High-dynamic aircraft inertia / satellite integrated navigation method

Abstract

The invention discloses a high-dynamic aircraft inertia / satellite integrated navigation method, relates to the technical field of integrated navigation, and aims to solve the problems that in a high-dynamic environment, a GPS signal of a carrier is extremely easy to lose lock or even lose, navigation errors of an INS (inertial navigation system) are accumulated and increased along with time, and the navigation precision of the carrier is seriously influenced. The position and the speed of the carrier are obtained, and high-precision navigation of the carrier can be realized during the interruption of a GPS signal in a high-dynamic environment. When the GPS signal is normally received, training the neural network by using the GPS signal; when the GPS signal is lost, the trained neural network is used to predict the position and speed of the carrier, and the output of the neural network is used to replace the GPS signal as the input of the Kalman filter. The navigation problem of the carrier when the GPS signal is lost in a high-dynamic environment is effectively solved, the navigation error of the INS is avoided, and the navigation precision of the carrier is improved.

Description

Technical Field

[0001] The invention relates to the technical field of combined navigation, and in particular to an inertial / satellite combined navigation method for a high-dynamic aircraft. Background Art

[0002] The combined navigation system established by the Global Positioning System (GPS) and the Inertial Navigation System (INS) is the most common combined navigation mode in engineering. This method effectively utilizes the advantages of high GPS accuracy and strong INS autonomy. However, in a high-dynamic environment, the GPS signal of the carrier is very easy to lose lock or even lose. At this time, the navigation error of INS will accumulate and increase over time, seriously affecting the navigation accuracy of the carrier. How to achieve high-precision navigation when the GPS signal is lost in a high-dynamic environment is a problem that technicians in this field urgently need to solve. Summary of the invention

[0003] The purpose of the present invention is to propose a high-dynamic aircraft inertial / satellite combined navigation method to solve the problem that in a high-dynamic environment, the GPS signal of the carrier is very easy to lose lock or even be lost. At this time, the navigation error of INS will accumulate and increase over time, seriously affecting the navigation accuracy of the carrier.

[0004] The technical solution adopted by the present invention to solve the above technical problems is:

[0005] A high-dynamic aircraft inertial / satellite integrated navigation method comprises the following steps:

[0006] Step 1: construct an INS / GPS integrated navigation model, wherein the INS / GPS integrated navigation model includes a state equation and a measurement equation;

[0007] Step 2: Obtain the inertial information measured by INS, and use the inertial information and state equation to calculate the error state variable X k Update the time and get the predicted value X of the current state kk-1 and the state error covariance P k|k-1 ;

[0008] Step 3: Determine whether GPS is available. If GPS is available, obtain the current GPS measurement, i.e., the three-directional position and speed information of the carrier. If GPS is not available, obtain the GPS measurement at the previous moment, and input the GPS measurement at the previous moment into the pre-trained network model to obtain the predicted measurement, and use the predicted measurement as the GPS measurement at the current moment.

[0009] Step 4: Use the current GPS measurement and combine it with the measurement equation to get the measurement Z k, then based on the quantity measurement Z k , the predicted value of the current state X kk-1 and the state error covariance P k|k-1 Estimate the current state and get the estimated state value X k and the error covariance estimate P k , and then the state estimate X k and the error covariance estimate P k Input the inertial navigation solution model and obtain the output attitude, velocity and position information.

[0010] Further, the state equation is expressed as:

[0011]

[0012] X k =[δα,Δp,Δv,Δb g ,Δb a ] T

[0013]

[0014]

[0015]

[0016] in, For X k The first derivative of k is the state quantity of the combined navigation filter, F is the system matrix, G is the noise input matrix, w is the zero-mean system noise, δα is the attitude error angle of the carrier, Δp is the position error of the carrier, Δv is the velocity error of the carrier, Δb g is the gyro constant drift error of the carrier, Δb a is the constant drift error of the carrier's accelerometer, Representation calculation The antisymmetric matrix of is the direction cosine matrix between the carrier system and the launching inertial system, is the specific force obtained after correction of the accelerometer, GM is the earth's gravitational constant, is the distance from the carrier to the center of the earth, x, y, z are the position coordinate components of the carrier in the inertial system, and R0 is the average radius of the earth.

[0017] Furthermore, the measurement equation is expressed as:

[0018] Z k =H k X k +V GPS

[0019]

[0020] Among them, Z k For quantity measurement, H k is the intermediate variable, V GPS is the measurement noise of the GPS system.

