Satellite positioning method and system for urban complex environment and deployment method

Abstract

The invention provides a satellite positioning method, a satellite positioning system and a satellite positioning deployment method for an urban complex environment, and the method comprises the steps: integrating the local structure information of a map into an original attention score through a satellite positioning correction model, and extracting more comprehensive environment features from a satellite observation map; a position correction value with higher precision is obtained; a weighted confrontation weight disturbance robust enhancement strategy is adopted to train the model, and disturbance factors are introduced during model optimization, so that the model can effectively find flat minimization, the influence of extreme input samples is avoided, and meanwhile, the environment generalization performance of the model is enhanced; a weight mixed soft pruning method is used for carrying out lightweight processing on the model, so that the reasoning speed at an embedded end is increased, and the model memory is reduced; in the deployment process, a method based on data-driven loose coupling is adopted to ensure that the model can quickly correct the initial position information, so that the positioning precision is improved, and the method is suitable for various chips.

Description

Technical Field

[0001] The present invention relates to the field of satellite positioning technology, and in particular to a satellite positioning method, system and deployment method for complex urban environments. Background Art

[0002] The Global Navigation Satellite System (GNSS) is one of the main methods currently widely used in positioning technology. In practical applications, GNSS measurement scenarios can be mainly divided into open environments (such as suburban highways) and urban environments (such as urban forests, urban canyons, and under viaducts).

[0003] In open environments, existing model-based methods have been able to achieve good positioning accuracy. However, in complex urban environments such as urban canyons and overpasses, GNSS signals are often affected by surrounding environmental factors, which makes them prone to interference. When the satellite signals reach the receiver, they are blocked and reflected multiple times, which is called the multipath effect. This causes a lot of noise in the GNSS signals received by the receiver. Traditional GNSS positioning methods rely on the receiver to receive signals from at least four satellites, and determine the position of the receiver by measuring the arrival time and combining triangulation. However, this method has extremely low accuracy in complex urban scenarios. Although mobile GNSS receivers such as smartphones have made great progress in accuracy and reliability, their receiver chips and receiving antennas are of lower quality than traditional mainstream GNSS signal receivers, resulting in more outliers in the measurements received by the mobile phone. In such harsh environments, the positioning accuracy of smartphones will be further reduced.

[0004] In order to solve the above technical problems, the existing technology has made improvements in the application and software. For example, in terms of hardware, the existing technology reduces the positioning error rate by adding dual-frequency pseudorange and carrier phase signal functions in the receiver and improving the receiver signal quality, but the improvement in positioning accuracy is not significant. In terms of software, with the widespread application of machine learning in the field of positioning in recent years, model-based or data-driven post-processing methods have improved the positioning accuracy of GNSS. In various mobile receiver positioning, model-based methods have been widely used, such as Kalman filtering (KF) in vehicle-mounted RTK and smartphones. These methods perform well in open environments and can provide high-precision positioning. However, in urban environments, complex electromagnetic interference and signal shielding problems make it difficult to completely eliminate observation errors through modeling, resulting in deficiencies in positioning accuracy and stability of these model methods. In contrast, data-driven deep learning methods can mine the impact of complex random noise on positioning through training on a large amount of data, solving the problem of unmodeled noise that cannot be handled by traditional mathematical model methods. In the absence of prior knowledge, these methods require a large amount of data for training. However, when there is a large deviation between the distribution of training data and test data, the positioning accuracy may be low due to overfitting of the training data and failure to adapt to new changes in data. In addition, existing data-driven methods usually only consider the values ​​of satellite measurements, while ignoring the location information of satellite distribution in the sky, and fail to fully explore the spatial structural relationship between different satellite measurements, thus affecting positioning accuracy.

[0005] In addition, the industry currently lacks modules that can provide high-precision positioning in complex urban environments. Most existing modules use model-based methods, and data-driven methods with higher positioning performance are still at the theoretical testing level. The actual application scenarios are not considered in model design, so they generally have problems such as large number of model parameters and long inference time. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a satellite positioning method, system and deployment method for complex urban environments. The present invention corrects the preliminary positioning results to obtain higher-precision positioning, and uses weighted hybrid soft pruning to lightweight the model to reduce the complexity of model reasoning and reduce the model's software and hardware requirements.

