A satellite positioning method, system and deployment method for urban complex environments

Abstract

The present invention provides a satellite positioning method, system and deployment method for urban complex environments. The present invention integrates the local structure information of the graph into the original attention scores through a satellite positioning correction model to extract more comprehensive environmental features from the satellite observation graph, obtain a higher-precision position correction value, and adopts a weighted adversarial perturbation robust enhancement strategy to train the model. A perturbation factor is introduced during model optimization, enabling the model to effectively find a flat minimum and avoid the influence of extreme input samples, while enhancing the environmental generalization performance of the model. The method of weight-mixed soft pruning is used to lightweight the model, improving the inference speed at the embedding end and reducing the model memory. During the deployment process, a data-driven loose coupling method is adopted to ensure that the model can quickly correct the initial position information, thereby improving the positioning accuracy and being applicable to multiple chips.

Description

Technical Field

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

[0002] The Global Navigation Satellite System (GNSS) is one of the main methods 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, etc.).

[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, resulting in easy interference. When the signals sent by satellites reach the receiver, they will be blocked and undergo multiple reflections, which is called the multipath effect; this will cause a large amount 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 smart phones have made great progress in terms of accuracy and reliability, due to the lower quality of their receiver chips and receiving antennas compared to traditional mainstream GNSS signal receivers, the number of outliers in the measurements received by the mobile phone will be more, and the positioning accuracy of the smart phone will be further reduced in such a harsh environment.

[0004] To solve the above technical problems, the prior art has made improvements in hardware and software aspects. For example, in the hardware aspect, the prior art has added dual-frequency pseudorange and carrier phase signal functions to the receiver to reduce the positioning error rate by improving the receiver signal quality, but the improvement in positioning accuracy is not significant. In the software aspect, with the wide application of machine learning in the positioning field in recent years, the post-processing methods based on models or data-driven have improved the positioning accuracy of GNSS. Among various mobile receiver positionings, model-based methods have been widely applied, such as vehicle-mounted RTK and Kalman filtering (KF) in 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 the positioning accuracy and stability of these model methods. In contrast, data-driven deep learning methods can, through the training of a large amount of data, uncover the impact of complex random noise on positioning and solve the problem of unmodeled noise that cannot be handled by traditional mathematical model methods. These methods require a large amount of data for training in the absence of prior knowledge, but when there is a large deviation between the training data and the test data distribution, due to overfitting to the training data and being unable to adapt to new changing data, the positioning accuracy may be low. Moreover, existing data-driven methods usually only consider the numerical values of satellite measurements and ignore the position information of satellites distributed in the sky, failing to fully exploit the spatial structure relationship between different satellite measurements, thus affecting the positioning accuracy.

[0005] In addition, there is currently a lack of high-precision positioning modules for urban complex environments in the industry. Most existing modules adopt model-based methods, and data-driven methods with higher positioning performance are almost still at the theoretical test level. The actual application scenarios are not considered in the model design, so they generally have problems such as a large number of model parameters and long inference time. Summary of the Invention

[0006] Aiming at the deficiencies of the prior art, the present invention provides a satellite positioning method, system and deployment method for urban complex environments. The present invention corrects the preliminary positioning result to obtain a higher-precision position positioning, and performs lightweight processing on the model through weighted mixed soft pruning to reduce the model inference complexity and the software and hardware requirements of the model.

[0007] The technical solution of the present invention is as follows: A satellite positioning method for urban complex environments includes the following steps:

[0008] S1), receiving multi-frequency multi-constellation satellite data in a real urban scene, and performing initial filtering and screening processing on it to obtain preprocessed 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), constructing a satellite positioning correction model and training it;

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

[0013] S6), combining the initial positioning estimation result and the position positioning correction result to obtain a 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), according to the initial positioning estimation result and the satellite observation value, the pseudorange residual PR and the line-of-sight vector LOS from each satellite to the receiver are calculated, and are concatenated with the satellite carrier-to-noise ratio C / N0 to obtain satellite measurement features.

