A vehicle navigation method based on LSTM neural network assistance

Abstract

The invention discloses a vehicle navigation method based on LSTM neural network assistance, comprising the following steps: S1. constructing the output signal characteristics of the sensor when the vehicle is running; S2. constructing the output signal characteristics of the GPS when the GPS is not interrupted, and obtaining the error data based on the extended Kalman filter processing, and subtracting the error data from the strapdown solution result to obtain navigation information; S3. constructing the LSTM neural network model, and training the LSTM neural network in the time period when the GPS is not interrupted; S4. when the GPS is interrupted, the LSTM neural network model outputs the pseudo position increment, and obtains the prediction error data based on the Kalman filter, and subtracts the error data from the strapdown solution result to obtain navigation information. The invention is based on the assistance of the LSTM neural network, and takes the specific force, angular velocity, and the speed and heading angle of the strapdown settlement of the MEMS IMU as input to predict the position increment when the GPS is interrupted, and uses this as the observation quantity of the EKF, so as to accurately estimate the trajectory of the vehicle.

Description

Technical Field

[0001] The invention relates to GPS navigation, and in particular to a vehicle navigation method assisted by an LSTM neural network. Background Art

[0002] Currently, the Global Positioning System (GPS) is the most readily available vehicle navigation technology and is regarded as the standard solution for a wide range of transportation applications. However, when satellite signals are intermittent or blocked in GPS-denied environments such as urban canyons and tunnels, standalone GPS cannot provide continuous and reliable positioning information. To address this problem, GPS is often integrated with an inertial navigation system (INS) to create an integrated INS / GPS positioning system that is more robust to environmental factors such as GPS signal blocking. INS is an autonomous system that uses an inertial measurement unit (IMU) to measure the linear acceleration and angular velocity of a land vehicle using three accelerometers and three gyroscopes. When the initial conditions of the navigation parameters are given, the IMU measurement data can be analyzed to provide the current position, velocity, and attitude. Therefore, INS can compensate for GPS interruption moments and maintain the continuity of the navigation solution. For the INS / GPS combined system, the navigation solution is traditionally achieved by using a Kalman filter (KF). The accuracy of the combined navigation based on the Kalman filter is mainly affected by two aspects. The first is the accuracy of the sensor. Because the dynamic model of the IMU is based on the three velocity errors, three position errors, and three attitude errors of the INS, which are collectively referred to as the system error state. In addition, there are some sensor errors, such as accelerometer bias and gyroscope drift, which are usually modeled as first-order GM process or AR modeling. However, IMU drift is non-stationary, the first-order GM process is only suitable for processing stationary random processes, and the AR model is model-sensitive and not suitable for processing odd-power processes, high-order processes or wide dynamic ranges.

[0003] In addition, the accuracy of integrated navigation is also affected by the filtering algorithm. Kalman filtering is the most common fusion implementation in GNSS / INS integrated navigation because of its optimal estimation advantage and simplicity in the recursive process. However, since the optimality of KF depends on the quality of the prior assumptions about the covariance matrix (Q and R), nonlinear approximation and insufficient knowledge of the system model may lead to unreliable results and even disagreements in complex environments. In order to better handle nonlinearity, some improvements of KF have been proposed, such as the extended Kalman filter (EKF) and the unscented Kalman filter (UKF). However, they still have some shortcomings in land vehicle GNSS / INS integrated navigation, such as the rapid decline in accuracy during long-term GNSS outages and the relatively poor ability to predict and model the nonlinear behavior of land vehicles in GNSS-challenged environments. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a vehicle navigation method assisted by an LSTM neural network. Based on the LSTM neural network, the specific force, angular velocity, and speed and heading angle of the strapdown settlement of MEMSIMU are used as input to predict the position increment when the GPS is interrupted, and this is used as the observation value of the EKF, so as to accurately estimate the trajectory of the vehicle.

