CN111221011A - GNSS positioning method and device based on machine learning - Google Patents
GNSS positioning method and device based on machine learning Download PDFInfo
- Publication number
- CN111221011A CN111221011A CN201811423812.1A CN201811423812A CN111221011A CN 111221011 A CN111221011 A CN 111221011A CN 201811423812 A CN201811423812 A CN 201811423812A CN 111221011 A CN111221011 A CN 111221011A
- Authority
- CN
- China
- Prior art keywords
- gnss
- scene
- support vector
- positioning
- vector machine
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S19/00—Satellite radio beacon positioning systems; Determining position, velocity or attitude using signals transmitted by such systems
- G01S19/38—Determining a navigation solution using signals transmitted by a satellite radio beacon positioning system
- G01S19/39—Determining a navigation solution using signals transmitted by a satellite radio beacon positioning system the satellite radio beacon positioning system transmitting time-stamped messages, e.g. GPS [Global Positioning System], GLONASS [Global Orbiting Navigation Satellite System] or GALILEO
- G01S19/42—Determining position
Landscapes
- Engineering & Computer Science (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Position Fixing By Use Of Radio Waves (AREA)
Abstract
The invention provides a GNSS positioning method based on machine learning, which comprises the following steps: carrying out initial positioning, and calculating a pseudo-range prior residual error and a Doppler prior residual error; performing quality control on the GNSS observed quantity based on a k-mean clustering algorithm in machine learning; training and learning the parameters of the support vector machine model for scene recognition through training samples, and performing scene recognition after obtaining the corresponding classification function of the support vector machine; training and learning the parameters of the support vector machine model for motion pattern recognition by taking the Doppler prior residual error of the tracked satellite and a known motion pattern as training samples to obtain a corresponding support vector machine classification function, and identifying the motion pattern by taking the Doppler prior residual error of the satellite as the input of the support vector machine model; and adjusting the GNSS observation quality control algorithm based on the result of scene recognition, and adjusting the GNSS positioning strategy based on the result of motion pattern recognition.
Description
Technical Field
The invention relates to the technical field of positioning, in particular to a GNSS positioning method and device based on machine learning.
Background
GNSS (Global Navigation Satellite System) positioning is a mature Navigation positioning means, and has been widely applied to various industries of human beings, and has a profound influence on the daily life of human beings. With the expansion of the application range of GNSS positioning, people also put higher demands on the accuracy of GNSS positioning.
In an open environment, the GNSS signal is good, and the positioning precision can basically meet the general positioning requirement; however, in a complex occlusion environment, due to the influence of multipath signals and the like, a pseudorange/doppler observed quantity often has a large error, and if the observed quantity with the error is directly applied to GNSS positioning calculation, a large offset exists in positioning. Therefore, it is very important to perform quality control on the observed quantity before GNSS positioning calculation.
The prior residual information of the pseudo range/Doppler observed quantity is important information for controlling the quality of the GNSS observed quantity, and the information can accurately reflect the accuracy of the GNSS observed quantity to a certain extent. Clustering operation is respectively carried out on all pseudo-range observed quantity prior residual errors and Doppler observed quantity prior residual errors by using a k-mean clustering algorithm in machine learning, observed quantities corresponding to prior residual error information which obviously deviates from other categories are removed, and therefore the problem that positioning errors are increased due to the fact that the observed quantity information with large errors is used for positioning is solved;
the GNSS positioning has a wide application range, including open scenes such as overhead scenes, under-overhead scenes with serious shielding, and high-building-side scenes with serious multipath. In the GNSS positioning algorithm, it is very important to identify and detect a specific scene first, because the observed quantities of different scenes have different characteristics, different observed quantity quality control calculation logics and different threshold values are required to perform quality control of the observed quantities. The motion state of the GNSS comprises a dynamic mode and a static mode, and different motion modes need different positioning strategies, so that the optimization of GNSS positioning performance can be realized.
The support vector machine is a machine learning algorithm based on a statistical theory, and the upper limit of a model generalization error is reduced and the sample error is reduced by adopting a structure risk minimization criterion, so that the generalization capability of the model is improved. The main advantages of the support vector machine algorithm are the solution of small samples and non-linearity problems. The average carrier-to-noise ratio of the GNSS observation, the number and distribution of the satellite observations, the prior residual information of the observations and the actual scene classification information are used as samples of the support vector machine, a learning model of the support vector machine is established, the support vector is obtained, and the support vector is further applied to scene recognition in GNSS positioning. In addition, dynamic/static mode detection is performed using doppler observation information and velocity information as input samples of the support vector machine. Scene recognition and motion mode detection lay a foundation for differentiation of positioning strategies adopted for different scenes and different motion modes in GNSS positioning, and the aim is to ensure the GNSS comprehensive positioning precision in each scene and each motion state.
The prior art has the following defects:
1) in a complex environment, due to the fact that GNSS observed quantity is influenced by shielding and multipath, the observed quantity has large errors, and GNSS positioning accuracy is poor;
2) the application scenes of GNSS positioning are diversified, GNSS observation quantities under different scenes have different characteristics, and scene identification is needed. GNSS positioning vehicles typically include dynamic and static motion patterns, with different motion patterns requiring different positioning strategies. Scene recognition and motion state detection in GNSS positioning based on the traditional method have certain difficulty in realization.
