A method for using GRU neural network to assist navigation during GNSS signal interruption

Abstract

The present invention relates to an algorithm for using a GRU neural network to assist navigation during a GNSS signal interruption. When the GNSS signal is available, the speed, heading and output of the IMU of the INS will be used as input features for training the GRU model, and the position increment of the GNSS is the output of the model. Once the GNSS signal is lost, the information of the INS will be fed into the GRU model to generate a pseudo-GNSS increment. After accumulating the pseudo-GNSS increment, the pseudo-GNSS position will be sent to the Kalman filter to correct the INS navigation solution result. In addition, considering that it is difficult to obtain a large amount of GNSS observation data with good observation in complex scenarios, the number of observable satellites in some areas may be 1 to 3. At this time, the use of a tightly coupled model can still perform combined navigation, which guarantees a more effective positioning result to a certain extent and can provide better training data for the neural network. Therefore, the present invention selects a GRU-assisted tightly coupled mode.

Description

Technical Field

[0001] The present invention relates to an artificial intelligence (AI) algorithm based on a gated recurrent unit (GRU) for improving the navigation accuracy of an inertial navigation system (INS) in a GNSS / INS tightly coupled integrated navigation system by estimating a pseudo GNSS position during a global navigation satellite system (GNSS) signal interruption. Background Art

[0002] At present, there are two main common navigation and positioning methods, namely the Strapdown Inertial Navigation System (SINS) and the Global Navigation Satellite System (GNSS). The Strapdown Inertial Navigation System can obtain the navigation information of the carrier through the data of its own inertial components such as accelerometers and gyroscopes, and can maintain high-precision navigation in a short time. With the development of micro-electromechanical system (MEMS) technology, the size of inertial components has gradually become smaller and the cost has become lower, which has also made the corresponding inertial navigation system more and more widely used. It is used in many products such as mobile phones, children's watches, and drones. However, due to the large noise and instability of MEMS sensors, it is difficult to remove the noise, so the calculation of a single MEMS inertial navigation system will diverge rapidly over time. GNSS has the characteristics of high accuracy, wide coverage, easy use, and all-weather. When the observation conditions are good, GNSS can meet the high-precision positioning needs of the carrier. However, once the signal is blocked and the carrier cannot receive the GNSS signal, it will result in the inability to navigate and position by relying on a single GNSS positioning module. Therefore, combining SINS and GNSS can build a low-cost, small-volume, high-precision, and easy-to-implement combined navigation system, which is also the most commonly used positioning technology in the industry.

[0003] However, in complex urban scenes, GNSS signals are easily blocked by objects such as buildings, trees, and tunnels. In this case, the traditional Kalman filter lacks GNSS observation correction, and the integrated navigation system will be simplified to pure inertial navigation solution. At this time, due to the limited accuracy of the output data of the MEMS sensor, the noise is very large, and the navigation positioning accuracy diverges rapidly, it is necessary to introduce new sensors for combined correction, such as odometers, lidar, etc., but this will increase the cost of additional devices. Therefore, many scholars have also studied the use of neural networks to assist combined navigation and suppress the divergence of SINS results. Specifically, when the GNSS signal is good, a suitable navigation data training model is selected; after the GNSS signal is lost, the trained neural network is used to output the pseudo-observation of GNSS, which is continued to be used for combined filtering to suppress the divergence of SINS errors. Therefore, in order to train a good model, a large amount of navigation data needs to be provided for training when the GNSS signal is good. However, in complex scenarios, it is difficult to obtain a large amount of GNSS observation data with good observations. The number of observable satellites in some areas may be 1 to 3. At this time, the use of a tightly coupled model can still perform combined navigation, which guarantees more effective positioning results to a certain extent and can provide better training data for the neural network. Summary of the invention

[0004] In order to solve the problems of the prior art and suppress the error divergence of INS after GNSS interruption, the present invention provides a method using GRU recurrent neural network to assist tight coupling.

