A Quasi-Verification Method for Gross Errors in GNSS Observations Based on Random Forest

Abstract

This invention discloses a method for calibrating gross errors in GNSS observations based on random forest. The method includes: collecting and extracting dynamic GNSS observation data and its key feature vectors to construct a training set for a random forest classifier; dividing the observation data into "calibrated observations" and "non-calibrated observations," and training a random forest classifier based on this classification; applying the random forest classifier to a real-time dynamic positioning system, obtaining a "calibration score" for each observation based on the real-time dynamic GNSS observation data; if the "calibration score" of an observation reaches a threshold, it is classified as a "calibrated observation"; if the number of "calibrated observations" meets the requirements, the true error of all observations is calculated; if the number does not meet the requirements, a portion of the "calibrated observations" are selected based on the least squares residuals until the number meets the requirements, and the true error of all observations is calculated; the standardized true error is used to replace the traditional least squares residuals for IGG3 robustness processing, and the robust positioning coordinates are output. This invention improves the accuracy and stability of real-time dynamic positioning systems in urban environments.

Description

Technical Field

[0001] This invention belongs to the field of positioning and navigation technology, specifically relating to a method for calibrating gross errors in GNSS observations based on random forest. Background Technology

[0002] Satellite navigation and positioning technology has become a key tool for acquiring, processing, and analyzing spatial information in modern society. Its importance is widely recognized and applied in various fields such as vehicle navigation, mobile phone positioning, land surveying, and drone flight. The development of Global Navigation Satellite Systems (GNSS) can be traced back to the late 1970s, with its core objective being to achieve global positioning, navigation, and time services through satellite positioning technology. Currently, four major global navigation systems have been established: the US GPS (Global Positioning System), Russia's GLONASS, Europe's Galileo, and China's BeiDou Navigation Satellite System (BDS). Each represents the development achievements of different countries and regions in the field of global navigation satellite systems, and their establishment and continuous improvement provide diverse and highly reliable positioning, navigation, and time services globally.

[0003] The factors affecting satellite navigation positioning accuracy can be mainly summarized into two aspects. First, space environment factors can disrupt satellite signals and reduce signal quality. These factors are closely related to the application environment. Space environment disturbances can be further classified into intentional and unintentional interference. Intentional interference includes the suppression of satellite signals by artificially generated high-intensity electromagnetic signals or false satellite signals, preventing the receiver from obtaining accurate navigation data, thus leading to positioning failure or positioning errors. Unintentional interference involves navigation signal disturbances caused by natural environmental factors, especially in urban environments. Due to factors such as building coverage, tree obstruction, and glass curtain wall reflections, signal quality deteriorates sharply. This includes, but is not limited to, signal phase delay, reduced carrier signal-to-noise ratio, and multipath propagation effects, all of which lead to reduced positioning accuracy. Second, positioning accuracy is directly affected by the user-end hardware configuration and signal processing algorithms. This includes factors such as antenna type, gain, quantization noise, filtering bandwidth, the performance of anti-interference algorithms, and processor computing power. These factors are closely related to the type of user-end equipment and determine the accuracy of the user-end signal processing module's estimation of signal observations at each level, thus directly affecting the quality of positioning performance.

[0004] The rapid pace of global urbanization poses challenges to satellite navigation systems. In complex urban environments, such as urban canyons, tree-lined streets, and elevated highways, satellite signals are easily blocked, resulting in fewer available satellites and low observation redundancy. Furthermore, they are susceptible to multipath effects and indirect signals, leading to a significant non-normal distribution of observation errors and the presence of numerous gross errors. Traditional gross error detection methods typically assume that the number of normal observations far exceeds the number of gross error observations, allowing for the detection and elimination of observations with gross errors using residual testing methods. However, in complex observation environments with fewer satellites, low observation redundancy, and frequent gross errors, conventional methods have significant limitations, easily leading to false positives and false negatives. Therefore, there is an urgent need to develop a method for efficiently and accurately detecting gross errors in observations under complex conditions to meet the high-precision positioning needs of modern urban life and various industries. Summary of the Invention

[0005] To address the aforementioned issues, this invention discloses a method for calibrating gross errors in GNSS observations based on random forests. This method introduces a random forest machine learning model to explore and mine more feature values ​​from dynamic GNSS observation data, and establishes a classifier to identify "calibrated observation values," thereby improving the efficiency of gross error calibration and the accuracy and stability of dynamic positioning in complex environments.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] The method for calibrating gross errors in GNSS observations based on random forests includes the following steps:

[0008] Step 1: Collect and extract dynamic GNSS observation data and its key feature vectors to construct a training set for a random forest classifier;

[0009] Step 2: Divide the observation data into "quasi-observations" and "non-quasi-observations", and train a random forest classifier based on this classification;

[0010] Step 3: Apply the random forest classifier to the real-time dynamic positioning system and obtain the "fit score" for each observation based on the real-time dynamic GNSS observation data.