Disclosure of Invention
The invention provides a GNSS positioning method and a GNSS positioning device based on machine learning, which solve the following technical problems:
1) on the basis of calculating pseudo-range/Doppler prior residual information, introducing a k-mean clustering algorithm in machine learning to perform quality control on observed quantity, and improving the positioning accuracy of the GNSS in a severe environment;
2) establishing a learning model by using different characteristics of GNSS observables under different scenes and adopting a support vector machine algorithm in machine learning to identify and detect the scenes in GNSS positioning;
3) the Doppler information of the GNSS is used as the input of the algorithm of the support vector machine to detect the motion state (static/dynamic) in the GNSS positioning;
4) on the basis of scene identification and motion mode detection, different positioning algorithm strategies are adopted for different scenes and different motion states, and the positioning performance of the GNSS under each scene and each motion state is ensured.
The technical scheme adopted by the invention is as follows:
after initial positioning, calculating pseudo range/Doppler prior residual error information of each observed quantity according to the recursion position at the moment, satellite information and input pseudo range/Doppler observed quantity. Clustering all observation quantity prior residual error information obtained by calculation by adopting a k-mean clustering algorithm, determining an abnormal cluster, and rejecting the observation quantity corresponding to the abnormal cluster according to a set threshold value;
the method comprises the steps of taking a satellite signal-to-noise ratio, the number of tracked satellite particles, the average height angle of all tracked satellites, the mean value and standard deviation of the prior residual error of all satellite pseudo-range observed quantities and a known scene label as training samples, carrying out training learning on support vector machine model parameters, obtaining support vector machine classification functions corresponding to an open scene, an overhead scene and a high-rise side scene, taking the satellite signal-to-noise ratio, the number of tracked satellite particles, the average height angle of all tracked satellites and the mean value and standard deviation of the prior residual error of all satellite pseudo-range observed quantities as input, and identifying and detecting the open scene, the overhead scene and the high-rise side scene in real time in a GNSS positioning process based on a support vector machine model obtained in a training learning stage;
taking Doppler observed quantities of all tracked satellites and a known motion mode as training samples, training and learning support vector machine model parameters, taking the Doppler observed quantities of the satellites as input after obtaining corresponding support vector machine classification functions, and detecting the motion mode in real time in the GNSS positioning process based on the support vector machine model obtained in the training and learning stage;
and carrying out targeted adjustment on related threshold values and positioning strategies in the quality control algorithm on the basis of the scene identification/motion mode.
The invention has the following beneficial technical effects:
1) after GNSS initial positioning, calculating prior residual information of pseudo-range/Doppler observed quantity based on the predicted position, speed and satellite ephemeris information, and laying a foundation for quality control of subsequent observed quantity;
2) clustering the prior residual errors of the pseudo-range/Doppler observed quantities by adopting a k-mean clustering algorithm in machine learning, setting a reasonable threshold, and removing observed quantities corresponding to prior residual error information deviating from other categories without being used for positioning calculation, so that the final positioning accuracy of the GNSS is improved;
3) and carrying out scene identification of GNSS positioning and carrier motion state detection by using a support vector machine algorithm in machine learning. The positioning scenes mainly comprise open environment/elevated scene/high-rise building side scene/tunnel scene and the like, and the motion state of the carrier mainly comprises a dynamic mode and a static mode.
4) On the basis of typical scene identification and motion state detection, different positioning strategies are adopted for different scenes and different motion modes in a GNSS positioning algorithm, so that the GNSS positioning performance under each scene is improved.
Drawings
FIG. 1 is a flow chart of an observed quantity quality control algorithm based on k-means clustering;
FIG. 2 is a flow diagram of support vector machine based open scene recognition;
FIG. 3 is a flow chart of overhead scene recognition based on a support vector machine;
FIG. 4 is a flow chart of the identification of a high-rise building-side scene based on a support vector machine;
FIG. 5 is a flow chart of motion pattern recognition based on a support vector machine;
FIG. 6 is a block diagram of a GNSS positioning apparatus based on machine learning.
Detailed Description
According to the invention, the GNSS positioning scene recognition is introduced into the support vector machine, so that the GNSS positioning scene recognition accuracy is improved, and a foundation is laid for the positioning strategy targeted adjustment based on different scenes; the support vector machine is introduced into GNSS motion mode identification, so that the detection accuracy of a static/dynamic mode in GNSS positioning is improved, and a foundation is laid for the targeted adjustment of positioning strategies based on different motion states; on the basis of GNSS positioning scene identification and motion mode detection, quality control logic, filter parameter setting and output control logic in a positioning algorithm are subjected to targeted adjustment.
The invention is further illustrated below with reference to the figures and examples.
The first embodiment is as follows:
the invention provides a GNSS positioning method based on machine learning, which comprises the following steps:
1) observed quantity quality control algorithm based on k-mean clustering is used for clustering pseudo-range prior residual error and Doppler prior residual error
The common observations in GNSS positioning are mainly pseudorange and doppler observations. In an open environment, errors of a pseudo range mainly include an ionosphere, troposphere delay, ephemeris error, satellite clock error and the like. These errors can be eliminated by means of a specific model or differential positioning. However, in a severe and complex environment, due to the influence of multipath signals on GNSS signals, large multipath errors exist in pseudorange and doppler observations, and the errors are difficult to eliminate through a model or a differential mode. If these observations with errors are directly used for positioning, positioning deviation will be caused, so it is necessary to detect these observations with errors by effective quality control means and eliminate them.