[0005] A method for using a GRU neural network to assist navigation during GNSS signal interruption, characterized in that it includes:

[0006] Step 1. When the Global Navigation Satellite System (GNSS) signal is available, the AI ​​module will work in training mode. The speed, heading and IMU output of the INS will be used as input features for training the GRU neural network. The input feature parameters are selected as the gyroscope output. , the accelerometer output , INS speed , INS heading angle , output selection is GNSS position increment , the GNSS position increment is the output of the model.

[0007] Step 2: When GNSS signal is lost during GNSS interruption, the AI ​​module will work in prediction mode, and the information of INS will be fed into the GRU neural network to generate pseudo GNSS increments. After accumulating pseudo GNSS increments, the pseudo GNSS position will be sent to KF to correct the INS navigation solution.

[0008] In the above method of using a GRU neural network to assist navigation during a GNSS signal interruption, the GRU neural network includes two gate structures, an update gate and a reset gate. The update gate can change the degree to which the current moment is affected by the state of the previous moment, and the output is The value of the update gate is proportional to the amount of information retained. The smaller the update gate output, the less state information is retained. The reset gate is used to discard some hidden state information. The output is ,The value of the reset gate is inversely proportional to the amount of discarded information. The smaller the value is, the more past state information is discarded.

[0009] In the above method of using a GRU neural network to assist navigation during a GNSS signal interruption, when the GNSS signal is available, the AI ​​module will work in training mode to train the GRU neural network, and the input is the three-axis accelerometer data output by the IMU. , three-axis gyroscope data , the speed information obtained by INS and heading angle information . Therefore, when the GNSS signal is good, the GRU neural network finds the mapping relationship between the IMU, INS information and the GNSS increment information, so that the GRU neural network obtains functions similar to those of the GNSS module. The gate structure described in power 2 discards some of the useless state information in the past, avoids the problems of gradient explosion and gradient disappearance, and retains the advantage of continuous trajectory prediction. The combined navigation model during training is a tightly coupled model, and combined navigation can continue when the number of GNSS satellites is 1-3.

[0010] In the above method of using a GRU neural network to assist navigation during a GNSS signal interruption, when the GNSS is interrupted, the system switches to the prediction mode, and the GNSS module can no longer provide position information, so there is no GNSS module in the prediction model. At this time, the IMU and INS modules continue to input navigation information to the GRU neural network, and the trained GRU neural network predicts the GNSS signal based on the input.

[0011] Increment The incremental information is integrated to obtain a predicted GNSS position information. The predicted GNSS position information is pseudo GNSS position information, which is used to simulate the positioning result when the GNSS signal is good. The result is compared with the position solved by INS. After the difference is made, it can be sent to the KF module for Kalman filtering. At this time, the errors of position, velocity and attitude can continue to be output as when the GNSS signal is good, which is used to feedback and correct the INS results.

[0012] Beneficial effects: The present invention proposes a method of using GRU recurrent neural network to assist tight coupling, which is used to reduce the cumulative INS error after the GNSS signal is interrupted, and improves the navigation accuracy during the GNSS signal interruption. Compared with the traditional static neural network MLP algorithm, the GRU algorithm has excellent performance when processing continuous processes, and can predict navigation results more accurately and stably. Compared with the LSTM algorithm, which is also a recurrent neural network, the GRU has similar performance and shorter training time. Therefore, during the GNSS signal interruption, the navigation solution developed using the GRU algorithm is more accurate and reliable. In addition, in complex scenarios, there may be many scenarios with 1 to 3 satellite observations. Loose coupling cannot obtain GNSS positioning information, while tight coupling can continue to perform combined navigation, and better training set data can be obtained. The effects of GRU-assisted tight coupling and GRU-assisted loose coupling are compared. The results show that the combined navigation results using tight coupling in the training set time period are more accurate, the trained network is better, the test set prediction error is smaller, and the actual assisted navigation results are better. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The above and other features of the present invention and methods for achieving these features will become more apparent with reference to the detailed description of the present invention in conjunction with the accompanying drawings. The present invention itself will also become easier to understand, in which:

[0014] Figure 1 is a GRU-assisted integrated navigation training model under tight coupling.