[0011] Step 4: If the "fit score" of the observation reaches the threshold, it is classified as a "fit observation". If the number of "fit observations" meets the requirements, the true error of all observations is calculated. If the number does not meet the requirements, a portion of the "fit observations" are selected based on the least squares residuals until the number meets the requirements, and the true error of all observations is calculated.

[0012] Step 5: Replace the traditional least squares residual with the standardized true error, perform IGG3 robustness processing, and output the robust positioning coordinates.

[0013] Furthermore, the GNSS observation sample features in the feature vector described in step 1 include:

[0014] The features obtained include elevation angle, signal-to-noise ratio (SNR), SNR attenuation, geometric contribution, signal integrity, dMW (Difference of Melbourne-Wübbena Combination), dGF (Difference of Geometry-Free Combination), carrier consistency, Doppler consistency, single-point positioning residual, LLI (Loss of Lock Indicator) flag, GDOP (Geometric Dilution of Precision), number of satellites, and satellite utilization. Elevation angle, SNR, signal integrity, LLI flag, and number of satellites can be directly obtained from RINEX files. GDOP, number of satellites, and satellite utilization are features shared by all samples in a given epoch. The others require some preprocessing. The feature acquisition method is as follows:

[0015] 1) Signal-to-noise ratio (SNR) attenuation: Using an ideal observation signal in an open environment, fit an elevation angle-SNR template function. The difference between the standard SNR and the actual SNR at the same elevation angle is the SNR attenuation, expressed as:

[0016]

[0017] Where S k The actual signal-to-noise ratio of satellite k. Let Att be the standard signal-to-noise ratio of satellite k at an elevation angle θ. k The signal-to-noise ratio attenuation of satellite k.

[0018] 2) Geometric contribution: Represents the component of the observation in the entire adjustment system, expressed as:

[0019] r k =R NDkk

[0020] Where r k Let r be the geometric contribution of satellite k to the system. k =0, then it indicates that the observation is necessary; if r k =1 indicates that the observation is completely redundant, R NDkk R is the adjustment factor matrix of the unequal pseudorange observation model. ND The element in the k-th row and k-th column. Adjustment factor matrix R. ND The calculation method is as follows:

[0021] R ND =IA ND(A ND T Q llND -1 A ND ) -1 A ND T Q llND -1

[0022] Where I is the identity matrix, A ND Design the matrix Q for the first-order Gauss-Markov model of unequal pseudorange observations. llND -1 These are the corresponding weight matrices. Specifically, their expressions are:

[0023] A ND =[-e x -e y -e z 1]

[0024] Q llND -1 =diag(f ELE (ele))

[0025] Among them, e x e y e z f represents the x, y, and z components of the direction vector from which the receiver points to the satellite, respectively. ELE (ele) represents a stochastic model related to elevation angle, and diag(·) represents the operation of converting a vector into a diagonal matrix.

[0026] 3) dMW and dGF: These are the epoch differences between observations without ionospheric combinations and without geometric combinations. Under conditions where there are no gross errors in the pseudorange and no cycle slips in the carrier wave, these two characteristics should remain within a relatively stable range. The non-difference calculation methods for observations without ionospheric combinations and without geometric combinations are as follows:

[0027]

[0028] GF=λ1L1-λ2L2

[0029] Where MW represents the ionospheric-free combination, GF represents the geometric-free combination, L represents the satellite carrier observation, P represents the satellite pseudorange observation, f represents the carrier frequency, λ represents the carrier wavelength, and the subscript represents the frequency point of the observation.