[0007] The technical solution of the present invention is: a satellite positioning method for complex urban environments, comprising the following steps:

[0008] S1), receiving multi-frequency and multi-constellation satellite data in a real urban scene, and performing initial filtering and screening processing on the data to obtain pre-processed satellite measurement data;

[0009] S2), predicting the preprocessed satellite measurement data to obtain an initial positioning estimation result;

[0010] S3), extracting satellite measurement features from the preprocessed satellite measurement data and constructing a satellite observation map;

[0011] S4), building a satellite positioning correction model and training it;

[0012] S5) The satellite positioning correction model uses the graph structure-aware attention mechanism to simultaneously learn the positional relationship between GNSS satellite measurement data and different satellite nodes in the satellite observation map to extract the optimal environment representation and then output the position correction result;

[0013] S6) Combining the initial positioning estimation result with the position positioning correction result to obtain the final positioning result.

[0014] Preferably, in step S1), after the satellite data is preliminarily filtered, the preliminarily filtered satellite data is screened by elevation angle and carrier-to-noise ratio.

[0015] Preferably, in step S2), the weighted least squares (WLS KF) combined Kalman filter is used to predict the initial positioning estimation result.

[0016] Preferably, in step S3), the pseudorange residual PR and the line-of-sight vector LOS from each satellite to the receiver are calculated based on the initial positioning estimation result and the satellite observation value, and are spliced ​​with the satellite carrier-to-noise ratio C / N0 to obtain the satellite measurement characteristics.

[0017] Preferably, in step S3), the satellite positions in the sky are used as nodes, the satellite measurement features are used as node features, and edges are constructed between the nodes of the satellite observation map based on the cosine similarity values ​​between different satellite measurement feature vectors.

[0018] Preferably, in step S4), the satellite positioning correction model adopts a graph transformer satellite image structure learning network based on a graph neural network GNN, including a subgraph structure extractor based on the graph neural network GNN, a structure-aware attention layer, and a feedforward network layer. The subgraph structure extractor extracts local structural information of each node in the satellite observation image, and integrates the extracted local structural information into the original attention score through the structure-aware attention layer to generate a more expressive node representation by considering the similarity of node attributes at different distances.

[0019] Preferably, in step S4), a training enhancement strategy based on weighted adversarial weighted perturbation is adopted to train the satellite positioning correction model. During training, a perturbation factor is introduced and the perturbation loss is calculated. The perturbation factor is then weighted with the original loss to obtain the final weighted loss of the current training batch, and then back-propagation is performed to obtain model parameters with stronger generalization and robustness.

[0020] Preferably, in step S5), local structure information is extracted from the satellite observation image by using a satellite positioning correction model, specifically:

[0021] S511), perform k-hop sampling on each node i in the satellite observation graph input into the satellite positioning correction model, and obtain a neighborhood node set S with a distance of k edges from the node i. i k ;

[0022] S512), aggregate and update S through the graph neural network GNN aggregation mechanism i k Subgraph node information of ;

[0023] S513), using sum pooling to obtain the graph representation of the subgraph and update it to the subgraph representation of node i.

[0024] Preferably, in step S5), after obtaining the local structure information, the graphic representation of the satellite observation image is updated by the structure-aware attention score.

[0025] Preferably, in step S5), the kernel-smoothed structure-aware attention layer integrates the extracted local structure information into the original attention score to generate a more expressive node representation by considering the similarity of node attributes at different distances.