[0005] The object of the present invention is to achieve the following technical solution: a vehicle navigation method based on LSTM neural network assistance, comprising the following steps:

[0006] S1. When the vehicle is running, the sensors installed on the vehicle are used to collect vehicle data, perform noise reduction and strapdown calculation, and construct the output signal characteristics of the sensor;

[0007] S2. When GPS is not interrupted, construct the GPS output signal characteristics according to the GPS module installed on the vehicle, and obtain the error data based on the extended Kalman filter, and make a difference with the strapdown solution result to obtain the navigation information;

[0008] S3. Construct an LSTM neural network model, and train the LSTM neural network using the output signal characteristics of the sensor and the output signal characteristics of the GPS during the period when the GPS is not interrupted;

[0009] S4. When GPS is interrupted, the LSTM neural network model outputs a pseudo position increment, and obtains the prediction error data based on the Kalman filter, which is subtracted from the strapdown solution result to obtain navigation information.

[0010] Furthermore, the sensor adopts an IMU sensor, which includes a three-axis accelerometer and a three-axis gyroscope, and the output signal includes three-axis accelerometer data f b and three-axis gyroscope data ω b The signal output by the GPS module includes position data and speed data. The signal output by the IMU sensor is a 100 Hz signal; The signal output by the GPS module is a 1 Hz signal.

[0011] The process of constructing the sensor output signal characteristics in step S1 includes:

[0012] S101. Collect vehicle data using sensors installed on the vehicle and output triaxial accelerometer data f b and three-axis gyroscope data ω b , and use WMRA for noise reduction, where WMRA refers to wavelet multi-resolution analysis;

[0013] S102. Integrate and average the noise-reduced data to match the output frequency of the GPS module:

[0014] The GPS module outputs a 1HZ signal, that is, it outputs a signal once per second, so the noise-reduced data within each second are superimposed:

[0015] Since the signal output by the sensor is 100HZ, that is, there are 100 data outputs per second, each time a three-axis accelerometer data and a three-axis gyroscope data are output;

[0016] The 100 three-axis accelerometer data after noise reduction per second are averaged, and the 100 three-axis gyroscope data after noise reduction per second are averaged, thus completing the integrated average.

[0017] S103. Perform INS strapdown calculation on the sensor output data after noise reduction to obtain vehicle speed information, attitude information and position information per second;

[0018] S104. Assume that the data output by the sensor in the tth second of vehicle operation is subjected to noise reduction and integrated averaging to obtain the three-axis accelerometer data and the three-axis gyroscope data, respectively. Assume that the speed information and attitude information obtained after the sensor output data is denoised and INS strapdown solution in the tth second of vehicle operation are Then the output signal characteristics of the sensor in the tth second are constructed as:

[0019]

[0020] The step S2 comprises:

[0021] The position signal output by the GPS module every second is subtracted from the position signal output in the previous second to calculate the GPS position increment as the GPS output signal feature;

[0022] For the tth second of vehicle operation, the output signal characteristics of GPS are recorded as:

[0023] y t =[P GPS (t)-P GPS (t-1)]

[0024] Where P GPS (t) represents the position data output by the GPS module at the tth second of vehicle operation. When t = 1, P GPS (t-1) Get the pre-saved position data output by the GPS module when the vehicle is stationary.

[0025] The step S3 comprises:

[0026] S301. Build LSTM neural network model;

[0027] S302. During the period when GPS is not interrupted, the output signal characteristics of the sensor in any second are used as the input of the LSTM neural network model, and the GPS output signal characteristics in that second are used as the target value of LSTM training, that is, the expected output of the LSTM neural network model, to train the constructed LSTM neural network model;

[0028] S302. During the period when GPS is not interrupted, the output signal characteristics of the sensor and the output signal characteristics of the GPS are used every second to continuously train the LSTM neural network according to step S302 to obtain a trained LSTM neural network.