[0021] Furthermore, the predicted value X of the current state kk-1 It is expressed as:

[0022] X kk-1 =Φ k|k-1 X k-1

[0023] Among them, Φ k|k-1 is the state transfer matrix, T represents the sampling time interval, and I represents the identity matrix.

[0024] Furthermore, the state error covariance P k|k-1 It is expressed as:

[0025]

[0026]

[0027]

[0028]

[0029]

[0030] Θ k =P kk-1 -Q

[0031]

[0032]

[0033] Among them, M k is the intermediate variable, P k-1 is the state error covariance of the previous moment, Γ k|k-1 is the noise driving matrix, Q is an intermediate variable, diag means constructing a diagonal matrix, are multiple time-varying adaptive factors, κ≥1 is the adjustment factor, T k ,N k ,Θ k are the noise-related variables in the design process, are the random drift noise variances of the gyroscope and accelerometer, respectively, and C k is the intermediate variable, δZk is the new information, α is the forgetting parameter, and k is the kth moment.

[0034] Furthermore, the state estimation value X k It is expressed as:

[0035] X k =X kk-1 +K k δZ k

[0036]

[0037]

[0038]

[0039]

[0040] Among them, X k is the estimated value of the state at time k, K k is the Kalman filter gain matrix, is the actual measured value, S k is the innovation covariance matrix, R k is the measurement noise variance.

[0041] Furthermore, the error covariance estimate P k It is expressed as:

[0042] P k =P kk-1 -K k H k P kk-1

[0043] Among them, P k is the error covariance estimate at time k.

[0044] Furthermore, the inertial navigation solution model includes an update equation for the attitude in the launch inertial system, an update equation for the velocity in the launch inertial system, and an update equation for the position of the velocity in the launch inertial system;

[0045] The update equation of the attitude in the launch inertial system is expressed as:

[0046]

[0047] q=[q1 q2 q3 q4] T

[0048]

[0049] in, is the first-order derivative of q, q is an intermediate variable, is the carrier angular velocity measured by the gyroscope, q1, q2, q3 and q4 are the carrier attitude quaternions;

[0050] The update equation of the velocity in the launch inertial system is expressed as:

[0051]

[0052] in, is the first-order derivative of V, V is the velocity vector of the carrier in the inertial system, V = [v x v y v z ] T , v x 、v y and v z are the x, y, and z axis components of the velocity in the launching inertial system, g is the gravitational acceleration, and f b is the specific force measured by the accelerometer;

[0053] The position update equation of the velocity in the launch inertial system is expressed as:

[0054]

[0055] in, is the first-order derivative of x, is the first-order derivative of y, is the first-order derivative of z, v x 、v y and v z is the position vector of the carrier in the inertia system.

[0056] Furthermore, the pre-trained network model is Transformer-LSTM.

[0057] Furthermore, the pre-trained network model is an improved Transformer-LSTM, and the improved Transformer-LSTM includes Transformer-LSTM and a self-attention mechanism;

[0058] The hidden layer state output h of the Transformer-LSTM k It is expressed as:

[0059] h k =o k εtanh(c k )

[0060]

[0061] ok =sigmoid(w ox x k +w oh h k-1 +b o )

[0062] f k =sigmoid(w fx x k +w fh h k-1 +b f )

[0063] i k =sigmoid(w ix x k +w ih h k-1 +b i )

[0064]

[0065] Among them, c k is the memory unit, i k is the input gate, f k For the forget gate, k is the output gate, c k-1 is the memory unit of the previous moment, w fx and w fh The input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the forget gate, b f is the bias coefficient of the forget gate, Update the value of the memory cell, w ix and w ih The input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the input gate, b i is the bias coefficient of the input gate, w cx and w ch The input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the memory unit, b c is the bias coefficient of the memory unit, w ox and w oh The input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the output gate, b o is the bias coefficient of the output gate;

[0066] Output the hidden layer state h kAs the input of the self-attention mechanism, the prediction result is obtained;

[0067] The self-attention mechanism is expressed as:

[0068]

[0069] Among them, Q, K, and V represent Query, Key, Value, and d respectively. k Indicates the dimension of the Key.

