A combined navigation and positioning method for extreme weather rescue

Abstract

The present invention proposes a combined navigation and positioning method for extreme weather rescue, comprising the following steps: step 1, constructing an original training data set; step 2, adding incremental data set training samples on the basis of the original training data set, constructing an incremental grid real-time correction model, and obtaining a real-time dynamically updated multipath heat map; step 3, assisting GNSS / IMU tight combination solution according to the real-time dynamically updated multipath heat map, and completing the combined navigation and positioning for extreme weather rescue. The present invention fully exploits the signal characteristics and multipath errors in the current positioning environment, continuously receives and accumulates new data, improves sample diversity, and solves the problem of single sample degradation; the present invention also improves the accuracy and adaptability of the grid multipath error prediction model based on machine learning in urban environments affected by extreme weather.

Description

Technical Field

[0001] The invention relates to a combined navigation and positioning method, in particular to a combined navigation and positioning method for extreme weather rescue. Background Art

[0002] This section merely provides background information related to the present disclosure and is not necessarily prior art.

[0003] Since coastal cities are easily affected by typhoons and rainstorms, emergency rescue of people in the disaster-stricken areas is critical once a natural disaster occurs. At the same time, high-precision, real-time, and reliable navigation and positioning technology is also needed to improve rescue efficiency. However, due to signal attenuation caused by extreme weather and the obstruction of tall buildings in complex urban environments, GNSS (Global Navigation Satellite System) signals are more susceptible to multipath interference and non-line-of-sight reception. The multipath effect results in low GNSS signal quality, which seriously restricts the positioning accuracy of GNSS and its combined navigation system. Therefore, accurate positioning in complex urban environments with extreme weather is of great research significance for the evacuation of rescue vehicles and personnel after natural disasters.

[0004] In view of the limitations of the above-mentioned navigation and positioning technologies, domestic and foreign researchers have carried out many valuable explorations, which are mainly divided into signal processing methods, antenna design methods and methods based on observation value modeling. Since the signal processing-based methods cannot effectively identify NLOS (non-line-of-sight) signals and reduce their impact, and the antenna line design methods are heavy and expensive, these two methods are not suitable for the widespread promotion of urban positioning. The error correction method based on observation value modeling is based on mathematical statistics, machine learning and other means to establish a mathematical model between the original GNSS observation and the multipath error to obtain the predicted value of the multipath error, and use the predicted value to correct the error, thereby improving the navigation and positioning accuracy. At present, there are two main error correction methods based on observation value modeling: modeling based on reference point data and modeling based on regional grids.

[0005] Sun proposed a training regression model based on GBDT (Gradient Boosting Decision Tree), which predicts and corrects multipath errors by comprehensively considering signal strength, satellite elevation angle and pseudorange residual. The positioning test results in complex urban environments show that the three-dimensional positioning accuracy of the algorithm based on multipath error correction is more than 70% higher than that of the traditional positioning algorithm. Qin proposed a multipath error prediction algorithm based on bagged trees, which considers signal strength and satellite elevation angle to predict and correct multipath errors, mainly improving the GNSS positioning accuracy in leaf-blocked environments, and the three-dimensional RMSE of the positioning results is improved by more than 80%. In addition to using the multipath prediction prediction value to directly correct the error, the prediction value is also used in combination with Kalman filtering to adjust the filter parameters. Zhang uses signal strength and satellite elevation angle as the input of the bagged tree model to predict multipath errors, and proposes a GNSS measurement noise covariance update strategy based on the PDOP value and the average value of the predicted multipath error. This strategy can adaptively adjust the Kalman filter parameters in the traditional GNSS / IMU loose combination algorithm. Compared with the traditional GNSS / IMU loose combination algorithm, the horizontal and three-dimensional positioning accuracy of this method are improved by 28.15% and 43.10% respectively. Sun proposed a measurement noise covariance update scheme for the GNSS / IMU tight combination algorithm in urban areas by considering signal strength, satellite elevation angle and coordinate information, using an integrated bagged regression tree model to predict multipath error, and constructing an adaptive factor based on the multipath error prediction value. The algorithm can improve the three-dimensional positioning accuracy of 9.21m, which is 55% higher than the traditional GNSS / IMU tight combination algorithm based on extended Kalman filter, and 15% higher than the multipath error correction method based on extended Kalman filter, proving the effectiveness of the adaptive filtering algorithm based on multipath error prediction in improving the positioning accuracy of GNSS / IMU integrated navigation. However, the method based on reference point data modeling still has problems: the positioning scheme that predicts multipath error for all urban scenes with the same model is not fine enough, and the accuracy of multipath error prediction is not enough.

[0006] The method of assisting navigation and positioning based on regional grid multipath modeling divides the positioning area into grid units, and establishes a multipath error model in each grid unit to achieve refined modeling of multipath effects. While effectively improving the navigation positioning accuracy and stability, it also has the advantages of simple installation and low cost. Sun proposed a multipath error prediction and correction strategy based on regional grids to improve the GNSS positioning accuracy in urban environments. This method uses signal strength, satellite elevation angle and regional prior position information as input features, and uses the machine learning random forest algorithm to train the regression model to predict multipath errors and use the correction pseudorange. The dynamic test results in the urban environment show that compared with the traditional positioning method, the positioning accuracy of the algorithm in the horizontal direction is improved by 41.50%, and the positioning accuracy in the three-dimensional direction is improved by 63.38%. Lee introduced a nonlinear regression model to estimate the high-elevation clock error first, and trained the extracted unbiased multipath through SVM to generate a nonlinear multipath map model based on the azimuth and elevation angle of each satellite in the urban depth area. This model effectively improves the positioning accuracy of the urban depth area by 58.4% and 77.7% in the vertical direction. Since the previously constructed grid or map models are divided on a two-dimensional plane, it is impossible to accurately correct the elevation of the user. Sun et al. also extended the grid prediction model from the two-dimensional plane to the three-dimensional layout, and proposed a navigation and positioning method for complex urban environments based on three-dimensional grid multipath modeling, which fully considered the reflection environment at different elevations and improved the positioning accuracy in complex interchange environments and rugged roads. Although the regional grid division modeling for multipath in complex urban environments has effectively improved the positioning accuracy to a certain extent, the method still has limitations, including: due to the signal attenuation in complex urban environments, especially under the influence of extreme weather, the prior position information is inaccurate, and it is easy to match the wrong grid prediction model. The current environment is affected by extreme weather and has a large difference from the pre-trained environment. The pre-trained multipath prediction model cannot be dynamically adjusted according to the signal characteristics in the current environment. The model lacks adaptability to extreme environments and the real-time nature of the positioning process, which leads to larger positioning errors, and the rescue vehicle cannot be accurately positioned, which reduces the rescue efficiency in urban environments during extreme weather.