[0029] In step S2, the error data is obtained based on the extended Kalman filter processing, and the difference is made with the strapdown solution result to obtain the navigation information as follows:

[0030] A1. Let the state matrix In the state matrix, each quantity represents position error, velocity error, attitude error, accelerometer error and gyroscope error. Each quantity contains data in the x, y and z directions in the vehicle coordinate system. is a 15-dimensional matrix;

[0031] Initialize the measurement matrix H t , at any time when GPS is not interrupted, H t All remain unchanged:

[0032]

[0033] Initialize the state matrix when the vehicle is stationary is an all-zero matrix, denoted by Initialize the state noise covariance matrix of the vehicle at rest to the preset matrix P0(+);

[0034] A2. If the GPS is not interrupted at the tth second of vehicle operation, the position data output by the GPS module is P GPS (t), the speed data is v GPS (t), the position data obtained by INS strapdown solution is P INS (t), speed data v obtained by INS strapdown solution INS (t), then calculate the observation value z of the extended Kalman filter at the tth second of vehicle operation t for:

[0035] z t = {P GPS (t)-P INS (t),v GPS(t)-v INS (t)};

[0036] A3. Let the noise covariance matrix of the known observation at the tth second be R t , the known process noise covariance at the t-1 second is Q t-1 , the known noise driving matrix at the t-1 second is G t-1 , where R t , Q t-1 , G t-1 Does not change over time;

[0037] Starting from t=1, the state matrix correction value is calculated based on the state matrix at t-1 seconds. Get the state matrix correction value at the tth second The process is as follows:

[0038]

[0039] Among them, φ t / t-1 represents the vehicle dynamic transfer matrix from time t-1 to time t, P t (-) represents the predicted value of the state noise covariance matrix at time t, P t (+) represents the correction value of the state noise covariance matrix at time t, P t-1 (+) represents the correction value of the state noise covariance matrix at time t-1, K t represents the Kalman filter gain matrix at time t;

[0040] P is the state noise covariance matrix, which represents the noise level of the state (IMU); R is the observation noise covariance matrix, which represents the noise level of the observation (GPS or LSTM). P is determined by the process noise covariance Q. Q and R are the prior information, that is, the known noise level of the IMU or GPS / LSTM; G is determined by the selected sensor noise level.

[0041] A4. Use the vehicle speed information, attitude information and position information obtained by INS strapdown settlement at the tth second, minus the updated information at the tth second. The velocity error, attitude error and position error contained in are used to obtain the required navigation information.

[0042] The step S4 comprises:

[0043] S401. When GPS is interrupted, the signal characteristics output by the sensor are sent to the trained neural network to obtain the predicted GPS position increment:

[0044] If the GPS is interrupted at the tth second of vehicle operation, the sensor output signal characteristic is Send it to the trained neural network and get the predicted output denoted as Δp ANN ;

[0045] S402. Construct the observation z of Kalman filter at time t t :

[0046] z t ={Δp ANN};

[0047] S403. Starting from t=1, obtain the state matrix correction value at the tth second according to step A3 In the process of processing according to step A4, the measurement matrix in step A3 is replaced by H t =[I 3×3 0 3×3 0 3×3 0 3×3 0 3×3 ], the matrix R t Replace with Right now:

[0048]

[0049] P t (+) = P t (-)-K t H t P t (-)

[0050] in,

[0051]

[0052] v t Defined as the innovation sequence,

[0053] c vt is the theoretical innovation sequence covariance, is the covariance of the actual new information sequence within the window size N;

[0054] S404. Using the vehicle speed information, attitude information and position information obtained by the INS strapdown settlement within the tth second, minus the state matrix correction value of the tth second The velocity error, attitude error and position error contained in are used to obtain the required navigation information.

[0055] The beneficial effects of the present invention are as follows: the present invention is based on the assistance of LSTM neural network, takes the specific force, angular velocity, and speed and heading angle of strapdown settlement of MEMS IMU as input to predict the position increment when GPS is interrupted, and uses this as the observation quantity of EKF, so as to accurately estimate the trajectory of the vehicle. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the method of the present invention;

[0057] Figure 2 This is a schematic diagram of data processing when GPS is not interrupted;

[0058] Figure 3 This is a schematic diagram of data processing when GPS is interrupted;

[0059] Figure 4 It is a schematic diagram of the comparison results in the embodiment. DETAILED DESCRIPTION

[0060] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0061] like Figure 1 As shown, a vehicle navigation method based on LSTM neural network assistance includes the following steps:

[0062] S1. When the vehicle is running, the sensors installed on the vehicle are used to collect vehicle data, perform noise reduction and strapdown calculation, and construct the output signal characteristics of the sensor;

[0063] S2. When GPS is not interrupted, construct the GPS output signal characteristics according to the GPS module installed on the vehicle, and obtain the error data based on the extended Kalman filter, and make a difference with the strapdown solution result to obtain the navigation information;

[0064] S3. Construct an LSTM neural network model, and train the LSTM neural network using the output signal characteristics of the sensor and the output signal characteristics of the GPS during the period when the GPS is not interrupted;

[0065] S4. When GPS is interrupted, the LSTM neural network model outputs a pseudo position increment, and obtains the prediction error data based on the Kalman filter, which is subtracted from the strapdown solution result to obtain navigation information.

[0066] Furthermore, the sensor adopts an IMU sensor, which includes a three-axis accelerometer and a three-axis gyroscope, and the output signal includes three-axis accelerometer data f b and three-axis gyroscope data ω b ; The signal output by the GPS module includes position data and speed data. The signal output by the IMU sensor is a 100HZ signal; The signal output by the GPS module is a 1HZ signal. In an embodiment of the present application, the IMU sensor adopts MEMS IMU;

[0067] The process of constructing the sensor output signal characteristics in step S1 includes:

[0068] S101. Collect vehicle data using sensors installed on the vehicle and output triaxial accelerometer data f b and three-axis gyroscope data ω b , and use WMRA for noise reduction, where WMRA refers to wavelet multi-resolution analysis; (using wavelet soft threshold processing, wavelet basis function is sym6, and the decomposition scale is 6 for the best effect)

[0069] S102. Integrate and average the noise-reduced data to match the output frequency of the GPS module:

[0070] The GPS module outputs a 1HZ signal, that is, it outputs a signal once per second, so the noise-reduced data within each second are superimposed:

[0071] Since the signal output by the sensor is 100HZ, that is, there are 100 data outputs per second, each time a three-axis accelerometer data and a three-axis gyroscope data are output;

[0072] The 100 three-axis accelerometer data after noise reduction per second are averaged, and the 100 three-axis gyroscope data after noise reduction per second are averaged, thus completing the integrated average.

[0073] S103. Perform INS strapdown calculation on the sensor output data after noise reduction to obtain vehicle speed information, attitude information and position information per second;

[0074] S104. Assume that the data output by the sensor in the tth second of vehicle operation is subjected to noise reduction and integrated averaging to obtain the three-axis accelerometer data and the three-axis gyroscope data, respectively. Assume that the speed information and attitude information obtained after the sensor output data is denoised and INS strapdown solution in the tth second of vehicle operation are Then the output signal characteristics of the sensor in the tth second are constructed as:

[0075]

[0076] The step S2 comprises:

[0077] The position signal output by the GPS module every second is subtracted from the position signal output in the previous second to calculate the GPS position increment as the GPS output signal feature;

[0078] For the tth second of vehicle operation, the output signal characteristics of GPS are recorded as:

[0079] y t =[P GPS(t)-P GPS (t-1)]

[0080] Where P GPS (t) represents the position data output by the GPS module at the tth second of vehicle operation. When t = 1, P GPS (t-1) Get the pre-saved position data output by the GPS module when the vehicle is stationary.

[0081] The step S3 comprises:

[0082] S301. Build LSTM neural network model;

[0083] S302. During the period when GPS is not interrupted, the output signal characteristics of the sensor in any second are used as the input of the LSTM neural network model, and the GPS output signal characteristics in that second are used as the target value of LSTM training, that is, the expected output of the LSTM neural network model, to train the constructed LSTM neural network model;

[0084] S302. During the period when GPS is not interrupted, the output signal characteristics of the sensor and the output signal characteristics of the GPS are used every second to continuously train the LSTM neural network according to step S302 to obtain a trained LSTM neural network.