[0015] Figure 2 is a GRU-assisted integrated navigation prediction model under tight coupling.

[0016] Figure 3 is a navigation device.

[0017] Figure 4 shows the position errors of different algorithms when the GNSS is interrupted for 200 seconds.

[0018] Figure 5 is a comparison of map trajectories of different algorithms when the GNSS is interrupted for 200 seconds.

[0019] Figure 6 is a comparison of the straight line effects of GRU-assisted tight coupling and GRU-assisted loose coupling.

[0020] Figure 7 is a straight line trajectory diagram of the GRU-assisted tight coupling and GRU-assisted loose coupling effects.

[0021] Figure 8 is a comparison of the GRU-assisted tight coupling and GRU-assisted loose coupling effects on curved road sections.

[0022] Figure 9 is a trajectory diagram of a curved road section with GRU-assisted tight coupling and GRU-assisted loose coupling effects.

[0023] Figure 10 is a comparison of the training effects of GRU and LSTM. DETAILED DESCRIPTION

[0024] First, the principle of the method of the present invention is introduced.

[0025] When GNSS signals are available, the velocity, heading, and output of the IMU from the INS are used as input features for training the GRU neural network, and the GNSS position increment is the output of the model.

[0026] Once the GNSS signal is lost, the information of the INS will be fed into the GRU neural network to generate pseudo-GNSS increments. After accumulating the pseudo-GNSS increments, the pseudo-GNSS position will be sent to the KF to correct the INS navigation solution. And compared with the GRU-assisted loosely coupled mode, the GRU-assisted tightly coupled mode can continue combined navigation when the number of satellites is 1 to 3, which improves the accuracy of the training data and the effect of network training.

[0027] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is: using a recurrent neural network based on GRU units as an additional model to suppress the error divergence of INS after GNSS interruption, and the combined navigation model in the training model is selected as tightly coupled. The GRU recurrent neural network can train neural networks in both the time dimension and the network depth, and can better capture the temporal dynamic performance of input data compared to static neural networks. Compared with LSTM, which is also a recurrent neural network, the GRU network has similar training performance and faster training speed.

[0028] At present, there are three main models for using neural network-assisted integrated navigation when GNSS is lost: Model, Model and Models can play a role in suppressing the divergence of inertial navigation accuracy. The model selects INS information as input and the position errors of GNSS and INS as output; The model selects INS information as input and the state vector of Kalman filter as output; The model selects INS information as input and GNSS position increment information as output. The dimension of the output vector of the second model is larger than that of the other two models, the prediction effect is also poor, and it may be mixed with filtering errors, so it is generally less used. The output vector of the first model contains both INS information and GNSS information, while the output vector of the third model is only related to GNSS position information, which can reduce additional mixing errors. Therefore, the present invention adopts Model. It can be directly derived from some INS information, as shown in Formula 1.

[0029]

[0030] In formula (1) is the direction cosine matrix, is the specific force, i.e. the accelerometer output, is the angular velocity of the e system relative to the i system, is the angular velocity of the n system relative to the e system, For speed, is the gravity vector.

[0031] Further analysis of the gravity vector Mainly related to longitude and latitude. and The expression is shown in formula (2). , the meridian radius of the Earth and the radius of the circle In the vehicle environment, the changes are slow, so the two are mainly influence, therefore influence The main factors are , , .and Determined by the attitude angle, as shown in formula (3), the ideal attitude matrix differential equation is as shown in formula (4).

[0032]

[0033]

[0034] (4)

[0035] therefore Attitude Angle (roll angle), (pitch angle), (Heading angle), (gyroscope output) influence, but in a vehicle environment, , It is generally close to 0 and changes very little, so the input feature parameters of the final training model in this paper are selected as , , , , the output selection is .