[0007] In summary, the prior art has the following defects:

[0008] (1) The existing GNSS / IMU combined navigation scheme based on machine learning methods for regional grid multipath modeling has the problems of fixed initial training data set and single sample library. Especially in complex urban environments under extreme weather conditions, the signal attenuation is severe, and the degraded training samples lead to poor prediction results. It is impossible to improve the combined navigation positioning accuracy assisted by regional grid multipath modeling, which affects the precise positioning of rescue vehicles and reduces rescue efficiency.

[0009] (2) The grid multipath error prediction model constructed by the existing navigation and positioning scheme based on regional grid multipath modeling in complex urban environments is static. However, the complex urban scenes in extreme weather are becoming increasingly different from the previous ones. The multipath error prediction model based on offline data training lacks adaptability to the current environment and the real-time positioning process, which will lead to insufficient prediction accuracy of the model and fail to guarantee the reliability of the combined navigation and positioning system used for extreme weather rescue.

[0010] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0011] Purpose of the invention: The technical problem to be solved by the present invention is to provide a combined navigation and positioning method for extreme weather rescue in view of the deficiencies in the prior art.

[0012] In order to solve the above technical problems, the present invention discloses a combined navigation and positioning method for extreme weather rescue, comprising the following steps:

[0013] Step 1, construct the original training data set;

[0014] Step 2: Add incremental data set training samples to the original training data set, build an incremental grid real-time correction model, and obtain a real-time dynamically updated multipath heat map;

[0015] Step 3: Perform auxiliary GNSS / IMU tight combination solution based on the real-time dynamically updated multipath heat map to complete the combined navigation and positioning for extreme weather rescue.

[0016] Furthermore, the construction of the original training data set described in step 1 specifically includes:

[0017] Step 1-1, calibrate the training data set;

[0018] Step 1-2, constructing a regional grid layout;

[0019] Step 1-3, dividing the sample data in the training data set according to the satellite constellation types of GPS satellites and Beidou satellites;

[0020] According to the reference position corresponding to the sample data, the grid layer to which the sample data belongs is determined according to the three-dimensional grid layout described in step 1-2, and the original training data set is obtained, which are respectively expressed as:

[0021] GPS original training data set [δ,C / N 0 ,θ e ,θ a ,Δρ] GPS_gridand BeiDou original training dataset [δ,C / N 0 ,θ e ,θ a ,Δρ] BDS_grid .

[0022] Furthermore, the calibration training data set described in step 1-1 is as follows:

[0023] Step 1-1-1, the positioning carrier repeatedly drives along a predetermined route in an urban area to collect prior data for constructing an original training data set;

[0024] Step 1-1-2, differential positioning is performed through a high-precision integrated navigation and positioning system and base station GNSS observation data to obtain reference coordinate information;

[0025] Step 1-1-3, collect the original observations output by the GNSS receiver, and extract the corresponding input features in the original observations, the features include: pseudorange residual δ, carrier-to-noise ratio C / N 0 、Satellite altitude angle θ e and the satellite azimuth θ a , a set of characteristic variables is expressed as: [δ,C / N 0 ,θ e ,θ a ];

[0026] Step 1-1-4, invert the reference coordinate information obtained in step 1-1-2 and the original GNSS observations collected in step 1-1-3 to obtain the pseudorange error value Δρ, as follows:

[0027] Δρ=c(Δδt u -Δδt s )+ΔI+ΔT+ε

[0028] Where c is the speed of light in a vacuum; Δδt s represents the residual after correction of satellite clock error; Δδt u represents the residual after correction of the receiver clock error; ΔI represents the residual after incomplete correction of ionospheric delay; ΔT represents the residual after incomplete correction of tropospheric delay;

[0029] And the above pseudo-range error is regarded as a multipath error;

[0030] Step 1-1-5, calibrate the pseudorange error value obtained in step 1-1-4 with the characteristic variable extracted in step 1-1-3 to obtain the labeled sample data to form the training data set [δ, C / N 0 ,θ e ,θ a ,Δρ].

[0031] Furthermore, the construction of the regional grid layout described in step 1-2 specifically includes:

[0032] Divide the two-dimensional plane grid, that is, according to the geographic information of the urban area and the GNSS location data, take the smallest square A of the area containing the location data as the modeling area, one side of the square is parallel to the E axis in the ENU coordinate system, divide the hexagonal grid inside the square area, and divide it in units of the ENU coordinates of the hexagon, set the side length of the hexagon to l, and take the center point of the hexagonal grid as o;

[0033] The two-dimensional plane grid divided above is extended into grid columns in the height direction, and each grid column is divided into n grid layers in turn according to the selected height h to complete the construction of the three-dimensional grid layout.

