Satellite geometric configuration change monitoring method and device, electronic equipment and readable medium

Abstract

The invention relates to a satellite geometric configuration change monitoring method and device, electronic equipment and a readable medium. The method comprises the following steps: inputting target observation data collected by a GNSS receiver into a pre-trained bidirectional LSTM network model; the bidirectional LSTM network model is obtained by carrying out joint training by adopting an environment classification loss function and a geometric quality regression loss function; calculating key features of the geometric configuration change of the satellite based on the target observation data through a bidirectional LSTM network model; capturing a front-and-back dependency relationship of a time sequence formed by the key features through a bidirectional LSTM network model so as to monitor the geometric configuration change of the satellite; the environment where the GNSS receiver is located and the deterioration degree of the satellite geometric configuration are predicted based on the change of the satellite geometric configuration through the bidirectional LSTM network model, so that a data processing mode matched with the GNSS receiver is adopted for positioning. According to the scheme provided by the invention, the problem of deficiencies in processing the change of the geometric configuration of the satellite in the prior art can be effectively solved.

Description

Technical Field

[0001] The present application relates to the field of satellite navigation technology, and in particular to a method, device, electronic device and readable medium for monitoring changes in satellite geometric configuration. Background Art

[0002] Satellite geometry changes refer to the changes in the satellite's geometric parameters, such as its position and attitude in space, over time. These changes can affect the positioning accuracy and reliability of the GNSS (Global Navigation Satellite System).

[0003] Related technologies primarily analyze changes in satellite geometry using geometric models and statistical methods. For example, DOP (Dilution of Precision) is used to assess the impact of satellite geometry on positioning accuracy. However, these methods are limited in handling complex dynamic environments and multipath effects. They cannot accurately capture subtle changes in satellite geometry and their temporal dependencies. Their ability to monitor and predict changes in satellite geometry in real time is limited, making them unable to meet the requirements of high-precision positioning. Summary of the Invention

[0004] In order to solve or partially solve the problems existing in the related art, the present application provides a satellite geometric configuration change monitoring method, device, electronic device and readable medium, which can effectively solve the deficiencies in the related art when dealing with satellite geometric configuration changes.

[0005] A first aspect of the present application provides a method for monitoring changes in satellite geometric configuration, the method comprising: Obtaining target observation data collected by a GNSS receiver and inputting the target observation data into a pre-trained bidirectional LSTM network model; wherein the bidirectional LSTM network model is jointly trained using a multi-task loss function, wherein the multi-task loss function includes an environment classification loss function and a geometric quality regression loss function; Calculating key features of satellite geometric configuration changes based on the target observation data through the bidirectional LSTM network model; The bidirectional LSTM network model is used to capture the front-back dependency of the time series composed of the key features to monitor the changes in the satellite geometric configuration; Predicting the environment of the GNSS receiver and the degradation degree of the satellite geometry based on the change of the satellite geometry by using the bidirectional LSTM network model; Positioning is performed using a data processing method that matches the environment in which the GNSS receiver is located and the degree of degradation of the satellite geometry.

[0006] A second aspect of the present application provides a device for monitoring changes in satellite geometric configuration, the device comprising: A target observation data acquisition module is used to obtain target observation data collected by a GNSS receiver and input the target observation data into a pre-trained bidirectional LSTM network model; wherein the bidirectional LSTM network model is jointly trained using a multi-task loss function, wherein the multi-task loss function includes an environmental classification loss function and a geometric quality regression loss function; A key feature calculation module, configured to calculate key features of satellite geometric configuration changes based on the target observation data using the bidirectional LSTM network model; A satellite geometry change monitoring module is used to capture the front-end dependency of the time series composed of the key features through the bidirectional LSTM network model to monitor the satellite geometry change; a multi-task prediction module, configured to predict the environment of the GNSS receiver and the degradation degree of the satellite geometry based on the change of the satellite geometry using the bidirectional LSTM network model; The data processing module is used to perform positioning using a data processing method that matches the environment and the degradation degree.

[0007] A third aspect of the present application provides an electronic device, including: processor; and The memory stores executable codes thereon, and when the executable codes are executed by the processor, the processor is caused to execute the method described above.

[0008] A fourth aspect of the present application provides a computer-readable storage medium having executable code stored thereon. When the executable code is executed by a processor of an electronic device, the processor is caused to execute the method described above.

[0009] The technical solution provided by this application may include the following beneficial results: The solution provided in this application obtains target observation data collected by a GNSS receiver and inputs the target observation data into a pre-trained bidirectional LSTM network model; wherein the bidirectional LSTM network model is obtained by joint training using a multi-task loss function, and the multi-task loss function includes an environmental classification loss function and a geometric quality regression loss function; the bidirectional LSTM network model is used to calculate the key features of the satellite's geometric configuration changes based on the target observation data; the bidirectional LSTM network model is used to capture the front-to-back dependencies of the time series composed of the key features to monitor the satellite's geometric configuration changes; the bidirectional LSTM network model is used to predict the environment of the GNSS receiver and the degradation degree of the satellite's geometric configuration based on the changes in the satellite's geometric configuration; and positioning is performed using a data processing method that matches the environment of the GNSS receiver and the degradation degree of the satellite's geometric configuration. This application uses a bidirectional LSTM network model to capture the forward and backward dependencies of the time series composed of key features of satellite geometric configuration changes, which can improve the accuracy and real-time performance of satellite geometric configuration change monitoring. Moreover, by jointly training the bidirectional LSTM network model with multiple tasks, the comprehensive performance and generalization ability of the bidirectional LSTM network model can be improved, so that the bidirectional LSTM network model can achieve good prediction results in different scenarios. Therefore, based on the monitored satellite geometric configuration changes, the bidirectional LSTM network model can accurately and quickly predict the environment in which the GNSS receiver is located and the degradation degree of the satellite geometric configuration, thereby breaking the limitations of related technologies in dealing with complex dynamic environments and multipath effects, and enhancing the positioning accuracy and reliability of the GNSS system in complex environments.

[0010] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The above and other objects, features and advantages of the present application will become more apparent by describing in more detail exemplary embodiments of the present application in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the present application.

[0012] Figure 1 1 is a flow chart of a method for monitoring changes in satellite geometric configuration according to an embodiment of the present application; Figure 2 1 is another flow chart of a method for monitoring changes in satellite geometric configurations according to an embodiment of the present application; Figure 3 This is a training flow chart of a bidirectional LSTM network model shown in an embodiment of the present application; Figure 4 1 is a schematic structural diagram of a satellite geometric configuration change monitoring device shown in an embodiment of the present application; Figure 5 It is a structural diagram of an electronic device shown in an embodiment of the present application. DETAILED DESCRIPTION

[0013] The following describes embodiments of the present application in more detail with reference to the accompanying drawings. Although the accompanying drawings illustrate embodiments of the present application, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to make the present application more thorough and complete, and to fully convey the scope of the present application to those skilled in the art.

[0014] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. As used in this application and the appended claims, the singular forms "a," "an," "the," and "the" are intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0015] It should be understood that although the terms "first", "second", "third", etc. may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0016] Related technologies mainly use geometric models and statistical methods to analyze changes in satellite geometry, such as using the DOP indicator to evaluate the impact of satellite geometry distribution on positioning accuracy. However, these methods have the following problems: 1) Related technologies have difficulty handling complex dynamic environments and multipath effects, resulting in reduced positioning accuracy; 2) Related technologies are insufficient in capturing subtle changes in satellite geometry and time dependencies; 3) The relevant technologies have limited real-time monitoring and prediction capabilities for changes in satellite geometric configurations and cannot meet the needs of high-precision positioning.

[0017] In response to the above problems, an embodiment of the present application provides a method for monitoring changes in satellite geometric configuration. By capturing the front-to-back dependencies of the time series composed of key features of satellite geometric configuration changes through a bidirectional LSTM network model, the accuracy and real-time performance of monitoring changes in satellite geometric configuration can be improved. Moreover, by jointly training the bidirectional LSTM network model with multiple tasks, the comprehensive performance and generalization ability of the bidirectional LSTM network model can be improved, so that the bidirectional LSTM network model can achieve good prediction results in different scenarios. Therefore, the bidirectional LSTM network model can accurately and quickly predict the environment of the GNSS receiver and the degradation degree of the satellite geometric configuration based on the monitored changes in the satellite geometric configuration, thereby breaking the limitations of related technologies in dealing with complex dynamic environments and multipath effects, and enhancing the positioning accuracy and reliability of the GNSS system in complex environments.

[0018] The technical solutions of the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0019] Figure 1 It is a flow chart of a method for monitoring changes in satellite geometric configuration shown in an embodiment of the present application.

[0020] See also Figure 1 The satellite geometric configuration change monitoring method of the present application includes: S110, obtaining target observation data collected by the GNSS receiver, and inputting the target observation data into a pre-trained bidirectional LSTM network model; wherein the bidirectional LSTM network model is obtained by joint training using a multi-task loss function, and the multi-task loss function includes an environment classification loss function and a geometric quality regression loss function.