The pseudorange/Doppler observed quantity prior residual of the GNSS is important basis and information for carrying out observed quantity quality control, wherein the computation method of the pseudorange prior residual is as follows:
(1) calculating the linear distance between the satellite and the user according to the satellite position and the predicted position of the user at the moment, namely calculating the pseudo distance:
in the formula (1)
Is the position coordinate of the ith satellite, (x)u，yu，zu) Is based on the previous oneAnd the user position and the speed at the moment are recurrently obtained to obtain the user position coordinate at the moment.
And (2) subtracting the calculated pseudo range from the pseudo range observed quantity, and compensating the local clock error to obtain the prior residual error of the pseudo range:
ρ of the formula (2)iIn order to compensate pseudo-range observed quantity of the ith satellite with troposphere error, ionosphere error and satellite clock error, dT is local clock error;
the doppler observations a priori residuals may be obtained with a similar calculation method.
When a GNSS positioning carrier enters a severe complex environment from an open environment, the error of pseudo-range/Doppler observed quantity can be immediately reflected on pseudo-range/Doppler prior residual error, and the greater the error of the observed quantity is, the more the corresponding prior residual error generally deviates from the prior residual error of the observed quantity of other satellites, so that the quality control of the observed quantity can be performed according to the prior residual error checking value of the observed quantity. The k-means clustering algorithm in machine learning is adopted to cluster the prior residuals of all satellite pseudo ranges/Doppler observations at a certain moment, the satellite pseudo ranges/Doppler observations are divided into three cluster classes, the pseudo range observations corresponding to clusters which are far away from other cluster classes and have fewer members in the cluster classes are considered to have larger errors, and the clusters are considered to be removed.
The k-means clustering algorithm is a basic and should be extensive algorithm in a clustering analysis method. The algorithm firstly needs to appoint the category number k and select k points as initial clustering centers. Based on a given clustering objective function, the algorithm adopts an iterative updating method, each iteration process is carried out in the direction of reducing the objective function value, and the final clustering result is that the objective function value obtains a minimum value, so that the optimal clustering effect is achieved.
The specific process of applying the k-means clustering algorithm to pseudo-range observed quantity quality control is shown in fig. 1, and the specific steps are detailed as follows:
(1) setting the category of the cluster as 3 categories;
(2) assuming pseudorange apriori residuals for all satellitesN data objects of composition are xiI is 1, 2 … n, and the maximum value, the minimum value and the intermediate value are respectively selected as the initial clustering center m of each classj，j＝1，2，3；
(3) Calculating the distance between each data object and the cluster center, and if the following conditions are met:
d(xi，mj)＝min{d(xi，mj)}，j＝1，2，3 (3)
then the data is classified into the category corresponding to the cluster center: x is the number ofi∈wj。
(4) Computing a sum of squared errors based criterion function J corresponding to the 3 new cluster centersCThe value of (c):
(5) judgment JCIf it is less than a certain threshold or it is relative to last JCIf the change amount of (2) is less than a certain threshold value, the algorithm is terminated, otherwise, the steps (3), (4) and (5) are repeated;
(6) and taking the number of data and the average carrier-to-noise ratio corresponding to each category as a basis and a judgment standard, and performing quality division on three prior residual cluster categories obtained based on k-means clustering in the steps, wherein the observation error corresponding to the category with the most data objects and the highest average carrier-to-noise ratio is considered to be the smallest and set as the category A, the observation error corresponding to the category with the least data objects and the lowest average carrier-to-noise ratio is considered to be the largest and set as the category C, and the rest category is the category B.
(7) Setting a proper threshold, and when the difference value of the mean values of the data objects (prior residuals) in the class C and all the data objects (prior residuals) in the class A is greater than a set threshold 1 (rejection threshold), considering that a large error exists in the corresponding pseudo-range observed quantity, and removing the large error; when the difference value of the mean value of the data objects (prior residuals) in the class C and all the data objects (prior residuals) in the class A is larger than a set threshold value 2 (weight reduction threshold value), the corresponding pseudo-range observed quantity is considered to have a large error, a flag bit is set, and the pseudo-range observed quantity is used for subsequent positioning calculation after weight reduction processing is carried out. The principle and steps of applying the k-means clustering algorithm to the Doppler observed quantity quality control are similar to those described above, and are not described in detail.
2) Support vector machine applied to GNSS positioning for scene recognition
The specific application scenarios of GNSS positioning are wide and various, and typical scenarios mainly include: open environment, high-rise scene with serious shielding and high-building side scene with serious multipath. The observed quantity of the GNSS in different scenes can show different characteristics, for example, the observed quantity in an open environment has better quality and smaller error, while the observed quantity in a multipath environment is seriously influenced by multipath and has larger error. Therefore, if the scene identification and detection can be carried out in real time in the GNSS positioning process, the quality control logic and the positioning strategy can be adjusted in a targeted manner according to the characteristics of the observed quantity in different scenes on the basis, and the comprehensive positioning performance of the GNSS in different scenes can be improved.
The signal-to-noise ratio intensity, the number and the distribution of tracked satellite particles and the characteristics of the observation amount prior residual error under three typical scenes of GNSS positioning are obviously different, for example, the signal-to-noise ratio is strongest in an open environment, the number of satellites is the most, and the observation amount prior residual error is generally smaller; and the signal-to-noise ratio of the scene satellite under the overhead with serious shielding is lower, the number of tracked satellites is less, and the altitude angle of the tracked satellite is lower. The number of satellites and the signal-to-noise ratio of a high-rise side scene with serious multipath are between the two scenes, but due to the influence of multipath signals, the prior residual error of the observed quantity is generally large. Therefore, based on the characteristics of the analysis, the GNSS positioning scene identification and detection are carried out by taking the signal-to-noise ratio of the satellite, the number of the tracked satellite particles, the average height angle of all tracked satellites, and the average value and standard deviation of the prior residuals of all satellite pseudo-range observations as input.