[0085] like Figure 2 As shown, in step S2, error data is obtained based on extended Kalman filtering, and the difference between the error data and the strapdown solution result is made to obtain navigation information as follows:

[0086] A1. Let the state matrix In the state matrix, each quantity represents position error, velocity error, attitude error, accelerometer error and gyroscope error. Each quantity contains data in the x, y and z directions in the vehicle coordinate system. is a 15-dimensional matrix;

[0087] Initialize the measurement matrix H t , at any time when GPS is not interrupted, H t All remain unchanged:

[0088]

[0089] Initialize the state matrix x when the vehicle is stationary t is an all-zero matrix, denoted by Initialize the state noise covariance matrix of the vehicle at rest to the preset matrix P0(+);

[0090] A2. If the GPS is not interrupted at the tth second of vehicle operation, the position data output by the GPS module is P GPS(t), the speed data is v GPS (t), the position data obtained by INS strapdown solution is P INS (t), speed data v obtained by INS strapdown solution INS (t), then calculate the observation value z of the extended Kalman filter at the tth second of vehicle operation t for:

[0091] z t = {P GPS (t)-P INS (t),v GPS (t)-v INS (t)};

[0092] A3. Let the noise covariance matrix of the known observation at the tth second be R t , the known process noise covariance at the t-1 second is Q t-1 , the known noise driving matrix at the t-1 second is G t-1 , where R t , Q t-1 , G t-1 Does not change over time;

[0093] Starting from t=1, the state matrix correction value is calculated based on the state matrix at t-1 seconds. Get the state matrix correction value at the tth second The process is as follows:

[0094]

[0095] Among them, φ t / t-1 represents the vehicle dynamic transfer matrix from time t-1 to time t, P t (-) represents the predicted value of the state noise covariance matrix at time t, P t (+) represents the correction value of the state noise covariance matrix at time t, P t-1 (+) represents the correction value of the state noise covariance matrix at time t-1, K t represents the Kalman filter gain matrix at time t;

[0096] A4. Use the vehicle speed information, attitude information and position information obtained by INS strapdown settlement at the tth second, minus the updated information at the tth second. The velocity error, attitude error and position error contained in are used to obtain the required navigation information.

[0097] like Figure 3 As shown, step S4 includes:

[0098] S401. When GPS is interrupted, the signal characteristics output by the sensor are sent to the trained neural network to obtain the predicted GPS position increment:

[0099] If the GPS is interrupted at the tth second of vehicle operation, the sensor output signal characteristic is Send it to the trained neural network and get the predicted output denoted as Δp ANN ;

[0100] S402. Construct the observation z of Kalman filter at time t t :

[0101] z t ={Δp ANN};

[0102] S403. Starting from t=1, obtain the state matrix correction value at the tth second according to step A3 In the process of processing according to step A4, the measurement matrix in step A3 is replaced by H t =[I 3×3 0 3×3 0 3×3 0 3×3 0 3×3 ], the matrix R t Replace with Right now:

[0103]

[0104] P t (+) = P t (-)-K t H t P t (-)

[0105] in,

[0106]

[0107] v t Defined as the innovation sequence,

[0108] c vt is the theoretical innovation sequence covariance, is the covariance of the actual new information sequence within the window size N;

[0109] S404. Using the vehicle speed information, attitude information and position information obtained by the INS strapdown settlement within the tth second, minus the state matrix correction value of the tth second The velocity error, attitude error and position error contained in are used to obtain the required navigation information.

[0110] In the embodiments of the present application, the proposed navigation strategy is verified by road tests. The experimental equipment uses MTi-G (a low-cost MEMSIMU with a built-in GNSS chip), the MEMS data sampling rate is 100HZ, the GPS data sampling rate is 1HZ, and the error characteristics of MTi-G are shown in the table. In order to verify the improvement of the positioning accuracy of the traditional GPS / INS combined navigation system by this strategy, the collected MEMSIMU data and GNSS data (position and speed) are analyzed and fused based on EKF. The error characteristics of the MTi-G MEMSIMU are shown in the following table:

[0111]

[0112] The LSTM is trained through a 1800-second period of uninterrupted GPS signal travel time, and then a 30-second interruption is simulated to verify the performance of the proposed scheme.