[0021] The embodiments of the present application are applicable to various GNSS application scenarios that require high-precision positioning, such as autonomous driving, drone navigation, precision agriculture, etc. Specifically, in the embodiments of the present application, it can be applied to a GNSS system, which may include a space satellite constellation, a GNSS receiver, a data preprocessing module, a training server, and a pre-trained bidirectional LSTM network model, wherein multiple satellites in the space satellite constellation are used to transmit radio signals, the GNSS receiver is used to collect observation data, the data preprocessing module is used to preprocess the observation data, and the training server is used to construct and train the bidirectional LSTM network model, wherein the bidirectional LSTM network model is obtained by joint training using a multi-task loss function (including an environment classification loss function and a geometric quality regression loss function).

[0022] In practical applications, the GNSS system can receive radio signals transmitted by satellites through a GNSS receiver to parse and obtain target observation data from the radio signals. The target observation data can then be preprocessed through a data preprocessing module so that the preprocessed target observation data can be input into a pre-trained bidirectional LSTM network model.

[0023] Preprocessing of target observation data can include spatiotemporal data alignment. Specifically, there is a time delay between the satellite's launch time and the GNSS receiver's reception time. Since the GNSS receiver is installed on the user's terminal device, the GNSS receiver can calculate the user's position, speed, time, and other information based on this time delay. To obtain accurate time delays, the GNSS system can synchronize the satellite's launch time and the GNSS receiver's reception time to the same time reference through a data preprocessing module. Generally speaking, the satellite's clock accuracy is greater than the GNSS receiver's clock accuracy. Therefore, the data preprocessing module can use the satellite's clock as a reference and utilize a 30-second sliding window to compensate for the GNSS receiver's clock jumps, thereby achieving spatiotemporal data alignment.

[0024] S120 calculates the key features of satellite geometric configuration changes based on target observation data through a bidirectional LSTM network model.

[0025] Each frame of target observation data may include data from multiple satellites, and the data from each satellite may include at least one of pseudorange, carrier phase, Doppler shift, C / N0, satellite elevation angle, satellite azimuth angle, satellite system type, frequency band identifier, and health status. Furthermore, in an embodiment of the present application, pseudorange, carrier phase, Doppler shift, C / N0, satellite elevation angle, and satellite azimuth angle may be used as mandatory features, and satellite system type, frequency band identifier, and health status may be used as optional features. In other words, the data from each satellite must include pseudorange, carrier phase, Doppler shift, C / N0, satellite elevation angle, and satellite azimuth angle. On this basis, auxiliary information such as satellite system type, frequency band identifier, and health status may be introduced.

[0026] It should be noted that pseudorange refers to the geometric distance obtained by the GNSS receiver by measuring the satellite signal propagation time (i.e., delay) multiplied by the speed of light, and carrier phase refers to the phase difference of the carrier signal measured by the GNSS receiver, which reflects the instantaneous distance change between the satellite and the GNSS receiver. Therefore, pseudorange and carrier phase are both measures of the distance between the satellite and the GNSS receiver. Since pseudorange is susceptible to multipath effects and noise, while carrier phase is insensitive to these effects, the accuracy of carrier phase is higher than that of pseudorange. In the embodiment of the present application, the method of smoothing pseudorange with carrier phase can be used to improve the accuracy of pseudorange, so that the key features calculated subsequently using pseudorange can be more accurate. The Doppler effect refers to the change in carrier frequency caused by the relative motion between the satellite and the GNSS receiver; C / N0 (Carrier-to-Noise Density Ratio) refers to the ratio of signal power to noise power spectral density; satellite elevation angle refers to the vertical angle of the satellite relative to the horizontal plane of the GNSS receiver; satellite azimuth angle refers to the horizontal direction angle of the satellite on the horizontal plane; satellite system type refers to the satellite system to which the satellite belongs. Among them, GNSS systems can include Beidou, Galileo, GPS (Global Positioning System), GLONASS and other satellite systems; frequency band identifier refers to the frequency band number of the satellite signal; health status refers to the flag indicating whether the satellite is available.

[0027] The bidirectional LSTM network model receives preprocessed target observation data and calculates key features of satellite geometry changes based on the target observation data. These key features can include at least one of satellite geometry and satellite signal quality features. Satellite geometry features can include at least one of the satellite azimuth angle change rate and the satellite elevation angle entropy value. Satellite signal quality features can include the Multipath Perception Index (MPI). These key features can comprehensively reflect changes in satellite geometry and the quality of satellite signals. The satellite azimuth angle change rate can reflect the satellite's motion relative to the GNSS receiver, helping to determine whether the satellite is moving rapidly or experiencing an anomaly. The satellite elevation angle entropy value can assess the uniformity and stability of the satellite geometric distribution by measuring the dispersion of satellite elevation angles. The Multipath Combination Indicator (MPI) is a comprehensive approach to assessing satellite signal quality, detecting multipath effects in satellite signals to determine their purity and reliability.

[0028] S130 uses a bidirectional LSTM network model to capture the forward and backward dependencies of the time series composed of key features to monitor changes in satellite geometric configuration.

[0029] The bidirectional LSTM network model consists of two LSTMs (Long Short-Term Memory) networks. One LSTM processes the forward time series, and the other processes the reverse time series. Therefore, the bidirectional LSTM network model can capture the forward and backward dependencies of the time series composed of key features. Each frame of target observation data carries a corresponding timestamp. Therefore, based on these timestamps, forward and reverse time series can be obtained. The forward time series is composed of key features sorted in forward time, and the reverse time series is composed of key features sorted in reverse time.

[0030] Because key features (such as the satellite azimuth rate of change and the satellite elevation entropy) can comprehensively reflect changes in satellite geometry, the bidirectional LSTM network model, by capturing the dependencies between the time series formed by these key features (such as the satellite azimuth rate of change and the satellite elevation entropy), can more comprehensively understand the changes in satellite geometry, thereby accurately and rapidly monitoring changes in satellite geometry. Furthermore, because key features (such as the MPI) can also comprehensively reflect the quality of satellite signals, the bidirectional LSTM network model, by capturing the dependencies between the time series formed by these key features (such as the MPI), can also more comprehensively understand the changes in satellite signal quality, thereby accurately and rapidly monitoring changes in satellite signal quality.

[0031] S140, predicting the environment of the GNSS receiver and the degradation degree of the satellite geometry based on the change of the satellite geometry through a bidirectional LSTM network model.

[0032] The environment in which a GNSS receiver operates can be dynamic. For example, a GNSS receiver can be mounted on a vehicle, drone, or other device. When a vehicle is in motion or a drone is in flight, the environment in which the GNSS receiver operates is subject to dynamic changes. Furthermore, when a GNSS receiver is located in a complex environment, such as an urban canyon, it is susceptible to interference such as multipath, signal obstruction, reduced satellite visibility, and NLOS, which can affect the positioning accuracy and reliability of the GNSS system.

[0033] It should be noted that the multipath effect refers to the satellite signal reflecting from the surface of a tall building and reaching the GNSS receiver through different paths and delays, resulting in the superposition of the reflected signal and the direct signal; signal obstruction and reduced satellite visibility refer to the fact that the field of view of the GNSS receiver is limited due to obstruction by tall buildings, resulting in a reduction in the number of visible satellites; NLOS (Non-Line-Of-Sight) means that the GNSS receiver cannot directly receive satellite signals, but receives satellite signals through reflection, refraction and other means.

[0034] In this regard, in the process of training the bidirectional LSTM network model, in order to simultaneously optimize the environmental classification prediction task and the geometric degradation prediction task, the embodiment of the present application can construct a multi-task loss function, which includes at least an environmental classification loss function and a geometric quality regression loss function, so that the bidirectional LSTM network model can be jointly trained using the environmental classification loss function and the geometric quality regression loss function, so that the trained bidirectional LSTM network model can accurately and quickly predict the environment in which the GNSS receiver is located and the degradation degree of the satellite geometric configuration based on the changes in the satellite geometric configuration, wherein the degradation degree can be represented by a degradation index, which reflects the degree of influence of the satellite geometric distribution on the positioning accuracy of the GNSS system.

[0035] In the specific implementation, the bidirectional LSTM network model determines the satellite geometry at each moment based on changes in the satellite geometry, and determines the satellite signal quality at each moment based on changes in the satellite signal quality. Then, based on the satellite geometry at each moment, it predicts the environment in which the GNSS receiver is located, and based on the satellite signal quality at each moment, it predicts the degradation of the satellite geometry at each moment.

[0036] S150: Positioning is performed using a data processing method that matches the environment in which the GNSS receiver is located and the degradation degree of the satellite geometry.

[0037] Different scenarios correspond to different data processing methods. The GNSS system can determine the current scenario of the GNSS receiver based on the environment in which the GNSS receiver is located and the degradation of the satellite geometric configuration, so as to adopt a data processing method that matches the current scenario for positioning.

[0038] In one example, assume that the bidirectional LSTM network model predicts that the environment in which the GNSS receiver is located at time T0~T1 is open sky. Since the GNSS receiver is subject to less and weak interference (such as multipath effect, NLOS, etc.) in the open sky environment, the satellite signal quality in the open sky environment is good. Therefore, the satellite geometric configuration predicted by the bidirectional LSTM network model has a low degradation degree at time T0~T1. In this case, the GNSS system can use data processing method A for positioning at time T0~T1. Data processing method A is suitable for scenarios with open sky environments and low degradation.

