Multipath signal recognition method in urban dynamic environment based on graph contrast learning
Through the graph comparison learning method, multipath signals are identified in complex environments, which solves the problems of high redundancy dimensions of training samples and large demand for tag data, and realizes high-precision multipath signal recognition, reducing labor costs and tag dependence.
Patent Information
- Application Number
- CN202411087455.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-09
AI Technical Summary
In complex environments, the existing multipath signal recognition methods have the problem of high redundancy dimensions of training samples and require a large amount of label data, which leads to low recognition accuracy and poor generalization of the model, making it difficult to achieve high-precision multipath signal recognition.
Using a graph comparison learning method, we use satellite measurement data to model graph data, build a sky satellite feature map, and use a graph comparison learning network framework to enhance the asymmetric model to perform unsupervised pre-training to obtain the graph after the updated parameters. Transformer encoder performs feature extraction and multipath signal prediction, and combines multipath signal prediction value and real tag to obtain the final identification result.
The dimensionality reduction and unsupervised feature extraction of Beidou satellite signals are realized, which reduces the dependence on label data, improves the accuracy and generalization of multipath signal recognition, and is suitable for high-precision multipath signal recognition in complex environments.
Smart Images

Figure CN118626939B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of satellite signal recognition, and in particular to a multipath signal recognition method in a city dynamic environment based on graph comparison learning. Background Art
[0002] Precise satellite navigation and positioning technology is one of the core technologies of intelligent driving. The satellite navigation system can realize autonomous positioning. In an open environment, there are more observable satellites, the distribution range of observation satellites in space is large, and the dynamic positioning accuracy is higher. In a circular route under the obstruction of buildings and trees, due to the influence of the external environment, the number of observation satellites is relatively reduced, and the geometric distribution range of satellites in space is relatively small. The positioning results will produce sudden errors due to the influence of observation satellites, which reduces the dynamic positioning accuracy. High-precision satellite positioning in urban dynamic scenes is an international problem that needs to be solved urgently.
[0003] Although the Beidou Navigation Satellite System (BDS) has the characteristics of high precision and short messages, it can provide high-precision and high-reliability positioning, navigation, and timing services to all types of users around the world all day and all day. However, vehicle navigation in urban canyon environments with many high-rise buildings has higher requirements for the Beidou Navigation Satellite System (BDS) positioning service. However, due to the multipath effect, the satellite navigation positioning is offset, which leads to inaccurate vehicle positioning. Normally, in an open environment without building obstructions, the BDS receiver receives a direct signal (LOS) from the Beidou satellite, at which time the BDS can provide centimeter-level positioning services. However, in an urban canyon environment with many buildings, the multipath effect seriously reduces the positioning accuracy.
[0004] At present, there are two major methods for dealing with multipath effects: hardware and software. Hardware-based methods include anti-multipath antennas or improving receiver hardware performance. Such methods are not very effective in suppressing multipath effects, and the need for hardware modification limits their application scenarios. Software-based methods such as the sidereal day filter method and the multipath semi-celestial mapping method have problems such as instability or huge computational complexity, and are difficult to meet the needs of high-precision and fast positioning.
[0005] Artificial intelligence technology provides a new way to solve the problem of multipath signal feature recognition and extraction in complex environments. It suppresses the multipath effect through machine learning algorithms. For example, the initial position of the robust estimator and the satellite visibility determined by the support vector machine are used to improve the shadow matching method. However, this method has the problem of difficulty in building a training database and requires a large amount of labeled data. In addition, the existing deep learning algorithms based on artificial intelligence technology also have problems such as low recognition accuracy and poor model generalization, which directly leads to the difficulty in accurately identifying non-line-of-sight signals in complex environments.
[0006] Therefore, there is an urgent need for an effective high-precision multipath signal recognition method that can be applied in complex environments. Summary of the invention
[0007] In view of the shortcomings of the prior art, the present invention provides a multipath signal recognition method in an urban dynamic environment based on graph contrast learning. The present invention can solve the technical problems in the existing multipath recognition method that the Beidou satellite signal training samples have high redundant dimensions and require a large amount of label data.