[0113] like Figure 4 The figure shows the vehicle trajectory under a 30-second interruption. During the 30-second interruption, the latitude drift of the pure inertial navigation system reached 140m and the longitude drift reached 160m, while the LSTM neural network effectively limited the drift of the inertial navigation system. The maximum latitude error of the LSTM system was 10m, an improvement of 93%, and the maximum longitude error was 12.8m, an improvement of 92%.

[0114] The above is a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not depart from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A vehicle navigation method based on LSTM neural network assistance, characterized in that: The following steps are involved: S1. When the vehicle is running, the sensors installed on the vehicle are used to collect vehicle data, perform noise reduction and strapdown calculation, and construct the output signal characteristics of the sensor; S2. When GPS is not interrupted, construct the GPS output signal characteristics according to the GPS module installed on the vehicle, and obtain the error data based on the extended Kalman filter, and make a difference with the strapdown solution result to obtain the navigation information; S3. Construct an LSTM neural network model, and train the LSTM neural network using the output signal characteristics of the sensor and the output signal characteristics of the GPS during the period when the GPS is not interrupted; S4. When GPS is interrupted, the LSTM neural network model outputs a pseudo position increment, and obtains the prediction error data based on the Kalman filter, which is subtracted from the strapdown solution result to obtain navigation information; The step S4 comprises: S401. When GPS is interrupted, the signal characteristics output by the sensor are sent to the trained neural network to obtain the predicted GPS position increment: If the GPS is interrupted at the tth second of vehicle operation, the sensor output signal characteristic is Send it to the trained neural network and get the predicted output denoted as Δp ANN ; in, The three-axis accelerometer data and three-axis gyroscope data are obtained after noise reduction and integration averaging of the data output by the sensors within the tth second of vehicle operation. They are the speed information and attitude information obtained by denoising the data output by the sensor within the tth second of the vehicle's operation and performing INS strapdown solution; S402. Construct the observation z of Kalman filter at time t t : z t ={Δp ANN }; S403. Assume that the noise covariance matrix of the known observation at the tth second is R t , the known process noise covariance at the t-1 second is Q t-1 , the known noise driving matrix at the t-1 second is G t-1 ; Starting from t=1, the state matrix correction value is calculated based on the state matrix at t-1 seconds. Get the state matrix correction value at the tth second Let the measurement matrix be H t =[I 3×3 0 3×3 0 3×3 0 3×3 0 3×3 ],set up get: Among them, φ t / t-1 represents the vehicle dynamic transfer matrix from time t-1 to time t, P t (-) represents the predicted value of the state noise covariance matrix at time t, P t (+) represents the correction value of the state noise covariance matrix at time t, P t-1 (+) represents the correction value of the state noise covariance matrix at time t-1, K t represents the Kalman filter gain matrix at time t; in, v t Defined as the innovation sequence, c vt is the theoretical innovation sequence covariance, is the covariance of the actual new information sequence within the window size N; S404. Using the vehicle speed information, attitude information and position information obtained by the INS strapdown settlement within the tth second, minus the state matrix correction value of the tth second The velocity error, attitude error and position error contained in are used to obtain the required navigation information.

2. The vehicle navigation method based on LSTM neural network assistance according to claim 1 is characterized in that: The sensor adopts an IMU sensor, which includes a three-axis accelerometer and a three-axis gyroscope. The output signal includes three-axis accelerometer data f b and three-axis gyroscope data ω b ; The signal output by the GPS module includes position data and speed data.

3. The vehicle navigation method based on LSTM neural network assistance according to claim 2 is characterized in that: The signal output by the IMU sensor is a 100HZ signal; the signal output by the GPS module is a 1HZ signal.