The support vector machine is a general learning method of a statistical theory, and is a tool with great potential classification and regression. The support vector machine minimizes the structural risk as a criterion, minimizes the error of a sample point, and reduces the complexity of the model, namely reduces the generalization error of the model, thereby improving the generalization capability of the model; meanwhile, the support vector machine skillfully adopts a kernel function to map data to a high-dimensional space to solve the problem of inseparability of linearity in an original space and avoid the problems of computation complexity and dimension disaster; the support vector machine finally converts the problem into a quadratic programming problem for solving a linear constraint through conversion, and a unique global optimal solution exists. Currently, support vector machines have been practically studied and have shown superiority over conventional learning algorithms in many fields, such as pattern recognition, regression estimation, text classification, data mining, etc.
The support vector machine is a two-classification model in popular terms, a basic model of the support vector machine is defined as a linear classifier with the maximum interval on a feature space, and a learning strategy of the support vector machine is interval maximization and can be finally converted into the solution of a convex quadratic programming problem. Models of support vector machines include linear separable models and nonlinear separable models:
(1) linear separable model of support vector machine
Assume a set of training samples { (x)1，y1)，…(xN，yN)，x∈RnAnd y belongs to R, and the classification function with the classification interval of epsilon is set as follows:
y＝f(x)＝wTx+b
the requirements are satisfied:
the function is normalized, all samples meet | f (x) | equal to or more than 1, and the sample closest to the interface meets, so that the classification interval is 2/| | w |. Therefore, the maximum classification interval is the minimum | | w | |, and the interface is required to correctly implement classification, and the following requirements are met: y isi(w.xi+b)≥1，i＝1，…n。
Therefore, the support vector machine seeks an optimal interface, which is a quadratic programming problem requiring the solution of:
the above problem can be changed to an optimization problem of dual variables through lagrange duality, that is, an optimal solution of the original problem is obtained by solving a dual problem equivalent to the original problem. After lagrange transformation, the problem to be solved can be transformed into:
α thereiniIs a Lagrange multiplier, the optimization problem is solved to obtain an optimal solution α*＝[α1，α2，…αN]T. The value of the interface is thus calculated:
and solving w and b, namely finally obtaining the interface function.
(2) Nonlinear support vector machine
Most problems encountered in reality are nonlinear problems, and for the nonlinear problems, a processing method of a support vector machine is to introduce a kernel function, solve the nonlinear problems in an original space by mapping data to a high-dimensional space, only involve an inner product operation form among samples, and avoid the complexity of high-dimensional operation. The kernel function is equivalent to the original classification function:
mapping into:
in addition to solving the non-linear problem with the kernel function, the support vector model introduces a relaxation variable ξiTo allow the existence of misclassified samples due to outlier points, thereby improving the generalization capability of the model. Combining the above factors to form the optimal super-linearity under the nonlinear conditionThe objective function of the plane is:
where C is a parameter used to control the weight between the two terms in the objective function ("find hyperplane with largest classification interval" and "guarantee minimum deviation of data points"). Using the lagrange method and the dual principle, the above objective function can be transformed into:
the corresponding optimal hyperplane equation is:
the decision function is g (x) sign (f (x)).
Commonly used kernel functions are:
linear kernel function: k (x)1，x2)＝<x1，x2>
Polynomial kernel function: k (x)1，x2)＝(<x1，x2>+R)d
Gaussian radial basis kernel function: k (x)1，x2)＝exp(-||x1-x2||2/2σ2)
(3) Specific implementation of GNSS positioning scene recognition based on support vector machine
Fig. 2, fig. 3, and fig. 4 show flowcharts of the application of the support vector machine to GNSS positioning in open scene, under-overhead scene, and high-rise building-side scene recognition, respectively. When the support vector machine is applied to GNSS positioning scene recognition, the method specifically comprises two stages, namely a training and learning stage to obtain model parameters (classification functions), wherein the process is generally carried out off-line; and the second is a specific application stage, namely, the specific identification of the scene is carried out in real time in the positioning process according to the support vector machine model and the parameters obtained in the learning stage.
And in the training and learning stage, the actual drive test data is used as a training sample, the training sample comprises the number of satellites tracked at each time point, the average carrier-to-noise ratio of the tracked satellites, the average altitude angle of the tracked satellites, the average value and the standard deviation of pseudo-range observed quantity prior residuals of all the tracked satellites are used as input samples of the support vector machine, and the known specific scene label at each time point is used as an output sample for training the support vector machine. Considering the nonlinear relationship between the input and the output, the kernel function selects a gaussian radial basis function:
K(x1，x2)＝exp(-||x1-x2||2/2σ2) And selecting the appropriate slack variable ξiAnd a parameter C, learning the support vector machine model in an off-line manner to obtain an optimal learning parameter, namely a classification function of the support vector machine:
since three typical scenarios need to be performed: the identification of scenes in a wide environment, under an overhead and beside a high building needs to be carried out, so that the learning of the support vector machine models needs to be carried out on the three scenes respectively to obtain corresponding three support vector machine classification functions.
