A differential GNSS scene recognition method and system based on ConvLSTM

Abstract

The present invention provides a differential GNSS scene recognition method and system based on ConvLSTM, including: collecting multi-constellation GNSS observation values ​​in common urban navigation scenarios, and extracting multiple scene features; performing differential processing on scene features, including differences between epochs and differences between mobile stations and base stations, and converting the differential features into multi-channel feature images; firstly, using the differential feature images between epochs to train a binary ConvLSTM neural network, constructing a scene mutation recognition model, which can identify when a scene mutation occurs, and divide the data between adjacent mutation points into segments in turn; secondly, using the differential feature images between mobile stations and base stations to train a multi-classification ConvLSTM neural network, constructing a scene recognition model, which can identify specific scene categories. After inputting the GNSS data of the scene to be identified, the model can divide the data set into independent scene time periods after identifying the scene mutation points, and finally identify the scene categories of each time period.

Description

Technical Field

[0001] The present invention relates to the technical fields of scene recognition, satellite navigation and high-precision positioning, and in particular to a differential GNSS scene recognition method and system based on ConvLSTM. Background Art

[0002] In complex urban environments, high-precision positioning is still one of the difficulties that GNSS needs to overcome. Due to the attenuation, reflection and shielding of GNSS signals by urban buildings, the availability of satellites is significantly reduced. Considering that different scenes have significantly different interference characteristics to GNSS signals, scene recognition technology can be used to deeply analyze the characteristics, optimize GNSS terminal algorithms, and improve the accuracy and reliability of navigation.

[0003] At present, scene recognition mainly adopts image recognition and GNSS signal discrimination. Image recognition relies on low-level features, which makes it difficult to fully describe the environment. It is also sensitive to image quality and has limited recognition rate. GNSS signal discrimination selects a variety of suitable GNSS feature parameters for machine learning recognition. However, the signal is easily affected by common mode noise and errors, and urban scenes are complex and changeable. When scene recognition is performed, the boundaries are often blurred at the scene switching point, making it difficult to accurately identify. There are problems such as weak anti-interference ability, insufficient generalization ability, low recognition accuracy, and high computing cost.

[0004] Therefore, it is urgent to provide a method that can model and perform scene recognition using only GNSS observation data. This method should have high scene recognition accuracy, strong generalization ability, and less computational complexity, and can solve the problem of fuzzy recognition at scene switching. Summary of the invention

[0005] Purpose of the invention: In view of the above-mentioned defects of the prior art, the present invention provides a differential GNSS scene recognition method based on ConvLSTM. First, multi-constellation GNSS observations in common urban navigation scenarios are collected, and a variety of scene features are extracted; the scene features are differentially processed, including the difference between epochs and the difference between the mobile station and the base station, and the features are converted into multi-channel feature images; the differential feature images between epochs are used to train a binary classification ConvLSTM neural network, and a scene mutation recognition model is constructed, which can identify when the scene mutates, and divide the data between adjacent mutation points into segments in turn; secondly, the differential feature images between the mobile station and the base station are used to train a multi-classification ConvLSTM neural network, and a scene recognition model is constructed, which can identify the specific category of the scene. After inputting the GNSS data of the scene to be identified, the model can divide the data set into independent scene time periods after identifying the scene mutation points, and finally identify the scene category of each time period. The present invention utilizes differentials not only to reduce common-mode noise and errors, but also the inter-epoch differentials can highlight and enhance the signal change characteristics when the scene switches, and improve the problem of recognition ambiguity before and after the scene switch; the inter-station differentials can reflect the change rules within the scene, and use ConvLSTM to deeply explore the hidden spatiotemporal correlation characteristics of features, thereby improving the accuracy and continuity of scene recognition, and providing effective support for navigation and high-precision positioning in complex urban environments.

[0006] To achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a differential GNSS scene recognition method based on ConvLSTM, comprising the following steps:

[0007] Using the GNSS raw observation values ​​of multiple satellites at multiple identical epochs of the mobile station and the base station under n types of preset scenarios, generating a scene mutation identification feature map and a scene recognition feature map about the feature data, and using the scene mutation identification feature map to train a binary ConvLSTM neural network to obtain a scene mutation identification model for identifying the mutation epoch point of the GNSS raw observation value under the scenario to be identified;

[0008] Using the scene recognition feature map to train a multi-classification ConvLSTM neural network to obtain a scene recognition model for identifying the scene corresponding to the GNSS raw observation value between adjacent mutation epoch points in the scene to be recognized;

[0009] The characteristic data include: satellite elevation angle El and satellite azimuth angle Az and at least one of pseudorange P, Doppler frequency shift d, carrier-to-noise ratio SNR, signal loss lock indicator LLI, and visible satellite number SatNum.

[0010] The n preset scenes include open space scene, dense tree shade scene, single-sided building scene and double-sided building scene.

[0011] Further, generating a scene mutation recognition feature map and a scene recognition feature map about the feature data includes the following steps:

[0012] A1. Project the satellite position onto a circular image plane to obtain a circular satellite position map, project the circular satellite position map onto a circumscribed square, fill the area outside the circular satellite position map in the square with a grayscale value of 0, and divide the circular satellite position map in the square into Thiessen polygons based on the discrete points using the satellite positions in the circular satellite position map as discrete points according to the Thiessen polygon algorithm;

[0013] A2. Perform inter-epoch differential processing on various residual characteristic data of the mobile station, and perform inter-station differential processing on various residual characteristic data of the same satellite observed by the mobile station and the base station; perform data cleaning and normalization on the data after inter-epoch differential processing and inter-station differential processing to obtain inter-epoch differential data and inter-station differential data of each satellite; various residual characteristic data are characteristic data other than satellite elevation angle El and satellite azimuth angle Az;

[0014] A3. Then, the inter-epoch differential data and inter-station differential data of each satellite are interpolated into pixel values, and the pixel values ​​corresponding to the satellite positions are filled in the Thiessen polygons to which the discrete points corresponding to the satellite positions belong, so as to obtain the scene mutation recognition feature map and the scene recognition feature map.