[0070] The beneficial effects of the present invention are:

[0071] This application predicts the measurements of the GPS network, namely the position and speed of the carrier, through a neural network, and can achieve high-precision navigation of the carrier during GPS signal interruption in a high-dynamic environment. When the GPS signal is received normally, the neural network is trained using the GPS signal; when the GPS signal is lost, the position and speed of the carrier are predicted using the trained neural network, and its output replaces the GPS signal as the input of the Kalman filter. This effectively solves the navigation problem of the carrier when the GPS signal is lost in a high-dynamic environment, thereby avoiding the navigation error of the INS and improving the navigation accuracy of the carrier. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is a schematic diagram of a high dynamic integrated navigation system when GPS is effective;

[0073] Figure 2 It is a high dynamic integrated navigation system when GPS is lost;

[0074] Figure 3 This is a flow chart of an embodiment of the present application;

[0075] Figure 4 It is the flight trajectory diagram of the carrier aircraft;

[0076] Figure 5 This is the longitude settlement error diagram;

[0077] Figure 6 This is the latitude settlement error map. DETAILED DESCRIPTION

[0078] It should be particularly noted that, in the absence of conflict, the various embodiments disclosed in this application can be combined with each other.

[0079] Specific implementation method 1: A high-dynamic aircraft inertial / satellite integrated navigation method described in this implementation method includes the following steps:

[0080] Step 1: construct an INS / GPS integrated navigation model, wherein the INS / GPS integrated navigation model includes a state equation and a measurement equation;

[0081] Step 2: Obtain the inertial information measured by INS, and use the inertial information and state equation to calculate the error state variable X k Update the time and get the predicted value X of the current state kk-1 and the state error covariance P k|k-1 ;

[0082] Step 3: Determine whether GPS is available. If GPS is available, obtain the current GPS measurement, i.e., the three-directional position and speed information of the carrier. If GPS is not available, obtain the GPS measurement at the previous moment, and input the GPS measurement at the previous moment into the pre-trained network model to obtain the predicted measurement, and use the predicted measurement as the GPS measurement at the current moment.

[0083] Step 4: Use the current GPS measurement and combine it with the measurement equation to get the measurement Z k , then based on the quantity measurement Z k , the predicted value of the current state X kk-1 and the state error covariance P k|k-1 Estimate the current state and get the estimated state value X k and the error covariance estimate P k , and then the state estimate X k and the error covariance estimate P k Input the inertial navigation solution model and obtain the output attitude, velocity and position information.

[0084] Specific implementation method 2: This implementation method is a further description of the specific implementation method 1. The difference between this implementation method and the specific implementation method 1 is that the state equation is expressed as:

[0085]

[0086] X k =[δα,Δp,Δv,Δb g ,Δb a ] T

[0087]

[0088]

[0089]

[0090] in, For X k The first derivative of kis the state quantity of the combined navigation filter, F is the system matrix, G is the noise input matrix, w is the zero-mean system noise, δα is the attitude error angle of the carrier, Δp is the position error of the carrier, Δv is the velocity error of the carrier, Δb g is the gyro constant drift error of the carrier, Δb a is the constant drift error of the carrier's accelerometer, Representation calculation The antisymmetric matrix of is the direction cosine matrix between the carrier system and the launching inertial system, is the specific force obtained after correction of the accelerometer, GM is the earth's gravitational constant, is the distance from the carrier to the center of the earth, x, y, z are the position coordinate components of the carrier in the inertial system, and R0 is the average radius of the earth.

[0091] Specific implementation method three: This implementation method is a further description of specific implementation method two. The difference between this implementation method and specific implementation method two is that the measurement equation is expressed as:

[0092] Z k =H k X k +V GPS

[0093]

[0094] Among them, Z k For quantity measurement, H k is the intermediate variable, V GPS is the measurement noise of the GPS system.

[0095] Specific implementation method 4: This implementation method is a further explanation of specific implementation method 3. The difference between this implementation method and specific implementation method 3 is that the predicted value X of the current state kk-1 It is expressed as:

[0096] X kk-1 =Φ k|k-1 X k-1

[0097] Among them, Φ k|k-1 is the state transfer matrix, T represents the sampling time interval, and I represents the identity matrix.

