Adaptive robust robust information fusion method for GNSS / INS tightly integrated navigation in combination with neural network

Abstract

The invention relates to an adaptive robust robust information fusion method for GNSS / INS tightly integrated navigation in combination with a neural network, belongs to the field of integrated navigation, and solves the problem of robust information fusion. The method comprises the following steps: establishing and training a neural network model used for distribution fitting of kernel parameters and errors of each channel of the multi-channel heteronuclear generalized correlation entropy Kalman filter; in the GNSS / INS tight integration navigation system INS recursion error estimation process, when the running time reaches the length of a close time observation window, predicting each channel kernel function parameter corresponding to each satellite in the filter; in the generalized correlation entropy Kalman filtering, performing INS error state estimation and recursion error correction by adopting the predicted kernel function parameters of each channel; and when the running time reaches the length of the next close time observation window, predicting kernel function parameters of each channel again. The method can deal with general noise of which the overall distribution is in a non-Gaussian form, and has robust capability.

Description

Technical Field

[0001] The invention belongs to the field of integrated navigation, and in particular relates to an adaptive robust anti-error information fusion method for GNSS / INS tightly integrated navigation combined with a neural network. Background Art

[0002] In actual navigation applications, both GNSS and INS have their own limitations. The combined navigation scheme of INS and GNSS positioning information can make full use of the complementarity between the various information sources and overcome the limitations of navigation using a single information source. This is also the development trend of navigation technology today, that is, the use of multi-source information fusion navigation and positioning technology using multiple sensors. The GNSS / INS combined navigation system can combine the observation information of GNSS with the recursive information of INS to obtain navigation results with high accuracy, fast refresh rate and strong stability.

[0003] In the research of GNSS / INS integrated navigation system, the information fusion algorithm that fuses the two navigation information is the key technology. The mainstream information fusion algorithm is Kalman filtering, but its robustness is poor. Its performance is optimal only when the noise is Gaussian distribution. In actual application scenarios, there are often tunnels, canyons, urban building shielding and unknown electromagnetic interference, which leads to the GNSS observations often containing non-Gaussian noise. In the current research on information fusion algorithms for dealing with non-Gaussian noise, there is a method that introduces correlation entropy into Kalman filtering as a robust factor to replace the second-order moment, thereby improving the robustness of the information fusion algorithm, which is called maximum correlation entropy Kalman filtering (MCKF). Furthermore, in order to deal with more diverse and complex non-Gaussian noise, there is a maximum generalized correlation entropy Kalman filtering (GMCKF) based on the generalized Gaussian kernel function. The generalized Gaussian kernel function introduces shape parameters and kernel bandwidth to fit more complex non-Gaussian noise, but due to the increase in parameters and the great influence of parameters on system performance, how to choose appropriate parameters has become an urgent problem to be solved. The generalized Gaussian kernel introduced by GMCKF can not only show good performance in systems where the overall noise distribution is non-Gaussian, such as Laplace distribution, but also fit the noise distribution with thick tail characteristics by adjusting the kernel bandwidth, thereby showing robustness in systems with low observation quality and abnormal observation values. However, the increase in parameters and the coupling between parameters make it difficult to determine the parameters, and the statistical characteristics of GNSS observation noise may also change with the movement of the dynamic carrier. Therefore, it is necessary to adjust the shape parameters of the kernel function in real time to adapt to different overall distributions, and also adjust the kernel bandwidth to perform robustness processing on channels with impulse noise. In addition, in the GNSS / INS integrated navigation system, the observation error characteristics of different GNSS observation channels (observations of different satellites) may be significantly different, their overall distributions may be completely different, and some channels contain obvious impulse noise while other channels do not, so it is necessary to perform differentiated processing on different observation channels. Summary of the invention

[0004] In view of the above analysis, the present invention aims to disclose an adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with a neural network, which is used for information fusion of the GNSS / INS tightly integrated navigation system. It can not only cope with general noise with a non-Gaussian overall distribution, but also cope with obvious outlier anomalies (i.e., impulse noise), and has robustness. On the basis of GMCKF, the kernel parameters of each channel are independent, and a neural network is used to dynamically estimate the shape parameters and kernel bandwidth, and it has the ability to adapt to time-varying noise.