[0015] Further, using the scene mutation recognition feature graph to train a binary classification ConvLSTM neural network includes the following steps: marking a mutation label according to whether a scene change occurs in the scene mutation recognition feature graph, 0 indicates that the scene has not changed, and 1 indicates that the scene has mutated, arranging the marked scene mutation recognition feature graph into a time series according to time to obtain a scene mutation recognition data set; dividing the scene mutation recognition data set into a training set and a validation set according to a preset ratio; using the scene mutation recognition data set to train a binary classification ConvLSTM neural network to construct a scene mutation recognition model;

[0016] Using scene recognition feature maps to train a multi-classification ConvLSTM neural network includes the following steps: marking scene labels according to the scenes corresponding to the scene recognition feature maps, arranging the marked scene recognition feature maps into a time series by time, and obtaining a scene recognition data set; dividing the scene recognition data set into a training set and a validation set according to a preset ratio; using the scene mutation recognition data set to train a multi-classification ConvLSTM neural network and construct a scene mutation recognition model.

[0017] Further, identifying the mutation epoch points of the GNSS raw observation values ​​of the mobile station to multiple satellites at multiple epochs in the scene to be identified includes:

[0018] Obtaining GNSS raw observation values ​​of multiple satellites at multiple epochs from the mobile station and the base station in the scene to be identified and processing them according to steps A1 to A3 to obtain a scene mutation identification feature map to be identified and a scene mutation identification feature map to be identified, inputting the scene mutation identification feature map to be identified into the scene mutation identification model to obtain a scene mutation epoch point;

[0019] The data of the scene recognition feature graph to be recognized between adjacent mutation epoch points are divided into segments in turn to obtain data of several independent scene time periods.

[0020] Further, identifying the scenes corresponding to the GNSS raw observation values ​​of the mobile station and the base station on multiple satellites at multiple epochs in the scene to be identified between adjacent mutation epoch points includes: inputting the independent scene period data into the scene recognition model to obtain a scene recognition result.

[0021] Furthermore, A2 specifically includes:

[0022] A21. Define a matrix set:

[0023] The GNSS raw observation values ​​of multiple satellites collected by the mobile station under n preset scenarios at multiple epochs are converted into a mobile station feature matrix set:

[0024] M rover ={M rover (P), M rover (d), M rover (SNR), M rover (LLI),

[0025] M rover (SatNum)};

[0026] Each feature matrix is ​​an m*n matrix, where m is the number of satellites, n is the number of epochs, and the element M of the matrix rover (P) ij represents the pseudorange value of the mobile station's i-th satellite at the j-th epoch, M rover (d) ij represents the Doppler shift of the i-th satellite at the j-th epoch, M rover (SNR) ij Indicates the carrier-to-noise ratio of the i-th satellite at the j-th epoch of the mobile station, and the signal loss indicator M rover (LLI) ij Indicates the signal loss flag of the mobile station's i-th satellite at the j-th epoch, M rover (SatNum) j represents the number of visible satellites observed by the mobile station at the jth epoch;

[0027] The original GNSS observation values ​​of the base station on multiple satellites at multiple epochs collected under n types of preset scenarios are converted into a set of base station feature matrices:

[0028] M base ={M base (P), M base (d), M base (SNR), M base (LLI),

[0029] M base (SatNum)};

[0030] Each feature matrix is ​​an m*n matrix, where m is the number of satellites, n is the number of epochs, and the element M of the matrix base (P) ij represents the pseudorange value of the base station observing the i-th satellite at the j-th epoch, M base (d) ij It indicates the Doppler shift of the ith satellite observed by the base station at the jth epoch, M base (SNR) ij M represents the carrier-to-noise ratio of the base station observing the i-th satellite at the j-th epoch. base (LLI) ij Indicates the signal loss flag of the base station observing the i-th satellite at the j-th epoch, M base (SatNum) j represents the number of visible satellites observed by the base station at the jth epoch;

[0031] A22. Perform inter-epoch difference processing on the observation data of the mobile station, that is, calculate the difference between the scene characteristic values ​​of the mobile station observing the same satellite between adjacent epochs:

[0032] To M rover (P) Perform the above-mentioned differential processing to obtain D epoch (P), where D epoch (P) ij For: D epoch (P) ij =M rover (P) (i，j+1) -M rover (P) (i，j) (j=1, 2, ..., n-1);

[0033] To M rover (d) Perform the above-mentioned differential processing to obtain D epoch (d), where D epoch (d) ij For: D epoch (d) ij =M rover (d)(i，j+1) -M rover (d) (i，j) (j=1, 2, ..., n-1);

[0034] To M rover (SNR) is subjected to the above differential processing to obtain D epoch (SNR), where D epoch (SNR) ij For: D epoch (SNR) ij =M rover (SNR) (i，j+1) -M rover (SNR) (i，j) (j=1, 2, ..., n-1);

[0035] To M rover (LLI) performs the above-mentioned differential processing to obtain D epoch (LLI), where D epoch (LLI) ij For: D epoch (LLI) ij =M rover (LLI) (i，j+1) -M rover (LLI) (i，j) (j=1, 2, ..., n-1);

[0036] To M rover (SatNum) performs the above-mentioned differential processing to obtain D epoch (SatNum), where D epoch (SatNum) j For: D epoch (SatNum) j =M rover (SNR) (j+1) -M rover (SatNum) (j) (j=1,2,...,n-1); get the set:

[0037] D epoch {D epoch (P), D epoch (D),D epoch (SNR), D epoch (LLI), D epoch (SatNum)};

[0038] A23. Perform inter-station differential processing on various characteristic data of the same satellite observed by the mobile station and the base station, that is, calculate the difference between the observation values ​​of the same satellite observed by the mobile station and the base station at the same time:

[0039] To M rover (P) and M base (P) Perform the above-mentioned differential processing to obtain D rb (P), where D rb (P) ij For: D rb (P) ij =M rover (P) (i，j) -M base (P) (i，j) ;

[0040] To M rover (d) and M base (d) Perform the above-mentioned differential processing to obtain D rb (d), where D rb (d) ij For: D rb (d) ij =M rover (d) (i，j) -M base (d) (i，j) ;