[0098] Specific implementation mode 5: This implementation mode is a further explanation of the specific implementation mode 4. The difference between this implementation mode and the specific implementation mode 4 is that the state error covariance P k|k-1 It is expressed as:

[0099]

[0100]

[0101]

[0102]

[0103]

[0104] Θ k =P kk-1 -Q

[0105]

[0106]

[0107] Among them, M k is the intermediate variable, P k-1 is the state error covariance of the previous moment, Γ k|k-1 is the noise driving matrix, Q is an intermediate variable, diag means constructing a diagonal matrix, are multiple time-varying adaptive factors, κ≥1 is the adjustment factor, T k ,N k ,Θ k are the noise-related variables in the design process, are the random drift noise variances of the gyroscope and accelerometer, respectively, and C k is the intermediate variable, δZ k is the new information, α is the forgetting parameter, and k is the kth moment.

[0108] Specific implementation method 6: This implementation method is a further explanation of the specific implementation method 5. The difference between this implementation method and the specific implementation method 5 is that the state estimation value X k It is expressed as:

[0109] X k =X kk-1 +K k δZ k

[0110]

[0111]

[0112]

[0113]

[0114] Among them, X k is the estimated value of the state at time k, K kis the Kalman filter gain matrix, is the actual measured value, S k is the innovation covariance matrix, R k is the measurement noise variance.

[0115] Specific implementation method 7: This implementation method is a further explanation of specific implementation method 6. The difference between this implementation method and specific implementation method 6 is that the error covariance estimation value P k It is expressed as:

[0116] P k =P kk-1 -K k H k P kk-1

[0117] Among them, P k is the error covariance estimate at time k.

[0118] Specific implementation eight: This implementation is a further explanation of specific implementation seven. The difference between this implementation and specific implementation seven is that the inertial navigation solution model includes an update equation for the attitude in the launch inertial system, an update equation for the velocity in the launch inertial system, and an update equation for the position of the velocity in the launch inertial system;

[0119] The update equation of the attitude in the launch inertial system is expressed as:

[0120]

[0121] q=[q1 q2 q3 q4] T

[0122]

[0123] in, is the first-order derivative of q, q is an intermediate variable, is the carrier angular velocity measured by the gyroscope, q1, q2, q3 and q4 are the carrier attitude quaternions;

[0124] The update equation of the velocity in the launch inertial system is expressed as:

[0125]

[0126] in, is the first-order derivative of V, V is the velocity vector of the carrier in the inertial system, v y and v z are the x, y, and z axis components of the velocity in the launching inertial system, g is the gravitational acceleration, and f b is the specific force measured by the accelerometer;

[0127] The position update equation of the velocity in the launch inertial system is expressed as:

[0128]

[0129] in, is the first-order derivative of x, is the first-order derivative of y, is the first-order derivative of z, v x 、v y and v z is the position vector of the carrier in the inertia system.

[0130] Specific implementation method nine: This implementation method is a further explanation of specific implementation method one. The difference between this implementation method and specific implementation method one is that the pre-trained network model is Transformer-LSTM.

[0131] Specific implementation method ten: This implementation method is a further explanation of specific implementation method one. The difference between this implementation method and specific implementation method one is that the pre-trained network model is an improved Transformer-LSTM, and the improved Transformer-LSTM includes Transformer-LSTM and self-attention mechanism;

[0132] The hidden layer state output h of the Transformer-LSTM k It is expressed as:

[0133] h k =o k εtanh(c k )

[0134]

[0135] o k =sigmoid(w ox x k +w oh h k-1 +b o )

[0136] f k =sigmoid(w f xx k +w fh h k-1 +b f )

[0137] i k =sigmoid(w ix x k +w ih hk-1 +b i )

[0138]

[0139] Among them, c k is the memory unit, i k is the input gate, f k For the forget gate, k is the output gate, c k-1 is the memory unit of the previous moment, w fx and w fh The input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the forget gate, b f is the bias coefficient of the forget gate, Update the value of the memory cell, w ix and w ih The input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the input gate, b i is the bias coefficient of the input gate, w cx and w ch The input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the memory unit, b c is the bias coefficient of the memory unit, w ox and w oh The input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the output gate, b o is the bias coefficient of the output gate;

[0140] Output the hidden layer state h k As the input of the self-attention mechanism, the prediction result is obtained;

[0141] The self-attention mechanism is expressed as:

[0142]

[0143] Among them, Q, K, and V represent Query, Key, Value, and d respectively. k Indicates the dimension of the Key.