[0017] Preferably, in step S3), the positions of the satellites in the sky are used as nodes, and 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 graph 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 feed-forward network layer. The subgraph structure extractor extracts the local structure information of each node from the satellite observation map, and the structure-aware attention layer integrates the extracted local structure information into the original attention score to consider the node attribute similarity at different distances to generate a more expressive node representation.

[0019] Preferably, in step S4), a training enhancement strategy based on weighted adversarial weight perturbation is adopted to train the satellite positioning correction model. During training, a perturbation factor is introduced and the perturbation loss is calculated, and then it is weighted with the original loss to obtain the final weighted loss of the current training batch, and then backpropagation 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 map through the satellite positioning correction model, specifically:

[0021] S511): For each node i in the satellite observation map input to the satellite positioning correction model, k-hop sampling is performed to obtain a set S of neighborhood nodes at a distance of k edges from node i i k ;

[0022] S512): Aggregate and update the subgraph node information of S through the graph neural network GNN aggregation mechanism i k ;

[0023] S513): Use 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 graph representation of the satellite observation map is updated through the structure-aware attention score.

[0025] Preferably, in step S5), the structure-aware attention layer based on kernel smoothing integrates the extracted local structure information into the original attention score and generates a more expressive node representation by considering the node attribute similarity 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 combined with the initial position estimate Pos Init to obtain the corrected positioning result Pos Pred , that is:

[0027]

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

[0029] An acquisition module for acquiring multi-frequency and multi-constellation satellite data in real urban scenarios;

[0030] A preprocessing module for performing preliminary filtering and screening on the multi-frequency and multi-constellation satellite data;

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

[0032] The satellite observation map construction module is used to extract satellite measurement features and construct a satellite observation map;

[0033] The graph feature extraction module is used to extract the local structure information of the satellite observation map and integrate the extracted local structure information into the original attention score, consider the node attribute similarity at different distances, generate a more expressive node representation, and then output a 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 the positions of satellites in the sky as nodes and the measurement features of satellites as node features, and constructs edges between nodes based on the cosine similarity values between different satellite measurement feature vectors.

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

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

[0038] Preferably, the present invention also provides a deployment method for a satellite positioning system, including the following steps:

[0039] A1), Compressing the model size based on the weight mixing soft pruning technique;

[0040] A2), Deploying the satellite positioning correction model to the microprocessor MCU of the positioning module in a loosely coupled manner, and writing 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 can integrate the local structure information of the graph into the original attention score, helping the graph transformer extract more comprehensive environmental features from the satellite map; obtaining a higher-precision position correction value;

[0043] 2. The present invention adopts a weighted adversarial weight perturbation (W-AWP) robust enhancement strategy, introduces a perturbation factor during model optimization, enables the model to effectively find flat minimization, avoids the influence of extreme input samples, and simultaneously enhances the environmental generalization performance of the model;

[0044] 3. The present invention uses the method of weighted mixed soft pruning 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 method based on loose coupling of the model and data-driven 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 various chips. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is the process framework diagram of the method of the present invention;

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

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

[0049] Figure 4 is the process schematic diagram of the satellite positioning correction model positioning in Embodiment 3 of the invention. DETAILED DESCRIPTION OF THE INVENTION

[0050] The following further describes the specific embodiments of the present invention with reference to the drawings:

[0051] Embodiment 1

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

[0053] S1). Receive multi-frequency and multi-constellation satellite data in a real urban scene, and perform initial filtering and screening processing on it to obtain preprocessed satellite measurement data;

[0054] In actual situations, in this embodiment, the receiver receives satellite signals and demodulates them to extract information such as pseudorange, carrier phase, and satellite position; then the initial filtering simply filters out incorrect signals.

[0055] And in this embodiment, after the initial filtering process of the satellite data, the satellite data after the initial filtering is then screened through the elevation angle and carrier-to-noise ratio to obtain satellite measurement data with higher quality.

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

[0057] In this embodiment, the weighted least squares WLS KF combined Kalman filtering method is used to predict the initial positioning estimation result.

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

[0059] Among them, the extraction of the satellite measurement features is as follows: calculate the pseudorange residual PR and the line-of-sight vector LOS from each satellite to the receiver according to the initial positioning estimation result and the satellite observation value, splice them with the satellite carrier-to-noise ratio C / N0 to obtain the satellite measurement features, and then construct a satellite observation map based on the extracted satellite measurement features.

