Methods, devices, equipment and media for classifying and identifying train satellite positioning observation scenarios

Abstract

This specification provides a method, apparatus, device, and medium for classifying and identifying train satellite positioning observation scenes. The method includes acquiring satellite position observation information and generating a satellite observation feature matrix and a satellite signal observation category feature matrix based on a three-dimensional electronic track map database; training an observation category classification model based on an SVM model; dividing the elevation and azimuth angles of each satellite at observation epochs into equal-angle intervals to generate a sky visibility feature matrix; dividing the sky visibility feature matrix into L×L blocks to obtain a block matrix; flattening each block matrix to obtain a block feature vector; and using the processed block feature vectors to train an observation scene category classification model based on a ViT model to obtain scene category classification results. This addresses the problem of inaccurate identification of complex scenes along train routes.

Description

Technical Field

[0001] This invention relates to the field of train satellite positioning technology, and in particular to a method, device, equipment and medium for classifying and identifying train satellite positioning observation scenarios. Background Technology

[0002] The environment along railway lines is complex, and can be broadly categorized into open areas, cuttings, viaducts, tunnels, stations, and urban canyons. Satellite-based train positioning systems are susceptible to obstruction from buildings or other objects along the tracks, leading to reflections and blockages during satellite signal propagation, thus affecting positioning accuracy. Existing railway scene classification methods mostly rely on cluster analysis, requiring manual identification of cluster characteristics. Therefore, they can only classify and identify a subset of typical scenes.

[0003] Currently, satellite positioning receivers typically cannot distinguish between reflected and direct signals, and will use reflected signals for position calculations, leading to a degradation in positioning accuracy in restricted environments. Therefore, it is necessary to classify the railway lines into different scenarios, quantitatively analyze the degree of limitation in satellite positioning observation, and predict and assess train positioning accuracy.

[0004] In view of this, there is an urgent need to provide a method that can identify complex scenarios along train routes in order to improve the stability of the optimization level of satellite positioning optimization methods under different scenarios. Summary of the Invention

[0005] To overcome the problems existing in related technologies, this disclosure provides a method, device, equipment and medium for classifying and identifying train satellite positioning observation scenarios, so as to solve the technical problem of inaccurate identification of complex scenarios along train operation lines in related technologies.

[0006] This specification provides one or more embodiments of a method for classifying and identifying train satellite positioning observation scenarios, acquiring satellite observation data at each observation epoch, and extracting satellite position observation information under limited scenarios from the observation data;

[0007] Based on satellite observation data and a 3D electronic orbital map database, a satellite observation feature matrix and a satellite signal observation category feature matrix are generated. These are used to train an observation category classification model based on an SVM model. The observation categories include line-of-sight signals and non-line-of-sight signals.

[0008] The elevation and azimuth angles of each satellite at each observation epoch are divided into equal angular intervals to generate an identity matrix; the signal observation category information of each satellite in the row and column of the identity matrix at each observation epoch is averaged to obtain the signal observation category value; and the sky visibility feature matrix is ​​generated by filling the empty element positions in the identity matrix through bilinear interpolation.

[0009] The sky visibility feature matrix is ​​divided into blocks of size L×L to obtain a block matrix. Each block matrix is ​​then flattened to obtain a block feature vector. The scene category to which each block feature vector belongs and its corresponding position encoding in the sky visibility feature matrix are then marked.

[0010] The processed block feature vectors are used to train an observation scene category classification model based on the ViT model to obtain scene category classification results.

[0011] This specification provides one or more embodiments of a train satellite positioning observation scene classification and identification device, including:

[0012] The data acquisition module is used to acquire satellite observation data at each observation epoch and extract satellite position observation information under limited scenarios from the observation data;

[0013] The first model training module is used to generate a satellite observation feature matrix and a satellite signal observation category feature matrix based on satellite observation data and a three-dimensional electronic orbit map database. It is used to train an observation category classification model based on the SVM model training observation data. The observation categories include line-of-sight signals and non-line-of-sight signals.

[0014] The sky visibility feature matrix generation module is used to divide the elevation and azimuth angles of each satellite at the observation epoch into equal angular intervals to generate an identity matrix; to perform mean processing on the signal observation category information of the satellite in the row and column of the identity matrix at each observation epoch to obtain the signal observation category value; and to generate the sky visibility feature matrix by filling the empty element positions in the identity matrix through bilinear interpolation.

[0015] The encoding module is used to divide the sky visibility feature matrix into blocks of size L×L to obtain block matrices, flatten each block matrix to obtain block feature vectors, and mark the scene category to which each block feature vector belongs and its corresponding position in the sky visibility feature matrix.

[0016] The second model training module is used to train the observed scene category classification model based on the ViT model using the processed block feature vectors to obtain the scene category classification results.

[0017] This specification provides one or more embodiments of a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the train satellite positioning observation scene classification and identification method described above.

[0018] This specification provides one or more embodiments of a computer-readable storage medium storing a computer program, characterized in that, when executed by a processor, the computer program implements the steps of the train satellite positioning observation scene classification and identification method described above.