[0026] Preferably, in step S6), after obtaining the structure-aware attention score, the position correction value is obtained through the fully connected layer and the average pooling layer. And estimate Pos based on the initial position Init Combined to get the corrected positioning result Pos Pred ,Right now:

[0027]

[0028] Preferably, the present invention further provides a satellite positioning system for complex urban environments, comprising:

[0029] The acquisition module is used to collect multi-frequency and multi-constellation satellite data in real urban scenes;

[0030] A pre-processing module is used to perform preliminary filtering and screening on multi-frequency and multi-constellation satellite data;

[0031] A preliminary positioning prediction module is used to predict the initial positioning estimation result of the receiver;

[0032] Satellite observation map construction module, used to extract satellite measurement features and construct satellite observation maps;

[0033] The graph feature extraction module is used to extract the local structural information of the satellite observation image and integrate the extracted local structural information into the original attention score to generate a more expressive node representation by considering the similarity of node attributes at different distances, and then output the position correction value;

[0034] The prediction module is used to combine the initial positioning estimation result and the position correction value to obtain the final corrected positioning result.

[0035] Preferably, the satellite observation map construction module uses satellite positions in the sky as nodes and satellite measurement features as node features, and constructs edges between nodes based on cosine similarity values ​​between different satellite measurement feature vectors.

[0036] Preferably, the graph feature extraction module adopts a graph transformer satellite image structure learning network based on a graph neural network GNN, including a subgraph structure extractor based on the graph neural network GNN, a structure-aware attention layer, and a feedforward network layer. The local structure information of each node in the satellite observation image is extracted through the subgraph structure extractor, and the extracted local structure information is integrated into the original attention score through the structure-aware attention layer to generate a more expressive node representation by considering the similarity of node attributes at different distances.

[0037] Preferably, a training enhancement strategy based on weighted adversarial weight perturbation is adopted to train the graph transformer satellite image structure learning network based on the graph neural network GNN.

[0038] Preferably, the present invention further provides a method for deploying a satellite positioning system, comprising the following steps:

[0039] A1) Compress the model size based on weighted hybrid soft pruning technology;

[0040] A2) Deploy the satellite positioning correction model to the microprocessor MCU of the positioning module in a loosely coupled manner, and write an application program for the satellite positioning correction model according to the data flow direction of the positioning process.

[0041] The beneficial effects of the present invention are:

[0042] 1. The satellite positioning correction model provided by the present invention integrates the local structural information of the image into the original attention score, helping the image transformer to extract more comprehensive environmental features from the satellite image and obtain a higher-precision position correction value;

[0043] 2. The present invention adopts the weighted adversarial weight perturbation (W-AWP) robust enhancement strategy and introduces a perturbation factor during model optimization, so that the model can effectively find a flat minimum, avoid the influence of extreme input samples, and enhance the environmental generalization performance of the model;

[0044] 3. The present invention uses a weighted hybrid soft pruning method to lightweight the model, thereby improving the inference speed at the embedding end and reducing the model memory requirement;

[0045] 4. During the deployment process, the present invention adopts a model-based and data-driven loosely coupled method to ensure that the model can quickly correct the position information output by the initial model-based positioning module, thereby improving the positioning accuracy and being applicable to a variety of chips. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of the method of the present invention;

[0047] Figure 2 This is a flow chart of satellite positioning correction model deployment in Example 3 of the present invention;

[0048] Figure 3 This is a flow chart of the satellite positioning correction model inference application function (API) operation in Example 3 of the present invention;

[0049] Figure 4 This is a schematic diagram of the process of satellite positioning correction model positioning in embodiment 3 of the invention. DETAILED DESCRIPTION

[0050] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings:

[0051] Example 1

[0052] like Figure 1 As shown, this embodiment provides a satellite positioning method for complex urban environments, comprising the following steps:

[0053] S1), receiving multi-frequency and multi-constellation satellite data in a real urban scene, and performing initial filtering and screening processing on the data to obtain pre-processed satellite measurement data;

[0054] In actual situations, this embodiment receives and demodulates satellite signals through a receiver to extract information such as pseudorange, carrier phase, and satellite position; then initial filtering is performed to simply filter out erroneous signals.

[0055] In addition, in this embodiment, after the satellite data is preliminarily filtered, the preliminarily filtered satellite data is screened by elevation angle and carrier-to-noise ratio to obtain satellite measurement data with higher quality.