[0036] The input of the GRU neural network training model is the three-axis gyroscope data of the IMU, the three-axis accelerometer data, the three-axis speed and heading angle information solved by the INS, with a total dimension of 10 dimensions, and the output is the GNSS increment information. During training, due to the different frequencies of the input IMU sensor data, speed, heading angle and other INS information and GNSS increment information, the IMU data and INS information frequency is 200 times per second, while the GNSS increment information can only be output once per second, so during training, it is necessary to integrate the high-frequency IMU and INS information, and use the IMU and INS information within 1 second as the input layer input to correspond to the low-frequency GNSS increment. In addition, the GRU neural network has a time dimension, which is reflected in the data as the existence of the time step parameter. Combined with the above-mentioned training data features of 10 dimensions, the data length of the training set is 3000 seconds, and the sampling frequency is 200HZ, it can be seen that the data size of the input layer is (3000, time step, 2000); at the same time, the data at the output end is the corresponding GNSS increment information, with a size of (3000, 3). The amount of training data is huge, so it is necessary to set the appropriate number of hidden layer neurons and time step parameters, otherwise it will affect the speed and accuracy of training. In order to speed up the convergence speed during training and prevent overfitting, the training data is normalized before training, and the prediction results must be denormalized when using the prediction model for prediction. Experiments were conducted on the adjustment of several related parameters, so the final time step parameter was selected as 4 and the number of hidden layer units was selected as 128.

[0037] The working modes of the AI ​​module include training mode and prediction mode. When the GNSS signal is good, the GRU neural network is trained and the input is the three-axis accelerometer data output by the IMU. , three-axis gyroscope data , the speed information obtained by INS and heading angle information Therefore, when the GNSS signal is good, the GRU neural network can find the mapping relationship between the navigation information such as IMU and INS and the GNSS increment information. When the GNSS signal is lost, the GNSS module can no longer provide position information, so there is no GNSS module in the prediction model. At this time, the IMU and INS modules continue to input navigation information to the GRU neural network, and the trained GRU neural network predicts the GNSS increment based on the input. By integrating the incremental information, we can get a predicted GNSS position information, also called pseudo GNSS position information, which is used to simulate the positioning result when the GNSS signal is good. After the difference is made, it is sent to the KF module for Kalman filtering, and the errors of position, velocity and attitude are continuously outputted to correct the INS results. In this way, when the GNSS is lost, the integrated navigation system that could only use pure INS solution can be corrected with pseudo-GNSS information, thus suppressing the error divergence rate of the inertial navigation.

[0038] The present case will be described below with reference to the accompanying drawings.

[0039] Attached Figure 1 It is a GRU-assisted integrated navigation training model under tight coupling. When the GNSS signal is good, the GRU neural network is trained and the input is the three-axis accelerometer data output by the IMU. , three-axis gyroscope data , the speed information obtained by INS and heading angle information Therefore, when the GNSS signal is good, the GRU neural network can find the mapping relationship between the navigation information such as IMU and INS and the GNSS increment information. However, since the GNSS provides the original observation under tight coupling and does not directly give the positioning result, the position increment of the combined result is directly input into the GRU neural network during training.

[0040] Attached Figure 2 It is a tightly coupled GRU-assisted integrated navigation prediction model. When GNSS is interrupted, the system switches to prediction mode. In this case, GNSS no longer outputs position information, but , , These values ​​can also be obtained from IMU and INS, and combined with the network trained in training mode, the predicted position increment at each moment can be obtained. The position increment at each moment After accumulation, the GNSS pseudo-measurement value at each moment can be obtained. This pseudo-position observation obtained by the neural network will replace the original GNSS signal as a new observation and continue to participate in the Kalman filtering process to suppress the rapid divergence of the error of pure inertial navigation.

[0041] Attached Figure 3 It is a navigation device. The data acquisition device is M39, the IMU sampling frequency is 200HZ, and the reference device is SPAN-CPT. The comparison of the IMU performance parameters of the two is shown in Table 1.