[0019] This disclosure provides a method, device, equipment, and medium for classifying and identifying train satellite positioning observation scenarios. Its advantages lie in using a support vector machine model to train observation data, generating LOS signal and non-line-of-sight (NLOS) signal identification decision functions, forming visibility identification decision results for all satellites corresponding to their numbers at each observation epoch. Furthermore, it improves upon the traditional one-dimensional azimuth-angle sky visibility boundary sequence condition by dividing the elevation and azimuth angles into equal angular intervals to generate an identity matrix. Based on the signal observation category information of the observed satellites in the matrix's row and column at each observation epoch, the observation category value is obtained through mean processing, and empty element positions are filled using bilinear interpolation. This constructs a two-dimensional sky visibility feature matrix. By increasing the dimensionality, it can supplement the identification of scenarios with discontinuous sky visibility features. Finally, the scene category classification result is obtained through an observation scene category classification model based on the ViT model. This method can identify complex scenarios along train operation lines, improving the stability of satellite positioning optimization methods under different scenarios. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of a first scenario of a train satellite positioning observation scene classification and identification method provided in one or more embodiments of this specification;

[0022] Figure 2 A block diagram of a train satellite positioning observation scene classification and identification device provided in one or more embodiments of this specification;

[0023] Figure 3 This is a schematic diagram of the structure of a computer device provided for one or more embodiments of this specification. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this invention.

[0025] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings.

[0026] Method Implementation Examples

[0027] According to embodiments of the present invention, a method for classifying and identifying train satellite positioning observation scenarios is provided, such as... Figure 1 The diagram shown is a flowchart of the train satellite positioning observation scene classification and identification method provided in this embodiment. The train satellite positioning observation scene classification and identification method according to this embodiment includes:

[0028] Step S1: Obtain satellite observation data for each observation epoch, and extract satellite position observation information under limited scenarios from the observation data, including satellite navigation system category, satellite number, satellite elevation angle, azimuth angle, and signal-to-noise ratio.

[0029] In this embodiment, satellite position observation information data under limited scenarios is extracted from the NMEA data acquired by the satellite receiver during train operation. satellite This includes satellite navigation system, satellite azimuth, satellite elevation, and satellite signal-to-noise ratio. A confined scenario refers to a situation, excluding open environments, where environmental features obstruct satellite signals during reception, resulting in LOS and NLOS signals. Specific satellite position observation information is as follows:

[0030]

[0031] Where, SYS = [sys 11 ,sys 12 ,…,sys 1l ,…,sys kl ] T The satellite navigation system is designated as a number, with values ​​ranging from 1-GPS / 2-BDS / 3-GLONASS;

[0032] AZI = [azi] 11 ,azi 12 ,…,azi 1l ,…,azi kl ] TThis corresponds to the azimuth angle of the observed satellite in the satellite navigation system;

[0033] ELE = [ele 11 ,ele 12 ,…,ele 1l ,…,ele kl ] T This corresponds to the elevation angle of the observed satellite in the satellite navigation system;

[0034] SNR=[snr 11 ,snr 12 ,…,snr 1l ,…,snr kl ] T This refers to the signal-to-noise ratio of observed satellite signals under the corresponding satellite navigation system.

[0035] Step S2: Based on satellite observation data and a three-dimensional electronic orbital map database, generate a satellite observation feature matrix M and a satellite signal observation category feature matrix T, which are used to train an observation category classification model based on the observation data. The observation categories include line-of-sight signals and non-line-of-sight signals.

[0036] In this embodiment, the three-dimensional electronic map database includes track route information, track key point information, and three-dimensional environmental information; the satellite signal propagation path is determined using the three-dimensional electronic map, and Data is generated. satellite The corresponding satellite observation feature matrix M and satellite signal observation category feature matrix T, where the satellite observation feature matrix M = [M1 … M k ] T M k The satellite observation feature matrix for the k-th observation epoch is as follows:

[0037]

[0038] Among them, ele kl ,azi kl and SNR kl These correspond to the satellite elevation angle, satellite azimuth angle, and satellite signal-to-noise ratio of the l-th satellite at the k-th observation epoch, respectively.

[0039] The satellite signal observation category feature matrix T is specifically represented as follows:

[0040] T = [t] 11 … t 1l … t kl ] T ;

[0041] Among them, t klLet be the signal observation category of the l-th satellite at the k-th observation epoch. The observation category includes two types: line-of-sight (LOS) signals and non-line-of-sight (NLOS) signals.

[0042] In this embodiment, the classification model is one that can classify both LOS and NLOS signals, such as the Naive Bayes classification model or SVM (Support Vector Machine). This embodiment preferably uses an SVM model as the base model for the observation category classification, as detailed below:

[0043] The observed feature matrix M and observed category matrix T are input into the SVM model as training and validation data. After iterative training on the training set, the optimal performance on the validation set is determined, and then the model parameters are determined, including the minimized weight parameter w, bias term b, Lagrange multiplier α, regularization parameter C, and kernel function k(x,x').

