Autonomous Navigation System and Method Based on Fault Detection and Neural Network Assistance

Abstract

The present invention discloses an autonomous navigation system and method based on fault detection and neural network assistance, which relates to the field of navigation technology. The system includes an INS system, a GNSS system, a fault detection module, an RFR-GRU neural network module, and an adaptive Kalman filter module; the method includes: S1, obtaining information; S2, judging the state of the GNSS system. When the GNSS system is in an effective state, the navigation system enters the training mode, which specifically includes the following steps: S2.1, obtaining accurate positioning information; S2.2, training the RFR-GRU neural network model; when the GNSS system is in a failure state, the navigation system enters the prediction mode, which specifically includes the following steps: S2.3, predicting the positioning information. Beneficial effects: The present invention improves the robustness and positioning accuracy of the system, can accurately and effectively determine whether the GNSS system is in an effective state, and then determines whether the autonomous navigation system works in the training mode or the prediction mode, so as to maintain the navigation positioning accuracy.

Description

Technical Field

[0001] This invention patent relates to the technical field of navigation, and particularly relates to an autonomous navigation system and method based on fault detection and neural network assistance. Background Art

[0002] The integrated navigation technology of Inertial Navigation System (INS) / Global Navigation Satellite System (GNSS) fully combines the advantages of INS's autonomous and all-weather operation and GNSS's high-precision measurement. The navigation method based on INS / GNSS integration is a commonly used method for providing high-precision attitude, speed, position and other parameter measurements for mobile carriers such as vehicles. However, in the actual application process, 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 the attenuation or even loss of lock of the GNSS signal. During this satellite denial period, GNSS cannot provide observation information, and the integrated navigation system can only operate in pure INS mode, and the error accumulates over time and diverges without being suppressed, affecting the navigation accuracy.

[0003] Currently, the main methods to improve the performance of the INS / GNSS integrated navigation system under satellite denial conditions include integrated navigation methods based on tight integration and ultra-tight integration technologies, adding other sensor information to assist INS navigation methods, and artificial intelligence-assisted INS navigation methods based on neural networks. However, although the tight integration and ultra-tight integration technologies can maintain good system navigation performance under complex conditions such as multipath effect and external interference, their integration process is extremely complex, and when the satellite is completely unavailable, this method still cannot fundamentally solve this 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] Using artificial intelligence technology represented by neural networks to assist the integrated navigation method is relatively feasible and efficient. Various theoretical models and algorithms based on neural networks are introduced into the 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 that when the satellite signal is available, a neural network module is constructed for training; when the satellite is denied, the trained neural network module is used to predict the specific parameter information required during the GNSS denial period, so that the integrated navigation system can keep working normally.

[0005] However, existing methods based on neural network-assisted navigation often use only a single neural network structure for prediction, and a large number of training samples are required when training the neural network. In practical applications, the samples collected for neural network training are few and incomplete. In this case, using only a single-structure neural network for training is likely to result in problems such as overfitting of network training, local optimization, and instability of the prediction model, leading to unsatisfactory prediction results. In addition, when performing neural network-assisted navigation, it does not consider when to use neural network prediction intervention, that is, it does not consider how to determine when GNSS information is unavailable and then enable the neural network to enter 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 the robustness and positioning accuracy of the system.

[0007] Another object of the present invention is to provide an autonomous navigation method based on fault detection and neural network assistance, which ensures the 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;

[0009] 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;

[0010] The GNSS system is used to obtain measurement information of speed and position, and transmit the measurement information to the fault detection module;

[0011] The fault detection module is used to determine whether the GNSS system is in an effective state; when it is in an effective state, transmit the measurement information of the GNSS system to the RFR-GRU neural network module and the adaptive Kalman filter module;

[0012] The RFR-GRU neural network module is used to obtain pseudo-GNSS measurement information based on the position, speed, and attitude obtained by the INS system;

[0013] When it is in an invalid state, the RFR-GRU neural network module sends the pseudo-GNSS measurement information to the adaptive Kalman filter module;

[0014] 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, and further obtain accurate positioning information.

