NLOS identification-based single Beidou terminal precision single-point positioning enhancement method

Abstract

The invention provides a single Beidou terminal precision single-point positioning enhancement method based on NLOS identification. The method comprises the following steps: acquiring satellite GNSS measurement data of a plurality of positions or scenes and constructing a satellite sky map; constructing an NLOS signal identification model based on a graph neural network; identifying a satellite signal classification result of the satellite sky map through an NLOS signal identification model; constructing a random model based on a satellite signal classification result; and calculating the position of the user according to the GNSS precise ephemeris, the precise clock error, the differential code deviation, the GNSS measurement data and the ephemeris data. According to the invention, the precision single-point positioning precision of GNSS satellite positioning in urban complex environment areas and the generalization ability of NLOS signals in different time and place data sets can be improved; according to the invention, an appropriate random model is distributed according to the identified signal, and finally the random model is used to calculate the position of the user, so that the GNSS satellite positioning precision in the urban complex environment is improved.

Description

Technical Field

[0001] The invention relates to the field of satellite positioning technology, and in particular to a single Beidou terminal precise single-point positioning enhancement method based on NLOS recognition. Background Art

[0002] GNSS precise point positioning (PPP) is a typical technical means for high-precision positioning. The random model is an important part of PPP positioning and has a great influence on the results of precise point positioning. At present, the random model is mainly obtained based on empirical formulas. However, in complex scenes such as cities, tree-lined areas, and overpasses, GNSS satellite signals are easily blocked by buildings, tree shades, bridges, etc., resulting in the reception of non-line-of-sight (NLOS) signals. At this time, the empirical formula often cannot accurately reflect the actual error level of the observation value, which will affect the PPP positioning results.

[0003] Traditional methods use vector tracking algorithms to mitigate the impact of NLOS signal reception and multipath effects. If the NLOS signal persists, its signal strength may be equal to that of the direct signal, thus affecting the effectiveness of the vector tracking algorithm. Some other methods use three-dimensional map-assisted positioning technology to identify NLOS signals, which greatly increases costs and has poor versatility.

[0004] In recent years, machine learning-based methods have been widely used in global navigation satellite system (GNSS) positioning to mitigate the interference of NLOS signals. Such methods, such as support vector machine (SVM) and decision tree (DT), use machine learning models to classify NLOS signals and remove identified NLOS signals from the observation equation to eliminate NLOS interference. Using deep learning models to identify NLOS signals is a direct and effective way to mitigate NLOS interference and improve GNSS positioning accuracy. However, machine learning-based methods still have the following defects:

[0005] 1) Machine learning methods are excellent in extracting hidden features, but existing machine learning-based NLOS signal recognition methods have poor generalization capabilities on different time and location data sets and are difficult to adapt to changing environments. 2) In complex urban scenarios, directly eliminating NLOS signals may reduce the number of satellites available to the positioning algorithm, resulting in reduced positioning accuracy. If all NLOS signals are directly multiplied by a value to reduce their weights, without considering that NLOS signals will also have different effects, the final random model of the NLOS signal will be inaccurate, which will also lead to reduced positioning accuracy. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a single Beidou terminal precise single-point positioning enhancement method based on NLOS recognition, which effectively improves the accuracy and generalization ability of NLOS satellite signal recognition and improves the GNSS satellite positioning accuracy in complex urban environments.

[0007] The technical solution of the present invention is: a single Beidou terminal precise single-point positioning enhancement method based on NLOS recognition, comprising the following steps:

[0008] S1), obtaining satellite GNSS measurement data and ephemeris data for multiple locations or scenes, and obtaining a sky map for the current location;

[0009] S2), constructing a satellite sky map based on the satellite data obtained in step S1);

[0010] S3), constructing a NLOS signal recognition model based on a graph neural network; and inputting the satellite sky map into the pre-trained NLOS signal recognition model based on a graph neural network to obtain the satellite signal classification result;

[0011] S4), constructing a random model based on the satellite signal classification results, and obtaining the variance of the pseudorange and phase observation values ​​of each satellite by inputting all satellites of a single epoch into the random model;

[0012] S5) Calculate the user's position based on the GNSS precise ephemeris, precise clock error, differential code deviation, and GNSS measurement data and ephemeris data.

