Low-cost GNSS high-precision navigation method in complex urban environment

Abstract

This invention relates to the field of satellite navigation and positioning and multi-source information fusion technology, specifically a low-cost, high-precision GNSS navigation method for complex urban environments. The method constructs a dual-channel adaptive weighting model based on scene classification results. Pseudorange observations are weighted using a nonlinear smooth reduction based on the Sigmoid function, while carrier phase observations are weighted using a two-way reward / penalty system based on signal-to-noise ratio and environment type. This invention achieves a smooth transition of pseudorange weights through the Sigmoid nonlinear function, avoiding filter divergence caused by hard cutoff. A unique open-field carrier reward mechanism improves positioning accuracy in ideal environments. Furthermore, without increasing hardware costs, only algorithm upgrades can significantly improve the navigation performance of low-cost chips in complex urban environments, demonstrating high engineering application value.

Description

Technical Field

[0001] This invention relates to the field of satellite navigation and positioning and multi-source information fusion technology, specifically a low-cost, high-precision GNSS navigation method for complex urban environments. Background Technology

[0002] Global Navigation Satellite System (GNSS) has become an indispensable spatiotemporal reference sensor in modern intelligent transportation systems, autonomous vehicles, and vehicle-to-everything (V2X) technologies. With advancements in semiconductor and radio frequency technologies, low-cost, multi-frequency, multi-system GNSS chips, such as the u-blox F9P and Sinan K8 series, have become increasingly common. However, compared to expensive measurement-grade receivers, low-cost chips, constrained by cost control, typically employ linearly polarized or low-gain patch antennas, resulting in poor internal clock crystal stability and weaker anti-interference and multipath suppression capabilities of the baseband processor.

[0003] In open, unobstructed environments, low-cost chips can achieve centimeter-level positioning using real-time dynamic carrier phase differential (RTK) technology. However, vehicles spend most of their time navigating complex urban environments, such as tree-lined avenues, high-rise canyons, under overpasses, and tunnel entrances. In these scenarios, GNSS signal propagation faces two core challenges: First, multipath propagation. Satellite signals enter the receiver antenna after being reflected from buildings, ground, water surfaces, or vehicle surfaces. Reflected signals have path delays and phase flips compared to direct signals, and these two factors superimpose at the antenna end, causing systematic deviations in observations (especially pseudorange observations) at the meter or even ten-meter level. Although the multipath error of carrier phase is relatively small (usually at the centimeter level), its fluctuations can severely disrupt the consistency of integer ambiguity. Second, there is non-line-of-sight (NLOS) propagation. In deep urban canyons, the direct path of the satellite is completely blocked by buildings, and the receiver only locks onto signals reflected from glass curtain walls or metal surfaces. In this case, the observation distance includes not only geometric distance but also a significant path extension. NLOS signals typically have a high signal-to-noise ratio (SNR), which can easily deceive the receiver's tracking loop, resulting in uncontrollable drift of positioning results by tens or even hundreds of meters.

[0004] Existing positioning strategies face significant technical bottlenecks in addressing the aforementioned challenges: First, the methods for constructing stochastic models are too crude. Traditional methods generally use elevation angle weighted models, whose mathematical expression is usually as follows: This model is based on the assumptions of "atmospheric isotropy" and "multipath propagation mainly originates from ground reflection," assuming that the lower the elevation angle, the greater the error. However, in urban environments, this assumption often fails—satellites at high elevation angles are often blocked by tall buildings on both sides (resulting in NLOS), while low-elevation-angle satellites along the street direction may actually be direct signals (LOS). If the traditional elevation angle weighting method continues to be used, the filter will assign too high a weight to the poor-quality high-elevation-angle NLOS signals, directly leading to positioning divergence.

[0005] Secondly, robust mechanisms lack environmental adaptability. Existing robust Kalman filters mainly rely on post-hoc residual detection or threshold rejection based on signal-to-noise ratio (e.g., (i.e., elimination). This "hard cut-off" strategy can easily lead to a sharp reduction in the number of available satellites to below 4 in complex environments, causing a drastic deterioration in the geometric precision factor (DOP) and interruption of positioning continuity. In addition, for low-cost antennas, the SNR of normal signals is inherently low, and simple threshold cutting can easily result in the accidental deletion of available signals.

[0006] Furthermore, machine learning-assisted localization methods lack physical depth. Although recent studies have attempted to use neural networks to identify NLOS (Normally Unknown Oscillators), two drawbacks exist: first, most directly input raw observation data into the network without physical model-based feature separation, leading to difficulty in network convergence; second, existing training set construction does not consider the orbital characteristics of specific constellations, such as geostationary orbit satellites in the BeiDou Navigation Satellite System, which are stationary relative to the ground, resulting in static multipath features. Including such data in the training set would interfere with the model's ability to learn and generalize to dynamic onboard environments.