[0041] To M rover (SNR) and M base (SNR) is subjected to the above differential processing to obtain D rb (SNR), where D rb (SNR) ij For: D rb (SNR) ij =M rover (SNR) (i，j) -M base (SNR) (i，j) ;

[0042] To M rover (LLI) and M base (LLI) performs the above-mentioned differential processing to obtain D rb (LLI), where D rb (LLI) ij For: D rb (LLI) ij =M rover (LLI) (i，j) -M base (LLI) (i，j) ;

[0043] To M rover (SatNum) and M base (SatNum) performs the above-mentioned differential processing to obtain D rb(SatNum), where D rb (SatNum) ij For: D rb (SatNum) j =M rover (SatNum) (j) -M base (SatNum) (j) ; Get the collection: D rb {D rb (P), D rb (D),D rb (SNR), D rb (LLI), D rb (SatNum)};

[0044] A24 normalization processing:

[0045] The normalization process is mathematically expressed as follows:

[0046]

[0047] Where max(D) and min(D) refer to the maximum and minimum values ​​of the observations after inter-epoch difference processing or inter-station difference processing, respectively.

[0048] Furthermore, the satellites involved in the difference include observed satellites and unobserved satellites in the system. For the unobserved satellites, the mean of the observed values ​​of the open scene is taken as the background value.

[0049] Furthermore, the scene mutation identification feature map includes an inter-epoch differential pseudorange map, an inter-epoch differential Doppler frequency shift map, an inter-epoch differential carrier-to-noise ratio map, an inter-epoch differential signal loss identification map, and an inter-epoch differential visible satellite number map for identifying scene mutations;

[0050] The scene recognition feature map includes the inter-station differential pseudo-range map, the inter-station differential Doppler frequency shift map, the inter-station differential carrier-to-noise ratio map, the inter-station differential signal loss identification map and the inter-station differential visible satellite number map for identifying the scene category;

[0051] The satellites corresponding to the discrete points include observed satellites and unobserved satellites in the system. When setting the pixel value of the Thiessen polygon, the mean of the observed values ​​of the open scene is taken as the background value for the Thiessen polygon image part where the satellite is not observed.

[0052] Furthermore, the binary classification ConvLSTM neural network and the multi-classification ConvLSTM neural network include an input layer, a convolutional layer, a plurality of stacked ConvLSTM layers, a fully connected layer, and an output layer connected in sequence.

[0053] Furthermore, the ConvLSTM neural network is trained through the data set to obtain the scene mutation recognition model and the scene recognition model, which specifically includes:

[0054] The hyperparameters of the model are configured, cross entropy is used as the loss function, the gradient is updated with the Adam optimizer, and the scene mutation recognition model and the scene recognition model are obtained after the training is performed until the preset iteration termination condition is met;

[0055] The preset iteration termination condition is: the training accuracy and the verification accuracy both reach the corresponding preset accuracy standards, and the difference between the training accuracy and the verification accuracy is within a preset range.

[0056] A ConvLSTM-based differential GNSS scene recognition system comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the ConvLSTM-based differential GNSS scene recognition method are implemented.

[0057] Beneficial effects of the present application: The present invention provides a differential GNSS scene recognition method based on ConvLSTM. Through the technical solution of the method, the original GNSS observations of each scene are first collected and the scene features are extracted, and the inter-epoch and inter-station differential processing is performed and converted into a multi-channel image to form a data set, and a scene mutation recognition model and a scene recognition model based on ConvLSTM are constructed; when the original GNSS observations of an unknown scene are input, the inter-epoch difference data set is input into the scene mutation recognition model, which can identify the moment of scene mutation in the time series, and thereby divide the collected original observation data into independent scene time periods, and the inter-station difference data set is input into the scene recognition model, and finally the scene category corresponding to each time period is identified. Compared with the prior art, the urban road scene recognition method based on ConvLSTM GNSS differential:

[0058] (1) Improved recognition accuracy of scene switching and enhanced signal characteristics in different scenes:

[0059] First, the epoch difference can be used to identify the sudden change moment of scene switching. Urban road scenes are complex and changeable. When the scene changes significantly, the epoch difference may change significantly. By monitoring the changes in the epoch difference, the sudden change point can be highlighted, making it easier for the model to capture the key information of scene switching, thereby improving the accuracy of recognition.

[0060] Second, before and after the mutation moment, the difference between the base station and the mobile station can be used to further analyze the change pattern in the scene. The base station can observe more satellite information and has a fixed position, which can be used as a reference point. By comparing the observation differences between the base station and the mobile station, the changes in the scene where the mobile station is located can be inferred. Each scene is blocked to different degrees, and the reflected signal or signal attenuation is received to different degrees. During the movement of the mobile station, the difference in GNSS signal characteristics changes with the scene. By analyzing the change pattern of the difference, it can help the model further extract the dynamic characteristics of the scene and adapt to the characteristics related to the scene.

[0061] (2) Mining of spatiotemporal correlation characteristics and change patterns:

[0062] In complex and ever-changing scenes, the performance of different features is often different. For example, in a scene blocked by buildings, GNSS signals are usually reflected to produce a multipath effect; in a scene with dense shade, satellite signals are not only blocked and affected by multipath signals, but also show some flickering characteristics. These features are particularly significant within a specific time window. The present invention introduces the ConvLSTM neural network, which learns the long-term dependencies in the spatiotemporal sequence through the memory function of the LSTM unit, can accurately identify the scene mutation points, and deeply understand the law of changes in the features in the scene over time. ConvLSTM can also further explore the spatial distribution characteristics based on satellite position information, so as to more accurately identify the key points and scene categories in the scene, and improve the accuracy of recognition.