[0015] Further, the INS system includes an IMU device and an inertial navigation mechanical arrangement device;

[0016] The IMU device includes an accelerometer and a gyroscope, which are respectively used to measure acceleration and angular velocity, and send the measurement data to the inertial navigation mechanical arrangement device;

[0017] The inertial navigation mechanical arrangement device is used to calculate the position, velocity and attitude by using the mechanical arrangement algorithm according to the measurement data.

[0018] Further, the accelerometer bias and gyroscope bias can also be obtained through the adaptive Kalman filter module and sent to the IMU device for bias correction.

[0019] The second object of the present invention is implemented by the following technical solution: An autonomous navigation method based on fault detection and neural network assistance, which includes the following steps:

[0020] S1. Obtain information: Use the INS system to obtain the positioning information of speed, position and attitude; use the GNSS system to obtain the measurement information of speed and position;

[0021] S2. Judge the state of the GNSS system: Use the fault detection module to judge whether the GNSS system in step S1 is in an effective state;

[0022] When the GNSS system is in an effective state, the navigation system enters the training mode, which specifically includes the following steps:

[0023] S2.1. Obtain accurate positioning information: Establish a combined navigation adaptive Kalman filter model, and obtain the 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;

[0024] S2.2. Train the RFR-GRU neural network model: Use the positioning information and the measurement information in step S1 as the input features of the RFR-GRU neural network model to train the RFR-GRU neural network model; find out the relationship between the positioning information provided by the INS system and the measurement information provided by the GNSS system;

[0025] When the GNSS system is in a failure state, the navigation system enters the prediction mode, which specifically includes the following steps:

[0026] S2.3. Predictive 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 and the pseudo GNSS measurement information are 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, and then accurate positioning information is obtained.

[0027] Further, the algorithm of the adaptive Kalman filter model in step S2.1 includes a state equation and an observation equation, which are specifically described as follows:

[0028] (1)

[0029] where k is the discrete time, is the state vector of the system, which can be set as . 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 gyroscope drift and accelerometer bias of the INS system respectively, is the observation vector of the system. In this model, represents the difference between the positions and the difference between the velocities of GNSS and INS, that is, ; are the system structure parameters, which respectively represent the state one-step transfer matrix, the system noise distribution matrix, and the observation matrix; is the system noise vector, is the observation noise vector, and it satisfies , ; is the Kronecker function, and are respectively and 's covariance matrices;

[0030] State prediction:

[0031] (2)

[0032] Calculate the state one-step prediction covariance: Introduce an adaptive factor constructed based on the innovation sequence to suppress the influence of various complex noises on the filtering effect, and adjust the prior prediction covariance in real time;

[0033] (3)

[0034] (4)

[0035] Observation update: Construct the measurement innovation

[0036] (5)

[0037] Filter gain:

[0038] (6)

[0039] Corresponding state quantity update part:

[0040] (7)

[0041] State estimation covariance:

[0042] (8)

[0043] Where is the adaptive factor.

[0044] Furthermore, the method for the fault detection module in step S2 to judge whether the GNSS system is in an effective state specifically includes the following steps:

[0045] Use formula (5) to obtain the first n innovation sequences and calculate the mean of the first n innovation sequences:

[0046] (9)

[0047] (10)

[0048] Where is the judgment threshold, and its value is 1.

[0049] Furthermore, the adaptive factor The algorithm is as follows: Use the residual vector to construct the adaptive factor. The residual vector can reflect the error of the predicted state vector Construct the error discrimination statistic:

[0050] (11)

[0051] Where represents the trace of the innovation covariance matrix, and the innovation covariance matrix can be determined by the following formula

[0052] (12)

[0053] Thus, construct the adaptive factor :

[0054] (13)

[0055] Where c is a constant with a value of 1.

[0056] Furthermore, the RFR-GRU neural network model in step S2.2 is a weighted fusion model of a random forest regression model RFR and a gated recurrent unit GRU. Specifically, RFR is used to optimize GRU to improve the model accuracy of GRU during the training of GRU with small samples, and the RFR-GRU neural network outputs pseudo GNSS measurement information.