[0042] Table 1 Comparison of IMU performance parameters between M39 and SPAN-CPT

[0043]

[0044] Attached Figure 4are the errors of different algorithms when GNSS is interrupted for 200 seconds. The maximum errors of latitude, longitude and horizontal direction of pure inertial navigation solution are 394.51m, 294.66m and 492.40m respectively; the maximum errors of latitude, longitude and horizontal direction of MLP-assisted combined navigation are 370.23m, 166.78m and 406.06m respectively; the maximum errors of latitude, longitude and horizontal direction of GRU-assisted combined navigation are 86.88m, 69.07m and 87.14m respectively. The results of GRU-assisted combined navigation show the expected effect. Compared with pure inertial navigation solution, the position accuracy in all directions has been greatly improved, and the maximum position errors in latitude, longitude and horizontal directions have decreased by 78.0%, 76.6% and 82.3% respectively. At the same time, compared with the results of MLP-assisted combined navigation, the results are also consistent. Figure 5-6 The same conclusion is reached, the maximum position errors in latitude, longitude and horizontal directions are reduced by 76.5%, 58.6% and 78.5% respectively.

[0045] Attached Figure 5 This is a comparison of map trajectories of different algorithms when GNSS is interrupted for 200 seconds. Different colored icons are used to represent the navigation trajectories of different algorithms. The purple icon represents the real trajectory, the red icon represents the trajectory of GRU-assisted integrated navigation, the green icon represents the trajectory of MLP-assisted integrated navigation, and the blue icon represents the trajectory of pure inertial navigation. It can be seen from the figure that the road section used for the experiment is complex enough, including a long straight road and three turns. The difference between the three algorithms in the initial straight section is not obvious, and they all maintain a lane-level accuracy. However, after the second bend, the trajectories of the other three methods that lose GNSS signals have obvious differences from the real trajectory, among which the trajectory of GRU-assisted integrated navigation is the closest, the trajectory of MLP-assisted integrated navigation is suboptimal, and the trajectory of pure inertial navigation is the worst. After the third bend, the errors of the other two algorithms diverge seriously, and the trajectories are far from the real trajectory. Only the trajectory of GRU-assisted integrated navigation is still close to the real trajectory trend.

[0046] Attached Figure 6The figure shows the comparison of the effects of GRU-assisted tight coupling and GRU-assisted loose coupling on a straight section. The maximum errors of the three algorithms in the latitude direction are 296.33m, 24.44m and 8.77m respectively. The maximum errors in the longitude direction are 68.84m, 9.39m and 8.95m respectively. The maximum errors in the horizontal direction are 304.22m, 25.46m and 12.26m respectively. The results of GRU-assisted tight coupling and loose coupling are significantly improved compared with the pure inertial navigation solution. Comparing the two GRU-assisted combined navigation results, the performance of the two is similar in the longitude direction, while the GRU-assisted tight coupling navigation error is smaller in the latitude direction. The GNSS signal of this straight section was lost for 100 seconds. Compared with the pure inertial navigation solution, the maximum horizontal error of GRU-assisted tight coupling decreased by 95.97%; at the same time, compared with GRU-assisted loose coupling, the maximum horizontal error of GRU-assisted tight coupling also decreased by 51.8%.

[0047] Attached Figure 7 The following are the trajectory diagrams of the straight-line sections with GRU-assisted tight coupling and GRU-assisted loose coupling effects. The trajectory calculated by pure inertial navigation is far from the real straight-line trajectory. Since the time correlation of the straight-line section trajectory is stronger and the data of the straight-line sections in the training set is relatively rich, the two GRU-assisted combined navigation models work well on the straight-line sections. Further comparison shows that the GRU-assisted tight coupling trajectory is more consistent with the real trajectory.

