CN112505737A - GNSS/INS combined navigation method based on Elman neural network online learning assistance - Google Patents

Abstract

The invention discloses a GNSS/INS combined navigation method based on Elman neural network online learning assistance, which includes two parts: (1) implementing a conventional GNSS/MEMS-INS combined navigation algorithm, and designing a Kalman filter for GNSS signals Fusion with inertial navigation data, and output the fused navigation data; (2) Design a neural network model on the basis of step (1), and then combine the inertial navigation data obtained in step (1) with the output of the Kalman filter. The data are respectively used as the sample input and sample output of the training neural network to train the neural network model. The present invention predicts the output error of the inertial navigation system when the GNSS signal of the UAV navigation system is lost, and uses the error data for the inertial navigation system. The output of the system is compensated and corrected, so that the inertial navigation system can output accurate navigation data with the aid of the neural network algorithm when the GNSS signal is lost.

Description

An online learning-assisted GNSS/INS integrated navigation method based on Elman neural network

technical field

The invention belongs to the technical field of combined navigation methods, in particular to a GNSS/INS combined navigation method based on Elman neural network online learning assistance.

Background technique

In terms of navigation technology, the most widely used and mature navigation methods are inertial navigation and satellite navigation. The advantage of GNSS satellite navigation is that it has the characteristics of global, all-weather, and long-term positioning accuracy, but the disadvantage is that the signal is easily interfered and blocked. In a strong electromagnetic environment and when there is a high-rise building block, the signal quality becomes poor, and the output frequency is limited, generally 1-10Hz, and the output is discontinuous. On the UAV system, the shortcomings of GNSS satellite navigation are highlighted. The INS inertial navigation system is a fully autonomous navigation method, so it has strong concealment and anti-interference ability, and the output information is continuous, and the positioning accuracy is high in a short time. However, due to the characteristics of MEMS-INS devices, gyroscopes and accelerometers have errors such as initial bias and random drift. With the accumulation of time, the errors will become larger and larger, and the positioning accuracy will be poor for a long time, which cannot be accurately reflected in the end. Attitude and position information of the drone.

In order to overcome the defects of the two types of navigation, the usual practice is to fuse the satellite navigation and inertial navigation signals through Kalman filtering, and use their respective advantages to make up for their respective shortcomings. However, under some special environmental conditions, such as signal blocking areas and environments with many obstructions, satellite signals may be lost. At this time, the navigation system can only rely on pure inertial navigation. With the passage of time, the navigation data The error will increase. Therefore, it is necessary to study a method that can replace the role of GNSS and cooperate with inertial navigation to complete the output of navigation data when the GNSS signal is lost.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present invention discloses a GNSS/INS integrated navigation method based on Elman neural network online learning assistance, and proposes a method for predicting the output error of the inertial navigation system of the UAV navigation system when the GNSS signal is lost. , and use the error data to compensate and correct the output of the inertial navigation system. In order to realize the loss of GNSS signal in the navigation system, the inertial navigation system can output accurate navigation data with the assistance of neural network algorithm.

For achieving the above object, technical scheme of the present invention is as follows:

An online learning-assisted GNSS/INS integrated navigation method based on Elman neural network, comprising the following steps:

1. Implement a conventional GNSS/MEMS-INS integrated navigation algorithm, design a Kalman filter to fuse GNSS signals and inertial navigation data, and output the fused navigation data;

2. Design the neural network model on the basis of step 1, and then use the inertial navigation data obtained in step 1 and the data output by the Kalman filter as the sample input and sample output of training the neural network, respectively, to train the neural network model, When the GNSS signal is lost, the trained neural network model is used to predict the inertial navigation output error, and the predicted error is used to compensate and correct the inertial navigation.