[0057] Furthermore, the specific modeling method of the RFR-GRU neural network model includes the following steps:

[0058] 1) Construct the original training sample set S: Use the positioning information and the measurement information set in step S1 as the original training sample S, and there are T samples in the original training sample S.

[0059] 2) Construct sub-training sets: Randomly sample from the original training sample set S in step 1) through Bootstrap Sampling, perform N samplings in total, construct N sub-training sets, and each sub-training set has T samples;

[0060] 3) Train the GRU model: Train a GRU model for each sub-training set in step 2); Each GRU model gives a predicted value, and the predicted values of the N GRU models are arithmetically averaged to obtain the final regression result output;

[0061] 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 result.

[0062] Furthermore, the GRU model specifically includes an update gate ( ) and a reset gate ( ):

[0063] Among them, the update gate ( ) is used to capture the long-term dependence relationship in the time series:

[0064] (14)

[0065] The reset gate ( ) is used to capture the short-term dependence relationship in the time series:

[0066] (15)

[0067] Among them, is the state output at the previous moment t-1, is the input data at the current moment t, and both gates are determined by the input data at the current moment t and the output state at the previous moment t-1 ; the sigmoid function converts the data into values within the range of 0-1;

[0068] Integrated update gate and reset gate output:

[0069] (16)

[0070] where, , and are the weight matrices from the output state at the previous moment t-1 to the update gate, reset gate, and the output state at the current moment t respectively; , and are the weight matrices from the input data to the update gate, reset gate, and the output state respectively; , and are the offset vectors of the corresponding structures, and the function is a non-linear activation function.

[0071] Advantages of the present invention:

[0072] 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 relationship between the positioning information based on INS and the measurement information of GNSS is found using this neural network module; when the satellite loses lock and is in a failure state, the pseudo-GNSS measurement information output by this neural network module is used to replace the original GNSS measurement information for integrated navigation, improving the robustness and positioning accuracy of the system.

[0073] 2. The present invention provides an autonomous navigation method based on fault detection and neural network assistance, in which the first n innovation sequences are obtained by constructing the innovation of the measurement, the mean value of the first n innovation sequences is calculated, and it is judged whether the mean value is ≤1. If ≤1, it is determined that the GNSS system is in an effective state; if >1, it is determined that the GNSS system is in a failure state, which can accurately and effectively determine whether the GNSS system is in an effective state, and then determine whether the autonomous navigation system operates in the training mode or the prediction mode.

[0074] 3. The present invention provides an autonomous navigation method based on fault detection and neural network assistance, in which an adaptive factor constructed based on the innovation sequence is introduced into the algorithm of the adaptive Kalman filter model to suppress the influence of various complex noises on the filtering effect, and the prior prediction covariance is adjusted in real time to suppress the influence of the change of noise characteristics on the integrated filtering accuracy.

[0075] 4. The present invention provides an autonomous navigation method based on fault detection and neural network assistance. The GRU is optimized by a random forest regression model, enabling it to achieve good prediction accuracy even under small-sample training conditions. Moreover, the GRU can fully consider and utilize historical data, improving the prediction accuracy of the neural network model. It has the advantages of a simple structure and few parameters, and can effectively predict pseudo-GNSS measurement information in the event of GNSS failure, thereby maintaining the navigation and positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0077] Figure 1 An autonomous navigation system based on fault detection and neural network assistance in Training Mode for Embodiment 2 of the present invention.

[0078] Figure 2 An autonomous navigation system based on fault detection and neural network assistance in Prediction Mode for Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0079] The present invention will be further described in detail through embodiments below.

[0080] 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;

[0081] 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 judge whether the GNSS system is in an effective state; when in an effective state, transmit the measurement information of the GNSS system 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, and further obtain accurate positioning information.