[0048] Attached Figure 8 The figure shows the comparison of the GRU-assisted tight coupling and GRU-assisted loose coupling effects on curved road sections. The maximum errors of the three algorithms in the latitude direction are 361.60m, 45.61m and 39.74m respectively. The maximum errors in the longitude direction are 285.06m, 118.72m and 20.51m respectively. The maximum errors in the horizontal direction are 460.45m, 124.39m and 44.33m respectively. Similar to the speed error, the two GRU-assisted combined navigation models have similar performance in the latitude direction, while the GRU-assisted tight coupling has a significant performance improvement in the longitude direction. Reflected in the horizontal position error, compared with the pure inertial navigation solution, the maximum error of the GRU-assisted tight coupling decreased by 90.3%; at the same time, compared with the GRU-assisted loose coupling, the maximum horizontal error of the GRU-assisted tight coupling also decreased by 64.4%.

[0049] Attached Fig. 9The following is a trajectory diagram of the curved road section with GRU-assisted tight coupling and GRU-assisted loose coupling. The real trajectory contains two straight sections and a turn, and the trajectories of the three algorithms diverge to varying degrees. Among them, the trajectory of pure inertial navigation solution diverges the most seriously, and there is no obvious turn at all. The huge drift of speed makes the position of pure inertial navigation solution far different from the actual position; the trajectory of GRU-assisted loose coupling can see the turn, and the heading offset is not as serious as that of pure inertial navigation solution; the trajectory of GRU-assisted tight coupling is closest to the real trajectory, the corners of the turn are also the clearest, and the heading after the turn is more in line with the real trajectory.

[0050] Fig.10 The figure shows the training effect comparison between GRU and LSTM. The LSTM training result has a latitude prediction error MSE of 3.13e-6°, a longitude prediction error MSE of 3.22e-6°, and an epoch time of 9.11 seconds. The GRU training result has a latitude prediction error MSE of 3.12e-6°, a longitude prediction error MSE of 3.17e-6°, and an epoch time of 7.17 seconds, which proves that the training performance of GRU and LSTM is similar, but the training time is shorter.