[0039] In another example, suppose the bidirectional LSTM network model predicts that the environment in which the GNSS receiver is located at time T1~T2 is an urban canyon. Since the GNSS receiver is subject to more and more severe interference (such as multipath effect, NLOS, etc.) when in the urban canyon environment, the satellite signal quality in the urban canyon environment is poor. Therefore, the satellite geometric configuration predicted by the bidirectional LSTM network model has a high degree of degradation at time T1~T2. In this case, the GNSS system can use data processing method B for positioning at time T1~T2. Data processing method B is suitable for scenarios with urban canyon environments and high degradation.

[0040] It can be seen that the embodiments of the present application use different data processing methods to perform positioning in different scenarios, thereby ensuring the positioning accuracy and reliability of the GNSS system in different scenarios.

[0041] As can be seen from this example, the solution provided by this application obtains the target observation data collected by the GNSS receiver and inputs the target observation data into a pre-trained bidirectional LSTM network model; wherein, the bidirectional LSTM network model is obtained by joint training using a multi-task loss function, and the multi-task loss function includes an environmental classification loss function and a geometric quality regression loss function; the bidirectional LSTM network model is used to calculate the key features of the satellite geometric configuration changes based on the target observation data; the bidirectional LSTM network model is used to capture the front-end and back-end dependencies of the time series composed of key features to monitor the satellite geometric configuration changes; the bidirectional LSTM network model is used to predict the environment of the GNSS receiver and the degradation degree of the satellite geometric configuration based on the satellite geometric configuration changes; and positioning is performed using a data processing method that matches the environment of the GNSS receiver and the degradation degree of the satellite geometric configuration. This application uses a bidirectional LSTM network model to capture the forward and backward dependencies of the time series composed of key features of satellite geometric configuration changes, which can improve the accuracy and real-time performance of satellite geometric configuration change monitoring. Moreover, by jointly training the bidirectional LSTM network model with multiple tasks, the comprehensive performance and generalization ability of the bidirectional LSTM network model can be improved, so that the bidirectional LSTM network model can achieve good prediction results in different scenarios. Therefore, based on the monitored satellite geometric configuration changes, the bidirectional LSTM network model can accurately and quickly predict the environment in which the GNSS receiver is located and the degradation degree of the satellite geometric configuration, thereby breaking the limitations of related technologies in dealing with complex dynamic environments and multipath effects, and enhancing the positioning accuracy and reliability of the GNSS system in complex environments.

[0042] Figure 2 This is another flowchart of the satellite geometric configuration change monitoring method shown in this application.

[0043] See also Figure 2The satellite geometric configuration change monitoring method of the present application includes: S210, obtaining target observation data collected by a GNSS receiver, and inputting the target observation data into a pre-trained bidirectional LSTM network model; wherein the bidirectional LSTM network model is obtained by joint training using a multi-task loss function, the multi-task loss function includes an environment classification loss function and a geometric quality regression loss function, the bidirectional LSTM network model includes a satellite system embedding layer and a bidirectional LSTM core layer, and the bidirectional LSTM core layer includes a forward LSTM layer, a reverse LSTM layer, and an LSTM feature fusion layer.

[0044] In practical applications, the GNSS system can receive radio signals transmitted by satellites through a GNSS receiver to parse and obtain target observation data from the radio signals. The target observation data can then be preprocessed through a data preprocessing module so that the preprocessed target observation data can be input into a pre-trained bidirectional LSTM network model. The bidirectional LSTM network model is obtained by joint training using a multi-task loss function (such as an environmental classification loss function and a geometric quality regression loss function).

[0045] See also Figure 3 The bidirectional LSTM network model can include an input layer, a feature engineering module, a satellite system embedding layer, a bidirectional LSTM core layer, a self-attention mechanism layer, and an output layer. The bidirectional LSTM core layer can include a forward LSTM layer, a backward LSTM layer, and an LSTM feature fusion layer. The output layer can include an environmental classification prediction output layer and a geometric degradation prediction output layer. The input layer is used to receive preprocessed target observation data and convert it into a format suitable for network processing.

[0046] S220 calculates the key features of satellite geometric configuration changes based on target observation data through a bidirectional LSTM network model.

[0047] The feature engineering module is used to calculate key features of satellite geometry changes based on target observation data. Each frame of target observation data may include data from multiple satellites, and each satellite's data may include at least one of pseudorange, carrier phase, Doppler shift, C / N0, satellite elevation angle, satellite azimuth angle, satellite system type, frequency band identifier, and health status. Key features may include at least one of satellite geometry and satellite signal quality features. Satellite geometry may include at least one of satellite azimuth angle change rate and satellite elevation angle entropy. Satellite signal quality features may include MPI. These key features comprehensively reflect changes in satellite geometry and the quality of satellite signals.

[0048] In one embodiment, calculating key features of satellite geometric configuration changes based on target observation data using a bidirectional LSTM network model may include: The satellite azimuth angle change rate is calculated based on the satellite azimuth angle, carrier phase and Doppler frequency shift through a bidirectional LSTM network model; and / or, the satellite elevation angle entropy is calculated based on the satellite elevation angle, C / N0, satellite system type and frequency band identifier through a bidirectional LSTM network model; and / or, the MPI is calculated based on the pseudorange and satellite signal frequency through a bidirectional LSTM network model.

[0049] The feature engineering module calculates each satellite's azimuth rate of change based on its azimuth, carrier phase, and Doppler shift. The azimuth rate of change reflects the satellite's motion relative to the GNSS receiver, helping to determine whether the satellite is moving rapidly or experiencing an anomaly. Furthermore, auxiliary information such as the satellite system type, frequency band identifier, and health status can affect the characteristics and trajectory of the satellite signal, and thus the azimuth rate of change. Therefore, the feature engineering module can also consider these auxiliary information when calculating the azimuth rate of change.

[0050] Auxiliary information such as satellite system type and frequency band identifier determines the characteristics and propagation mode of satellite signals, which in turn affects the calculation and analysis of satellite elevation angle entropy. Therefore, the feature engineering module can calculate the satellite elevation angle entropy of each satellite based on the satellite elevation angle, C / N0, satellite system type and frequency band identifier of each satellite. The satellite elevation angle entropy can evaluate the uniformity and stability of satellite geometric distribution by measuring the degree of dispersion of satellite elevation angles.

[0051] The feature engineering module can calculate the MPI of each satellite signal based on the pseudorange and satellite signal frequency of each satellite. The satellite signal frequency (i.e., frequency point) is the operating frequency of the satellite transmitting the radio signal. The frequency point of each satellite is fixed. For example, the calculation formula of MPI is as follows: MPI = P1-(f1 2 ×P2+f2 2 ×P1) / (f1 2 + f2 2 ) Here, P1 and P2 represent the pseudoranges of different satellites, and f1 and f2 represent the frequencies of different satellites. MPI is a comprehensive method for evaluating satellite signal quality. It detects the multipath effect in satellite signals to determine their purity and reliability.

[0052] S230 , mapping identifiers of different satellite systems to embedding vectors of fixed dimensions through a satellite system embedding layer.

[0053] In order to process data from different satellite systems, the embodiment of the present application can introduce a satellite system embedding layer (such as Figure 3 The target observation data may also include identifiers of different satellite systems (such as BeiDou, Galileo, GPS, and GLONASS). The satellite system embedding layer is used to map the identifiers of different satellite systems (such as BeiDou, Galileo, GPS, and GLONASS) to an embedding vector of fixed dimension, so that the bidirectional LSTM network model can learn the differences and commonalities between different satellite systems, thereby enhancing the adaptability of the bidirectional LSTM network model to different satellite systems.

[0054] S240, concatenates the embedding vector and the key features at the time step through the satellite system embedding layer to obtain the input feature vector.

[0055] After obtaining the embedding vector, the satellite system embedding layer concatenates the embedding vector with the key features calculated by the feature engineering module (such as satellite azimuth angle change rate, satellite elevation angle entropy, and MPI) at the time step to form a complete input feature vector.

[0056] S250, through the forward LSTM based on the forward time series in the input feature vector, captures the impact of the past time step on the current moment and obtains the forward expression vector of the key feature.

[0057] like Figure 3 As shown, the bidirectional LSTM core layer can include a forward LSTM layer, which is used to process the forward time series. Specifically, the forward time series includes key features sorted in forward time. Therefore, the forward LSTM captures the impact of past time steps on the current moment based on the key features sorted in forward time in the input feature vector, thereby obtaining a forward expression vector for the key features.

[0058] Since the forward expression vector is an integrated reflection of the process information from the beginning of the sequence to the current time step by the forward LSTM, the forward expression vector can represent the feature representation formed by the comprehensive impact of the forward time series on the current moment.

[0059] In one embodiment, the forward LSTM is used to capture the impact of past time steps on the current moment based on the forward time series in the input feature vector to obtain a forward expression vector of the key features, which may include: Through the forward LSTM, based on the key features from the start time to the end time in the input feature vector, the changing pattern of the key features with the forward time sequence is captured to obtain the forward expression vector of the key features.