[0060] In this embodiment, the positions of the satellites in the sky are used as nodes, and the measurement features of the satellites are used as node features, and edges are constructed between the nodes based on the cosine similarity values between different satellite measurement feature vectors; specifically:

[0061] The cosine similarity is used to calculate the similarity between different satellites. The calculation formula for the similarity between satellites is:

[0062]

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

[0064] S4), construct a satellite positioning correction model and train it;

[0065] In this embodiment, the satellite positioning correction model adopts a graph transformer satellite map 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 feed-forward network layer. The subgraph structure extractor extracts the local structure information of each node from the satellite observation map, and the structure-aware attention layer integrates the extracted local structure information into the original attention score to consider the node attribute similarity at different distances and generate a more expressive node representation.

[0066] In practical applications, due to regional environmental differences between different cities, as well as between urban canyons and highways in the same city, there are often significant deviations in data distribution. In addition, in some scenarios with severe occlusion, satellite measurements may be strongly interfered with. These factors can lead to the training of models with low generalization and low robustness. 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, and then it is weighted with the original loss to obtain the final weighted loss of the current training batch, and then backpropagation is performed to obtain model parameters with stronger generalization and robustness. The specific steps are as follows:

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

[0068] S42), Copy the weight w of the current satellite positioning correction model train to the proxy model. The proxy model calculates the loss and updates the weight w p , then calculates the weight difference w' between the proxy model and the current satellite positioning correction model to obtain data perturbation, and at the same time uses the perturbation intensity factor K to weight the weight difference w' to obtain the final adversarial weight perturbation w awp ; The final adversarial weight perturbation w awp is expressed as:

[0069] w awp =K·w' = K·(w p -w train );

[0070] S43), Add the adversarial weight perturbation w awp to the original satellite positioning correction model weight and calculate the perturbation loss loss awp ;

[0071] S44), Obtain the final loss loss of the current batch according to the training loss and the perturbation loss all and update the satellite positioning correction model parameters. The final loss loss all is expressed as:

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

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

[0074] S5), the satellite positioning correction model uses the graph structure-aware attention mechanism to learn the position relationships between GNSS satellite measurement data and different satellite nodes in the satellite observation graph simultaneously, so as to extract the optimal environmental representation and then output the position positioning correction result;

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

[0076] In this embodiment, the subgraph structure extractor is first used to extract the local structure information of each node from the satellite observation graph. The specific steps are as follows:

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

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

[0079]

[0080] In the formula, represents the node representation of node i in the k-th layer of GNN; obtained by the non-linear activation function σ, is the representation of the neighborhood node u in the (k - 1)-th layer of GNN; AGGREGATE k is the aggregator of the k-th layer; is the representation of node i in the (k - 1)-th layer of GNN; CONCAT is the connection function; 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 of GNN;

[0081] S513), use 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) represents the k-hop neighborhood of node i including itself; φ(i, G) represents the subgraph calculation function; represents the graph neural network function used for subgraph extraction in the k-th layer.

[0084] After obtaining the local structure information of the satellite observation map, in this embodiment, the structure-aware attention layer SA-attention based on kernel smoothing is used to calculate the attention scores. Then, the extracted local structure information is integrated into the original attention scores to generate more expressive node representations by considering the node attribute similarities at different distances. For node v, its structure-aware attention score SA-attn(v) is as follows:

[0085]

[0086] In the formula, f(x) = W v x is a linearization function, V is the node set of the satellite observation map; κ exp is an asymmetric exponential kernel on R d ×R d ; R d represents the dimension of the vector input to the exponential kernel; it is linearly obtained through the learnable parameters W Q and W K ; x v and x u 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 map after subgraph extraction, and W Q and W K , W V are learnable weight parameters.