[0005] The present invention discloses an adaptive robust anti-error information fusion method for GNSS / INS tightly integrated navigation combined with a neural network, comprising: Step S1: for a GNSS / INS tightly integrated navigation system that uses a multi-channel heterogeneous kernel generalized correlation entropy Kalman filter for INS recursive error estimation, a neural network model for fitting the kernel parameters of each channel with the error distribution is established and trained; Step S2: in the INS recursive error estimation process of the GNSS / INS tightly integrated navigation system, when the running time reaches a near-time observation window length, the error sample probability distribution histogram sequence is established by using the posterior errors of the observation quantities of each satellite in the near-time observation window, and the neural network model is input to predict the kernel function parameters of each channel corresponding to each satellite in the filter; Step S3, using the predicted kernel function parameters of each channel, generalized correlation entropy Kalman filtering is performed to realize INS error state estimation and correct recursive error; when the running time reaches the next near-time observation window length, return to step S2 to re-predict the kernel function parameters of each channel.

[0006] Furthermore, the neural network model is a CNN-GRU hybrid neural network structure; the input of the model is a probability distribution histogram sequence of error samples, and the model output is a shape parameter and kernel bandwidth of a generalized Gaussian kernel function.

[0007] Furthermore, the training data set of the neural network model and the process of preprocessing the data set include: 1) Generate a two-dimensional target grid of kernel function shape parameters and kernel bandwidth of a multi-channel heterogeneous kernel generalized correlation entropy Kalman filter with a certain step size; 2) Substitute each target parameter in the two-dimensional target grid into the noise probability density function fitted by the generalized Gaussian kernel function to obtain the probability density function of the random variable; 3) Using the Markov chain Monte Carlo algorithm to obtain a sample sequence of random variables that conform to the probability distribution of the probability density function, and using multiple sample sequences to form a sample data matrix; 4) Convert the sample data matrix into a training sample matrix consisting of a sample probability distribution histogram sequence.

[0008] Furthermore, the sample data matrix is ​​converted into a training sample matrix consisting of a sample probability distribution histogram sequence, including: (1) Set the detection step size; (2) Divide the range of the random variable into segments at equal intervals with the set detection step size; divide the observation error range into multiple segments; (3) For each sample sequence in the sample data matrix generated by each set of target parameters in the two-dimensional target grid, the number of samples that fall within the observation error range of each segment is counted, and the proportion relative to the total number is calculated to generate a histogram sequence.

[0009] Further, in step S2, The process of predicting the kernel function parameters of each observation channel includes: 1) Set the length of the near-time observation window to When the GNSS / INS tightly integrated navigation system runs for a time greater than or equal to the length of a near-time observation window, a corrected position at each moment is obtained. ; 2) According to the corrected position at each moment Calculate the Observation channels approaching The sample sequence of posterior error of the observation at each moment; 3) Use the Observation channels approaching The observed error sample sequence at each moment establishes the error sample probability distribution histogram sequence; 4) The error sample probability distribution histogram sequence of the observation channel is input into the neural network model to predict the shape parameters of the channel kernel function and core bandwidth .

[0010] Furthermore, when the filtering process runs to At the moment, Observation channel approaching The observation error sequence at each moment is ;in, ; for Moment The a posteriori error of the satellite observations is for Time to The pseudoranges observed by the satellites, For the Satellites in The optimal estimate of the pseudorange at time is obtained by Three-dimensional coordinates of satellite position at the moment And the corrected 3D coordinates of the user position It is obtained by taking the difference modulo the vector.

[0011] Furthermore, during the INS recursive error estimation process, the state vector of the GNSS / INS tightly integrated navigation system is established as: ; in, , , , , , and They are the position vector error, velocity error, misalignment angle, gyroscope bias, accelerometer bias, gyroscope scale factor error and accelerometer scale factor error of the INS state recursion result in the ECEF coordinate system; The observed quantity is: ; in, for Time to The pseudoranges observed by the observable satellites, For the The three-dimensional coordinates of the satellite positions, is the three-dimensional coordinate of the recursive position of INS at time k; i=1,…,m, where m is the number of observable satellites.