[0044] The above parameters form a decision function f(x), which outputs satellite signal observation category information for each observation epoch, i.e., line-of-sight (LOS) signal or non-line-of-sight (LOS) signal. To facilitate distance calculation in the subsequent classification model, this embodiment sets p when the satellite signal type is LOS signal. i =1, setting p when the satellite signal type is non-line-of-sight (NLOS). i =-1.

[0045] In this preferred embodiment, since it is impossible to judge all satellite signals using the three-dimensional electronic map method, Data is used instead. satellite The elevation angle (ELE), signal-to-noise ratio (SNR), and satellite signal observation category feature matrix T are used as inputs to the SVM model to generate a training dataset R = {(x1,y1),(x2,y2),L(x...}. N ,y N Training observation category classification models based on support vector machines (SVMs) for LOS and NLOS is performed. The specific classification process of the SVM model is as follows.

[0046] Given a dataset R and a hyperplane wx+b=0 for separating LOS and NLOS signals, the hyperplane is defined relative to the sample point (x i y i The geometric distance interval is:

[0047]

[0048] The minimum geometric distance interval of the hyperplane with respect to all satellite sample points is:

[0049]

[0050] The problem of solving the maximum LOS and NLOS signal segmentation plane problem using the SVM model is transformed into the following constrained optimization problem:

[0051]

[0052] Solving the maximum hyperplane problem in an SVM model can be represented as a constrained optimization problem:

[0053]

[0054] sty i (w·x i +b)≥1, i=1,2,...N;

[0055] The constrained original objective function is transformed into an unconstrained Lagrange objective function using Lagrange multipliers:

[0056]

[0057] Where, α i ≥0 represents a Lagrange multiplier.

[0058] For the LOS and NLOS signals in the training set, the following KKT (Karush-Kuhn-Tucker) conditions are satisfied:

[0059]

[0060] Choosing a penalty function C>0, a Gaussian kernel function K(x,z), and a Lagrange objective function L(w,b,α), construct and solve a convex quadratic programming problem:

[0061]

[0062] 0≤α i ≤C, i=1,2,…N;

[0063] Obtain the optimal solution to the planning problem Choose α * One of the components Meet the conditions The bias parameter b is calculated. * :

[0064]

[0065] Substituting the Gaussian kernel function K(x,z) into the equation yields the classification decision function:

[0066]

[0067] Here, sign(*) is a classification function that classifies values ​​greater than 0 as 1 and values ​​less than 0 as -1.

[0068] Step S3: Divide the elevation and azimuth angles of each satellite at equal angular intervals to generate an identity matrix; perform mean processing on the signal observation category information of the satellite in the row and column of the identity matrix at each observation epoch to obtain the signal observation category value; and fill the empty element positions in the identity matrix with bilinear interpolation to generate a sky visibility feature matrix.

[0069] In this embodiment, to construct the sky visibility feature matrix, the elevation and azimuth angles of the sky area above the satellite receiver antenna are divided into equal intervals of q° to generate an identity matrix. The satellite position matrix in the sky is obtained as follows:

[0070]

[0071] Where 1≤i≤m, 1≤j≤n, and In this preferred embodiment, q can take values ​​of 5, 10, or 15.

[0072] In this embodiment, after dividing the elevation and azimuth angles of each satellite at the observation epoch into equal angular intervals, there will be cases where two satellites appear in the same position. For example, satellite A has an elevation angle of 61° and an azimuth angle of 71°, and satellite B has an elevation angle of 69° and an azimuth angle of 79°. These two satellites are in the same position in the matrix, so it is necessary to form a vector and take the average value, as follows.

[0073] Based on the satellite's azimuth and elevation information, if the position p is determined... ij Given satellite distribution, a support vector machine is used to determine the satellite propagation paths, obtaining the label value t∈{+1, -1} for the propagation path type, indicating whether it is a line-of-sight (line-of-sight) or non-line-of-sight (non-line-of-sight) signal. For the satellite positions p in the sky... ij The data contains K historical satellite observations, forming a satellite signal propagation vector. The mean value of the vector components is then calculated to obtain the element s in the sky visibility feature matrix. ij The constructed sky visibility feature matrix S is as follows:

[0074]

[0075] Where 1≤i≤m, 1≤j≤n, and

[0076] The operation of averaging the components of a vector is specifically as follows: For the position p of a satellite in the sky... ij The corresponding vector length ||I ij||>1, after averaging, the signal category value is obtained and assigned to s. ij ; for the position of satellites in the sky p ij The corresponding vector length ||I ij ||=1,s ij Take the vector value.

[0077] Firstly, the identity matrix encompasses all possible satellite locations. However, each observation involves approximately 20 satellites, resulting in many locations lacking satellite signal category information. Therefore, further investigation is needed.

[0078] Step S4: Divide the sky visibility feature matrix into blocks of size L×L to obtain block matrices. Flatten each block matrix to obtain block feature vectors and their corresponding position codes in the sky visibility feature matrix. According to the scene category corresponding to each sky visibility feature matrix, mark the scene category to which each block feature vector belongs to, and obtain the processed sky visibility feature vector sequence.