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

[0090] After obtaining the structure-aware attention scores, the position correction value is obtained through a fully connected layer and an average pooling layer and combined with the initial position estimate Pos[[ID=4,5]] Init to obtain the corrected positioning result Pos Pred , that is:

[0091]

[0092] Embodiment 2

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

[0094] An acquisition module for acquiring multi-frequency and multi-constellation satellite data in real urban scenes;

[0095] A preprocessing module for performing preliminary filtering and screening on the multi-frequency and multi-constellation satellite data;

[0096] A preliminary positioning prediction module for predicting the initial positioning estimation result of the receiver;

[0097] A satellite observation map construction module for extracting satellite measurement features and constructing a satellite observation map;

[0098] A graph feature extraction module for extracting local structural information of the satellite observation map and integrating the extracted local structural information into the original attention score to consider the node attribute similarity at different distances to generate a more expressive node representation, and then outputting a position correction value;

[0099] A prediction module for combining the initial positioning estimation result and the position correction value to obtain a final corrected positioning result.

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

[0101] Preferably in this embodiment, the satellite observation map construction module uses the positions of satellites in the sky as nodes and the measurement features of satellites as node features, and constructs edges between nodes based on the cosine similarity values between different satellite measurement feature vectors.

[0102] Preferably in this embodiment, the graph feature extraction module uses a graph transformer satellite map 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 feed-forward network layer. The subgraph structure extractor extracts the local structural information of each node of the satellite observation map, and the structure-aware attention layer integrates the extracted local structural information into the original attention score to consider the node attribute similarity at different distances to generate a more expressive node representation.

[0103] Preferably in this embodiment, a training enhancement strategy based on weighted adversarial weight perturbation is used to train the graph transformer satellite map structure learning network based on the graph neural network GNN.

[0104] Embodiment 3

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

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

[0107] A11), Setting the sparsity target S of the satellite positioning correction model t , the sparsity target S tThe proportion of weights to be pruned in the network layer weight matrix W of the representation model;

[0108] A12), Prune the weights using the lowest L2 norm as the criterion, and define the weight threshold as W t , and the weights lower than the weight threshold W t will be pruned;

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

[0110]

[0111] where d (i) represents the input feature dimension of L i , represents the output feature dimension of L i ; S t is the sparsity target;

[0112] A14), Refer to the size of the lowest L2 norm and sort according to the sparsity target to obtain the maximum weight value W h to be pruned and the minimum weight value W l to be retained, that is:

[0113]

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

[0115] Then the weight threshold is obtained as:

[0116]

[0117] A15), Perform sparse training on the trained satellite positioning correction model to achieve weight mixing, and gradually integrate the weights of the trained satellite positioning correction model to the target value 0, thereby obtaining the pruned and sparsely trained model.

[0118] In this embodiment, in sparse training, satellite measurement features are used as the training set. Among them, the measurement features of the satellite include pseudorange residuals (PR), carrier-to-noise ratio (C / N0), and the LOS vector from the satellite to the receiver. The pseudorange residual refers to the difference between the actually measured pseudorange value 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 give the learning module an index to measure the reliability of the signal. The LOS vector is the unit direction vector from the receiver to the satellite, which contains information such as the elevation angle and azimuth angle of the satellite.

[0119] The satellite positioning correction model is trained on the server side. Among them, the training configuration adopted in this embodiment is: a high-performance RTX3090 graphics card with a memory of 24G. The proposed positioning correction model is implemented based on the artificial intelligence framework PyTorch and uses CUDA version 11.6 for hardware acceleration. During the training process, the training gradient loss Loss is used as the index for training completion. After 100 iterations when 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, a weight factor α i is introduced to adjust the weights to be pruned. In the pruning training parameter update stage, the original, non-parametric weights are used to adjust the weights w to be pruned to the target value, that is:

[0121]

[0122] where w new represents the weights 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 expression of the weight factor α i is:

[0124]

[0125] where E i represents the current training round, E init represents the number of iterations of normal training, and E knee represents the number of iterations of sparse training.

[0126] A2) Incorporate the satellite positioning correction model part through a loose coupling method.