[0082] The INS system includes an IMU device and an inertial navigation mechanical arrangement device;

[0083] The IMU device includes an accelerometer and a gyroscope, which are respectively used to measure acceleration and angular velocity, 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 a mechanical arrangement algorithm.

[0084] The adaptive Kalman filter module can also obtain the accelerometer bias and the gyroscope bias, and send them to the IMU device for bias correction.

[0085] In the present invention, by constructing an RFR-GRU neural network module, when the satellite signal is intact, the relationship between the positioning information based on INS and the measurement information of GNSS is found by using this neural network module; when the satellite is out of lock and in an invalid state, the pseudo-GNSS measurement information output by this neural network module is used to replace the original GNSS measurement information for integrated navigation, improving the robustness and positioning accuracy of the system.

[0086] Embodiment 2: An autonomous navigation method based on fault detection and neural network assistance, which includes the following steps:

[0087] 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.

[0088] S2. Judge the state of the GNSS system: Use the fault detection module to judge whether the GNSS system in step S1 is in an effective state.

[0089] The method for the fault detection module to determine whether the GNSS system is in an effective state specifically includes the following steps:

[0090] Use formula (5) to obtain the first n innovation sequences, and calculate the mean value of the first n innovation sequences:

[0091] (9)

[0092] (10)

[0093] Where is the judgment threshold, and its value is 1.

[0094] As Figure 1 shown, when the GNSS system is in an effective state, the navigation system enters the training mode, which specifically includes the following steps:

[0095] S2.1. Obtain accurate positioning information: Establish a combined navigation adaptive Kalman filter model, and obtain the 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.

[0096] Where the algorithm of the adaptive Kalman filter model includes a state equation and an observation equation, and is specifically described as follows:

[0097] (1)

[0098] Where k is the discrete time, is the state vector of the system, which can be set as . 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 respectively the gyroscope drift and accelerometer bias of the INS system, is the observation vector of the system. In this model represents the difference between the positions and the difference between the velocities of GNSS and INS, that is ; are the system structure parameters, which respectively represent the state one-step transfer matrix, the system noise distribution matrix, and the observation matrix; is the system noise vector, is the observation noise vector, and satisfies , ; is the Kronecker function, and respectively and covariance matrices

[0099] State prediction:

[0100] (2)

[0101] Calculate the one-step state prediction covariance: Introduce an adaptive factor constructed based on the innovation sequence to suppress the influence of various complex noises on the filtering effect, and adjust the prior prediction covariance in real time;

[0102] (3)

[0103] (4)

[0104] Observation update: Construct the measurement innovation

[0105] (5)

[0106] Filter gain:

[0107] (6)

[0108] The corresponding state quantity update part:

[0109] (7)

[0110] State estimation covariance:

[0111] (8)

[0112] where is the adaptive factor; The adaptive factor algorithm is as follows: Use the residual vector to construct the adaptive factor. The residual vector can reflect the error of the predicted state vector . Construct the error discrimination statistic:

[0113] (11)

[0114] where represents the trace of the innovation covariance matrix, and the innovation covariance matrix can be determined by the following formula

[0115] (12)

[0116] Thus, construct the adaptive factor :

[0117] (13)

[0118] Where c is a constant with a value of 1.

[0119] S2.2. Train the RFR-GRU neural network model: By using the positioning information and the measurement information in step S1 as the input features of the RFR-GRU neural network model, train the RFR-GRU neural network model; find the relationship between the positioning information provided by the INS system and the measurement information provided by the GNSS system.

[0120] 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, use RFR to optimize GRU to improve the model accuracy of GRU during the training of GRU with small samples, and use the RFR-GRU neural network to output pseudo GNSS measurement information.

[0121] The specific modeling method of the RFR-GRU neural network model includes the following steps:

[0122] 1) Construct the original training sample set S: Use the positioning information and the measurement information set in step S1 as the original training sample S, and there are T samples in the original training sample S.