[0051] Finally, it should be noted that the above specific implementation methods are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A method for using a GRU neural network to assist navigation during GNSS signal interruption, characterized in that: include Step 1. When the GNSS signal is available, the AI ​​module will work in training mode. The speed, heading and IMU output of INS will be used as input features for training the GRU neural network. The input feature parameters are selected as gyroscope output. , the accelerometer output , INS speed , INS heading angle , output selection is GNSS position increment , the position increment of GNSS is the output of the model; Step 2: When GNSS signal is lost due to GNSS interruption, the AI ​​module will work in prediction mode, and the information of INS will be fed into the GRU neural network to generate pseudo GNSS increments; After accumulating pseudo-GNSS increments, the pseudo-GNSS position will be sent to the KF to correct the INS navigation solution; When GNSS signals are available, the velocity, heading, and output of the IMU from the INS are used as input features for training the GRU neural network, and the position increment from the GNSS is the output of the model; Once the GNSS signal is lost, the information from the INS will be fed into the GRU neural network to generate pseudo GNSS increments; After accumulating pseudo-GNSS increments, the pseudo-GNSS position will be sent to KF to correct the INS navigation solution results. Compared with the GRU-assisted loose coupling mode, the GRU-assisted tight coupling mode continues combined navigation when the number of satellites is 1 to 3, which improves the accuracy of training data and the effect of network training. A recurrent neural network based on GRU units is used as an additional model to suppress the error divergence of INS after GNSS interruption. The integrated navigation model in the training model is selected as tightly coupled. The GRU recurrent neural network trains neural networks in both the time dimension and the network depth, and can better capture the temporal dynamic performance of input data compared with static neural networks. Compared with LSTM, which is also a recurrent neural network, the GRU network has similar training performance and faster training speed. There are currently three models that use neural network-assisted integrated navigation when GNSS is lost: Model, Model and Models all play a role in suppressing the divergence of inertial navigation accuracy; The model selects INS information as input and the position errors of GNSS and INS as output; The model selects INS information as input and the state vector of Kalman filter as output; The model selects INS information as input and GNSS position increment information as output; the dimension of the output vector of the second model is larger than that of the other two, the prediction effect is also poor, and the error of filtering is mixed in; the output vector of the first model contains both INS information and GNSS information, while the output vector of the third model is only related to GNSS position information, which reduces additional mixed errors, so it is adopted Model; It is directly derived from some INS information, as shown in formula (1); （1） In formula (1) is the direction cosine matrix, is the specific force, i.e. the accelerometer output, is the angular velocity of the e system relative to the i system, is the angular velocity of the n system relative to the e system, For speed, is the gravity vector; Further analysis of the gravity vector Related to longitude and latitude; and The expression is shown in formula (2). Due to the latitude , the meridian radius of the Earth and the radius of the circle In the vehicle environment, the changes are slow, so the two are influence, therefore influence The main factors are , , ;and Determined by the attitude angle, as shown in formula (3), the ideal attitude matrix differential equation is as shown in formula (4); （2） （3） （4） therefore Attitude Angle , Pitch angle , heading angle and gyroscope output However, in the vehicle environment, , It is approximately 0 and changes very little, so the input feature parameters of the final training model are selected as , , , , the output selection is ; The input of the GRU neural network training model is the three-axis gyroscope data of the IMU, the three-axis accelerometer data, the three-axis velocity and heading angle information solved by the INS, with a total dimension of 10 dimensions, and the output is GNSS increment information; During training, the frequencies of the input IMU sensor data, speed, heading angle INS information and GNSS increment information are different. The frequency of IMU data and INS information is 200 times per second, while the GNSS increment information can only be output once per second. Therefore, it is necessary to integrate the high-frequency IMU and INS information during training, and use the IMU and INS information within 1 second as the input layer to correspond to the low-frequency GNSS increment. In addition, the GRU neural network has a time dimension, which is reflected in the data as the existence of the time step parameter. The training data feature is 10 dimensions, the data length of the training set is 3000 seconds, and the sampling frequency is 200HZ. It can be seen that the data size of the input layer is (3000, time step, 2000). At the same time, the data at the output end is the corresponding GNSS increment information, and the size is (3000, 3). The amount of training data is huge, so it is necessary to set the appropriate number of hidden layer neurons and time step. step parameter, otherwise it will affect the speed and accuracy of training; in order to speed up the convergence speed during training and prevent overfitting, the training data was normalized before training, and the prediction results should be denormalized when using the prediction model for prediction; the final time step parameter was selected as 4, and the number of hidden layer units was selected as 128; The working modes of the AI ​​module include training mode and prediction mode. When the GNSS signal is good, the GRU neural network is trained and the input is the three-axis accelerometer data output by the IMU. , three-axis gyroscope data , the speed information obtained by INS and heading angle information ; Therefore, when the GNSS signal is good, the GRU neural network finds the mapping relationship between the IMU, INS navigation information and the GNSS increment information; when the GNSS signal is lost, the GNSS module can no longer provide position information, so there is no GNSS module in the prediction model; at this time, the IMU and INS modules continue to input navigation information to the GRU neural network, and the trained GRU neural network predicts the GNSS increment based on the input. ; Integrate the incremental information to obtain a predicted GNSS position information, also called pseudo GNSS position information, which is used to simulate the positioning result when the GNSS signal is good; compare the result with the position solved by INS After the difference is made, it is sent to the KF module for Kalman filtering, and the errors of position, velocity and attitude are continuously outputted to provide feedback and correction to the INS results. In this way, when the GNSS is lost, the combined navigation system, which could only use pure INS solution, can be corrected with pseudo-GNSS information, thus suppressing the error divergence rate of the inertial navigation.

2. The method for using a GRU neural network to assist navigation during GNSS signal interruption according to claim 1, characterized in that: The GRU neural network consists of two gate structures, the update gate and the reset gate. The update gate changes the degree to which the current moment is affected by the state of the previous moment, and the output is ; The value of the update gate is proportional to the amount of information retained. The smaller the update gate output, the less state information is retained; the function of the reset gate is to discard some hidden state information, and the output is ,The value of the reset gate is inversely proportional to the amount of discarded information. The smaller the value is, the more past state information is discarded.

Images