In the specific application stage of scene identification, at each moment in the real-time GNSS positioning process, the number of tracked satellites, the average carrier-to-noise ratio of the tracked satellites, the average altitude angle of the tracked satellites, and the mean value and standard deviation of pseudo-range observed quantity prior residuals of all the tracked satellites are used as input, and according to a support vector machine model (classification function) obtained in the training and learning stage, the specific positioning scene is identified in real time, including the identification of three typical scenes:
(1) based on the classification function of the support vector machine corresponding to the open environment recognition obtained in the learning stage, and calculating the value of the classification function according to the input, and simultaneously obtaining the value of the corresponding decision function: g (x) sign (f (x)). If g (x) is 1, it is determined as an open scene, and if g (x) is-1, it is determined as an open scene;
(2) based on the classification function of the support vector machine corresponding to the scene recognition under the overhead acquired in the learning stage, the value of the classification function is calculated according to the input, and the value corresponding to the decision function is acquired at the same time: g (x) sign (f (x)). If g (x) is 1, determining as an elevated scene, and if g (x) is-1, determining as a non-elevated scene;
(3) based on a support vector machine classification function which is obtained in a learning stage and corresponds to the high-rise building side scene recognition, calculating the value of the classification function according to input, and simultaneously obtaining the value of a corresponding decision function: g (x) sign (f (x)). If g (x) is 1, it is determined as a high-rise side scene, and if g (x) is-1, it is determined as a non-high-rise side scene.
3) Motion state detection of support vector machine applied to GNSS positioning
The motion state of the GNSS positioning carrier includes a static mode and a dynamic mode, and for different motion modes, different positioning strategies can be adopted in the GNSS positioning process, for example, a corresponding constraint equation is added in the static mode, which is beneficial to improving the positioning accuracy of GNSS.
Considering that the doppler observables in the GNSS observables are information closely related to the velocity of the carrier, the motion state detection of the GNSS positioning carrier is performed with the doppler observables of the GNSS as input to the support vector machine. FIG. 5 is a flow chart illustrating the application of a support vector machine to motion pattern detection for GNSS positioning. The method specifically comprises two stages, namely a training stage and an application stage.
In the learning training stage, actual drive test data is used as a training sample, Doppler observed quantities of all satellites at each moment are used as input samples, the known motion state (static/dynamic) of the positioning carrier at each moment is used as an output sample, and in consideration of the nonlinear relation between input and output, a kernel function selects a Gaussian radial basis function:
K(x1，x2)＝exp(-||x1-x2||2/2σ2)，
selecting proper parameters sigma and relaxation variables ξiAnd a parameter C, learning the support vector machine model in an off-line manner to obtain the optimal support vector machine model parameter, namely the classification function of the support vector machine:
in the phase of detecting the motion state based on the support vector machine, that is, in the GNSS real-time positioning process, all GNSS doppler observations at each time are used as the input of the support vector machine model (classification function) obtained in the learning phase, and the motion state is detected:
i.e. calculating the value of the classification function according to the input and obtaining the value corresponding to the decision function: g (x) sign (f (x)). If g (x) is 1, the mode is determined as the static mode, and if g (x) is-1, the mode is determined as the dynamic mode.
4) Positioning strategy adjustment based on scene recognition/motion detection mode detection
Because GNSS observations in different scenes have different characteristics, on the basis of scene identification, an observation processing strategy needs to be adjusted in a targeted manner, including:
(1) adjusting the relevant threshold value in the quality control algorithm according to different scene modes;
(2) adjusting the parameter settings of related filters in the positioning filter according to different scene modes, wherein the parameter settings comprise a state driving noise matrix, a measurement noise matrix and the like;
(3) and performing targeted setting on the output position integrity confidence coefficient calculation logic according to the scene recognition result.
On the basis of the motion pattern detection, the positioning strategy is adjusted in a targeted manner, and the method comprises the following steps:
(1) the setting of relevant thresholds in the quality control algorithm is different under different motion states (static/dynamic);
(2) the initial point positioning output strategies are different under different motion states (static/dynamic);
(3) in the static mode, the speed/position constraint needs to be increased, and the positioning precision in the static state is ensured.
Example two:
the present invention further provides a GNSS positioning apparatus based on machine learning, as shown in fig. 6, including:
the observation quantity preprocessing module is used for preprocessing the original observation quantity input and laying a foundation for a subsequent processing module;
the satellite information calculation module is mainly used for calculating information such as satellite position, speed, direction cosine and the like according to the ephemeris;
the observation quality control module is used for performing quality control on the GNSS observation, detecting and rejecting the observation with errors, and reserving the observation with good quality for subsequent positioning and resolving;
the scene identification module is used for identifying specific scenes of each stage in the GNSS continuous positioning process, including open scenes, off-shelf scenes, high-rise building-side scenes and the like;
the motion state detection module is used for detecting the motion state of the GNSS application carrier, including static and dynamic modes;
and the positioning filter module is used for performing position/speed calculation by utilizing observed quantity information such as pseudo range and the like and information such as satellite position and the like entering the filter.