[0079] In this embodiment, the sky visibility feature matrix is ​​divided into blocks of size L×L to obtain a block matrix, which is specifically achieved through the following steps:

[0080] Because the ViT (Vision Transformer) model differs from traditional convolutional methods, the size of its block cuts must be divisible by the size of the original matrix to ensure that each block has a consistent size. Therefore, the value of L must be a divisor of n and m. For example, when q is 10, then m = 9 and n = 36; when L is 3, the sky visibility feature matrix is ​​divided into L×L blocks, resulting in 36 block feature matrices s. k .

[0081] In this embodiment, due to the characteristics of the ViT model, the sky visibility feature matrix needs to be resized and adjusted to a square matrix. Furthermore, due to the inconsistent sizes in the two directions during position encoding, the model's feature extraction may be biased to one side when adding position encoding to the final features. Therefore, to ensure the matrix size of the input ViT model meets the requirements and to reduce the number of rows in the matrix, and to ensure the block cutting balances the size and number of blocks as much as possible, the obtained sky visibility feature matrix is ​​first resized and adjusted. Specifically, this includes the following steps:

[0082] The portion of the sky visibility feature matrix with azimuth angles greater than 180° is extracted, arranged in reverse order, and then concatenated along the major axis with the portion with azimuth angles less than 180°. This ensures that the element s in the sky visibility feature matrix... ij (j≤18) is adjusted to element s (i+9)j Sky visibility feature matrix elements s ij (j>18) is readjusted to element s(9-i+1)(36-j+1) Finally, the transformed sky visibility feature matrix S is obtained. 18×18 .

[0083] In this embodiment, the block feature vector is obtained by flattening each block matrix, specifically including the following steps:

[0084] For each block matrix s k Transform it into a one-dimensional vector:

[0085]

[0086] Then for each one-dimensional vector s′ k Embedding is performed using a linear transformation (i.e., a fully connected layer), where the parameters of the linear transformation are given. Where D is the embedding dimension, and each one-dimensional vector s′ k It will then be embedded as a D-dimensional vector: e = E(s′) k ).

[0087] In this embodiment, since the ViT model does not have the ability to perceive the position of elements in the sky visibility feature matrix S corresponding to the input vector sequence, it is necessary to add position encoding to clarify the relative position of each block matrix. Specifically, the position encoding function is PE, and each vector e will be added with its corresponding position encoding PE(k), where k is the position index of the block matrix.

[0088] In this embodiment, each block matrix s is obtained through the above steps. k After the corresponding vector, each block matrix s k Each element corresponds to a scene category. This is then converted into a scene category vector cls using one-hot encoding. The cls vector is then appended to the beginning of the feature vector sequence to obtain a feature vector sequence containing the scene category.

[0089] Step S5: Use the sky visibility feature vector sequence to train the observation scene category classification model based on the ViT model to obtain the scene category classification result.

[0090] In this embodiment, the specific steps for training the ViT model to obtain the observation scene category classification model are as follows: after cleaning the satellite observation data, import it into the sky visibility feature training set generated by the support vector machine model with determined parameters after training; then input the processed sky visibility feature training set into the ViT model for iterative training; and determine the parameter weights in ViT based on the optimal results of the validation set, thereby determining the final network model.

[0091] In this embodiment, the specific processing steps for inputting the sequence S′ processed in step S4 into the ViT model include the following steps:

[0092] Step S51: The feature vector sequence S′ is input into the Encoder Block, and each vector in the sequence is encoded, denoted as f. E (S i ')=W E S i '+S i ', W E To be with S i In this embodiment, row vectors of equal length are used to encode each vector in the feature vector sequence S′, and the sequence is fed into the Encoder Block layer a total of 4 times to obtain a new sequence S” = f E 4 (S').

[0093] Step S52, new sequence S” = f E 4 (S') Input to the MLP layer, and compare each vector in the sequence S” with the MLP layer weight vector W. M Multiply the sequences to transform the sequence S” into a 10×1 vector, and then pass it through a linear transformation layer to obtain an unnormalized probability vector p of size n×1.

[0094] Step S52: Perform Softmax normalization on the probability vector p using the following formula to obtain the normalized probability vector p′ corresponding to the scene category;

[0095]

[0096] In this example, the i-th element of the probability vector p′ corresponds to the predicted probability of the i-th scene category. The scene category corresponding to the highest probability is output. In this embodiment, the ViT model uses a linear transformation as a classification head to classify the first element of the model output. Let the parameters of the linear transformation be... Where K is the number of categories of typical railway scenarios, the final classification output is z = Wy0, which gives the category information of the corresponding scenario, i.e. the prediction result of the ViT model.

