Single point positioning method based on neural network and fusing meteorological data and satellite ephemeris

Abstract

The application discloses a single point positioning method based on a neural network and fusing meteorological data and satellite ephemeris, relates to the cross technical field of satellite navigation and artificial intelligence, and aims at the problems of the single point positioning of the global navigation satellite system being restricted by the troposphere delay and the error accumulation of the traditional step-by-step solving architecture. The application fuses meteorological data and satellite ephemeris data, constructs a neural network architecture, extracts the correlation characteristics of temperature, air pressure, humidity, the satellite elevation angle and the azimuth angle by the meteorological branch, extracts the inherent characteristics of the satellite ephemeris data by the ephemeris branch, the attention fusion layer is used for dynamically weighting and fusing the output characteristics of the two branches, the output layer regresses and outputs the three-dimensional coordinates of the user and the receiver clock difference, and the atmospheric physical constraint is embedded in the neural network training process to ensure that the network output result conforms to the atmospheric physical law. The method breaks through the traditional step-by-step process, eliminates the error accumulation, realizes the sub-meter positioning precision in the dynamic atmospheric environment, and provides a new scheme for high-precision single point positioning.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of satellite navigation and artificial intelligence, and in particular to a single-point positioning method based on neural networks and integrating meteorological data and satellite ephemeris. Background Technology

[0002] Global Navigation Satellite Systems (GNSS) point positioning has long been hampered by meter-level errors caused by tropospheric delay. Traditional solutions employ a "step-by-step calculation" architecture: first, the signal delay is calculated using empirical models (such as Saastamoinen) or meteorological sensors, and then the coordinates are solved by substituting them into the observation equations. However, this architecture has a fundamental flaw—the progressive propagation of errors makes it difficult to overcome the accuracy ceiling: on the one hand, the lack of meteorological parameters results in wet delay correction residuals as high as 30%; on the other hand, the positioning calculation stage amplifies previous errors, with planar errors reaching 4.8 meters when satellite geometry deteriorates.

[0003] In recent years, neural networks have been introduced into the field of positioning in an attempt to alleviate this problem, but existing technologies have fallen into new predicaments: modified neural networks only replace traditional delay models and still need to be connected in series with geometric solution modules, which cannot avoid the effect of error accumulation; geometric end-to-end models directly output coordinates but ignore the physical relationship between meteorological disturbances and positioning results, and perform out of control in dynamic atmospheric environments; multimodal fusion attempts to simply splice meteorological data and ephemeris into the input, but the convergence failure rate exceeds 40% due to the mismatch of spatiotemporal scales of heterogeneous features.

[0004] None of the above solutions have overcome the core challenge: the deep physical coupling between meteorological parameters (temperature / pressure / humidity) and satellite ephemeris (position / clock error) (atmospheric refraction alters the signal path, and ephemeris configuration determines error sensitivity), resulting in long-term fragmented processing of the two in technical implementation. Establishing a data-driven end-to-end mapping that simultaneously satisfies atmospheric physical constraints has become the final barrier hindering high-precision single-point positioning. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a single-point positioning method based on neural networks and integrating meteorological data and satellite ephemeris. By deeply coupling meteorological data and satellite ephemeris data, the neural network directly outputs the user's three-dimensional coordinates, eliminating the error accumulation of traditional step-by-step calculations. Furthermore, atmospheric physical constraints are embedded in the neural network to achieve sub-meter positioning accuracy under dynamic meteorological conditions.

[0006] To achieve the above objectives, the present invention adopts the following technical solution, including:

[0007] A single-point positioning method based on neural networks and fusing meteorological data with satellite ephemeris includes the following steps:

[0008] S1, a meteorological sensor is integrated into the single-point positioning terminal for real-time collection of meteorological data, including temperature T, air pressure P, and humidity RH; the single-point positioning terminal also acquires satellite ephemeris data, including satellite coordinates (x...). s ,y s ,z s The data includes clock error δt, pseudorange observation ρ, satellite elevation angle θ, and azimuth angle φ; meteorological data and satellite ephemeris data are processed synchronously, and the meteorological data and satellite ephemeris data are merged to construct a spatiotemporal joint feature tensor;