[0034] Furthermore, the construction of the incremental grid real-time correction model described in step 2 specifically includes:

[0035] Step 2-1: Construct a GNSS / IMU integrated navigation solution feedback update mechanism, as follows:

[0036] Match and call the existing three-dimensional grid prediction model to obtain the multipath error prediction value of the current epoch and determine whether it meets the following conditions:

[0037]

[0038] in, Represents the absolute value of the multipath error prediction value at the current epoch, It represents the mean absolute pseudorange error value of GNSS data in open areas;

[0039] If satisfied, the original grid prediction model and prediction rules obtained after training in step 2-2-2 are fed back to update, as follows:

[0040] The temporal and spatial information of the current epoch positioning carrier are respectively recorded as the current epoch T i and the current position P i , where the current position P i Including the coordinate information of the ENU coordinate system where the current epoch positioning carrier is located, as the real-time data for correcting the grid prediction model at the current time i;

[0041] Otherwise, do not update;

[0042] Step 2-2, based on the three-dimensional grid layout, multipath error modeling of incremental SVR is performed grid by grid, that is, a multipath heat map that is dynamically updated in real time is constructed.

[0043] Further, the construction of the real-time dynamically updated multipath heat map described in step 2-2 is as follows:

[0044] Step 2-2-1, construct incremental data set training samples, as follows:

[0045] According to the spatial coordinates of the current epoch positioning carrier, search for the grid layer to which it belongs in the three-dimensional grid layout, and extract the input features of the observation quantity, including: pseudorange residual δ, carrier-to-noise ratio C / N 0 、Satellite altitude angle θ e and the satellite azimuth θ a , get the incremental training samples, expressed as: [δ,C / N 0 ,θ e ,θ a ] c ;

[0046] Invert and calibrate according to the methods in steps 1-1-4 and 1-1-5 to obtain the incremental data set, expressed as: [δ, C / N 0 ,θ e ,θ a ,Δρ] c ;

[0047] According to the method in steps 1-3, the incremental data set is divided according to the GPS and BDS constellation types and grids to obtain the GPS incremental data set and the Beidou incremental data set, which are expressed as: [δ, C / N 0 ,θ e ,θ a ,Δρ] c_GPS_grid and [δ,C / N 0 ,θ e ,θ a ,Δρ] c_BDS_grid ;

[0048] Step 2-2-2, perform incremental SVR multipath error training for each grid to obtain the original grid prediction model, specifically including:

[0049] Determine the grid layer to which the sample data in the original training data set belongs. Assume that the center point of the hexagon in the three-dimensional grid layout is o, and the x-axis and y-axis correspond to the E-axis and the N-axis in the ENU coordinate system respectively. The judgment is made by the following two conditions:

[0050] Condition 1, y o ≥l or

[0051] Condition 2,

[0052] Only when the above two equations are satisfied, the position of the sample data (x o ,y o ) in the grid column;

[0053] Traverse each sample in the original training data set and divide it into the corresponding grid column. According to the height data of each sample in the original training data set, divide it into the corresponding grid layer, and then perform SVR regression training on each grid layer to obtain the SVR prediction rule, that is, the original grid model prediction rule;

[0054] Step 2-2-3, update the original grid prediction model and prediction rules, that is, according to the grid layer to which the incremental data set training samples described in step 2-2-1 belong, update and optimize the SVR prediction rules of the corresponding grid layer, that is, match the incremental data set training samples one by one to the grid layer to which they belong, introduce multiple new training samples and perform incremental training together with the original training samples in the grid layer, and obtain dynamically updated SVR prediction rules, as follows:

[0055] Suppose the training sample of the incremental dataset is X c , let the Lagrange multiplier initialization condition Δa c =0, use the current SVR prediction rule to determine whether it violates the above KKT conditions, and make the new optimization problem meet the KKT conditions again, that is, update the current SVR prediction rule. The update process is as follows:

[0056] According to the KKT condition, define the deviation coefficient θ i And the sample error h(X i ) are:

[0057]

[0058] Combined with the current SVR prediction rule, and based on whether there is a support vector and whether there is a boundary support vector, the incremental data set training samples are divided into three parts:

[0059] E set, that is, the boundary support vector set:

[0060]

[0061] S set, that is, the support vector set:

[0062]

[0063] R set, that is, the set of non-support vectors:

[0064] R={i|θ i =0}

[0065] During the incremental training process, the above conditions are used to screen the newly added samples. When the newly added samples belong to the R set, the samples are directly discarded. When the samples belong to the E set or the S set, the samples are retained.

[0066] The final optimized SVR prediction rule is as follows:

[0067]

[0068] Among them, f′ is the updated prediction function, which is the final regression curve determined by the support vector obtained by balance judgment. is the multipath error prediction value updated by batch incremental mode;

[0069] Step 2-2-4, completes the construction of a real-time dynamically updated multipath heat map, wherein the real-time dynamically updated multipath heat map includes: the original grid prediction model and prediction rules, and the updated grid prediction model and prediction rules.

[0070] Further, the SVR regression training of each grid layer described in step 2-2-2 specifically includes:

[0071] Assume that the number of samples in the original training data set is N, and the preset deviation threshold between the predicted value of the prediction function f(X) and the true value Y of the multipath error is ε. The goal of the SVR regression training is to minimize the error between the predicted value and the true value of the prediction function, which is achieved by solving the following optimization problem:

[0072]

[0073] st(ω T ·X+b)-Y i ≤ε+ξ i

[0074]

[0075]

[0076] Among them, C is the regularization coefficient and the loss function is X i represents the i-th group of input features, X = [δ, C / N 0 ,θ e ,θ a ],Y i represents the true value of the multipath error corresponding to the i-th group of features, ξ i and represents the introduced slack variable, J represents the number of features, ω represents the weight vector, b represents the bias, and the sum represents traversing all data points;

[0077] After introducing the Lagrange coefficient a and converting the above optimization problem into a dual form, we get:

[0078]

[0079]

[0080]

[0081] According to the following KKT conditions:

[0082] a i (f(X i )-Y i -ε-ξ i )=0

[0083]

[0084]

[0085] After nonlinear mapping, the SVR prediction rule is:

[0086]

[0087] Among them, K(,) represents the kernel function operation, a i , is the Lagrange multiplier.