[0007] In summary, there is an urgent need for a high-precision navigation method that can deeply integrate physical features and data-driven models, accurately perceive the microscopic observation environment, and achieve smooth, continuous adaptive weighting, so as to enhance the application potential of low-cost GNSS chips in complex urban scenarios. Summary of the Invention

[0008] The purpose of this invention is to provide a low-cost, high-precision GNSS navigation method for complex urban environments to address the problems mentioned in the background. This invention proposes a closed-loop processing architecture of "physical feature extraction—deep environmental perception—nonlinear adaptive weighting." This method first utilizes a specialized software kernel to extract clean multipath and residual features, and strategically removes geostationary orbit satellite data from the BeiDou Navigation Satellite System during the data cleaning stage. Secondly, it constructs a TodyNet neural network to capture the temporal fluctuation features of the signal. Finally, based on the classification results, it constructs a refined stochastic model based on the Sigmoid function and a reward / penalty mechanism, significantly improving positioning accuracy while ensuring filter stability.

[0009] To address the aforementioned technical problems, this invention provides the following technical solution: a low-cost, high-precision GNSS navigation method for complex urban environments, comprising: S1. Collect and analyze the raw observation data stream of the vehicle-mounted low-cost GNSS chip to obtain pseudorange, carrier phase and signal-to-noise ratio data of multiple constellations and multiple frequency points; S2. Based on the multi-path extraction algorithm and the least squares algorithm, extract the multi-path error and pseudo-range residual, and perform strategic data removal for specific constellation orbits during the feature extraction process; S3. Construct a multi-dimensional physical feature vector containing multipath error, pseudorange residual, signal-to-noise ratio and number of visible satellites, and construct a temporal feature tensor containing historical information based on a sliding time window; S4. Input the vehicle driving trajectory after manually dividing the scene and the obtained temporal feature tensor into the pre-trained TodyNet deep neural network model, and output the environmental scene classification result of the current epoch. S5. Based on the scene classification results, a dual-channel adaptive weighting model is constructed. The pseudorange observations are weighted using a nonlinear smooth reduction based on the Sigmoid function, and the carrier phase observations are weighted using a two-way reward and punishment system based on the signal-to-noise ratio and the environment type. S6. Based on the corrected observation weights obtained in step S5, reconstruct the observation noise covariance matrix, substitute it into the extended Kalman filter equation, output high-precision position information, and return to step S1 to iterate and execute the processing of the next epoch.

[0010] Furthermore, the raw observation data stream is acquired using a vehicle-mounted low-cost GNSS receiver through an active vehicle-mounted antenna; The specific constellation orbit refers to the geostationary orbit satellites of the BeiDou Navigation Satellite System.

[0011] Furthermore, a specific implementation of S2 includes: A dual-frequency linear combination algorithm is used to separate pure multipath errors from the original observations; the algorithm formulas are as follows: in, This represents the pseudorange observation value corresponding to the process of the signal propagating from the corresponding observation satellite to the receiver numbered i; The carrier phase observation value is the value observed during the process of the signal propagating from the corresponding observation satellite to the receiver numbered i. This represents the geometric distance the signal travels from the corresponding observation satellite to the corresponding receiver; The speed of light in a vacuum; Let r be the receiver clock bias; The clock bias of the observation satellite numbered s; is the ionospheric delay corresponding to receiver number i; T is the tropospheric delay between the corresponding receiver and the corresponding observation satellite; and These are the pseudorange multipath error and carrier phase multipath error at frequency point i, respectively. The carrier wavelength is the wavelength corresponding to the signal as it propagates from the corresponding observation satellite to the receiver numbered i. Let i be the integer ambiguity corresponding to receiver number i; and The measurement noise for pseudorange and carrier phase measurements at frequency i is due to the multipath error of the carrier phase. Multipath error much smaller than pseudorange Therefore, it can be ignored in the calculation. To eliminate geometric distance... Receiver clock bias Satellite clock bias and first-order ionospheric delay Construct the following two-frequency linear combination: in, For frequency point i, the multipath combination observation value; For the observed coefficients, the , and Let i and j be the carrier frequencies, respectively. From the above, the multipath estimation expression can be derived as follows: in, It is a composite bias term consisting of integer ambiguity and hardware delay; This is the residual noise resulting from the combination of pseudorange and phase noise without an ionosphere. Because within a continuous, cycle-slip-free observation arc... Keep it constant, while It exhibits an oscillating characteristic. Therefore, by analyzing... By applying moving average filtering or piecewise mean subtraction, the multipath effect oscillation component can be separated. .

[0012] By combining the purified multipath error separated from the original observations, the least squares algorithm is used to pre-determine the coordinates, and the post-hoc pseudorange residuals are calculated.

[0013] in, The results are the calculation results of the corresponding post-hoc pseudorange residuals; For the corresponding actual observed pseudorange value, To calculate the geometric distance of the signal obtained from the inverse coordinate calculation from the corresponding observation satellite to the corresponding receiver.

[0014] Furthermore, in the process of constructing the multidimensional physical feature vector in S3, for each visible satellite, there is a unique multidimensional physical feature vector corresponding to each epoch, and the multidimensional physical feature vectors of the current epoch and several past historical moments are combined into a time-series input matrix to construct a time-series feature tensor containing historical information with the corresponding duration as the sliding time window length.

[0015] For any visible satellite, the multidimensional physical eigenvector corresponding to epoch t is denoted as... The Let be the matrix consisting of the multipath error, pseudorange residual, signal-to-noise ratio, and number of visible satellites at epoch t; and let be the temporal feature tensor containing historical information with a sliding time window length of T corresponding to the visible satellites at epoch t. The The This refers to the multidimensional physical feature vector of the corresponding visible satellite at time t-T+1; This represents the multidimensional physical feature vector of the corresponding visible satellite at time t-T+2.