[0009] S2, Design a neural network; the network architecture adopts a two-branch physical constraint design, including a meteorological branch, an ephemeris branch, an attention fusion layer, and an output layer; the input of the neural network is a spatiotemporal joint feature tensor; the meteorological branch is used to extract the correlation features between temperature, air pressure, humidity and satellite elevation and azimuth angles; the ephemeris branch is used to encode satellite spatial configuration and clock bias characteristics, extracting the intrinsic features of satellite ephemeris data; the attention fusion layer is used to dynamically weight and fuse the output features of the two branches; the output layer is used to regress the fused features output by the attention fusion layer, outputting the user's three-dimensional coordinates (x, y, y). u ,y u ,z u and receiver clock bias δt u ;

[0010] S3, train the neural network. During the training process, atmospheric physical constraints are embedded to ensure that the output results conform to the laws of atmospheric physics. The trained neural network is then used for single-point localization.

[0011] Preferably, in step S1, the meteorological data is filtered by moving average to remove random noise, and outliers and missing values ​​are supplemented by interpolation; a multi-mode GNSS receiver is used to receive satellite broadcast ephemeris, analyze the satellite broadcast ephemeris, and extract satellite ephemeris data; and millisecond-level time synchronization between meteorological data and satellite ephemeris data is achieved through hardware-level PPS pulse signals, with a time synchronization error of less than 1ms.

[0012] Preferably, in step S1, a spatiotemporal joint feature tensor is constructed. for:

[0013]

[0014] Where N is the number of visible satellites; K = 10 dimensions;

[0015] The input spatiotemporal joint feature tensor X is normalized by a BN normalization layer to obtain the spatiotemporal joint feature tensor after BN normalization.

[0016] Preferably, in step S2, the processing procedure for the meteorological branch includes:

[0017] S211, 1D convolutional layer sliding extraction of correlation features between T, P, RH and θ, φ:

[0018]

[0019] in, This is the output of the 1D convolutional layer; kernel=3 indicates that the convolutional kernel size is set to 3.

[0020] S212, Introducing nonlinearity by using the ReLU activation function:

[0021]

[0022] in, for The output after the ReLU activation function;

[0023] S213, Suppress overfitting using Dropout layers:

[0024]

[0025] Among them, F meteo This refers to the characteristics output by the meteorological branch.

[0026] Preferably, in step S2, the processing of the ephemeris branch includes:

[0027] S221, acquire ephemeris data, including pseudorange observation ρ and satellite coordinates (x). s ,y s ,z s The clock difference δt constitutes the ephemeris branch vector X. eph =[ρ,x s ,y s ,z s ,δt s The ephemeris branch vector x is processed by a BN normalization layer. eph Standardization is performed to obtain the ephemeris branch vector after BN normalization layer processing.

[0028] S222, satellite space configuration encoded via a fully connected layer FC:

[0029]

[0030] Among them, W eph Let b be the weight matrix. eph For bias, for The output after passing through the fully connected layer;

[0031] S223, Introducing nonlinearity by using the ReLU activation function:

[0032]

[0033] in, for The output after the ReLU activation function;

[0034] S224, Suppress overfitting using Dropout layers:

[0035]

[0036] Among them, F eph The features output by the ephemeris branch.

[0037] Preferably, the processing procedure of the attention fusion layer includes:

[0038] S231, the feature F output by the meteorological branch meteo Feature F of the ephemeris branch output eph After dimensional adaptation, the features are concatenated to obtain the concatenated feature X. fusion The spliced ​​features X are processed by a BN normalization layer. fusion Standardization is performed to obtain the spliced ​​features after processing by the BN normalization layer.

[0039] S232, the attention weights α are calculated through a fully connected layer and a sigmoid activation:

[0040]

[0041] Among them, W a Let b be the weight matrix. a For the bias, σ is the Sigmoid function;

[0042] S233, features are fused using attention weight α:

[0043] F fuse =α·F meteo +(1-α)·F eph ;

[0044] Among them, F fuse The fused features are output by the attention fusion layer.

[0045] Preferably, the processing procedure of the output layer includes:

[0046] The fused feature F is directly mapped from the attention fusion layer output by a fully connected layer.fuse Location results:

[0047] [x u ,y u ,z u ,δt u ] = FC out (F fuse W out ,b out ) = W out ·F fuse +b out ;

[0048] Among them, W out b represents the output layer weights. out For bias, (x u ,y u ,z u ) represents the user's three-dimensional coordinates, δt u This refers to the receiver clock bias.