[0097] The method provided in this embodiment utilizes a feature model to train observation data, generating a decision function for identifying LOS signals and non-line-of-sight (NLOS) signals. This results in visibility identification decision results for all satellites with corresponding numbers at each epoch. Furthermore, it improves upon the traditional one-dimensional azimuth sky visibility boundary sequence condition by dividing the elevation and azimuth angles into equal angular intervals to generate an identity matrix. Based on the signal observation category information of the observed satellites in the matrix's row and column at each observation epoch, the observation category value is obtained through mean processing. Empty element positions are filled using bilinear interpolation to obtain the sky visibility feature matrix. This construction of a two-dimensional sky visibility feature matrix, by increasing the dimensionality, can supplement the identification of scenarios with discontinuous sky visibility features. Finally, a scene category classification result is obtained through an observation scene category classification model based on the ViT model. This method can identify complex scenarios along train routes, improving the stability of satellite positioning optimization methods under different scenarios.

[0098] Device Examples

[0099] According to embodiments of the present invention, a train satellite positioning observation scene classification and identification device is provided, such as... Figure 1 The diagram shown is a block diagram of the train satellite positioning observation scene classification and identification device provided in this embodiment. The train satellite positioning observation scene classification and identification device according to this embodiment includes:

[0100] The data acquisition module 10 is used to acquire satellite observation data at each observation epoch and extract satellite position observation information under limited scenarios from the observation data, including satellite navigation system category, satellite number, satellite elevation angle, azimuth angle, and signal-to-noise ratio.

[0101] The first model training module 20 is used to generate a satellite observation feature matrix M and a satellite signal observation category feature matrix T based on satellite observation data and a three-dimensional electronic orbital map database. It is used to train an observation category classification model based on the observation data trained by the classification model. The observation categories include line-of-sight signals and non-line-of-sight signals.

[0102] In this embodiment, the three-dimensional electronic map database includes track route information, track key point information, and three-dimensional environmental information; the satellite signal propagation path is determined using the three-dimensional electronic map, and Data is generated. satellite The corresponding satellite observation feature matrix M and satellite signal observation category feature matrix T, where the satellite observation feature matrix M = [M1 … M n ] T M n The satellite observation feature matrix for the nth observation epoch is as follows:

[0103]

[0104] Among them, ele kl ,azi kl and SNR kl These correspond to the satellite elevation angle, satellite azimuth angle, and satellite signal-to-noise ratio of the l-th satellite at the k-th observation epoch, respectively.

[0105] The satellite signal observation category feature matrix T is specifically represented as follows:

[0106] T = [t] 11 … t 1l … t kl ] T ;

[0107] Among them, t kl Let be the signal observation category of the l-th satellite at the k-th observation epoch. The observation category includes two types: line-of-sight (LOS) signals and non-line-of-sight (NLOS) signals.

[0108] In this embodiment, the SVM model is preferably used as the base model for the observation category classification model, as detailed below:

[0109] The observed feature matrix M and observed category matrix T are input into the SVM model as training and validation data. After iterative training on the training set, the optimal performance on the validation set is determined, and then the model parameters are determined, including the minimized weight parameter w, bias term b, Lagrange multiplier α, regularization parameter C, and kernel function k(x,x').

[0110] The above parameters form a decision function f(x), which outputs satellite signal observation category information for each observation epoch, i.e., line-of-sight (LOS) signal or non-line-of-sight (LOS) signal. To facilitate distance calculation in the subsequent classification model, this embodiment sets p when the satellite signal type is LOS signal. i =1, setting p when the satellite signal type is non-line-of-sight (NLOS). i =-1.

[0111] In this preferred embodiment, since it is impossible to judge all satellite signals using the three-dimensional electronic map method, Data is used instead. satellite The elevation angle (ELE), signal-to-noise ratio (SNR), and satellite signal observation category feature matrix T are used as inputs to the SVM model to generate a training dataset R = {(x1,y1),(x2,y2),L(x...}. N ,y N Train LOS and NLOS observation category classification models based on support vector machines.

[0112] The sky visibility feature matrix generation module 30 is used to divide the elevation and azimuth angles of each satellite at the observation epoch into equal angular intervals to generate an identity matrix; to perform mean processing on the signal observation category information of the satellite in the row and column of the identity matrix at each observation epoch to obtain the signal observation category value; and to generate the sky visibility feature matrix by filling the empty element positions in the identity matrix through bilinear interpolation.

[0113] In this embodiment, the sky visibility feature matrix generation module 30 includes an initial matrix generation submodule 301 and a matrix filling submodule 302; wherein...

[0114] The initial matrix generation submodule 301 is configured to perform the following steps:

[0115] Divide the elevation and azimuth angles of the sky region above the satellite receiver antenna into equal intervals of q° to generate an identity matrix. The satellite position matrix in the sky is obtained as follows:

[0116]

[0117] Where 1≤i≤m, 1≤j≤n, and In this preferred embodiment, q can take values ​​such as 5, 10, and 15.