[0030] 4) Carrier consistency and Doppler consistency: This refers to the consistency between the carrier wave and pseudorange at a certain satellite frequency, and the consistency between Doppler and pseudorange. Under normal circumstances, these two characteristics should not vary significantly. Their calculation methods are as follows:

[0031] LP=P(t)-P(t-1)-λL(t)+λL(t-1)

[0032]

[0033] Where LP stands for carrier consistency, DP stands for Doppler consistency, D stands for Doppler observation, t and t-1 represent two adjacent epochs, and dt is their time difference.

[0034] 5) Single-point location residuals: The residuals obtained by substituting the unequal pseudorange observations into the first-order Gauss-Markov model and applying the least squares method are the single-point location residuals. Specifically, the expression for solving the least squares observation residuals is:

[0035] V ND =-Rl ND

[0036] Among them, V ND Let l be the least squares residual vector of the non-difference pseudorange observation model. ND The corresponding unequal pseudorange observation can be specifically represented as:

[0037] l ND =P-ρ+v ion +v trop

[0038] Where P is the pseudorange observation value of the satellite, ρ is the satellite-to-ground distance from the receiver to the satellite, and v ion For the corresponding ionospheric correction, v trop The tropospheric correction in response.

[0039] Furthermore, in step 2, the sample labels in the training set are based on the true error of the observations, and the sample correlation, i.e., the correlation between sample features and labels, is based on the distance between the true error and the classification threshold, and further includes:

[0040] When the true error of an observation is less than or equal to 1, the sample is labeled "quasi-observable"; when the true error of an observation is greater than 1, the sample is labeled "non-quasi-observable". The classification threshold is 1. The further the true error of an observation is from 1, the greater the correlation between the samples. Specifically:

[0041]

[0042] Where w is the sample correlation factor, e is the error, thresh is the threshold for selecting quasi-observations, and a, b, and c are the coefficients of the piecewise function, which can be adjusted according to the actual situation. This function is called the observation error and sample correlation function.

[0043] Furthermore, the method for obtaining the "accuracy score" of observations based on real-time dynamic GNSS observation data described in step 3 further includes:

[0044] The dynamic GNSS observation data and its key feature vectors are collected and extracted using the same method as in step 1, and used as input to the trained random forest classifier. The output is the probability of the observation being a "quasi-observable" and the probability of it being a "non-quasi-observable". The probability of selecting a "quasi-observable" is used as the quasi-score of the observation.

[0045] Furthermore, the method for screening quasi-observations and calculating true error described in step 4 further includes:

[0046] First, iterate through all the observations' "fit scores" and select observations whose "fit scores" reach a threshold as "fit observations". If the number of "fit observations" determined in this iteration already meets the requirement for calculating the true error, then calculate the true error. Next, iterate through all the observations' least squares residuals, select observations whose least squares residuals reach a threshold, and add them to the "fit observations" selected in the previous iteration, until the number of "fit observations" meets the requirement for calculating the true error.

[0047] The true error is calculated using the following expression:

[0048]

[0049] in, For the true error estimate, l DD For double-difference pseudorange observations, Q llDD -1 Let R be the corresponding weight matrix. DD The corresponding adjustment factor matrices are calculated using methods similar to those used in the unequal pseudorange observation model, i.e.

[0050]

[0051] A DD =[-Δe x -Δe y -Δe z ]

[0052] Q llDD -1 =diag(f ELE (ele))+diag(f ELE (ele ref )·1)

[0053] Among them, A DDDesign a matrix for a first-order Gauss-Markov model of double-difference pseudorange observations. Δ is the inter-satellite station double difference operator, and ele is the inter-satellite single difference operator. ref This represents the altitude angle of the reference star, and 1 represents an all-1 vector. Additionally, in the above formula, there is also...

[0054]

[0055] Among them, A r For A DD The submatrix belonging to the "quasi-observation" part, Q r -1 Let W be the corresponding weight matrix, and W be the transition matrix of size r×n, and let A be the weight matrix. r =WA DD .

[0056] Furthermore, the IGG3 robustness treatment method described in step 5 further includes:

[0057] The standardized expression for the true error is:

[0058]

[0059] in Let be the standardized true error of the i-th observation. To estimate the true error, Q llii Let be the corresponding variance. The expression for the variance inflation factor using the IGG3 robust model is:

[0060]

[0061] Where, γ ii Let k be the variance inflation factor for the i-th observation, and k0 and k1 are the robust parameters of IGG3. Their recommended values ​​are k0 = 1.5 and k1 = 2.5.