[0013] Preferably, in step S2), constructing a satellite sky map specifically includes the following steps:

[0014] S21), select satellite node features;

[0015] S22), constructing an adjacency matrix based on neighborhood attributes and constellation attributes, connecting satellite nodes at the same time;

[0016] S23), selecting edge features;

[0017] S24), combining the node features, the adjacency matrix and the edge feature vector to construct a sky satellite map.

[0018] Preferably, in step S21), the satellite node characteristics include elevation angle, azimuth angle, carrier-to-noise ratio and pseudorange residual.

[0019] Preferably, in step S22), the neighborhood attribute refers to judging whether the connection is based on the distance between satellites, and the distance between satellites is calculated based on three-dimensional coordinates converted from elevation angle and azimuth angle;

[0020] The constellation attribute refers to the establishment of a connection relationship between satellites in the same constellation.

[0021] Preferably, in step S23), the edge features include the relative position of the elevation angle and the azimuth angle difference, the relative signal quality between the carrier-to-noise ratio and the pseudorange residual, and the spherical distance.

[0022] Preferably, in step S3), the NLOS signal recognition model based on graph neural network includes a multi-layer graph attention-sparse adaptation layer and a classifier. The graph attention-sparse adaptation layer calculates the attention weights between satellite nodes through a multi-head attention mechanism, captures the complex nonlinear relationship between nodes, updates satellite features, helps to learn the environmental representation of adjacent satellites, is beneficial to non-line-of-sight satellite detection tasks, and outputs the probability of the satellite being NLOS through the classifier.

[0023] Preferably, in step S3), the graph attention-sparse adaptation layer includes a graph attention module and a sparse adapter; for a given node feature, the graph attention module calculates the multi-head attention between satellite nodes according to the adjacency matrix; and obtains the attention score by the dot product of the query vector and the key vector, multiplies the obtained attention score with the learned value vector, and then concatenates the information from all heads to obtain the environment representation of the central node; after the edge feature is projected into the latent space, it is added to the attention score, the representation of the relationship between satellites is adjusted, the topological relationship modeling is promoted, and the graph attention mechanism is further assisted to learn the environment information and improve the model performance. The sparse adapter selects useful environment information from the environment representation of the central node to update the satellite features of the l+1 layer; by learning the weight vector to screen the key environment representation, the influence of input data noise and redundant information is reduced, so that the satellite can adaptively select useful environment information according to its own characteristics, improve the availability of environment representation, and enhance the generalization performance of the model.

[0024] Preferably, in step S4), a random model is constructed based on the satellite signal classification result, which specifically includes the following steps:

[0025] S41), classify the satellites according to the NLOS identification results, and obtain the number of NLOS satellites N NLOS 、N number of LOS satellites LOS ;

[0026] S42), calculate the position precision loss value PDOP of all LOS satellites; if the PDOP value of the LOS satellite is less than the first preset value; then the variance of all NLOS satellites is *A; otherwise, the variance of all NLOS satellites is *B;

[0027] S43), if the PDOP value of the LOS satellite is less than the second preset value; then the variance of all satellites is *C; otherwise, the variance of all NLOS satellites is *D.

[0028] The beneficial effects of the present invention are:

[0029] 1. The present invention can improve the precision of precise single-point positioning of GNSS satellite positioning in urban complex environments; and improve the generalization ability of NLOS signals on different time and location data sets; first identify the NLOS signal, then assign a suitable random model according to the identified signal, and finally use the random model to calculate the user's position, so as to improve the precision of GNSS satellite positioning in urban complex environments;

[0030] 2. The present invention uses sky satellite images to represent the relative relationship between satellites, combines edge feature encoding in the graph transformer, projects edge features into the latent space and adds them to the attention score, which promotes topological relationship modeling and further assists the graph attention mechanism in learning environmental information;

[0031] 3. In order to enhance the generalization performance of the model, the present invention designs a sparse adapter, which filters key environmental representations by learning weight vectors, reduces the impact of input data noise and redundant information, improves the availability of environmental representations, and enhances the generalization performance of the model;

[0032] 4. The present invention uses the results of NLOS identification to signal mark each observation value involved in the solution. For the satellite value of each epoch, it is necessary to comprehensively consider the characteristic values ​​of all satellites and then reasonably determine the random model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the process of the present invention;