Further, a GNSS/INS combined navigation method based on Elman neural network online learning assistance described in the present invention specifically includes the following steps:

S1, IMU and GNSS equipment are installed, and the IMU coordinate axis is matched with the carrier axis; the LGM851 vehicle-mounted integrated navigation locator based on Beidouxingtong is used as the navigation calculation unit, LoRa communication is used to ensure data security, and the serial data cable is used to connect the computer and the integrated navigation. system, using computer to process integrated navigation data in real time;

S2, INS and GNSS antenna external parameter calibration, use the total station to calibrate the installation external parameters of INS and GNSS receiver antenna, and input it into the system as the initial value;

S3. Install the combined equipment on the trolley, test it on the experimental site, and start data collection and processing, which is divided into the following steps:

Initial alignment of IMU with GNSS assistance under S3-1 dynamics

转换到地理坐标系下ⁿV_G，The initial alignment is done by dynamic motion. Since GNSS can give the speed of the carrier relative to the local geographic coordinate system, assuming that the position ^e P _GNSS and the speed ^e V _GNSS of the GNSS receiver under ECEF have been obtained, the speed is calculated by the formula

Converted to ⁿ V _G in the geographic coordinate system,

The yaw and pitch angles are determined by projecting the velocity in the geographic coordinate system, and the formula is as follows:

Since it is aimed at the vehicle carrier, roll is directly assigned a value of 0, and a larger initial variance is given at the same time.

S3-2 Inertial Navigation Solution Based on ECEF Coordinate System

加速度矢量

使用组合系统获取GNSS接收机的位置和速度^eP_k+1,G和^eV_k+1,G。S3-2-1. In order to better initialize the IMU, firstly move the carrier slowly, the neural network learning part will not start temporarily, after the initialization is completed, start the neural network learning part; read the carrier parameters collected by the inertial navigation module, including the angular velocity information

acceleration vector

The position and velocity of the GNSS receiver, ^e P _k+1,G and ^e V _{k+1,G ,} are obtained using a combined system.

S3-2-2, obtain the real-time rotation quaternion ^e q=[q ₀ q ₁ q ₂ q ₃ ] ^T by solving the following quaternion differential equation;

和步骤S3-2-2求解的姿态四元数

求解下式的微分方程得到载体在导航坐标系下的三个方向上的速度信息：S3-2-3, the carrier acceleration vector obtained according to step S3-2-1

and the attitude quaternion solved in step S3-2-2

Solve the differential equation of the following formula to obtain the speed information of the carrier in the three directions in the navigation coordinate system:

为地球自转角速度，^eg为重力在地理坐标系下的加速度；where V=[V _E V _N V _U ] ^T is the velocity in the Middle East, North and Sky directions of the geographic coordinate system, respectively,

is the angular velocity of the earth's rotation, and ^e g is the acceleration of gravity in the geographic coordinate system;

S3-2-4, obtain the position parameters of inertial navigation output by the following formulas;

^e P _k+1,INS ≈ ^e P _k,EKF +( ^e V _k,EKF + ^e V _k+1,INS )Δt/2 (5)

S3-3 Constraint Update Assisted by Elman Neural Network

When the GNSS signal is good, the neural network is connected to the system to predict the output error of the inertial navigation and compensate and correct the output of the inertial navigation.

S3-3-1 Elman Neural Network Design

The mathematical model of Elman neural network is:

为输入层与隐含层连接权值；

为承接层与隐含层连接权值；

为隐含层与输出层连接权值。Elman神经网络的隐含层可采用式(7)所示的双极性Sigmoid函数作为激励函数，输出层采用Pureline激活函数，可得到式(8)：Among them: y(k) is the output of the output layer; x(k-1) is the external input of the input layer; u ^c (k) is the output of the successor layer; u(k) is the output of the hidden layer;

Connect the weights for the input layer and the hidden layer;

is the connection weight between the successor layer and the hidden layer;

Connect the weights for the hidden layer and the output layer. The hidden layer of the Elman neural network can use the bipolar sigmoid function shown in formula (7) as the excitation function, and the output layer adopts the Pureline activation function, and the formula (8) can be obtained:

f(a)=(1- ^e -a)/(1+ ^e -a) (7)

From equations (6) to (8), it can be deduced:

和

分别表示k-1、k-2时刻的网络连接权值。u^c(k)与其历史时间的连接权值有关，体现了动态递归的特点。where: f(a)∈(-1,1);

and

Represent the network connection weights at moments k-1 and k-2, respectively. u ^c (k) is related to the connection weight of its historical time, which reflects the characteristics of dynamic recursion.

The Elman neural network uses the BP algorithm to correct the weights, and the learning index function uses the error sum of squares function:

S3-3-2 conducts online learning based on the output of inertial sensors and EKF;

和

如公式(11)。S3-3-2-1 uses the position and velocity output by GNSS to correct the predicted data obtained by the inertial sensor in step S1, and the filtered data can be obtained by Kalman filtering.

and

as formula (11).