Example three:
the invention also provides a memory, in which a computer program is stored, the computer program performing the steps of:
performing initial positioning, calculating pseudo-range prior residual and Doppler prior residual of the GNSS observed quantity after the initial positioning is finished, and introducing a k-mean clustering algorithm in machine learning to perform quality control on the GNSS observed quantity on the basis of the prior residual checking calculation;
taking the signal-to-noise ratio of the satellites, the number of tracked satellite particles, the average height angle of all tracked satellites, the mean value and standard deviation of pseudo-range prior residuals of all satellites and known scene labels as training samples, carrying out training learning on support vector machine model parameters, obtaining a support vector machine classification function corresponding to a scene to be identified, and then taking the signal-to-noise ratio of the satellites, the number of tracked satellite particles, the average height angle of all tracked satellites and the mean value and standard deviation of all satellite pseudo-range prior residuals as input to carry out scene identification;
training and learning support vector machine model parameters by taking Doppler prior residual errors of all tracked satellites and a known motion mode as training samples, obtaining corresponding support vector machine classification functions, and identifying the motion mode by taking the Doppler prior residual errors of the satellites as input;
the GNSS observation quality control strategy and the positioning strategy are adjusted based on scene identification, and the GNSS positioning method is adjusted based on motion pattern identification.
Although the present invention has been described with reference to the preferred embodiments, it is not intended to limit the present invention, and those skilled in the art can make variations and modifications of the present invention without departing from the spirit and scope of the present invention by using the methods and technical contents disclosed above.
Claims (13)
1. A GNSS positioning method based on machine learning is characterized in that the GNSS positioning method comprises the following steps:
performing initial positioning, calculating pseudo-range prior residual and Doppler prior residual of GNSS observed quantity after the initial positioning is finished, introducing a k-mean clustering algorithm in machine learning to perform quality control on the GNSS observed quantity on the basis, eliminating the observed quantity with larger error, and avoiding using the observed quantity for final positioning calculation;
taking the signal-to-noise ratio of the satellites, the number of tracked satellite particles, the average height angle of all tracked satellites, the mean value and standard deviation of pseudo-range prior residuals of all satellites and known scene labels as training samples, carrying out training learning on support vector machine model parameters, obtaining a support vector machine classification function corresponding to a scene to be identified, and then taking the signal-to-noise ratio of the satellites, the number of tracked satellite particles, the average height angle of all tracked satellites and the mean value and standard deviation of all satellite pseudo-range prior residuals as input to carry out scene identification;
training and learning support vector machine model parameters by taking Doppler prior residuals of all tracked satellites and a known motion mode as training samples, obtaining a support vector machine classification function corresponding to the motion mode, and identifying the motion mode by taking the Doppler prior residuals of the satellites as input;
the GNSS observation amount is adjusted based on scene identification, and the GNSS positioning method is adjusted based on motion pattern identification.
2. The machine-learning-based GNSS positioning method of claim 1, wherein after initial positioning, the quality of GNSS observations is controlled using a k-mean-based clustering algorithm on the basis of calculating pseudorange and doppler prior residuals.
3. The method as claimed in claim 2, wherein the method calculates the value of the classification function based on the classification function of the support vector machine corresponding to the scene to be identified, and obtains the value of the corresponding decision function, and performs the scene identification according to the value of the decision function.
4. The method as claimed in claim 2, wherein the method calculates the value of the classification function based on the classification function of the support vector machine corresponding to the motion pattern, and obtains the value of the corresponding decision function, and performs motion pattern recognition according to the value of the decision function.
5. The GNSS positioning method based on machine learning of claim 2, wherein a quality control algorithm based on k-mean clustering is used to cluster pseudo-range prior residuals and doppler prior residuals of GNSS observations obtained by calculation, determine an abnormal cluster class, and reject GNSS observations corresponding to the abnormal cluster class according to a set threshold.
6. The machine-learning-based GNSS positioning method of claim 5, wherein the scenes to be identified include an open scene, an under-overhead scene, and a high-rise side scene.
7. The method as claimed in claim 6, wherein the open scene, the under-elevated scene and the high-rise building scene are identified in real time during the GNSS positioning process based on the SVM mode obtained during the learning stage of the scene recognition training.
8. The method as claimed in claim 6, wherein the motion pattern recognition is performed in real time during the GNSS positioning process based on the support vector machine pattern obtained during the training and learning phase of the motion pattern recognition, and the motion pattern includes a static pattern and a dynamic pattern.
9. The machine-learning-based GNSS positioning method of claim 6, wherein based on the results of scene recognition and motion pattern recognition, a position solution and a velocity solution are performed by a positioning filter, and a solution result is output.
10. The method as claimed in claim 9, wherein the adjusting of the control strategy and the positioning strategy for the quality of GNSS observations based on scene recognition comprises the following steps:
adjusting a threshold value in a quality control algorithm according to different scenes;
adjusting parameter settings in the positioning filter according to different scenes;
and setting the output position integrity confidence coefficient calculation logic according to the scene recognition result.
11. The method as claimed in claim 8, wherein the adapting of the GNSS positioning method based on motion pattern recognition specifically comprises:
setting a threshold value in a quality control algorithm according to different motion modes;
outputting initial point positioning according to different motion state modes;
and the constraints of speed and position are increased in the static mode, so that the positioning accuracy in the static mode is ensured.
