Adaptive Robust Anti-Robust Information Fusion Method for GNSS / INS Tightly Coupled Navigation Combined with Neural Network

Abstract

The present invention relates to an adaptive robust anti - bias information fusion method for GNSS / INS tightly - coupled navigation combined with a neural network, belonging to the field of integrated navigation, and solving the problem of anti - bias information fusion. In the method, a neural network model for fitting the kernel parameters and error distribution of each channel of a multi - channel heterogeneous generalized correlation entropy Kalman filter is established and trained. During the INS recursive error estimation process of the GNSS / INS tightly - coupled navigation system, when the running time reaches a length of a neighboring time observation window, the kernel function parameters of each channel corresponding to each satellite in the filter are predicted. In the generalized correlation entropy Kalman filter, the predicted kernel function parameters of each channel are used for INS error state estimation and recursive error correction. When the running time reaches the length of the next neighboring time observation window, the kernel function parameters of each channel are predicted again. The present invention can cope with general noise with a non - Gaussian overall distribution and has anti - bias ability.

Description

Technical Field

[0001] The present invention belongs to the field of integrated navigation, and particularly relates to an adaptive robust anti-robust 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 respective limitations. The combined navigation scheme using the positioning information of INS and GNSS can make full use of the complementarity between the information sources and overcome the limitations when using a single information source for navigation. This is also the trend of the development of navigation technology today, that is, the multi-source information fusion navigation and positioning technology using multiple sensors. The GNSS / INS integrated navigation system can combine the observation information of GNSS and the recursive information of INS to obtain a navigation result with high accuracy, fast refresh rate and strong stability.

[0003] In the research of GNSS / INS integrated navigation systems, the information fusion algorithm for fusing the two types of navigation information is a key technology. The mainstream information fusion algorithm is the Kalman filter, but its robustness is poor, and its performance is optimal only when the noise is Gaussian distributed. However, in actual application scenarios, there are often tunnels, canyons, urban building blockages, and unknown electromagnetic interferences, etc., resulting in the GNSS observations often containing non-Gaussian noise. In the current research on information fusion algorithms for dealing with non-Gaussian noise, the correlation entropy is introduced into the Kalman filter as a robust factor to replace the second moment, thereby improving the robustness of the information fusion algorithm, which is called the maximum correlation entropy Kalman filter (MCKF). Further, in order to deal with more diverse and complex non-Gaussian noise, there is the maximum generalized correlation entropy Kalman filter (GMCKF) based on the generalized Gaussian kernel function. The generalized Gaussian kernel function introduces the shape parameter and the kernel bandwidth to fit more complex non-Gaussian noise. However, due to the increase in parameters and the great influence of the parameters on the system performance, how to select 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 can fit the noise distribution with heavy-tailed characteristics by adjusting the kernel bandwidth, thereby showing the 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. Moreover, with the movement of the dynamic carrier, the statistical characteristics of the GNSS observation noise may also change. Therefore, it is necessary to adjust the shape parameter of the kernel function in real time to adapt to different overall distributions, and at the same time, adjust the kernel bandwidth to perform robustness processing on the 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 have obvious differences, their overall distributions may be completely different, and some channels contain obvious impulse noise while other channels do not. Therefore, it is necessary to perform differential 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 and resistant information fusion method for GNSS / INS tightly integrated navigation combined with a neural network, which is used for the information fusion of the GNSS / INS tightly integrated navigation system. It can not only deal with general noise with a non-Gaussian overall distribution, but also deal with obvious outlier anomalies (i.e., impulse noise), has the robustness, and on the basis of GMCKF, makes the kernel parameters of each channel independent, and uses a neural network to dynamically estimate the shape parameter and the kernel bandwidth, and has the adaptive ability to time-varying noise.

[0005] The present invention discloses an adaptive robust and resistant information fusion method for GNSS / INS tightly integrated navigation combined with a neural network, including:

[0006] Step S1: For a GNSS / INS tightly coupled navigation system that uses a multi-channel heteronuclear generalized correlation entropy Kalman filter for INS recursive error estimation, establish and train a neural network model for fitting the kernel parameters and error distribution of each channel.

[0007] Step S2: During the INS recursive error estimation process of the GNSS / INS tightly coupled navigation system, when the running time reaches the length of a near-time observation window, use the posterior errors of the observations of each satellite within the near-time observation window to establish a sequence of error sample probability distribution histograms, and input them into the neural network model to predict the kernel function parameters of each channel corresponding to each satellite in the filter.

[0008] Step S3: Use the predicted kernel function parameters of each channel to perform generalized correlation entropy Kalman filtering to achieve INS error state estimation and correct the recursive error; when the running time reaches the length of the next near-time observation window, return to Step S2 to re-predict the kernel function parameters of each channel.

[0009] Further, the neural network model is a CNN-GRU hybrid neural network structure; the input of the model is the sequence of error sample probability distribution histograms, and the output of the model is the shape parameter and kernel bandwidth of the generalized Gaussian kernel function.

[0010] Further, the training data set of the neural network model and the process of preprocessing the data set include:

[0011] 1) Generate a two-dimensional target grid of the shape parameters and kernel bandwidths of each kernel function of the multi-channel heteronuclear generalized correlation entropy Kalman filter with a certain step size.

[0012] 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.

[0013] 3) Use the Markov chain Monte Carlo algorithm for the probability density function to obtain a sample sequence of random variables that conform to this probability distribution, and form a sample data matrix using multiple such sample sequences.

[0014] 4) Convert the sample data matrix into a training sample matrix composed of a sequence of sample probability distribution histograms.

[0015] Further, converting the sample data matrix into a training sample matrix composed of a sequence of sample probability distribution histograms includes:

[0016] (1) Set the detection step size.

[0017] (2) Segment the range of the random variable at equal intervals with the set detection step size; divide the range of the observation error into multiple segments.

[0018] (3) For each set of target parameters in the two-dimensional target grid, count the number of samples in each sample sequence of the generated sample data matrix that fall within each observation error range, and calculate the proportion relative to the total number to generate a histogram sequence.

[0019] Further, in step S2, the process of predicting the kernel function parameters of the th observation channel includes:

[0020] 1) Set the length of the adjacent time observation window to . When the GNSS / INS tightly coupled navigation system runs for greater than or equal to one adjacent time observation window length, obtain the corrected position at each moment .

[0021] 2) Calculate the posterior error sample sequence of the observations at the th moment adjacent to the th observation channel according to the corrected positions at each moment .

[0022] 3) Use the observation error sample sequence at the th moment adjacent to the th observation channel to establish a histogram sequence of the error sample probability distribution.

[0023] 4) Input the histogram sequence of the error sample probability distribution of the th observation channel into the neural network model to predict the shape parameter and the kernel bandwidth of this channel.

[0024] Further, when the filtering process runs to the th moment, the observation error sequence at the th moment adjacent to the th observation channel is ; where

[0025] ;

[0026] is the posterior error of the observation of the th satellite at the th moment, is the pseudorange of the observation of the th satellite at the th moment, is the optimal estimate of the pseudorange of the th satellite at the th moment, through the three-dimensional coordinates of the satellite position at the th moment and the three-dimensional coordinates It is obtained by taking the modulus of the difference of vectors.

[0027] Furthermore, in the INS recursive error estimation process, the state vector of the GNSS / INS tightly coupled navigation system established is:

[0028] ;

[0029] Among them, , , , , , and are respectively the position vector error, velocity error, misalignment angle, gyroscope zero bias, accelerometer zero bias, gyroscope scale factor error, and accelerometer scale factor error of the INS state recursive result in the ECEF coordinate system;

[0030] The observation quantity is:

[0031] ;

[0032] Among them, is the pseudorange observed for the th observable satellite at time , is the three-dimensional coordinate of the position of the th satellite, is the three-dimensional coordinate of the INS recursive position at time k; i = 1,..., m, where m is the number of observable satellites.

[0033] Furthermore, the process of performing INS state recursive error includes:

[0034] 1) Perform a one-step state prediction to obtain the covariance matrix of the one-step estimate at time :

[0035] 2) Perform Cholesky decomposition according to the covariance matrix of the one-step estimate at time and the covariance matrix of the observation noise;

[0036] 3) Based on the Cholesky decomposition result, perform iterative calculation of the filtering residual according to the one-step estimate at time ; represents the number of iterations, and the initial value is 1;

[0037] 4) Calculate based on the filtering residual of the th iteration The generalized correlation entropy weighted values of the observation channels in the current iteration form a generalized correlation entropy weighted matrix ;

[0038] During the iteration process, use the prediction process of the observation channel kernel function parameters in step S2 to update the shape parameters and kernel bandwidths of the kernel functions of each channel in the weighted matrix ;

[0039] 5) Calculate the Kalman gain using the generalized correlation entropy weighted matrix ; ;

[0040] 6) Perform state estimation according to the Kalman gain to obtain the value of the current iteration ; when the difference between the value of the current iteration and the value of the previous iteration is greater than the threshold, return to step 3) and execute again; otherwise, obtain the best estimated value of the state quantity at the current moment ; when the difference between the value of the current iteration and the value of the previous iteration is greater than the threshold, return to step 3) and execute again; otherwise, obtain the best estimated value of the state quantity at the current moment ; when the difference between the value of the current iteration and the value of the previous iteration is greater than the threshold, return to step 3) and execute again; otherwise, obtain the best estimated value of the state quantity at the current moment ; when the difference between the value of the current iteration and the value of the previous iteration is greater than the threshold, return to step 3) and execute again; otherwise, obtain the best estimated value of the state quantity at the current moment ; .

[0041] Furthermore, at time in the current iteration, the weighted matrix :

[0042] ;

[0043] wherein, is the th channel kernel function,

[0044] the kernel function expression is:

[0045] ;

[0046] is the shape parameter of the kernel function of the th observation channel is the kernel bandwidth. When the system running time is less than the first adjacent time observation window, the shape parameters of each channel are set to 2, is set to ; when it is not the first adjacent time observation window, the shape parameters and kernel bandwidth predicted by the previous adjacent time observation window are adopted

[0047] Furthermore, in the feedback correction of the carrier motion state using the estimated error value,

[0048] use the best estimated value of the state quantity at time in 、 、 correct the position of the INS recurrence state 、speed 、direction cosine matrix to obtain the position, speed, and direction cosine matrix corrected by the estimation error at the current moment 、speed 、direction cosine matrix ;

[0049] ;

[0050] 。

[0051] The present invention can achieve the following beneficial effects:

[0052] The adaptive robust anti-robust information fusion method for GNSS / INS tightly coupled navigation combining neural networks disclosed in the present invention uses a multi-channel heteronuclear generalized maximum correlation entropy Kalman filter to implement combined navigation state estimation, and different shape parameters and kernel bandwidths are used for different observation channels. A convolutional neural (CNN) and gated recurrent unit (GRU) hybrid network is used to monitor the current noise statistical characteristics in real time, focusing on its overall distribution shape and heavy-tailed characteristics, so as to estimate the kernel function shape parameters and kernel bandwidths of each observation channel, and achieve the adaptability of the filter to dynamic time-varying noise. Moreover, the kernel functions of each GNSS observation channel not only have independent shape parameters but also independent kernel bandwidths. Therefore, this method can not only cope with general non-Gaussian distributed noise but also cope with obvious outlier anomalies, and the utilization of the observed quantities of each channel will not interfere with each other;

[0053] Using the CNN-GRU hybrid neural network as a dynamic estimation method for kernel parameters and using the histogram data of noise samples as feature samples can estimate multiple parameters simultaneously, with strong generalization ability and more accurate estimation, can fully extract the features in the samples, and can make full use of the correlation information between samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The drawings are only for the purpose of showing specific embodiments and are not considered to be a limitation of the present invention. Throughout the drawings, the same reference signs denote the same components;

[0055] Figure 1 is a flowchart of the adaptive robust anti-robust information fusion method in the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0056] The following will specifically describe the preferred embodiments of the present invention in conjunction with the drawings, wherein the drawings form a part of this application and are used together with the embodiments of the present invention to explain the principle of the present invention.

[0057] An embodiment of the present invention discloses an adaptive robust anti-robust information fusion method for GNSS / INS tightly coupled navigation combined with a neural network, as Figure 1 shown, including:

[0058] Step S1: For a GNSS / INS tightly coupled navigation system that uses a multi-channel heteronuclear generalized correlation entropy Kalman filter for INS recursive error estimation, establish and train a neural network model for fitting the kernel parameters and error distribution of each channel;

[0059] Step S2: During the INS recursive error estimation process of the GNSS / INS tightly coupled navigation system, when the running time reaches the length of a near-time observation window, use the posterior error of the observations of each satellite within the near-time observation window to establish a sequence of error sample probability distribution histograms, and input the neural network model to predict the kernel function parameters of each channel corresponding to each satellite in the filter;

[0060] Step S3: Use the predicted kernel function parameters of each channel to perform generalized correlation entropy Kalman filtering to achieve INS error state estimation and correct the recursive error; when the running time reaches the length of the next near-time observation window, return to Step S2 to re-predict the kernel function parameters of each channel.

[0061] Specifically, the neural network model for fitting the kernel parameters and error distribution in Step S1 is a CNN-GRU hybrid neural network structure; the input of the model is a sequence of error sample probability distribution histograms, and the output of the model is the shape parameter and kernel bandwidth of the generalized Gaussian kernel function.

[0062] Specifically, since the characteristics of the parameters to be estimated are hidden in a large amount of data in the network structure, in the network structure, first use a convolutional layer (Conv) to extract data features through a convolutional kernel; then perform downsampling on the data through a pooling layer (Pooling) to reduce the data scale, focus on the most effective features, remove redundant information and compress the features, improve the network training and feedforward operation speed, and then use a dropout layer (Dropout) to randomly deactivate some neurons during the model training process to prevent overfitting. Multiple convolutional layers make the feature extraction more complete, and finally use a fully connected layer (Flatten) to map the feature matrix to a feature vector and enter the GRUmodel part for further processing.

[0063] The neurons within each layer of the convolutional neural network are independent of each other, which is suitable for the case where the input data is independent. It performs poorly on time series data with related contexts. Since the features reflecting the sample distribution are hidden in the correlation of the histogram data, it is necessary to take targeted processing. GRU is a type of recurrent neural network (RNN) that is good at capturing time-related information in data sequences and has fewer parameters and longer-term memory capabilities than the long short-term memory network (LSTM). The GRU network is connected in series after the CNN network, and the deep features extracted by the CNN network are used for further fitting. Then, it is output to the dropout layer to further prevent overfitting during training. Finally, the fully connected layer (FC1) matches the dimension with the output data dimension to obtain the neural network output.

[0064] The specific network parameters are configured as shown in the following table:

[0065]

[0066] Specifically, before training the neural network model, the process of generating the training dataset and preprocessing the dataset includes:

[0067] 1) Generate a two-dimensional target grid of the shape parameters and kernel bandwidths of the kernel functions of the multi-channel heteronuclear generalized correlation entropy Kalman filter at a certain step size;

[0068] Among them, the two-dimensional target grid is:

[0069] ;

[0070] Among them, ; The shape parameter is set to have a value range of , and a uniform sampling method is used. The kernel bandwidth is set to have a value range of , and a non-linear sampling method is used. Dense sampling is performed on smaller values, and sparse sampling is performed on larger values. The sampling intervals for dense sampling and sparse sampling can be adaptively adjusted according to requirements.

[0071] 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;

[0072] Substitute each target parameter in into the noise probability density function fitted by the generalized Gaussian kernel function to obtain the probability density function of the random variable e is:

[0073] ;

[0074] Among them, is a normalization parameter. After determining the parameter, a numerical solution can be obtained through 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 , with the corresponding unit being meters.

[0075] 3) Obtain a sample sequence of the random variable e that conforms to this distribution from the probability density function through the Markov chain Monte Carlo algorithm; form a sample data matrix using multiple said sample sequences ;

[0076] ;

[0077] The above formula is a sample data matrix generated by a set of target parameters in, and is composed of mutually independent sample sequences of length . Each sample sequence (i.e., each row in ) has samples, and a total of such sample sequences are generated.

[0078] The setting of needs to consider the size of the filter adjacent time window in the subsequent steps and is determined according to the performance of the operation platform. Here it is set to 500. Since it is offline training,

[0079] Since directly inputting the above data into the network for training has a large amount of data, and the network may eventually assign different weights to data at different positions, which is contradictory to the facts, the sample data must be processed into the form of a histogram.

[0080] 4) Convert the sample data matrix into a training sample matrix composed of a sample probability distribution histogram sequence;

[0081] (1) Set the detection step size;

[0082] Set the detection step size ; It should be set according to the scale of the generated sample sequence , and should be set as small as possible to ensure accuracy.

[0083] (2) Segment at equal intervals within the range of the random variable with the set detection step size; divide the observation error range into multiple segments;

[0084] in the range of the random variable e at intervals as the step size;

[0085] Divide into n segments, ;

[0086] (3) For each sample sequence in the sample data matrix generated by each set of target parameters in the two-dimensional target grid, count the number of samples falling within the range of each segment of the observation error, calculate the proportion relative to the total number, and generate a histogram sequence;

[0087] For each set of target parameters in the generated sample data matrix count the number of samples in each sample sequence (i.e., each row in) falling within each segment, and calculate the proportion relative to the total number , and generate a histogram sequence for each sample sequence in , for example:

[0088] ;

[0089] Then convert the sample data matrix to:

[0090] ;

[0091] which is the training sample matrix corresponding to the target parameter , correspondingly, the above process also needs to be implemented for each pair of target parameters in to obtain the final model training set data.

[0092] Specifically, during the training of the neural network model, the input of the model is the histogram sequence in the training sample matrix, that is, the row vector of, and the output is a 2*1 vector, and the corresponding training target is .

[0093] When training the network, set the batch size to 4 and the learning rate to 0.05, but this does not mean that these parameters are the only choices and can be adjusted appropriately according to the training situation. Use the backpropagation algorithm for iterative optimization, automatically adjust the neuron parameters, stop training after the loss converges, package the model, and use it for subsequent filtering steps.

[0094] When using GNSS / INS tightly coupled navigation, use the INS sensor data to recursively calculate the carrier motion state (such as position, speed, etc.) to obtain the position after pure inertial navigation state recursion and speed , direction cosine matrix ; The recursive method can adopt existing public methods.

[0095] However, due to the existence of sensor errors, the deduced state also contains errors; the errors in the deduced motion state are set as the state variables of the information fusion algorithm , and an error model is established. Then, the predicted value of the error by this model is fused with the navigation information brought by the GNSS observation through Kalman filtering to obtain the optimal estimated value of the error in the current motion state, and this value is used to correct the error in the state, so as to obtain a more accurate current motion state of the carrier: position , speed , direction cosine matrix .

[0096] Specifically, the process of predicting the kernel function parameters of the i-th observation channel in step S2 includes:[[]]

[0097] 1) Set the length of the adjacent time observation window as . When the GNSS / INS tightly coupled navigation system runs for greater than or equal to one adjacent time observation window length, a corrected position at each moment is obtained ;

[0098] 2) Calculate the posterior error sample sequence of the observations at the adjacent -th moment of the -th observation channel according to the corrected positions at each moment ;

[0099] When the filtering process runs to moment, the observation error sequence at the adjacent -th moment of the -th observation channel is ; where

[0100] ;

[0101] is the posterior error of the observation of the -th satellite at moment, is the pseudo-range of the observation of the -th satellite at moment, is the optimal estimated value of the pseudo-range of the -th satellite at moment, which is obtained by taking the modulus of the difference between the three-dimensional coordinates of the satellite position at moment and the three-dimensional coordinates of the corrected user position .

[0102] 3) Use the observation error sample sequence of the th observation channel near the th moment to establish a sequence of error sample probability distribution histograms;

[0103] Use the method in sub-step 4) of step S1 to establish a sequence of sample probability distribution histograms for the

[0104] 4) Input the sequence of error sample probability distribution histograms of the th observation channel into the neural network model to predict the shape parameter and kernel bandwidth of the kernel function of this channel.

[0105] Specifically, in step S3, the state vector of the GNSS / INS tightly coupled navigation system established is:

[0106] ;

[0107] Among them, , , , , , and are respectively the position vector error, velocity error, misalignment angle, gyroscope zero bias, accelerometer zero bias, gyroscope scale factor error, and accelerometer scale factor error of the INS state recursion result in the ECEF coordinate system;

[0108] According to the error perturbation theory, establish the discrete state equation and state equation of the INS error as:

[0109] ;

[0110] ;

[0111] Among them, , are respectively the , state vectors at the time; is the system noise at the time,

[0112] is the state transition matrix; is the observable quantity at the time, is the observation noise at the time; is the observation matrix at the

[0113] Among them, the state transition matrix is as follows:

[0114]

[0115] Among them, is the system discrete time interval, is the angular velocity vector of the Earth's rotation, is the direction cosine matrix from the vehicle coordinate system b to the ECEF coordinate system, is the specific force output by the vehicle accelerometer, , , , are the correlation time parameters of the gyro bias error, accelerometer bias error, gyro scale factor error, and accelerometer scale factor error modeled as a first-order Markov process, respectively;

[0116] The observation quantity is as follows:

[0117] ;

[0118] Among them, is at time the pseudo-range of the th observable satellite, is the geometric distance calculated based on the position deduced by the INS at time and the position of the ; is the number of observable satellites;

[0119] The observation matrix at time is calculated from the known optimal state estimate at time and the satellite position solved from the ephemeris, and is a known quantity.

[0120] The relationship between the satellite pseudo-range observation quantity and the INS-derived state error (i.e., the state quantity X) is established through the observation equation; that is, The observation quantity at time

[0121]

[0122] Among them, is at time the pseudo-range of the th observable satellite, is the three-dimensional coordinate of the position of the The three-dimensional coordinate of the position for INS recursion at time k; i = 1, …, m, where m is the number of observable satellites.

[0123] In the embodiment of the present invention, the multi-channel heteronuclear generalized correlation entropy Kalman filter is adaptively robust based on a neural network, and it is a generalized maximum correlation entropy Kalman filter with adaptive robustness based on a neural network (NN-ARGMCKF). The filter introduces generalized correlation entropy into the Kalman filter, replacing the mean square value of the error as the optimization cost; adopts a multi-channel heteronuclear structure to isolate the shape parameters of the kernel function in different observation channels, and also has an independent kernel bandwidth. Therefore, this method can not only cope with general non-Gaussian distributed noise but also cope with obvious outlier anomalies, and the utilization of the observed values of each channel will not interfere with each other; and uses a CNN-GRU hybrid neural network as a dynamic estimation method for kernel parameters, uses the histogram data of noise samples as feature samples, can estimate multiple parameters simultaneously, has strong generalization ability and more accurate estimation, can fully extract the features in the samples, and make full use of the correlation information between the samples to achieve a better filtering effect.

[0124] Specifically, in step S3, the process of estimating the INS state recursion error by NN-ARGMCKF includes:

[0125] 1) Perform a one-step state prediction to obtain The one-step estimate at time Of the covariance matrix :

[0126]

[0127]

[0128] Is The optimal estimate of the error state at time, Is obtained using And the state transition matrix Calculated The one-step estimate of the error state at time, Is the covariance matrix of the one-step estimate , Is The estimated error covariance matrix at time, Is The system noise covariance matrix at time;

[0129] Here And Will be used later to calculate The optimal estimated value of the error state at time ;

[0130] 2) According to the one-step estimation at a moment and the covariance matrix and the covariance matrix of the observation noise perform Cholesky decomposition;

[0131] ;

[0132] is the covariance matrix of the observation noise in the observation equation.

[0133] 3) According to the Cholesky decomposition result, based on the one-step estimation at a moment perform iterative calculation of the filtering residual ;

[0134] ;

[0135] where , is the observed quantity at a moment; n is the dimension of the state vector;

[0136] , is the observation matrix at a moment;

[0137] When step 3) is first executed at the same moment, take in step 1), and then calculate according to the iterative value; represents the number of iterations, and the initial value is 1.

[0138] 4) Calculate the generalized correlation entropy weighted value of each observed quantity channel for the -th iteration of the filtering residual, and form the generalized correlation entropy weighted matrix ;

[0139] During the iteration process, use the observation channel kernel function parameter prediction process in step S2 to update the shape parameter and kernel bandwidth of each channel kernel function in the weighted matrix ;

[0140] At a moment the weighted matrix for the

[0141] -th iteration:

[0142] where is the -th channel kernel function,

[0143] Kernel function The expression is:

[0144] ;

[0145] is the shape parameter of the kernel function for the th observation channel is the kernel bandwidth. When the system running time is less than the first adjacent time observation window, the shape parameters of each channel are set to 2, is set to ; When it is not the first adjacent time observation window, the shape parameter and kernel bandwidth predicted by the previous adjacent time observation window are adopted.

[0146] Each observation channel adopts different shape parameters to enable it to face the problem of different non-Gaussian noises in different channels, while each observation channel adopts different kernel bandwidths to effectively control the influence of the channel with larger outliers in the filter, while the normal channels can be unaffected, showing stronger robustness and outlier resistance.

[0147] 5) Calculate the Kalman gain using the generalized correlation entropy weighted matrix ; ;

[0148] ;

[0149] ;

[0150] where, is the observation error covariance matrix weighted by the correlation entropy for each observation channel;

[0151] 6) Perform state estimation according to the Kalman gain to obtain the th iteration value ; When the difference between the th iteration value and the th iteration value is greater than the threshold, return to step 3) and execute again; otherwise, obtain the best estimated value of the state quantity at the current moment .

[0152] Specifically, ;

[0153] is the iteration value obtained in the th iteration, and is calculated from the th iteration result.

[0154] At this time, it is judged that if:

[0155]

[0156] Jump to step 3) and execute again, otherwise obtain the best estimate of the state quantity at the current moment ; is the set threshold value.

[0157] In the feedback correction of the carrier motion state using the estimated error value,

[0158] Use the best estimate of the state quantity at the moment in , , to correct the position of the INS recursive state , speed , and direction cosine matrix , and obtain the position , speed , and direction cosine matrix after being corrected by the estimated error at the current moment;

[0159] ;

[0160] .

[0161] In summary, the adaptive robust anti - difference information fusion method for GNSS / INS tight integration navigation combined with neural network in the embodiment of the present invention uses a multi - channel heteronuclear 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. The convolutional neural (CNN) and gated recurrent unit (GRU) hybrid network is used to monitor the current noise statistical characteristics in real time, focusing on its overall distribution shape and heavy - tail characteristics, so as to estimate the kernel function shape parameters and kernel bandwidths of each observation channel, and realize the adaptability of the filter to dynamic time - varying noise. And,

[0162] The kernel functions of each GNSS observation channel not only have independent shape parameters, but also have independent kernel bandwidths. Therefore, this method can not only cope with general non - Gaussian distributed noise, but also cope with obvious outlier anomalies, and the utilization of the observed quantities of each channel will not interfere with each other;

[0163] Using the CNN - GRU hybrid neural network as the dynamic estimation method of kernel parameters, and using the histogram data of noise samples as feature samples, it can estimate multiple parameters simultaneously, has strong generalization ability and more accurate estimation, can fully extract the features in the samples, and can make full use of the correlation information between samples.

[0164] In summary, the above are only the preferred specific embodiments 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 those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. An adaptive robust anti-robust information fusion method for GNSS / INS tightly integrated navigation combined with neural network, characterized in that Including: Step S1: For a GNSS / INS tightly coupled navigation system that uses a multi-channel heteronuclear generalized correlation entropy Kalman filter for INS recursive error estimation, establish and train a neural network model for fitting the kernel parameters and error distribution of each channel. Step S2: During the INS recursive error estimation process of the GNSS / INS tightly coupled navigation system, when the running time reaches the length of a neighboring time observation window, use the posterior errors of the observations of each satellite within the neighboring time observation window to establish a sequence of error sample probability distribution histograms, and input them into the neural network model to predict the kernel function parameters of each channel corresponding to each satellite in the filter. Step S3: Use the predicted kernel function parameters of each channel to perform generalized correlation entropy Kalman filtering to achieve INS error state estimation and correct the recursive error; when the running time reaches the length of the next neighboring time observation window, return to Step S2 to re-predict the kernel function parameters of each channel. In step S2, the process of predicting the kernel function parameters of the th observation channel includes: 1) Set the length of the near-time observation window to , and when the GNSS / INS tightly coupled navigation system operates for greater than or equal to one near-time observation window length, obtain a corrected position at each moment ; 2) According to the corrected positions at each moment Calculate the posterior error sample sequence of the observed quantities of the observation channels at the adjacent moments; 3) Using the observation error sample sequences of the th observation channel near the time instants to establish a sequence of error sample probability distribution histograms; 4) Input the error sample probability distribution histogram sequence of the th observation channel into the neural network model to predict the shape parameter and kernel bandwidth of this channel.

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

3. The adaptive robust anti-interference information fusion method for GNSS / INS tightly coupled navigation combined with a neural network according to claim 2, 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 the shape parameters and kernel bandwidths of each kernel function of the multi-channel heteronuclear generalized correlation entropy Kalman filter at 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) Use the Markov chain Monte Carlo algorithm for the probability density function to obtain a sample sequence of random variables that conform to this probability distribution, and form a sample data matrix using multiple such sample sequences. 4) Convert the sample data matrix into a training sample matrix composed of a sequence of sample probability distribution histograms.

4. The adaptive robust anti-interference information fusion method for GNSS / INS tightly coupled navigation combined with a neural network according to claim 3, characterized in that converting the sample data matrix into a training sample matrix composed of a sequence of sample probability distribution histograms includes: (1) Set the detection step size. (2) Segment the range of the random variable 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 group of target parameters in the two-dimensional target grid, count the number of samples falling within each segment of the observation error range, and calculate the proportion relative to the total number to generate a histogram sequence.

5. The adaptive robust anti-interference information fusion method for GNSS / INS tightly coupled navigation combined with a neural network according to claim 1, characterized in that When the filtering process runs to moment, the observation error sequence of the th moment is ; where ; be the posterior error of the observation of the be the pseudorange observed for the satellite at be the optimal estimated value of the pseudorange of the satellite at obtained by taking the modulus of the difference between the three-dimensional coordinates of the satellite position at and the three-dimensional coordinates of the user position after correction ​ 6. The adaptive robust anti-robust information fusion method for GNSS / INS tightly coupled navigation combined with a neural network according to claim 1, wherein In the INS recursive error estimation process, the state vector of the GNSS / INS tightly coupled navigation system established is: ； Among them, , , , , , and are respectively the position vector error, velocity error, misalignment angle, gyroscope zero bias, accelerometer zero 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.

7. The adaptive robust anti-robust information fusion method for GNSS / INS tightly coupled navigation combined with a neural network according to claim 6, wherein The process of performing INS state recursive error includes: 1) Perform one-step state prediction to obtain One-step estimation at Covariance matrix of : 2) According to one-step estimation at the moment of the covariance matrix and the covariance matrix of the observation noise perform Cholesky decomposition; 3) Based on the Cholesky decomposition result, perform one-step estimation at time instant and calculate the iterative filtering residuals ; Let represent the number of iterations, with an initial value of 1; 4) Calculate, according to the filtering residuals of the th iteration, the generalized correlation entropy weighted values of each observable channel of the th iteration, and form a generalized correlation entropy weighted matrix During the iteration process, the shape parameters and kernel bandwidths of the kernel functions of each channel in the weighted matrix are updated by using the prediction process of the observation channel kernel function parameters in step S2. 5) Utilize the generalized correlation entropy weighted matrix Calculate the Kalman gain ; 6) According to the Kalman gain perform state estimation to obtain the value of the ith iteration; when the difference between the value of the ith iteration and the value of the (i - 1)th iteration is greater than the threshold, return to step 3) and execute again; otherwise, obtain the best estimated value of the state quantity at the current moment .

8. The adaptive robust anti-robust information fusion method for GNSS / INS tightly coupled navigation combined with a neural network according to claim 7, wherein moment weighted matrix for the th iteration ; Among them, is the channel kernel function Kernel function The expression is: ； is the shape parameter of the kernel function for the th observation channel, is the kernel bandwidth. When the system running time is less than the first adjacent time observation window, the shape parameters of each channel are set to 2, is set to ; when it is not the first adjacent time observation window, the shape parameter and kernel bandwidth predicted by the previous adjacent time observation window are adopted.

9. The adaptive robust anti-robust information fusion method for GNSS / INS tightly coupled navigation combined with a neural network according to claim 7, wherein In the feedback correction of the carrier motion state using the estimated error value Use The best estimated value of the moment state quantity in , , to correct the position of the INS recursive state , speed , and direction cosine matrix , and obtain the position , speed , and direction cosine matrix after being corrected by the estimation error at the current moment; ； 。

Images