[0008] The technical solution of the present invention is: a method for identifying multipath signals in a dynamic urban environment based on graph contrast learning, comprising the following steps:
[0009] S1) Obtain satellite measurement data;
[0010] S2), perform graph data modeling based on the measurement data to obtain a sky satellite feature map;
[0011] S3) Using the sky satellite feature map to perform unsupervised pre-training on the graph contrast learning network framework based on the asymmetric model enhancement method; and obtaining the graph Transformer encoder with updated parameters;
[0012] S4), using the graph Transformer encoder with updated parameters to perform feature extraction processing on the sky satellite feature graph, and output the multipath signal prediction value;
[0013] S5) Combine the multipath signal prediction value with the true label to obtain the final multipath signal recognition result.
[0014] Preferably, in step S1), the measurement data is derived from global satellite navigation signals from multiple different locations and orbits, including data collected by GPS, Galileo satellite system, BeiDou 2 (BDS-2) and BeiDou 3 system (BDS-3).
[0015] Preferably, in step S1), the measurement data includes four satellite characteristics: azimuth angle, altitude angle, carrier signal-to-noise ratio and pseudorange residual.
[0016] Preferably, in step S1), each satellite feature is standardized, that is, each satellite feature is converted into a distribution with a mean of 0 and a standard deviation of 1, and the calculation formula is:
[0017]
[0018] in, is the satellite characteristics before conversion, is the transformed satellite feature, represents the mean, Represents variance.
[0019] Preferably, in step S2), graph data modeling is performed based on the measurement data to obtain a sky satellite characteristic graph, which specifically includes the following steps:
[0020] S21), for the satellite measurements of different constellations received by the receiver at a certain moment, a sky satellite characteristic map is proposed;
[0021] S22), taking each satellite position as a node in the graph, and then splicing the pre-selected altitude angle, azimuth angle, carrier-to-noise ratio and pseudo-range residual features as node features of the sky satellite feature graph;
[0022] S23) Based on the edge construction method of cosine similarity, an adjacency matrix describing the spatial structure information of the graph is constructed as the edge of the sky satellite feature graph; edge connections are established based on the proximity features of satellites. Assuming that the satellites in the same period are distributed on a hemisphere centered on the receiver, the elevation angle and azimuth angle are used to represent the position of the satellite on the spherical surface, and the spherical distance between the satellites is calculated. When the spherical distance between the satellites is Less than the set threshold At the same time, the satellites in the same constellation are connected; each element in the adjacency matrix is constructed as follows:
[0023]
[0024] in, and For satellite nodes, that is, when two satellites belong to the same constellation Or its spatial distance is less than the set threshold , establish edge connections.
[0025] Preferably, in step S3), the pre-training process of the graph contrast learning network frame based on the asymmetric model enhancement method is:
[0026] S31), inputting the constructed sky satellite feature map into the constructed graph comparison learning network frame based on the asymmetric model enhancement method;
[0027] S32), determine the hyperparameters of the network framework, select the training optimizer and loss function, and then start training;
[0028] S33), determine whether the training reaches the stop condition;
[0029] S34), if yes, then end the pre-training and obtain a trained classifier; if no, use the back propagation algorithm BP to update the weights and biases of the Transformer encoder and then return to step S33).
[0030] Preferably, in step S3), during the pre-training process, for the satellite features and the edge-connected sky satellite image, two views are first generated based on simple data enhancement; then the two enhanced views are input into two graph Transformer encoders with shared parameters but different propagation depths to obtain different representations of the same sample in the feature space; and the graph contrast loss between the two different representations is calculated, similar image features are represented as close vectors, and dissimilar image features are represented as far away vectors; and the parameters of the graph Transformer encoder are updated using the back propagation algorithm BP.