[0088] Further, the step 3 described in which the GNSS / IMU tight combination solution is assisted according to the real-time dynamically updated multipath heat map includes:

[0089] Step 3-1, in the extreme weather rescue state, extract the input features and position precision factor PDOP value from the original observations of the GNSS receiver, and obtain the initial position according to the IMU mechanical arrangement;

[0090] Step 3-2, calling the real-time dynamically updated multipath heat map, matching the initial position to the corresponding grid layer, performing position calculation, and obtaining the specific position of the carrier in the extreme weather rescue state.

[0091] Furthermore, the position solution described in step 3-2 is specifically performed as follows:

[0092] If the position precision factor PDOP value is greater than or equal to the threshold, no correction is performed, and the GNSS / IMU tightly combined Kalman filter is directly used to calculate the position solution, and this position solution is used as the final solution;

[0093] If the PDOP value is less than the threshold, the pseudorange error is predicted using the grid prediction model, and the incremental grid real-time correction model constructed in step 2 is used for the next correction.

[0094] Further, the step 3-2 of predicting pseudorange errors using the grid prediction model specifically includes:

[0095] Extract the multipath error value obtained by the adjacent grid according to the grid prediction model, and assign weight W to the adjacent grid n , and the final multipath error prediction value is obtained after dynamic weighting processing It is expressed as follows:

[0096]

[0097] Where n represents the grid prediction model number of the adjacent grid, and m represents the number of grid prediction models of the adjacent grid;

[0098] The multipath error prediction value is used to construct the adaptive factor f(·) to adjust the GNSS / IMU tight combination measurement noise matrix R k , as follows:

[0099] f(·)=F(·) 2

[0100]

[0101] Where X = [δ, C / N 0 ,θ e ,θ a ];

[0102] Adjust the measurement noise matrix R k Finally, the carrier position is obtained by Kalman filtering.

[0103] Beneficial effects:

[0104] (1) The present invention aims at the problem that the single sample degradation affects the prediction effect during the offline training process of the existing regional grid multipath modeling scheme based on the machine learning method, and constructs incremental data set training samples. By incrementally processing the training samples, the sample degradation problem is solved. When the current positioning result and the previous incremental data are in the same grid and belong to the same constellation type, the positioning results of multiple epochs are used to use the incremental training samples as batch training data samples together with the original training data set to perform incremental SVR (Support Vector Regression) training on a grid-by-grid basis, fully exploit the signal characteristics and multipath errors in the current positioning environment, continuously receive and accumulate new data, and improve sample diversity.

[0105] (2) In order to solve the problem that the existing multipath error prediction model based on machine learning method is fixed, resulting in insufficient prediction accuracy and real-time performance, the present invention proposes a feedback mechanism for updating the grid prediction model using the spatiotemporal information of the current epoch positioning result. When the quality of the grid prediction model cannot meet the demand, the feedback mechanism will call the positioning result of the epoch and extract the spatiotemporal information therein, and perform real-time correction and optimization on the grid prediction model accordingly. At the same time, in the model prediction process, the continuously updated and optimized grid model and the adjacent grids are dynamically weighted, and the prediction of a single grid model is transformed into the prediction using multiple grid models, so as to realize the adaptive update of the grid prediction model, thereby improving the accuracy and adaptability of the grid multipath error prediction model based on machine learning in urban environments affected by extreme weather. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0107] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0108] Figure 2 It is a schematic diagram of incremental SVR training in the present invention.

[0109] Figure 3 It is a schematic diagram of the method for determining the grid attribution of training samples in the present invention. DETAILED DESCRIPTION

[0110] The core content of the present invention is a combined navigation and positioning method for extreme weather rescue, which includes the following three stages: (1) constructing the original training data set; (2) constructing the incremental grid real-time correction model; (3) real-time dynamic update of the multipath heat map-assisted GNSS / IMU tight combination solution. The overall process is as follows: Figure 1 As shown:

[0111] (1) Constructing the original training dataset

[0112] 1) Calibrate the training data set

[0113] The positioning carrier repeatedly drives along the predetermined route in the urban area to collect prior data for the construction of the original training data set. Then, the reference coordinate information is obtained through differential positioning between the high-precision integrated navigation and positioning system and the base station GNSS observation data. At the same time, the original observations output by the GNSS receiver are collected, and the corresponding input features in the original observations are extracted, including: pseudorange residual δ, carrier-to-noise ratio C / N 0 、Satellite altitude angle θ e and the satellite azimuth θ a , a set of characteristic variables is expressed as [δ,C / N 0 ,θe ,θ a ].

[0114] The pseudo-range error value obtained by inverting the reference coordinate information and the original GNSS observations is calibrated with the extracted characteristic variables, and the pseudo-range error value is expressed as:

[0115] Δρ=c(Δδt u -Δδt s )+ΔI+ΔT+ε

[0116] Where c is the speed of light in a vacuum; Δδt s represents the residual after correction of satellite clock error; Δδt u represents the residual after the receiver clock error is corrected; ΔI represents the residual incompletely corrected in the ionospheric delay; ΔT represents the residual incompletely corrected in the tropospheric delay; since the error caused by the multipath effect in ε is much larger than other errors, and the four remaining residuals can also be ignored compared with the multipath error, the error caused by the multipath effect in the pseudorange error is obtained, and the multipath error calibration can be completed. The above formula can be regarded as the multipath error. Each set of characteristic variables is calibrated with the corresponding multipath error to obtain the labeled sample data, which constitutes the training data set [δ,C / N 0 ,θ e ,θ a ,Δρ].