[0049] Preferably, in step S3, the atmospheric physical constraint is implemented through a loss function to ensure that the output results conform to the atmospheric physical laws of pressure-elevation gradient and temperature-elevation gradient;

[0050] Total loss function during neural network training for:

[0051]

[0052] in, These represent coordinate loss, air pressure-elevation gradient loss, and temperature-elevation gradient loss, respectively; k1, k2, and k3 are the corresponding weights.

[0053]

[0054] Where n is the number of samples. The output of the neural network contains the 3D coordinates of the i-th user. The actual 3D coordinates of the i-th user;

[0055]

[0056] Where m is the number of adjacent samples, z uj z uj-1 Output elevation for adjacent epochs, P j P j-1 To calibrate the gas pressure for adjacent epochs;

[0057]

[0058] Where k is the number of adjacent samples, z ulz ul-1 To output the elevation for adjacent epochs, T l T l-1 Temperatures of adjacent epochs;

[0059] During neural network training, the Adam gradient descent optimization algorithm is used to iteratively update the network parameters until the total loss function is reached. Convergence occurs or the maximum number of iterations is reached.

[0060] This invention provides a computer program product, which includes a computer program / instruction that, when executed by a processor, implements the aforementioned single-point positioning method based on neural networks and fusing meteorological data and satellite ephemeris.

[0061] The present invention provides an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the single-point positioning method based on neural networks and fusing meteorological data and satellite ephemeris.

[0062] The advantages of this invention are:

[0063] (1) This invention breaks through the error accumulation of traditional step-by-step calculation, skips intermediate steps such as tropospheric delay correction and geometric calculation, and directly outputs the positioning result through end-to-end neural network, eliminating the defect of error propagation at each level. The planar positioning error is reduced to 0.87 meters and the elevation error is 1.23 meters, achieving sub-meter level accuracy.

[0064] (2) The present invention can dynamically adapt to complex meteorological environments. It adopts an attention fusion mechanism to dynamically adjust the weight of meteorological and ephemeris features according to weather conditions, solves the problem of spatiotemporal scale mismatch of heterogeneous features, and maintains high accuracy under extreme weather conditions.

[0065] (3) This invention ensures the rationality of the results by embedding atmospheric physical constraints. It constrains the pressure-elevation and temperature-elevation gradient relationships through loss functions, ensuring that the output results conform to atmospheric physical laws and avoiding the physical logic loss of control of traditional data-driven models.

[0066] (4) The present invention integrates a lightweight meteorological sensor into a single-point positioning terminal, which can collect data autonomously without external dependence. The micro neural network design reduces computing power overhead and is suitable for real-time positioning needs in scenarios such as emergency rescue, disaster relief, and field exploration. It has low dependence and high practicality.

[0067] (5) This invention extracts key meteorological and ephemeris features through a dual-branch structure and achieves deep physical coupling through an attention mechanism, which solves the problem of high convergence failure rate caused by simple feature splicing in the prior art and significantly improves model stability.

[0068] (6) The fusion mechanism of the present invention can automatically increase the meteorological weight (α→1) in typhoon weather, focusing on the influence of meteorological branch features on positioning; and focus on ephemeris features (α→0) in clear weather, dynamically adjusting the feature weight according to different meteorological environments to improve the adaptability and accuracy of positioning. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of the method of the present invention.

[0070] Figure 2 This is a flowchart of an embodiment of the present invention.

[0071] Figure 3 This is a diagram showing the training effect of the model of the present invention. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] like Figure 1 As shown, the single-point positioning method of the present invention, which is based on neural networks and integrates meteorological data and satellite ephemeris, integrates lightweight meteorological sensors (specifically temperature, pressure, and humidity sensors) into the single-point positioning terminal. It utilizes autonomously collected real-time meteorological data combined with station spatiotemporal data (i.e., satellite ephemeris data) to train a micro-neural network, skipping intermediate steps such as tropospheric delay correction and geometric calculation. This breakthrough achieves high-precision positioning with no external dependence and low computational overhead. The specific process of the method of the present invention is as follows:

[0074] S1, synchronous processing of multi-source data.

[0075] The multi-source data includes: meteorological data such as temperature T, air pressure P, and humidity RH (sampling rate ≥ 1Hz) collected in real time by a lightweight meteorological sensor integrated in the single-point positioning terminal; ephemeris data such as satellite coordinates and clock differences obtained by the single-point positioning terminal; and latitude, longitude, and altitude data of the station used for calibration during neural network training.