[0127] It is signed onto the microprocessor MCU of the positioning module, and an application program of the satellite positioning correction model is written according to the data flow direction in 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 onto 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: first, convert the satellite positioning correction model into the ONNX framework compatible with multiple platforms, and then use the RKNN toolchain provided by Rockchip to further convert the ONNX framework into the RKNN model that can be called by RK3568. Use the Huawei Cloud ModelBox framework to encapsulate the inference model at the embedded end, then write the model call API function at the embedded end to call the satellite positioning correction model for position correction result inference, and finally write the application program of the positioning correction APP according to the data flow of the positioning correction process. The transplantation process of the satellite positioning correction model and the call and inference process on the RK3568 platform are shown in Process Figure 2 and Process Figure 3 are given, and the data flow for reference to write the positioning correction application program APP is shown in Figure 4 is given.

[0129] The above embodiments and the descriptions in the specification only illustrate the principles and the best embodiments of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed.

Claims

1. A satellite positioning method for urban complex environments, characterized in that It includes the following steps: S1), Receive multi-frequency and multi-constellation satellite data in a real urban scene, and perform initial filtering and screening on it to obtain preprocessed satellite measurement data; S2), Predict the preprocessed satellite measurement data to obtain an initial positioning estimation result; S3), Extract satellite measurement features from the preprocessed satellite measurement data and construct a satellite observation map; S4), Construct a satellite positioning correction model and train it; Adopt a training enhancement strategy based on weighted adversarial weight perturbation to train the satellite positioning correction model. Introduce a perturbation factor during training and calculate the perturbation loss, then weight it with the original loss to obtain the final weighted loss of the current training batch, and then perform backpropagation to obtain model parameters with stronger generalization and robustness; specifically, it includes the following steps: S41), calculate the batch training loss during the training process train ; S42), copy the weight w of the current satellite positioning correction model train to the proxy model, and the proxy model calculates the loss and updates the weight w p , then calculate the weight difference w' between the proxy model and the current satellite positioning correction model to obtain the data perturbation, and at the same time use the perturbation intensity factor K to weight the weight difference w' to obtain the final adversarial weight perturbation w awp ; the final adversarial weight perturbation w awp is expressed as: w awp = K·w′ = K·(w p - w train ); S43), add the adversarial right perturbation w awp to the original satellite positioning correction model weights and calculate the perturbation loss loss awp ; S44) Obtain the final loss loss of the current batch based on the training loss and the perturbation loss all and update the satellite positioning correction model parameters, the final loss loss all which is expressed as: loss all =(1 - λ)·loss train + λ·loss awp ; where λ represents the weight ratio of the perturbation loss in calculating the total loss; loss train represents the batch training loss; loss awp represents the perturbation loss; S5), The satellite positioning correction model uses a graph structure-aware attention mechanism to simultaneously learn the positional relationships between different satellite nodes in GNSS satellite measurement data and the satellite observation map to extract the optimal environmental representation and then output a position positioning correction result; S6), Combine the initial positioning estimation result with the position positioning correction result to obtain the final positioning result.

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

3. A satellite positioning method for urban complex environments according to claim 1, characterized in that: In step S4), the satellite positioning correction model adopts a graph transformer satellite graph structure learning network based on the graph neural network GNN, including a sub-graph structure extractor based on the graph neural network GNN, a structure-aware attention layer, and a feed-forward network layer. The sub-graph structure extractor extracts the local structure information of each node from the satellite observation map, and the structure-aware attention layer integrates the extracted local structure information into the original attention score to consider the node attribute similarity at different distances and generate a more expressive node representation.

4. The satellite positioning method for urban complex environment according to claim 3, characterized in that: In step S5), extract local structure information from the satellite observation map through the satellite positioning correction model. Specifically: S511), perform k-hop sampling on each node i in the satellite observation map input to the satellite positioning correction model to obtain a set of neighborhood nodes that are k edges away from node i S512), aggregate and update the subgraph node information through the aggregation mechanism of the graph neural network GNN, that is: ​ wherein, represents the node representation of node i in the k-th layer of the GNN; obtained by the non-linear activation function σ, is the representation of the neighboring node u in the (k-1)-th layer of the GNN; AGGREGATE k is the aggregator in the k-th layer; is the representation of node i in the (k-1)-th layer of the GNN; CONCAT is the concatenation function; it is used to concatenate 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), Use sum pooling to obtain the graph representation of the sub-graph and update it as the sub-graph representation of the central node i, that is: where x i = φ(i, G) is the node representation of node i with respect to its subgraph G; N k (i) represents the k-hop neighborhood of node i including itself; φ(i, G) represents the subgraph calculation function; represents the graph neural network function used for extracting the subgraph at the k-th layer.