[0034] Figure 2 A framework diagram for constructing a satellite sky map for the present invention;

[0035] Figure 3 This is a structural framework diagram of the NLOS signal recognition model based on graph neural network of the present invention;

[0036] Figure 4 This is a structural framework diagram of the graph attention-sparse adaptation layer of the present invention;

[0037] Figure 5 This is a structural framework diagram of the graph attention module of the present invention;

[0038] Figure 6 It is a structural framework diagram of the sparse adapter of the present invention. DETAILED DESCRIPTION

[0039] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings:

[0040] like Figure 1As shown, this embodiment provides a single Beidou terminal precise single point positioning enhancement method based on NLOS recognition, including the following steps:

[0041] S1), obtaining satellite GNSS measurement data and ephemeris data for multiple locations or scenes, and obtaining a sky map for the current location;

[0042] This embodiment uses a satellite receiver to record satellite GNSS measurement data of multiple locations or scenes, and uses a fisheye camera to capture the sky map of the current location. The position of each satellite projected onto the sky map is calculated based on the satellite's altitude and azimuth in the satellite GNSS measurement data. Whether the satellite is an NLOS signal is determined by whether the satellite projection position overlaps with a building, thereby marking the satellite data samples and finally constructing a sample data set for model pre-training.

[0043] S2) constructing a satellite sky map based on the satellite data obtained in step S1); specifically comprising the following steps:

[0044] S21), selecting satellite node features; wherein the satellite node features include elevation angle el, azimuth angle az, carrier-to-noise ratio C / N0 and pseudorange residual ε; wherein the pseudorange residual ε is expressed as:

[0045] ε=pr-D·Δx

[0046] Where ε represents the pseudorange residual, Δx represents the initial guess state, which consists of the initial guess position bias and the receiver clock bias; pr is the difference between the pseudorange measurement and the initial state; and D is the measurement matrix consisting of the unit line-of-sight vector between the satellite and the receiver.

[0047] S22), constructing an adjacency matrix based on neighborhood attributes and constellation attributes, connecting satellite nodes at the same time;

[0048] The neighborhood attribute refers to determining whether the connection is based on the distance between satellites, and the distance between satellites is calculated based on three-dimensional coordinates converted from elevation angles and azimuth angles. In order to calculate the distance between two satellites on the spherical surface, this embodiment first uses the elevation angle el and the azimuth angle az to obtain their three-dimensional Cartesian coordinates:

[0049]

[0050] Where (x, y, z) represents the three-dimensional coordinates of the satellite in Euclidean space, and r represents the radius of the mapping sphere;

[0051] Then the three-dimensional Euclidean coordinates are used to calculate the spherical distance s between the two satellites. ij , calculate the angle α between the two satellites according to the cosine formula ij ,Right now:

[0052]

[0053] In the formula, the positions of the two satellites are represented by P i (x i ,y i ,z i ) and P j (x j ,y j ,z j );

[0054] Since the positions of the two satellites P i and P j On the same sphere, the spherical distance between them is s ij Defined as:

[0055]

[0056] Finally, the spherical distance s is obtained ij for:

[0057]

[0058] When the distance between two satellites ij When it is less than the threshold γ, this embodiment considers that there is an edge connection between the two satellites.

[0059] In this embodiment, the constellation attribute refers to the establishment of a connection relationship between satellites in the same constellation.

[0060] According to the neighborhood property and constellation property, the values ​​in the adjacency matrix A are defined as:

[0061]

[0062] Where C refers to the constellation attribute, and v represents the satellite node on the sky satellite map.

[0063] S23), select edge features; wherein the edge features include the relative position of the elevation angle and the azimuth angle difference, the relative signal quality between the carrier-to-noise ratio and the pseudo-range residual, the spherical distance s ij , the spherical distance s ij Used to provide more spatial and signal quality information between satellites and assist in non-line-of-sight detection.

[0064] Among them, the relative position of the elevation angle and azimuth angle difference is expressed as:

[0065]

[0066] In the formula, Δel ij represents the elevation angle el of the i-th target satellite iThe elevation angle el with respect to the jth adjacent satellite j The difference between ij represents the azimuth angle az of the i-th central satellite i The azimuth angle az to the jth adjacent satellite j The difference between .