12. A machine learning based GNSS positioning apparatus, the apparatus comprising:
the observation quantity preprocessing module is used for preprocessing the original GNSS observation quantity;
the satellite information calculation module is used for calculating a pseudo-range prior residual error and a Doppler prior residual error according to the preprocessed GNSS observed quantity;
the observation quality control module is used for performing quality control on the GNSS observation quantity, and detecting and rejecting the GNSS observation quantity with errors;
the scene recognition module is used for performing scene recognition, and specifically comprises: taking the signal-to-noise ratio of the satellites, the number of tracked satellite particles, the average height angle of all tracked satellites, the mean value and standard deviation of pseudo-range prior residuals of all satellites and known scene labels as training samples, carrying out training learning on support vector machine model parameters, obtaining a support vector machine classification function corresponding to a scene to be identified, and then taking the signal-to-noise ratio of the satellites, the number of tracked satellite particles, the average height angle of all tracked satellites and the mean value and standard deviation of all satellite pseudo-range prior residuals as input to carry out scene identification;
the motion state detection module is used for performing motion pattern recognition, and specifically comprises: training and learning support vector machine model parameters by taking Doppler prior residual errors of all tracked satellites and a known motion mode as training samples, obtaining corresponding support vector machine classification functions, and identifying the motion mode by taking the Doppler prior residual errors of the satellites as input;
and the positioning filter module is used for performing position calculation and speed calculation based on the GNSS observed quantity and the satellite position information.
13. A memory storing a computer program, the computer program performing the steps of:
performing initial positioning, calculating pseudo-range prior residual and Doppler prior residual of the GNSS observed quantity after the initial positioning is finished, and controlling the quality of the GNSS observed quantity by adopting a k-mean-based clustering algorithm;
taking the signal-to-noise ratio of the satellites, the number of tracked satellite particles, the average height angle of all tracked satellites, the mean value and standard deviation of pseudo-range prior residuals of all satellites and known scene labels as training samples, carrying out training learning on support vector machine model parameters, obtaining a support vector machine classification function corresponding to a scene to be identified, and then taking the signal-to-noise ratio of the satellites, the number of tracked satellite particles, the average height angle of all tracked satellites and the mean value and standard deviation of all satellite pseudo-range prior residuals as input to carry out scene identification;
training and learning support vector machine model parameters by taking Doppler prior residual errors of all tracked satellites and a known motion mode as training samples, obtaining corresponding support vector machine classification functions, and identifying the motion mode by taking the Doppler prior residual errors of the satellites as input;
the quality control strategy of the GNSS observation is adjusted based on scene identification, and the GNSS positioning method is adjusted based on motion pattern identification.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811423812.1A CN111221011B (en) | 2018-11-26 | 2018-11-26 | GNSS positioning method and device based on machine learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811423812.1A CN111221011B (en) | 2018-11-26 | 2018-11-26 | GNSS positioning method and device based on machine learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111221011A true CN111221011A (en) | 2020-06-02 |
| CN111221011B CN111221011B (en) | 2022-04-19 |
Family
ID=70807703
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811423812.1A Active CN111221011B (en) | 2018-11-26 | 2018-11-26 | GNSS positioning method and device based on machine learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111221011B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114199262A (en) * | 2020-08-28 | 2022-03-18 | 阿里巴巴集团控股有限公司 | Method for training position recognition model, position recognition method and related equipment |
| CN114252892A (en) * | 2020-09-10 | 2022-03-29 | 阿里巴巴集团控股有限公司 | Training method of machine learning model, interference detection method and device |
| CN114265127A (en) * | 2021-12-23 | 2022-04-01 | 上海海洋大学 | GNSS-R snow detection method based on support vector machine |
| CN114488222A (en) * | 2020-10-28 | 2022-05-13 | 阿里巴巴集团控股有限公司 | Positioning method, positioning device and storage medium |
| WO2022161229A1 (en) * | 2021-01-27 | 2022-08-04 | 腾讯科技（深圳）有限公司 | Error model calibration method and apparatus, electronic device, error model-based positioning method and apparatus, terminal, computer-readable storage medium, and program product |
| WO2024219212A1 (en) * | 2023-04-19 | 2024-10-24 | ソニーグループ株式会社 | Program, information processing device, and information processing method |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100033370A1 (en) * | 2003-12-02 | 2010-02-11 | Gmv Aerospace And Defence, S.A. | Gnss navigation solution integrity in non-controlled environments |
| CN105044744A (en) * | 2015-06-30 | 2015-11-11 | 南京航空航天大学 | Non-cooperative navigation satellite identification method based on Doppler frequency shift characteristic |
| CN106535133A (en) * | 2016-11-11 | 2017-03-22 | 天津大学 | Indoor telephone traffic accurate location method based on machine learning in cellular network |
| CN107966724A (en) * | 2017-11-27 | 2018-04-27 | 南京航空航天大学 | Satellite positioning method in a kind of urban canyons based on 3D city models auxiliary |
| US10024974B1 (en) * | 2014-10-24 | 2018-07-17 | Mediatek Inc. | Method and apparatus applicable to positioning in NLOS environment |