5. A satellite positioning method for urban complex environments according to claim 4, characterized in that: In step S5), after obtaining the local structure information, update the graph representation of the satellite observation map through the structure-aware attention score; The structure-aware attention layer based on kernel smoothing integrates the extracted local structure information into the original attention score to consider the node attribute similarity at different distances and 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 a linearization function, and \(V\) is the node set of the satellite observation map; \(\kappa\) exp is \(R\) d \(\times R\) d an asymmetric exponential kernel on \(R\), and \(R\) d represents the dimension of the input exponential kernel vector; obtained by linearizing through the learnable parameters \(W\) Q and \(W\) K ; \(x\) v , \(x\) u are the node representations of nodes \(v\) and \(u\) with respect to their subgraph \(G\), respectively.

6. The satellite positioning method for urban complex environment according to claim 5, wherein: In step S6), after obtaining the structure-aware attention score, a position correction value is obtained through a fully connected layer and an average pooling layer and combined with the initial position estimate Pos Init to obtain the corrected positioning result Pos Pred , that is:

7. A satellite positioning system for complex urban environments, characterized in that The system is used for the method described in any one of claims 1-6. The system includes: An acquisition module for acquiring multi-frequency and multi-constellation satellite data in a real urban scene; A preprocessing module for preliminarily filtering and screening the multi-frequency and multi-constellation satellite data; A preliminary positioning prediction module for predicting the initial positioning estimation result of the receiver; A satellite observation map construction module for extracting satellite measurement features and constructing a satellite observation map; The map feature extraction module is used to extract the local structure information of the satellite observation map, integrate the extracted local structure information into the original attention score, consider the node attribute similarity at different distances, generate a more expressive node representation, and then output a 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.

8. A deployment method for a satellite positioning system, characterized in that, The method described above is used to deploy the system described in claim 7 to a chip, and the deployment method includes the following steps: A1), Compressing the model size based on the weighted hybrid soft pruning technique; A2), Deploying the satellite positioning correction model to the microcontroller MCU of the positioning module in a loosely coupled manner, and writing an application program for the satellite positioning correction model according to the data flow direction during the positioning process.

9. The deployment method according to claim 8, wherein In step A1), the following steps are specifically included: A11), set the sparse target S of the satellite positioning correction model t , the sparse target S t represents the weight ratio to be pruned in the network layer weight matrix W of the model; A12) Prune the weights using the lowest norm l2 as the standard, and define the weight threshold as W t , and the weights lower than the weight threshold W t will be pruned; A13)、For network layer L i , define the number of initial weights N orig and the number of pruned sparse weights N kept respectively as: where d (i) represents the input feature dimension of L i , and represents the output feature dimension of L i ; S t is a sparse target; A14), referring to the minimum l2 norm size, and sorting according to the sparse objective to obtain the maximum weight value W to be pruned h and the minimum weight value W to be retained l , that is: Where W is the weight matrix of the network layer; N kept The number of pruned sparse weights; N orig Is the initial number of weights; topk represents sorting and taking the largest or smallest k items, and largest represents the sorting method used. largest = True means taking the largest N kept Elements, largest = False means taking the smallest N orig -N kept Elements; Then the weight threshold is obtained as: A15), Performing sparse training on the trained satellite positioning correction model to achieve weighted mixing, gradually integrating the weights of the trained satellite positioning correction model into the target value 0, so as to obtain a pruned and sparsely trained model; In sparse training, a weight factor α is introduced i to adjust the weights to be pruned; in the pruning training parameter update phase, the original, non-parametric weights are used to adjust the weights w to be pruned to the target value, i.e.: where w new represents the weights of sparse training updates, and w t represents the weight threshold; α i is the weight factor; w represents the weight value before pruning in the weight matrix of the current network layer; Among them, the weight factor α in the i-th training round i has the following expression: Where, E i represents the i-th training round, E init represents the number of iterations for normal training, and E knee represents the number of iterations for sparse training.

Images