[0117] 2) Construct regional grid layout

[0118] According to the geographic information of the urban area and the GNSS position data, the smallest square A in the area containing the position data is taken as the modeling area. One side of the square is parallel to the E-axis in the ENU coordinate system. The hexagonal grid is divided into units of the ENU coordinates of the hexagon inside the square area. The side length of the hexagon is set to l according to the actual situation and positioning accuracy, and the center point of the hexagonal grid is taken as o.

[0119] Since the GNSS / IMU signal modeling based on the three-dimensional grid can fully consider the reflection environment at different elevations, adding the elevation dimension to the real-time dynamic update multipath grid proposed in this method can improve the positioning performance in rugged urban roads, complex interchange environments, and drone application scenarios. Therefore, it is also necessary to extend the above-mentioned two-dimensional plane grid in the height direction into grid columns, and each grid column is divided into n grid layers according to the selected height h.

[0120] GPS (Global Positioning System) and BDS (BeiDou Navigation Satellite System) have different orbit repetition periods, so the training data needs to be divided according to the satellite constellation type. On this basis, it is also necessary to determine the grid to which the training sample belongs according to the divided grid layout based on the corresponding reference position of the sample. Finally, the original training data set is constructed and expressed as: [δ,c / N 0 ,θ e ,θ a ,Δρ] GPS_grid and [δ,C / N 0 ,θ e ,θ a ,Δρ] BDS_grid .

[0121] (2) Constructing an incremental grid real-time correction model

[0122] 1) GNSS / IMU integrated navigation solution feedback update mechanism (IMU, Inertial Measurement Unit)

[0123] Since the pseudo-range error value output by the three-dimensional grid prediction model can effectively reflect the current grid positioning quality, when executing the matching and calling of the existing three-dimensional grid prediction model step, the current epoch error prediction value is extracted and compared with the average absolute pseudo-range error value of the GNSS data in the open area to evaluate the quality of the grid prediction model. If the epoch error prediction value satisfies the following formula, it is determined that the grid prediction model needs to be updated using the feedback of the calculated carrier position result. Otherwise, it is determined that the grid quality meets the conditions and the grid prediction model does not need to be updated. The original grid prediction model is continued to be used for the next step of positioning solution.

[0124]

[0125] in, Represents the absolute value of the multipath error prediction value at the current epoch, It represents the mean absolute pseudorange error of GNSS data in open areas.

[0126] Furthermore, when it is determined that the grid prediction model is not accurate enough, the temporal and spatial information contained in the position information of the epoch carrier is extracted and recorded as the current epoch T i and the current position P i , current position P i Including the coordinate information of the ENU coordinate system where the epoch carrier is located, which is used as real-time data for correcting the grid prediction model at the current time i.

[0127] 2) Multipath error modeling for incremental SVR on a grid-by-grid basis

[0128] 1. Use the extracted spatiotemporal information to construct incremental data set training samples

[0129] First, according to the spatial coordinate data of the epoch positioning carrier corresponding to the grid prediction model used for positioning, the input features of the observation quantity are extracted, including: pseudorange residual δ, carrier-to-noise ratio C / N 0 、Satellite altitude angle θ e and the satellite azimuth θ a , the incremental training sample can be expressed as [δ,C / N 0 ,θ e ,θ a ] c .

[0130] The pseudorange inversion process of the original data set training sample is the same. The pseudorange error value obtained by inverting the reference coordinate information and the spatial coordinate data of the epoch positioning result is calibrated with the extracted characteristic variables. After calibrating each set of characteristic variables with the corresponding multipath error, the incremental training sample [δ, C / N 0 ,θ e ,θ a ,Δρ] c .

[0131] The incremental data set also needs to be divided according to the GPS and BDS constellation types and grids before the next step of training. Therefore, the incremental data sets are expressed as: [δ, C / N 0 ,θ e ,θ a ,Δρ] c_GPS_grid and [δ,C / N 0 ,θ e ,θ a ,Δρ] c_BDS_grid .

[0132] On this basis, the incremental training samples are processed as batch samples for training. The relationship between samples can be screened from a large number of data samples to obtain representative samples and avoid the dynamic change problem of data distribution in a single sample. Therefore, the incremental training samples need to be matched into multiple new samples according to the grid layer and constellation type as batch training samples.

[0133] 2. Perform grid-by-grid incremental SVR multipath error model training

[0134] Determine the grid layer to which the sample data in the original training data set belongs. Assume that the center point of the hexagon of the three-dimensional grid layout is o, and the x-axis and y-axis correspond to the E-axis and N-axis in the ENU coordinate system respectively. The following two conditions are used to determine whether it is located inside the hexagonal grid:

[0135] y o ≥l or

[0136]

[0137] Only when both of the above two equations are satisfied, the position is determined to be within the grid column, and each sample in the original training data set is traversed, and the training samples are divided into the corresponding grid columns. According to the height data of each training sample in the original training data set, the training samples are divided into the corresponding grid layers, and then SVR regression training is performed for each grid layer to obtain the SVR prediction rule, that is, the original grid model prediction rule, such as Figure 2 shown.