[0062] Finally, the variance of the original observations is amplified by a new variance inflation factor, and the least squares method is used again to solve the double-difference pseudorange observation model in step 4 to obtain more accurate GNSS positioning results.

[0063] Accordingly, this application also discloses a terminal, including a processor and a storage medium;

[0064] The storage medium is used to store instructions;

[0065] The processor is used to operate according to the instructions to execute the steps of the above-described random forest-based GNSS observation gross error calibration method.

[0066] Accordingly, this application also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned random forest-based GNSS observation gross error fitting verification method.

[0067] The beneficial effects of this invention are as follows:

[0068] This invention presents a GNSS observation gross error fitting verification method based on random forest. By introducing a random forest machine learning model, it explores and mines more GNSS feature values ​​and establishes a classifier for identifying "fitted observations". This method extends existing gross error verification methods by combining random forest with fitting verification methods, proposing a GNSS observation gross error fitting verification method based on random forest identification. This provides a more adaptive gross error verification means for different scenarios in complex environments, ultimately improving the positioning accuracy and stability of the positioning model. Attached Figure Description

[0069] Figure 1 This is a flowchart of the GNSS observation gross error fitting verification method based on random forest in this invention.

[0070] Figure 2 This is a schematic diagram showing how the model indicators of the present invention change under different test thresholds.

[0071] Figure 3 This is a schematic diagram comparing the positioning trajectories of the present invention's method (RF-QUAD) with two other traditional methods (LS-IGG3 and LS-QUAD) on a certain road section.

[0072] Figure 4 yes Figure 3 The corresponding error sequence diagram. Detailed Implementation

[0073] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention.

[0074] As shown in the figure, the GNSS observation gross error fitting verification method based on random forest described in this invention includes the following steps:

[0075] Step 1: Collect and extract dynamic GNSS observation data and its key feature vectors to construct a training set for a random forest classifier. The GNSS observation sample features in the feature vectors include:

[0076] The features obtained include elevation angle, signal-to-noise ratio (SNR), SNR attenuation, geometric contribution, signal integrity, dMW (Difference of Melbourne-Wübbena Combination), dGF (Difference of Geometry-Free Combination), carrier consistency, Doppler consistency, single-point positioning residual, LLI (Loss of Lock Indicator) flag, GDOP (Geometric Dilution of Precision), number of satellites, and satellite utilization. Elevation angle, SNR, signal integrity, LLI flag, and number of satellites can be directly obtained from RINEX files. GDOP, number of satellites, and satellite utilization are features shared by all samples in a given epoch. Other features requiring preprocessing are obtained through the following methods:

[0077] 1) Signal-to-noise ratio (SNR) attenuation: Using an ideal observation signal in an open environment, fit an elevation angle-SNR template function. The difference between the standard SNR and the actual SNR at the same elevation angle is the SNR attenuation, expressed as:

[0078]

[0079] Where S k The actual signal-to-noise ratio of satellite k. Let Att be the standard signal-to-noise ratio of satellite k at an elevation angle θ. k The signal-to-noise ratio attenuation of satellite k.

[0080] 2) Geometric contribution: Represents the component of the observation in the entire adjustment system, expressed as:

[0081] r k =R NDkk (2)

[0082] Where r k Let r be the geometric contribution of satellite k to the system. k =0, then it indicates that the observation is necessary; if r k =1 indicates that the observation is completely redundant, R NDkk R is the adjustment factor matrix of the unequal pseudorange observation model. ND The element in the k-th row and k-th column. Adjustment factor matrix R. ND The calculation method is as follows:

[0083] R ND =IA ND (A ND T Q llND -1 A ND )-1 A ND T Q llND -1 (3)

[0084] Where I is the identity matrix, A ND Design the matrix Q for the first-order Gauss-Markov model of unequal pseudorange observations. llND -1 These are the corresponding weight matrices. Specifically, their expressions are:

[0085] A ND =[-e x -e y -e z 1] (4)

[0086] Q llND -1 =diag(f ELE (ele)) (5)

[0087] Among them, e x e y e z f represents the x, y, and z components of the direction vector from which the receiver points to the satellite, respectively. ELE (ele) represents a stochastic model related to elevation angle, and diag(·) represents the operation of converting a vector into a diagonal matrix.