The GNSS measurement status in the EKF is updated to:

和

分别是惯性传感器输出值；^eP_k+1,G和^eV_k+1,G分别是GNSS位置和速度的输出值；

和

分别为卡尔曼滤波解。in

and

are the output values of the inertial sensor, respectively; ^e P _k+1,G and ^e V _k+1,G are the output values of the GNSS position and velocity, respectively;

and

are the Kalman filter solutions.

和

分别作为Elman神经网络的输入输出加入到训练样本去在线学习，使得网络具有预测速度和位置变化量的能力，再根据前一时刻的位置和速度获得本历元的位置、旋转四元数和速度，当训练误差满足一定的阈值设置是时，表示训练模型可靠，可以辅助系统进行预报；S3-3-2-2 simultaneously trains the neural network model simultaneously. The neural network model is designed on the basis of step S3-3-2-1, and then the relative change value between epochs between the inertial navigation data obtained in step S3-3-2-1 and the output data of the Kalman filter, when the GNSS When the number of available satellites is large, set the

and

As the input and output of the Elman neural network, they are added to the training samples for online learning, so that the network has the ability to predict the speed and position change, and then obtain the position, rotation quaternion and speed of the current epoch according to the position and speed of the previous moment. , when the training error meets a certain threshold setting, it means that the training model is reliable and can assist the system in forecasting;

In order to ensure the timeliness of the system, the fixed window method is used to limit the number of samples. When the number of samples reaches a certain number, in order to ensure the timeliness of the samples, the older samples are removed from the window.

分别是惯性传感器输出值的历元间变化量；

分别是卡尔曼滤波后输出值的历元间变化量；

和

分别是k时刻惯性传感器输出值；

和

分别为k时刻卡尔曼滤波输出解。in

are the inter-epoch variation of the output value of the inertial sensor, respectively;

are the inter-epoch variation of the output value after Kalman filtering, respectively;

and

are the output values of the inertial sensor at time k, respectively;

and

are the output solutions of the Kalman filter at time k, respectively.

S3-3-3 When the online learning meets the threshold and can predict, classify according to whether the satellite signal is blocked or not;

作为Elman神经网络进行预测，输出神经网络预测数据为

和

作为观测约束进行EKF更新。S3-3-3-1 When the GNSS satellite signal is blocked, the INS can work normally due to its properties; the GNSS data at the moment before the failure is the initial value of the filter, which is equivalent to the accurate value, so it can avoid the INS accuracy due to navigation Divergence caused by long time; when the GNSS signal is lost, the trained neural network model is used to predict the output error of inertial navigation, and the prediction error is used to compensate and correct the inertial navigation.

As Elman neural network for prediction, the output neural network prediction data is

and

EKF update as observation constraint.

和

求得的k+1时刻的位置、旋转四元数和速度的预测值为：Neural network predicts the amount of change

and

The obtained predicted values of the position, rotation quaternion and velocity at time k+1 are:

The constraints are updated to:

和

分别是根据惯性传感器在步骤S3-1输出的状态向量；

和

分别为神经网络求得的k+1时刻的位置、旋转四元数和速度的预测值。in,

and

are respectively the state vector outputted by the inertial sensor in step S3-1;

and

are the predicted values of the position, rotation quaternion and velocity at time k+1 obtained by the neural network, respectively.

和

及利用EKF求出的

和

加入到神经网络的在线学习系统中进行实时学习训练。S3-3-3-2 When the GNSS signal is normal, there are two strategies available: 1. Directly use the position, velocity and rotation quaternion at the moment of k+1 output by the inertial sensor as the state vector update, and then use the GNSS output 2. Input the position, velocity and rotation quaternion variation of inertial sensor output k+1 time into Elman neural network, and input the output of Elman neural network into Elman neural network. The position, rotation quaternion and velocity change are used as pseudo-observed values to update the state vector of the EKF, as in formula (16); then the constraint update is performed according to the position and velocity output by GNSS, as in formula (17); requested

and

and obtained using EKF

and

Join the online learning system of the neural network for real-time learning and training.