[0138] For the original sample set, SVR can be understood as solving a convex quadratic programming problem in multiple coefficients equal to the number of training samples N. The obtained prediction function f(X) should be as close as possible to the true value Y. Assuming that the tolerable deviation between f(X) and Y is ε, and Y is a continuous value in SVR, the corresponding optimization problem is expressed as:

[0139]

[0140] Among them, C is the regularization coefficient and the loss function is Introducing the slack variable ξ i and SVR can be further expressed as the following optimization problem:

[0141]

[0142] st(ω T ·X+b)-Y i ≤ε+ξ i

[0143]

[0144]

[0145] After introducing the Lagrange coefficient a and converting the optimization problem into a dual form, we get:

[0146]

[0147]

[0148] Satisfy the following KKT conditions (Karush-Kuhn-Tucker optimization conditions), which are necessary and sufficient conditions for solving convex quadratic programming problems:

[0149] a i (f(Xi )-Y-ε-ξ i )=0

[0150]

[0151]

[0152] After considering nonlinear mapping, the SVR prediction rule is:

[0153]

[0154] Among them, K(,) represents the kernel function operation, a i , is the Lagrange multiplier and b is the deviation term.

[0155] Likewise, if Figure 3 As shown in Figure 1, according to the grid layer to which the incremental data set training samples belong, the prediction rules of the corresponding grid layer can be updated and optimized, and the batch incremental learning samples are matched one by one to the grid layer where the incremental data is located, and incremental training is performed together with the original training samples in the grid layer. c When it arrives, first let Δa c =0, and then use the current SVR prediction rule to determine whether it violates the above KKT condition; if it does not violate the KKT condition, it is determined that it does not affect the SVR prediction rule and the model does not need to be updated; if it violates the KKT condition, the sample may also cause the transfer of historical sample sets, and ultimately still achieve the purpose of re-satisfying the KKT condition. The current SVR prediction model is updated accordingly, and the update process is expressed as follows:

[0156] The SVR dual form is written in Lagrangian form and optimized to derive the KKT condition as follows:

[0157]

[0158] Since according to the KKT condition, a i , At most one is non-zero and both are non-negative, so the bias coefficient and sample error can be defined as:

[0159]

[0160] Combined with the SVR original training sample prediction rule, the regression problem sample set can be divided into five parts:

[0161]

[0162] On this basis, the sample set can be further synthesized into the following three parts by determining whether the vector is supported and whether the boundary supports the vector:

[0163] (1) E set: boundary support vector set

[0164]

[0165] (2) S set: support vector set

[0166]

[0167] (3) R set: non-support vector set

[0168] R={i|θ i =0}

[0169] When the support vector regression machine learning process is over, all samples will enter one of these three sets. Therefore, in the process of incremental learning, the above conditions can be used to screen new samples. When a new sample belongs to the R set, it can be directly discarded; only when the sample belongs to the E set or the S set, the sample is retained, thereby ensuring that the conditions are met when the new sample is added, completing the SVR incremental grid real-time correction model training process, and obtaining the new model prediction rules obtained based on the batch incremental samples.

[0170] When multiple Lagrange multipliers are selected for optimization, the multiplier that violates the KKT condition the most is selected, and the new Lagrange multiplier value updates the model parameters as follows:

[0171]

[0172] The final optimized SVR prediction rule is as follows:

[0173]

[0174] Among them, f′ is the updated prediction function, which is the final regression curve determined by the support vector obtained by balance judgment. is the multipath error prediction value updated by batch incremental mode;

[0175] 3. Complete the construction of multipath heat map with real-time dynamic update

[0176] At this point, the real-time dynamic update multipath heat map can be constructed, including: the original grid multipath error prediction model and grid model prediction rules, as well as the optimized grid prediction model and rules. The regional grid prediction model can be dynamically updated in real time based on the original positioning results.

[0177] (3) Real-time dynamic update of multipath heat map assisted GNSS / IMU integrated navigation solution

[0178] The input features and PDOP (Position Dilution of Precision) values ​​are extracted from the original observations of the GNSS receiver, and according to the initial position obtained by the IMU mechanical arrangement, the real-time dynamically updated multipath heat map is called to match the initial position to the corresponding grid layer and the corresponding three-dimensional grid prediction model.

[0179] The following formula is used to determine whether the grid model is available according to the test data quality. If the PDOP value is greater than or equal to the set threshold, no correction is performed, and the GNSS / IMU tightly combined Kalman filter is directly used to solve the position solution, and the position solution is used as the final solution; if the PDOP value is less than the threshold, the grid multipath error model is used to predict the pseudorange error, and the next step of correction is performed.

[0180] PDOP <k

[0181] If the prediction step is executed, the multipath error value obtained by dynamically updating the prediction model of the adjacent grid is extracted, and the weight W is assigned to the adjacent grid. n , and the final pseudorange error prediction value is obtained after dynamic weighting processing It is expressed as follows:

[0182]

[0183] Among them, n represents the adjacent grid prediction model number, m represents the number of adjacent grid prediction models, and the weight value W is assigned n The size increases with the number of iterations of the corresponding grid prediction model.

[0184] The prediction results obtained by the multipath error prediction model based on the regional grid change continuously with the improvement of grid accuracy. Therefore, the pseudorange error prediction value can be used to construct the adaptive factor f(·) to adjust the GNSS / IMU tight combination measurement noise matrix R k .

[0185] f(·)=F(·) 2

[0186]

[0187] Where X = [δ, C / N 0 ,θ e ,θ a ].

[0188] Adjust the measurement noise matrix R k The carrier position can then be calculated based on the Kalman filter, and the positioning result of each epoch can be output, thus completing the combined navigation positioning solution based on the grid real-time incremental feedback correction.