[0056] S2), predicting the preprocessed satellite measurement data to obtain an initial positioning estimation result;

[0057] This embodiment uses a weighted least squares (WLS KF) combined with a Kalman filter to predict the initial positioning estimation result.

[0058] S3), extracting satellite measurement features from the preprocessed satellite measurement data, and constructing a satellite observation map based on the extracted satellite measurement features;

[0059] The satellite measurement feature extraction is as follows: the pseudorange residual PR and the line-of-sight vector LOS from each satellite to the receiver are calculated according to the initial positioning estimation result and the satellite observation value, and the satellite measurement features are obtained by splicing with the satellite carrier-to-noise ratio C / N0, and then the satellite observation map is constructed according to the extracted satellite measurement features.

[0060] This embodiment uses satellite positions in the sky as nodes, satellite measurement features as node features, and constructs edges between nodes based on cosine similarity values ​​between different satellite measurement feature vectors; specifically:

[0061] The similarity between different satellites is calculated based on cosine similarity. The calculation formula for the similarity between satellites is:

[0062]

[0063] In the formula, s i and j are the measurement characteristics of the i-th and j-th satellites respectively; η represents s i and j similarity is the cosine similarity function; where the satellite measurement characteristics s = [PR, LOS, C / N0]; PR is the pseudorange residual; C / N0 is the carrier signal-to-noise ratio; LOS is the unitized line-of-sight vector.

[0064] S4), building a satellite positioning correction model and training it;

[0065] In this embodiment, the satellite positioning correction model adopts a graph transformer satellite image structure learning network based on the graph neural network GNN, including a subgraph structure extractor based on the graph neural network GNN, a structure-aware attention layer, and a feedforward network layer. The subgraph structure extractor extracts the local structural information of each node in the satellite observation image, and the extracted local structural information is integrated into the original attention score through the structure-aware attention layer to generate a more expressive node representation by considering the similarity of node attributes at different distances.

[0066] In practical applications, data distribution often deviates greatly due to regional environmental differences between different cities and between urban canyons and highways in the same city. In addition, in some scenes with severe occlusion, satellite measurements may be strongly disturbed. These factors will lead to the training of low generalization and low robustness models. In order to solve the above technical problems, this embodiment adopts a training enhancement strategy based on weighted adversarial weight perturbation to train the satellite positioning correction model. During training, a perturbation factor is introduced and the perturbation loss is calculated. The final weighted loss of the current training batch is then weighted with the original loss and back-propagated to obtain model parameters with stronger generalization and robustness. Specifically, the following steps are included:

[0067] S41) Calculate the batch training loss during training train ;

[0068] S42), the weight w of the current satellite positioning correction model train Copied to the proxy model, the proxy model calculates the loss and updates the weight w p , and then calculate the weight difference w′ between the proxy model and the current satellite positioning correction model to obtain the data disturbance, and use the disturbance intensity factor K to weight the weight difference w′ to obtain the final adversarial weight disturbance w awp ; The final confrontation right disturbance w awp It is expressed as:

[0069] w awp =K·w′=K·(w p -w train );

[0070] S43) Disturb the confrontation right w awp Add to the original satellite positioning correction model weights and calculate the disturbance loss loss awp ;

[0071] S44) Get the final loss loss of the current batch based on the training loss and perturbation loss all And update the satellite positioning correction model parameters, the final loss loss all Expressed as:

[0072] loss all =(1-λ)·loss train +λ·loss awp ;

[0073] In the formula, λ represents the weight ratio of perturbation loss when calculating the total loss; loss train Represents batch training loss; loss awp represents the disturbance loss.

[0074] S5) The satellite positioning correction model uses the graph structure-aware attention mechanism to simultaneously learn the positional relationship between GNSS satellite measurement data and different satellite nodes in the satellite observation map to extract the optimal environment representation and then output the position correction result;

[0075] The satellite positioning correction model of this embodiment can combine local and global information when aggregating information to extract a more comprehensive environmental representation from the satellite measurement map.