EKF status updated to:

The GNSS measurement status in the EKF is updated to:

和

分别是根据输出的神经网络预测变化量

和

求得的k+1时刻的位置、旋转四元数和速度的预测值；^eP_k+1,G和^eV_k+1,G分别是GNSS位置和速度的输出值。

and

are the predicted changes based on the output neural network

and

The obtained predicted values of position, rotation quaternion and velocity at time k+1; ^e P _k+1,G and ^e V _k+1,G are the output values of GNSS position and velocity, respectively.

The beneficial effects of the present invention are:

The GNSS/INS combined navigation method based on the Elman neural network online learning assistance described in the present invention solves the problem in some special environment conditions, such as the signal blocking area, the environment with many occluders or the occlusion time is long. This paper proposes a method to predict the output error of the inertial navigation system of the UAV navigation system when the GNSS signal is lost, and use the error data to compensate and correct the output of the inertial navigation system. In order to realize the loss of GNSS signal in the navigation system, the inertial navigation system can output accurate navigation data with the assistance of neural network algorithm.

Description of drawings

Figure 1 is a flow chart of GNSS/INS integrated navigation based on genetic neural network (when there are enough available GNSS satellites).

Figure 2 is a flow chart of GNSS/INS integrated navigation based on genetic neural network (when the number of GNSS satellites is insufficient).

Fig. 3 is the structure diagram of Elman neural network.

Figure 4 is a dynamic quasi-analytical diagram.

Figure 5 is a diagram of an integrated navigation device.

Detailed ways

The present invention will be further clarified below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and not to limit the scope of the present invention.

A GNSS/INS combined navigation method based on Elman neural network online learning assistance described in the present invention specifically includes the following steps:

S1, IMU and GNSS equipment installation, in order to reduce the error of the lever arm, the IMU coordinate axis should be matched with the carrier axis as much as possible; as shown in Figure 5, it is a diagram of the integrated navigation hardware system based on GNSS/INS. The invention adopts the LGM851 vehicle-mounted integrated navigation locator based on Beidouxingtong as the navigation calculation unit, adopts LoRa communication to ensure data security, connects the computer and the integrated navigation system with a serial data line, and uses the computer to process the integrated navigation data in real time.

S2, INS and GNSS antenna external parameter calibration, use the total station to calibrate the installation external parameters of INS and GNSS receiver antenna, and input it into the system as the initial value;

S3. Install the combined equipment on the trolley, test it at the experimental site of Sipailou campus of Southeast University, and start data collection and processing, which is divided into the following steps:

Initial alignment of IMU with GNSS assistance under S3-1 dynamics

(可以直接求出)转换到地理坐标系下ⁿV_G。具体如图4所示。In the GNSS/INS integrated navigation system, the initial alignment is a more critical step. The inertial navigation system uses the dead reckoning algorithm to achieve continuous autonomous navigation to obtain full navigation parameters (position, speed and attitude information). Therefore, the first The acquisition of the initial state of the epoch is particularly critical. At the same time, since GNSS can easily give the position and velocity (if the receiver can output the Doppler frequency shift, the instantaneous velocity can be directly obtained through Doppler integration; if not, it can be obtained through the position difference between receiver epochs) average velocity), so the critical task of initial alignment falls to attitude determination. Therefore, the initial alignment of the IMU mainly refers to the attitude initialization, and the attitude accuracy becomes the main factor affecting the navigation accuracy. For low-precision MEMS IMU static coarse alignment initialization, because the gyroscope measurement accuracy is low, it cannot sense the angular velocity of the earth's rotation, so the coarse alignment can only rely on dynamic motion to complete the initial alignment. Because GNSS can give the speed of the carrier relative to the local geographic coordinate system (n system). Assuming that the position ^e P _GNSS and the velocity ^e V _GNSS of the GNSS receiver under ECEF have been obtained, the velocity can be calculated by the formula

(It can be directly obtained) Convert to ⁿ V _G under the geographic coordinate system. Specifically as shown in Figure 4.

Determine the yaw and pitch angles by projecting the velocity under the n system. The schematic diagram and formula are as follows:

As for the vehicle carrier, roll can be directly assigned to 0, and at the same time give a larger initial variance.

S3-2 Inertial Navigation Solution Based on ECEF Coordinate System

加速度矢量

使用组合系统获取GNSS接收机的位置和速度^eP_k+1,G和^eV_k+1,G。S3-2-1. In order to better initialize the IMU, first move the carrier slowly, and the neural network learning part will not be started temporarily. After the initialization is completed, start the neural network learning part. Read the carrier parameters collected by the inertial navigation module, including angular velocity information

acceleration vector

The position and velocity of the GNSS receiver, ^e P _k+1,G and ^e V _{k+1,G ,} are obtained using a combined system.