[0088] 3) dMW and dGF: These are the epoch differences between observations without ionospheric combinations and without geometric combinations. Under conditions where there are no gross errors in the pseudorange and no cycle slips in the carrier wave, these two characteristics should remain within a relatively stable range. The non-difference calculation methods for observations without ionospheric combinations and without geometric combinations are as follows:

[0089]

[0090] GF=λ1L1-λ2L2 (7)

[0091] Where MW represents the ionospheric-free combination, GF represents the geometric-free combination, L represents the satellite carrier observation, P represents the satellite pseudorange observation, f represents the carrier frequency, λ represents the carrier wavelength, and the subscript represents the frequency point of the observation.

[0092] 4) Carrier consistency and Doppler consistency: This refers to the consistency between the carrier wave and pseudorange at a certain satellite frequency, and the consistency between Doppler and pseudorange. Under normal circumstances, these two characteristics should not vary significantly. Their calculation methods are as follows:

[0093] LP=P(t)-P(t-1)-λL(t)+λL(t-1) (8)

[0094]

[0095] Where LP stands for carrier consistency, DP stands for Doppler consistency, D stands for Doppler observation, t and t-1 represent two adjacent epochs, and dt is their time difference.

[0096] 5) Single-point location residuals: The residuals obtained by substituting the unequal pseudorange observations into the first-order Gauss-Markov model and applying the least squares method are the single-point location residuals. Specifically, the expression for solving the least squares observation residuals is:

[0097] V ND =-Rl ND (10)

[0098] Among them, V ND Let l be the least squares residual vector of the non-difference pseudorange observation model. ND The corresponding unequal pseudorange observation can be specifically represented as:

[0099] l ND =P-ρ+v ion +v trop (11)

[0100] Where P is the pseudorange observation value of the satellite, ρ is the satellite-to-ground distance from the receiver to the satellite, and v ion For the corresponding ionospheric correction, v trop The tropospheric correction in response.

[0101] In step 2: the observed data is divided into "quasi-observations" and "non-quasi-observations," and a random forest classifier is trained based on this classification. The sample labels in the training set are based on the true error of the observations. Sample correlation, i.e., the correlation between sample features and labels, is based on the distance between the true error and the classification threshold, and further includes:

[0102] When the true error of an observation is less than or equal to 1, the sample is labeled "quasi-observable"; when the true error of an observation is greater than 1, the sample is labeled "non-quasi-observable". The classification threshold is 1. The further the true error of an observation is from 1, the greater the correlation between the samples. Specifically:

[0103]

[0104] Where w is the sample correlation factor, e is the error, thresh is the threshold for selecting quasi-observations, and a, b, and c are the coefficients of the piecewise function, which can be adjusted according to the actual situation. This function is called the observation error and sample correlation function.

[0105] Step 3: Apply the random forest classifier to the real-time dynamic positioning system, and obtain the "fit score" for each observation based on the real-time dynamic GNSS observation data. The trained random forest classifier outputs the "fit score" for the real-time dynamic GNSS observation data, further including:

[0106] The dynamic GNSS observation data and its key feature vectors are collected and extracted using the same method as in step 1, and used as input to the trained random forest classifier. The output is the probability of the observation being a "quasi-observable" and the probability of it being a "non-quasi-observable". The probability of selecting a "quasi-observable" is used as the quasi-score of the observation.

[0107] In step 4: If the "fit score" of an observation reaches a threshold, it is classified as a "fit observation." If the number of "fit observations" meets the requirement, the true error of all observations is calculated. If the number does not meet the requirement, a portion of the "fit observations" are selected based on the least squares residuals until the number meets the requirement, and the true error of all observations is calculated. The process of determining "fit observations" based on the "fit score" and least squares residuals, and calculating the true error based on the determined "fit observations," further includes:

[0108] First, iterate through all the observations' "fit scores" and select observations whose "fit scores" reach a threshold as "fit observations". If the number of "fit observations" determined in this iteration already meets the requirement for calculating the true error, then calculate the true error. Next, iterate through all the observations' least squares residuals, select observations whose least squares residuals reach a threshold, and add them to the "fit observations" selected in the previous iteration, until the number of "fit observations" meets the requirement for calculating the true error.