[0076] This embodiment first extracts the local structure information of each node from the satellite observation image by using a subgraph structure extractor, which specifically includes the following steps:

[0077] S511), perform k-hop sampling on each node i in the satellite observation graph input into the satellite positioning correction model, and obtain a set of neighboring nodes with a distance of k edges from the node i.

[0078] S512), aggregate and update through the graph neural network GNN aggregation mechanism The subgraph node information, namely:

[0079]

[0080] In the formula, represents the node representation of node i in the kth layer of GNN; obtained by the nonlinear activation function σ, is the representation of the neighborhood node u in the k-1 layer GNN; AGGREGATE k is the aggregator at level k; is the representation of the k-1 layer GNN of node i; CONCAT is the connection function; it is used to connect the obtained neighborhood representation and the representation of node i in the previous layer; W k is the weight matrix of the k-th layer GNN;

[0081] S513), using sum pooling to obtain the graph representation of the subgraph and update it to the subgraph representation of the central node i, that is:

[0082]

[0083] In the formula, x i =φ(i,G) is the node representation of node i for its subgraph G; N k (i) indicates that node i contains its own k-hop neighborhood; φ(i,G) indicates the subgraph calculation function; Represents the graph neural network function used to extract subgraphs at the kth layer.

[0084] After obtaining the local structural information of the satellite observation image, this embodiment calculates the attention score through the structure-aware attention layer SA-attention based on kernel smoothing, and then integrates the extracted local structural information into the original attention score to generate a more expressive node representation by considering the similarity of node attributes at different distances. For node v, its structure-aware attention score SA-attn(v) is:

[0085]

[0086] Where f(x) = W v x is the linearization function, V is the node set of the satellite observation graph; κ exp For R d ×R d The asymmetric exponential kernel on R d The dimension of the vector representing the input exponential kernel; through the learnable parameter W Q and W K Linearization gives: x v 、x u They are the node representations of nodes v and u for their subgraph G respectively.

[0087]

[0088] Among them, x and x′ are different node features in the satellite observation graph after subgraph extraction, W Q and W K , W V are learnable weight parameters.

[0089] S6) Combining the initial positioning estimation result with the position positioning correction result to obtain the final positioning result.

[0090] After obtaining the structure-aware attention score, the position correction value is obtained through the fully connected layer and the average pooling layer. And estimate Pos based on the initial position Init Combined to get the corrected positioning result Pos Pred ,Right now:

[0091]

[0092] Example 2

[0093] This embodiment provides a satellite positioning system for complex urban environments, including:

[0094] The acquisition module is used to collect multi-frequency and multi-constellation satellite data in real urban scenes;

[0095] A pre-processing module is used to perform preliminary filtering and screening on multi-frequency and multi-constellation satellite data;

[0096] A preliminary positioning prediction module is used to predict the initial positioning estimation result of the receiver;

[0097] Satellite observation map construction module, used to extract satellite measurement features and construct satellite observation maps;

[0098] The graph feature extraction module is used to extract the local structural information of the satellite observation image and integrate the extracted local structural information into the original attention score to generate a more expressive node representation by considering the similarity of node attributes at different distances, and then output the position correction value;

[0099] The prediction module is used to combine the initial positioning estimation result and the position correction value to obtain the final corrected positioning result.

[0100] In this embodiment, the preliminary positioning prediction module uses a weighted least squares (WLS KF) combined Kalman filter method to predict the initial positioning estimation result.

[0101] As a preferred embodiment of the present invention, the satellite observation map construction module uses satellite positions in the sky as nodes and satellite measurement features as node features, and constructs edges between nodes based on cosine similarity values ​​between different satellite measurement feature vectors.

[0102] As a preferred embodiment of this embodiment, the graph feature extraction module adopts a graph transformer satellite image structure learning network based on the graph neural network GNN, including a subgraph structure extractor based on the graph neural network GNN, a structure-aware attention layer, and a feedforward network layer. The satellite observation image is extracted through the subgraph structure extractor to extract local structural information of each node, and the extracted local structural information is integrated into the original attention score through the structure-aware attention layer to generate a more expressive node representation by considering the similarity of node attributes at different distances.