[0067] The relative signal quality between the carrier-to-noise ratio and the pseudorange residual is expressed as:

[0068]

[0069] in, represents the carrier-to-noise ratio of the i-th target satellite The carrier-to-noise ratio with the jth adjacent satellite The difference between ij represents the i-th central satellite pr i pr with the jth adjacent satellite j The difference between .

[0070] S24), combining the node features, the adjacency matrix and the edge feature vector to construct a sky satellite map SSG. The sky satellite map SSG constructed in this embodiment is as follows: Figure 2 As shown, the expression of the sky satellite image SSG is:

[0071] SSG=(V,A,E)

[0072]

[0073] Among them, V represents the node feature set of n satellites captured in the same epoch, where, is the selected satellite feature; A represents the adjacency matrix; E represents the edge feature vector set, is the selected edge feature.

[0074] S3), constructing a NLOS signal recognition model based on a graph neural network; and inputting the satellite sky map into the pre-trained NLOS signal recognition model based on a graph neural network to obtain the satellite signal classification result;

[0075] In this embodiment, Figure 3 As shown, the NLOS signal recognition model based on graph neural network includes a multi-layer graph attention-sparse adaptation layer and a classifier. The graph attention-sparse adaptation layer calculates the attention weights between satellite nodes through a multi-head attention mechanism, captures the complex nonlinear relationship between nodes, updates satellite features, helps to learn the environmental representation of adjacent satellites, is beneficial to non-line-of-sight satellite detection tasks, and outputs the probability of the satellite being NLOS through the classifier.

[0076] like Figure 4 As shown, the graph attention-sparse adaptation layer includes a graph attention module and a sparse adapter; for a given node feature set like Figure 5 As shown, the graph attention module calculates the multi-head attention between satellite nodes according to the adjacency matrix A; in this embodiment, the calculation of the m-th head attention is as follows:

[0077]

[0078] in, The query vector representing the i-th target satellite of the m-th head attention of the l-th network layer; The key vector representing the jth neighboring satellite of the mth head attention of the lth network layer; The value vector representing the jth neighboring satellite of the mth head attention of the lth network layer; The weight vectors representing the query vector, key vector, and value vector of the m-th head attention of the l-th network layer respectively; They represent the learnable biases of the query vector, key vector, and value vector of the m-th head attention of the l-th network layer respectively; represents the node characteristics of the i-th target satellite at the l-th network layer; Node characteristics of the jth neighboring satellite at the lth network layer.

[0079] In this embodiment, first, a learnable parameter matrix and a learnable bias Convert the node features of the central satellite into a query vector

[0080] Then, use different trainable parameters and Convert node features of neighboring satellites into key vectors Sum value vector

[0081] To capture the relationship between a satellite and its neighbors, the attention score is obtained by calculating the dot product of the query vector and the key vector: Finally, the output is represented as a weighted sum of the value vectors based on the attention weights, and the information from multiple heads is fused together for message passing.

[0082]

[0083] The attention score obtained With the learned value vector Multiply and then concatenate the information from all heads to obtain the environment representation of the central node

[0084]

[0085] in, represents the attention score of the i-th target satellite and the j-th neighboring satellite of the m-th head attention of the l-th network layer; T represents the transpose of the matrix; softmax represents the activation function; d represents the hidden dimension of each attention layer, ‖ represents the tensor concatenation operation; M is the number of heads; N(i) represents the set of neighboring nodes of the central node, and the adjacency matrix A containing edge connection information must be taken as an additional input; The environment representation of the central node;

[0086] Then, by taking the satellite features h of the previous layer l Environment representation with central node Add together to get the updated satellite characteristics h l+1 .

[0087] After projecting the edge features into the latent space, they are added to the attention score to adjust the representation of the relationship between satellites, promote topological relationship modeling, further assist the graph attention mechanism in learning environmental information, and improve model performance; specifically:

[0088] The edge features are first projected into a latent space, namely:

[0089]

[0090] in, Represents the implicit edge features learned by the encoding layer; is the linear projection matrix, if l=1, k=5; otherwise, k=d; ReLu represents the nonlinear activation function, which ensures that the output of the model is positive; encoding represents the encoding operation; represents the implicit edge features of the i-th target satellite and the j-th adjacent satellite learned by the encoding layer at the l-th layer;

[0091] Then, the implicit edge feature Adding this to the attention score, we get:

[0092]

[0093] in, is a learnable matrix that transforms The projection is the weight value, and ‖ means concatenating the outputs of M heads.