[0109] The true error is calculated using the following expression:

[0110]

[0111] in, For the true error estimate, l DD For double-difference pseudorange observations, Q llDD -1 Let R be the corresponding weight matrix. DD The corresponding adjustment factor matrices are calculated using methods similar to those used in the unequal pseudorange observation model, i.e.

[0112] R DD =IA DD (A DD T Q llDD -1 A DD )-1 A DD T Q llDD -1 (14)

[0113]

[0114] A DD =[-Δe x -Δe y -Δe z (16)

[0115] Q llDD -1 =diag(f ELE (ele))+diag(f ELE (ele ref )·1) (17)

[0116] Among them, A DD Design a matrix for a first-order Gauss-Markov model of double-difference pseudorange observations. Δ is the inter-satellite station double difference operator, and ele is the inter-satellite single difference operator. ref Let represent the altitude angle of the reference star, and 1 represent a vector consisting entirely of 1s. Additionally, in equation (13), there is also...

[0117]

[0118] Among them, A r For A DD The submatrix belonging to the "quasi-observation" part, Q r -1 Let W be the corresponding weight matrix, and W be the transition matrix of size r×n, and let A be the weight matrix. r =WA DD .

[0119] Step 5: Replace the traditional least squares residual with the standardized true error, perform IGG3 robustness processing, and output the robust positioning coordinates. This process of replacing the traditional least squares residual with the standardized true error and performing IGG3 robustness processing further includes:

[0120] The standardized expression for the true error is:

[0121]

[0122] in Let be the standardized true error of the i-th observation. To estimate the true error, Q llDDii For Q llDDThe i-th row and i-th column. The expression for the variance inflation factor using the IGG3 robust model is:

[0123]

[0124] Where, γ ii Let k be the variance inflation factor for the i-th observation, and k0 and k1 are the robust parameters of IGG3. Their recommended values ​​are k0 = 1.5 and k1 = 2.5.

[0125] Finally, the variance of the original observations is amplified by a new variance inflation factor, and the least squares method is used again to solve the double-difference pseudorange observation model of equations (15), (16), and (17) to obtain more accurate GNSS positioning results.

[0126] Figure 2 The diagram illustrates the changes in the model metrics of this invention under different test thresholds. It can be seen that when the test threshold is 0.7, the accuracy and recall of the random forest classification model reach a balanced level. Therefore, this invention uses this threshold as the test threshold for "quasi-observables" and "non-quasi-observables".

[0127] Figure 3 To compare the positioning trajectories of the present invention's method (RF-QUAD) with two other traditional methods (LS-IGG3 and LS-QUAD) for a certain road segment, Figure 4 The corresponding error sequence diagram is shown. It can be seen that the positioning trajectory of the method of this invention is more concentrated, and there are fewer positioning points that significantly deviate from the trajectory, proving the effectiveness of this invention in GNSS observation gross error detection and positioning accuracy improvement.

[0128] The beneficial effect of this invention lies in that, compared with the prior art, it proposes a random forest-based method for GNSS observation gross error fitting verification. Based on pre-acquired and extracted dynamic GNSS observation data and its key feature vectors, a random forest classifier is trained for a binary classification problem of "fitted observations" and "non-fitted observations." Combining the classifier with traditional fitting verification methods improves the verification efficiency of observation gross errors and the accuracy and stability of dynamic positioning in complex environments.

[0129] This invention can be a system, method, and / or computer program product.

[0130] Based on the spirit of this invention, those skilled in the art will readily conceive of a computer program product derived from the aforementioned ground-enhanced positioning dynamic three-dimensional non-uniform grid point clustering method. The computer program product may include a computer-readable storage medium loaded with computer-readable program instructions for causing a processor to implement various aspects of this disclosure. Specifically, this application also includes a terminal comprising a processor and a storage medium; the storage medium stores instructions; the processor operates according to the instructions to execute the steps of the aforementioned ground-enhanced positioning dynamic three-dimensional non-uniform grid point clustering method.