[0189] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and when the computer program is executed by the data processing unit, the invention content of a combined navigation and positioning method for extreme weather rescue provided by the present invention and some or all of the steps in each embodiment can be executed. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0190] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on such an understanding, the technical solutions in the embodiments of the present invention can be essentially or partly contributed to the prior art in the form of computer programs, i.e., software products, which can be stored in a storage medium and include several instructions for enabling a device including a data processing unit (which can be a personal computer, a server, a single-chip microcomputer, an MCU or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0191] The present invention provides a method and idea for a combined navigation and positioning method for extreme weather rescue. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A combined navigation and positioning method for extreme weather rescue, characterized in that: The following steps are involved: Step 1, construct the original training data set; Step 2: Add incremental data set training samples to the original training data set, build an incremental grid real-time correction model, and obtain a real-time dynamically updated multipath heat map; Step 3: Perform auxiliary GNSS / IMU tight combination solution based on the real-time dynamically updated multipath heat map to complete the combined navigation and positioning for extreme weather rescue; The construction of the original training data set described in step 1 specifically includes: Step 1-1, calibrate the training data set; Step 1-2, constructing a regional grid layout; Step 1-3, dividing the sample data in the training data set according to the satellite constellation types of GPS satellites and Beidou satellites; And according to the reference position corresponding to the sample data, according to the regional grid layout described in step 1-2, the grid layer to which the sample data belongs is determined, and the original training data set is obtained, which is represented as: GPS original training data set [δ,C / N0,θ e ,θ a ,Δρ] GPS_grid and BeiDou original training data set [δ,C / N0,θ e ,θ a ,Δρ] BDS_grid ; The calibration training data set described in step 1-1 is as follows: Step 1-1-1, the positioning carrier repeatedly drives along a predetermined route in an urban area to collect prior data for constructing an original training data set; Step 1-1-2, differential positioning is performed through a high-precision integrated navigation and positioning system and base station GNSS observation data to obtain reference coordinate information; Step 1-1-3, collect the original observations output by the GNSS receiver, and extract the corresponding input features in the original observations, which include: pseudorange residual δ, carrier-to-noise ratio C / N0, satellite altitude angle θ e and the satellite azimuth θ a , a set of characteristic variables is expressed as: [δ, C / N0, θ e ,θ a ]; Step 1-1-4, invert the reference coordinate information obtained in step 1-1-2 and the original GNSS observations collected in step 1-1-3 to obtain the pseudorange error value Δρ, as follows: Δρ=c(Δδt u -Δδt s )+ΔI+ΔT+ε Where c is the speed of light in a vacuum; Δδt s represents the residual after correction of satellite clock error; Δδt u represents the residual after correction of the receiver clock error; ΔI represents the residual after incomplete correction of ionospheric delay; ΔT represents the residual after incomplete correction of tropospheric delay; And the above pseudo-range error is regarded as a multipath error; Step 1-1-5, calibrate the pseudorange error value obtained in step 1-1-4 with the characteristic variable extracted in step 1-1-3 to obtain labeled sample data to form a training data set [δ, C / N0, θ e ,θ a ,Δρ]; The construction of the regional grid layout described in step 1-2 specifically includes: Divide the two-dimensional plane grid, that is, according to the geographic information of the urban area and the GNSS location data, take the smallest square A of the area containing the location data as the modeling area, one side of the square is parallel to the E axis in the ENU coordinate system, divide the hexagonal grid inside the square area, and divide it in units of the ENU coordinates of the hexagon, set the side length of the hexagon to l, and take the center point of the hexagonal grid as o; The two-dimensional plane grid divided above is extended into grid columns in the height direction, and each grid column is divided into n grid layers according to the selected height h in turn, so as to complete the construction of the three-dimensional grid layout; The construction of the incremental grid real-time correction model described in step 2 specifically includes: Step 2-1: Construct a GNSS / IMU integrated navigation solution feedback update mechanism, as follows: Match and call the existing three-dimensional grid prediction model to obtain the multipath error prediction value of the current epoch and determine whether it meets the following conditions: in, Represents the absolute value of the multipath error prediction value at the current epoch, It represents the average absolute pseudorange error value of GNSS data in open areas; If satisfied, the original grid prediction model and prediction rules obtained after training in step 2-2-2 are fed back to update, as follows: The temporal and spatial information of the current epoch positioning carrier are respectively recorded as the current epoch T i and the current position P i , where the current position P i Including the coordinate information of the ENU coordinate system where the current epoch positioning carrier is located, as the real-time data for correcting the grid prediction model at the current time i; Otherwise, do not update; Step 2-2, based on the three-dimensional grid layout, multipath error modeling of incremental SVR is performed grid by grid, that is, a multipath heat map that is dynamically updated in real time is constructed; The construction of the real-time dynamically updated multipath heat map described in step 2-2 is as follows: Step 2-2-1, construct incremental data set training samples, as follows: According to the spatial coordinates of the current epoch positioning carrier, search for the grid layer to which it belongs in the three-dimensional grid layout, and extract the input features of the observation quantity, including: pseudorange residual δ, carrier-to-noise ratio C / N0, satellite altitude angle θ e and the satellite azimuth θ a , get the incremental training samples, expressed as: [δ,C / N0,θ e ,θ a ] c ; Invert and calibrate according to the methods in steps 1-1-4 and 1-1-5 to obtain the incremental data set, expressed as: [δ, C / N0, θ e ,θ a ,Δρ] c ; According to the method in steps 1-3, the incremental data set is divided according to the GPS and BDS constellation types and grids to obtain the GPS incremental data set and the Beidou incremental data set, which are expressed as: [δ, C / N0, θ e ,θ a ,Δρ] c_GPS_grid and [δ,C / N0,θ e ,θ a ,Δρ] c_BDS_grid ; Step 2-2-2, perform incremental SVR multipath error training for each grid to obtain the original grid prediction model, specifically including: Determine the grid layer to which the sample data in the original training data set belongs. Assume that the center point of the hexagon in the three-dimensional grid layout is o, and the x-axis and y-axis correspond to the E-axis and the N-axis in the ENU coordinate system respectively. The judgment is made by the following two conditions: Condition 1, y o ≥l or Condition 2, Only when the above two equations are satisfied, the position of the sample data (x o ,y o ) in the grid column; Traverse each sample in the original training data set and divide it into the corresponding grid column. According to the height data of each sample in the original training data set, divide it into the corresponding grid layer, and then perform SVR regression training on each grid layer to obtain the SVR prediction rule, that is, the original grid model prediction rule; Step 2-2-3, update the original grid prediction model and prediction rules, that is, according to the grid layer to which the incremental data set training samples described in step 2-2-1 belong, update and optimize the SVR prediction rules of the corresponding grid layer, that is, match the incremental data set training samples one by one to the grid layer to which they belong, introduce multiple new training samples and perform incremental training together with the original training samples in the grid layer, and obtain dynamically updated SVR prediction rules, as follows: Assume that the training sample of the incremental data set is X c , let the Lagrange multiplier initialization condition Δa c =0, use the current SVR prediction rule to determine whether it violates the KKT condition, and make the new optimization problem meet the KKT condition again, that is, update the current SVR prediction rule. The update process is as follows: According to the KKT condition, define the deviation coefficient θ i And the sample error h(X i ) are: Combined with the current SVR prediction rule, and based on whether there is a support vector and whether there is a boundary support vector, the incremental data set training samples are divided into three parts: E set, that is, the boundary support vector set: S set, that is, the support vector set: R set, that is, the set of non-support vectors: R={i|θ i =0} During the incremental training process, the above conditions are used to screen the newly added samples. When the newly added samples belong to the R set, the samples are directly discarded. When the samples belong to the E set or the S set, the samples are retained. The final optimized SVR prediction rule is as follows: Among them, f′ is the updated prediction function, which is the final regression curve determined by the support vector obtained by balance judgment. is the multipath error prediction value after updating in batch increment mode, J represents the number of features; Step 2-2-4, completing the construction of a real-time dynamically updated multipath heat map, wherein the real-time dynamically updated multipath heat map includes: an original grid prediction model and prediction rules, and an updated grid prediction model and prediction rules; Among them, KKT conditions: a i (f(X i )-Y i -e-x i )=0 Among them, a i , is the Lagrange multiplier, C is the regularization coefficient, X i represents the i-th group of input features, Y i represents the true value of the multipath error corresponding to the i-th group of features, ξ i and represents the introduced slack variables.