[0118] Based on the satellite's azimuth and elevation information, if the position p is determined... ij Given satellite distribution, a support vector machine is used to determine the satellite propagation paths, obtaining the label value t∈{+1, -1} for the propagation path type, indicating whether it is a line-of-sight (line-of-sight) or non-line-of-sight (non-line-of-sight) signal. For the satellite positions p in the sky... ij The data contains K historical satellite observations, forming a satellite signal propagation vector. The mean value of the vector components is then calculated to obtain the element s in the sky visibility feature matrix. ij The constructed sky visibility feature matrix S is as follows:

[0119]

[0120] Where 1≤i≤m, 1≤j≤n, and

[0121] The operation of averaging the components of a vector is specifically as follows: For the position p of a satellite in the sky... ij The corresponding vector length ||I ij ||>1, after averaging, the signal category value is obtained and assigned to s. ij ; for the position of satellites in the sky p ij The corresponding vector length ||I ij ||=1,s ij Take the vector value.

[0122] Matrix filling submodule 302 is configured to perform the following steps:

[0123] Based on the existing location category information, bilinear interpolation is used to fill in locations without satellites. For locations that cannot meet the bilinear interpolation conditions, the nearest element value is used for filling, as detailed below:

[0124] If the position of the satellite in the sky is p ij There is no satellite distribution, i.e., the vector length ||I ij If ||=0, then bilinear interpolation is used to calculate and determine s. ij For s that do not satisfy bilinear interpolation ij Let its value be the nearest s. ac The value is determined as follows:

[0125]

[0126] For s that do not satisfy bilinear interpolation ij For any assigned value s ac And |(ia,jc)| min , making s ij =s ac .

[0127] In this embodiment, the sky visibility feature matrix obtained through the above processing is used to obtain a sky visibility matrix based on the actual satellite observations in each region. With the sky visibility matrix, the actual LOS / NLOS of the observations can be obtained directly through the matrix correspondence.

[0128] The encoding module 40 is used to divide the sky visibility feature matrix into blocks of size L×L to obtain block matrices, and to flatten each block matrix to obtain block feature vectors and their corresponding position codes in the sky visibility feature matrix. Based on the scene category corresponding to each sky visibility feature matrix and the scene category corresponding to each block feature vector, the processed sky visibility feature vector sequence is obtained.

[0129] In this embodiment, the value of L must be a divisor of n and m, and the sky visibility feature matrix is ​​divided into blocks of L×L.

[0130] In this embodiment, the matrix size of the input ViT model is made to meet the requirements, and the number of rows in the matrix is ​​reduced. The segmentation of blocks is designed to balance the size and number of segments after segmentation. First, the obtained sky visibility feature matrix is ​​resized and adjusted. A matrix conversion module 60 is set up to extract the portion of the sky visibility feature matrix with an azimuth angle greater than 180°, arrange it in reverse order, and then concatenate it along the major axis with the portion with an azimuth angle less than 180°. This ensures that the elements s in the sky visibility feature matrix are within the correct order.ij (j≤18) is adjusted to element s (i+9)j Sky visibility feature matrix elements s ij (j>18) is readjusted to element s (9-i+1)(36-j+1) Finally, the transformed sky visibility feature matrix S is obtained. 18×18 .

[0131] In this embodiment, the encoding module 40 includes a vector conversion submodule 401 and a tagging module 402, wherein...

[0132] The vector transformation submodule 401 is used to flatten the block matrix obtained by dividing it into blocks of size L×L to obtain the block feature vector;

[0133] The labeling module 402 is used to label the scene category to which each block feature vector belongs and its corresponding position encoding in the sky visibility feature matrix.

[0134] In this embodiment, the vector transformation submodule 401 performs a flattening operation on each block matrix to obtain the block feature vector, specifically including the following steps:

[0135] For each block matrix s k Transform it into a one-dimensional vector:

[0136]

[0137] Then for each one-dimensional vector s′ k Embedding is performed using a linear transformation (i.e., a fully connected layer), where the parameters of the linear transformation are given. Where D is the embedding dimension, and each one-dimensional vector s′ k It will then be embedded as a D-dimensional vector: e = E(s' k ).

[0138] The labeling module 402 labels the scene category to which each feature vector belongs and its corresponding position encoding in the sky visibility feature matrix. This is achieved through the following steps:

[0139] Based on the obtained block matrices s k The corresponding vectors, each block matrix s k Each element corresponds to a scene category. This is then converted into a scene category vector cls using one-hot encoding. The cls vector is then appended to the beginning of the feature vector sequence to obtain a feature vector sequence containing the scene category.

[0140] In this embodiment, since the Transformer model does not have the ability to perceive the position of the elements in the sky visibility feature matrix S corresponding to the input vector sequence, it is necessary to add position encoding to clarify the relative position of each block matrix. Specifically, the position encoding function is PE, and each vector e will be added with its corresponding position encoding PE(k), where k is the position index of the block matrix.

[0141] The second model training module 50 is used to train an observation scene category classification model based on the ViT model by using the sky visibility feature vector sequence to obtain scene category classification results.