4. The vehicle navigation method based on LSTM neural network assistance according to claim 1 is characterized in that: The process of constructing the sensor output signal characteristics in step S1 includes: S101. Collect vehicle data using sensors installed on the vehicle and output triaxial accelerometer data f b and three-axis gyroscope data ω b , and use WMRA for noise reduction, where WMRA refers to wavelet multi-resolution analysis; S102. Integrate and average the noise-reduced data to match the output frequency of the GPS module: The GPS module outputs a 1HZ signal, that is, it outputs a signal once per second, so the noise-reduced data within each second are superimposed: Since the signal output by the sensor is 100HZ, that is, there are 100 data outputs per second, each time a three-axis accelerometer data and a three-axis gyroscope data are output; The 100 three-axis accelerometer data after noise reduction per second are averaged, and the 100 three-axis gyroscope data after noise reduction per second are averaged, thus completing the integrated average. S103. Perform INS strapdown calculation on the sensor output data after noise reduction to obtain vehicle speed information, attitude information and position information per second; S104. Assume that the data output by the sensor in the tth second of vehicle operation is subjected to noise reduction and integrated averaging to obtain the triaxial accelerometer data and triaxial gyroscope data, respectively. Assume that the speed information and attitude information obtained after the sensor output data is denoised and INS strapdown solution in the tth second of vehicle operation are Then the output signal characteristics of the sensor in the tth second are constructed as:

5. The vehicle navigation method based on LSTM neural network assistance according to claim 1, characterized in that: The step S2 comprises: The position signal output by the GPS module every second is subtracted from the position signal output in the previous second to calculate the GPS position increment as the GPS output signal feature; For the tth second of vehicle operation, the output signal characteristics of GPS are recorded as: y t =[P GPS (t)-P GPS (t-1)] Where P GPS (t) represents the position data output by the GPS module at the tth second of vehicle operation. When t = 1, P GPS (t-1) Get the pre-saved position data output by the GPS module when the vehicle is stationary.

6. The vehicle navigation method based on LSTM neural network assistance according to claim 1, characterized in that: The step S3 comprises: S301. Build LSTM neural network model; S302. During the period when GPS is not interrupted, the output signal characteristics of the sensor in any second are used as the input of the LSTM neural network model, and the GPS output signal characteristics in that second are used as the target value of LSTM training, that is, the expected output of the LSTM neural network model, to train the constructed LSTM neural network model; S302. During the period when GPS is not interrupted, the output signal characteristics of the sensor and the output signal characteristics of the GPS are used every second to continuously train the LSTM neural network according to step S302 to obtain a trained LSTM neural network.

7. The vehicle navigation method based on LSTM neural network assistance according to claim 1, characterized in that: In step S2, the error data is obtained based on the extended Kalman filter processing, and the difference is made with the strapdown solution result to obtain the navigation information as follows: A1. Let the state matrix In the state matrix, each quantity represents position error, velocity error, attitude error, accelerometer error and gyroscope error. Each quantity contains data in the x, y and z directions in the vehicle coordinate system. is a 15-dimensional matrix; Initialize the measurement matrix H t , at any time when GPS is not interrupted, H t All remain unchanged: Initialize the state matrix when the vehicle is stationary is an all-zero matrix, denoted by Initialize the state noise covariance matrix of the vehicle at rest to the preset matrix P0(+); A2. If the GPS is not interrupted at the tth second of vehicle operation, the position data output by the GPS module is P GPS (t), the speed data is v GPS (t), the position data obtained by INS strapdown solution is P INS (t), speed data v obtained by INS strapdown solution INS (t), then calculate the observation value z of the extended Kalman filter at the tth second of vehicle operation t for: z t ={P GPS (t)-P INS (t),v GPS (t)-v INS (t)}; A3. Let the noise covariance matrix of the known observation at the tth second be R t , the known process noise covariance at the t-1 second is Q t-1 , the known noise driving matrix at the t-1 second is G t-1 ; Starting from t=1, the state matrix correction value is calculated based on the state matrix at t-1 seconds. Get the state matrix correction value at the tth second The process is as follows: Among them, φ t / t-1 represents the vehicle dynamic transfer matrix from time t-1 to time t, P t (-) represents the predicted value of the state noise covariance matrix at time t, P t (+) represents the correction value of the state noise covariance matrix at time t, P t-1 (+) represents the correction value of the state noise covariance matrix at time t-1, K t represents the Kalman filter gain matrix at time t; A4. Use the vehicle speed information, attitude information and position information obtained by INS strapdown settlement at the tth second, minus the updated information at the tth second. The velocity error, attitude error and position error contained in are used to obtain the required navigation information.

Images