[0123] 2) Construct sub-training sets: Randomly sample from the original training sample set S in step 1) through Bootstrap Sampling, perform N samplings in total, construct N sub-training sets, and each sub-training set has T samples;

[0124] 3) Train the GRU model: Train a GRU model for each sub-training set in step 2); each GRU model gives a predicted value, and perform arithmetic averaging on the predicted values of the N GRU models to obtain the final regression result output;

[0125] 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 result.

[0126] The GRU model specifically includes an update gate ( ) and a reset gate ( ):

[0127] Among them, the update gate ( ) is used to capture the long-term dependence relationship in the time series:

[0128] (14)

[0129] The reset gate ( )Used to capture short-term dependencies in time series:

[0130] (15)

[0131] Among them, is the state output at the previous moment t - 1, is the input data at the current moment t. Both gates are determined by the input data at the current moment t and the output state at the previous moment t - 1; the sigmoid function converts the data into values within the range of 0 - 1;

[0132] Integrate the update gate and reset gate outputs:

[0133] (16)

[0134] Among them, , and are the weight matrices from the output state at the previous moment t - 1 to the update gate, reset gate, and the output state at the current moment t respectively; , and are the weight matrices from the input data to the update gate, reset gate, and the output state respectively; , and are the offset vectors of the corresponding structures respectively, and the function is a non-linear activation function.

[0135] As Figure 2 shown, when the GNSS system is in a failure state, the navigation system enters the prediction mode, which specifically includes the following steps:

[0136] S2.3. Predict the positioning information: By inputting the positioning information in step S1 into the RFR-GRU neural network model in step S2.2, obtaining pseudo-GNSS measurement information, and transmitting the positioning information in step S1 and the pseudo-GNSS measurement information 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, and then obtaining accurate positioning information.

[0137] Experimental verification:

[0138] The autonomous navigation method disclosed in Embodiment 2 was compared with various different navigation methods. An in-vehicle navigation test was carried out using a prototype INS / GNSS integrated navigation system, and INS / GNSS data was collected during the operation of the in-vehicle test platform to evaluate the proposed method. In this test, a GNSS signal interruption test on the road was designed. The specific scheme was as follows: when the INS / GNSS integrated navigation system was working normally, the adaptive Kalman filter integrated navigation algorithm under available satellite signals was executed, and the in-vehicle platform continued to run for 400 s. In this case, the samples collected for training the neural network were less. Subsequently, the GNSS signal was interrupted, and the in-vehicle platform continued to run for 200 s under satellite denial conditions. That is, the total running time of the in-vehicle test was 600 s, and 200 s of it was under GNSS denial conditions. In this verification, the first 400 s were used as the neural network training data, and then the GNSS was interrupted for 200 s. The neural network model started to predict and assist the adaptive Kalman filter for integrated navigation. At the same time, in this verification, pure inertial navigation, other LSTM and GRU networks, and the RFR-GRU neural network assisted navigation method without a fault detection module were selected for comparison, and the results are shown in Table 1.

[0139] Table 1 Comparison of test results

[0140]

[0141] 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 detection module, the autonomous navigation method disclosed in Embodiment 2 of the present invention more effectively improved the problem of divergence of the speed and position errors of the navigation system during satellite denial, and improved the robustness and positioning accuracy of the system.

[0142] The above is the preferred implementation manner of the present invention. For those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can still be made, and these improvements and refinements should also be regarded as the protection scope 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 further 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 an effective state; when in the effective state, it transmits the measurement information of the GNSS system 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, speed, and attitude obtained 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. RFR is used to optimize GRU to improve the model accuracy of GRU during the training of small samples for GRU, and the RFR-GRU neural network outputs pseudo-GNSS measurement information; When 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, and further obtain accurate positioning information; Wherein the autonomous navigation method of the autonomous navigation system 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. Judge the state of the GNSS system: Use the fault detection module to judge whether the GNSS system in step S1 is in an effective state; When the GNSS system is in an effective state, the navigation system enters the training mode, which specifically includes the following steps: S2.

1. Obtain accurate positioning information: Establish a combined navigation adaptive Kalman filter model, and 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 further obtain accurate positioning information; S2.