[0103] As a preferred embodiment of this embodiment, a training enhancement strategy based on weighted adversarial weight perturbation is adopted to train the graph transformer satellite image structure learning network based on the graph neural network GNN.

[0104] Example 3

[0105] This embodiment provides a method for deploying a satellite positioning system, including the following steps:

[0106] A1) Compressing the size of the satellite positioning correction model based on weighted hybrid soft pruning technology; specifically including the following steps:

[0107] A11), set the sparse target S of the satellite positioning correction model t , sparse target S tRepresents the proportion of weights that need to be pruned in the network layer weight matrix W of the model;

[0108] A12), use the lowest norm l2 as the standard to prune the weights, and define the weight threshold as W t , below the weight threshold W t The weights of will be pruned;

[0109] A13) For the network layer L i , define the initial weight number N orig and the number of sparse weights after pruning N kept They are:

[0110]

[0111] Where, d (i) Indicates L i The input feature dimension is Indicates L i The output feature dimension of S t For sparse targets;

[0112] A14), refer to the minimum norm l2 size, and sort according to the sparse target to obtain the maximum weight value W that needs to be pruned h and the minimum weight value W that needs to be retained l ,Right now:

[0113]

[0114] Where W is the network layer weight matrix; N kept The number of sparse weights after pruning; N orig is the number of initial weights; topk represents the sorting function, largest represents the sorting method used; largest = true means taking the largest N from W kept elements, largest = false means taking the smallest N from W orig -N kept elements;

[0115] Then the weight threshold is obtained as:

[0116]

[0117] A15), sparse training is performed on the trained satellite positioning correction model to achieve weight mixing, and the weights of the trained satellite positioning correction model are gradually integrated to the target value 0, thereby obtaining a pruned sparse training model.

[0118] In this embodiment, in sparse training, satellite measurement features are used as training sets, wherein the satellite measurement features include pseudorange residuals (Pseudorange Residuals, PR), carrier signal-to-noise ratio (C / N0), and LOS vectors from the satellite to the receiver; wherein the pseudorange residual refers to the difference between the pseudorange value actually measured and the expected value, which can reflect potential errors in measurement or positioning estimation; the carrier-to-noise ratio reflects the quality of the received signal, and can provide the learning module with an indicator that can measure the reliability of the signal; the LOS vector refers to the unit direction vector from the receiver to the satellite, which contains information such as the satellite's altitude angle and azimuth angle.

[0119] The satellite positioning correction model is trained on the server side. The training configuration used in this embodiment is: a high-performance RTX3090 graphics card with 24G memory; the proposed positioning correction model is implemented based on the artificial intelligence framework PyTorch, and hardware acceleration is achieved using the 11.6 version of CUDA. During the training process, the training gradient loss Loss is used as the indicator of training completion. After 100 iterations of Loss tends to be stable, the training is terminated and the model weight file with the suffix .pth is saved;

[0120] In addition, in sparse training, the weight factor α is introduced i To adjust the weights that need to be pruned; in the pruning training parameter update phase, the original, non-parametric weights are used to adjust the weights w that need to be pruned to the target value, that is:

[0121]

[0122] In the formula, w new represents the weight updated by sparse training, w t represents the weight threshold; α i is the weight factor; w represents the weight value before pruning in the current network layer weight matrix;

[0123] Among them, the weight factor α i The expression is:

[0124]

[0125] In the formula, E i represents the current training round, E init represents the number of iterations of normal training, E knee Indicates the number of iterations of sparse training.

[0126] A2) The satellite positioning correction model is implemented in a loosely coupled manner.

[0127] Deploy it to the microprocessor MCU of the positioning module, and write an application for the satellite positioning correction model according to the data flow direction of the positioning process.