[0094] The node and edge features are fused through the graph attention module, namely:

[0095]

[0096] In the formula, H and Z represent node features and edge features respectively; GNE-Attention represents the graph attention module, and ER represents the output of the graph attention module.

[0097] like Figure 6 As shown, the sparse adapter is represented by the environment of the central node Useful environmental information is selected to update the satellite features of the l+1 layer; by learning the weight vector to screen the key environmental representation, the impact of input data noise and redundant information is reduced, so that the satellite can adaptively select useful environmental information according to its own characteristics, improve the availability of environmental representation, and enhance the generalization performance of the model, as follows:

[0098]

[0099] in, represents the weight vector of the upper and lower projections of the i-th target satellite in the l-th network layer; GeLU represents the nonlinear activation function; sigmoid represents the activation function; They represent the learnable weight matrices of the upper and lower projections of the lth network layer respectively. Compress the central satellite into a low-dimensional space to obtain Then an up-projection matrix is ​​applied so that the weight vector is aligned with the environment representation of the center node They have the same shape; they use the nonlinear activation function GeLU to reduce noise interference.

[0100] By adding the weight vector Applied to the original environment representation ER to obtain the sparse environment representation SER, that is:

[0101]

[0102] Satellite characteristics of layer l+1 The updates are as follows:

[0103]

[0104] in, is a learnable residual connection matrix that is used to retain some information from the previous layer and prevent over-smoothing. In this way, the sparse adapter is derived from the environment representation of the central node. Useful environmental information is selected to update the satellite features of the l+1 layer.

[0105] S4) Based on the satellite signal classification results, a random model is constructed, and the pseudorange of each satellite is obtained by inputting all satellites in a single epoch into the random model. and the variance of the phase observations

[0106] In this embodiment, a random model is constructed based on the satellite signal classification results, which specifically includes the following steps:

[0107] S41), classify the satellites according to the NLOS identification results, and obtain the number of NLOS satellites N NLOs 、N number of LOS satellites LOS ;

[0108] S42), calculating the position loss of precision PDOP of all LOS satellites; if the PDOP value of the LOS satellite is less than the first preset value 3.0; then the variance of all NLOS satellites is * 1000; otherwise, the variance of all NLOS satellites is * 10;

[0109] S43), if the PDOP value of the LOS satellite is less than the second preset value 4.0; then the variance of all satellites is * 100; otherwise, the variance of all NLOS satellites is * 3.0.

[0110] In this embodiment, the calculation expression of the position loss of precision value PDOP is as follows:

[0111]

[0112] Among them, Q xx , Q yy , Q zz are the diagonal elements of the covariance matrix.

[0113] The covariance matrix can be calculated by the following steps:

[0114] Assume that in GNSS positioning, the observation equation is expressed as:

[0115] Z=HX+V;

[0116] Where: Z is the observation value vector, which contains observation information such as pseudorange and phase; H is the observation matrix, whose elements are related to the geometric position of the satellite and the observation type; X is the state vector to be determined, including the three-dimensional coordinates and clock error of the receiver; V is the observation noise vector;

[0117] According to the least squares principle, the optimal estimate of the state vector X is for:

[0118]

[0119] Where W is the weight matrix of the observations, which is a diagonal matrix whose diagonal elements are related to the accuracy of the observations; T represents the transpose operation of the matrix;

[0120] Covariance matrix Q = (H T W -1 H)-1 , from which we can get Q xx , Q yy , Q zz , and then calculate the PDOP.

[0121] S5) Calculate the user's position based on the GNSS precise ephemeris, precise clock error, and GNSS measurement data and ephemeris data.

[0122] This embodiment uses a non-difference non-combined PPP algorithm to calculate the user location, which is as follows:

[0123]

[0124] In the formula, r and s represent the receiver and satellite respectively, represents the original pseudorange observation value, represents the original observed value of the phase; represents the distance between the satellite and the phase center of the receiver antenna; c represents the speed of light; dt r and dt s Represent the receiver and satellite precise clock errors, T r represents the tropospheric delay error, represents the ionospheric delay error; λ represents the wavelength; represents the phase ambiguity; represents pseudorange multipath; represents phase multipath; represents the influence of pseudorange observation noise and other unmodeled errors; represents the effect of phase observation noise and other unmodeled errors.