[0060] In one example, the satellite azimuth angle change rate is a dynamic time series feature, and its changing trend needs to consider the association between past time steps and future time steps at the same time. Therefore, the forward LSTM can capture the change pattern of the satellite azimuth angle with the forward time sequence (such as the satellite orbit motion characteristics) based on the satellite azimuth angle change rate from the start time to the end time in the input feature vector, thereby obtaining a forward expression vector of the satellite azimuth angle change rate.

[0061] In another example, the satellite elevation angle entropy reflects the stability of the satellite's geometric distribution. The bidirectional LSTM network model can distinguish between periodic fluctuations under normal conditions (such as satellite constellation switching) and abnormal fluctuations (such as a sudden drop in entropy due to sudden multipath interference) by analyzing the entropy change pattern of the previous and next time windows. Therefore, the forward LSTM can capture the change pattern of the satellite elevation angle entropy with the forward time sequence based on the satellite elevation angle entropy from the start time to the end time in the input feature vector, thereby obtaining a forward expression vector of the satellite elevation angle entropy. This forward expression vector can reflect whether the change pattern of the satellite elevation angle entropy with the forward time sequence is a periodic fluctuation under normal conditions (such as satellite constellation switching) or an abnormal fluctuation (such as a sudden drop in entropy due to sudden multipath interference).

[0062] In another example, MPI is a signal quality feature, and its time series is coupled with geometric features. For example, during forward propagation, the current value of MPI affects the interpretation of geometric features in subsequent time steps (for example, a high MPI indicates multipath risk, and the trust in the satellite elevation angle entropy value needs to be reduced). Therefore, the forward LSTM can capture the change pattern of MPI with the forward time sequence based on the MPI from the start time to the end time in the input feature vector, thereby obtaining the forward expression vector of MPI.

[0063] S260, through the reverse LSTM based on the reverse time series in the input feature vector, captures the impact of future time steps on the current moment and obtains the reverse expression vector of the key features.

[0064] like Figure 3 As shown in Figure 1, the bidirectional LSTM core layer can also include a reverse LSTM layer, which is used to process reverse time series. Specifically, the reverse time series includes key features sorted in reverse time. Therefore, the reverse LSTM captures the impact of future time steps on the current moment based on the key features sorted in reverse time in the input feature vector, thereby obtaining a reverse expression vector for the key features.

[0065] Since the reverse expression vector is an integrated reflection of the process information from the end of the sequence to the current time step by the reverse LSTM, the reverse expression vector can represent the feature representation formed by the comprehensive impact of the reverse time series on the current moment.

[0066] In one embodiment, the reverse LSTM is used to capture the impact of future time steps on the current moment based on the reverse time series in the input feature vector to obtain the reverse expression vector of the key features, which may include: Through the reverse LSTM, based on the key features from the end time to the start time in the input feature vector, the changing pattern of the key features in the reverse time sequence is captured to obtain the reverse expression vector of the key features.

[0067] In one example, the satellite azimuth angle change rate is a dynamic time series feature, and its changing trend needs to consider the association between past time steps and future time steps at the same time. Therefore, the reverse LSTM can capture the change pattern of the satellite azimuth angle along the reverse time sequence (such as mutations caused by occlusion) based on the satellite azimuth angle change rate from the end time to the start time in the input feature vector, thereby obtaining the reverse expression vector of the satellite azimuth angle change rate.

[0068] In another example, the satellite elevation angle entropy reflects the stability of the satellite's geometric distribution. The bidirectional LSTM network model can distinguish between periodic fluctuations under normal conditions (such as satellite constellation switching) and abnormal fluctuations (such as a sudden drop in entropy due to sudden multipath interference) by analyzing the entropy change pattern of the previous and next time windows. Therefore, the reverse LSTM can capture the change pattern of the satellite elevation angle entropy in the reverse time sequence based on the satellite elevation angle entropy from the end time to the start time in the input feature vector, thereby obtaining a reverse expression vector of the satellite elevation angle entropy. This reverse expression vector can reflect whether the change pattern of the satellite elevation angle entropy in the reverse time sequence is a periodic fluctuation under normal conditions (such as satellite constellation switching) or an abnormal fluctuation (such as a sudden drop in entropy due to sudden multipath interference).

[0069] In another example, MPI is a signal quality feature, and its time series is coupled with geometric features. For example, during backpropagation, the geometric degradation results of subsequent time steps (such as increased positioning error) can correct the weight distribution of the current MPI. Therefore, the reverse LSTM can capture the change pattern of MPI in the reverse time sequence based on the MPI from the end time to the start time in the input feature vector, thereby obtaining the reverse expression vector of MPI.

[0070] S270, through the LSTM feature fusion layer, fuses the forward expression vector and the reverse expression vector to obtain the front-end dependency relationship of the time series to monitor the changes in the satellite's geometric configuration.

[0071] After obtaining the forward expression vector and the reverse expression vector of the key feature, the LSTM feature fusion layer can fuse the forward expression vector and the reverse expression vector, for example, splicing the hidden states in the two directions, so as to obtain the fused overall information. The overall information can reflect the front-end dependency of the time series, for example, the front-end dependency of the forward time series, or the front-end dependency of the reverse time series. Then the LSTM feature fusion layer uses the overall information as input for subsequent processing, for example, passing the overall information through operations such as the fully connected layer and activation function to determine the degree of attention to the key feature.

[0072] In the specific implementation, the hidden layer dimension of the bidirectional LSTM core layer is set to 64, and two LSTM layers (i.e., a forward LSTM layer and a backward LSTM layer) are stacked to increase the network depth and expressiveness of the bidirectional LSTM model. Input data is organized based on time steps, batch size, and feature dimensions to ensure that the bidirectional LSTM model can efficiently process large-scale time series data. The LSTM feature fusion layer can dynamically adjust its focus on key features through a bidirectional gating structure consisting of an input gate and a forget gate. When the overall information of a key feature has a high weight, the LSTM feature fusion layer can increase its focus on this key feature through the input gate; when the overall information of a key feature has a low weight, the LSTM feature fusion layer can reduce its focus on this key feature through the forget gate.

[0073] Among them, the key features with high attention have regular abnormal changes that can represent the changes in the satellite's geometric configuration. Therefore, the LSTM feature fusion layer can monitor the changes in the satellite's geometric configuration based on the key features with high attention.

[0074] In one embodiment, the bidirectional LSTM network model further includes a self-attention mechanism layer; the method may further include: The self-attention mechanism layer calculates the similarity between the key features of different time steps in the input feature vector; the self-attention mechanism layer performs weighted processing on the front and back dependencies of the time series based on the similarity to obtain the target attention output.

[0075] In order to further enhance the ability of the bidirectional LSTM model to capture key features, the embodiment of the present application can introduce a self-attention mechanism layer (such as Figure 3 The self-attention mechanism layer calculates the correlation between different time steps in the input feature vector, and then weights the dependencies between the time series based on the correlation between different time steps to obtain the final target attention output, so as to assign weights to the hidden state of each time step.

[0076] In a specific implementation, the self-attention mechanism layer can adopt a multi-head attention mechanism to convert the hidden state of the input into a query vector (Q), a key vector (K), and a value vector (V), and then calculate the attention score of each time step based on the query vector (Q), key vector (K), and value vector (V) of each time step, so as to perform a weighted summation of the attention scores of all time steps to obtain the final target attention output, and then assign the target attention output to the hidden state of each time step.

[0077] It should be noted that the length of the time series processed by the self-attention mechanism layer is greater than the length of the time series processed by the bidirectional LSTM core layer. For example, the bidirectional LSTM core layer can capture the front-to-back dependency of the time series composed of the key features of 100 frames of data, while the self-attention mechanism layer can capture the correlation between the key features of different frames of data from the overall data (such as the key features of 1000 frames of data). For example, the self-attention mechanism layer can capture the correlation between the key features of the first frame of data and the 1000th frame of data. Therefore, the introduction of the self-attention mechanism layer can enable the bidirectional LSTM model to dynamically focus on the most important features related to changes in satellite geometric configuration, thereby improving the sensitivity and capture ability of the bidirectional LSTM model to changes in satellite geometric configuration.

[0078] In addition, in order to estimate and suppress the noise components in the data in real time, the embodiment of the present application can introduce an adaptive noise suppression module into the bidirectional LSTM network model ( Figure 3 Not shown). Specifically, an embodiment of the present application can add a noise estimation and suppression subnetwork to each time step of the bidirectional LSTM network model. The subnetwork is used to dynamically adjust the strength of noise suppression based on current data and historical data. At the same time, an embodiment of the present application can fuse satellite signal strength, Doppler frequency shift, and satellite geometric features, and utilize the correlation between them to improve the effect of noise suppression. Specifically, the processing results of the subnetwork and satellite signal strength, Doppler frequency shift, and satellite geometric features can be used as input, and feature fusion and noise suppression can be performed through shared weights and cross-connections.

[0079] S280 predicts the degradation of the GNSS receiver environment and satellite geometry based on changes in satellite geometry using a bidirectional LSTM network model.

