阅读说明：本技术 一种gnss/ins紧组合滤波器及导航方法 (GNSS/INS tight combination filter and navigation method ) 是由 肖凯 孙付平 张伦东 朱新慧 李万里 何劢航 于 2020-06-16 设计创作，主要内容包括：本发明涉及一种GNSS/INS紧组合滤波器及导航方法,属于卫星导航技术领域。本发明在现有GNSS/INS紧组合滤波器的基础上增加了一个深度神经网络模型,利用该深度神经网络模型根据GNSS的伪距误差和INS上一时刻的状态误差预测当前时刻INS的状态误差,得到当前时刻INS的状态误差对Kalman滤波器进行修正,用修正后的Kalman滤波器输出的INS状态误差校正值对当前时刻的INS的状态量进行校正,将校正后的INS状态量作为当前时刻的定位信息,以此实现GNSS/INS的紧组合导航。本发明在利用深度神经网络模型对Kalman滤波器进行修正时,充分考虑了卫星定位的误差,大大提高了GNSS/INS紧组合滤波器导航定位的精度。(The invention relates to a GNSS/INS tight combination filter and a navigation method, and belongs to the technical field of satellite navigation. The invention adds a deep neural network model on the basis of the prior GNSS/INS tight combination filter, predicts the state error of the current time INS according to the pseudo-range error of the GNSS and the state error of the last time of the INS by using the deep neural network model, obtains the state error of the current time INS, corrects the Kalman filter, corrects the state quantity of the current time INS by using the corrected INS state error correction value output by the Kalman filter, and uses the corrected INS state quantity as the positioning information of the current time, thereby realizing the tight combination navigation of the GNSS/INS. When the Kalman filter is corrected by using the deep neural network model, the method fully considers the error of satellite positioning, and greatly improves the navigation positioning precision of the GNSS/INS tight combination filter.)

1. A GNSS/INS tight combination filter, characterized in that, the tight combination filter includes Kalman filter and deep neural network model; the Kalman filter is used for carrying out real-time filtering calculation according to the pseudo range and the Doppler data of the GNSS and the INS predicted satellite-ground distance and relative speed; the deep neural network model is used for predicting according to the pseudo-range error of the GNSS and the state error of the INS to obtain a position error, a speed error and an attitude error, and correcting the Kalman filter based on the obtained position error, speed error and attitude error.

2. The GNSS/INS tightly-combined filter of claim 1, wherein the deep neural network model comprises an input layer, a hidden layer and an output layer, wherein the hidden layer comprises a first hidden layer and a second hidden layer; the neurons of the input layer are position errors, speed errors and attitude errors of the INS and pseudo-range observation errors of each satellite; the first hidden layer is divided into a pseudo-range error part and an INS error part which are independent from each other, the pseudo-range error part is fully connected with neurons corresponding to pseudo-range observation errors of all satellites in the input layer, the INS error part of the first hidden layer is fully connected with neurons corresponding to position errors, speed errors and attitude errors of INS in the input layer, and the neurons in the second hidden layer are fully connected with the neurons in the first hidden layer; and the neurons of the output layer are position errors, speed errors and attitude errors of the INS.

3. The GNSS/INS tightly-combined filter of claim 2, wherein the number of neurons in the second hidden layer is 3, which is used to represent an implicit relationship between the INS predicted position and the satellite-to-ground distance.

4. The GNSS/INS tightly-combined filter of claim 2, wherein the number of neurons in the second hidden layer is 6, which is used to represent an implicit relationship between the INS predicted speed and the satellite-to-ground distance.

5. The GNSS/INS tight combination filter of claim 2 wherein the data of the output layer further includes zero bias error values for accelerometers and gyros.

6. The GNSS/INS tightly combined filter of any of claims 2-5, wherein the reference output values used in the deep neural network model training are ideal output values or state estimation output values of a Kalman filter.

7. The GNSS/INS tightly combined filter of any of claims 2-5, wherein the state equation and the measurement equation of the Kalman filter are respectively:

wherein [ r, v, ψ, b_a,b_g]The inner elements respectively represent a position error, a speed error, an attitude error, an acceleration zero offset error and a gyro zero offset error;a direction cosine matrix representing a direction from the carrier b to the geocentric earth-fixed e system;

8. A navigation method of a GNSS/INS tight combination is characterized in that a Kalman filter in the GNSS/INS tight combination is corrected by adopting a deep neural network model, the deep neural network model is used for predicting according to a pseudo-range error of a GNSS and a state error of an INS to obtain a position error, a speed error and an attitude error, the Kalman filter is corrected based on the obtained position error, speed error and attitude error, and a positioning signal of the INS is corrected by an INS state error correction quantity output by the corrected Kalman filter.