[0125] The above embodiments and descriptions are only for illustrating the principles and best embodiments of the present invention. Without departing from the spirit and scope of the present invention, the present invention may be subject to various changes and improvements, all of which fall within the scope of the present invention to be protected.

Claims

1. A single Beidou terminal precise single point positioning enhancement method based on NLOS recognition, characterized in that: The steps include: S1), obtaining satellite GNSS measurement data and ephemeris data for multiple locations or scenes, and a sky map of the current location; S2), constructing a satellite sky map based on the satellite data obtained in step S1); S3) constructing a NLOS signal recognition model based on graph neural network; and identifying the satellite sky map through the pre-trained NLOS signal recognition model to obtain the satellite signal classification result; S4), constructing a random model based on the satellite signal classification results, and obtaining the variance of the pseudorange and phase observation values ​​of each satellite by inputting all satellites of a single epoch into the random model; S5) Calculate the user's position based on the GNSS precise ephemeris, precise clock error, differential code deviation, and GNSS measurement data and ephemeris data.

2. The method for enhancing single Beidou terminal precise point positioning based on NLOS recognition according to claim 1 is characterized in that: In step S2), a satellite sky map is constructed, which specifically includes the following steps: S21), selecting satellite node features, wherein the satellite node features include elevation angle, azimuth angle, carrier-to-noise ratio and pseudorange residual; S22), constructing an adjacency matrix based on neighborhood attributes and constellation attributes, connecting satellite nodes at the same time; S23), selecting edge features, wherein the edge features include relative positions of elevation angle and azimuth angle differences, relative signal quality between carrier-to-noise ratio and pseudorange residual, and spherical distance; S24), combining the node features, the adjacency matrix and the edge feature vector to construct a sky satellite map.

3. The method for enhancing single Beidou terminal precise point positioning based on NLOS recognition according to claim 1 is characterized in that: In step S3), the NLOS signal recognition model based on graph neural network includes a multi-layer graph attention-sparse adaptation layer and a classifier. The graph attention-sparse adaptation layer calculates the attention weights between satellite nodes through a multi-head attention mechanism, captures the complex nonlinear relationship between nodes, updates satellite features, and outputs the probability of the satellite being NLOS through the classifier.

4. The method for enhancing single Beidou terminal precise point positioning based on NLOS recognition according to claim 3 is characterized in that: In step S3), the graph attention-sparse adaptation layer includes a graph attention module and a sparse adapter.

5. The method for enhancing single Beidou terminal precise point positioning based on NLOS recognition according to claim 4 is characterized in that: In step S3), for a given node feature set The graph attention module calculates the multi-head attention between satellite nodes according to the adjacency matrix A; the calculation of the m-th head attention is as follows: in, The query vector representing the i-th target satellite of the m-th head attention of the l-th network layer; The key vector representing the jth neighboring satellite of the mth head attention of the lth network layer; The value vector representing the jth neighboring satellite of the mth head attention of the lth network layer; The weight vectors representing the query vector, key vector, and value vector of the m-th head attention of the l-th network layer respectively; They represent the learnable biases of the query vector, key vector, and value vector of the m-th head attention of the l-th network layer respectively; represents the node characteristics of the i-th target satellite at the l-th network layer; Node characteristics of the jth neighboring satellite at the lth network layer; Through a learnable parameter matrix and a learnable bias Convert the node features of the central satellite into a query vector Then, use different trainable parameters and Convert node features of neighboring satellites into key vectors Sum value vector 6. The method for enhancing single Beidou terminal precise point positioning based on NLOS recognition according to claim 5 is characterized in that: In step S3), in order to capture the relationship between the satellite and its adjacent satellites, the attention score is obtained by calculating the dot product of the query vector and the key vector The output is then represented as a weighted sum of the value vectors based on the attention weights, and the information from multiple heads is fused together for message passing, i.e.: The attention score obtained With the learned value vector Multiply and then concatenate the information from all heads to obtain the environment representation of the central node in, The environment representation of the central node; represents the attention scores of the i-th target satellite and the j-th neighboring satellite of the m-th head attention of the l-th network layer; T represents the transpose of the matrix; softmax represents the activation function; d represents the hidden dimension of each attention layer, || represents the tensor concatenation operation; M is the number of heads; N(i) represents the set of neighboring nodes of the central node, and the adjacency matrix A containing edge connection information must be taken as an additional input; Then, by taking the satellite features h of the previous layer l Environment representation with central node Add together to get the updated satellite characteristics h l+1 .