[0012] Furthermore, the process of performing INS state recursion error includes: 1) Perform one-step state prediction to obtain One-step estimation The covariance matrix of : 2) According to One-step estimation The covariance matrix of and the covariance matrix of the observation noise Perform Cholesky decomposition; 3) According to the Cholesky decomposition results, based on One-step estimation Filter residual Iterative calculation of ; Indicates the number of iterations, the initial value is 1; 4) According to The filter residual of the iteration is calculated The generalized correlation entropy weighted values ​​of each observation channel of the iteration constitute the generalized correlation entropy weighted matrix ; In the iterative process, the observation channel kernel function parameter prediction process in step S2 is used to perform the weighting matrix Update the shape parameters and kernel bandwidth of each channel kernel function; 5) Using the generalized correlation entropy weighted matrix Calculate Kalman gain ; 6) According to Kalman gain Perform state estimation to obtain Iteration value ; When the The iteration value is the same as the When the difference of the iteration value is greater than the threshold, return to step 3) and execute again; otherwise, get the current moment The best estimate of the state quantity .

[0013] Further, time The weight matrix of the iteration : ; in, For the Channel kernel function, Kernel Function The expression is: ; For the The shape parameter of the kernel function for each observation channel is the kernel bandwidth. When the system operation time is less than the first near-time observation window, the shape parameters of each channel are Set to 2, Set to ; When it is not the first near-time observation window, the shape parameters and kernel bandwidth predicted by the previous near-time observation window are used.

[0014] Furthermore, the estimated error value is used to perform feedback correction on the carrier motion state. use The best estimate of the state quantity at the moment In , , Correction of the position of the INS recursive state ,speed , direction cosine matrix , get the current position after the estimated error correction ,speed , direction cosine matrix ; ; .

[0015] The present invention can achieve the following beneficial effects: The present invention discloses an adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with a neural network, which uses a multi-channel heterogeneous kernel generalized maximum correlation entropy Kalman filter to realize the combined navigation state estimation, and adopts different shape parameters and kernel bandwidths for different observation channels. A convolutional neural network (CNN) and a gated recurrent unit (GRU) hybrid network are used to monitor the current noise statistical characteristics in real time, focusing on its overall distribution shape and thick-tail characteristics, so as to estimate the kernel function shape parameters and kernel bandwidth of each observation channel, and realize the filter's adaptation to dynamic time-varying noise. In addition, the kernel function of each GNSS observation channel not only has an independent shape parameter, but also has an independent kernel bandwidth. Therefore, this method can not only deal with general noise of non-Gaussian distribution, but also deal with obvious outliers, and the utilization of the observations of each channel will not interfere with each other. The CNN-GRU hybrid neural network is used as a dynamic estimation method for kernel parameters, and the histogram data of noise samples is used as feature samples. It can estimate multiple parameters at the same time, has strong generalization ability and more accurate estimation, can fully extract features in samples, and make full use of the correlation information between samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components; Figure 1 The figure is a flow chart of the adaptive robust information fusion method in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used to illustrate the principles of the present invention together with the embodiments of the present invention.

[0018] An embodiment of the present invention discloses an adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with a neural network, such as Figure 1 As shown, including: Step S1: for a GNSS / INS tightly integrated navigation system that uses a multi-channel heterogeneous kernel generalized correlation entropy Kalman filter for INS recursive error estimation, a neural network model for fitting the kernel parameters of each channel with the error distribution is established and trained; Step S2: in the INS recursive error estimation process of the GNSS / INS tightly integrated navigation system, when the running time reaches a near-time observation window length, the error sample probability distribution histogram sequence is established by using the posterior errors of the observation quantities of each satellite in the near-time observation window, and the neural network model is input to predict the kernel function parameters of each channel corresponding to each satellite in the filter; Step S3, using the predicted kernel function parameters of each channel, generalized correlation entropy Kalman filtering is performed to realize INS error state estimation and correct recursive error; when the running time reaches the next near-time observation window length, return to step S2 to re-predict the kernel function parameters of each channel.

[0019] Specifically, the neural network model for fitting kernel parameters and error distribution in step S1 is a CNN-GRU hybrid neural network structure; the input of the model is a probability distribution histogram sequence of error samples, and the output of the model is the shape parameters and kernel bandwidth of the generalized Gaussian kernel function.

[0020] Specific network structure Since the features of the parameters to be estimated are hidden in a large amount of data, in the network structure, the convolution layer (Conv) is first used to extract data features through the convolution kernel; then the data is downsampled through the pooling layer (Pooling), thereby reducing the data scale, focusing on the most effective features, removing redundant information and compressing the features, improving the network training and feedforward operation speed, and then the inactivation layer (Dropout) is used to randomly inactivate some neurons during the model training process to prevent overfitting. Multiple convolution layers make feature extraction more complete, and finally the fully connected layer (Flatten) is used to map the feature matrix to a feature vector, and enter the GRU model part for further processing.

[0021] The neurons in each layer of the convolutional neural network are independent of each other, which is suitable for situations where the input data are independent of each other. It is not very effective for time series data with contextual connections. The features that reflect the distribution of samples are hidden in the correlation of the histogram data, so it is necessary to take targeted processing. GRU is a recurrent neural network (RNN) that is good at capturing time-related information in data sequences and has fewer parameters and longer-term memory than the long short-term memory network (LSTM). After the GRU network is connected in series to the CNN network, the deep features extracted by the CNN network are used for further fitting, and then output to the deactivation layer (Dropout) to further prevent overfitting of training. Finally, the fully connected layer (FC1) matches the dimension with the output data dimension to obtain the neural network output.

[0022] The specific network parameter configuration is shown in the following table:

[0023] Specifically, before training the neural network model, the process of generating a training data set and preprocessing the data set includes: 1) Generate a two-dimensional target grid of kernel function shape parameters and kernel bandwidth of a multi-channel heterogeneous kernel generalized correlation entropy Kalman filter with a certain step size; Among them, the two-dimensional target grid for: ; in, ; Shape parameters The value range is set to , using uniform sampling, kernel bandwidth The value range is set to , using nonlinear sampling, dense sampling of smaller values ​​and sparse sampling of larger values; the sampling intervals of dense sampling and sparse sampling can be adjusted according to needs.

[0024] 2) Substitute each target parameter in the two-dimensional target grid into the noise probability density function fitted by the generalized Gaussian kernel function to obtain the probability density function of the random variable; Will Each target parameter Substitute the noise probability density function fitted by the generalized Gaussian kernel function to obtain the probability density function of the random variable e for: ; in, is the normalization parameter, and After the parameters are calculated, the numerical solution can be obtained by integral normalization operation, and then the general range of the system observation error is set. , which depends on the accuracy of the actual device, and is set here to , the corresponding unit is meter.

[0025] 3) The probability density function The sample sequence of the random variable e that conforms to the distribution is obtained by the Markov chain Monte Carlo algorithm; a sample data matrix is ​​formed by using a plurality of the sample sequences. ; ; The above formula is A set of target parameters The sample data matrix generated by The lengths of the independent The sample sequence is composed of Each row in samples, a total of Such a sample sequence.

[0026] The setting of needs to consider the size of the filter near time window in the subsequent steps, which is determined according to the performance of the computing platform. Here it is set to 500. Since it is offline training, The value of can be appropriately large, depending on the tolerable training time cost, and is set to 200 here.

[0027] Since directly inputting the above data into the network for training will result in a large amount of data, and the network may eventually assign different weights to data in different positions, which is inconsistent with the facts, the sample data must be processed into the form of a histogram.

[0028] 4) Transform the sample data matrix Convert to a training sample matrix consisting of a sample probability distribution histogram sequence; (1) Set the detection step size; Set the detection step size ; The sample sequence should be generated according to The scale should be set as small as possible to ensure accuracy.

[0029] (2) Divide the range of the random variable into segments at equal intervals with the set detection step size; divide the observation error range into multiple segments; In the range of random variable e Inner The step length is divided into equally spaced segments; Will Divided into n segments, ; (3) For each sample sequence in the sample data matrix generated by each set of target parameters in the two-dimensional target grid, the number of samples falling within the observation error range of each segment is counted, and the proportion relative to the total number is calculated to generate a histogram sequence; right Each set of target parameters The generated sample data matrix Each sample sequence in Count the number of samples that fall within each segment and calculate their proportion relative to the total number ,for Generate a histogram sequence for each sample sequence in , for example: ; Then the sample data matrix Translates to: ; The target parameter The corresponding training sample matrix, accordingly, the above process also needs to be Each target parameter pair in is implemented to obtain the final model training set data.

[0030] Specifically, when training the neural network model, the model input is the histogram sequence in the training sample matrix, that is, The row vector of is 2*1, and the corresponding training target is .

[0031] When training the network, the batch size is set to 4 and the learning rate is set to 0.05, but this does not mean that these parameters are the only choice and can be adjusted appropriately according to the training situation. The back propagation algorithm is used for iterative optimization to automatically adjust the neuron parameters. After the loss converges, the training is stopped and the model is packaged for the subsequent filtering step.

[0032] When using GNSS / INS tightly integrated navigation, the INS sensor data is used to recursively infer the carrier's motion state (such as position, speed, etc.) to obtain the position after pure inertial navigation state recursion. ,speed , direction cosine matrix ; The recursive method can adopt the existing public method.

[0033] However, due to the existence of sensor errors, the reported state also contains errors; the error in the reported motion state is set as the state quantity of the information fusion algorithm , and establish an error model, then fuse the model's predicted error value with the navigation information brought by the GNSS observation through Kalman filtering to obtain the optimal estimate of the error in the current motion state, and use this value to correct the error in the state to obtain a more accurate current carrier motion state: position ,speed , direction cosine matrix .

[0034] Specifically, the process of predicting the kernel function parameters of the i-th observation channel in step S2 includes: 1) Set the length of the near-time observation window to When the GNSS / INS tightly integrated navigation system runs for a time greater than or equal to the length of a near-time observation window, a corrected position at each moment is obtained. ; 2) According to the corrected position at each moment Calculate the Observation channels approaching The sample sequence of posterior error of the observation at each moment; When the filtering process runs to At the moment, Observation channel approaching The observation error sequence at each moment is ;in, ; for Moment The a posteriori error of the satellite observations is for Time to The pseudoranges observed by the satellites, For the Satellites in The optimal estimate of the pseudorange at time is obtained by Three-dimensional coordinates of satellite position at the moment And the corrected 3D coordinates of the user position It is obtained by taking the difference modulo the vector.

[0035] 3) Use the Observation channels approaching The observed error sample sequence at each moment establishes the error sample probability distribution histogram sequence; Using the method of sub-step 4) in step S1, a sample probability distribution histogram sequence of the i-th observation channel is established; 4) The error sample probability distribution histogram sequence of the observation channel is input into the neural network model to predict the shape parameters of the channel kernel function and core bandwidth .

[0036] Specifically, in step S3, the state vector of the established GNSS / INS tightly integrated navigation system is: ; in, , , , , , and They are the position vector error, velocity error, misalignment angle, gyroscope bias, accelerometer bias, gyroscope scale factor error and accelerometer scale factor error of the INS state recursion result in the ECEF coordinate system; According to the error disturbance theory, the discrete state equation and state equation of INS error are established as follows: ; ; in, , They are , The state vector at the moment; for The system noise at the moment, is the state transfer matrix; for The observed quantity at time, for Observe the noise at all times; for The moment observation matrix.

[0037] Among them, the state transfer matrix for:

[0038] in, is the system discrete time interval, is the Earth's rotation angular velocity vector, is the direction cosine matrix from the carrier coordinate system b to the ECEF coordinate system, is the specific force output by the carrier accelerometer, , , , The gyro bias error, accelerometer bias error, gyro scale factor error, and accelerometer scale factor error are modeled as the relevant time parameters of the first-order Markov process respectively; Observable for: ; in, for Time to The pseudoranges observed by the observable satellites, Based on At this moment, INS delivers the position and The geometric distance calculated from the positions of the observable satellites; ; is the number of observable satellites; Time observation matrix , from the known Optimal estimation of the state at the moment and calculated from the ephemeris The satellite position at this moment is calculated and is a known quantity.

[0039] The relationship between the satellite pseudorange observation and the INS recursive state error (i.e., state quantity X) is established through the observation equation; Observable quantity at a time:

[0040] in, for Time to The pseudoranges observed by the observable satellites, For the The three-dimensional coordinates of the satellite's position, is the three-dimensional coordinate of the position recursively derived by INS at time k; i=1,…,m, where m is the number of observable satellites.

[0041] The multi-channel heterogeneous kernel generalized correlation entropy Kalman filter in the embodiment of the present invention is adaptively robust based on a neural network. It is a generalized maximum correlation entropy Kalman filter (NN-ARGMCKF) based on a neural network for adaptive robustness. The filter introduces generalized correlation entropy into Kalman filtering, replacing the mean square value of the error as the optimization cost; a multi-channel heterogeneous kernel structure is adopted to achieve mutual isolation of kernel function shape parameters in different observation channels, and an independent kernel bandwidth is also provided. Therefore, the method can not only cope with general noise of non-Gaussian distribution, but also cope with obvious outliers, and the utilization of observations of each channel will not interfere with each other; and a CNN-GRU hybrid neural network is used as a dynamic estimation method for kernel parameters, and histogram data of noise samples are used as feature samples, so that multiple parameters can be estimated at the same time, generalization ability is strong and estimation is more accurate, features in samples can be fully extracted, and correlation information between samples can be fully utilized to achieve better filtering effect.

[0042] Specifically, in step S3, the process of estimating the INS state recursive error by NN-ARGMCKF includes: 1) Perform one-step state prediction to obtain One-step estimation The covariance matrix of :

[0043]

[0044] for The optimal estimate of the error state at time , For use and the state transition matrix Computational One-step estimate of the error state at time , One-step estimate The covariance matrix of for The moment estimation error covariance matrix, for The system noise covariance matrix at time instant;.

[0045] Here and It will be used to calculate later The optimal estimate of the error state at the moment ; 2) According to One-step estimation The covariance matrix of and the covariance matrix of the observation noise Perform Cholesky decomposition; ; is the covariance matrix of the observation noise in the observation equation.

[0046] 3) According to the Cholesky decomposition results, based on One-step estimation Filter residual Iterative calculation of ; ; in, , for The moment observation; n is the dimension of the state vector; , for The moment observation matrix; When step 3) is executed for the first time at the same time, Take the , and then calculate according to the iterative value; Represents the number of iterations, with an initial value of 1.

[0047] 4) According to The filter residual of the iteration is calculated The generalized correlation entropy weighted values ​​of each observation channel of the iteration constitute the generalized correlation entropy weighted matrix ; In the iterative process, the observation channel kernel function parameter prediction process in step S2 is used to perform the weighting matrix Update the shape parameters and kernel bandwidth of each channel kernel function; time The weight matrix of the iteration : ; in, For the Channel kernel function, Kernel Function The expression is: ; For the The shape parameter of the kernel function for each observation channel is the kernel bandwidth. When the system operation time is less than the first near-time observation window, the shape parameters of each channel are Set to 2, Set to ; When it is not the first near-time observation window, the shape parameters and kernel bandwidth predicted by the previous near-time observation window are used.

[0048] Each observation channel uses different shape parameters This allows it to handle the problem of differentiated non-Gaussian noise in different channels, and each observation channel uses a different kernel bandwidth The influence of channels containing large outliers in the filter can be effectively controlled, while normal channels are not affected, showing stronger robustness and anti-error capability.

[0049] 5) Using the generalized correlation entropy weighted matrix Calculate Kalman gain ; ; ; in, is the observation error covariance matrix of each observation channel weighted by relevant entropy; 6) According to Kalman gain Perform state estimation to obtain Iteration value ; When the The iteration value is the same as the When the difference of the iteration value is greater than the threshold, return to step 3) and execute again; otherwise, get the current moment The best estimate of the state quantity .

[0050] Specifically, ; For the The iteration value obtained by the first iteration is The result of the iteration is calculated.

[0051] At this time, if:

[0052] Jump to step 3) and re-execute, otherwise get the best estimate of the state quantity at the current moment ; The threshold to set.

[0053] The estimated error value is used to perform feedback correction on the carrier motion state. use The best estimate of the state quantity at the moment In , , Correction of the position of the INS recursive state ,speed , direction cosine matrix , get the current position after the estimated error correction ,speed , direction cosine matrix ; ; .

[0054] In summary, the adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with a neural network in an embodiment of the present invention uses a multi-channel heterogeneous kernel generalized maximum correlation entropy Kalman filter to achieve combined navigation state estimation, and adopts different shape parameters and kernel bandwidths for different observation channels. A hybrid network of convolutional neural network (CNN) and gated recurrent unit (GRU) is used to monitor the current noise statistical characteristics in real time, focusing on its overall distribution shape and thick-tail characteristics, so as to estimate the kernel function shape parameters and kernel bandwidth of each observation channel, and realize the filter's adaptation to dynamic time-varying noise. And, The kernel function of each GNSS observation channel has not only independent shape parameters but also independent kernel bandwidth. Therefore, this method can not only deal with general noise of non-Gaussian distribution, but also deal with obvious outliers, and the utilization of observations of each channel will not interfere with each other. The CNN-GRU hybrid neural network is used as a dynamic estimation method for kernel parameters, and the histogram data of noise samples is used as feature samples. It can estimate multiple parameters at the same time, has strong generalization ability and more accurate estimation, can fully extract features in samples, and make full use of the correlation information between samples.

[0055] To sum up, the above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. An adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with a neural network, characterized in that: include: Step S1: for a GNSS / INS tightly integrated navigation system that uses a multi-channel heterogeneous kernel generalized correlation entropy Kalman filter for INS recursive error estimation, a neural network model for fitting the kernel parameters of each channel with the error distribution is established and trained; Step S2: in the INS recursive error estimation process of the GNSS / INS tightly integrated navigation system, when the running time reaches a near-time observation window length, the error sample probability distribution histogram sequence is established by using the posterior errors of the observation quantities of each satellite in the near-time observation window, and the neural network model is input to predict the kernel function parameters of each channel corresponding to each satellite in the filter; Step S3, using the predicted kernel function parameters of each channel, generalized correlation entropy Kalman filtering is performed to realize INS error state estimation and correct recursive error; when the running time reaches the next near-time observation window length, return to step S2 to re-predict the kernel function parameters of each channel.

2. The adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with a neural network according to claim 1 is characterized in that: The neural network model is a CNN-GRU hybrid neural network structure; the input of the model is an error sample probability distribution histogram sequence, and the model output is the shape parameter and kernel bandwidth of a generalized Gaussian kernel function.

3. The adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with neural network according to claim 2 is characterized in that: The training data set of the neural network model and the process of preprocessing the data set include: 1) Generate a two-dimensional target grid of kernel function shape parameters and kernel bandwidth of a multi-channel heterogeneous kernel generalized correlation entropy Kalman filter with a certain step size; 2) Substitute each target parameter in the two-dimensional target grid into the noise probability density function fitted by the generalized Gaussian kernel function to obtain the probability density function of the random variable; 3) Using the Markov chain Monte Carlo algorithm to obtain a sample sequence of random variables that conform to the probability distribution of the probability density function, and using multiple sample sequences to form a sample data matrix; 4) Convert the sample data matrix into a training sample matrix consisting of a sample probability distribution histogram sequence.

4. The adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with neural network according to claim 3 is characterized in that: Convert the sample data matrix into a training sample matrix consisting of a sample probability distribution histogram sequence, including: (1) Set the detection step size; (2) Divide the range of the random variable into segments at equal intervals with the set detection step size; divide the observation error range into multiple segments; (3) For each sample sequence in the sample data matrix generated by each set of target parameters in the two-dimensional target grid, the number of samples that fall within the observation error range of each segment is counted, and the proportion relative to the total number is calculated to generate a histogram sequence.

5. The adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with neural network according to claim 1, characterized in that: In step S2, The process of predicting the kernel function parameters of each observation channel includes: 1) Set the length of the near-time observation window to When the GNSS / INS tightly integrated navigation system runs for a time greater than or equal to the length of a near-time observation window, a corrected position at each moment is obtained. ; 2) According to the corrected position at each moment Calculate the Observation channels approaching The sample sequence of posterior error of the observation at each moment; 3) Use the Observation channels approaching The observed error sample sequence at each moment establishes the error sample probability distribution histogram sequence; 4) The error sample probability distribution histogram sequence of the observation channel is input into the neural network model to predict the shape parameters of the channel kernel function and core bandwidth .

6. The adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with neural network according to claim 5 is characterized in that: When the filtering process runs to At the moment, Observation channel approaching The observation error sequence at each moment is ;in, ; for Moment The a posteriori error of the satellite observations is for Time to The pseudoranges observed by the satellites, For the Satellites in The optimal estimate of the pseudorange at time is obtained by Three-dimensional coordinates of satellite position at the moment And the corrected 3D coordinates of the user position It is obtained by taking the difference modulo the vector.

7. The adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with neural network according to claim 5 is characterized in that: During the INS recursive error estimation process, the state vector of the established GNSS / INS tightly integrated navigation system is: ； in, , , , , , and They are the position vector error, velocity error, misalignment angle, gyroscope bias, accelerometer bias, gyroscope scale factor error and accelerometer scale factor error of the INS state recursion result in the ECEF coordinate system; The observed quantity is: ； in, for Time to The pseudoranges observed by the observable satellites, For the The three-dimensional coordinates of the satellite's position, is the three-dimensional coordinate of the position recursively derived by INS at time k; i=1,…,m, where m is the number of observable satellites.

8. The adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with neural network according to claim 7 is characterized in that: The process of performing INS state recursion error includes: 1) Perform one-step state prediction to obtain One-step estimation The covariance matrix of : 2) According to One-step estimation The covariance matrix of and the covariance matrix of the observation noise Perform Cholesky decomposition; 3) According to the Cholesky decomposition results, based on One-step estimation Filter residual Iterative calculation of ; Indicates the number of iterations, the initial value is 1; 4) According to The filter residual of the iteration is calculated The generalized correlation entropy weighted values ​​of each observation channel of the iteration constitute the generalized correlation entropy weighted matrix ; In the iterative process, the observation channel kernel function parameter prediction process in step S2 is used to perform the weighting matrix Update the shape parameters and kernel bandwidth of each channel kernel function; 5) Using the generalized correlation entropy weighted matrix Calculate Kalman gain ; 6) According to Kalman gain Perform state estimation to obtain the Iteration value ; When the The iteration value is the same as the When the difference of the iteration value is greater than the threshold, return to step 3) and execute again; otherwise, get the current moment The best estimate of the state quantity .

9. The adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with neural network according to claim 8, characterized in that: time The weight matrix of the iteration : ; in, For the Channel kernel function, Kernel Function The expression is: ； For the The shape parameter of the kernel function for each observation channel is the kernel bandwidth. When the system operation time is less than the first near-time observation window, the shape parameters of each channel are Set to 2, Set to ; When it is not the first near-time observation window, the shape parameters and kernel bandwidth predicted by the previous near-time observation window are used.

10. The adaptive robust information fusion method for GNSS / INS tightly integrated navigation combined with neural network according to claim 8, characterized in that: The estimated error value is used to perform feedback correction on the carrier motion state. use The best estimate of the state quantity at the moment In , , Correction of the position of the INS recursive state ,speed , direction cosine matrix , get the current position after the estimated error correction ,speed , direction cosine matrix ; ； 。

Images