2. The combined navigation and positioning method for extreme weather rescue according to claim 1, characterized in that: The incremental SVR multipath error training for each grid as described in step 2-2-2 specifically includes: Assume that the number of samples in the original training data set is N, and the preset deviation threshold between the predicted value of the prediction function f(X) and the true value Y of the multipath error is ε. The goal of the SVR regression training is to minimize the error between the predicted value and the true value of the prediction function, which is achieved by solving the following optimization problem: st(ω T ·X+b)-Y i ≤e+ξ i Among them, C is the regularization coefficient and the loss function is X i represents the i-th group of input features, X = [δ, C / N0, θ e ,θ a ],Y i represents the true value of the multipath error corresponding to the i-th group of features, ξ i and represents the introduced slack variable, J represents the number of features, ω represents the weight vector, b represents the bias, and the sum represents traversing all data points; After introducing the Lagrange coefficient a and converting the above optimization problem into a dual form, we get: According to the following KKT conditions: a i (f(X i )-Y i -e-x i )=0 After nonlinear mapping, the SVR prediction rule is: Among them, K(,) represents the kernel function operation, a i , is the Lagrange multiplier.

3. The combined navigation and positioning method for extreme weather rescue according to claim 2, characterized in that: The step 3 described in the embodiment of the invention is to perform the assisted GNSS / IMU tight combination solution based on the real-time dynamically updated multipath heat map, specifically including: Step 3-1, in the extreme weather rescue state, extract the input features and position precision factor PDOP value from the original observations of the GNSS receiver, and obtain the initial position according to the IMU mechanical arrangement; Step 3-2, calling the real-time dynamically updated multipath heat map, matching the initial position to the corresponding grid layer, performing position calculation, and obtaining the specific position of the carrier in the extreme weather rescue state.

4. The combined navigation and positioning method for extreme weather rescue according to claim 3, characterized in that: The position solution described in step 3-2 is as follows: If the position precision factor PDOP value is greater than or equal to the threshold, no correction is performed, and the GNSS / IMU tightly combined Kalman filter is directly used to calculate the position solution, and this position solution is used as the final solution; If the PDOP value is less than the threshold, the pseudorange error is predicted using the grid prediction model, and the incremental grid real-time correction model constructed in step 2 is used for the next correction.

5. The combined navigation and positioning method for extreme weather rescue according to claim 4, characterized in that: The step 3-2 of predicting pseudorange errors by using the grid prediction model specifically includes: Extract the multipath error value obtained by the adjacent grid according to the grid prediction model, and assign weight W to the adjacent grid n , and the final multipath error prediction value is obtained after dynamic weighting processing It is expressed as follows: Where n represents the grid prediction model number of the adjacent grid, and m represents the number of grid prediction models of the adjacent grid; The multipath error prediction value is used to construct the adaptive factor f(·) to adjust the GNSS / IMU tight combination measurement noise matrix R k , as follows: f(·)=F(·) 2 Where, X=[δ,C / N0,θ e ,i a ]; Adjust the measurement noise matrix R k Finally, the carrier position is obtained by Kalman filtering.

Images