[0131] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0132] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0133] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0134] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A method for calibrating gross errors in GNSS observations based on random forest, characterized in that, Includes the following steps: Step 1: Collect and extract dynamic GNSS observation data and its key feature vectors to construct a training set for a random forest classifier; Step 2: Divide the observation data into "quasi-observations" and "non-quasi-observations", and train a random forest classifier based on this classification; Step 3: Apply the random forest classifier to the real-time dynamic positioning system and obtain the "fit score" for each observation based on the real-time dynamic GNSS observation data. Step 4: If the "fit score" of the observation reaches the threshold, it is classified as a "fit observation". If the number of "fit observations" meets the requirements, the true error of all observations is calculated. If the number does not meet the requirements, a portion of the "fit observations" are selected based on the least squares residuals until the number meets the requirements, and the true error of all observations is calculated. Step 5: Replace the traditional least squares residual with the standardized true error, perform IGG3 robustness processing, and output the robust positioning coordinates.

2. The GNSS observation gross error fitting verification method based on random forest according to claim 1, characterized in that, In step 1, dynamic GNSS observation data and its key feature vectors are collected and extracted. The GNSS observation sample features in the feature vectors include: The features obtained include elevation angle, signal-to-noise ratio (SNR), SNR attenuation, geometric contribution, signal integrity, dMW, dGF, carrier consistency, Doppler consistency, single-point positioning residual, LLI flag, GDOP, number of satellites, and satellite utilization. Elevation angle, SNR, signal integrity, LLI flag, and number of satellites are obtained directly from RINEX files. GDOP, number of satellites, and satellite utilization are features shared by all samples in a given epoch. The others require some preprocessing. The feature acquisition method is as follows: 1) Signal-to-noise ratio (SNR) attenuation: Using an ideal observation signal in an open environment, fit an elevation angle-SNR template function. The difference between the standard SNR and the actual SNR at the same elevation angle is the SNR attenuation, expressed as: Where S k The actual signal-to-noise ratio of satellite k. Let Att be the standard signal-to-noise ratio of satellite k at an elevation angle θ. k The signal-to-noise ratio attenuation for satellite k; 2) Geometric contribution: Represents the component of the observation in the entire adjustment system, expressed as: r k ＝R NDkk Where r k Let r be the geometric contribution of satellite k to the system. k =0, which indicates that the observation is necessary; if r k =1 indicates that the observation is completely redundant, R NDkk R is the adjustment factor matrix of the unequal pseudorange observation model. ND The element in the k-th row and k-th column; the adjustment factor matrix R ND The calculation method for R is as follows: ND =IA ND (A ND T Q llND -1 A ND ) -1 A ND T Q llND -1 Where I is the identity matrix, A ND Design the matrix Q for the first-order Gauss-Markov model of unequal pseudorange observations. llND -1 These are the corresponding weight matrices; specifically, their expressions are: TO ND =[-e x -And y -And z 1] Q llND -1 =diag(f ELE (he)) Among them, e x e y e z f represents the x, y, and z components of the direction vector from which the receiver points to the satellite, respectively. ELE (ele) represents a stochastic model related to elevation angle, and diag(·) represents the operation of converting a vector into a diagonal matrix; 3) dMW and dGF: These are the epoch differences between observations without ionospheric combinations and without geometric combinations. Under the condition that there are no gross errors in the pseudorange and no cycle slips in the carrier, these two characteristics should remain within a relatively stable range. The non-difference calculation methods for observations without ionospheric combinations and without geometric combinations are as follows: GF=λ1L1-λ2L2 Where MW represents the ionospheric-free combination, GF represents the geometric-free combination, L represents the satellite carrier observation, P represents the satellite pseudorange observation, f represents the carrier frequency, λ represents the carrier wavelength, and the subscript represents the frequency point of the observation. 4) Carrier consistency and Doppler consistency: This refers to the consistency between the carrier wave and pseudorange at a certain satellite frequency, and the consistency between Doppler and pseudorange. Under normal circumstances, these two characteristics should not vary significantly. Their calculation methods are as follows: LP=P(t)-P(t-1)-λL(t)+λL(t-1) Where LP stands for carrier consistency, DP stands for Doppler consistency, D stands for Doppler observation, t and t-1 represent two adjacent epochs, and dt is their time difference. 5) Single-point location residuals: After substituting the non-difference pseudorange observations into the first-order Gauss-Markov model, the residuals obtained by applying the least squares method are the single-point location residuals; specifically, the expression for solving the least squares observation residuals is: V ND =-Rl ND Among them, V ND Let l be the least squares residual vector of the non-difference pseudorange observation model. ND The corresponding non-difference pseudorange observations are specifically represented as follows: l ND =P-ρ+v ion +v trop Where P is the pseudorange observation value of the satellite, ρ is the satellite-to-ground distance from the receiver to the satellite, and v ion For the corresponding ionospheric correction, v trop The tropospheric correction in response.

3. The GNSS observation gross error fitting verification method based on random forest according to claim 2, characterized in that, In step 2, the sample labels in the training set are based on the true error of the observations. Sample correlation, i.e., the correlation between sample features and labels, is based on the distance between the true error and the classification threshold, and further includes: When the true error of an observation is less than or equal to 1, the sample is labeled "quasi-observable"; when the true error of an observation is greater than 1, the sample is labeled "non-quasi-observable". The classification threshold is 1. The further the true error of an observation is from 1, the greater the correlation between the samples. Specifically: Where w is the sample correlation factor, e is the error, thresh is the threshold for selecting quasi-observations, and a, b, and c are the coefficients of the piecewise function, which are adjusted according to the actual situation; this function is called the observation error and sample correlation function.

4. The GNSS observation gross error fitting verification method based on random forest according to claim 3, characterized in that, In step 3, the trained random forest classifier is used to output a "fit score" for the real-time dynamic GNSS observation data, which further includes: The dynamic GNSS observation data and its key feature vectors are collected and extracted using the same method as in step 1. These are then used as input to the trained random forest classifier. The output consists of the probability that the observation is a "quasi-observable" and the probability that it is a "non-quasi-observable". The probability of selecting a "quasi-observable" is taken as the quasi-observable score.

5. The GNSS observation gross error fitting verification method based on random forest according to claim 4, characterized in that, In step 4, the "prototyped observation" is determined based on the "prototyped score" and the least squares residual, and the true error is calculated based on the determined "prototyped observation," which further includes: First, iterate through all the observations' "fit scores" and select observations whose "fit scores" reach a threshold as "fit observations". If the number of "fit observations" determined in this iteration already meets the requirement for calculating the true error, then calculate the true error. Next, iterate through all the observations' least squares residuals and select observations whose least squares residuals reach a threshold, adding them to the "fit observations" selected in the previous iteration, until the number of "fit observations" meets the requirement for calculating the true error. The true error is calculated using the following expression: in, For the true error estimate, l DD For double-difference pseudorange observations, Q llDD -1 Let R be the corresponding weight matrix. DD The corresponding adjustment factor matrices are calculated using methods similar to those used in the unequal pseudorange observation model, i.e. R DD =I-A DD (A DD T Q llDD -1 A DD ) -1 A DD T Q llDD -1 l DD =Δ▽P-Δ▽ρ+Δ▽v ion +Δ▽v trop A DD =[-Δe x -Δe y -Δe z ] Q llDD -1 =diag(f ELE (he))+diag(f ELE (he ref )·1) Among them, A DD Design the matrix for the first-order Gauss-Markov model of double-difference pseudorange observations, where Δ▽ is the inter-station double-difference operator, Δ is the inter-satellite single-difference operator, and ele ref The altitude angle of the reference star is represented by , and 1 represents a vector consisting entirely of 1s; additionally, in the above formula, there is also Among them, A r For A DD The submatrix belonging to the "quasi-observation" part, Let W be the corresponding weight matrix, and W be the transition matrix of size r×n, and let A be the weight matrix. r =WA DD .

6. The method for GNSS observation gross error fitting verification based on random forest according to claim 5, characterized in that, In step 5, the standardized true error is used instead of the traditional least squares residual, and IGG3 robustness processing is performed. This further includes: The standardized expression for the true error is: in Let be the standardized true error of the i-th observation. To estimate the true error, Q llDDii For Q llDD The expression for the variance inflation factor using the IGG3 robust model is: (The first row and the second column are not provided in the original text.) Where, γ ii Let k be the variance inflation factor for the i-th observation, and k0 and k1 be the IGG3 robustness parameters. Their recommended values ​​are k0 = 1.5 and k1 = 2.

5. Finally, the variance of the original observations is amplified by the calculated variance inflation factor, and the least squares method is used again to solve the double-difference pseudorange observation model in step 4 to obtain a more accurate GNSS positioning result.

7. A terminal, characterized in that: Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the GNSS observation gross error fitting verification method based on random forest according to any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the steps of the GNSS observation gross error fitting verification method based on random forest as described in any one of claims 1-6.

Images