[0142] The device provided in this embodiment uses a support vector machine model to train observation data, generates LOS signal and non-line-of-sight (NLOS) signal identification decision function, and forms the visibility identification decision result for all satellites corresponding to their numbers at each observation epoch. It further improves upon the traditional one-dimensional azimuth sky visibility boundary sequence condition by dividing the elevation and azimuth angles into equal angular intervals to generate an identity matrix. Based on the signal observation category information of the observed satellites in the matrix's row and column at each observation epoch, the observation category value is obtained through mean processing, and empty element positions are filled using bilinear interpolation. This constructs a two-dimensional sky visibility feature matrix. By increasing the dimensionality, it can supplement the identification of scenarios with discontinuous sky visibility features. Finally, the scene category classification result is obtained through an observation scene category classification model based on the ViT model. This method can identify complex scenarios along train operation lines, improving the stability of the satellite positioning optimization method's optimization level under different scenarios.

[0143] The method provided in this embodiment uses a classification model to train observation data, generates LOS and NLOS signal identification decision functions, and forms visibility identification decision results for all satellites corresponding to their numbers at each epoch. It further improves upon the traditional one-dimensional azimuth sky visibility boundary sequence condition by dividing the elevation and azimuth angles into equal angular intervals to generate an identity matrix. Based on the signal observation category information of the observed satellites in the matrix's row and column at each observation epoch, the observation category value is obtained through mean processing, and empty element positions are filled using bilinear interpolation to obtain the sky visibility feature matrix. This construction of a two-dimensional sky visibility feature matrix, by increasing the dimensionality, can supplement the identification of scenarios with discontinuous sky visibility features. Finally, the scene category classification result is obtained through an observation scene category classification model based on the ViT model. This device can identify complex scenarios along train operation lines, improving the stability of the satellite positioning optimization method's optimization level under different scenarios.

[0144] The embodiments of the present invention are device embodiments corresponding to the above method embodiments. The specific operations of each module processing step can be understood with reference to the description of the method embodiments, and will not be repeated here.

[0145] like Figure 3 As shown, the present invention also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the train satellite positioning observation scene classification and identification method in the above embodiments, or when the computer program is executed by a processor, it implements the train satellite positioning observation scene classification and identification method in the above embodiments. When the computer program is executed by the processor, it implements the following method steps:

[0146] Step S1: Obtain satellite observation data for each observation epoch, and extract satellite position observation information under limited scenarios from the observation data, including satellite navigation system category, satellite number, satellite elevation angle, azimuth angle, and signal-to-noise ratio.

[0147] Step S2: Based on satellite observation data and a three-dimensional electronic orbital map database, generate a satellite observation feature matrix M and a satellite signal observation category feature matrix T, which are used to train an observation category classification model based on the observation data. The observation categories include line-of-sight signals and non-line-of-sight signals.

[0148] Step S3: Divide the elevation and azimuth angles of each satellite at equal angular intervals to generate an identity matrix; perform mean processing on the signal observation category information of the satellite in the row and column of the identity matrix at each observation epoch to obtain the signal observation category value; and fill the empty element positions in the identity matrix with bilinear interpolation to generate a sky visibility feature matrix.

[0149] Step S4: Divide the sky visibility feature matrix into blocks of size L×L to obtain block matrices. Flatten each block matrix to obtain block feature vectors and their corresponding position codes in the sky visibility feature matrix. According to the scene category corresponding to each sky visibility feature matrix, mark the scene category to which each block feature vector belongs to, and obtain the processed sky visibility feature vector sequence.

[0150] Step S5: Use the sky visibility feature vector sequence to train the observation scene category classification model based on the ViT model to obtain the scene category classification result.

[0151] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0152] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and the contents not described in detail in the specification of the present invention are well known to those skilled in the art.

Claims

1. A method for classifying and identifying train satellite positioning observation scenes, characterized in that, Includes the following steps: Acquire satellite observation data at each observation epoch and extract satellite position observation information under limited scenarios from the observation data; Based on satellite observation data and a 3D electronic orbital map database, a satellite observation feature matrix and a satellite signal observation category feature matrix are generated to train an observation category classification model based on an SVM model. The observation categories include line-of-sight signals and non-line-of-sight signals. The elevation and azimuth angles of each satellite at each observation epoch are divided into equal angular intervals to generate an identity matrix; the signal observation category information of each satellite in the row and column of the identity matrix at each observation epoch is averaged to obtain the signal observation category value; and the sky visibility feature matrix is ​​generated by filling the empty element positions in the identity matrix through bilinear interpolation. The sky visibility feature matrix is ​​divided into blocks of size L×L to obtain a block matrix, where L is a divisor of n and m. m = , n = , q To divide the elevation and azimuth angles of each satellite at the observation epoch, the block matrix is ​​flattened to obtain the block feature vectors and their corresponding position codes in the sky visibility feature matrix. Based on the scene category corresponding to each sky visibility feature matrix, the scene category to which each block feature vector belongs is marked, and the processed sky visibility feature vector sequence is obtained. The sequence of sky visibility feature vectors is used to train an observation scene category classification model based on the ViT model to obtain scene category classification results.