| CN108363079A (en) * | 2018-01-30 | 2018-08-03 | 上海交通大学 | A kind of GNSS pseudorange double difference localization methods and system towards portable intelligent device |
| CN108828640A (en) * | 2018-07-02 | 2018-11-16 | 中国人民解放军战略支援部队信息工程大学 | A kind of satellite navigation and positioning observation weighs method and device surely |
-
2018
- 2018-11-26 CN CN201811423812.1A patent/CN111221011B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100033370A1 (en) * | 2003-12-02 | 2010-02-11 | Gmv Aerospace And Defence, S.A. | Gnss navigation solution integrity in non-controlled environments |
| US10024974B1 (en) * | 2014-10-24 | 2018-07-17 | Mediatek Inc. | Method and apparatus applicable to positioning in NLOS environment |
| CN105044744A (en) * | 2015-06-30 | 2015-11-11 | 南京航空航天大学 | Non-cooperative navigation satellite identification method based on Doppler frequency shift characteristic |
| CN106535133A (en) * | 2016-11-11 | 2017-03-22 | 天津大学 | Indoor telephone traffic accurate location method based on machine learning in cellular network |
| CN107966724A (en) * | 2017-11-27 | 2018-04-27 | 南京航空航天大学 | Satellite positioning method in a kind of urban canyons based on 3D city models auxiliary |
| CN108363079A (en) * | 2018-01-30 | 2018-08-03 | 上海交通大学 | A kind of GNSS pseudorange double difference localization methods and system towards portable intelligent device |
| CN108828640A (en) * | 2018-07-02 | 2018-11-16 | 中国人民解放军战略支援部队信息工程大学 | A kind of satellite navigation and positioning observation weighs method and device surely |
Non-Patent Citations (3)
| Title |
|---|
| LI-TA HSU: "GNSS multipath detection using a machine learning approach", 《2017 IEEE 20TH INTERNATIONAL CONFERENCE ON INTELLIGENT TRANSPORTATION SYSTEMS (ITSC)》 * |
| N. V. MIKHAILOV 等: "Identification and Mitigation of Multipath in GNSS Receivers Using Cluster Analysis Methods", 《GYROSCOPY AND NAVIGATION》 * |
| 颜文娴 等: "组合导航系统容错技术研究进展", 《全球定位系统》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114199262A (en) * | 2020-08-28 | 2022-03-18 | 阿里巴巴集团控股有限公司 | Method for training position recognition model, position recognition method and related equipment |
| CN114252892A (en) * | 2020-09-10 | 2022-03-29 | 阿里巴巴集团控股有限公司 | Training method of machine learning model, interference detection method and device |
| CN114488222A (en) * | 2020-10-28 | 2022-05-13 | 阿里巴巴集团控股有限公司 | Positioning method, positioning device and storage medium |
| WO2022161229A1 (en) * | 2021-01-27 | 2022-08-04 | 腾讯科技（深圳）有限公司 | Error model calibration method and apparatus, electronic device, error model-based positioning method and apparatus, terminal, computer-readable storage medium, and program product |
| EP4206741A4 (en) * | 2021-01-27 | 2024-08-14 | Tencent Tech Shenzhen Co Ltd | Error model calibration method and apparatus, electronic device, error model-based positioning method and apparatus, terminal, computer-readable storage medium, and program product |
| US12130366B2 (en) | 2021-01-27 | 2024-10-29 | Tencent Technology (Shenzhen) Company Limited | Error model calibration method and apparatus, electronic device, error model-based positioning method and apparatus, terminal, computer-readable storage medium, and program product |
| CN114265127A (en) * | 2021-12-23 | 2022-04-01 | 上海海洋大学 | GNSS-R snow detection method based on support vector machine |
| WO2024219212A1 (en) * | 2023-04-19 | 2024-10-24 | ソニーグループ株式会社 | Program, information processing device, and information processing method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111221011B (en) | 2022-04-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111221011B (en) | GNSS positioning method and device based on machine learning | |
| CN108573496B (en) | Multi-target tracking method based on LSTM network and deep reinforcement learning | |
| CN112308881A (en) | Ship multi-target tracking method based on remote sensing image | |
| CN110533691B (en) | Target tracking method, device and storage medium based on multiple classifiers | |
| Jouaber et al. | Nnakf: A neural network adapted kalman filter for target tracking | |
| CN113705590B (en) | Steel structure intelligent polishing control method based on artificial intelligence | |
| CN108168564A (en) | A kind of Data Association based on LHD grey relational grades | |
| CN102853836B (en) | Feedback weight fusion method based on track quality | |
| CN106803265A (en) | Multi-object tracking method based on optical flow method and Kalman filtering | |
| WO2020099024A1 (en) | Apparatus and method for characterizing an object based on measurement samples from one or more location sensors | |
| CN110954132A (en) | Method for carrying out navigation fault identification through GRNN (generalized regression neural network) assisted adaptive Kalman filtering | |
| Liyun et al. | Improved Faster R-CNN algorithm for defect detection in powertrain assembly line | |
| CN102063625B (en) | Improved particle filtering method for multi-target tracking under multiple viewing angles | |
| CN112862858A (en) | Multi-target tracking method based on scene motion information | |
| CN108871365B (en) | State estimation method and system under course constraint | |
| CN111274529B (en) | Robust Gao Sini Weisal PHD multi-expansion target tracking algorithm | |
| CN114137562B (en) | Multi-target tracking method based on improved global nearest neighbor | |
| CN113971688B (en) | Anchor-free multi-target tracking method for enhancing ID re-identification | |
| CN102663773A (en) | Dual-core type adaptive fusion tracking method of video object | |
| CN112880659B (en) | Fusion positioning method based on information probability | |
| CN110412531B (en) | Amplitude information-based receiving station path optimization method under clutter condition | |
| CN115932913B (en) | Satellite positioning pseudo-range correction method and device | |
| CN111487613A (en) | Radar/ESM (electronic stability management) track robust association method based on hierarchical clustering | |
| CN116047480A (en) | External parameter calibration method from laser radar to attitude sensor | |
| Lu et al. | Vision-based real-time road detection in urban traffic |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |