CN115166804A

CN115166804A - GNSS/INS tightly-combined positioning method for predicting measurement noise based on machine learning

Info

Publication number: CN115166804A
Application number: CN202210783091.5A
Authority: CN
Inventors: 叶飞; 焦洪宇; 陈庆樟; 马春; 朱海民
Original assignee: Changshu Institute of Technology
Current assignee: Changshu Institute of Technology
Priority date: 2022-07-05
Filing date: 2022-07-05
Publication date: 2022-10-11

Abstract

The invention discloses a GNSS/INS tight combination positioning method for forecasting measurement noise based on machine learning, which comprises the following steps: in an off-line state, adopting GNSS/INS observation parameters to construct training set data, and using GRU as a noise prediction model to carry out model training; in an online state, a state model is built through a GNSS/INS observation value, meanwhile, the GNSS/INS observation value is used as input of a trained noise prediction model to obtain predicted measurement noise, a measurement model is built through the predicted measurement noise, and finally, the state model and the measurement model are used for filtering estimation processing to output a positioning result. The method can better reflect the change of the environment according to the measured noise array obtained by the regression prediction of the environmental parameters, more effectively adjust the filter gain, realize the self-adaptive adjustment of the noise covariance array and improve the accuracy of actual estimation in the urban environment.

Description

GNSS/INS tightly-combined positioning method for measuring noise prediction based on machine learning

Technical Field

The invention relates to a tight combination positioning method of GNSS/INS, in particular to a tight combination positioning method of GNSS/INS for measuring noise prediction based on machine learning.

Background

The GNSS/INS tightly-combined positioning realizes the estimation of the positioning error by means of a nonlinear filtering algorithm. Kalman filtering can realize a certain regulation effect through a filtering gain array, and the regulation performance is related to the precision of a model. In vehicle-mounted application in urban environment, the change of the measurement noise is not easy to be modeled accurately, and the deviation of an inertia device is not a fixed value, so that the INS device error and the GNSS measurement noise are difficult to be obtained accurately. This means that in a changing dynamic environment, a fixed filter noise model designed using conventional methods such as simple elevation weighting or parameter tuning is deficient. General gaussian filtering adjusts the filter gain through a noise covariance matrix and affects the filtering result. According to the fixed quantity measurement noise array designed according to the parameter adjustment, when complex shielding and interference are faced, the noise parameter can not be adjusted due to the fact that the noise parameter does not change along with environmental factors. Meanwhile, too complex measurement noise changes in the urban environment also cause difficulty in accurately constructing a measurement noise model in a numerical modeling mode in practical application.

Disclosure of Invention

In view of the above-mentioned defects of the prior art, the present invention provides a GNSS/INS tightly-combined positioning method for performing measurement noise prediction based on machine learning, which solves the problem that the positioning accuracy is affected due to inaccurate construction of a measurement noise model.

The technical scheme of the invention is as follows: a GNSS/INS tight combination positioning method for performing measurement noise prediction based on machine learning comprises the following steps: 1. in an off-line state, adopting GNSS/INS observation parameters to construct training set data, and using GRU as a noise prediction model to carry out model training; 2. in an online state, a state model is built through a GNSS/INS observation value, meanwhile, the GNSS/INS observation value is used as input of a trained noise prediction model to obtain predicted measurement noise, a measurement model is built through the predicted measurement noise, and finally, the state model and the measurement model are used for filtering estimation processing to output a positioning result.

Furthermore, the GNSS/INS observation parameters include an altitude angle, an azimuth angle, a carrier observation value, and a signal-to-noise ratio, the altitude angle, the azimuth angle, the carrier observation value, and the signal-to-noise ratio constitute a feature quantity of a training set of the noise prediction model, and a label quantity of the training set of the noise prediction model is constructed by a deviation between a positioning reference value and a positioning result.

Further, the deviation between the positioning reference value and the positioning result is obtained by the following method:

the first three terms of the state vector are taken to represent the real three-dimensional positioning error, P _Ref Denotes a positioning reference value, P _est Representing a positioning result estimated by a GNSS/INS tight combination algorithm to obtain a measurement residual vector:

wherein the content of the first and second substances,

is the measured residual error, Z, of the k epoch inverse calculation _k And the measurement residual vector constitutes the label quantity of the training set of the noise prediction model.

Further, the training set data is subjected to data processing before the noise prediction model is trained, and the data processing comprises deleting data with the number of visible satellites in the training set data being less than 4.

Furthermore, when the measurement model is constructed by the predicted measurement noise, the measurement is obtained by inputting the GNSS/INS observation value into the trained noise prediction modelNoise error

Computing a measured noise covariance matrix

In which only the diagonal parameters are retained.