[0063] (3) Low hardware resource consumption: No need to rely on additional sensor data, the current scene category can be accurately identified based on GNSS observations alone. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A method flow chart of a preferred embodiment of the present invention;

[0065] Figure 2 A schematic diagram of a preset scene of a preferred embodiment of the present invention;

[0066] Figure 3 A sky map of each scene in a preferred embodiment of the present invention;

[0067] Figure 4 A pseudorange matrix of a preferred embodiment of the present invention;

[0068] Figure 5 This is an example of the drawing process of Thiessen polygons;

[0069] Figure 6 is a pixel value interpolated after the Doppler frequency shift is differentiated and normalized in a preset scenario in a preferred embodiment of the present invention; the horizontal axis corresponds to the satellite number, and the vertical axis corresponds to different epochs;

[0070] Figure 7 Schematic diagram of the ConvLSTM network of the present invention. DETAILED DESCRIPTION

[0071] The technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings and specific examples. The embodiments described below are only a part of the embodiments of the present invention, rather than all of the embodiments, and are only used to illustrate the present invention, but are not used to limit the scope of the present invention. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0072] The BDS, GPS, GLONASS and GALILEO recorded in the following embodiments are all global satellite navigation systems, BDS is the Beidou Satellite Navigation System, GPS is the Global Positioning System established and operated by the United States, GLONASS is the Global Satellite Navigation System established and operated by Russia, and GALILEO is the Galileo Satellite Navigation System. The L1 frequency is approximately 1575.42 MHz. The L1 signal is used by the GPS, GLONASS and GALILEO systems. The wavelength of the L1 signal is longer, so it is easily affected by ionospheric delay, but it is the main signal for implementing basic navigation services. The frequency of the L2 signal is approximately 1227.60 MHz, which generally provides higher positioning accuracy than L1 because it has a shorter wavelength and is less sensitive to ionospheric delay. The L2 signal is used by the GPS and GALILEO systems and is generally used for civil aviation and high-precision positioning services. The frequency of the L5 signal is approximately 1176.45 MHz, which is a new frequency band mainly used for civil aviation and safety-critical applications. The wavelength of the L5 signal is longer than L2 but shorter than L1, so it provides better performance in terms of ionospheric delay. The L5 signal is used by the GPS and GALILEO systems, especially the GALILEO system, which provides two L5 signals: E5a and E5b.

[0073] This embodiment relates to a differential GNSS scene recognition method based on ConvLSTM, which can identify scenes where GNSS signals on urban roads are blocked to varying degrees. This embodiment selects four types of scenes: open space, dense shade, single-sided building, and double-sided building. The scene diagram is shown in the figure. Figure 2 As shown in (a), (b), (c), and (d), where:

[0074] Figure 2 (a) shows an open scene: an open area with no obstacles or only a few obstacles blocking the signal around the GNSS receiver. In this scenario, the GNSS receiver receives all direct signals, the received satellites are well distributed, and are less affected by multipath.

[0075] Figure 2 (b) shows a dense tree shade scenario: this refers to an area where there are no buildings on either side of the GNSS receiver, but there are dense trees above that block the signal. In this scenario, the GNSS receiver receives some direct and reflected signals, and some observations may show irregular flickering characteristics within a certain period of time.

[0076] Figure 2 (c) shows a one-sided building scenario: This refers to a positioning scenario where a GNSS receiver is in an urban canyon, city block, or other high-rise building-dense area, where only one side has buildings larger than six stories, while the other side has no or low buildings. In this case, the GNSS receiver receives low-angle direct signals, building-reflected signals, and ground-reflected signals. Signal reflection or refraction leads to NLOS propagation, which is greatly affected by multipath.

[0077] Figure 2 (d) shows a double-sided building scenario: This refers to a positioning scenario where a GNSS receiver is located in an urban canyon, a city block, or other high-rise building-dense area, with buildings higher than six stories high on both sides. In this case, it is difficult for the GNSS receiver to receive direct signals, and most of the signals received are reflected signals from the buildings. The NLOS propagation caused by the reflection or refraction of the signal is more serious, and the multipath effect is serious.

[0078] like Figure 1 As shown, the specific method steps of a preferred embodiment of the present invention include:

[0079] Model training steps: Perform two differential processing on the collected known scene data, and build a scene mutation recognition model and a scene recognition model by training the ConvLSTM network. The specific steps are as follows:

[0080] 1.1) Use the GNSS receivers of the mobile station and the base station to collect GNSS raw observation values ​​for multiple common-view satellites at multiple identical epochs in n types of preset scenarios. The n types of preset scenarios include: open space, dense shade, single-sided building and double-sided building scenes. In this embodiment, the GNSS receiver is compatible with the frequency points (L1, L2, L5) of multiple systems (BDS, GPS, GLONASS and GALILEO) for data collection, and the sampling interval is 1s. In order to ensure the training effect of the subsequent scene recognition model, the amount of data collected in each type of scene is required to be sufficient. In this embodiment, continuous sampling is performed every 10 minutes in a type of preset scene, and a large amount of multi-constellation GNSS data with scene labels is obtained through multi-day collection, and the total time exceeds 600 minutes.

[0081] 1.2) Six types of feature data with strong correlation with the scene are extracted from the GNSS original observation data, including: satellite position (satellite elevation angle El and satellite azimuth angle Az), pseudorange P, Doppler frequency shift d, carrier-to-noise ratio SNR, signal loss lock indicator LLI, and number of visible satellites SatNum. Satellite position (satellite elevation angle El and satellite azimuth angle Az) and at least one of pseudorange P, Doppler frequency shift d, carrier-to-noise ratio SNR, signal loss lock indicator LLI, and number of visible satellites SatNum can also be extracted.

[0082] The satellite position is determined by El and Az. The elevation angle is the angle of the satellite above the receiver, and the azimuth angle is the angle measured clockwise from the north of the receiver to the satellite in the horizontal plane of the receiver.

[0083] The pseudorange P is a value obtained by multiplying the signal propagation time from the satellite to the receiver by the speed of light (or the speed of radio wave propagation). P = true distance + satellite clock error + receiver clock error + ionospheric delay + tropospheric delay + multipath effect.

[0084] True Distance: The actual straight-line distance between the satellite and the receiver.

[0085] Satellite clock error: The error between the satellite clock and the GPS time standard.

[0086] Receiver clock error: The error between the receiver clock and the GPS time standard.

[0087] Ionospheric delay: The propagation time delay of a signal due to changes in electron density when it passes through the ionosphere.