[0016] Furthermore, the TodyNet deep neural network model in S4 includes: One-dimensional convolutional layers (1D-CNN) are used to extract high-frequency fluctuation features of the signal within a local time window (such as rapid SNR fluctuations caused by foliage occlusion); the relevant algorithm formulas are as follows: in, The output feature map of the convolutional layer; ReLU() is the modified linear unit activation function; The weight matrix of the convolution kernel; This is the convolution operator; This is the bias term for the convolutional layer.

[0017] A bidirectional long short-term memory (Bi-LSTM) layer is used to extract long-term temporal dependency features of signals and distinguish between transient signal jumps and persistent building occlusions. The LSTM unit controls the information flow through input gates, forget gates, and output gates. in, The cell state of the LSTM unit at time t is used for long-term memory; This represents the hidden state of the LSTM unit at time t, used for short-time output. The Forgotten Gate controls how much historical information is retained; This is the input gate, which controls how much new information is written. As an output gate, it controls how much information is output; This is the Hadamard product, which is the product of corresponding matrix elements. It is the hyperbolic tangent activation function.

[0018] The classification output module, constructed from a fully connected layer and a Softmax classification layer, outputs the probability distributions for three types of environmental scenes: open areas, general occlusion, and severe occlusion. The relevant algorithm formulas are as follows: in, A probability distribution vector for classifying environmental scenes; and These are the weights and biases of the fully connected layer, respectively. This is the final state vector output by the bidirectional long short-term memory layer; This is a normalized exponential function that converts the output into probability values.

[0019] Furthermore, the training process of the TodyNet deep neural network model includes: Based on the vehicle's driving trajectory, three types of scene labels were manually marked: open, moderately obscured, and severely obscured. The labeled data is divided into training set, validation set and test set according to the proportions; The model is trained by adjusting the training parameters, and the model file is saved for subsequent inference.

[0020] Specifically, based on the features of the GNSS observation data extracted in step S3 and the vehicle's trajectory in the collected GNSS observation data, the vehicle's trajectory is manually divided into three regions according to the vehicle's travel time: open, general, and severely obstructed environments. At this point, the data corresponding to each epoch includes: the number of visible satellites, signal-to-noise ratio, multipath value, pseudorange residual, and the vehicle's current environment. Then, this data is trained using the TodyNet deep neural network model, with the data for the three scenarios divided into training, validation, and test sets at proportions of 60%, 20%, and 20%, respectively. For the TodyNet deep neural network model, the configuration of model training parameters can be adjusted, including EPOCHS, BATCH_SIZE, and LEARNING_RATE. EPOCHS refers to the process of completely traversing the entire training dataset once, meaning all data is learned by the model in one round. BATCH_SIZE refers to the number of data points given to the model for training in each round, corresponding to each sample in step S3 containing T seconds of historical data across four feature dimensions. LEARNING_RATE refers to the model's learning rate, which controls the update magnitude of model parameters during training. By appropriately adjusting these parameters, the learning efficiency and classification accuracy of the TodyNet deep neural network model can be improved.

[0021] After the model finishes running, the trained model will be automatically saved, and the test set results and a confusion matrix will be output. The output results include the classification accuracy on the test set; the confusion matrix shows the classification details, including how many environments were correctly classified and which class was assigned to the incorrectly classified environments.

[0022] The model file trained and saved using the TodyNet deep neural network model can be directly used in step S3 to avoid manually dividing the vehicle's trajectory in step S4 if new GNSS vehicle observation data is subsequently collected. The system directly outputs the scene where the vehicle's trajectory is located based on newly acquired GNSS observation data using a pre-trained model file.

[0023] Furthermore, in S5, a dual-channel adaptive weighting model is constructed to obtain the corrected pseudorange observation weights and corrected carrier phase observation weights for pseudorange observations and carrier phase observations, respectively. The pseudorange observation weights are obtained by nonlinear smoothing weight reduction based on the Sigmoid function. Traditional methods typically employ a "step-wise" weight reduction approach, which can easily lead to abrupt weight changes and filter oscillations. This invention utilizes the smooth saturation characteristics of the sigmoid function to construct a continuous decay model; and sets a multipath tolerance threshold for the current scenario. and steepness Based on the current satellite multipath error values The normalized bias index x is calculated, and then the Sigmoid scaling factor is calculated to obtain the corrected pseudorange observation weights. The calculation formulas involved are as follows: in, is the minimum retained weight factor; e is the base of the natural logarithm. When the multipath error of the observations exceeds the tolerance threshold of the current scene, this formula drives the pseudorange weights to decay smoothly according to the Sigmoid curve, avoiding a hard cut-off of the corrected pseudorange weights.

[0024] The carrier phase observation weights are obtained by a two-way weighting method based on signal-to-noise ratio and environment type to obtain the corrected carrier phase observation weights; the calculation formulas involved are as follows: in, The corrected carrier phase observation weights; Weighting of the basic elevation angle; The signal-to-noise ratio of the current signal; This is a poor-quality threshold. A high-quality threshold; This refers to the current scene being labeled as open. A reward multiplier is applied to open areas. This strategy, while suppressing poor-quality signals, utilizes a reward mechanism to significantly improve positioning accuracy in ideal environments.