S3-2-2, obtain the real-time rotation quaternion ^e q=[q ₀ q ₁ q ₂ q ₃ ] ^T by solving the following quaternion differential equation;

和步骤S3-2-2求解的姿态四元数

求解下式的微分方程得到载体在导航坐标系下的三个方向上的速度信息：S3-2-3, the carrier acceleration vector obtained according to step S3-2-1

and the attitude quaternion solved in step S3-2-2

Solve the differential equation of the following formula to obtain the speed information of the carrier in the three directions in the navigation coordinate system:

为地球自转角速度，^eg为重力在地理坐标系下的加速度；where V=[V _E V _N V _U ] ^T is the velocity in the Middle East, North and Sky directions of the geographic coordinate system, respectively,

is the angular velocity of the earth's rotation, and ^e g is the acceleration of gravity in the geographic coordinate system;

S3-2-4, the position parameters of the inertial navigation output can be obtained respectively by the following formulas.

^e P _k+1,INS ≈ ^e P _k,EKF +( ^e V _k,EKF + ^e V _k+1,INS )Δt/2 (22)

S3-3 Constraint Update Assisted by Elman Neural Network

When the GNSS signal is good, the neural network is connected to the system to predict the output error of the inertial navigation and compensate and correct the output of the inertial navigation. Figure 2.

S3-3-1 Elman Neural Network Design

Elman neural network is a typical dynamic recurrent neural network. It is based on the basic structure of BP network, adding a successor layer to the hidden layer as a one-step delay operator to achieve the purpose of memory, so that the system has the ability to adapt The ability of time-varying characteristics enhances the global stability of the network. It has stronger computing power than the feedforward neural network and can also be used to solve fast optimization problems. The Elman neural network structure is shown in Figure 1. There are four layers in total, namely the input layer, the hidden layer, the successor layer and the output layer. Specifically as shown in Figure 3. The mathematical model of Elman neural network is:

为输入层与隐含层连接权值；

为承接层与隐含层连接权值；

为隐含层与输出层连接权值。Elman神经网络的隐含层可采用式(7)所示的双极性Sigmoid函数作为激励函数，输出层采用Pureline激活函数，可得到式(8)：Among them: y(k) is the output of the output layer; x(k-1) is the external input of the input layer; u ^c (k) is the output of the successor layer; u(k) is the output of the hidden layer;

Connect the weights for the input layer and the hidden layer;

is the connection weight between the successor layer and the hidden layer;

Connect the weights for the hidden layer and the output layer. The hidden layer of the Elman neural network can use the bipolar sigmoid function shown in formula (7) as the excitation function, and the output layer adopts the Pureline activation function, and the formula (8) can be obtained:

f(a)=(1- ^e -a)/(1+ ^e -a)

(twenty four)

From equations (6) to (8), it can be deduced:

和

分别表示k-1、k-2时刻的网络连接权值。u^c(k)与其历史时间的连接权值有关，体现了动态递归的特点。where: f(a)∈(-1,1);

and

Represent the network connection weights at moments k-1 and k-2, respectively. u ^c (k) is related to the connection weight of its historical time, which reflects the characteristics of dynamic recursion.

The Elman neural network uses the BP algorithm to correct the weights, and the learning index function uses the error sum of squares function:

S3-3-2 conducts online learning based on the output of inertial sensors and EKF;

和

如公式(11)。S3-3-2-1 uses the position and velocity output by GNSS to correct the predicted data obtained by the inertial sensor in step S1, and the filtered data can be obtained by Kalman filtering.

and

as formula (11).

The GNSS measurement status in the EKF is updated to:

和

分别是惯性传感器输出值；^eP_k+1,G和^eV_k+1,G分别是GNSS位置和速度的输出值；

和

分别为卡尔曼滤波解。in

and

are the output values of the inertial sensor, respectively; ^e P _k+1,G and ^e V _k+1,G are the output values of the GNSS position and velocity, respectively;

and

are the Kalman filter solutions.