[0128] The microprocessor MCU used in this embodiment is the Rockchip RK3568 chip. First, the satellite positioning correction model needs to be transplanted to the RK3568 chip. Considering that the chip cannot directly run Python scripts, the satellite positioning correction model under the PyTorch framework needs to be converted into the RKNN framework supported by RK3568. Specifically, the satellite positioning correction model is first converted to the multi-platform compatible ONNX framework, and then the RKNN tool chain provided by Rockchip is used to further convert the ONNX framework into an RKNN model callable by RK3568. The Huawei Cloud ModelBox framework is used to encapsulate the inference model on the embedded end, and then the model is written on the embedded end to call the API function to call the satellite positioning correction model to perform position correction result inference. Finally, the positioning correction APP application is written according to the data flow of the positioning correction process. The transplantation process of the satellite positioning correction model and the calling reasoning process on the RK3568 platform are described in the process. Figure 2 and Process Figure 3 Given, the data flow used for reference in writing the positioning correction application APP is as follows Figure 4 Given.

[0129] 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 satellite positioning method for complex urban environments, characterized in that: The following steps are involved: S1), receiving multi-frequency and multi-constellation satellite data in a real urban scene, and performing initial filtering and screening processing on the data to obtain pre-processed satellite measurement data; S2), predicting the preprocessed satellite measurement data to obtain an initial positioning estimation result; S3), extracting satellite measurement features from the preprocessed satellite measurement data and constructing a satellite observation map; S4), building a satellite positioning correction model and training it; S5) The satellite positioning correction model uses the graph structure-aware attention mechanism to simultaneously learn the positional relationship between GNSS satellite measurement data and different satellite nodes in the satellite observation map to extract the optimal environment representation and then output the position correction result; S6) Combining the initial positioning estimation result with the position positioning correction result to obtain the final positioning result.

2. The satellite positioning method for complex urban environments according to claim 1, characterized in that: The weighted least squares (WLS) KF combined Kalman filter is used to predict the initial positioning estimation result.

3. The satellite positioning method for complex urban environments according to claim 1, characterized in that: In step S4), the satellite positioning correction model adopts a graph transformer satellite image structure learning network based on a graph neural network GNN, including a subgraph structure extractor based on the graph neural network GNN, a structure-aware attention layer, and a feedforward network layer. The subgraph structure extractor extracts the local structural information of each node in the satellite observation image, and the extracted local structural information is integrated into the original attention score through the structure-aware attention layer to generate a more expressive node representation by considering the similarity of node attributes at different distances.

4. The satellite positioning method for complex urban environments according to claim 3, characterized in that: In step S4), a training enhancement strategy based on weighted adversarial weight perturbation is used to train the satellite positioning correction model. During training, a perturbation factor is introduced and the perturbation loss is calculated. The perturbation factor is then weighted with the original loss to obtain the final weighted loss of the current training batch, and then back-propagation is performed to obtain model parameters with stronger generalization and robustness.

5. The satellite positioning method for complex urban environments according to claim 4, characterized in that: In step S5), local structure information is extracted from the satellite observation image through the satellite positioning correction model, specifically: S511), perform k-hop sampling on each node i in the satellite observation graph input into the satellite positioning correction model, and obtain a set of neighboring nodes with a distance of k edges from the node i. S512), aggregate and update through the graph neural network GNN aggregation mechanism The subgraph node information is: In the formula, represents the node representation of node i in the kth layer of GNN; obtained by the nonlinear activation function σ, is the representation of the neighborhood node u in the k-1 layer GNN; AGGREGATE k is the aggregator at level k; is the representation of the k-1 layer GNN of node i; CONCAT is the connection function; it is used to connect the obtained neighborhood representation and the representation of node i in the previous layer; W k is the weight matrix of the k-th layer GNN; S513), sum pooling is performed to obtain the graph representation of the subgraph and update it to the subgraph representation of the central node i, that is: In the formula, x i =φ(i,G) is the node representation of node i for its subgraph G; N k (i) indicates the k-hop neighborhood of node i including itself; φ(i,G) indicates the subgraph calculation function; Represents the graph neural network function used to extract subgraphs at the kth layer.