[0076] For meteorological data, a moving average filtering method is used to remove random noise, and outliers and missing values ​​are processed. For ephemeris data, a multi-mode GNSS receiver is used to receive satellite broadcast ephemeris, parse the satellite broadcast ephemeris, and extract satellite ephemeris data, such as satellite coordinates, clock bias, and health status.

[0077] By fusing input vectors, a spatiotemporal joint feature tensor is constructed.

[0078]

[0079] Where N is the number of visible satellites; K = 10 dimensions, representing temperature T, air pressure P, humidity RH, satellite elevation angle θ, satellite azimuth angle φ, pseudorange observation value ρ, and satellite coordinates (x, y, y). s ,y s ,z s ), clock difference δt.

[0080] The input spatiotemporal joint feature tensor X is standardized using a Batch Normalization (BN) layer, with the following formula:

[0081]

[0082] Where, μ BN The mean, Let be the variance, γ and β be the learnable scaling and offset parameters, respectively, and ε be a small constant to prevent division by zero, where X is the spatiotemporal joint feature tensor. This is the spatiotemporal joint feature tensor after processing by the BN normalization layer.

[0083] S2, Design an end-to-end neural network with physical constraints: Design a micro neural network suitable for this localization task. The network adopts a hybrid structure of fully connected layers combined with convolutional neural networks.

[0084] The neural network architecture adopts a dual-branch physical constraint design, including a meteorological branch, an ephemeris branch, an attention fusion layer, and an output layer.

[0085] Meteorological Branch: Features are extracted using 1D convolutional layers with a kernel size of 3. This branch aims to extract the nonlinear correlation between temperature, air pressure, humidity, and satellite elevation / azimuth angles. The input consists of partial features containing meteorological data and satellite angle information. After 1D convolution, it captures feature patterns related to meteorological factors and satellite signal propagation path angles, providing crucial meteorological information for subsequent fusion.

[0086] Ephemeris Branch: Encodes satellite spatial configuration and clock bias characteristics through a fully connected layer. The input consists of ephemeris-related features such as satellite coordinates and clock bias. After processing by the fully connected layer, the spatial distribution and clock bias information of the satellites are encoded to reflect the inherent characteristics of the satellite ephemeris data, providing key information in the ephemeris dimension for fusion.

[0087] Attention Fusion Layer: The features from the two branches are dynamically weighted and fused through the attention fusion layer.

[0088] Output layer: The output layer uses physical enhancement constraints to ensure that the output conforms to the laws of atmospheric physics.

[0089] The specific handling methods for meteorological branches are as follows:

[0090] S211, 1D convolutional layer sliding extraction of correlation features between T, P, RH and θ, φ:

[0091]

[0092] in, This represents the output of T, P, RH and θ, φ through a 1D convolutional layer; kernel=3 indicates that the convolutional kernel size is set to 3.

[0093] S212, the ReLU activation function is applied to the output of the 1D convolutional layer to introduce nonlinearity:

[0094]

[0095] in for The output after the ReLU activation function;

[0096] S213, then suppress overfitting using the Dropout layer:

[0097]

[0098] Among them, F meteo This represents the characteristics of the meteorological branch output; p = 0.2 indicates that the Dropout rate is 0.2.

[0099] The specific handling method for ephemeris branches is as follows:

[0100] S221, Input ephemeris parameters are pseudorange observation value ρ, satellite coordinates (x... s ,y s ,z s The clock difference δt constitutes the ephemeris branch vector X. eph =[ρ,x s ,y s ,z s ,δt s The ephemeris branch vector X for all inputs is processed by a BN normalization layer. eph Standardization is performed, and the formula is as follows:

[0101]

[0102] Where, μ BN The mean, Let ε be the variance, γ and β be the learnable scaling and offset parameters, respectively, and ε be a small constant to prevent division by zero. The ephemeris branch vector after processing by the BN normalization layer;

[0103] S222, after normalization, uses a fully connected layer FC-coded satellite space configuration:

[0104]

[0105] in, It is a weight matrix. It's a bias, D in D is the input dimension. out For output dimensions; for The output after passing through the fully connected layer;

[0106] S223, the ReLU activation function is applied to the output of the fully connected layer (FC) to introduce nonlinearity:

[0107]

[0108] in, for The output after the ReLU activation function;

[0109] S224, then use the Dropout layer to suppress overfitting:

[0110]

[0111] Among them, F eph The features output by the ephemeris branch.