2. The train satellite positioning observation scene classification and identification method as described in claim 1, characterized in that, The specific steps for generating the satellite observation feature matrix and satellite signal observation category feature matrix based on satellite observation data and a three-dimensional electronic orbital map database are as follows: By determining the satellite signal propagation path using a 3D electronic map, a satellite observation feature matrix and a satellite signal observation category feature matrix corresponding to the satellite position observation information are generated, as detailed below: Satellite observation feature matrix M n For the first n The satellite observation feature matrix for each observation epoch is as follows: ； in, , and Corresponding to the number k The first observation epoch l The satellite's elevation angle, azimuth angle, and signal-to-noise ratio; The satellite signal observation category feature matrix T is: ； in, t kl For the first k The first observation epoch l The satellite's signal observation categories include two types: LOS signals and NLOS signals.

3. The train satellite positioning observation scene classification and identification method as described in claim 1, characterized in that, The specific training of the observation category classification model based on the SVM model is as follows: Based on the hyperplane wx+b=0 for segmenting LOS and NLOS signals, the hyperplane is defined relative to the sample points ( x i , y i The geometric distance interval is: ； The minimum geometric distance interval of the hyperplane with respect to all satellite sample points is: ； The problem of solving the maximum LOS and NLOS signal segmentation plane problem using the SVM model is transformed into the following constrained optimization problem: ； Solving the maximum hyperplane problem in an SVM model can be represented as a constrained optimization problem: ； The constrained original objective function is transformed into an unconstrained Lagrange objective function using Lagrange multipliers: ； in, For Lagrange multipliers; For the LOS and NLOS signals in the training set, the following KKT conditions are satisfied: ； Choose a penalty function Gaussian kernel function and Lagrange objective function Construct and solve the convex quadratic programming problem: ； Obtain the optimal solution to the planning problem ,choose One of the components Meet the conditions The bias parameters are calculated. : ； Gaussian kernel function Substituting the values, we obtain the classification decision function: 。 4. The train satellite positioning observation scene classification and identification method as described in claim 1, characterized in that, The elevation and azimuth angles of each satellite at the observation epoch were respectively... q ° interval division, where, q The value can be 5, 10, or 15.

5. The train satellite positioning observation scene classification and identification method as described in claim 1, characterized in that, It also includes the following steps: The portion of the sky visibility feature matrix with azimuth angles greater than 180° is extracted, arranged in reverse order, and then concatenated along the major axis with the portion with azimuth angles less than 180°, so that the elements in the sky visibility feature matrix are... s ij ( j ≤18) are adjusted to elements s （i+9）j Elements in the sky visibility feature matrix s ij ( j >18) Re-adjusted as an element s （9-i+1）（36-j+1） Obtain the transformed sky visibility feature matrix. S 18×18 ,in, ,and m = , n = , q The angles for dividing the elevation and azimuth of each satellite at the observation epoch.

6. The train satellite positioning observation scene classification and identification method as described in claim 5, characterized in that, The block matrix s k Transform it into a one-dimensional vector: thing( )=( ) T ; For each one-dimensional vector Embedding is performed using a linear transformation, where the parameters of the linear transformation are... ,in, For the embedding dimension, each one-dimensional vector Embedded as a 3D vector: .

7. The train satellite positioning observation scene classification and identification method as described in claim 6, characterized in that, The block matrix s k The scene categories corresponding to the sky visibility feature matrices are converted into scene category vectors using one-hot encoding. cls and class vector cls Adding to the head of the feature vector sequence yields a feature vector sequence containing the scene category.

8. A train satellite positioning observation scene classification and identification device, characterized in that, include: The data acquisition module is used to acquire satellite observation data at each observation epoch and extract satellite position observation information under limited scenarios from the observation data; The first model training module is used to generate satellite observation feature matrix and satellite signal observation category feature matrix based on satellite observation data and a three-dimensional electronic orbit map database. It is used to train an observation category classification model based on an SVM model, where the observation categories include line-of-sight signal and non-line-of-sight signal categories. The sky visibility feature matrix generation module is used to divide the elevation and azimuth angles of each satellite at the observation epoch into equal angular intervals to generate an identity matrix; to perform mean processing on the signal observation category information of the satellite in the row and column of the identity matrix at each observation epoch to obtain the signal observation category value; and to generate the sky visibility feature matrix by filling the empty element positions in the identity matrix through bilinear interpolation. The encoding module is used to divide the sky visibility feature matrix into blocks of size L×L to obtain a block matrix, where L is a divisor of n and m. m = , n = , q To divide the elevation and azimuth angles of each satellite at the observation epoch, and to obtain the position encoding of the block feature vector in the sky visibility feature matrix by flattening each block matrix, and to mark the scene category to which each block feature vector belongs according to the scene category corresponding to each sky visibility feature matrix, the processed sky visibility feature vector sequence is obtained. The second model training module is used to train an observation scene category classification model based on the ViT model by using the sequence of sky visibility feature vectors to obtain scene category classification results.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the train satellite positioning observation scene classification and identification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the train satellite positioning observation scene classification and identification method as described in any one of claims 1 to 7.

Images