Compared with the prior art, the invention has the advantages that:

the method comprises the steps of firstly, using a GRU network offline to train a mapping model between characteristic parameters such as signal-to-noise ratio, altitude angle and the like and observation residual errors, then using online predicted observation residual errors to replace traditional parameter adjustment or numerical modeling measurement noise parameters in the process of constructing a measurement model, realizing dynamic updating of a measurement noise covariance matrix, and finally outputting a final result through filtering estimation. The measured noise array obtained according to the regression prediction of the environmental parameters can better reflect the change of the environment, more effectively adjust the filter gain, realize the self-adaptive adjustment of the noise covariance array and improve the accuracy of actual estimation in the urban environment.

Drawings

FIG. 1 is a schematic diagram illustrating steps of a GNSS/INS tight integrated positioning method for performing measurement noise prediction based on machine learning according to an embodiment of the present invention.

Fig. 2 is a 0922 dataset GNSS double-difference wide-lane observations in an implementation.

Fig. 3 illustrates 0418 data set GNSS double-difference wide-lane observations in an implementation.

FIG. 4 is a graph of the mean absolute error of the model training loss function.

FIG. 5 is a validation set model training residual statistics.

FIG. 6 shows the matching of the verification set prediction value and the true value.

Figure 7 is a 1207 data set satellite number for an implementation.

Figure 8 is a 1207 dataset ambiguity fraction portion of an implementation.

Fig. 9 is a probability distribution of a fraction of ambiguity.

FIG. 10 shows the measured residual calculated using the positioning reference.

FIG. 11 is a machine learning predicted metrology residual.

Fig. 12 is an E-direction positioning error curve obtained based on the GRU refined tight-combining positioning algorithm.

FIG. 13 is a plot of N-direction positioning error obtained based on a GRU-refined tight-fit positioning algorithm.

FIG. 14 is a U-direction positioning error curve obtained based on a GRU improved tight-combination positioning algorithm.

FIG. 15 is an E-direction positioning error curve obtained based on a conventional tight-combining positioning algorithm.

Fig. 16 is an N-direction positioning error curve obtained based on a conventional tight-combining positioning algorithm.

Fig. 17 is a U-direction positioning error curve obtained based on a conventional tight-combining positioning algorithm.

Detailed Description

The present invention is further illustrated by the following examples, which are not to be construed as limiting the invention thereto.

The embodiment of the invention relates to a GNSS/INS tight combination positioning method for predicting measurement noise based on machine learning.

Specifically, referring to fig. 1, the GNSS/INS tightly-integrated positioning method for performing measurement noise prediction based on machine learning is divided into two loops. The two loops are not executing at the same time. On the premise of ensuring that the hardware equipment is not changed, a model training loop is firstly carried out. The loop collects data for training in complex road conditions, and meanwhile, a conventional GNSS/INS tight combination algorithm is operated, and a training set is built by utilizing observation data and solution data. In the initial stage of training set construction, the scene types need to be considered as much as possible, and observation environment elements contained in complex scenes are ensured to be contained in the observation environment. After parameter selection, the acquired observation data are arranged into a training set with a specific format, and the process is also called data cleaning.

According to the machine learning rule, the selected parameters are divided into feature quantities and label quantities. The characteristic quantity is composed of a plurality of rows of data, and the label quantity is composed of a row. The regression task is therefore a multiple regression. The characteristic quantity represents the attribute of a learning object, and in the GNSS/INS tightly combined positioning applied in the urban environment, the characteristic quantity means observation parameters related to the environment. The label quantity is a reference value of the learning result and a target parameter to be predicted, and is referred to as measurement noise in a close-coupled positioning application. The interference of buildings and the like in cities on GNSS signals causes the signal quality to be reduced, and the signal-to-noise ratio or the carrier-to-noise ratio can be used for representing the signal quality. The signal quality may affect the ranging error and reduce the positioning accuracy. However, merely selecting the signal quality parameter is not sufficient, especially in complex physical environments, and other variables need to be taken into account. In the invention, the altitude angle, the azimuth angle, the carrier wave observation value and the signal-to-noise ratio are selected as characteristic quantities to participate in the machine learning process.

The parameter to be predicted in the present invention is the metrology noise used to construct the metrology model. The current common fixed observation noise design cannot cope with the observation value change caused by the environmental change when applied in the urban environment. Therefore, in order to cope with the change in the satellite observation value under different environments, the training data is selected by first classifying the composition of the training data. The training data composition is divided into three types according to the shielding condition: severe occlusion, sporadic occlusion, and complete patency. The characteristics of severe occlusion are: the number of visible satellites is small, and the satellites are mainly shielded by urban buildings; the scattered shielding is characterized in that: 4-6 visible stars still exist after being shielded, most of the shielding comes from an overpass, a side overhead or a street tree, and the lowest positioning requirement is basically met; the characteristics of the whole opening are as follows: hardly occluded, with sufficient visible stars and with a good geometrical distribution. Under the condition of complete shielding, as no satellite observation participates in resolving, the weight of the influence of the measurement model on the error estimation result is very small in the filtering process, and the significance of adjusting the measurement noise is not large, so that the parameters during complete shielding do not participate in training.

Under the condition of off-line training, a measurement error is constructed through the deviation between the positioning reference value and the positioning result, so that a label quantity is formed, and the label quantity for training is constructed through residual errors. In the case of a reference value, the exact positioning error, i.e. the true value, can be determined from the coordinate reference value (true value)

Here, the

The first three terms of the state vector are taken to represent the real three-dimensional positioning error, P _Ref Denotes a positioning reference value, P _est And the positioning result estimated by the GNSS/INS tight combination algorithm is shown. At this time, a measurement residual vector may be obtained:

wherein the content of the first and second substances,

is the measured residual error, Z, of the k epoch inverse calculation _k Is the measurement vector for the k epoch.

The physical meaning of residual characterization is the equivalent measurement deviation between the actual measurement value and the reference value inversely calculated from the positioning result. The residual error needs to indirectly reflect the modeling accuracy degree from the result, and the reasonability of the design of the random noise model is judged through the presented error change trend, so that the residual error analysis is more suitable for correcting the measurement noise item. And after the feature quantity and the label quantity are selected to complete the construction of the training set, using the GRU to train the model. The model training process is carried out in an off-line state, and the model training is considered to be finished after the training index is reached.

After the offline training is over, the second solution loop is executed online. The tightly combined GNSS/INS algorithm running in the online loop is an improved algorithm utilizing machine learning. In the actual measurement operation process, the observation value acquired by the system participates in the conventional combined positioning modeling including a state model and a measurement model on one hand, and on the other hand, corresponds to the parameter selection during the training set construction in the off-line loop, and establishes a test set on line. The difference between the test set and the training set is that the test set does not need the label quantity, and only has the characteristic quantity. The measurement noise prediction model which is constructed by an offline circuit can be used for outputting a single epoch to the test set on line to predict the measurement noise of the current epoch. If post-processing is needed, the test set can be stored first, and then multi-epoch prediction can be performed off-line.

Different from the conventional tight combination modeling, the method is characterized in that the measurement model construction does not come from relatively fixed altitude angle or signal-to-noise ratio definite weight, but is predicted by using a machine learning method. In a complex changing urban environment, the corresponding noise parameter is predicted according to the parameter related to the environment. In the GNSS/INS tight combination positioning modeling process, the noise item of the measurement model is constructed by using the measurement noise predicted in real time instead of the traditional empirical model. Using predicted metrology noise error

Computing a measured noise covariance matrix

In which only the diagonal parameters are retained. Then according to the measured noise covariance matrix

And carrying out filtering estimation and outputting a result. The measured noise array obtained according to the regression prediction of the environmental parameters can better reflect the change of the environment and can be more effectively adjustedAnd filtering gain, realizing the self-adaptive adjustment of the noise covariance matrix and improving the accuracy of actual estimation in the urban environment. In the application process, the observation value range improvement method of the improved algorithm collocation can achieve better effect in urban environment. And after the online estimation result is output, ending the resolving loop.

The verification and performance analysis for the method of the invention are as follows:

the collected actual measurement dynamic data is used for testing and verifying, and an actual measurement dynamic data set (collected in 22 days of 9 months and indicated by 0922 as a number hereinafter) contains rich urban complex environments and is suitable for training models to test other data collected under the urban environment. Since only the BDS and Galileo observation data are included in the 0922 data, and the test set includes GPS data, another piece of data is added for training. Supplemental data was collected at 18 months 4 (hereinafter 0418 numbered) and included the BDS and GPS data as a supplement to the 0922 dataset. Fig. 2 and 3 show GNSS double-difference wide-lane observations for the 0922 dataset and the 0418 dataset, respectively. As can be seen from the figure, the 0922 data set observation environment has a large influence on GNSS observations, which results in frequent observation changes and is a complex road condition. And 0418, the observation environment is better in data concentration, and the observed value changes stably, so that the road condition is open. In addition, as can also be seen from the satellite numbers, the satellite systems involved in the two sets of data sets can be complementary.

After the training set is assembled, a piece of typical urban canyon data (hereinafter indicated by 1207) is used as a test set to test the positioning performance of the algorithm. The three sections of data have different acquisition times, but the common characteristic is that hardware equipment is completely the same. The following table counts the PDOP fraction at different levels for visible satellites in the training set and test set. The PDOP reflects satellite distribution, is related to visual satellite number, altitude, and azimuth, and is suitable for expressing applicable associations of training sets and test sets. From the statistical results, the number of samples is sufficient for model training, although the percentage is different.

Different data set PDOP hierarchical statistics

The training process adopts a threshold limiting mode, namely, the training is finished when the model training reaches the precision threshold. In order to verify the effectiveness of model training, the model is subjected to verification testing by using a verification set. The training set adopts the label quantity to calibrate the model training result, and the verification process firstly uses the characteristic quantity to predict and then uses the label quantity as the true value to evaluate the accuracy of the prediction result. A small portion of the training set is typically used as the validation set. The testing process uses the measurement residuals as the label amount for the training set. Training is performed after the training set is prepared. The effect of model training is shown in fig. 4-6.

The mean absolute error of the model training loss function is shown in fig. 4. It can be seen from the figure that both the training set loss function and the validation set loss function reach stable convergence. The difference between the measured error predicted by the validation set and the label quantity can be predicted, and the residual error shown in fig. 5 shows better gaussian distribution. Shown in fig. 6 is the relationship between the prediction result and the ideal straight line, i.e., the euclidean distance. The model training achieves higher accuracy as can be seen from the characteristic that the prediction result in the graph is in good accordance with the ideal straight line.

After model training is completed, positioning accuracy testing is performed using the urban canyon data. The test set used 1207 data set. From the number of co-view satellites shown in fig. 7, frequent data occlusion exists in the test road segment, and the total number of observable satellites in part of epochs is less than 4. According to statistics, the PDOP value of more than 90% of epochs is distributed in a range of 3-10. The wide-lane observation ambiguity decimals and their distribution are shown in fig. 8 and 9.

The metrology noise is predicted using the trained model in the previous section, and the metrology model is updated based on the predicted metrology noise. Fig. 10 and 11 show measurement residuals obtained by performing inverse estimation using reference values when the machine learning method is not used, and measurement noises predicted by using the machine learning method, respectively. The comparison of the two figures shows that the predicted value is very close to the calculated residual.

Fig. 12-17 show positioning error curves in an urban canyon. Fig. 15 to 17 show error curves before machine learning is used, and fig. 12 to 14 show error curves predicted by machine learning. From the comparison in the figures, the divergence amplitude of the positioning error curve in the urban canyon is obviously reduced after the improvement of the machine learning method is used. And respectively using a machine learning-based improved close-combination positioning algorithm and a conventional close-combination positioning algorithm without machine learning to perform positioning calculation on the road section, wherein the positioning errors obtained by the two algorithms and the positioning accuracy statistics of the road section are shown in the following table. According to the analysis of the influence of the urban canyon shielding direction on the positioning accuracy, the positioning statistical accuracy in the E direction is superior to that in the N direction, and the improvement range shown in the table also conforms to the rule.

Improved algorithm based on GRU is compared with conventional algorithm

According to the experimental result, a tight combination positioning algorithm of a measuring noise model constructed by machine learning instead of a traditional empirical model is used, and the improvement range of 14.02% of a plane and 20.49% of a height is achieved in the urban complex environment test.

Claims

1. A tight combination positioning method of GNSS/INS for measuring noise prediction based on machine learning is characterized by comprising the following steps: 1. in an off-line state, adopting GNSS/INS observation parameters to construct training set data, and using GRU as a noise prediction model to perform model training; 2. in an online state, a state model is built through a GNSS/INS observation value, meanwhile, the GNSS/INS observation value is used as input of a trained noise prediction model to obtain predicted measurement noise, a measurement model is built through the predicted measurement noise, and finally, the state model and the measurement model are used for filtering estimation processing to output a positioning result.

2. The GNSS/INS tightly-combined positioning method for machine learning-based metrology noise prediction as claimed in claim 1, wherein the GNSS/INS observation parameters include altitude, azimuth, carrier observation and signal-to-noise ratio, the altitude, azimuth, carrier observation and signal-to-noise ratio constitute the feature quantities of the training set of the noise prediction model, and the tag quantities of the training set of the noise prediction model are constructed by the deviation between the positioning reference value and the positioning result.

3. The GNSS/INS tight-combination positioning method based on machine learning measurement noise prediction according to claim 2, wherein the deviation between the positioning reference value and the positioning result is obtained by:

wherein the content of the first and second substances,

4. The method of claim 1, wherein the training set data is subjected to data processing prior to training the noise prediction model, and the data processing comprises deleting data of less than 4 visible satellites in the training set data.

5. The tight-coupled GNSS/INS positioning method for machine-learning based metrology noise prediction as claimed in claim 2 wherein the metrology model built from predicted metrology noise is derived from GNSS/INS observations input into the trained noise prediction model

Computing a measured noise covariance matrix

In which only the diagonal parameters are retained.