[0025] Furthermore, in step S6, after updating the high-precision position information using the extended Kalman filter equation, it also includes iterative steps for standardized residual detection and outlier weight readjustment until all residuals meet the threshold or the maximum number of iterations is reached.

[0026] The diagonal elements of the reconstructed observation noise covariance matrix consist of the pseudorange variance and carrier variance obtained from the calculation results in step S5. Substituting the reconstructed observation noise covariance matrix into the extended Kalman filter equation, the Kalman gain is calculated and the state vector (consisting of position, velocity, clock error, and ambiguity) is updated: in, The Kalman gain matrix determines the weight of the observations in the state update. For the reconstructed observation noise covariance matrix; The prediction error covariance matrix; To design a matrix that describes the geometric relationship between observations and state variables; for The transpose of the matrix; This is the estimated state vector value at time k, which is the final positioning result; The predicted state value at time k; This is the observation vector, i.e., the actual measured value; The innovation vector reflects the difference between the actual observations at the current epoch and the theoretical observations based on the state predictions.

[0027] After completing one filter update, the standardized residuals of each observation are calculated. If there are outliers whose standardized residuals exceed a preset threshold, the weight of that observation is further reduced, and the process returns to the Kalman filter update step to recalculate until all residuals meet the test conditions or the maximum number of iterations is reached, finally outputting a high-precision positioning result.

[0028] Compared with the prior art, the beneficial effects achieved by the present invention are: (1) This invention accurately identifies various occlusion scenarios in urban canyons by introducing the TodyNet deep neural network model; and eliminates the interference of static multipath on dynamic model training by removing geostationary satellites in the data preprocessing stage. (2) This invention achieves a smooth transition of pseudorange weights through the Sigmoid nonlinear function, avoiding filter divergence caused by hard cut-off; it improves positioning accuracy in ideal environments through the original open-ground carrier reward mechanism; at the same time, it can significantly improve the navigation performance of low-cost chips in complex urban environments without increasing hardware costs, and has extremely high engineering application value. Attached Figure Description

[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall technical process of a low-cost, high-precision GNSS navigation method for complex urban environments according to the present invention. Figure 2 This is a schematic diagram of the temporal feature extraction structure of the TodyNet deep neural network model in the low-cost, high-precision GNSS navigation method for complex urban environments of the present invention; Figure 3 This is a schematic diagram of the confusion matrix generated by the TodyNet deep neural network model in a low-cost, high-precision GNSS navigation method for complex urban environments according to the present invention. Figure 4This is a schematic diagram of the solution results of the elevation angle weighting model used in the low-cost, high-precision GNSS navigation method for complex urban environments of the present invention; Figure 5 This is a schematic diagram of the solution results using the model proposed in this invention in a low-cost, high-precision GNSS navigation method for complex urban environments. Detailed Implementation

[0030] 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.

[0031] Please see Figures 1-5 The present invention provides a technical solution: such as Figure 1 As shown, this embodiment provides a low-cost, high-precision GNSS navigation method for complex urban environments, including: S1. Collect and analyze the raw observation data stream of the vehicle-mounted low-cost GNSS chip to obtain pseudorange, carrier phase and signal-to-noise ratio data of multiple constellations and multiple frequency points; the raw observation data stream is acquired by the vehicle-mounted low-cost GNSS receiver through an active vehicle-mounted antenna; S2. Based on the multi-path extraction algorithm and the least squares algorithm, extract the multi-path error and pseudorange residual (multi-source error feature extraction), and perform strategic data removal for specific constellation orbits during the feature extraction process; the specific constellation orbits are geostationary orbit satellites of the BeiDou Navigation Satellite System.

[0032] The specific implementation of S2 includes: A dual-frequency linear combination algorithm is used to separate pure multipath errors from the original observations; the algorithm formulas are as follows: in, This represents the pseudorange observation value corresponding to the process of the signal propagating from the corresponding observation satellite to the receiver numbered i; The carrier phase observation value is the value observed during the process of the signal propagating from the corresponding observation satellite to the receiver numbered i. This represents the geometric distance the signal travels from the corresponding observation satellite to the corresponding receiver; The speed of light in a vacuum; Let r be the receiver clock bias; The clock bias of the observation satellite numbered s; is the ionospheric delay corresponding to receiver number i; T is the tropospheric delay between the corresponding receiver and the corresponding observation satellite; and These are the pseudorange multipath error and carrier phase multipath error at frequency point i, respectively. The carrier wavelength is the wavelength corresponding to the signal as it propagates from the corresponding observation satellite to the receiver numbered i. Let i be the integer ambiguity corresponding to receiver number i; and The measurement noise for pseudorange and carrier phase measurements at frequency i is due to the multipath error of the carrier phase. Multipath error much smaller than pseudorange Therefore, it can be ignored in the calculation. To eliminate geometric distance... Receiver clock bias Satellite clock bias and first-order ionospheric delay Construct the following two-frequency linear combination: in, For frequency point i, the multipath combination observation value; For the observed coefficients, the , and Let i and j be the carrier frequencies, respectively. From the above, the multipath estimation expression can be derived as follows: in, It is a composite bias term consisting of integer ambiguity and hardware delay; This is the residual noise resulting from the combination of pseudorange and phase noise without an ionosphere. Because within a continuous, cycle-slip-free observation arc... Keep it constant, while It exhibits an oscillating characteristic. Therefore, by analyzing... By applying moving average filtering or piecewise mean subtraction, the multipath effect oscillation component can be separated. .

[0033] By combining the purified multipath error separated from the original observations, the least squares algorithm is used to pre-determine the coordinates, and the post-hoc pseudorange residuals are calculated.

[0034] in, The results are the calculation results of the corresponding post-hoc pseudorange residuals; For the corresponding actual observed pseudorange value, To calculate the geometric distance of the signal obtained from the inverse coordinate calculation from the corresponding observation satellite to the corresponding receiver.

[0035] In this embodiment, a data file generated by an algorithm is used; a Python script is written to automatically identify the satellite PRN number and forcibly remove geostationary orbit satellite data numbered C01 to C05 in the Beidou system to prevent their static multipath characteristics from polluting the training set; at the same time, the least squares method is used to perform standard point positioning pre-calculation to obtain the post-verification pseudorange residual.

[0036] S3. Construct a multi-dimensional physical feature vector containing multipath error, pseudorange residual, signal-to-noise ratio and number of visible satellites, and construct a temporal feature tensor containing historical information based on a sliding time window; In the process of constructing the multidimensional physical feature vector in S3, for each visible satellite, there is a unique multidimensional physical feature vector corresponding to each epoch. The multidimensional physical feature vectors of the current epoch and several past historical moments are combined into a time-series input matrix to construct a time-series feature tensor containing historical information with the corresponding duration as the sliding time window length.

[0037] For any visible satellite, the multidimensional physical eigenvector corresponding to epoch t is denoted as... The Let be the matrix consisting of the multipath error, pseudorange residual, signal-to-noise ratio, and number of visible satellites at epoch t; and let be the temporal feature tensor containing historical information with a sliding time window length of T corresponding to the visible satellites at epoch t. The The This refers to the multidimensional physical feature vector of the corresponding visible satellite at time t-T+1; This represents the multidimensional physical feature vector of the corresponding visible satellite at time t-T+2.

[0038] S4. Input the vehicle driving trajectory after manually dividing the scene and the obtained temporal feature tensor into the pre-trained TodyNet deep neural network model, and output the environmental scene classification result of the current epoch. like Figure 2 As shown, the TodyNet deep neural network model in S4 includes: One-dimensional convolutional layers (1D-CNN) are used to extract high-frequency fluctuation features of the signal within a local time window (such as rapid SNR fluctuations caused by foliage occlusion); the relevant algorithm formulas are as follows: in, The output feature map of the convolutional layer; ReLU() is the modified linear unit activation function; The weight matrix of the convolution kernel; This is the convolution operator; This is the bias term for the convolutional layer.

[0039] A bidirectional long short-term memory (Bi-LSTM) layer is used to extract long-term temporal dependency features of signals and distinguish between transient signal jumps and persistent building occlusions. The LSTM unit controls the information flow through input gates, forget gates, and output gates. in, The cell state of the LSTM unit at time t is used for long-term memory; This represents the hidden state of the LSTM unit at time t, used for short-time output. The Forgotten Gate controls how much historical information is retained; This is the input gate, which controls how much new information is written. As an output gate, it controls how much information is output; This is the Hadamard product, which is the product of corresponding matrix elements. It is the hyperbolic tangent activation function.

[0040] The classification output module, constructed from a fully connected layer and a Softmax classification layer, outputs the probability distributions for three types of environmental scenes: open areas, general occlusion, and severe occlusion. The relevant algorithm formulas are as follows: in, A probability distribution vector for classifying environmental scenes; and These are the weights and biases of the fully connected layer, respectively. This is the final state vector output by the bidirectional long short-term memory layer; This is a normalized exponential function that converts the output into probability values.

[0041] The training process of the TodyNet deep neural network model includes: Based on the vehicle's driving trajectory, three types of scene labels were manually marked: open, moderately obscured, and severely obscured. The labeled data is divided into training set, validation set and test set according to the proportions; The model is trained by adjusting the training parameters, and the model file is saved for subsequent inference.

[0042] Specifically, based on the features of the GNSS observation data extracted in step S3 and the vehicle's trajectory in the collected GNSS observation data, the vehicle's trajectory is manually divided into three regions according to the vehicle's travel time: open, general, and severely obstructed environments. At this point, the data corresponding to each epoch includes: the number of visible satellites, signal-to-noise ratio, multipath value, pseudorange residual, and the vehicle's current environment. Then, this data is trained using the TodyNet deep neural network model, with the data for the three scenarios divided into training, validation, and test sets at proportions of 60%, 20%, and 20%, respectively. For the TodyNet deep neural network model, the configuration of model training parameters can be adjusted, including EPOCHS, BATCH_SIZE, and LEARNING_RATE. EPOCHS refers to the process of completely traversing the entire training dataset once, meaning all data is learned by the model in one round. BATCH_SIZE refers to the number of data points given to the model for training in each round, corresponding to each sample in step S3 containing T seconds of historical data across four feature dimensions. LEARNING_RATE refers to the model's learning rate, which controls the update magnitude of model parameters during training. By appropriately adjusting these parameters, the learning efficiency and classification accuracy of the TodyNet deep neural network model can be improved.