7. The method for enhancing single Beidou terminal precise point positioning based on NLOS recognition according to claim 6 is characterized in that: In step S3), the representation of the relationship between satellites is adjusted by projecting the edge features into the latent space and adding them to the attention score. Specifically: The edge features are first projected into a latent space, namely: in, Represents the implicit edge features learned by the encoding layer; is the linear projection matrix, if l=1, k=5; otherwise, k=d; ReLu represents the nonlinear activation function, which ensures that the output of the model is positive; encoding represents the encoding operation; represents the implicit edge features of the i-th target satellite and the j-th adjacent satellite learned by the encoding layer at the l-th layer; Then, the implicit edge feature Adding this to the attention score, we get: in, is a learnable matrix that transforms The projection is the weight value, and || means concatenating the outputs of M heads; Finally, the node and edge features are fused through the graph attention module, namely: In the formula, H and Z represent node features and edge features respectively; GNE-Attention represents the graph attention module, and ER represents the output of the graph attention module.

8. The method for enhancing single Beidou terminal precise point positioning based on NLOS recognition according to claim 7 is characterized in that: The sparse adapter is represented by the environment of the central node Select useful environmental information to update the satellite features of layer l+1; By learning the weight vector to filter key environmental representations, the impact of input data noise and redundant information is reduced, so that the satellite can adaptively select useful environmental information according to its own characteristics, improve the availability of environmental representation, and enhance the generalization performance of the model. The details are as follows: in, represents the weight vector of the upper and lower projections of the i-th target satellite in the l-th network layer; GeLU represents the nonlinear activation function; sigmoid represents the activation function; Respectively represent the learnable weight matrices of the upper and lower projections of the lth network layer; Compress the central satellite into a low-dimensional space to obtain Then an up-projection matrix is ​​applied so that the weight vector is aligned with the environment representation of the center node Have the same shape; use nonlinear activation function GeLU to reduce noise interference; By adding the weight vector Applied to the original environment representation ER to obtain the sparse environment representation SER, that is: Satellite characteristics of layer l+1 The updates are as follows: in, is a learnable residual connection matrix that is used to retain some information from the previous layer and prevent over-smoothing. In this way, the sparse adapter is derived from the environment representation of the central node. Useful environmental information is selected to update the satellite features of the l+1 layer.

9. The method for enhancing single Beidou terminal precise point positioning based on NLOS recognition according to claim 1, characterized in that: In step S4), a random model is constructed based on the satellite signal classification results, which specifically includes the following steps: S41), classify the satellites according to the NLOS identification results, and obtain the number of NLOS satellites N NLOS 、Number of LOS satellites LOS ; S42), calculate the position precision loss value PDOP of all LOS satellites; if the PDOP value of the LOS satellite is less than the first preset value; then the variance of all NLOS satellites is *A; otherwise, the variance of all NLOS satellites is *B; S43), if the PDOP value of the LOS satellite is less than the second preset value; then the variance of all satellites is *C; otherwise, the variance of all NLOS satellites is *D.

10. The method for enhancing single Beidou terminal precise point positioning based on NLOS recognition according to claim 1, characterized in that: In step S5), the non-difference non-combined PPP algorithm is used to calculate the user position, as follows: In the formula, r and s represent the receiver and satellite respectively, represents the original pseudorange observation value, represents the original observed value of the phase; represents the distance between the satellite and the phase center of the receiver antenna; c represents the speed of light; dt r and dt s Represent the receiver and satellite precise clock errors, T r represents the tropospheric delay error, represents the ionospheric delay error; λ represents the wavelength; represents the phase ambiguity; represents pseudorange multipath; represents phase multipath; represents the influence of pseudorange observation noise and other unmodeled errors; represents the effect of phase observation noise and other unmodeled errors.

Images