6. A satellite positioning method for complex urban environments according to claim 5, characterized in that: In step S5), after obtaining the local structure information, the image representation of the satellite observation image is updated by the structure-aware attention score; The kernel-smoothed structure-aware attention layer integrates the extracted local structural information into the original attention score and considers the similarity of node attributes at different distances to generate a more expressive node representation; for node v, its structure-aware attention score SA-attn(v) is: Where f(x) = W v x is the linearization function, V is the node set of the satellite observation graph; κ exp For R d ×R d The asymmetric exponential kernel on R d represents the dimension of the input exponential kernel vector; through the learnable parameter W Q and W K Linearization gives: x v 、x u They are the node representations of nodes v and u for their subgraph G respectively.

7. The satellite positioning method for complex urban environments according to claim 6, characterized in that: In step S6), after obtaining the structure-aware attention score, the position correction value is obtained through the fully connected layer and the average pooling layer And estimate Pos based on the initial position Init Combined to get the corrected positioning result Pos Pred ,Right now:

8. A satellite positioning system for complex urban environments, characterized in that: The system is used for the method described in any one of claims 1 to 7, and the system comprises: The acquisition module is used to collect multi-frequency and multi-constellation satellite data in real urban scenes; A pre-processing module is used to perform preliminary filtering and screening on multi-frequency and multi-constellation satellite data; A preliminary positioning prediction module is used to predict the initial positioning estimation result of the receiver; Satellite observation map construction module, used to extract satellite measurement features and construct satellite observation maps; The graph feature extraction module is used to extract the local structural information of the satellite observation image and integrate the extracted local structural information into the original attention score to generate a more expressive node representation by considering the similarity of node attributes at different distances, and then output the position correction value; The prediction module is used to combine the initial positioning estimation result and the position correction value to obtain the final corrected positioning result.

9. A method for deploying a satellite positioning system, characterized in that: The method is used to deploy the system according to claim 8 to a chip, and the deployment method comprises the following steps: A1) Compress the model size based on weighted hybrid soft pruning technology; A2) Deploy the satellite positioning correction model to the microprocessor MCU of the positioning module in a loosely coupled manner, and write an application program for the satellite positioning correction model according to the data flow direction of the positioning process.

10. The deployment method according to claim 9, characterized in that: Step A1) specifically includes the following steps: A11), set the sparse target S of the satellite positioning correction model t , sparse target S t Represents the proportion of weights that need to be pruned in the network layer weight matrix W of the model; A12), use the lowest norm l2 as the standard to prune the weights, and define the weight threshold as W t , below the weight threshold W t The weights of will be pruned; A13) For the network layer L i , define the initial weight number N prig and the number of sparse weights after pruning N kept They are: Where, d (i) Indicates L i The input feature dimension is Indicates L i The output feature dimension of S t For sparse targets; A14), refer to the minimum norm l2 size, and sort according to the sparse target to obtain the maximum weight value W that needs to be pruned h and the minimum weight value W that needs to be retained l ,Right now: Where W is the network layer weight matrix; N kept The number of sparse weights after pruning; N orig is the number of initial weights; topk means sorting and taking the largest or smallest k items, largest means the sorting method used, and largest=true means taking the largest N from W kept elements, largest = false means taking the smallest N from W orig -N kept elements; Then the weight threshold is obtained as: A15), performing sparse training on the trained satellite positioning correction model to achieve weight mixing, and gradually integrating the weight of the trained satellite positioning correction model to a target value of 0, thereby obtaining a pruned sparse training model; In sparse training, the weight factor α is introduced i To adjust the weights that need to be pruned; in the pruning training parameter update phase, the original, non-parametric weights are used to adjust the weights w that need to be pruned to the target value, that is: In the formula, w new represents the weight updated by sparse training, e t represents the weight threshold; α i is the weight factor; w represents the weight value before pruning in the current network layer weight matrix; Among them, the weight factor α of the i-th training round is i The expression is: In the formula, E i represents the i-th round of training, E init represents the number of iterations of normal training, E knee Indicates the number of iterations of sparse training.

Images