[0031] Preferably, in step S3), the InfoNCE Loss loss function is used for graph contrast learning to train and optimize the graph Transformer encoder. For each sample, the distance between the positive sample and the negative sample is calculated, and the distance between the positive sample and the negative sample is minimized and maximized as the optimization target. The expression of the InfoNCE Loss loss function is:
[0032]
[0033] In the formula, represents the temperature coefficient; represents the encoding representation vector of the query sample, represents the encoding vector of the positive sample corresponding to the query sample, Represents the encoding vector of the total sample corresponding to the query sample; the numerator represents the similarity between positive examples, and the denominator represents the similarity between positive examples and negative examples.
[0034] Preferably, in step S34), by randomly changing the number of propagation operators in the encoder in each epoch, multiple graph Transformer encoders with different propagation depths are obtained during the training process.
[0035] Preferably, in step S32), the hyper parameters of the network framework that need to be determined are:
[0036] 1) The number of propagation layers of the Transformer encoder of the input graph of the two enhanced views;
[0037] 2) The length of the feature vector in the Transformer and the number of heads in its multi-head attention module;
[0038] 3) Number of fully connected neural network layers after fusion , and the number of neurons in its hidden layer.
[0039] Preferably, in step S32), the optimizer uses the first-order moment estimation and the second-order moment estimation of the Adam optimizer to dynamically adjust the learning rate of each parameter; and through bias correction, ensure that the learning rate of each iteration is within a certain range.
[0040] Preferably, in step S33), the total number of training rounds of the network framework is set to M, and the back propagation algorithm BP is used to obtain the gradient to update the weights and biases of the network framework, and the update formula is:
[0041]
[0042] in, represents the weight, Indicates bias, is the learning rate, It is the process of finding the partial derivative of the loss on the corresponding parameters.
[0043] Preferably, in step S4), the altitude angle, carrier signal-to-noise ratio, pseudorange residual and edge connection are input as input data into the trained graph Transformer encoder, and after feature transformation, attention weight calculation, message passing and feature update, the output features of the aggregated node information are finally obtained.
[0044] Preferably, step S4) specifically includes the following steps:
[0045] S41), combine multiple self-attention mechanism sub-networks to form a multi-head attention mechanism layer;
[0046] S42), stitching together the information of the heads in the multi-head attention mechanism layer, obtaining the updated information of the central node after the aggregation of the neighboring nodes, and finally performing a linear transformation to obtain the output multipath prediction value.
[0047] As a preference, in step S41), for a given node feature ; Calculate the multi-head attention mechanism between satellite nodes based on the edge connection, where The attention mechanism of a layer is calculated as follows:
[0048]
[0049]
[0050] In the formula, is the query vector; is the key vector; , are the trainable weights corresponding to the query vector and the key vector; is the central feature; , is the bias term corresponding to the query vector and key vector; is an adjacent feature; is the attention weight matrix; is a nonlinear activation function; represents the number of neurons in the hidden layer of each attention head; T is the transpose operation.
[0051] Through different trainable parameters , , , , the central feature and adjacent features Convert to query vectors respectively and key vector ; In addition, the adjacency matrix containing edge information is required as an additional input; query Vector transpose and key Applying the result of vector multiplication Function, you can get the attention weight matrix .
[0052] Preferably, in step S42), after completing the multi-head attention mechanism operation of the graph, the information learned by multiple heads is fused for message passing to obtain the update result of the next layer:
[0053]
[0054] in, is the value vector; is the trainable weight corresponding to the value vector; is an adjacent feature; is the bias term corresponding to the value vector; The output features after the central node aggregates the updated information of the adjacent nodes; is the trainable weight corresponding to the output feature; represents a tensor concatenation operation, M is the total number of heads.
[0055] By obtaining the attention weight matrix and the learned value vector Multiply them, and then perform multi-head information connection operation to obtain the central node update information after the neighboring nodes are aggregated.
[0056] Preferably, in step S5), a pre-trained graph Transformer encoder is used in the classifier to extract features, and the binary cross entropy loss between the multipath signal prediction value output by the graph Transformer encoder and the true label is calculated. The closer the prediction result is to 1, the closer the value of the loss function is to zero. The calculation formula of the binary cross entropy loss is as follows:
[0057]
[0058] In the formula, Represents a binary label 0 or 1, Indicates that the output belongs to The probability of the label; N Indicates the sample size.