[0043] After the model finishes running, the trained model will be automatically saved, and the test set results and a confusion matrix will be output. The output results include the classification accuracy on the test set; the confusion matrix shows the classification details, including how many environments were correctly classified and which class was assigned to the incorrectly classified environments.

[0044] The model file trained and saved using the TodyNet deep neural network model can be directly used in step S3 to avoid manually dividing the vehicle's trajectory in step S4 if new GNSS vehicle observation data is subsequently collected. The system directly outputs the scene where the vehicle's trajectory is located based on newly acquired GNSS observation data using a pre-trained model file.

[0045] In this embodiment, the vehicle driving trajectory of the collected GNSS observation data is manually divided into scenes and the time sequence input matrix obtained by S3 is input into the pre-trained TodyNet deep neural network model. The model uses 1D-CNN layer to capture high-frequency jumps in the signal (such as leaf occlusion), uses Bi-LSTM layer to capture bidirectional long temporal dependencies of the signal (such as building occlusion), and finally outputs the environmental scene classification result (open area, general occlusion, severe occlusion) of the current epoch through Softmax layer. In the specific implementation process: In order to train a TodyNet deep neural network model suitable for complex urban environments, this embodiment constructs a GNSS observation feature dataset containing multiple scenarios, a TodyNet model training configuration, and classification results, specifically including: Data Acquisition: Three typical scenarios—open, partially obstructed, and severely obstructed—were selected, and dynamic data acquisition was conducted using vehicles equipped with low-cost GNSS chips. The acquisition frequency was 1Hz, with a cumulative acquisition time exceeding 3 hours, yielding approximately 11,000 epochs of data. Vehicle trajectory scene segmentation: Combine vehicle trajectory with map post-processing and manually segment into three scenes, including open, moderately occluded, and severely occluded. Feature extraction and tensor construction: Core features for each epoch are extracted, including multipath error, pseudorange residual, signal-to-noise ratio, and number of visible satellites. A sliding time window is constructed with a window size of 10 seconds, allowing the TodyNe model to predict the current state using time-series data from the current epoch and the past 9 epochs. The data for each epoch includes: number of visible satellites, signal-to-noise ratio, multipath value, pseudorange residual, and the vehicle's current environment. TodyNet deep neural network model training configuration; The TodyNet deep neural network model has a total of 471,536 parameters and a total of 9,896 data samples. The data for three scenarios are divided into a 60% training set (5,931 samples), a 20% validation set (1,977 samples), and a 20% test set (1,978 samples). The model parameters EPOCHS are set to 600, BATCH_SIZE to 32, and LEARNING_RATE to 0.0001.

[0046] Classification results: After the model finishes running, the trained model will be automatically saved, and the test set results and a confusion matrix will be output, such as... Figure 3 As shown.

[0047] from Figure 3 The results show that in severely occluded scenarios, out of 822 samples, the model successfully identified 819; in moderately occluded scenarios, out of 991 samples, the model successfully identified 975; and in open area scenarios, out of 165 samples, the model successfully identified 156. In a total of 1978 test samples, the overall recognition accuracy reached 98.58%. This fully verifies the accuracy of the TodyNet model proposed in this invention for environmental perception in complex urban road conditions.

[0048] S5. Based on the scene classification results, a dual-channel adaptive weighting model is constructed. The pseudorange observations are weighted using a nonlinear smooth reduction based on the Sigmoid function, and the carrier phase observations are weighted using a two-way reward and punishment system based on the signal-to-noise ratio and the environment type. In S5, a dual-channel adaptive weighting model is constructed to obtain the corrected pseudorange observation weights and corrected carrier phase observation weights for pseudorange observations and carrier phase observations, respectively. The pseudorange observation weights are obtained by nonlinear smoothing weight reduction based on the Sigmoid function. Traditional methods typically employ a "step-wise" weight reduction approach, which can easily lead to abrupt weight changes and filter oscillations. This invention utilizes the smooth saturation characteristics of the sigmoid function to construct a continuous decay model; and sets a multipath tolerance threshold for the current scenario. and steepness Based on the current satellite multipath error values The normalized bias index x is calculated, and then the Sigmoid scaling factor is calculated to obtain the corrected pseudorange observation weights. The calculation formulas involved are as follows: in, is the minimum retained weight factor; e is the base of the natural logarithm. When the multipath error of the observations exceeds the tolerance threshold of the current scene, this formula drives the pseudorange weights to decay smoothly according to the Sigmoid curve, avoiding a hard cut-off of the corrected pseudorange weights.

[0049] The carrier phase observation weights are obtained by a two-way weighting method based on signal-to-noise ratio and environment type to obtain the corrected carrier phase observation weights; the calculation formulas involved are as follows: in, The corrected carrier phase observation weights; Weighting of the basic elevation angle; The signal-to-noise ratio of the current signal; This is a poor-quality threshold. A high-quality threshold; This refers to the current scene being labeled as open. A reward multiplier is applied to open areas. This strategy, while suppressing poor-quality signals, utilizes a reward mechanism to significantly improve positioning accuracy in ideal environments.

