Autonomous navigation system and method based on fault detection and neural network assistance

Abstract

The invention discloses an autonomous navigation system and method based on fault detection and neural network assistance, and relates to the technical field of navigation, the system comprises an INS system, a GNSS system, a fault detection module, an RFR-GRU neural network module and an adaptive Kalman filtering module; the method comprises the following steps: S1, acquiring information, S2, judging the state of a GNSS (Global Navigation Satellite System), and when the GNSS is in an effective state, enabling the navigation system to enter a training mode, and specifically comprises the following steps: S2.1, acquiring accurate positioning information, and S2.2, training an RFR-GRU neural network model; and when the GNSS system is in a failure state, the navigation system enters a prediction mode, and the method specifically comprises the following steps: S2.3, predicting positioning information. The method has the beneficial effects that the robustness and the positioning precision of the system are improved, whether the GNSS system is in an effective state can be accurately and effectively judged, then whether the autonomous navigation system works in a training mode or a prediction mode is determined, and the navigation positioning precision is kept.

Description

Technical Field

[0001] The patent of this invention relates to the field of navigation technology, and specifically to an autonomous navigation system and method based on fault detection and neural network assistance. Background Art

[0002] The inertial navigation system (INS) / global navigation satellite system (GNSS) combined navigation technology fully integrates the advantages of INS autonomy, all-weather operation and GNSS high-precision measurement. The navigation method based on the INS / GNSS combination is a commonly used navigation method that provides high-precision attitude, speed, position and other parameter measurements for mobile carriers such as vehicles. However, in actual application, due to the inherent defects of GNSS, such as multipath effect and poor anti-interference ability, the INS / GNSS system will inevitably be interfered by the external environment (such as high-rise buildings, mountains, trees, electromagnetic interference and various tunnels) during the operation of the carrier, resulting in GNSS signal attenuation or even loss of lock. During this satellite denial period, GNSS cannot provide observation information, and the combined navigation system can only operate in pure INS mode. The error accumulates over time and diverges uninhibitedly, affecting the navigation accuracy.

[0003] At present, the main methods to improve the performance of INS / GNSS integrated navigation system under satellite denial conditions include integrated navigation methods based on tight combination and ultra-tight combination technology, methods to assist INS navigation by adding other sensor information, and methods to assist INS navigation based on artificial intelligence. However, although tight combination and ultra-tight combination technology can maintain good system navigation performance under complex conditions such as multipath effects and external interference, their combination process is extremely complicated, and when the satellite is completely unavailable, this method still cannot fundamentally solve the problem. The method of adding other sensors to assist INS can effectively suppress the navigation error of INS, and adding other sensors will inevitably increase the cost and complexity of the navigation system.

[0004] The integrated navigation method assisted by artificial intelligence technology represented by neural networks is relatively feasible and efficient. Various theoretical models and algorithms based on neural networks have been introduced into integrated navigation to improve the continuity of the integrated navigation system and improve the INS / GNSS navigation accuracy under satellite denial conditions. The main idea of ​​this type of solution is to build a neural network module for training when satellite signals are available; when satellites are denied, use the trained neural network module to predict the specific parameter information required during GNSS denial, so that the integrated navigation system can keep working normally.

[0005] However, existing methods based on neural network-assisted navigation often only use a single neural network structure for prediction, and a large number of training samples are required when training the neural network. In actual applications, the samples collected for neural network training are few and incomplete. In this case, only using a single-structure neural network for training is prone to problems such as overfitting of network training, local optimization, and unstable prediction model, resulting in unsatisfactory prediction results. In addition, when performing neural network-assisted navigation, it is not considered when to use neural network prediction intervention, that is, it is not considered how to judge when GNSS information is unavailable, so that the neural network enters the prediction mode. Summary of the invention

[0006] An object of the present invention is to provide an autonomous navigation system based on fault detection and neural network assistance, which improves system robustness and positioning accuracy.

[0007] The second objective of the present invention is to provide an autonomous navigation method based on fault detection and neural network assistance to ensure navigation positioning accuracy.

[0008] The first object of the present invention is implemented by the following technical solution: an autonomous navigation system based on fault detection and neural network assistance, which includes: an INS system, a GNSS system, a fault detection module, an RFR-GRU neural network module and an adaptive Kalman filter module; The INS system is used to obtain positioning information of speed, position and attitude, and transmit it to the RFR-GRU neural network module and the adaptive Kalman filter module; The GNSS system is used to obtain speed and position measurement information and transmit the measurement information to the fault detection module; The fault detection module is used to determine whether the GNSS system is in a valid state; when in a valid state, the measurement information of the GNSS system is transmitted to the RFR-GRU neural network module and the adaptive Kalman filter module; The RFR-GRU neural network module is used to obtain pseudo GNSS measurement information based on the position, velocity and attitude acquired by the INS system; In the invalid state, the RFR-GRU neural network module sends the pseudo GNSS measurement information to the adaptive Kalman filter module; The adaptive Kalman filter module is used to fuse the positioning information and the measurement information or the pseudo GNSS measurement information to obtain a compensation value of the positioning information of the INS system, thereby obtaining accurate positioning information.

[0009] Furthermore, the INS system includes an IMU device and an inertial navigation mechanical arrangement device; The IMU device includes an accelerometer and a gyroscope, which are used to measure acceleration and angular velocity respectively, and send the measurement data to the inertial navigation mechanical arrangement device; The inertial navigation mechanical arrangement device is used to calculate the position, speed and attitude according to the measurement data by using the mechanical arrangement algorithm.

[0010] Furthermore, the accelerometer bias and the gyroscope bias can be obtained through the adaptive Kalman filter module and sent to the IMU device for correction.

[0011] The second objective of the present invention is implemented by the following technical solution: an autonomous navigation method based on fault detection and neural network assistance, which comprises the following steps: S1. Obtain information: Use the INS system to obtain positioning information of speed, position and attitude; use the GNSS system to obtain measurement information of speed and position; S2. Determine the status of the GNSS system: use a fault detection module to determine whether the GNSS system in step S1 is in a valid state; When the GNSS system is in an effective state, the navigation system enters a training mode, which specifically includes the following steps: S2.1. Obtain accurate positioning information: Establish a combined navigation adaptive Kalman filter model, obtain a compensation value of the positioning information of the INS system by fusing the positioning information and the measurement information in step S1, and then obtain accurate positioning information; S2.2, training the RFR-GRU neural network model: training the RFR-GRU neural network model by using the positioning information and the measurement information in step S1 as input features of the RFR-GRU neural network model; finding the relationship between the positioning information provided by the INS system and the measurement information provided by the GNSS system; When the GNSS system is in a failed state, the navigation system enters a prediction mode, which specifically includes the following steps: S2.3, predicting positioning information: by inputting the positioning information in step S1 into the RFR-GRU neural network model in step S2.2, pseudo GNSS measurement information is obtained, and the positioning information in step S1 is transmitted to the adaptive Kalman filter model in step S2.1 for data fusion to obtain the compensation value of the positioning information of the INS system, thereby obtaining accurate positioning information.

[0012] Furthermore, the algorithm of the adaptive Kalman filter model in step S2.1 includes the state equation and the observation equation, which are described in detail as follows: (1) Where k is the discrete time, is the state vector of the system, which can be set to . is the position error vector of the INS system, is the velocity error vector of the INS system, is the attitude angle error vector of the INS system, and are the gyro drift and accelerometer bias of the INS system, is the observation vector of the system. In this model It represents the difference in position and speed between GNSS and INS, that is, ; are system structure parameters, representing the state one-step transfer matrix, system noise allocation matrix, and observation matrix respectively; is the system noise vector, is the observation noise vector and satisfies , ; is the Kronecker function, and They are and The covariance matrix of Status prediction: (2) Calculate the state one-step prediction covariance: introduce an adaptive factor based on the new information sequence to suppress the impact of various complex noises on the filtering effect, and adjust the prior prediction covariance in real time; (3) (4) Observation update: construct measurement updates (5) Filter gain: (6) Corresponding state quantity update part: (7) State estimation covariance: (8) in is the adaptive factor.

[0013] Furthermore, the method in step S2 for the fault detection module to determine whether the GNSS system is in a valid state specifically comprises the following steps: Use formula (5) to get the first n new information sequences, and calculate the average of the first n new information sequences: (9) (10) in, is the judgment threshold, and its value is 1.

[0014] Furthermore, the adaptive factor The algorithm is as follows: Using the residual vector To construct the adaptive factor, the residual vector Can reflect the predicted state vector The error of constructing error discriminant statistics: (11) in represents the trace of the innovation covariance matrix, which can be determined by the following formula (12) The adaptive factor is constructed from : (13) Where c is a constant whose value is 1.

[0015] Furthermore, the RFR-GRU neural network model in step S2.2 is a weighted fusion model of the random forest regression model RFR and the gated recurrent unit GRU. Specifically, the GRU is optimized using RFR to improve the model accuracy of the GRU when training the GRU with small samples, and the RFR-GRU neural network is used to output pseudo GNSS measurement information.

[0016] Furthermore, the specific modeling method of the RFR-GRU neural network model includes the following steps: 1) Constructing an original training sample set S: taking the positioning information and the measurement information set in step S1 as the original training sample S, wherein the original training sample S has T samples.

[0017] 2) Construct sub-training sets: From the original training sample set S in step 1), random sampling is performed through bootstrap sampling, a total of N samplings are performed to construct N sub-training sets, each with T samples; 3) Training the GRU model: train a GRU model for each sub-training set in step 2); each GRU model gives a prediction value, and the prediction values ​​of N GRU models are arithmetic averaged to obtain the final regression result output; 4) Test the GRU model: Use the original training sample set S in step 1) as the test set to test the GRU model trained in step 3) and output the final test results.

[0018] Furthermore, the GRU model specifically includes an update gate ( ) and reset gate ( ): Among them, the update gate ( ) is used to capture long-term dependencies in time series: (14) Reset Gate( ) is used to capture short-term dependencies in time series: (15) in, is the state output at the previous time t-1, is the input data at the current time t, and both gates are controlled by the input data at the current time t and the output state at the previous time t-1 To decide; the sigmoid function converts the data into values ​​in the range of 0-1; Combined update gate and reset gate outputs: (16) in, , and are the weight matrices of the output state at the previous time t-1 to the update gate, the reset gate, and the output state at the current time t; , and They are the weight matrices for input data to the update gate, reset gate, and output state respectively; , and are the offset vectors of the corresponding structures, and the functions is a non-linear activation function.

[0019] Advantages of the present invention: 1. The present invention provides an autonomous navigation system based on fault detection and neural network assistance. By constructing an RFR-GRU neural network module, when the satellite signal is intact, the neural network module is used to find the relationship between the INS positioning information and the GNSS measurement information; when the satellite is locked and in a failed state, the pseudo GNSS measurement information output by the neural network module is used to replace the original GNSS measurement information for combined navigation, thereby improving the robustness and positioning accuracy of the system.

[0020] 2. The present invention provides an autonomous navigation method based on fault detection and neural network assistance, in which the first n new information sequences are obtained by constructing measurement new information, the average of the first n new information sequences is calculated, and it is determined whether the average is ≤1. If it is ≤1, the GNSS system is determined to be in a valid state. If it is >1, the GNSS system is determined to be in an invalid state. It can accurately and effectively determine whether the GNSS system is in a valid state, and then decide whether the autonomous navigation system is working in training mode or prediction mode. 3. The present invention provides an autonomous navigation method based on fault detection and neural network assistance, in which an adaptive factor based on the construction of the new information sequence is introduced into the algorithm of the adaptive Kalman filter model to suppress the influence of various complex noises on the filtering effect, adjust the prior prediction covariance in real time, and suppress the influence of changes in noise characteristics on the accuracy of the combined filter.

[0021] 4. The present invention provides an autonomous navigation method based on fault detection and neural network assistance, in which the GRU is optimized by a random forest regression model so that it can still achieve good prediction accuracy under small sample training conditions; and the GRU can fully consider and utilize historical data to increase the prediction accuracy of the neural network model, and has the advantages of simple structure and few parameters. It can effectively predict pseudo-GNSS measurement information when the GNSS fails, thereby maintaining navigation positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1 This is an autonomous navigation system based on fault detection and neural network assistance in training mode according to Example 2 of the present invention.

[0024] Figure 2 This is an autonomous navigation system based on fault detection and neural network assistance in the prediction mode of Example 2 of the present invention. DETAILED DESCRIPTION

[0025] The present invention will be further described in detail below by way of examples.

[0026] Embodiment 1: An autonomous navigation system based on fault detection and neural network assistance, comprising: an INS system, a GNSS system, a fault detection module, an RFR-GRU neural network module and an adaptive Kalman filter module; The INS system is used to obtain positioning information of speed, position and attitude, and transmit it to the RFR-GRU neural network module and the adaptive Kalman filter module (AKF); the GNSS system is used to obtain measurement information of speed and position, and transmit the measurement information to the fault detection module; the fault detection module is used to determine whether the GNSS system is in a valid state; when in a valid state, the measurement information of the GNSS system is transmitted to the RFR-GRU neural network module and the adaptive Kalman filter module; the RFR-GRU neural network module is used to obtain pseudo-GNSS measurement information according to the position, speed and attitude obtained by the INS system; when in an invalid state, the RFR-GRU neural network module sends the pseudo-GNSS measurement information to the adaptive Kalman filter module; the adaptive Kalman filter module is used to fuse the positioning information and the measurement information or the pseudo-GNSS measurement information to obtain a compensation value of the positioning information of the INS system, thereby obtaining accurate positioning information.

[0027] The INS system includes an IMU device and an inertial navigation mechanical arrangement device; The IMU device includes an accelerometer and a gyroscope, which are used to measure acceleration and angular velocity respectively, and send the measurement data to the inertial navigation mechanical arrangement device; the inertial navigation mechanical arrangement device is used to calculate the position, speed and attitude according to the measurement data using a mechanical arrangement algorithm.

[0028] The accelerometer bias and gyroscope bias can also be obtained through the adaptive Kalman filter module and sent to the IMU device for correction.

[0029] The present invention constructs an RFR-GRU neural network module. When the satellite signal is intact, the neural network module is used to find the relationship between the INS positioning information and the GNSS measurement information. When the satellite is locked and in a failed state, the pseudo GNSS measurement information output by the neural network module is used to replace the original GNSS measurement information to perform combined navigation, thereby improving the robustness and positioning accuracy of the system.

[0030] Embodiment 2: An autonomous navigation method based on fault detection and neural network assistance, comprising the following steps: S1. Obtain information: Use the INS system to obtain positioning information of speed, position and attitude; use the GNSS system to obtain measurement information of speed and position.

[0031] S2. Determine the status of the GNSS system: use the fault detection module to determine whether the GNSS system in step S1 is in a valid state.

[0032] The method for the fault detection module to determine whether the GNSS system is in a valid state specifically comprises the following steps: Use formula (5) to get the first n new information sequences, and calculate the average of the first n new information sequences: (9) (10) in, is the judgment threshold, and its value is 1.

[0033] like Figure 1 As shown, when the GNSS system is in an effective state, the navigation system enters a training mode, which specifically includes the following steps: S2.1. Obtain accurate positioning information: Establish a combined navigation adaptive Kalman filter model, obtain a compensation value of the positioning information of the INS system by fusing the positioning information and the measurement information in step S1, and then obtain accurate positioning information.

[0034] Among them, the algorithm of the adaptive Kalman filter model includes the state equation and the observation equation, which are described in detail as follows: (1) Where k is the discrete time, is the state vector of the system, which can be set to . is the position error vector of the INS system, is the velocity error vector of the INS system, is the attitude angle error vector of the INS system, and are the gyro drift and accelerometer bias of the INS system, is the observation vector of the system. In this model It represents the difference in position and speed between GNSS and INS, that is, ; are system structure parameters, representing the state one-step transfer matrix, system noise allocation matrix, and observation matrix respectively; is the system noise vector, is the observation noise vector and satisfies , ; is the Kronecker function, and They are and The covariance matrix of .

[0035] Status prediction: (2) Calculate the state one-step prediction covariance: introduce an adaptive factor based on the new information sequence to suppress the impact of various complex noises on the filtering effect, and adjust the prior prediction covariance in real time; (3) (4) Observation update: construct measurement updates (5) Filter gain: (6) Corresponding state quantity update part: (7) State estimation covariance: (8) in is the adaptive factor; adaptive factor The algorithm is as follows: Using the residual vector To construct the adaptive factor, the residual vector Can reflect the predicted state vector The error of constructing error discriminant statistics: (11) in represents the trace of the innovation covariance matrix, which can be determined by the following formula (12) The adaptive factor is constructed from : (13) Where c is a constant whose value is 1.

[0036] S2.2. Training the RFR-GRU neural network model: training the RFR-GRU neural network model by using the positioning information and the measurement information in step S1 as input features of the RFR-GRU neural network model; finding the relationship between the positioning information provided by the INS system and the measurement information provided by the GNSS system.

[0037] The RFR-GRU neural network model is a weighted fusion model of the random forest regression model RFR and the gated recurrent unit GRU. Specifically, the GRU is optimized using RFR to improve the model accuracy of the GRU when training the GRU with small samples, and the RFR-GRU neural network is used to output pseudo GNSS measurement information.

[0038] The specific modeling method of the RFR-GRU neural network model includes the following steps: 1) Constructing an original training sample set S: taking the positioning information and the measurement information set in step S1 as the original training sample S, wherein the original training sample S has T samples.

[0039] 2) Construct sub-training sets: From the original training sample set S in step 1), random sampling is performed through bootstrap sampling, a total of N samplings are performed to construct N sub-training sets, each with T samples; 3) Training the GRU model: train a GRU model for each sub-training set in step 2); each GRU model gives a prediction value, and the prediction values ​​of N GRU models are arithmetic averaged to obtain the final regression result output; 4) Test the GRU model: Use the original training sample set S in step 1) as the test set to test the GRU model trained in step 3) and output the final test results.

[0040] The GRU model specifically includes an update gate ( ) and reset gate ( ): Among them, the update gate ( ) is used to capture long-term dependencies in time series: (14) Reset Gate( ) is used to capture short-term dependencies in time series: (15) in, is the state output at the previous time t-1, is the input data at the current time t, and both gates are controlled by the input data at the current time t and the output state at the previous time t-1 To decide; the sigmoid function converts the data into values ​​in the range of 0-1; Combined update gate and reset gate outputs: (16) in, , and are the weight matrices of the output state at the previous time t-1 to the update gate, the reset gate, and the output state at the current time t; , and They are the weight matrices for input data to the update gate, reset gate, and output state respectively; , and are the offset vectors of the corresponding structures, and the functions is a non-linear activation function.

[0041] like Figure 2 As shown, when the GNSS system is in a failed state, the navigation system enters a prediction mode, which specifically includes the following steps: S2.3, predicting positioning information: by inputting the positioning information in step S1 into the RFR-GRU neural network model in step S2.2, pseudo GNSS measurement information is obtained, and the positioning information in step S1 is transmitted to the adaptive Kalman filter model in step S2.1 for data fusion to obtain the compensation value of the positioning information of the INS system, thereby obtaining accurate positioning information.

[0042] Experimental verification: The autonomous navigation method disclosed in Example 2 is compared with various navigation methods. The vehicle navigation test was carried out using the INS / GNSS integrated navigation system prototype, and the INS / GNSS data was collected when the vehicle test platform was running to evaluate the proposed method. This test designed a GNSS signal interruption test on the road. The specific scheme is: when the INS / GNSS integrated navigation system is working normally, the adaptive Kalman filter combined navigation algorithm is executed when the satellite signal is available, and the vehicle platform continues to run for 400s. In this case, the samples collected for training the neural network are relatively small, and then the GNSS signal is interrupted, and the vehicle platform continues to run for 200s under satellite denial conditions. That is, the total running time of the vehicle test is 600 seconds, of which 200 seconds are run under GNSS denial. This verification uses the first 400s as the neural network training data, and then the GNSS is interrupted for 200s, and the neural network model begins to predict and assist the adaptive Kalman filter for combined navigation. At the same time, this verification selects pure inertial navigation, other LSTM and GRU networks, and RFR-GRU neural network assisted navigation methods without fault detection for comparison, and the results are shown in Table 1.

[0043] Table 1 Comparison of test results

[0044] It can be seen from the test results in Table 1 that, compared with the methods of pure INS, LSTM+AKF, GRU+AKF and RFR-GRU+AKF without a fault-free detection module, the autonomous navigation method disclosed in Example 2 of the present invention more effectively improves the problem of velocity and position error divergence of the navigation system during satellite denial, and improves the robustness and positioning accuracy of the system.

[0045] The above are preferred embodiments of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. An autonomous navigation system based on fault detection and neural network assistance, comprising: INS system and GNSS system; characterized in that it also includes: a fault detection module, an RFR-GRU neural network module and an adaptive Kalman filter module; The fault detection module is used to determine whether the GNSS system is in a valid state; when in a valid state, the measurement information of the GNSS system is transmitted to the RFR-GRU neural network module and the adaptive Kalman filter module; The RFR-GRU neural network module is used to obtain pseudo GNSS measurement information based on the position, velocity and attitude acquired by the INS system; The RFR-GRU neural network model of the RFR-GRU neural network module is a weighted fusion model of a random forest regression model RFR and a gated recurrent unit GRU. The RFR is used to optimize the GRU to improve the model accuracy of the GRU when training the GRU with a small sample, and the RFR-GRU neural network is used to output pseudo GNSS measurement information; In the invalid state, the RFR-GRU neural network module sends the pseudo GNSS measurement information to the adaptive Kalman filter module; The adaptive Kalman filter module is used to fuse the positioning information and the measurement information or the pseudo GNSS measurement information to obtain a compensation value of the positioning information of the INS system, thereby obtaining accurate positioning information.

2. The autonomous navigation system based on fault detection and neural network assistance according to claim 1, characterized in that: The INS system includes an IMU device and an inertial navigation mechanical arrangement device; The IMU device includes an accelerometer and a gyroscope, which are used to measure acceleration and angular velocity respectively, and send the measurement data to the inertial navigation mechanical arrangement device; The inertial navigation mechanical arrangement device is used to calculate the position, speed and attitude according to the measurement data by using the mechanical arrangement algorithm.

3. The autonomous navigation system based on fault detection and neural network assistance according to claim 2 is characterized in that: The accelerometer bias and gyroscope bias can also be obtained through the adaptive Kalman filter module and sent to the IMU device for correction.

4. The autonomous navigation method of an autonomous navigation system based on fault detection and neural network assistance as claimed in any one of claims 1 to 3, characterized in that: It includes the following steps: S1. Obtain information: Use the INS system to obtain positioning information of speed, position and attitude; use the GNSS system to obtain measurement information of speed and position; S2. Determine the status of the GNSS system: use a fault detection module to determine whether the GNSS system in step S1 is in a valid state; When the GNSS system is in an effective state, the navigation system enters a training mode, which specifically includes the following steps: S2.

1. Obtain accurate positioning information: Establish a combined navigation adaptive Kalman filter model, obtain a compensation value of the positioning information of the INS system by fusing the positioning information and the measurement information in step S1, and then obtain accurate positioning information; S2.2, training the RFR-GRU neural network model: training the RFR-GRU neural network model by using the positioning information and the measurement information in step S1 as input features of the RFR-GRU neural network model; finding the relationship between the positioning information provided by the INS system and the measurement information provided by the GNSS system; When the GNSS system is in a failed state, the navigation system enters a prediction mode, which specifically includes the following steps: S2.3, predicting positioning information: by inputting the positioning information in step S1 into the RFR-GRU neural network model in step S2.2, pseudo GNSS measurement information is obtained, and the positioning information in step S1 is transmitted to the adaptive Kalman filter model in step S2.1 for data fusion to obtain the compensation value of the positioning information of the INS system, thereby obtaining accurate positioning information.

5. The autonomous navigation method according to claim 4, characterized in that: The algorithm of the adaptive Kalman filter model in step S2.1 includes the state equation and the observation equation, which are described in detail as follows: （1） Where k is the discrete time, is the state vector of the system, which can be set to ; is the position error vector of the INS system, is the velocity error vector of the INS system, is the attitude angle error vector of the INS system, and are the gyro drift and accelerometer bias of the INS system, is the observation vector of the system. In this model It represents the difference in position and speed between GNSS and INS, that is, ; are system structure parameters, representing the state one-step transfer matrix, system noise allocation matrix, and observation matrix respectively; is the system noise vector, is the observation noise vector and satisfies , ; is the Kronecker function, and They are and The covariance matrix of Status prediction: （2） Compute the state one-step forecast covariance: （3） （4） Observation update: construct measurement updates （5） Filter gain: （6） Corresponding state quantity update part: （7） State estimation covariance: （8） in is the adaptive factor.

6. An autonomous navigation method according to claim 5, characterized in that: The method in step S2 for the fault detection module to determine whether the GNSS system is in a valid state specifically comprises the following steps: Use formula (5) to get the first n new information sequences, and calculate the average of the first n new information sequences: （9） （10） in, is the judgment threshold, and its value is 1.

7. The autonomous navigation method according to claim 5, characterized in that: Adaptive Factor The algorithm is as follows: Using the residual vector To construct the adaptive factor, the residual vector Can reflect the predicted state vector The error of constructing error discriminant statistics: （11） in represents the trace of the innovation covariance matrix, which can be determined by the following formula （12） The adaptive factor is constructed from : （13） Where c is a constant whose value is 1.

8. The autonomous navigation method according to claim 4, characterized in that: The specific modeling method of the RFR-GRU neural network model in step S2.2 includes the following steps: 1) Constructing an original training sample set S: taking the positioning information and the measurement information set in step S1 as the original training sample S, wherein the original training sample S has T samples; 2) Construct sub-training sets: From the original training sample set S in step 1), random sampling is performed through bootstrap sampling, a total of N samplings are performed to construct N sub-training sets, each with T samples; 3) Training the GRU model: train a GRU model for each sub-training set in step 2); each GRU model gives a prediction value, and the prediction values ​​of N GRU models are arithmetic averaged to obtain the final regression result output; 4) Test the GRU model: Use the original training sample set S in step 1) as the test set to test the GRU model trained in step 3) and output the final test results.

9. An autonomous navigation method according to claim 8, characterized in that: The GRU model specifically includes an update gate ( ) and reset gate ( ): Among them, the update gate ( ) is used to capture long-term dependencies in time series: （14） Reset Gate( ) is used to capture short-term dependencies in time series: （15） in, is the state output at the previous time t-1, is the input data at the current time t, and both gates are controlled by the input data at the current time t and the output state at the previous time t-1 To decide; the sigmoid function converts the data into values ​​in the range of 0-1; Combined update gate and reset gate outputs: （16） in, , and are the weight matrices of the output state at the previous time t-1 to the update gate, the reset gate, and the output state at the current time t; , and They are the weight matrices for input data to the update gate, reset gate, and output state respectively; , and are the offset vectors of the corresponding structures, and the functions is a non-linear activation function.

Images