[0088] Tropospheric delay: The propagation time delay of a signal through the troposphere due to changes in the atmospheric refractive index.

[0089] Multipath effect: Signal reflections or diffraction reaching the receiver, causing the received signal to be received later than the directly transmitted signal.

[0090] The Doppler shift d is the difference between the frequency of the satellite signal detected by the receiver and the frequency at which the satellite transmitted it. This difference is caused by the relative motion between the satellite and the receiver. When the satellite and receiver are approaching each other, the received frequency increases; when they are moving away from each other, the received frequency decreases. The Doppler shift can be used to calculate the speed of the receiver relative to the satellite.

[0091] The carrier-to-noise ratio (SNR) is the ratio of the received signal strength to the background noise strength, usually expressed in decibels (dB). A high SNR means that the signal strength is much higher than the noise level, which usually means better signal quality and higher positioning accuracy. On the contrary, a low SNR means poor signal quality, which may affect the accuracy and reliability of positioning. SNR is a key indicator for evaluating the quality of satellite signals received by a receiver.

[0092] The signal loss lock indicator LLI represents the loss lock identifier, which ranges from 0 to 7 and is converted into a binary format of 000 to 111, that is, it has 3 bits. Bit 0 and bit 1 are only used for phase Q.

[0093] (1) 0 or space: normal or unknown;

[0094] (2) Bit 0 is set: the previous and current observations are out of lock, and a cycle slip may have occurred;

[0095] (3) Bit 1 is set: the receiver performs half-cycle modulus calculation, or the program cannot process half-cycle data and skips recording the observation value (valid only for the current epoch);

[0096] (4) Bit2 is set: Bit2 is set to 1 to indicate the observed value under anti-spoofing (AS) (which may be affected by increased noise).

[0097] 1.3) Generate a scene mutation recognition feature map and a scene recognition feature map corresponding to the feature data, including:

[0098] 1.31) Project the satellite positions onto the circular image plane to obtain a circular diagram of satellite positions to represent the satellite distribution characteristics.

[0099] x=R×cosEl×cosAz+R

[0100] y=R×cosEl×sinAz+R

[0101] Wherein, x and y represent the X-axis and Y-axis coordinate values ​​of the circular plane rectangular coordinate system respectively; El and Az represent the satellite elevation angle and azimuth angle respectively; R represents the circular image radius;

[0102] Figure 3 The sky map of each scene in a preferred embodiment of the present invention is shown in the figure of open space, dense shade, single-sided building, and double-sided building. In the above four scenes, due to different factors such as position and degree of shielding, it can be seen that the number of satellites and the distribution of satellites are very different.

[0103] 1.32) Perform inter-epoch differential processing on the pseudorange, Doppler shift, carrier-to-noise ratio, signal loss flag and number of visible satellites of each satellite in each epoch observed by the mobile station, and perform inter-station differential processing on the pseudorange, Doppler shift, carrier-to-noise ratio, signal loss flag and number of visible satellites of each satellite in each epoch observed by the mobile station and the base station. Monitoring the changes in the inter-epoch differences can highlight the scene mutation points to better identify the mutation moments of scene switching; before and after the mutation moments, by comparing the observation differences between the base station and the mobile station, the change rules within the scene can be further analyzed to infer the scene category in which the mobile station is located. The specific methods are as follows:

[0104] First, before doing the two difference processing, define the mobile station characteristic matrix set: M rover ={M rover (P), M rover (d), M rover (SNR), M rover (LLI), M rover (SatNum)}, where M rover (P), M rover (d), M rover (SNR), M rover (LLI), M rover (SatNum) is an m*n matrix, where m is the number of satellites and n is the number of epochs. rover (F) ij represents the scene observation value of the mobile station observing the i-th satellite at the j-th epoch. The scene observation value F is the pseudorange, Doppler frequency shift, carrier-to-noise ratio, signal loss flag and number of visible satellites; Figure 4 M rover (P) ij , i=28, j=20, the 19th satellite was not observed.

[0105] Define the base station feature matrix set M base ={M base (P), M base (d), M base (SNR), M base (LLI), M base (SatNum)}, as mentioned above, the element M of the matrix base (F) ij It represents the scene observation value of the base station observing the i-th satellite in the i-th epoch. The scene observation value F is the pseudorange, Doppler frequency shift, carrier-to-noise ratio, signal loss flag and number of visible satellites.

[0106] Finally, define the set D = {D epoch , D rb} is the post-difference feature.

[0107] I: The difference processing between epochs includes: calculating the difference between the scene observation values ​​of the mobile station observing the same satellite between adjacent epochs. rover (P), M rover (d), M rover (SNR), M rover (LLI), M rover (SatNum) is processed by inter-epoch difference to obtain the set D epoch ={D epoch (P), D epoch (d) D epoch (SNR), D epoch (LLI), D epoch (SatNum)}.

[0108] Taking the inter-epoch difference processing of the pseudorange matrix as an example, the pseudorange matrix M rover The n-1 epochs of (P) are subjected to epoch-by-epoch difference processing to obtain D epoch (P), where D epoch (P) ij for:

[0109] D epoch (P) ij =M rover (P) (i，j+1) -M rover (P) (i，j) (j=1,2,…,n-1)

[0110] Similarly, the matrix of the other observed values ​​after difference processing is obtained. As an example, for M rover (SatNum) performs the above-mentioned differential processing to obtain D epoch (SatNum), where D epoch (SatNum) j For: D epoch (SatNum) j =M rover (SNR) (j+1) -M rover (SatNum) (j) (j=1,2,...,n-1).

[0111] II: Differential processing between the mobile station and the base station includes: calculating the difference between the observation values ​​of the same satellite observed by the mobile station and the base station at the same time. rover (P) and M base (P), M rover (d) and M base (d), M rover (SNR) and M base(SNR), M rover (LLI) and M base (LLI), M rover (SatNum) and M base (SatNum) performs differential processing between the mobile station and the base station to obtain the set D rb ={D rb (P), D rb (d) D rb (SNR), D rb (LLI), D rb (SatNum)}.