[0050] In this embodiment, under severe occlusion conditions, the multipath tolerance threshold is set to 0.25 meters and the steepness is 3.0. For pseudorange observations exceeding the threshold, the weights are smoothly decayed to 0.01 using the Sigmoid function to maintain the observation geometry. In general obstruction environments, a signal-to-noise ratio (SNR) inferiority threshold of 28 dB is set, and carrier phase observations below this value are forcibly downweighted to prevent cycle slips or half-cycle slips from contaminating the calculation. In open environments, a reward mechanism is activated, with a high-quality threshold of 40dB and a reward multiplier of 10.0. High-quality carrier signals are given extremely high weight, locking the positioning trajectory near the true value.

[0051] S6. Based on the corrected observation weights obtained in step S5, reconstruct the observation noise covariance matrix, substitute it into the extended Kalman filter equation, output high-precision position information, and return to step S1 to iterate and execute the processing of the next epoch.

[0052] In step S6, after updating the high-precision position information using the extended Kalman filter equation, it also includes iterative steps for standardized residual detection and outlier weight readjustment until all residuals meet the threshold or the maximum number of iterations is reached.

[0053] The diagonal elements of the reconstructed observation noise covariance matrix consist of the pseudorange variance and carrier variance obtained from the calculation results in step S5. Substituting the reconstructed observation noise covariance matrix into the extended Kalman filter equation, the Kalman gain is calculated and the state vector (consisting of position, velocity, clock error, and ambiguity) is updated: in, The Kalman gain matrix determines the weight of the observations in the state update. For the reconstructed observation noise covariance matrix; The prediction error covariance matrix; To design a matrix that describes the geometric relationship between observations and state variables; for The transpose of the matrix; This is the estimated state vector value at time k, which is the final positioning result; The predicted state value at time k; This is the observation vector, i.e., the actual measured value; The innovation vector reflects the difference between the actual observations at the current epoch and the theoretical observations based on the state predictions.

[0054] After completing one filter update, the standardized residuals of each observation are calculated. If there are outliers whose standardized residuals exceed a preset threshold, the weight of that observation is further reduced, and the process returns to the Kalman filter update step to recalculate until all residuals meet the test conditions or the maximum number of iterations is reached, finally outputting a high-precision positioning result.

[0055] In this embodiment, to verify the effectiveness of the scene classification and adaptive weighting method proposed in this invention, GNSS vehicle-mounted data of a typical complex urban road section (including open, general, and severely obstructed environments) was collected using a low-cost receiver for field verification. The method of this invention is compared with the traditional elevation angle-based weighting model.

[0056] The software platforms used in the experiment included: (1) Reference software: The open-source high-precision data processing software RTKLIB (version 2.4.3 b34) was adopted. This software was developed by Tokyo University of Marine Science and Technology in Japan. It supports centimeter-level high-precision positioning and is suitable for real-time dynamic positioning and post-processing tasks. (2) Testing software: GREAT (GNSS+ REsearch, Application and Teaching) software was used. This software was designed and developed by the School of Geodesy and Geomatics of Wuhan University. It is a comprehensive software platform for spatial geodetic data processing, precise positioning and orbit determination, and multi-source fusion navigation. (3) Reference value acquisition: Using the post-processing kinematics mode of RTKLIB software and combined with the observation data of nearby base stations, the raw GNSS vehicle data was processed with high precision to obtain a position coordinate sequence with centimeter-level accuracy, which was used as the reference value for this experiment.

[0057] In terms of scheme design, this experiment set up two solution schemes under the GREAT software platform. Except for the stochastic model (weighting strategy), the other configurations (cutoff elevation angle, tropospheric model, ionospheric model, Kalman filter parameters, etc.) were kept consistent.

[0058] Option 1 (Control Group): The default elevation angle weighting model in GREAT software was used. This model only sets the weights of observations based on the satellite elevation angle and does not consider the specific impact of environmental obstruction on signal quality. Option 2 (Technical Solution of this Invention): Using GREAT software equipped with the algorithm of this invention, an adaptive stochastic model based on scene recognition is employed. This option first identifies the environmental scene using TodyNet, and then adaptively adjusts the variance of the observations using the Sigmoid function and a signal-to-noise ratio reward / penalty mechanism.

[0059] Regarding the experimental results and analysis, the solution results of Scheme 1 and Scheme 2 were compared epoch-by-epoch with the reference values ​​generated by RTKLIB. The ENU residual time series in the three directions of east (E), north (N), and sky (U) were calculated, and the standard deviations in the three directions were statistically analyzed. ,like Figure 4 and Figure 5 As shown.

[0060] When using the default elevation angle weighting model (Scheme 1) in GREAT software, the standard deviations of the east, north, and sky directions are... The elevation angles are 1.7828m, 2.0747m, and 4.8670m, respectively. This is because in complex urban environments, satellite signals at high elevation angles may still propagate non-line-of-sight through reflections from buildings, resulting in large errors in pseudorange observations, which the elevation angle model incorrectly assigns high weights to.

[0061] After using the adaptive stochastic model (Scheme 2) of the basic scenario proposed in this invention, the standard deviations in the three directions are... All three angles decreased, with the angles in the east, north, and sky directions decreasing to 1.4831m, 1.5675m, and 3.7707m respectively, representing improvements in accuracy of 16.81%, 24.45%, and 22.53% respectively.