[0080] In the process of training the bidirectional LSTM network model, in order to simultaneously optimize the environmental classification prediction task and the geometric degradation prediction task, the embodiment of the present application can construct a multi-task loss function, which includes at least an environmental classification loss function and a geometric quality regression loss function. Accordingly, the embodiment of the present application constructs an environmental classification prediction output layer and a geometric degradation prediction output layer (such as Figure 3 shown).

[0081] For the environmental classification prediction task, the environmental classification prediction output layer uses a fully connected layer to map the satellite geometric configuration changes obtained after the target observation data is processed by the bidirectional LSTM core layer and the self-attention mechanism layer to the corresponding environmental category, thereby obtaining the environmental category prediction result. The environmental category prediction result is used to reflect the environment in which the GNSS receiver is located (such as open sky, urban canyon, indoor environment, etc.).

[0082] For the geometric degradation prediction task, the geometric degradation prediction output layer adopts a regression structure to predict the degradation of the satellite geometric configuration, thereby obtaining the geometric degradation prediction result. The geometric degradation prediction result is used to reflect the impact of the satellite geometric distribution on the positioning accuracy of the GNSS system.

[0083] S290 uses a data processing method that matches the GNSS receiver's environment and the degradation of the satellite geometry for positioning.

[0084] Different scenarios correspond to different data processing methods. The GNSS system can determine the current scenario of the GNSS receiver based on the environment in which the GNSS receiver is located and the degradation of the satellite geometric configuration, so as to adopt a data processing method that matches the current scenario for positioning.

[0085] It can be seen that the embodiments of the present application use different data processing methods to perform positioning in different scenarios, thereby ensuring the positioning accuracy and reliability of the GNSS system in different scenarios.

[0086] In one embodiment, the bidirectional LSTM network model can be trained in the following manner: Acquire multiple frames of historical observation data collected by a GNSS receiver; divide the multiple frames of historical observation data into environmental categories to obtain historical observation data sets of different environmental categories; different environmental categories include simple environmental categories and complex environmental categories; in the first training phase, extract static data from the historical observation data set of the simple environmental category, and use the static data to train the bidirectional LSTM network model until the target loss value of the bidirectional LSTM network model in the first training phase is less than the preset loss value; in the second training phase, extract dynamic data from the historical observation data set of the complex environmental category, and use the dynamic data to train the bidirectional LSTM network model until the bidirectional The target loss value of the LSTM network model in the second training stage is less than the preset loss value; in the third training stage, the bidirectional LSTM network model is trained using full-scene mixed data until the target loss value of the bidirectional LSTM network model in the third training stage is less than the preset loss value, and a trained bidirectional LSTM network model is obtained; wherein the full-scene mixed data is a historical observation data set of different environmental categories; the weights of the environmental classification loss function in the first training stage, the second training stage and the third training stage respectively show a decreasing law; the weights of the geometric quality regression loss function in the first training stage, the second training stage and the third training stage respectively show an increasing law.

[0087] For the training of the bidirectional LSTM network model, the embodiment of the present application can first construct the network structure of the bidirectional LSTM network model through the training server, such as Figure 3 As shown, the bidirectional LSTM network model may include an input layer, a feature engineering module, a satellite system embedding layer, a bidirectional LSTM core layer, a self-attention mechanism layer and an output layer. The bidirectional LSTM core layer may include a forward LSTM layer, a reverse LSTM layer and an LSTM feature fusion layer. The output layer may include an environmental classification prediction output layer and a geometric degradation prediction output layer. In addition, the embodiment of the present application can obtain multiple frames of historical observation data collected by the GNSS receiver in different scenarios and different time periods. Each frame of historical observation data may include data from multiple satellites, and the data from each satellite may include at least one of pseudorange, carrier phase, Doppler shift, C / N0, satellite elevation angle, satellite azimuth angle, satellite system type, frequency band identifier and health status.

[0088] In an embodiment of the present application, each frame of historical observation data can be preprocessed by a data preprocessing module. The preprocessing of historical observation data may include spatiotemporal data alignment and data enhancement processing, wherein the specific operation of spatiotemporal data alignment can be referred to the above step S110 and will not be repeated here. Among them, for the data enhancement processing of historical observation data, as an example, the data preprocessing module can simulate the data missing situation that may occur in the real environment by randomly blocking the data of some satellites, so as to improve the anti-interference ability of the bidirectional LSTM network model. For example, a certain number of satellite data can be randomly selected and set to zero or a default value to simulate the effect of data loss; as another example, the data preprocessing module can generate new sample data by mixing data from different scenes or different time periods. This method can increase the diversity of data, so that the bidirectional LSTM network model can learn more complex patterns and features; as another example, the data preprocessing module can simulate irregular sampling in time by making a small perturbation to the time step of the time series composed of historical observation data, such as randomly inserting or deleting a small number of time steps.

[0089] After obtaining the pre-processed multi-frame historical observation data, the embodiment of the present application can input these historical observation data into the input layer, feature engineering module, satellite system embedding layer, bidirectional LSTM core layer, self-attention mechanism layer and output layer of the bidirectional LSTM network model in sequence for processing. Among them, the processing flow of the bidirectional LSTM network model for historical observation data can be specifically referred to the above-mentioned processing flow of the bidirectional LSTM network model for target observation data (i.e., steps S210 to S280), which will not be repeated here.

[0090] For the environmental classification prediction task, the environmental classification prediction output layer uses a fully connected layer to map the satellite geometric configuration changes obtained after the historical observation data is processed by the bidirectional LSTM core layer and the self-attention mechanism layer to the corresponding environmental category, thereby obtaining the environmental category prediction result. The environmental category prediction result is used to reflect the environment in which the GNSS receiver is located (such as open sky, urban canyon, indoor environment, etc.).

[0091] For the geometric degradation prediction task, the geometric degradation prediction output layer adopts a regression structure to predict the degradation of the satellite geometric configuration, thereby obtaining the geometric degradation prediction result. The geometric degradation prediction result is used to reflect the impact of the satellite geometric distribution on the positioning accuracy of the GNSS system.

[0092] It should be noted that each frame of historical observation data has been labeled, so each frame of historical observation data carries a true value label, which can include an environmental category label and a geometric degradation degree label.

[0093] The embodiment of the present application can utilize a multi-task joint training strategy and a curriculum learning strategy to realize the processing capability of a bidirectional LSTM network model in different scenarios.

[0094] In a multi-task joint training strategy, in order to simultaneously optimize the environmental classification prediction task and the geometric degradation prediction task, an embodiment of the present application can construct a multi-task loss function. The multi-task loss function can include an environmental classification loss function and a geometric quality regression loss function. The loss value of the environmental classification loss function can adopt a cross-entropy loss, and the loss value of the geometric quality regression loss function can adopt a mean square error loss. To ensure that the prediction results of the bidirectional LSTM network model meet expectations in a specific environment, the multi-task loss function can also include a consistency constraint loss function. An embodiment of the present application can achieve comprehensive optimization of the bidirectional LSTM network model by weighted summing the environmental classification loss (such as cross-entropy loss), geometric quality regression loss (such as mean square error loss), and consistency constraint loss. The environmental classification loss refers to the loss value of the environmental classification loss function, the geometric quality regression loss refers to the loss value of the geometric quality regression loss function, and the consistency constraint loss refers to the loss value of the consistency constraint loss function.

[0095] In the course learning strategy, the embodiment of the present application can divide the training process of the bidirectional LSTM network model into three training stages: the first training stage, the second training stage and the third training stage, and then the embodiment of the present application can configure the corresponding sample data set and the weight of the multi-task loss function for each training stage.

[0096] In a specific implementation, embodiments of the present application can classify multiple frames of historical observation data into environmental categories based on the environmental category label of each frame of historical observation data, thereby obtaining historical observation datasets of different environmental categories. The different environmental categories can include simple and complex environmental categories. Simple environmental categories can include open sky environments, for example, while complex environmental categories can include urban canyon environments, indoor environments, and the like. For the first training phase, embodiments of the present application can extract static data from the historical observation dataset of the open sky environment, using the static data in the open sky environment as the sample dataset for the first training phase. Static data refers to historical observation data collected when the GNSS receiver is stationary. For the second training phase, embodiments of the present application can extract dynamic data from the historical observation dataset of the urban canyon environment, using the dynamic data in the urban canyon environment as the sample dataset for the second training phase. Dynamic data refers to historical observation data collected when the GNSS receiver is in motion. For the third training phase, embodiments of the present application can use full-scene mixed data and adversarial samples as the sample dataset for the third training phase. The full-scene mixed data refers to historical observation datasets of different environmental categories (i.e., all historical observation data), and the adversarial samples refer to sample data generated by a generative adversarial network.

[0097] It should be noted that the weights of the environment classification loss function in the first, second, and third training stages follow a decreasing pattern, i.e., the weight of the environment classification loss in the first training stage > the weight of the environment classification loss in the second training stage > the weight of the environment classification loss in the third training stage. The weights of the geometric quality regression loss function in the first, second, and third training stages follow an increasing pattern, i.e., the weight of the geometric quality regression loss in the first training stage < the weight of the geometric quality regression loss in the second training stage < the weight of the geometric quality regression loss in the third training stage. Furthermore, the learning rate in the first training stage is > the learning rate in the second training stage > the learning rate in the third training stage.