[0112] Taking the differential processing between the mobile station and the base station of pseudorange P as an example, rover (P) and M base (P) Perform differential processing between the mobile station and the base station to obtain D rb (P), where D rb (P) ij for:

[0113] D rb (P) ij =M rover (P) ij -M base (P) ij (i=1,2,…,m)

[0114] Similarly, the matrix of the other observed values ​​after difference processing is obtained. As an example, for M rover (SatNum) and M base (SatNum) performs the above-mentioned differential processing to obtain D rb (SatNum), where D rb (SatNum) ij For: D rb (SatNum) j =M rover (SatNum) (j) -M base (SatNum) (j) .

[0115] The satellites involved in the difference include observed satellites and unobserved satellites in the system. For unobserved satellites, the mean of the observed values ​​of the open scene is taken as the background value.

[0116] 1.4) Construct scene mutation recognition dataset and scene recognition dataset. Set D is converted into an image and processed by Thiessen polygon interpolation. This is used as a feature image for subsequent neural network training. The specific process of dataset construction is as follows:

[0117] First, data preprocessing is performed, including data cleaning and normalization. Since the data set contains data of various dimensions, in order to better match and prevent training divergence, the training data is normalized without affecting the data time series characteristics to make the data interval uniform. The normalization process is mathematically expressed as follows:

[0118]

[0119] Where max(D) and min(D) refer to the maximum and minimum values ​​of the features after differential processing of all satellite observations.

[0120] Secondly, generate a feature map of the satellite position, divide the feature map of the satellite position into Thiessen polygons according to the Thiessen polygon algorithm, interpolate the feature data after planning unification into the Thiessen polygons to which the satellite position belongs, and obtain the scene mutation recognition feature map and scene recognition feature map of each feature data, specifically including:

[0121] The satellite position circular map is projected into its circumscribed square, and the area outside the satellite position circular map in the square is filled with a grayscale value of 0. According to the Thiessen polygon algorithm, the satellite positions in the satellite position circular map are used as discrete points. Based on the discrete points, the satellite position circular map in the square is divided into Thiessen polygons. Each Thiessen polygon contains only one discrete point. The points on the edge of the Thiessen polygon are equidistant from the discrete points on both sides. Any point in the Thiessen polygon is closest to the discrete point in this Thiessen polygon. The Thiessen polygon can be regarded as the coverage area of ​​the discrete points. The properties of the discrete points in the Thiessen polygon can be used to describe the properties of the Thiessen polygon.

[0122] like Figure 5 The division of Thiessen polygons includes: constructing a Delaunay triangulation of discrete points, calculating the center of the circumscribed circle of each triangle in the Delaunay triangulation, connecting the centers of the circumscribed circles of triangles adjacent to each discrete point in sequence to obtain Thiessen polygons, and for the connecting edges of triangles adjacent to the edge of the triangulation, making the perpendicular bisector of the connecting edge intersect with the circle outline of the satellite position circular map, and the perpendicular bisector, the circle outline and the line connecting the centers of the circumscribed circles of the triangles adjacent to the discrete points together constitute Thiessen polygons.

[0123] Finally, the D of each satellite epoch or D rb Interpolate to pixel values, fill the pixel values ​​corresponding to the satellite positions in the Thiessen polygons to which the discrete points corresponding to the satellite positions belong, and convert D epoch or D rbThe five-channel feature images include the inter-epoch differential pseudo-range map, inter-epoch differential Doppler frequency shift map, inter-epoch differential carrier-to-noise ratio map, inter-epoch differential signal loss identification map and inter-epoch differential visible satellite number map for identifying scene mutations; and the inter-station differential pseudo-range map, inter-station differential Doppler frequency shift map, inter-station differential carrier-to-noise ratio map, inter-station differential signal loss identification map and inter-station differential visible satellite number map for identifying scene categories.

[0124] like Figure 6 The figure shows the pixel values ​​interpolated after the Doppler frequency shift is differentiated and normalized in the single-sided building scene; the horizontal axis corresponds to the satellite number, and the vertical axis corresponds to different epochs. In addition, after the differential visible satellite numbers between stations and the differential visible satellite numbers between epochs are converted into pixel values ​​and filled into the Thiessen polygons to which the discrete points corresponding to the satellite positions belong, the SatNum values ​​are the same in one epoch, and the pixels of each Thiessen polygon are the same, but the SatNum values ​​are different between different epochs, and the SatNum values ​​observed by the mobile station and the base station are different, so the change of SatNum in a continuous time window is an important feature.

[0125] Among them, the satellites corresponding to the discrete points include observed satellites and unobserved satellites in the system. When setting the pixel value of the Thiessen polygon, the mean of the observed values ​​of the open scene is taken as the background value for the Thiessen polygon image part where the satellite is not observed.

[0126] Use feature images to form scene mutation recognition datasets and scene recognition datasets;

[0127] The scene mutation recognition dataset includes the differential feature images between epochs arranged in time and their corresponding mutation labels (0, 1), where 0 means that the scene has not changed and 1 means that the scene has changed suddenly;

[0128] The scene recognition dataset includes time-arranged inter-station differential feature images and their corresponding scene labels (1, 2, 3, 4), where 1 represents an open scene, 2 represents a dense tree-shaded scene, 3 represents a single-sided building scene, and 4 represents a double-sided building scene.

[0129] Finally, the scene mutation recognition dataset and the scene recognition dataset are divided according to the preset ratio, training set: validation set = 7:3.

[0130] The above mutation labels and scene labels are marked when the GNSS receiver collects GNSS raw observations. For example, one mutation label corresponds to one mutation label per second in the GNSS raw observations. The mutation label is 1 (the scene has changed) about 4 seconds before and after the scene switch, and the mutation label is 0 (the scene has not changed) in the rest of the time. The scene label is marked according to the scene.

[0131] 1.5) Train the ConvLSTM neural network based on the training set and the validation set to obtain the scene mutation recognition model and the scene recognition model.

[0132] The present invention uses a multi-layer ConvLSTM in the encoding network to extract the spatiotemporal feature information in the feature image. Figure 7 This is a schematic diagram of the ConvLSTM network, including sequentially connected input layers, convolutional layers, multiple stacked ConvLSTM layers, fully connected layers, and output layers. The time step T = 4s is set to fully and appropriately use the information in the time window. The multi-channel feature image time series is regarded as a period of sliding window size of 4s and is sequentially input into the network for training.