[0059] The beneficial effects of the present invention are:
[0060] 1. Considering the high redundant dimension of multipath training samples and the need for a large amount of label data, the present invention constructs a graph contrast learning framework to train the graph encoder; updates the graph encoder parameters by comparing the similarity between two feature spaces augmented by the same sample, thereby achieving dimensionality reduction and unsupervised feature extraction of BDS measurement signals;
[0061] 2. The present invention adopts the model enhancement method as data enhancement, focusing on the structural adjustment of the encoder rather than the input to retain most of the satellite image information;
[0062] 3. The present invention constructs a sky satellite map based on satellite nodes and satellite connections to learn the interactive representation of airspace satellites, and obtains a graph encoder through unsupervised pre-training of a graph comparison learning framework, thereby reducing the high dependence of tasks on labels and reducing labor costs;
[0063] 4. The present invention uses a multi-head attention method to learn the feature correlation between airspace satellites to obtain a representation related to the environment; by using the trained graph Transformer encoder to extract the features of the corresponding data set and comparing them with the true labels in an external classifier, an effective multipath signal recognition model with high accuracy and good generalization is achieved. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] Figure 1 It is a flow chart of the multipath signal identification method of the present invention;
[0065] Figure 2 Schematic diagram of the effect of the asymmetric model enhancement method in an embodiment of the present invention;
[0066] Figure 3 It is a schematic diagram of the flow chart of the training of the network framework for comparative learning in an embodiment of the present invention;
[0067] Figure 4 It is a schematic diagram of a framework for unsupervised pre-training of a comparative learning network framework in an embodiment of the present invention;
[0068] Figure 5 Schematic diagram of the structure of the Transformer module in the embodiment of the present invention. DETAILED DESCRIPTION
[0069] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings:
[0070] like Figure 1 As shown, this embodiment provides a method for identifying multipath signals in an urban dynamic environment based on graph contrast learning, comprising the following steps:
[0071] S1) Obtain satellite measurement data
[0072] In this embodiment, the measurement data is derived from global satellite navigation signals from multiple different locations and orbits, including data collected by GPS, Galileo satellite system, BeiDou 2 (BDS-2) and BeiDou 3 system (BDS-3). In addition, this embodiment collects satellite signals of BDS, GPS and GAL constellations at a frequency of 1 Hz in an urban canyon environment; the main satellite features collected are 4: azimuth, altitude, carrier signal-to-noise ratio and pseudorange residual.
[0073] Among them, the elevation angle is an important determinant of the quality of satellite signals; due to interference from buildings, trees or other obstacles, the signal strength and quality of low-elevation satellites are usually poor. Azimuth is a key environmental parameter that describes the positioning of the satellite together with the elevation angle. By using azimuth as one of the node features, the model can understand the degree of environmental obstacles in a specific direction. The carrier-to-noise ratio reflects the quality of the received signal and can give the learning module an indicator that can measure the reliability of the signal. The pseudorange residual represents the inconsistency between the pseudorange of the satellite observation value of a certain frequency and all pseudorange values. When the number of satellites is sufficient, it can represent the magnitude of the multipath error.
[0074] In addition, this embodiment also performs standardization processing on each satellite feature, that is, converting each satellite feature into a distribution with a mean of 0 and a standard deviation of 1, and the calculation formula is:
[0075]
[0076] in, is the satellite characteristics before conversion, is the transformed satellite feature, represents the mean, Represents variance.