[0098] In one example, in the first training stage, the learning rate is set to 1e-3, the environmental classification loss weight is set to 0.9, and the geometric quality regression loss weight is set to 0.1. The embodiment of the present application can input static data in an open sky environment into a bidirectional LSTM network model to obtain the environmental category prediction result and the geometric degradation prediction result output by the bidirectional LSTM network model for the static data in the open sky environment. Then, the embodiment of the present application can calculate the environmental classification loss between the environmental category prediction result and the environmental category label of the static data in the open sky environment, and can calculate the geometric degradation prediction result and the geometric degradation prediction result of the static data in the open sky environment. The geometric quality regression loss between degree labels, and the consistency constraint loss of static data in an open sky environment can be calculated. Then, the environmental classification loss, geometric quality regression loss and consistency constraint loss of the static data in the open sky environment can be weightedly summed to obtain the target loss value of the bidirectional LSTM network model in the first training stage. When the target loss value is not less than the preset loss value, the embodiment of the present application can use the target loss value through the optimizer to simultaneously update the parameters of the forward LSTM layer and the reverse LSTM layer until the target loss value of the bidirectional LSTM network model in the first training stage is less than the preset loss value, and the first training stage is ended. In the second training stage, the learning rate is adjusted to 5e-4, the environmental classification loss weight is adjusted to 0.7, and the geometric quality regression loss weight is adjusted to 0.3. The embodiment of the present application can input the dynamic data in the urban canyon environment into the bidirectional LSTM network model to obtain the environmental category prediction results and geometric degradation prediction results output by the bidirectional LSTM network model for the dynamic data in the urban canyon environment. Then, the embodiment of the present application can calculate the environmental classification loss between the environmental category prediction results of the dynamic data in the urban canyon environment and the environmental category label, and can calculate the geometric degradation prediction results of the dynamic data in the urban canyon environment and the geometric degradation label. The geometric quality regression loss between the signatures, and the consistency constraint loss of the dynamic data in the urban canyon environment can be calculated. Then, the environmental classification loss, geometric quality regression loss and consistency constraint loss of the dynamic data in the urban canyon environment can be weightedly summed to obtain the target loss value of the bidirectional LSTM network model in the second training stage. When the target loss value is not less than the preset loss value, the embodiment of the present application can use the target loss value through the optimizer to simultaneously update the parameters of the forward LSTM layer and the reverse LSTM layer until the target loss value of the bidirectional LSTM network model in the second training stage is less than the preset loss value, and the second training stage is ended.In the third training stage, the learning rate is further reduced to 1e-4, the environmental classification loss weight is adjusted to 0.5, and the geometric quality regression loss weight is adjusted to 0.5. The embodiment of the present application can input the full-scene mixed data and the adversarial sample into the bidirectional LSTM network model to obtain the environmental category prediction results and geometric degradation prediction results output by the bidirectional LSTM network model for the full-scene mixed data and the adversarial sample. Then, the embodiment of the present application can calculate the environmental classification loss between the environmental category prediction results of the full-scene mixed data and the adversarial sample and the environmental category label, and can calculate the geometric quality between the geometric degradation prediction results of the full-scene mixed data and the adversarial sample and the geometric degradation label. Regression loss, and the consistency constraint loss of the full-scene mixed data and the adversarial sample can be calculated, and then the environmental classification loss, geometric quality regression loss and consistency constraint loss of the full-scene mixed data and the adversarial sample can be weighted and summed to obtain the target loss value of the bidirectional LSTM network model in the third training stage. When the target loss value is not less than the preset loss value, the embodiment of the present application can use the target loss value through the optimizer to simultaneously update the parameters of the forward LSTM layer and the reverse LSTM layer until the target loss value of the bidirectional LSTM network model in the third training stage is less than the preset loss value, and the third training stage is ended to obtain a trained bidirectional LSTM network model.

[0099] As can be seen, the initial training of the bidirectional LSTM network model uses simple data, such as stable satellite data from open sky environments, to allow the bidirectional LSTM network model to learn basic features and patterns. As training progresses, more complex data, such as satellite data from complex environments such as urban canyons, is gradually introduced to increase the bidirectional LSTM network model's adaptability to complex scenarios.

[0100] As can be seen, the initial training of the bidirectional LSTM network model focuses on the environmental classification and prediction task, and therefore gives it a higher weight to help the bidirectional LSTM network model quickly learn the characteristics of satellite geometry in different environments. As training progresses, the weight of the geometric quality regression loss is gradually increased, allowing the bidirectional LSTM network model to accurately classify the environment while more precisely predicting the degradation of the satellite geometry.

[0101] In one embodiment, the method may further include: When the satellite geometric configuration change is greater than the preset change threshold or the prediction error of the bidirectional LSTM network model is greater than the preset error threshold, the online learning mechanism is triggered to incrementally update the parameters of the bidirectional LSTM network model and adjust the weight of the environmental classification loss function or the weight of the geometric quality regression loss function.

[0102] In order to make the bidirectional LSTM network model adapt to the long-term changes and dynamic characteristics of the environment, the embodiment of the present application introduces an online learning mechanism. The online learning mechanism dynamically adjusts the parameters of the bidirectional LSTM network model by real-time monitoring of environmental changes (such as the long-term drift of the satellite geometry, seasonal changes in the multipath effect, etc.), thereby improving the long-term adaptability and robustness of the bidirectional LSTM network model.

[0103] In a specific implementation, when the bidirectional LSTM network model detects that the change in satellite geometric configuration is greater than a preset change threshold (such as the satellite azimuth angle change rate exceeds the threshold, the satellite elevation angle entropy value fluctuates abnormally, etc.), the embodiment of the present application can trigger the online learning mechanism. Alternatively, when the prediction error of the environmental classification prediction result output by the bidirectional LSTM network model is greater than a preset error threshold, the embodiment of the present application can trigger the online learning mechanism. Alternatively, when the prediction error of the geometric degradation prediction result output by the bidirectional LSTM network model is greater than a preset error threshold, the embodiment of the present application can trigger the online learning mechanism.

[0104] After triggering the online learning mechanism, the embodiment of the present application can use a small batch gradient descent method to incrementally update the parameters of the bidirectional LSTM network model using the latest target observation data. In a specific implementation, the embodiment of the present application can extract small batch samples (such as the target observation data of the last 10 minutes) from the real-time data stream, and then calculate the target loss value of the bidirectional LSTM network model on the small batch samples, so as to use the target loss value to adjust the parameters of the bidirectional LSTM network model and update the weights of the bidirectional LSTM core layer and the self-attention mechanism layer. In addition, the embodiment of the present application can prevent the bidirectional LSTM network model from overfitting through regularization and constraint mechanisms (such as weight decay and gradient clipping).

[0105] During the online learning process, the embodiment of the present application can ensure that the bidirectional LSTM network model maintains stability during the incremental update process by retaining a snapshot of the historical parameters of the bidirectional LSTM network model and combining the weighted fusion of the new and old parameters. For example, the exponentially weighted moving average method is used to dynamically adjust the weights of the new and old parameters of the bidirectional LSTM network model. In addition, the embodiment of the present application can dynamically adjust the weights of the multi-task loss function. For example, when the prediction error of the environmental classification prediction result output by the bidirectional LSTM network model is greater than the preset error threshold, the embodiment of the present application can increase the environmental classification loss weight; or, when the prediction error of the geometric degradation prediction result output by the bidirectional LSTM network model is greater than the preset error threshold, the embodiment of the present application can increase the geometric quality regression loss weight, so that the bidirectional LSTM network model can adapt to environmental changes in real time.

[0106] As can be seen from this example, the solution provided by this application can improve the accuracy and real-time performance of satellite geometric configuration change monitoring by capturing the front-to-back dependencies of the time series composed of key features of satellite geometric configuration changes through a bidirectional LSTM core layer.

[0107] Furthermore, the solution provided in this application, by adopting multi-task loss functions (such as environmental classification loss function, geometric quality regression loss function and consistency constraint loss function, etc.) and curriculum learning strategies to jointly train the bidirectional LSTM network model, can improve the comprehensive performance and generalization ability of the bidirectional LSTM network model, so that the bidirectional LSTM network model can achieve good prediction results in different scenarios. Therefore, the bidirectional LSTM network model can accurately and quickly predict the environment of the GNSS receiver and the degradation degree of the satellite geometric configuration based on the monitored changes in the satellite geometric configuration, thereby breaking the limitations of related technologies in dealing with complex dynamic environments and multipath effects, and enhancing the positioning accuracy and reliability of the GNSS system in complex environments.

[0108] Furthermore, the solution provided in this application encodes the differences between different satellite systems into learnable embedding vectors through the satellite system embedding layer, which can enhance the cross-constellation generalization ability of the bidirectional LSTM network model, thereby improving the processing effect of the bidirectional LSTM network model on multi-constellation GNSS data.