[0133] The specific process of training the ConvLSTM neural network through the data set to obtain the scene mutation recognition model and the scene recognition model is as follows:

[0134] Configure the model's hyperparameters, use cross entropy as the loss function, and use the Adam optimizer for gradient update. Train until the preset iteration termination condition is met and then fix the scene mutation recognition model and the scene recognition model. The preset iteration termination condition is: both the training accuracy and the verification accuracy meet the corresponding preset accuracy standards, and the difference between the training accuracy and the verification accuracy is within the preset range.

[0135] Model execution step, scene recognition: the scene mutation recognition model and the scene recognition model obtained in the module training step are used to identify the GNSS data to be identified in turn.

[0136] 2.1) Perform inter-epoch differential processing on the collected unknown scene GNSS data, using the same inter-epoch differential method as described in 1.3). After differential processing, the scene mutation recognition model determined in the module training step is input to identify the mutation time points in the unknown scene GNSS data, which are used to define the boundaries between different scenes. Each two adjacent mutation points are defined as an independent time period, and each time period represents an independent scene. Based on the mutation time points, the collected scene data to be identified are divided in order to obtain multiple time period data sets.

[0137] 2.2) For the multiple time period data sets that have been divided, inter-station differential processing is performed, and the method is the same as the differential method between the mobile station and the base station described in 1.3). After differential processing, the inter-station differential data sets of multiple time periods are obtained. Each section of the inter-station differential data set is sequentially input into the scene recognition model determined in module 1 to identify the specific scene category of each time period.

[0138] Among them, the scene recognition result of a certain period of time obtained by the scene recognition model is actually a probability vector, that is, the probability of each epoch in the data set being identified as a different scene. Since a time period represents an independent scene, it is necessary to unify the results of scene category recognition at each epoch. That is, in the probability distribution vector, it is necessary to select the scene category with the highest probability and unify the scene recognition results of the same period.

[0139] Through the above method, combined with the scene mutation point recognition model and the scene recognition model, accurate scene category recognition results for GNSS data to be recognized of any length can be obtained.

[0140] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the steps of the above method are implemented when the processor executes the computer program.

[0141] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0142] Compared with the prior art, the method proposed in the present invention significantly enhances the performance regularity characteristics of GNSS observations in different complex scenarios through the inter-epoch difference and inter-station difference strategies of GNSS observations. It not only accurately identifies the mutation time point of scene switching and solves the problem of fuzzy scene recognition results when switching scenes, but also deeply mines the spatiotemporal correlation characteristics of features with the help of the ConvLSTM model, thereby effectively distinguishing the difference performances between scenes and realizing accurate classification of scene categories. And only GNSS satellite positions and observation value information are used as scene recognition features, which consumes less hardware resources, can realize low computing power, high precision, high intelligence, and strong generalization ability of scene recognition functions, and provide good support for satellite navigation and high-precision positioning.

[0143] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only for illustrating the principles of the present invention. Without departing from the spirit and scope of the present invention, the present application may have various changes and improvements, and these changes and improvements all fall within the scope of the present application for which protection is sought.

Claims

1. A differential GNSS scene recognition method based on ConvLSTM, characterized in that: The scene recognition method comprises the following steps: Using the GNSS raw observations of multiple satellites by the mobile station and the base station under four preset scenarios, namely, open space, dense shade, single-sided building, and double-sided building, at multiple identical epochs, a scene mutation identification feature map and a scene recognition feature map about feature data are generated; the scene mutation identification feature map is used to train a binary ConvLSTM neural network to obtain a scene mutation identification model for identifying mutation epoch points of the GNSS raw observations in the scene to be identified; the scene recognition feature map is used to train a multi-classification ConvLSTM neural network to obtain a scene recognition model for identifying the scene corresponding to the GNSS raw observations between adjacent mutation epoch points in the scene to be identified; Generating a scene mutation recognition feature map and a scene recognition feature map about feature data comprises the following steps: A1, projecting satellite positions onto a circular image plane to obtain a satellite position circular map, projecting the satellite position circular map onto a circumscribed square thereof, filling an area outside the satellite position circular map in the square with a grayscale value of 0, and dividing the satellite position circular map in the square into Thiessen polygons based on the discrete points using the satellite positions in the satellite position circular map as discrete points according to a Thiessen polygon algorithm; A2, perform inter-epoch differential processing on various residual characteristic data of the mobile station, and perform inter-station differential processing on various residual characteristic data of the same satellite observed by the mobile station and the base station; perform data cleaning and normalization processing on the data after inter-epoch differential processing and inter-station differential processing to obtain inter-epoch differential data and inter-station differential data of each satellite; various residual characteristic data are characteristic data except satellite elevation angle El and satellite azimuth angle Az; A3, interpolating the inter-epoch differential data and inter-station differential data of each satellite into pixel values, filling the pixel values ​​corresponding to the satellite positions in the Thiessen polygons to which the discrete points corresponding to the satellite positions belong, and obtaining the scene mutation recognition feature map and the scene recognition feature map; The characteristic data include: satellite elevation angle El and satellite azimuth angle Az as well as pseudorange P, Doppler frequency shift d, carrier-to-noise ratio , Signal lock indicator , Number of visible satellites At least one of .

2. The ConvLSTM-based differential GNSS scene recognition method according to claim 1, characterized in that: Using the scene mutation recognition feature graph to train a binary classification ConvLSTM neural network includes the following steps: marking a mutation label according to whether a scene change occurs in the scene mutation recognition feature graph, 0 indicates that the scene has not changed, and 1 indicates that the scene has mutated, arranging the marked scene mutation recognition feature graph into a time series according to time, and obtaining a scene mutation recognition data set; dividing the scene mutation recognition data set into a training set and a validation set according to a preset ratio; using the scene mutation recognition data set to train a binary classification ConvLSTM neural network to construct a scene mutation recognition model; Using scene recognition feature maps to train a multi-classification ConvLSTM neural network includes the following steps: marking scene labels according to the scenes corresponding to the scene recognition feature maps, arranging the marked scene recognition feature maps into a time series by time, and obtaining a scene recognition data set; dividing the scene recognition data set into a training set and a validation set according to a preset ratio; using the scene mutation recognition data set to train a multi-classification ConvLSTM neural network and construct a scene mutation recognition model.