[0077] S2) Model the graph data based on the measured data to obtain the sky satellite feature map
[0078] The specific steps include:
[0079] S21), in this embodiment, the satellite's altitude angle, azimuth angle, carrier-to-noise ratio and pseudo-range residual are used as node features of the sky satellite map; for the satellite measurements of different constellations received by the receiver at a certain moment, a sky satellite feature map is proposed;
[0080] S22), taking each satellite position as a node in the graph, and then splicing the altitude angle, azimuth angle, carrier-to-noise ratio and pseudo-range residual features as node features of the sky satellite feature graph;
[0081] S23) Based on the edge construction method of cosine similarity, an adjacency matrix describing the spatial structure information of the graph is constructed as the edge of the sky satellite feature graph; edge connections are established based on the proximity features of satellites. Assuming that the satellites in the same period are distributed on a hemisphere centered on the receiver, the elevation angle and azimuth angle are used to represent the position of the satellite on the spherical surface, and the spherical distance between the satellites is calculated. When the spherical distance between the satellites is Less than the set threshold At the same time, the satellites in the same constellation are connected; each element in the adjacency matrix is constructed as follows:
[0082]
[0083] in, and For satellite nodes, that is, when two satellites belong to the same constellation Or its spatial distance is less than the set threshold , establish edge connections.
[0084] S3) Use the sky satellite feature map to perform unsupervised pre-training on the graph contrast learning network framework based on the asymmetric model enhancement method; and obtain the graph Transformer encoder with updated parameters
[0085] This embodiment uses an asymmetric model enhancement method to focus on the implementation of encoders rather than their input, fully avoiding the loss of key information of the graph. By operating the encoder, a representation that is conducive to multipath identification is obtained. Using encoders with shared parameters but different propagation layers can reduce high-frequency noise. Contrastive learning can extract information shared between different perspectives, thereby filtering out task-irrelevant noise that only appears in a single perspective. Figure 2 As shown, To provide information related to downstream tasks, and The information contained in the two views generated by the encoder. When using exactly the same encoder, contrastive learning will maximize the mutual information of the two views, that is, the area of A and B is maximized, but the information related to the downstream task is only A, and the increase in area B will lead to the introduction of unnecessary noise; when using encoders of different depths, the size of area B can be significantly reduced, allowing the model to pay more attention to information related to downstream tasks.
[0086] like Figure 3 As shown, in this embodiment, the pre-training process of the graph contrast learning network frame based on the asymmetric model enhancement method is:
[0087] S31), inputting the constructed sky satellite feature map into the constructed graph comparison learning network frame based on the asymmetric model enhancement method;
[0088] S32), determine the hyperparameters of the network framework, select the training optimizer and loss function, and then start training;
[0089] S33), determine whether the training reaches the stop condition;
[0090] S34), if yes, then end the pre-training and obtain a trained classifier; if no, use the back propagation algorithm BP to update the weights and biases of the Transformer encoder and then return to step S33).
[0091] During the pre-training process, Figure 4 As shown, for the sky satellite image with satellite features and edge connections, two views are first generated based on simple data augmentation. and Represents two random proportions of node removal and edge perturbation. The goal is to provide more information to the model by increasing the diversity of data, thereby improving the generalization ability of the model. The two enhanced views are then input into two graph Transformer encoders that share parameters but have different propagation depths. In order to obtain different representations of the same sample in the feature space, The Transformer encoder is used to transmit node information in the propagation layer, which aims to merge the neighbor information of each node into the node features in order to better describe the properties and relationships of the nodes. The conversion layer representing the graph encoder performs nonlinear transformation on the node features, helping the network learn richer and more complex node features to improve the performance of the graph neural network. Finally, the graph contrast loss between the two different representations is calculated, similar image features are represented as close vectors, and dissimilar image features are represented as distant vectors. The back propagation algorithm BP is used to update the parameters of the graph Transformer encoder.
[0092] In this embodiment, the InfoNCE Loss loss function is used for graph contrast learning to train and optimize the graph Transformer encoder. For each sample, the distance between the positive sample and the negative sample is calculated, and the distance between the positive sample and the negative sample is minimized and maximized as the optimization goal.
[0093] By randomly changing the number of propagation operators in the encoder in each epoch, multiple graph Transformer encoders with different propagation depths are obtained during training.
[0094] In this embodiment, the hyper parameters of the network framework that need to be determined are:
[0095] 1) The number of propagation layers of the Transformer encoder of the input graph of the two enhanced views;
[0096] 2) The length of the feature vector in the Transformer and the number of heads in its multi-head attention module;
[0097] 3) Number of fully connected neural network layers after fusion , and the number of neurons in its hidden layer.
[0098] In this embodiment, the optimizer uses the first-order moment estimation and the second-order moment estimation of the Adam optimizer to dynamically adjust the learning rate of each parameter; through bias correction, it is ensured that the learning rate of each iteration is within a certain range.
[0099] In this embodiment, the total number of training rounds of the network framework is set to M, and the back propagation algorithm BP is used to obtain gradients to update the weights and biases of the network framework.
[0100] S4) Use the graph Transformer encoder with updated parameters to extract features from the sky satellite feature graph and output the multipath signal prediction value.
[0101] This embodiment uses the altitude angle, carrier signal-to-noise ratio, pseudorange residual and edge connection as input data to the trained graph Transformer encoder, and finally obtains the output features of the aggregated node information after feature transformation, attention weight calculation, message passing and feature update. The specific steps include:
[0102] S41), combine multiple self-attention mechanism sub-networks to form a multi-head attention mechanism layer, such as Figure 5 shown.
[0103] Among them, the expression of the self-attention mechanism in this embodiment is:
[0104]
[0105] in, , , , It is the characteristic information of multiple satellites at a certain moment. , and They are the query matrix, key matrix and value matrix, which are linear transformations of the node feature X and represent the projections of X in different directions; is a nonlinear activation function, which and After processing the inner product of , we get the correlation score matrix between each satellite feature. By performing the inner product between the result and V, we can get the attention map between multiple satellite features. is the scaling factor. When the sequence length is long, and The inner product is scaled to smooth The output of , thus obtaining a more appropriate score matrix.
[0106] For a given node feature ; Calculate the multi-head attention mechanism between satellite nodes based on the edge connection, where The attention mechanism of a layer is calculated as follows:
[0107]
[0108] In the formula, is the query vector; is the key vector; , are the trainable weights corresponding to the query vector and the key vector; is the central feature; , is the bias term corresponding to the query vector and key vector; is an adjacent feature; is the attention weight matrix; is a nonlinear activation function; represents the number of neurons in the hidden layer of each attention head; T is the transpose operation;
[0109] Through different trainable parameters , , , , the central feature and adjacent features Convert to query vectors respectively and key vector ; In addition, the adjacency matrix containing edge information is required as an additional input; query Vector transpose and key Applying the result of vector multiplication Function, you can get the attention weight matrix .
[0110] S42), stitching together the information of the heads in the multi-head attention mechanism layer, obtaining the updated information of the central node after the aggregation of the neighboring nodes, and finally performing a linear transformation to obtain the output multipath prediction value; the details are as follows:
[0111] After completing the multi-head attention mechanism operation of the graph, the information learned by multiple heads is fused for message passing to obtain the update result of the next layer:
[0112]
[0113] in, is the value vector; is the trainable weight corresponding to the value vector; is an adjacent feature; is the bias term corresponding to the value vector; The output features after the central node aggregates the updated information of the adjacent nodes; is the trainable weight corresponding to the output feature; ‖ represents the tensor connection operation, M is the total number of heads;
[0114] By obtaining the attention weight matrix and the learned value vector Multiply them, and then perform multi-head information connection operation to obtain the central node update information after the neighboring nodes are aggregated.
[0115] S5) Combine the multipath signal prediction value with the true label to obtain the final multipath signal recognition result
[0116] In this embodiment, a pre-trained graph Transformer encoder is used in the classifier to extract features, and the binary cross entropy loss between the multipath signal prediction value output by the graph Transformer encoder and the true label is calculated. The closer the prediction result is to 1, the closer the value of the loss function is to zero. The calculation formula of the binary cross entropy loss is as follows:
[0117]
[0118] In the formula, Represents a binary label 0 or 1, Indicates that the output belongs to The probability of the label; NIndicates the sample size.
[0119] This embodiment takes into account the problem that the multipath training samples have high redundant dimensions and require a large amount of labeled data, and constructs a graph contrast learning framework to train the graph encoder. By comparing the similarity between the two feature spaces augmented by the same sample, the graph encoder parameters are updated, thereby achieving dimensionality reduction and unsupervised feature extraction of BDS measurement signals. In addition, this embodiment takes into account that traditional data enhancement methods such as translation, rotation, scaling, and cropping may cause loss of graph information. This embodiment uses an asymmetric model enhancement method as the main means of data enhancement, and focuses on the structural adjustment of the encoder rather than the input to retain most of the satellite image information.