2. Train the RFR-GRU neural network model: Use the positioning information and the measurement information in step S1 as the input features of the RFR-GRU neural network model to train the RFR-GRU neural network model; find 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 failure state, the navigation system enters the prediction mode, which specifically includes the following steps: S2.

3. Predict positioning information: Input the positioning information in step S1 into the RFR-GRU neural network model in step S2.2 to obtain pseudo-GNSS measurement information, and transmit it to the adaptive Kalman filter model in step S2.1 together with the positioning information in step S1 for data fusion to obtain a compensation value of the positioning information of the INS system, and further obtain accurate positioning information; Among them, the algorithm of the adaptive Kalman filter model in step S2.1 includes a state equation and an observation equation, and the specific description is as follows: （1） where \(k\) is discrete time, is the state vector of the system, which can be set as ; 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 gyroscope drift and accelerometer bias of the INS system respectively, is the observation vector of the system. In this model represents the difference between the positions and the difference between the velocities of GNSS and INS, that is ; are the system structure parameters, which represent the state one-step transition matrix, the system noise distribution matrix, and the observation matrix respectively; is the system noise vector, is the observation noise vector, and satisfies , ; is the Kronecker function, and are respectively and 's covariance matrices;​​ State prediction: （2） Calculate the one-step prediction covariance of the state: （3） （4） Observation update: Construct the measurement innovation （5） Filter gain: （6） Corresponding state quantity update part: （7） State estimation covariance: （8） wherein is an adaptive factor; Among them, the method for the fault detection module in step S2 to judge whether the GNSS system is in an effective state specifically includes the following steps: Use formula (5) to obtain the first n innovation sequences, and calculate the mean of the first n innovation sequences: （9） （10） Among them, is the judgment threshold, and its value is 1.

2. The autonomous navigation system based on fault detection and neural network assistance according to claim 1, wherein 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 respectively used to measure acceleration and angular velocity, and send the measurement data to the inertial navigation mechanical arrangement device; The inertial navigation mechanical arrangement device is used to calculate the position, velocity and attitude by using the mechanical arrangement algorithm according to the measurement data.

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

4. An autonomous navigation system based on fault detection and neural network assistance according to claim 1, characterized in that, Adaptive factor The algorithm is as follows: Using the residual vector to construct the adaptive factor, and the residual vector can reflect the error of the predicted state vector Construct an error discrimination statistic: （11） where denotes the trace of the innovation covariance matrix, which can be determined by the following equation （12） Construct the adaptive factor accordingly : （13） Where c is a constant, and its value is 1.

5. An autonomous navigation system based on fault detection and neural network assistance according to claim 1, characterized in that, The specific modeling method of the RFR-GRU neural network model in step S2.2 includes the following steps: 1) Construct the original training sample set S: Use the positioning information and the measurement information set in step S1 as the original training sample S, and there are T samples in the original training sample set S; 2) Construct sub-training sets: Randomly sample from the original training sample set S in step 1) through Bootstrap Sampling. A total of N samplings are performed to construct N sub-training sets, and each sub-training set has T samples; 3) Train 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 the N GRU models are arithmetically 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 result.

6. The autonomous navigation system based on fault detection and neural network assistance according to claim 5, characterized in that, The specific GRU model includes an update gate ( ) and a reset gate ( ) as follows: Among them, the update gate ( ) is used to capture the long-term dependencies in the time series: （14） Reset Gate( ) is used to capture short-term dependencies in time series: （15） Among them, is the state output at the previous moment t-1, is the input data at the current moment t. Both gates are determined by the input data at the current moment t and the output state at the previous moment t-1; the sigmoid function converts the data into values within the range of 0-1; Comprehensive update gate and reset gate output: （16） Among them, , and are the weight matrices of the output state at the previous moment t - 1 to the update gate, the reset gate, and the output state at the current moment t, respectively; , and are the weight matrices of the input data to the update gate, the reset gate, and the output state, respectively; , and are the offset vectors of the corresponding structures, and the function is a non - linear activation function.

Images