[0109] Furthermore, the solution provided in this application performs weighted processing on the output of the bidirectional LSTM core layer through a self-attention mechanism layer to highlight important features related to changes in satellite geometric configuration, thereby improving the sensitivity and capture ability of the bidirectional LSTM network model to changes in satellite geometric configuration.

[0110] Furthermore, the solution provided in this application can improve the robustness, noise resistance and generalization ability of the bidirectional LSTM network model by performing data enhancement processing on sample data (i.e., historical observation data) during the training process of the bidirectional LSTM network model, thereby ensuring the reliability and stability of the bidirectional LSTM network model in extreme satellite conditions.

[0111] Furthermore, the solution provided in this application can capture the long-term changes in the satellite's geometric configuration in real time through an online learning mechanism, and dynamically adjust the parameters of the bidirectional LSTM network model so that the bidirectional LSTM network model can adapt to long-term environmental changes, ensuring that the bidirectional LSTM network model maintains high precision and high reliability during long-term operation, thereby improving the robustness and long-term stability of the bidirectional LSTM network model.

[0112] Corresponding to the aforementioned application function implementation method embodiment, the present application also provides a satellite geometric configuration change monitoring device, electronic equipment and corresponding embodiments.

[0113] Figure 4 It is a structural schematic diagram of a satellite geometric configuration change monitoring device shown in an embodiment of the present application.

[0114] See also Figure 4 The present application provides a device for monitoring changes in satellite geometric configuration, which may include: The target observation data acquisition module 410 is used to obtain target observation data collected by the GNSS receiver and input the target observation data into a pre-trained bidirectional LSTM network model. The bidirectional LSTM network model is jointly trained using a multi-task loss function, which includes an environment classification loss function and a geometric quality regression loss function. A key feature calculation module 420 is used to calculate key features of satellite geometric configuration changes based on target observation data using a bidirectional LSTM network model; Satellite geometry change monitoring module 430 is used to capture the front-end dependency of the time series composed of key features through a bidirectional LSTM network model to monitor satellite geometry changes; A multi-task prediction module 440 is configured to predict the GNSS receiver's environment and the degradation of the satellite geometry based on changes in the satellite geometry using a bidirectional LSTM network model; The data processing module 450 is used to perform positioning using a data processing method that matches the environment in which the GNSS receiver is located and the degradation degree of the satellite geometric configuration.

[0115] In one embodiment, the target observation data includes at least one of pseudorange, carrier phase, Doppler shift, C / N0, satellite elevation angle, satellite azimuth angle, satellite system type, frequency band identifier, and health status; the key feature includes at least one of satellite geometric features and satellite signal quality features; the satellite geometric feature includes at least one of satellite azimuth angle change rate and satellite elevation angle entropy; and the satellite signal quality feature includes MPI; the key feature calculation module 420 may include: A first calculation submodule is configured to calculate a satellite azimuth angle change rate based on the satellite azimuth angle, carrier phase, and Doppler shift using a bidirectional LSTM network model; and / or A second calculation submodule is configured to calculate the satellite elevation angle entropy value based on the satellite elevation angle, C / N0, satellite system type, and frequency band identifier through a bidirectional LSTM network model; and / or, The third calculation submodule is used to calculate the MPI based on the pseudorange and satellite signal frequency through a bidirectional LSTM network model.

[0116] In one embodiment, the bidirectional LSTM network model includes a satellite system embedding layer and a bidirectional LSTM core layer, wherein the bidirectional LSTM core layer includes a forward LSTM layer, a reverse LSTM layer, and an LSTM feature fusion layer; the satellite geometry change monitoring module 430 may include: A mapping submodule, for mapping identifiers of different satellite systems to embedding vectors of fixed dimension through a satellite system embedding layer; The splicing submodule is used to splice the embedding vector and the key features at the time step through the satellite system embedding layer to obtain the input feature vector; The first capture submodule is used to capture the impact of past time steps on the current moment based on the forward time series in the input feature vector through a forward LSTM to obtain a forward expression vector of the key features; and The second capture submodule is used to capture the impact of future time steps on the current moment based on the reverse time series in the input feature vector through a reverse LSTM, and obtain a reverse expression vector of the key features; wherein the forward expression vector is used to represent the feature representation formed by the comprehensive impact of the forward time series on the current moment, and the reverse expression vector is used to represent the feature representation formed by the comprehensive impact of the reverse time series on the current moment. The forward time series includes key features sorted by forward time, and the reverse time series includes key features sorted by reverse time. The fusion submodule is used to fuse the forward expression vector and the reverse expression vector through the LSTM feature fusion layer to obtain the front-end dependency relationship of the time series to monitor the changes in the satellite geometric configuration; wherein the time series is a forward time series or a reverse time series.

[0117] In one embodiment, the first capture submodule may include: The positive change law capturing unit is used to capture the change law of the key features along the positive time sequence based on the key features from the start time to the end time in the input feature vector through the positive LSTM, and obtain the positive expression vector of the key features.

[0118] In one embodiment, the second capturing submodule may include: The reverse change law capturing unit is used to capture the change law of the key features along the reverse time sequence based on the key features from the end time to the start time in the input feature vector through the reverse LSTM, and obtain the reverse expression vector of the key features.

[0119] In one embodiment, the bidirectional LSTM network model further includes a self-attention mechanism layer; the satellite geometry change monitoring module 430 may further include: The similarity calculation submodule is used to calculate the similarity between key features at different time steps in the input feature vector through the self-attention mechanism layer; The weighted processing submodule is used to perform weighted processing on the front-end dependency of the time series based on similarity through the self-attention mechanism layer to obtain the target attention output.

[0120] In one embodiment, the bidirectional LSTM network model can be trained by the following modules: The historical observation data acquisition module is used to obtain multiple frames of historical observation data collected by the GNSS receiver; The environment classification module is used to classify the multi-frame historical observation data into environmental categories to obtain historical observation data sets of different environmental categories; different environmental categories include simple environmental categories and complex environmental categories; A first training module is configured to extract static data from a historical observation dataset of a simple environment category in a first training phase, so as to train the bidirectional LSTM network model using the static data until a target loss value of the bidirectional LSTM network model in the first training phase is less than a preset loss value; A second training module is configured to extract dynamic data from a historical observation dataset of a complex environment category in a second training phase, so as to train the bidirectional LSTM network model using the dynamic data until a target loss value of the bidirectional LSTM network model in the second training phase is less than a preset loss value; A third training module is configured to train the bidirectional LSTM network model using full-scene mixed data in a third training phase until a target loss value of the bidirectional LSTM network model in the third training phase is less than a preset loss value, thereby obtaining a trained bidirectional LSTM network model; wherein the full-scene mixed data is a historical observation dataset of different environmental categories; The weights of the environment classification loss function in the first training stage, the second training stage, and the third training stage show a decreasing law; The weights of the geometric quality regression loss function in the first training stage, the second training stage, and the third training stage are increasing.

[0121] In one embodiment, the device may further include: The online learning module is used to trigger the online learning mechanism when the satellite geometric configuration change is greater than a preset change threshold or the prediction error of the bidirectional LSTM network model is greater than a preset error threshold, so as to incrementally update the parameters of the bidirectional LSTM network model and adjust the weight of the environmental classification loss function or the weight of the geometric quality regression loss function.

[0122] As can be seen from this example, the solution provided by this application obtains the target observation data collected by the GNSS receiver and inputs the target observation data into a pre-trained bidirectional LSTM network model; wherein, the bidirectional LSTM network model is obtained by joint training using a multi-task loss function, and the multi-task loss function includes an environmental classification loss function and a geometric quality regression loss function; the bidirectional LSTM network model is used to calculate the key features of the satellite geometric configuration changes based on the target observation data; the bidirectional LSTM network model is used to capture the front-end and back-end dependencies of the time series composed of key features to monitor the satellite geometric configuration changes; the bidirectional LSTM network model is used to predict the environment of the GNSS receiver and the degradation degree of the satellite geometric configuration based on the satellite geometric configuration changes; and positioning is performed using a data processing method that matches the environment of the GNSS receiver and the degradation degree of the satellite geometric configuration. This application uses a bidirectional LSTM network model to capture the forward and backward dependencies of the time series composed of key features of satellite geometric configuration changes, which can improve the accuracy and real-time performance of satellite geometric configuration change monitoring. Moreover, by jointly training the bidirectional LSTM network model with multiple tasks, the comprehensive performance and generalization ability of the bidirectional LSTM network model can be improved, so that the bidirectional LSTM network model can achieve good prediction results in different scenarios. Therefore, based on the monitored satellite geometric configuration changes, the bidirectional LSTM network model can accurately and quickly predict the environment in which the GNSS receiver is located and the degradation degree of the satellite geometric configuration, thereby breaking the limitations of related technologies in dealing with complex dynamic environments and multipath effects, and enhancing the positioning accuracy and reliability of the GNSS system in complex environments.

[0123] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated again here.

[0124] Figure 5 It is a structural diagram of an electronic device shown in an embodiment of the present application.

[0125] See also Figure 5 , the electronic device 500 includes a memory 510 and a processor 520.

[0126] The processor 520 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. Memory 510 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage. ROM may store static data or instructions required by processor 520 or other computer modules. Permanent storage may be a readable and writable storage device. Permanent storage may be a non-volatile storage device that retains stored instructions and data even when the computer is powered off. In some embodiments, the permanent storage device utilizes a mass storage device (e.g., a magnetic or optical disk, flash memory). In other embodiments, the permanent storage device may be a removable storage device (e.g., a floppy disk, optical drive). System memory may be a readable and writable storage device or a volatile readable and writable storage device, such as dynamic random access memory (DRAM). System memory may store some or all instructions and data required by the processor during operation. Furthermore, memory 510 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (e.g., DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), as well as magnetic disks and / or optical disks. In some embodiments, the memory 510 may include a readable and / or writable removable storage device, such as a compact disc (CD), a read-only digital versatile disc (e.g., DVD-ROM, double-layer DVD-ROM), a read-only Blu-ray disc, an ultra-density optical disc, a flash memory card (e.g., SD card, mini SD card, Micro-SD card, etc.), a magnetic floppy disk, etc. Computer-readable storage media do not include carrier waves and transient electronic signals transmitted wirelessly or wired.

[0127] The memory 510 stores executable codes. When the executable codes are processed by the processor 520 , the processor 520 may execute part or all of the above-mentioned methods.

[0128] In addition, the method according to the present application may also be implemented as a computer program or a computer program product, which includes computer program code instructions for executing some or all of the steps in the above method of the present application.

[0129] Alternatively, the present application can also be implemented as a computer-readable storage medium (or non-transitory machine-readable storage medium or machine-readable storage medium), which stores executable code (or computer program or computer instruction code) and, when executed by a processor of an electronic device (or server, etc.), enables the processor to perform part or all of the steps of the above-mentioned method according to the present application.

[0130] The present application also provides a computer program product, which includes computer instructions, and when the computer instructions are executed by a processor, the method described above is implemented.

[0131] The embodiments of the present application have been described above. The above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or improvements to the technology in the market, or to enable other persons skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for monitoring changes in satellite geometric configuration, characterized in that: The method comprises: Obtaining target observation data collected by a GNSS receiver and inputting the target observation data into a pre-trained bidirectional LSTM network model; wherein the bidirectional LSTM network model is jointly trained using a multi-task loss function, wherein the multi-task loss function includes an environment classification loss function and a geometric quality regression loss function; Calculating key features of satellite geometric configuration changes based on the target observation data through the bidirectional LSTM network model; The bidirectional LSTM network model is used to capture the front-back dependency of the time series composed of the key features to monitor the changes in the satellite geometric configuration; Predicting the environment of the GNSS receiver and the degradation degree of the satellite geometry based on the change of the satellite geometry by using the bidirectional LSTM network model; Positioning is performed using a data processing method that matches the environment in which the GNSS receiver is located and the degree of degradation of the satellite geometry.

2. The method according to claim 1, characterized in that The target observation data includes at least one of pseudorange, carrier phase, Doppler shift, C / N0, satellite elevation angle, satellite azimuth angle, satellite system type, frequency band identifier and health status; the key feature includes at least one of satellite geometric features and satellite signal quality features; the satellite geometric features include at least one of satellite azimuth angle change rate and satellite elevation angle entropy; the satellite signal quality features include MPI; The step of calculating key features of satellite geometric configuration changes based on the target observation data using the bidirectional LSTM network model includes: Calculating the satellite azimuth angle change rate based on the satellite azimuth angle, the carrier phase, and the Doppler shift through the bidirectional LSTM network model; and / or, Calculating the satellite elevation angle entropy value based on the satellite elevation angle, the C / N0, the satellite system type, and the frequency band identifier through the bidirectional LSTM network model; and / or, The MPI is calculated based on the pseudorange and satellite signal frequency through the bidirectional LSTM network model.

3. The method according to claim 1, characterized in that The bidirectional LSTM network model includes a satellite system embedding layer and a bidirectional LSTM core layer, wherein the bidirectional LSTM core layer includes a forward LSTM layer, a reverse LSTM layer, and an LSTM feature fusion layer. The bidirectional LSTM network model is used to capture the front-end dependency of the time series composed of the key features to monitor the changes in the satellite geometric configuration, including: Mapping identifiers of different satellite systems into embedding vectors of fixed dimension through the satellite system embedding layer; Concatenating the embedding vector and the key feature at a time step through the satellite system embedding layer to obtain an input feature vector; The forward LSTM is used to capture the influence of past time steps on the current moment based on the forward time series in the input feature vector, thereby obtaining a forward expression vector of the key feature; and the reverse LSTM is used to capture the influence of future time steps on the current moment based on the reverse time series in the input feature vector, thereby obtaining a reverse expression vector of the key feature; wherein the forward expression vector is used to characterize the feature representation formed by the comprehensive influence of the forward time series on the current moment, and the reverse expression vector is used to characterize the feature representation formed by the comprehensive influence of the reverse time series on the current moment, the forward time series includes the key features sorted in forward time, and the reverse time series includes the key features sorted in reverse time; The forward expression vector and the reverse expression vector are fused through the LSTM feature fusion layer to obtain the front-to-back dependency relationship of the time series to monitor the changes in the satellite geometric configuration; wherein the time series is the forward time series or the reverse time series.

4. The method according to claim 3, characterized in that The method of capturing the influence of past time steps on the current moment based on the forward time series in the input feature vector by the forward LSTM to obtain the forward expression vector of the key feature includes: Capturing the changing pattern of the key features along the forward time sequence based on the key features from the start time to the end time in the input feature vector through the forward LSTM to obtain a forward expression vector of the key features; The reverse LSTM is used to capture the impact of future time steps on the current moment based on the reverse time series in the input feature vector to obtain the reverse expression vector of the key feature, including: The reverse LSTM is used to capture the change pattern of the key feature in reverse time sequence based on the key feature from the end time to the start time in the input feature vector, and obtain the reverse expression vector of the key feature.

5. The method according to claim 3, characterized in that The bidirectional LSTM network model also includes a self-attention mechanism layer; the method further includes: Calculating the similarity between the key features at different time steps in the input feature vector through the self-attention mechanism layer; The self-attention mechanism layer performs weighted processing on the front-end dependency of the time series based on the similarity to obtain the target attention output.

6. The method according to claim 1, characterized in that The bidirectional LSTM network model is trained in the following way: Acquiring multiple frames of historical observation data collected by the GNSS receiver; Dividing the multiple frames of historical observation data into environmental categories to obtain historical observation data sets of different environmental categories; the different environmental categories include simple environmental categories and complex environmental categories; In a first training phase, extracting static data from the historical observation dataset of the simple environment category to train the bidirectional LSTM network model using the static data until a target loss value of the bidirectional LSTM network model in the first training phase is less than a preset loss value; In a second training phase, extracting dynamic data from the historical observation dataset of the complex environment category to train the bidirectional LSTM network model using the dynamic data until a target loss value of the bidirectional LSTM network model in the second training phase is less than a preset loss value; In the third training phase, the bidirectional LSTM network model is trained using full-scene mixed data until the target loss value of the bidirectional LSTM network model in the third training phase is less than the preset loss value, thereby obtaining a trained bidirectional LSTM network model; wherein the full-scene mixed data is a historical observation dataset of the different environment categories; The weights of the environment classification loss function in the first training stage, the second training stage, and the third training stage are in a decreasing order; The weights of the geometric quality regression loss function in the first training stage, the second training stage, and the third training stage are respectively increasing.

7. The method according to claim 6, characterized in that The method further comprises: When the change in the satellite geometric configuration is greater than a preset change threshold or the prediction error of the bidirectional LSTM network model is greater than a preset error threshold, the online learning mechanism is triggered to incrementally update the parameters of the bidirectional LSTM network model and adjust the weight of the environmental classification loss function or the weight of the geometric quality regression loss function.

8. A device for monitoring changes in satellite geometric configuration, characterized in that: The device comprises: A target observation data acquisition module is used to obtain target observation data collected by a GNSS receiver and input the target observation data into a pre-trained bidirectional LSTM network model; wherein the bidirectional LSTM network model is jointly trained using a multi-task loss function, wherein the multi-task loss function includes an environmental classification loss function and a geometric quality regression loss function; A key feature calculation module, configured to calculate key features of satellite geometric configuration changes based on the target observation data using the bidirectional LSTM network model; A satellite geometry change monitoring module is used to capture the front-end dependency of the time series composed of the key features through the bidirectional LSTM network model to monitor the satellite geometry change; a multi-task prediction module, configured to predict the environment of the GNSS receiver and the degradation degree of the satellite geometry based on the change of the satellite geometry using the bidirectional LSTM network model; The data processing module is used to perform positioning using a data processing method that matches the environment in which the GNSS receiver is located and the degradation degree of the satellite geometric configuration.

9. An electronic device, characterized in that: include: processor; as well as A memory having executable codes stored thereon, which, when executed by the processor, causes the processor to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having executable codes stored thereon, wherein when the executable codes are executed by a processor of an electronic device, the processor is caused to execute the method according to any one of claims 1 to 7.

Images