3. The ConvLSTM-based differential GNSS scene recognition method according to claim 1, characterized in that: The epoch points for identifying the mutation of the GNSS raw observation values ​​of multiple satellites at multiple epochs of the mobile station in the scene to be identified include: Obtaining the GNSS original observation values ​​of the mobile station and the base station for multiple satellites at multiple epochs in the scene to be identified and processing them according to steps A1 to A3 to obtain a scene mutation identification feature map to be identified and a scene mutation identification feature map to be identified, inputting the scene mutation identification feature map to be identified into the scene mutation identification model to obtain a scene mutation epoch point; The data of the scene recognition feature graph to be recognized between adjacent mutation epoch points are divided into segments in turn to obtain data of several independent scene time periods.

4. The ConvLSTM-based differential GNSS scene recognition method according to claim 3, characterized in that: Identifying the scenes corresponding to the GNSS raw observation values ​​of the mobile station and the base station on multiple satellites at multiple epochs in the scene to be identified between adjacent mutation epoch points includes: inputting the independent scene period data into the scene recognition model to obtain a scene recognition result.

5. The ConvLSTM-based differential GNSS scene recognition method according to claim 1, characterized in that: A2 specifically includes: A21. Define a matrix set: The GNSS raw observation values ​​of multiple satellites collected by the mobile station under n preset scenarios at multiple epochs are converted into a mobile station feature matrix set: ; Each feature matrix is ​​an m*n' matrix, where m is the number of satellites, n' is the number of epochs, and the elements of the matrix are represents the pseudorange value of the mobile station's ith satellite at the jth epoch, represents the Doppler shift of the i-th satellite at the j-th epoch of the mobile station, Indicates the carrier-to-noise ratio of the i-th satellite at the j-th epoch, and the signal loss flag Indicates the signal loss flag of the mobile station's i-th satellite at the j-th epoch, represents the number of visible satellites observed by the mobile station at the jth epoch; The original GNSS observation values ​​of the base station on multiple satellites at multiple epochs collected under n types of preset scenarios are converted into a set of base station feature matrices: ； Each feature matrix is ​​an m*n' matrix, where m is the number of satellites, n' is the number of epochs, and the elements of the matrix are represents the pseudorange value of the base station observing the i-th satellite at the j-th epoch, It indicates the Doppler frequency shift of the ith satellite observed by the base station at the jth epoch. represents the carrier-to-noise ratio of the ith satellite observed by the base station at the jth epoch, Indicates that the base station loses lock on the signal of the i-th satellite at the j-th epoch. represents the number of visible satellites observed by the base station at the jth epoch; A22. Perform inter-epoch difference processing on the observation data of the mobile station, that is, calculate the difference between the scene characteristic values ​​of the mobile station observing the same satellite between adjacent epochs: right Perform the above-mentioned differential processing to obtain ,in for: ; right Perform the above-mentioned differential processing to obtain ,in for: ; right Perform the above-mentioned differential processing to obtain ,in for: ; right Perform the above-mentioned differential processing to obtain ,in for: ; right Perform the above-mentioned differential processing to obtain ,in for: ; Get the collection: ; A23. Perform inter-station differential processing on various characteristic data of the same satellite observed by the mobile station and the base station, that is, calculate the difference between the observation values ​​of the same satellite observed by the mobile station and the base station at the same time: and Perform the above-mentioned differential processing to obtain ,in for: ; right and Perform the above-mentioned differential processing to obtain ,in for: ; right and Perform the above-mentioned differential processing to obtain ,in for: ; right and Perform the above-mentioned differential processing to obtain ,in for: ; right and Perform the above-mentioned differential processing to obtain ,in for: ; Get the collection: ; A24. Normalization processing: The normalization process is mathematically expressed as follows: ； in and They refer to the maximum and minimum values ​​of the observations after inter-epoch difference processing or inter-station difference processing, respectively.

6. The differential GNSS scene recognition method based on ConvLSTM as claimed in claim 1, characterized in that: The scene mutation identification feature map includes the inter-epoch differential pseudo-range map, the inter-epoch differential Doppler frequency shift map, the inter-epoch differential carrier-to-noise ratio map, the inter-epoch differential signal loss identification map and the inter-epoch differential visible satellite number map for identifying the scene mutation; The scene recognition feature map includes the inter-station differential pseudo-range map, the inter-station differential Doppler frequency shift map, the inter-station differential carrier-to-noise ratio map, the inter-station differential signal loss identification map and the inter-station differential visible satellite number map for identifying the scene category; The satellites corresponding to the discrete points include observed satellites and unobserved satellites in the system. When setting the pixel value of the Thiessen polygon, the mean of the observed values ​​of the open scene is taken as the background value for the Thiessen polygon image part where the satellite is not observed.

7. The differential GNSS scene recognition method based on ConvLSTM as claimed in claim 1, characterized in that: The binary classification ConvLSTM neural network and the multi-classification ConvLSTM neural network include an input layer, a convolutional layer, a plurality of stacked ConvLSTM layers, a fully connected layer, and an output layer connected in sequence.

8. The differential GNSS scene recognition method based on ConvLSTM as claimed in claim 1, characterized in that: The ConvLSTM neural network is trained by the data set to obtain the scene mutation recognition model and the scene recognition model, which specifically include: The hyperparameters of the model are configured, cross entropy is used as the loss function, the gradient is updated with the Adam optimizer, and the scene mutation recognition model and the scene recognition model are obtained after the training is performed until the preset iteration termination condition is met; The preset iteration termination condition is: the training accuracy and the verification accuracy both reach the corresponding preset accuracy standards, and the difference between the training accuracy and the verification accuracy is within a preset range.

9. A ConvLSTM-based differential GNSS scene recognition system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images