[0112] The attention fusion layer is processed in the following ways:

[0113] S231, the feature F output by the meteorological branch meteo Feature F of the ephemeris branch output eph After dimensional adaptation, the features are concatenated to obtain the concatenated feature X. fusion The spliced ​​features X are processed by a BN normalization layer. fusion Standardization is performed to obtain the spliced ​​features after processing by the BN normalization layer.

[0114]

[0115] Where, μ BN , Let ε be the mean and variance of the spliced ​​features, γ and β be learnable parameters, and ε be a small constant to prevent division by zero.

[0116] S232, the attention weights α are calculated through a fully connected layer and a sigmoid activation:

[0117]

[0118] Among them, W aIt is the weight matrix, b a It is the bias, and σ is the Sigmoid function;

[0119] S233, features are fused using attention weight α:

[0120] F fuse =α·F meteo +(1-α)·F eph ;

[0121] Among them, F fuse The fused features are output by the attention fusion layer.

[0122] Fusion feature F fuse Simultaneously, key meteorological and ephemeris information is retained to provide input for the output layer. The fusion mechanism of this invention can automatically increase meteorological weights (α→1) during typhoon weather, focusing on the impact of meteorological branch features on positioning; and focus on ephemeris features (α→0) during clear weather, dynamically adjusting feature weights according to different meteorological environments to improve the adaptability and accuracy of positioning.

[0123] The specific processing method for the output layer is as follows:

[0124] The fused feature F is directly mapped from the attention fusion layer output by a fully connected layer. fuse Location results:

[0125] [x u ,y u ,z u ,δt u ] = FC out (F fuse W out ,b out ) = W out ·F fuse +b out ;

[0126] Among them, W out It is the output layer weight, b out It is a bias, (x) u ,y u ,z u ) represents the user's three-dimensional coordinates, δt u This refers to the receiver clock bias.

[0127] S3, Training Neural Network: Atmospheric physical constraints are embedded during the training process to ensure that the output results conform to the laws of atmospheric physics. The trained neural network is used for single-point localization after training.

[0128] 70% of the spatiotemporal joint feature tensor data collected in step S1 is selected as training data, 20% as validation data, and 10% as test data.

[0129] The data is input into the neural network constructed in step S2 for training. The Adam gradient descent optimization algorithm is used to iteratively update the network parameters until the total loss function is reached. Convergence occurs or the maximum number of iterations is reached.

[0130] Atmospheric physical constraints are implemented through a loss function to ensure that the output conforms to the atmospheric physical laws of pressure-elevation gradient and temperature-elevation gradient. The total loss function during neural network training is... for:

[0131]

[0132] in, These are coordinate loss, air pressure-elevation gradient loss, and temperature-elevation gradient loss, respectively.

[0133] coordinate loss The L2 norm is used for calculation, and the PPK high-precision solution is used as the benchmark to measure the deviation between the network output coordinates and the true coordinates. The formula is as follows:

[0134]

[0135] Where n is the number of samples. The output of the neural network contains the 3D coordinates of the i-th user. Let be the actual 3D coordinates of the i-th user.

[0136] To ensure that the pressure-elevation gradient meets the constraints, the pressure-elevation gradient loss... The formula is:

[0137]

[0138] The standard atmospheric pressure vertical lapse rate is used to constrain the network output to conform to atmospheric physical laws. The formula, after discretization, can be expressed as:

[0139]

[0140] Where m is the number of adjacent samples, z uj z uj-1 Output elevation for adjacent epochs, P j P j-1 The gas pressure is calibrated for adjacent epochs.

[0141] To ensure that the temperature-elevation gradient satisfies the constraints, the temperature-elevation gradient loss... The formula is:

[0142]

[0143] Used to suppress the physical deviation caused by thermal expansion. The formula after discretization is as follows:

[0144]

[0145] Where k is the number of adjacent samples, z ul z ul-1 To output the elevation for adjacent epochs, T l T l-1 Temperatures of adjacent epochs.

[0146] Subsequently, the trained neural network is used for single-point localization. The test data is input into the trained neural network to output the user's three-dimensional coordinates and the receiver clock error.

[0147] Example 1

[0148] like Figure 2 The specific process of the embodiments of the present invention will be described in detail below.

[0149] Step 1, Hardware platform construction.