和

分别作为Elman神经网络的输入输出加入到训练样本去在线学习，假设杠臂已经过补偿，文中暂不考虑GNSS和INS之间时延误差的影响。使得网络具有预测速度和位置变化量的能力，再根据前一时刻的位置和速度获得本历元的位置、旋转四元数和速度，当训练误差满足一定的阈值设置是时，表示训练模型可靠，可以辅助系统进行预报；S3-3-2-2 simultaneously trains the neural network model simultaneously. The neural network model is designed on the basis of step S3-3-2-1, and then the relative change value between epochs between the inertial navigation data obtained in step S3-3-2-1 and the Kalman filter output data (ie k The relative variation of position, rotation quaternion and velocity between epochs and k+1, the inertial sensor can be solved according to the pre-integration, and the relative change after Kalman filtering can be solved according to formulas (12) and (13)), the specific The implementation steps are shown in Figure 1. When the number of available GNSS satellites is large, set the

and

As the input and output of the Elman neural network, they are added to the training samples for online learning. Assuming that the lever arm has been compensated, the influence of the delay error between GNSS and INS is not considered in this paper. Make the network have the ability to predict the speed and position change, and then obtain the position, rotation quaternion and speed of this epoch according to the position and speed of the previous moment. When the training error meets a certain threshold setting, it means that the training model is reliable. , which can assist the system in forecasting;

In order to ensure the timeliness of the system, the fixed window method is used to limit the number of samples. When the number of samples reaches a certain number, in order to ensure the timeliness of the samples, the older samples are removed from the window.

分别是惯性传感器输出值的历元间变化量；

分别是卡尔曼滤波后输出值的历元间变化量；

和

分别是k时刻惯性传感器输出值；

和

分别为k时刻卡尔曼滤波输出解。in

are the inter-epoch variation of the output value of the inertial sensor, respectively;

are the inter-epoch variation of the output value after Kalman filtering, respectively;

and

are the output values of the inertial sensor at time k, respectively;

and

are the output solutions of the Kalman filter at time k, respectively.

S3-3-3 When the online learning meets the threshold and can predict, classify according to whether the satellite signal is blocked or not;

作为Elman神经网络进行预测，输出神经网络预测数据为

和

作为观测约束进行EKF更新。S3-3-3-1 When the GNSS satellite signal is blocked, the INS can work normally due to its properties. The GNSS data at the moment before the failure is the initial value of the filter, which is equivalent to the accurate value, so the divergence of the INS accuracy caused by the long navigation time can be avoided. When the GNSS signal is lost, the trained neural network model is used to predict the inertial navigation output error, and the predicted error is used to compensate and correct the inertial navigation. The specific implementation steps are shown in Figure 2. First, the pre-integration based on the inertial sensor

As Elman neural network for prediction, the output neural network prediction data is

and

EKF update as observation constraint.

和

求得的k+1时刻的位置、旋转四元数和速度的预测值为：Neural network predicts the amount of change

and

The obtained predicted values of the position, rotation quaternion and velocity at time k+1 are:

The constraints are updated to:

和

分别是根据惯性传感器在步骤S3-1输出的状态向量；

和

分别为神经网络求得的k+1时刻的位置、旋转四元数和速度的预测值。in,

and

are respectively the state vector outputted by the inertial sensor in step S3-1;

and

are the predicted values of the position, rotation quaternion and velocity at time k+1 obtained by the neural network, respectively.

和

及利用EKF求出的

和

加入到神经网络的在线学习系统中进行实时学习训练。如图1。S3-3-3-2 When the GNSS signal is normal, we have two strategies to apply. 1. Directly use the position, velocity and rotation quaternion at the moment of k+1 output by the inertial sensor as the state vector update, and then use the position and velocity output by GNSS to update the constraints, as in step S3-3-2; 2. The position, velocity, and rotation quaternion changes at time k+1 output by the inertial sensor are input into the Elman neural network, and the position, rotation quaternion, and velocity changes output by the Elman neural network are used as pseudo-observed values. Update, such as formula (16); then update the constraints according to the position and velocity output by GNSS, such as formula (17); finally, use the inertial sensor to calculate

and

and obtained using EKF

and

Join the online learning system of the neural network for real-time learning and training. Figure 1.