9. The method of claim 8, wherein the deep neural network model comprises an input layer, a hidden layer and an output layer, wherein the hidden layer comprises a first hidden layer and a second hidden layer; the neurons of the input layer are position errors, speed errors and attitude errors of the INS and pseudo-range observation errors of each satellite; the first hidden layer is divided into a pseudo-range error part and an INS error part which are independent from each other, wherein the pseudo-range error part of the first hidden layer is fully connected with neurons corresponding to pseudo-range observation errors of all satellites in the input layer, the INS error part of the first hidden layer is fully connected with neurons corresponding to position errors, speed errors and attitude errors of the INS in the input layer, and each neuron in the second hidden layer is fully connected with each neuron in the first hidden layer; and the neurons of the output layer are position errors, speed errors and attitude errors of the INS.

Technical Field

The invention relates to a GNSS/INS tight combination filter and a navigation method, and belongs to the technical field of satellite navigation.

Background

In order to obtain Navigation parameters of interest to the user, a satellite/inertial integrated Navigation System (GNSS/integrated Navigation System) needs to perform state estimation. The most widely used state estimation method is the Kalman filter algorithm proposed by Kalman R E in 1972. The classical Kalman filtering algorithm establishes a state equation based on a kinetic equation of an INS system and an observation equation based on a GNSS observation model. The accuracy and reliability of the state estimation are determined by the degree of knowledge of the INS system dynamics model and the accuracy of the observed noise statistics. If all errors of the INS system and the GNSS system can be accurately modeled, only random errors are left for unmodeled errors, and ideal state estimation accuracy can be obtained through Kalman filtering by means of priori knowledge. However, colored noise and the like of an inertial sensing unit (IMU) are often difficult to accurately model, and particularly, for an IMU with low accuracy, the influence of the colored noise is significant. Furthermore, this disadvantage is more pronounced when the INS system is navigating alone after the satellite signal is lost. In the above situation, the effect of Kalman filtering is not ideal. The filtering algorithm based on artificial intelligence can accurately approximate the unmodeled error from the angle of a mathematical domain, and is very suitable for describing a nonlinear relation. The scenario that Kalman filtering is not applicable can be compensated.

Neural networks were indeed applied in various technical fields since the 80's of the 20 th century. The application of the neural network in the GNSS/INS combined navigation is realized by a filtering algorithm based on the neural network, and specifically, the application has the following two forms:

(1) and only using the neural network to perform fusion filtering of the satellite navigation subsystem and the inertial navigation subsystem. This form is limited by the inherent disadvantages of neural networks: the training time is long, the influence of uncertainty factors is not practical. Other technical approaches are often needed to further remedy the inherent deficiencies of neural networks.

(2) The neural network assists Kalman filtering. In this way, the neural network and Kalman filtering are stored in the GNSS/INS combined system structure, and the two cooperate with each other in two aspects: on one hand, training is carried out by using a state estimation output value of Kalman filtering to assist a neural network, and the trained neural network is used for correcting the Kalman filtering output value; on the other hand, when the inertial navigation system navigates alone, the neural network is utilized to predict the navigation parameters.

At present, a neural network assisted Kalman filtering form is mainly applied to a loose combination model, and a tight combination model can continuously navigate when the number of satellites is less than four, so that the reliability is higher. However, techniques for applying neural networks with tight combinations have not been developed. The publication number is CN110487271A, which discloses an Elman neural network assisted tight combination navigation method when GNSS signals are blocked, and the method applies a neural network to tight combination, but the training model adopted is still in a loose combination form based on position and speed, and only position and speed information is adopted to train the neural network, and only model errors of a system state equation are considered, and errors of satellite positioning are not considered. Particularly, for a low-cost satellite navigation receiver, a pseudo-range observation value output by the low-cost satellite navigation receiver implicitly contains satellite clock error, receiver clock error and various complex error sources of the receiver, and if the error of satellite positioning is not considered, the accuracy of the whole combined navigation is inevitably influenced.

Disclosure of Invention

The invention aims to provide a GNSS/INS tight combination filter and a combined navigation method, which are used for solving the problem of low positioning accuracy of the existing GNSS/INS tight combination navigation system.

The invention provides a tight combination filter for solving the technical problem, which comprises a Kalman filter and a deep neural network model; the Kalman filter is used for carrying out real-time filtering calculation according to the pseudo range and the Doppler data of the GNSS and the INS predicted satellite-ground distance and relative speed; the deep neural network model is used for predicting according to the pseudo-range error of the GNSS and the state error of the INS to obtain a position error, a speed error and an attitude error, and correcting the Kalman filter based on the obtained position error, speed error and attitude error.

The method is characterized in that a deep neural network model is added on the basis of the existing GNSS/INS tight combination filter, and the deep neural network model predicts the state error of the INS at the current moment based on the pseudo-range error and the INS state error of the GNSS; the invention corrects the original INS output position, speed and attitude result by using the state error of the current INS, and takes the corrected INS navigation result as the positioning information of the current time, thereby realizing the tightly-combined navigation of the GNSS/INS. When the Kalman filter is corrected by using the deep neural network model, the method fully considers the error of satellite positioning, and greatly improves the navigation positioning precision of the GNSS/INS tight combination filter.

Further, the deep neural network model comprises an input layer, a hidden layer and an output layer, wherein the hidden layer comprises a first hidden layer and a second hidden layer; the neurons of the input layer are position errors, speed errors and attitude errors of the INS and pseudo-range observation errors of each satellite; the first hidden layer is divided into a pseudo-range error part and an INS error part which are independent from each other, the pseudo-range error part is fully connected with neurons corresponding to pseudo-range observation errors of all satellites in the input layer, the INS error part of the first hidden layer is fully connected with neurons corresponding to position errors, speed errors and attitude errors of INS in the input layer, and the neurons in the second hidden layer are fully connected with the neurons in the first hidden layer; and the neurons of the output layer are position errors, speed errors and attitude errors of the INS.

Further, in order to better describe the relationship between the INS predicted position and the satellite-to-ground distance, the number of the neurons in the second hidden layer is 3, which is used for embodying the implicit relationship between the INS predicted position and the satellite-to-ground distance.

Further, in order to better describe the relationship between the INS predicted speed and the satellite-to-ground distance, the number of the neurons in the second hidden layer is 6, and the number of the neurons is used for embodying the implicit relationship between the INS predicted speed and the satellite-to-ground distance.

Further, in order to better improve the positioning accuracy of navigation, the data of the output layer further comprises zero offset values of an accelerometer and a gyroscope.

Further, the reference output value adopted during the deep neural network model training is an ideal output value or a state estimation output value of a Kalman filter.

Further, the state equation and the measurement equation of the Kalman filter are respectively:

wherein [ r, v, ψ, b_a,b_g]The inner elements respectively represent a position error, a speed error, an attitude error, an acceleration zero offset error and a gyro zero offset error;

a direction cosine matrix representing a direction from the carrier b to the geocentric earth-fixed e system;

representing an accelerometer specific force output;representing a gyro angular rate output; psi represents the attitude misalignment angle error;

representing angular rate of rotation of the earth

A constructed oblique symmetric array;_arepresenting accelerometer noise;_grepresenting gyroscope noise; p represents a pseudo-range observation;ρ_INSrepresents the distance (range-to-satellite) between the inertial navigation prediction receiver and the satellite; λ represents a carrier wavelength; d represents a doppler observation;representing the relative velocity of the inertial navigation prediction receiver and the satellite; e represents a direction cosine vector, and a specific expression is determined by a position corresponding item in a linearized pseudorange positioning equation; t is t _rRepresenting the receiver clock error;_Prepresenting pseudo-range observation random noise;_Drepresenting doppler observed random noise.

The invention also provides a navigation method of the GNSS/INS tight combination, which adopts a deep neural network model to correct the Kalman filter in the GNSS/INS tight combination, wherein the deep neural network model is used for predicting according to the pseudo-range error of the GNSS and the state error of the INS to obtain a position error, a speed error and an attitude error, and realizes the correction of the Kalman filter based on the obtained position error, speed error and attitude error, so as to correct the positioning signal of the INS by the INS state error correction quantity output by the corrected Kalman filter.

The method is characterized in that a deep neural network model is added on the basis of the existing GNSS/INS tight combination filter, and the deep neural network model predicts the state error of the INS at the current moment based on the pseudo-range error and the INS state error of the GNSS; the invention corrects the original INS output position, speed and attitude result by using the state error of the current INS, and takes the corrected INS navigation result as the positioning information of the current time, thereby realizing the tightly-combined navigation of the GNSS/INS. When the Kalman filter is corrected by using the deep neural network model, the method fully considers the error of satellite positioning, and greatly improves the navigation positioning precision of the GNSS/INS tight combination filter.

Further, the deep neural network model comprises an input layer, a hidden layer and an output layer, wherein the hidden layer comprises a first hidden layer and a second hidden layer; the neurons of the input layer are position errors, speed errors and attitude errors of the INS and pseudo-range observation errors of each satellite; the first hidden layer is divided into a pseudo-range error part and an INS error part which are mutually independent, the pseudo-range error part of the first hidden layer is fully connected with neurons corresponding to pseudo-range observation errors of all satellites in the input layer, the INS error part of the first hidden layer is fully connected with neurons corresponding to position errors, speed errors and attitude errors of INS in the input layer, and the neurons in the second hidden layer are fully connected with the neurons in the first hidden layer; and the neurons of the output layer are position errors, speed errors and attitude errors of the INS.

Drawings

FIG. 1 is a schematic diagram of a prior art tight combination GNSS/INS filter;

FIG. 2 is a schematic diagram of a GNSS/INS tight combination filter according to the present invention;

FIG. 3 is a schematic diagram of a current neuron;

FIG. 4 is a diagram of the operation of a current neuron;

FIG. 5 is a schematic structural diagram of a deep neural network model in a GNSS/INS tight combination filter according to the present invention.

Detailed Description

The following further describes embodiments of the present invention with reference to the drawings.

Embodiments of a GNSS/INS tightly combined filter

The structure of a conventional GNSS/INS tight combination filter is shown in fig. 1, and a Kalman filter is adopted, wherein the Kalman filter performs real-time filtering solution according to the output of an Inertial Navigation System (INS) (including the INS predicting the satellite-ground distance and the INS predicting the relative speed between a receiver and a satellite) and the output of the GNSS (including the observation pseudo-range and doppler data of the GNSS), outputs correction values of the position, speed and attitude error of the INS, and corrects the INS predicted position, speed and attitude output by the original inertial navigation system by the correction values.

The present invention adds a deep neural network model to the above structure, and as shown in fig. 2, the deep neural network model is used to predict a pseudo-range error and an INS state error of a GNSS to obtain a position error, a velocity error and an attitude error, and correct a Kalman filter based on the obtained position error, velocity error and attitude error. The pseudo-range error is obtained by subtracting the satellite-ground distance from the GNSS pseudo-range observation value, and the INS state error can use two modes. (1) If the INS state error output by the Kalman filter is adopted, another means is needed to determine the position, the speed and the attitude information of the carrier with higher precision so as to construct an output layer. For example, it is possible to employ: constructing a guide rail in advance, accurately measuring the attitude information of a carrier running on the guide rail, and reversely calculating the accurate position, speed and attitude information of the carrier by other technical means (such as distance measurement by using a total station) when a GNSS/INS combined task is carried out; and a high-precision INS system is configured to synchronously measure position, speed and attitude information with higher relative precision. (2) And using the INS error estimated value at the previous moment, and simultaneously using the output value of the Kalman filter at the current moment to construct an output layer. This mode is suitable for users who do not have other high-precision reference information to train the neural network. The simple forms of the state equation and the observation equation adopted by the Kalman filter are respectively as follows:

L＝Hx+e

Where H denotes an observation matrix, and x ═ r, v, ψ, b_a,b_g]Representing the state parameters to be estimated and e representing the observed noise vector.

The GNSS/INS tightly-combined navigation system aims to obtain navigation parameters required by a user, namely three navigation parameters of position, speed and attitude; as an error source having the largest influence on the inertial navigation system INS, the acceleration zero offset and the gyroscope zero offset also need to be estimated together.

Therefore, the state estimation is carried out by indirect filtering, the parameters are not directly estimated, but the state estimation is realized by estimating the error vector of the state parameters, and the influence of different parameter orders of magnitude on filtering calculation can be avoided. The state parameters are presented in the form of error vectors, namely x, and the state parameters are involved in calculation in the form of the error vectors, which is also beneficial to further neural network construction.

The inertial navigation system obtains a position, a speed and an attitude structure based on recursive operation, and an error state equation obtained according to an error equation of inertial navigation is as follows:

in the formula [ r, v, psi, b_a,b_g]^TThe inner elements respectively represent position error, speed error, attitude error, acceleration zero offset error and gyro zero offset error. A direction cosine matrix representing a direction from the carrier b to the geocentric earth-fixed e system;representing an accelerometer specific force output;representing a gyro angular rate output; psi represents the attitude misalignment angle error;representing angular rate of rotation of the earthA constructed oblique symmetric array; where H denotes an observation matrix, and x ═ r, v, ψ, b_a,b_g]Representing a state parameter to be estimated;_ais indicative of the noise of the accelerometer,_grepresenting gyroscope noise.

The GNSS/INS tight combination navigation system is tightly combined based on pseudo-range and Doppler observed values provided by a satellite navigation system, an observation equation of the GNSS/INS tight combination navigation system is converted from a pseudo-range positioning equation and a Doppler velocity measurement equation, and the GNSS/INS tight combination navigation system has the following form:

in the formula, P represents a pseudo-range observed value; rho_INSRepresents the distance (range-to-satellite) between the inertial navigation prediction receiver and the satellite; λ represents a carrier wavelength; d represents a doppler observation;

representing the relative velocity of the inertial navigation prediction receiver and the satellite; e represents a direction cosine vector, and a specific expression is determined by a position corresponding item in a linearized pseudorange positioning equation; t is t_rRepresenting the receiver clock error;_Prepresenting pseudo-range observation random noise;_Drepresenting doppler observed random noise.

And correcting the Kalman filter by using the deep neural network model, and constructing the deep neural network model by enhancing the core of the filtering effect. The basic unit of a neural network is a neuron. When the neuron works, the neuron body is provided with input and output. The essence of the neuron is the functional mapping of input to output, and the structure of the neuron is shown in fig. 3, where x represents the independent input values that enter the neuron together; y represents the output of the neuron. As shown in fig. 4, the neuron has a corresponding weight value for each value of x, denoted as w, and there is an offset value for the sum of the inputs, denoted as b. To ensure the output value is continuous, the output value needs to be processed by a first-order function, namely.

The weights (w1, w2, …, wn) and the inputs (x1, x2, …, xn) are expressed as vectors, and the mapping relation is expressed as a function (wx). According to different function mapping relations, the neurons are divided into a plurality of types, and the patent uses S-type neurons, namely the following functional expression exists between output y and input x:

in practical applications, a plurality of neurons are often used to form a hidden layer of a network, all input values form an input layer, and all output values form an output layer. The hidden layer accurately approximates the nonlinear relationship between the input and the output through the cooperation of a plurality of neurons. Neural networks can therefore be used to simulate any system with inputs and outputs. Here, the case of only 1 hidden layer is described, and if the hidden layer is more than 1, the neural network becomes a deep neural network.

The invention aims at a tightly-combined navigation system, so a deep neural network model must be constructed on the basis of tightly-combined observed quantities, and the deep neural network model adopted in the GNSS/INS tightly-combined filter of the embodiment comprises an input layer, a first hidden layer, a second hidden layer and an output layer, as shown in FIG. 5. The construction process of each layer is as follows:

1) Input layer

Because the original GNSS observation values used by the tight combination of the GNSS and the INS are pseudo-range observation values and Doppler observation values, the INS state parameters and the GNSS pseudo-range are basic parameters for constructing the inertial prediction of the earth-satellite (satellite antenna center and receiver antenna center) distance, and the INS state and the GNSS pseudo-range are bound to establish a connection at a certain layer of a neural network, the INS state errors and the GNSS pseudo-range errors are used as neurons of an input layer of the neural network, wherein the INS state errors comprise INS position errors, velocity errors and attitude errors, and the GNSS pseudo-range errors comprise pseudo-range errors of each satellite in the GNSS system.

2) Hidden layer

The hidden layer is the core of the deep neural network model and embodies the building philosophy of the deep neural network. According to the INS dynamic model and the error equation, the following steps are obtained: the INS state estimation error is closely related to the speed and attitude of the carrier, and when the carrier moves linearly at a constant speed, the characteristic can be described as follows: firstly, the speed and the posture are not changed, and the position increment is not changed; the position coordinate can be seen as linear change in a short time; and the increment of the geometric distance between the receiver and the satellite is regular. In this case, the INS state estimation error is small. If the carrier moves in a variable speed and a curve, the position increment is not changed linearly any more, and the geometric distance increment between the receiver and the satellite is not regular. The INS state estimation error under such carrier motion conditions increases significantly. Therefore, the link between the pseudo range observed value between the receiver and the satellite and the motion state of the carrier must be considered. Therefore, the link between the pseudo-range observation value and the motion state of the carrier must be embodied when the hidden layer is constructed. On the other hand, the observation quantities used for measurement updating in the tightly-combined observation equation are pseudo-range observation values and doppler observation values, so that the INS predicted satellite-ground distance and the pseudo-range observation values must be linked in a certain layer of the hidden layer, and only the last layer is required.

Therefore, for the above reasons, the hidden layer established in this embodiment has two layers, which are respectively a first hidden layer and a second hidden layer, where the first hidden layer is a non-fully-connected layer and is divided into a pseudo-range error portion and an INS error portion, which are independent of each other, the pseudo-range error portion is fully connected to the neurons corresponding to the pseudo-range observation errors of each satellite in the input layer, the INS error portion is fully connected to the neurons corresponding to the position errors, the velocity errors, and the attitude errors of the INS in the input layer, and each neuron in the second hidden layer is fully connected to each neuron in the first hidden layer, so as to establish the relationship between the two.

The number of neurons in the two hidden layers is selected according to actual needs. The fewer the neurons are, the lower the learning and training cost is; the more neurons, the more finely modeled the error. However, the marginal effect produced by too many neurons is not apparent. For the embodiment, the number of neurons in the first hidden layer is not less than the total number of observable satellites and INS state parameters; the number of neurons of the second hidden layer can be set to 3 or 6. The purpose of setting to 3 is to embody an implicit relationship between the INS predicted position and the satellite-to-ground distance; setting to 6 further embodies the implicit connection between the INS predicted speed and the satellite-to-ground distance.

3) Output layer

The construction of the output layer is related to the input layer, the form of which determines the form of the output layer. For the invention, the deep neural network model needs to keep consistent with the tightly combined observed quantity, so the invention takes the position error, the speed error and the attitude error as the neurons of the output layer. In addition, considering the influence of the inertial navigation system INS maximum error source, the accelerometer zero offset and the gyroscope zero offset can also be used as neurons of the output layer.

Before the INS state error of the inertial navigation system is predicted by adopting the deep neural network model, the deep neural network model needs to be trained, and the training set used in the training needs to determine the reference output for neuron comparison of an output layer of the deep neural network model. The reference output in the present invention may be an ideal output value or an output value of a Kalman filter.

Where ideal output values refer to relative true values. For example, it is possible to employ: (1) constructing a guide rail in advance, accurately measuring the attitude information of a carrier running on the guide rail, and reversely calculating the accurate position, speed and attitude information of the carrier by other technical means (such as distance measurement by using a total station) when a GNSS/INS combined task is carried out; (2) configuring a high-precision INS system to synchronously measure position, speed and attitude information with higher relative precision; (3) the Kalman filter outputs the value.

Method embodiment

The navigation method adopts a deep neural network model to correct the Kalman filter in the GNSS/INS tight combination, wherein the deep neural network model is used for predicting according to the pseudo-range error of the GNSS and the state error of the INS to obtain a position error, a speed error and an attitude error, and realizes the correction of the Kalman filter based on the obtained position error, speed error and attitude error, and corrects the positioning signal of the INS according to the INS state error correction quantity output by the corrected Kalman filter. The specific structure of the deep neural network model, the observation equation of the Kalman filter, and the system error equation have been described in detail in the embodiment of the filter, and are not described herein again.

When the Kalman filter is corrected by using the deep neural network model, the method fully considers the error of satellite positioning, and greatly improves the navigation positioning precision of the GNSS/INS tight combination filter.