[0150] This invention first constructs a hardware platform integrating a low-power meteorological sensor and a multi-mode GNSS receiver. A meteorological sensor, such as the Bosch BME280, is selected, which has an accuracy of ±0.5 hPa pressure, ±0.5°C temperature, and ±3% humidity, and can collect environmental data (T, P, and RH) in real time. It is paired with a multi-mode GNSS receiver to receive observational data such as satellite ephemeris data. Through hardware-level PPS pulse signals, the meteorological sensor data and satellite observations are synchronized at the millisecond level, with a synchronization error of less than 1 ms, ensuring consistency of multi-source data in the time dimension and laying the foundation for subsequent fusion processing.

[0151] Step 2, raw data processing.

[0152] First, meteorological data calibration is required. The air pressure data in the raw meteorological data needs to undergo piecewise linear calibration. Specifically, when the air pressure P... cal When the pressure is <800 hPa, the compensation formula is used: P cal =0.97P raw +2.1, when the air pressure P cal When P >800 hPa, cal =1.02P raw The calibration coefficients mentioned above were obtained through stepped pressure testing in a barometric chamber to improve the accuracy of the pressure data and provide more reliable meteorological parameters for subsequent positioning calculations. Temperature (T) and humidity (RH) data can be directly used for subsequent processing.

[0153] Next is the ephemeris data analysis and feature generation section. In the ephemeris data analysis stage, the elevation angle θ, azimuth angle φ, and satellite coordinates (x, y) of each visible satellite need to be calculated from the data obtained from the multi-mode GNSS receiver. s ,y s ,z s The system calculates the satellite's position and clock bias δt. Simultaneously, it combines the user's approximate coordinates (obtained from the previous epoch's solution) to jointly generate spatial geometric features. These features will be used to construct subsequent network input tensors to reflect the spatial geometric relationship between the satellite and the user terminal, providing ephemeris-level information support for positioning calculations.

[0154] Step 3: Construct the spatiotemporal joint feature tensor.

[0155] The processed temperature T and calibration pressure P cal Humidity RH, and the elevation angle θ, azimuth angle φ, pseudorange ρ, and geocentric coordinates (x) of each visible satellite. s ,y s ,z s The clock difference δt is integrated to form a 10-dimensional feature vector. If the number of visible satellites is N, then an N×10 dimensional matrix is ​​finally formed. The advantage of this tensor structure design is that it enables the same meteorological parameters to be shared across satellites, forcing the network to learn the global coupling effect of atmospheric disturbances on multipath signals during training and inference, fully exploring the correlation between meteorological data and ephemeris data, and providing more comprehensive input information for accurate positioning.

[0156] Step 4: Build the neural network architecture.

[0157] Step 5: Train the neural network.

[0158] In the training process, the preprocessed dataset is first divided into training, validation, and test sets in a 7:2:1 ratio. The Adam optimizer is used, with an initial learning rate of 0.001, decaying to 0.5 every 10 epochs. The batch size is set to 64, and the number of training epochs is 100. During forward propagation, the meteorology branch extracts meteorological features through a 1D convolutional layer, while the ephemeris branch extracts ephemeris features through a fully connected layer. The two feature sets are then weighted and fused by an attention layer before the coordinates are output by the fully connected layer. During training, the total loss function is used... To optimize the objective, network parameters are updated through backpropagation, where Measuring coordinate error, and The pressure-elevation and temperature-elevation gradient relationships are constrained separately. During training, the performance is evaluated on the validation set every 5 epochs. If the loss on the validation set does not decrease for 10 consecutive epochs, training is stopped. Finally, the model's positioning accuracy and physical constraint satisfaction are evaluated on the test set.

[0159] like Figure 3 The image shown is a diagram illustrating the training effect of the model according to the present invention.

[0160] Step 6: Evaluation of training results.

[0161] Take 10% as test data for prediction testing. Input the test data into the training model obtained in step 5 to obtain the output value. Use the following two metrics to evaluate the prediction error:

[0162]

[0163] Under rigorous evaluation on the test set, the model trained in this embodiment achieves a deep fit between high-precision positioning and physical laws. Its average planar positioning error is as low as 0.87 meters, and its elevation positioning error is 1.23 meters, significantly breaking through the meter-level error limit compared to traditional methods. This fully meets the sub-meter positioning requirements for scenarios such as emergency rescue, disaster relief, and field exploration. At the physical constraint level, the pressure-elevation gradient deviation is only 0.12 m / hPa, and the temperature-elevation gradient deviation is controlled at 0.003 m / ℃. This fully demonstrates that the model, through optimization of the physical constraint loss function, effectively learns and follows atmospheric physical laws. In dynamic meteorological environments, it not only outputs high-precision positioning results but also ensures that the results conform to actual physical characteristics, verifying the significant effectiveness of this method in improving positioning accuracy and ensuring physical rationality.

[0164] Step 7: The effects of the present invention will be explained below in conjunction with experimental simulation results.

[0165] To fully verify the effectiveness of the method of the present invention, the following comparison scheme is set up:

[0166] 1. Traditional step-by-step solution scheme: The Saastamoinen model is used to calculate the tropospheric delay, and the least squares method is used for coordinate solution.

[0167] 2. Modified Neural Network Scheme: Only the traditional delay model is replaced by a neural network, while the geometric solution module is still used in the subsequent process.

[0168] 3. Traditional end-to-end neural network solution: Meteorological data and ephemeris are simply concatenated and input into the neural network to directly output coordinates, without considering physical constraints and feature coupling.

[0169] The results of the positioning accuracy comparison are shown in Table 1:

[0170] Table 1 Comparison Results of Positioning Accuracy

[0171]

[0172] As shown in Table 1, in simulation tests under different meteorological conditions, the method of this invention has significantly lower horizontal, vertical, and three-dimensional positioning errors than other comparative schemes. In extreme meteorological simulation scenarios such as typhoon passage, the traditional step-by-step solution scheme has a horizontal error exceeding 6 meters, while the method of this invention can still maintain a high accuracy within 1.27 meters, verifying its superiority in dynamic atmospheric environments.

[0173] The results of the comparison of physical constraint satisfaction are shown in Table 2:

[0174] Table 2 Comparison Results of Physical Constraint Satisfaction

[0175]

[0176] As shown in Table 2, the method of this invention, through optimization of the physical constraint loss function, significantly outperforms other schemes in terms of pressure-elevation and temperature-elevation gradient deviations. This indicates that the method of this invention effectively learns and follows atmospheric physical laws, resulting in more physically reasonable output results. In contrast, other schemes, lacking a physical constraint mechanism, exhibit significant deviations in reflecting the physical correlation between meteorological parameters and positioning results.

[0177] In summary, the method of this invention outperforms traditional schemes and existing neural network positioning methods in terms of positioning accuracy, physical constraint satisfaction, and model performance, verifying its innovation and practicality, and providing an effective solution for high-precision single-point positioning technology.

[0178] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A single-point positioning method based on neural networks and fusing meteorological data with satellite ephemeris, characterized in that, Includes the following steps: S1 integrates a meteorological sensor in a single-point positioning terminal for real-time collection of meteorological data, including temperature. T air pressure P ,humidity RH ; The single-point positioning terminal simultaneously acquires satellite ephemeris data, including satellite coordinates. Clock difference pseudorange observations Satellite elevation angle and azimuth Simultaneous processing of meteorological data and satellite ephemeris data, merging meteorological data and satellite ephemeris data, and constructing a spatiotemporal joint feature tensor; S2, Design the neural network; The network architecture employs a dual-branch physical constraint design, comprising a meteorological branch, an ephemeris branch, an attention fusion layer, and an output layer. The neural network input is a spatiotemporal joint feature tensor. The meteorological branch extracts the correlation features between temperature, air pressure, humidity, and satellite elevation and azimuth angles. The ephemeris branch encodes satellite spatial configuration and clock bias characteristics, extracting intrinsic features from satellite ephemeris data. The attention fusion layer dynamically weights and fuses the output features from the two branches. The output layer regresses the fused features output from the attention fusion layer, outputting the user's three-dimensional coordinates. and receiver clock bias ; S3, train the neural network. During the training process, atmospheric physical constraints are embedded to ensure that the output results conform to the laws of atmospheric physics. The trained neural network is used for single-point localization. In step S3, the atmospheric physical constraints are implemented through a loss function to ensure that the output results conform to the atmospheric physical laws of pressure-elevation gradient and temperature-elevation gradient. Total loss function during neural network training for: ； in, These are coordinate loss, air pressure-elevation gradient loss, and temperature-elevation gradient loss, respectively. These are the corresponding weights; ； in, n For the sample size, The first output of the neural network i Three-dimensional coordinates of each user For the true first i Three-dimensional coordinates of each user; ； in, m The number of adjacent samples. , Output elevation for adjacent epochs. , To calibrate the gas pressure for adjacent epochs; ； in, k The number of adjacent samples. , Output elevation for adjacent epochs. , Temperatures of adjacent epochs.