EKF status updated to:

The GNSS measurement status in the EKF is updated to:

和

分别是根据输出的神经网络预测变化量

和

求得的k+1时刻的位置、旋转四元数和速度的预测值；^eP_k+1,G和^eV_k+1,G分别是GNSS位置和速度的输出值。

and

are the predicted changes based on the output neural network

and

The obtained predicted values of position, rotation quaternion and velocity at time k+1; ^e P _k+1,G and ^e V _k+1,G are the output values of GNSS position and velocity, respectively.

The technical means disclosed in the solution of the present invention are not limited to the technical means disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features.

Claims (2)

1. a GNSS/INS combined navigation method based on Elman neural network online learning assistance, is characterized in that: comprise the following steps: (1), realize the integrated navigation algorithm based on conventional GNSS/MEMS-INS, design Kalman filter to fuse GNSS signal and inertial navigation data, and output the fused navigation data; (2) Design a neural network model on the basis of step (1), and then use the inertial navigation data obtained in step (1) and the data output by the Kalman filter as the sample input and sample output of the training neural network, respectively. The neural network model is trained. When the GNSS signal is lost, the trained neural network model is used to predict the output error of the inertial navigation, and the prediction error is used to compensate and correct the inertial navigation. 2. a kind of GNSS/INS combined navigation method based on Elman neural network online learning assistance according to claim 1, is characterized in that: specifically comprises the following steps: S1, IMU and GNSS equipment are installed, and the IMU coordinate axis is matched with the carrier axis; the LGM851 vehicle-mounted integrated navigation locator based on Beidouxingtong is used as the navigation calculation unit, LoRa communication is used to ensure data security, and the serial data cable is used to connect the computer and the integrated navigation. system, using computer to process integrated navigation data in real time; S2, INS and GNSS antenna external parameter calibration, use the total station to calibrate the installation external parameters of INS and GNSS receiver antennas, and input them into the system as initial values; S3. Install the combined equipment on the trolley, test it on the experimental site, and start data collection and processing, which is divided into the following steps: Initial alignment of IMU with GNSS assistance under S3-1 dynamics The initial alignment is done by dynamic motion. Since GNSS can give the speed of the carrier relative to the local geographic coordinate system, assuming that the position ^e P _GNSS and the speed ^e V _GNSS of the GNSS receiver under ECEF have been obtained, the speed is calculated by the formula

Converted to ⁿ V _G in the geographic coordinate system,

The yaw and pitch angles are determined by projecting the velocity in the geographic coordinate system, and the formula is as follows:

Because for the vehicle carrier, the roll is directly assigned to 0, and a large initial variance is given at the same time; S3-2 Inertial Navigation Solution Based on ECEF Coordinate System S3-2-1. In order to better initialize the IMU, firstly move the carrier slowly, the neural network learning part will not start temporarily, after the initialization is completed, start the neural network learning part; read the carrier parameters collected by the inertial navigation module, including the angular velocity information

acceleration vector

Use the combined system to obtain the position and velocity of the GNSS receiver ^e P _k+1,G and ^e V _k+1,G , S3-2-2, obtain the real-time rotation quaternion ^e q=[q ₀ q ₁ q ₂ q ₃ ] ^T by solving the following quaternion differential equation; ^e q _k+1,INS = ^e q _k,EKF (I+[w ^× ]) (3) S3-2-3, the carrier acceleration vector obtained according to step S3-2-1

and the attitude quaternion solved in step S3-2-2

Solve the differential equation of the following formula to obtain the speed information of the carrier in the three directions in the navigation coordinate system:

where V=[V _E V _N V _U ] ^T is the velocity in the Middle East, North and Sky directions of the geographic coordinate system, respectively,

is the angular velocity of the earth's rotation, and ^e g is the acceleration of gravity in the geographic coordinate system; S3-2-4, obtain the position parameters of inertial navigation output by the following formulas; ^e P _k+1,INS ≈ ^e P _k,EKF +( ^e V _k,EKF + ^e V _k+1,INS )Δt/2 (5) S3-3 Constraint Update Assisted by Elman Neural Network When the GNSS signal is good, the neural network is connected to the system to predict the output error of the inertial navigation and compensate and correct the output of the inertial navigation, S3-3-1 Elman Neural Network Design The mathematical model of Elman neural network is:

Among them: y(k) is the output of the output layer; x(k-1) is the external input of the input layer; u ^c (k) is the output of the successor layer; u(k) is the output of the hidden layer; w ¹ is the input The connection weight between the layer and the hidden layer; w ² is the connection weight between the successor layer and the hidden layer; w ³ is the connection weight between the hidden layer and the output layer. The hidden layer of the Elman neural network adopts the formula shown in formula (7). The bipolar sigmoid function is used as the excitation function, and the output layer adopts the Pureline activation function, and equation (8) is obtained: f(a)=(1- ^e -a)/(1+ ^e -a) (7) y(k)=w ³ u(k) (8) It can be deduced from formulas (6) to (8): u ^c (k)=f(w ² (k-1)u ^c (k-1)+w ¹ (k-2)u ^c (k-2)) (9) where: f(a)∈(-1,1); w ² (k-1) and w ¹ (k-2) represent the network connection weights at moments k-1 and k-2, respectively, u ^c (k) It is related to the connection weight of its historical time, which reflects the characteristics of dynamic recursion. The Elman neural network uses the BP algorithm to correct the weights, and the learning index function uses the error sum of squares function:

S3-3-2 conducts online learning based on the output of inertial sensors and EKF; S3-3-2-1 uses the position and velocity output by GNSS to correct the predicted data obtained by the inertial sensor in step S1, and obtains the filtered value through Kalman filtering.

and

As in formula (11), The GNSS measurement status in the EKF is updated to:

in

and

are the output values of the inertial sensor, respectively; ^e P _k+1,G and ^e V _k+1,G are the output values of the GNSS position and velocity, respectively;

and

are the Kalman filter solutions, respectively, S3-3-2-2 simultaneously trains the neural network model, designs the neural network model on the basis of step S3-3-2-1, and then combines the inertial navigation data obtained in step S3-3-2-1 with the The relative change value between epochs of the output data of the Kalman filter. When the number of available satellites in GNSS is large, the

and

As the input and output of the Elman neural network, they are added to the training samples for online learning, so that the network has the ability to predict the speed and position change, and then obtain the position, rotation quaternion and speed of the current epoch according to the position and speed of the previous moment. , when the training error meets a certain threshold setting, it means that the training model is reliable, and the auxiliary system makes predictions; In order to ensure the timeliness of the system, the fixed window method is used to limit the number of samples. When the number of samples reaches a certain number, in order to ensure the timeliness of the samples, the older samples are removed from the window.

in

are the inter-epoch variation of the output value of the inertial sensor, respectively;

are the inter-epoch variation of the output value after Kalman filtering, respectively;

and

are the output values of the inertial sensor at time k, respectively;

and

are the Kalman filter output solutions at time k, respectively, S3-3-3 When the online learning meets the threshold for forecasting, classify according to whether the satellite signal is occluded or not; S3-3-3-1 When the GNSS satellite signal is blocked, the INS can work normally due to its properties; the GNSS data at the moment before the failure is the initial value of the filter, which is equivalent to the accurate value, so it can avoid the INS accuracy due to navigation Divergence caused by long time; when the GNSS signal is lost, the trained neural network model is used to predict the output error of inertial navigation, and the prediction error is used to compensate and correct the inertial navigation.

As Elman neural network for prediction, the output neural network prediction data is

and

EKF update as observation constraint, Neural network predicts the amount of change

and

The obtained predicted values of the position, rotation quaternion and velocity at time k+1 are:

The constraints are updated to:

in,

and

are respectively the state vector outputted by the inertial sensor in step S3-1;

and

are the predicted values of the position, rotation quaternion and velocity at time k+1 obtained by the neural network, respectively, S3-3-3-2 When the GNSS signal is normal, there are two strategies available: 1. Directly use the position, velocity and rotation quaternion at the moment of k+1 output by the inertial sensor as the state vector update, and then use the GNSS output 2. Input the position, velocity and rotation quaternion variation of inertial sensor output k+1 time into Elman neural network, and input the output of Elman neural network into Elman neural network. The position, rotation quaternion and velocity change are used as pseudo-observed values to update the state vector of the EKF, as in formula (16); then the constraint update is performed according to the position and velocity output by GNSS, as in formula (17); requested

and

and obtained using EKF

and

Join the online learning system of the neural network for real-time learning and training, EKF status updated to:

The GNSS measurement status in EKF is updated to:

and

are the predicted changes based on the output neural network

and

The obtained predicted values of position, rotation quaternion and velocity at time k+1; ^e P _k+1,G and ^e V _k+1,G are the output values of GNSS position and velocity, respectively.

Images