based on the deviation values, thus obtaining the overall absolute deviation of the monitoring station.

4. The GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method according to claim 1, characterized in that, The process of obtaining the target dataset by performing site-level robust normalization and hierarchical site-level sample equalization on the original dataset includes: All the GNSS multidimensional derived features are processed using site-level robust normalization on the original dataset to generate a transitional dataset; The transition dataset is subjected to hierarchical site-level synthetic minority class oversampling to achieve sample balancing and obtain the target dataset.

5. The GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method according to claim 4, characterized in that, The original dataset undergoes site-level robust normalization processing to generate a transitional dataset, comprising: The original dataset is divided into independent site subsets based on the monitoring site identifiers; For each GNSS derived feature in each station subset, the median and interquartile range of the corresponding feature are determined respectively; if the interquartile range of a certain feature of a certain station is equal to 0, the normalization value of that feature is set to 0; if the interquartile range of a certain feature of a certain station is greater than 0, the feature is standardized based on the median and interquartile range of the corresponding station. After standardizing the GNSS-derived features of all stations, a normalized transition dataset is obtained.

6. The GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method according to claim 4, characterized in that, The process of performing hierarchical site-level synthetic minority class oversampling on the transition dataset to achieve sample balancing and obtain the target dataset includes: A site feature statistical vector is constructed using a single site as a unit, and the distribution similarity between sites is determined by cosine similarity. Sites with similarity greater than or equal to a preset similarity threshold are identified as similar sites. For stations with no fewer than 5 anomalous samples, synthetic samples are generated by interpolation based on the anomalous samples of the station. For stations with fewer than 5 anomalous samples, anomalous samples from all similar stations of the station are merged, and a preset number of neighboring samples are selected and combined with Gaussian noise to generate synthetic samples that do not overlap with the original samples. After generating synthetic samples for each site, the total number of its abnormal samples is limited to a preset proportion of the number of normal samples at that site to obtain the target dataset.

7. The GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method according to claim 1, characterized in that, The input to the MLP reliability metric model is the GNSS multidimensional derived features after site-level robust normalization, and the output is the SGM solution coordinates of the corresponding baseline. Interval reliability metric score; the MLP reliability metric model includes a feature attention layer set between the input layer and the classifier module, and a classifier module and an output layer connected in series with the feature attention layer; The feature attention layer is an attention weighting module added to the front end of the network and built into a two-layer perceptron sub-network. It is used to adaptively generate attention weights for features of each dimension and obtain weighted features by weighting the input GNSS multidimensional derived features through Hadamard product. The classifier module is used to perform nonlinear mapping and deep feature extraction on the weighted features to obtain deep features; The output layer is used to receive the deep features, activate the deep features through an activation function, and output a reliable metric score corresponding to the baseline coordinate calculation.

8. The GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method according to claim 1, characterized in that, The construction of the adaptive weighted fusion loss function for the physical prior constraints includes: The physical constraint loss is constructed by including two types of physical penalty terms: normal matrix condition number constraint and carrier kurtosis anomaly constraint, as shown in the following equation: in, For the physical constraint loss, To assess the reliability of model predictions, For high reliability threshold, , For physical penalty weights, The condition number of the normal matrix. For carrier residual kurtosis; The adaptive weight loss is constructed as shown in the following equation: in, For a reliable and accurate metric score of the sample, For the reliability metric score predicted by a fully trained MLP reliability metric model, These are adaptive weighting coefficients; The physical constraint loss With the adaptive weight loss The weighted summation yields the adaptive weighted fusion loss function for the physical prior constraints, as shown in the following equation: in, The adaptive weighted fusion loss function is the physical prior constraints. This is the physical constraint balance coefficient.

9. The GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method according to claim 1, characterized in that, The optimized solution coordinates are obtained by adaptively adjusting the state noise of the Kalman filter in the SGM solution based on the reliability metric score, as shown in the following formula: in, For the optimized state noise, As the reference state noise, and This is used as an adjustment coefficient to control the sensitivity of observation noise covariance scaling. For the aforementioned reliability metric score, The current solution coordinates, These are the predicted values ​​of the prior state coordinates. To predict residual bias, This is the average value of the predicted residual time series vector.

10. A GNSS deformation monitoring coordinate reliability measurement, evaluation, and optimization system, characterized in that, include: GNSS reference station, used to collect GNSS data from the reference station; GNSS monitoring station, used to collect GNSS data from the monitoring station; The memory is used to store the program, the GNSS data of the base station, and the GNSS data of the monitoring station; A processor, coupled to the memory, is configured to execute a program stored in the memory to implement the steps in the GNSS deformation monitoring coordinate reliability quantification evaluation and optimization method according to any one of claims 1 to 9; The display, coupled to the processor, is used to output the reliability quantification score corresponding to the SGM solution result and the optimized monitoring station coordinates.

Images