2. The single-point positioning method based on neural networks and fusing meteorological data and satellite ephemeris as described in claim 1, characterized in that, In step S1, the meteorological data is filtered by moving average to remove random noise, and outliers and missing values ​​are supplemented by interpolation. A multi-mode GNSS receiver is used to receive satellite broadcast ephemeris data, analyze the satellite broadcast ephemeris data, and extract the satellite ephemeris data. The time synchronization between meteorological data and satellite ephemeris data is achieved at the millisecond level through hardware-level PPS pulse signals, with a time synchronization error of less than 1ms.

3. The single-point positioning method based on neural networks and fusing meteorological data and satellite ephemeris as described in claim 1, characterized in that, In step S1, the spatiotemporal joint feature tensor is constructed. for: ； in, N Number of visible satellites; K =10 dimensions; The input spatiotemporal joint feature tensor X is normalized by a BN normalization layer to obtain the spatiotemporal joint feature tensor after BN normalization. .

4. The single-point positioning method based on neural networks and fusing meteorological data and satellite ephemeris as described in claim 1, characterized in that, In step S2, the processing procedure for the meteorological branch includes: S211, 1D convolutional layer sliding extraction T, P, RH and , Association features: ； in, The output of a 1D convolutional layer; This indicates that the kernel size is set to 3; S212, Introducing nonlinearity by using the ReLU activation function: ； in, for The output after the ReLU activation function; S213, Suppress overfitting using Dropout layers: ； in, This refers to the characteristics output by the meteorological branch.

5. The single-point positioning method based on neural networks and fusing meteorological data and satellite ephemeris as described in claim 1, characterized in that, In step S2, the processing of the ephemeris branch includes: S221, Acquire ephemeris data, including pseudorange observations. Satellite coordinates Clock difference , forming the ephemeris branch vector The ephemeris branch vector is processed by a BN normalization layer. Standardization is performed to obtain the ephemeris branch vector after BN normalization layer processing. ; S222, satellite space configuration encoded via a fully connected layer FC: ； in, This is the weight matrix. For bias, for The output after passing through the fully connected layer; S223, Introducing nonlinearity by using the ReLU activation function: ； in, for The output after the ReLU activation function; S224, Suppress overfitting using Dropout layers: ； in, The features output by the ephemeris branch.

6. The single-point positioning method based on neural networks and fusing meteorological data and satellite ephemeris as described in claim 1, characterized in that, The processing steps of the attention fusion layer include: S231, the characteristics output by the meteorological branch Features of Ephemeris Branch Output After dimensional adaptation, the data is stitched together to obtain the stitched features. The splicing features are processed by a BN normalization layer. Standardization is performed to obtain the spliced ​​features after processing by the BN normalization layer. : S232, attention weights are calculated through a fully connected layer and a Sigmoid activation. : ； in, This is the weight matrix. For bias, For the Sigmoid function; S233, through attention weights Weighted fusion characteristics: ； in, The fused features are output by the attention fusion layer.

7. The single-point positioning method based on neural networks and fusing meteorological data and satellite ephemeris as described in claim 1, characterized in that, The processing procedure of the output layer includes: The fused features are directly mapped from the attention fusion layer through a fully connected layer. Location results: ； in, For output layer weights, For bias, For the user's three-dimensional coordinates, For receiver clock bias, The fused features are output by the attention fusion layer.

8. The single-point positioning method based on neural networks and fusing meteorological data and satellite ephemeris as described in claim 1, characterized in that, During neural network training, the Adam gradient descent optimization algorithm is used to iteratively update the network parameters until the total loss function is reached. Convergence occurs until the maximum number of iterations is reached.

9. A computer program product, characterized in that, It includes a computer program / instruction that, when executed by a processor, implements the single-point positioning method based on a neural network and fusing meteorological data and satellite ephemeris as described in any one of claims 1 to 8.

10. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the single-point positioning method based on neural networks and fusing meteorological data and satellite ephemeris as described in any one of claims 1 to 8.

Images