[0062] The results demonstrate that the TodyNet scene recognition and adaptive weighting strategy proposed in this invention improves the accuracy of GREAT software's positioning calculations using the traditional elevation angle weighting model in complex environments, making the results closer to the high-precision reference values ​​of RTKLIB. Compared to traditional methods, this invention achieves a 21.26% improvement in overall positioning accuracy through algorithm optimization alone, without changing the hardware configuration, verifying the technical advantages of this method in suppressing multipath effects and improving the reliability of low-cost terminal positioning.

[0063] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0064] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A low-cost, high-precision GNSS navigation method for complex urban environments, characterized in that: include: S1. Collect and analyze the raw observation data stream of the vehicle-mounted low-cost GNSS chip to obtain pseudorange, carrier phase and signal-to-noise ratio data of multiple constellations and multiple frequency points; S2. Based on the multi-path extraction algorithm and the least squares algorithm, extract the multi-path error and pseudo-range residual, and perform strategic data removal for specific constellation orbits during the feature extraction process; S3. Construct a multi-dimensional physical feature vector containing multipath error, pseudorange residual, signal-to-noise ratio and number of visible satellites, and construct a temporal feature tensor containing historical information based on a sliding time window; S4. Input the vehicle driving trajectory after manually dividing the scene and the obtained temporal feature tensor into the pre-trained TodyNet deep neural network model, and output the environmental scene classification result of the current epoch. S5. Based on the scene classification results, a dual-channel adaptive weighting model is constructed. The pseudorange observations are weighted using a nonlinear smooth reduction based on the Sigmoid function, and the carrier phase observations are weighted using a two-way reward and punishment system based on the signal-to-noise ratio and the environment type. S6. Based on the corrected observation weights obtained in step S5, reconstruct the observation noise covariance matrix, substitute it into the extended Kalman filter equation, output high-precision position information, and return to step S1 to iterate and execute the processing of the next epoch.

2. The low-cost, high-precision GNSS navigation method for complex urban environments according to claim 1, characterized in that: The raw observation data stream was acquired using a low-cost vehicle-mounted GNSS receiver via an active vehicle-mounted antenna; The specific constellation orbit refers to the geostationary orbit satellites of the BeiDou Navigation Satellite System.

3. The low-cost, high-precision GNSS navigation method for complex urban environments according to claim 1, characterized in that: The specific implementation of S2 includes: A dual-frequency linear combination algorithm is used to separate pure multipath errors from the original observations; By combining the purified multipath error separated from the original observations, the least squares algorithm is used to pre-determine the coordinates, and the post-hoc pseudorange residuals are calculated.

4. The low-cost, high-precision GNSS navigation method for complex urban environments according to claim 1, characterized in that: In the process of constructing the multidimensional physical feature vector in S3, for each visible satellite, there is a unique multidimensional physical feature vector corresponding to each epoch. The multidimensional physical feature vectors of the current epoch and several past historical moments are combined into a time-series input matrix to construct a time-series feature tensor containing historical information with the corresponding duration as the sliding time window length. For any visible satellite, the multidimensional physical eigenvector corresponding to epoch t is denoted as V. t The V t Let X be a matrix consisting of the multipath error, pseudorange residual, signal-to-noise ratio, and number of visible satellites at epoch t; and let X be the temporal feature tensor containing historical information with a sliding time window length of T corresponding to the visible satellites at epoch t. input The The This refers to the multidimensional physical feature vector of the corresponding visible satellite at time t-T+1; This represents the multidimensional physical feature vector of the corresponding visible satellite at time t-T+2.

5. The low-cost, high-precision GNSS navigation method for complex urban environments according to claim 1, characterized in that: The TodyNet deep neural network model in S4 includes: One-dimensional convolutional layers are used to extract high-frequency fluctuation features of signals within a local time window; A bidirectional long short-term memory layer is used to extract long-term temporal dependency features of the signal; The classification output module, constructed from a fully connected layer and a Softmax classification layer, is used to output the probability distribution of three types of environmental scenes: open area, general occlusion, and severe occlusion.

6. The low-cost, high-precision GNSS navigation method for complex urban environments according to claim 1, characterized in that: The training process of the TodyNet deep neural network model includes: Based on the vehicle's driving trajectory, three types of scene labels were manually marked: open, moderately obscured, and severely obscured. The labeled data is divided into training set, validation set and test set according to the proportions; The model is trained by adjusting the training parameters, and the model file is saved for subsequent inference.

7. The low-cost, high-precision GNSS navigation method for complex urban environments according to claim 1, characterized in that: In S5, a dual-channel adaptive weighting model is constructed to obtain the corrected pseudorange observation weights and corrected carrier phase observation weights for pseudorange observations and carrier phase observations, respectively. The pseudorange observation weights are obtained by nonlinear smoothing weight reduction based on the Sigmoid function. The carrier phase observation weights are obtained by a two-way weighting method based on signal-to-noise ratio and environment type.

8. The low-cost, high-precision GNSS navigation method for complex urban environments according to claim 1, characterized in that: In step S6, after updating the high-precision position information using the extended Kalman filter equation, it also includes iterative steps for standardized residual detection and outlier weight readjustment until all residuals meet the threshold or the maximum number of iterations is reached.

Images