[0120] This embodiment takes into account the problem that the Beidou satellite sequence signal training samples have high redundant dimensions and require a large amount of labeled data, and uses a graph comparison learning framework for feature extraction, which can achieve effective recognition of non-direct signals of unlabeled Beidou satellite long-term series collected in urban canyon environments. For the learning of hidden information in GNSS measurements, this embodiment constructs a sky satellite map based on satellite nodes and satellite connections to learn the interactive representation of airspace satellites, and obtains a graph encoder through unsupervised pre-training of the graph comparison learning framework, reducing the high dependence of the task on labels and reducing labor costs. Compared with the method of directly using the target satellite features for multipath identification, the model of multipath identification that introduces the graph comparison learning framework has the characteristics of high robustness and low label requirements, and the generalization performance of multipath signal identification is better when labels are missing.
[0121] 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 multipath signal recognition method in a dynamic urban environment based on graph contrast learning, characterized in that: The steps include: S1), obtaining satellite measurement data; S2), performing graph data modeling based on the measurement data to obtain a sky satellite feature map; S3), using sky satellite feature maps to perform unsupervised pre-training on a graph contrast learning network framework based on an asymmetric model enhancement method; And get the graph Transformer encoder with updated parameters; The pre-training process of the graph contrast learning network frame based on the asymmetric model enhancement method is as follows: S31), inputting the constructed sky satellite feature map into the constructed graph comparison learning network frame based on the asymmetric model enhancement method; S32), determine the hyperparameters of the network framework, select a training optimizer and a loss function, and then start training; S33), determine whether the training reaches the stop condition; S34), if yes, then end the pre-training and obtain a trained classifier; if no, use the back propagation algorithm BP to update the weights and biases of the Transformer encoder and then return to step S33); During the pre-training process, for satellite features and edge-connected sky satellite images, two views are first generated based on simple data augmentation; then the two augmented views are input into two graph Transformer encoders with shared parameters but different propagation depths to obtain different representations of the same sample in the feature space; and the graph contrast loss between the two different representations is calculated, similar image features are represented as close vectors, and dissimilar image features are represented as far-away vectors; the parameters of the graph Transformer encoder are updated using the back propagation algorithm BP; Use the InfoNCE Loss function for graph contrast learning and train the optimized graph Transformer encoder. For each sample, calculate the distance between the positive sample and the negative sample. Minimize the distance to the positive sample and maximize the distance to the negative sample as the optimization goal. By randomly changing the number of propagation operators in the encoder in each epoch, multiple graph Transformer encoders with different propagation depths are obtained during training; S4), using the graph Transformer encoder with updated parameters to perform feature extraction processing on the sky satellite feature graph, and output the multipath signal prediction value; S5), combining the multipath signal prediction value and the true label to obtain the final multipath signal recognition result.
2. The multipath signal recognition method in an urban dynamic environment based on graph contrast learning according to claim 1 is characterized in that: In step S1), the measurement data includes four satellite characteristics: azimuth, altitude, carrier signal-to-noise ratio and pseudorange residual.
3. The multipath signal recognition method in an urban dynamic environment based on graph contrast learning according to claim 1 is characterized in that: In step S2), the graph data modeling is performed based on the measurement data to obtain the sky satellite characteristic graph, which specifically includes the following steps: S21), for the satellite measurements of different constellations received by the receiver at a certain moment, a sky satellite characteristic map is proposed; S22), taking each satellite position as a node in the graph, and then splicing the altitude angle, azimuth angle, carrier-to-noise ratio and pseudorange residual features as node features of the sky satellite feature graph; S23) Based on the edge construction method of cosine similarity, an adjacency matrix describing the spatial structure information of the graph is constructed as the edge of the sky satellite feature graph; edge connections are established based on the proximity features of the satellites. Assuming that the satellites in the same period are distributed on a hemisphere centered on the receiver, the elevation angle and azimuth angle are used to represent the position of the satellite on the spherical surface, and the spherical distance between the satellites is calculated. When the spherical distance S between the satellites is ij When it is less than the set threshold γ, an edge connection is established; at the same time, connections are established between satellites in the same constellation; each element in the adjacency matrix is constructed as follows: Among them, v i and v j For satellite nodes, that is, when two satellites belong to the same constellation C or their spatial distance is less than the set threshold γ, an edge connection is established.
4. The method for multipath signal recognition in urban dynamic environment based on graph contrast learning according to claim 1 is characterized in that: In step S3), the InfoNCE Loss function L q The expression is: Where τ represents the temperature coefficient; q represents the encoding representation vector of the query sample, and k + represents the encoding vector of the positive sample corresponding to the query sample, k i represents the encoding vector of the i-th sample corresponding to the query sample; the numerator represents the similarity between positive examples, and the denominator represents the similarity between positive examples and negative examples.
5. The method for multipath signal recognition in urban dynamic environment based on graph contrast learning according to claim 1 is characterized in that: Step S4) specifically includes the following steps: S41), combining multiple self-attention mechanism sub-networks to form a multi-head attention mechanism layer; S42) splice the information of the heads in the multi-head attention mechanism layer to obtain the updated information of the central node after the neighboring nodes are aggregated, and finally perform linear transformation to obtain the output multipath prediction value.
6. The method for identifying multipath signals in a dynamic urban environment based on graph contrast learning according to claim 5 is characterized in that: In step S41), for a given node feature The multi-head attention mechanism between satellite nodes is calculated based on the edge connection, where the attention mechanism of the mth layer is calculated as follows: In the formula, is the query vector; is the key vector; are the trainable weights corresponding to the query vector and the key vector; is the central node feature; is the bias item corresponding to the query vector and key vector; is the adjacent node feature; is the attention weight matrix; softmax is the nonlinear activation function; d represents the number of neurons in the hidden layer of each attention head; T is the transposition operation; Through different trainable parameters The central node feature and adjacent node features Convert to query vectors respectively and key vector In addition, the adjacency matrix containing edge information is required as an additional input; the attention weight matrix can be obtained by applying the softmax function to the result of transposing the query vector and multiplying the key vector 7. The method for identifying multipath signals in a dynamic urban environment based on graph contrast learning according to claim 5 is characterized in that: In step S42), after completing the multi-head attention mechanism operation of the graph, the information learned by multiple heads is fused for message passing to obtain the update result of the next layer: in, is the value vector; is the trainable weight corresponding to the value vector; is the adjacent node feature; is the bias term corresponding to the value vector; The output features after the central node aggregates the updated information of the adjacent nodes; is the trainable weight corresponding to the output feature; ∥ represents the tensor connection operation, and M is the total number of heads; By obtaining the attention weight matrix and the learned value vector Multiply them, and then perform multi-head information connection operation to obtain the central node update information after the neighboring nodes are aggregated.
8. The method for multipath signal recognition in urban dynamic environment based on graph contrast learning according to claim 1 is characterized in that: In step S5), a pre-trained graph Transformer encoder is used in the classifier to extract features, and the binary cross entropy loss between the multipath signal prediction value output by the graph Transformer encoder and the true label is calculated. The closer the prediction result is to 1, the closer the value of the loss function is to zero. The binary cross entropy loss The calculation formula is as follows: In the formula, y represents the binary label 0 or 1, p represents the probability that the output belongs to the y label; N represents the number of samples.
Citation Information
Patent Citations
BOC signal anti-multipath method based on two-dimensional tracking structure
CN103439715A
Signal modulation format recognition method, system and device based on attention mechanism graph neural network and storage medium
CN114513394A
Navigation multi-feature GNSS deception jamming detection method, system and device and medium
CN115598669A
Multi-path signal identification method considering satellite time sequence and airspace interactivity
CN115952407A
Multi-path detection method for CDMA receivers
US20060140254A1