[0144] At present, in order to improve the navigation accuracy of the carrier when the GPS signal is unavailable, a common method is to add additional sensors, such as terrain assistance, visual odometer, etc., to correct the accumulated error of INS, but this method increases the cost of the integrated navigation system. Another method is to use time series to update the state, such as autoregression, recurrent neural network and LSTM. Recently, the self-attention mechanism has been widely used to predict time series data. The self-attention mechanism can more effectively handle this type of sequence data prediction problem by establishing global dependencies. This application proposes a high-dynamic integrated navigation method based on neural network assistance. First, an adaptive Kalman filter is used to reduce the impact of noise on the system. At the same time, a learning framework based on a multi-head self-attention mechanism and a long short-term memory network (Long Short Term Memory Network, LSTM) is proposed for predicting the position and speed of the carrier. Combining the adaptive Kalman filter with the Transformer-LSTM can achieve high-precision navigation of the carrier during the interruption of the GPS signal in a high-dynamic environment. When the GPS signal is received normally, the neural network is trained using the GPS signal; when the GPS signal is lost, the position and speed of the carrier are predicted using the trained neural network, and its output replaces the GPS signal as the input of the Kalman filter. It effectively solves the navigation problem of the carrier when the GPS signal is lost in a high dynamic environment.

[0145] S1. Constructing the inertial navigation solution model

[0146] The strapdown inertial navigation system has three orthogonally mounted gyroscopes and accelerometers. By measuring the carrier's angular velocity and acceleration, the position, velocity and attitude information of the carrier are calculated. Let q be the carrier's attitude quaternion,

[0147]

[0148] Then the update equation of the attitude in the launch inertial system is:

[0149]

[0150] In the formula, is the carrier angular velocity measured by the gyroscope,

[0151]

[0152] The update equation of the velocity in the launch inertial system is:

[0153]

[0154] In the formula, V = [v x v y v z ]T is the velocity vector of the carrier in the emission-inertial system, g is the gravitational acceleration, and the direction cosine matrix between this system and the emission-inertial system is It can be obtained by converting the attitude quaternion, f b is the specific force measured by the accelerometer.

[0155] At the same time, the position update equation under the inertial system is:

[0156]

[0157] In the formula, x, y, z are the position vectors of the carrier in the inertial system. By integrating equations (1.2), (1.4), and (1.5), the real-time position, velocity, and attitude information of the carrier can be obtained by using inertial navigation.

[0158] S2. Construct INS / GPS integrated navigation model, including state equation and measurement equation;

[0159] Consider the following integrated navigation system model in the launch inertial system, select the carrier's attitude error angle δα, position error Δp, velocity error Δv, gyro constant drift error Δb g and the accelerometer constant drift error Δb a As the state quantity X of the integrated navigation filter k ,Right now:

[0160]

[0161] Among them, the superscript T represents the transpose of the matrix, and k is the time series.

[0162] Then the state equation obtained according to the INS error propagation equation is:

[0163]

[0164] Where w is the system noise with zero mean, and the variance of w is:

[0165]

[0166] in, are the random drift noise variances of the gyroscope and accelerometer, respectively, and I represents the identity matrix.

[0167] The system matrix F and the noise input matrix G are:

[0168]

[0169]

[0170] In the formula, is the direction cosine matrix between the carrier system and the launching inertial system, is the specific force obtained after correction of the accelerometer, Representation calculation The antisymmetric matrix of

[0171]

[0172] Among them, x, y, z are the position coordinate components of the carrier in the inertial system, R0 is the average radius of the earth, is the distance from the carrier to the center of the earth, and GM is the gravitational constant of the earth.

[0173] The difference between the position and speed information output by GPS and the position and speed information output by INS is taken as the quantity measurement, that is:

[0174]

[0175] Among them, x GPS ,y GPS ,z GPS and v xGPS ,v yGPS ,v zGPS They represent the three-directional position and speed information of the carrier output by GPS, respectively. INS ,y INS ,z INS and v xINS ,v yINS ,v zINS Respectively represent the carrier position and velocity information output by the inertial navigation system, V GPS is the measurement noise of the GPS system.

[0176] Then the measurement equation of the integrated navigation system is:

[0177] Z k =H k X k +V GPS (1.13)

[0178] in,

[0179]

[0180] S3. Adaptive Kalman Navigation Filter Settlement

[0181] Step 3.1: Time update. First, based on the inertial information measured by INS (including the angular velocity measured by the gyroscope, the direction cosine matrix between the carrier system and the launch inertial system, and the specific force obtained by the accelerometer after correction) and the state equation, the error state variable X k Update the time, assuming the state estimate at the previous moment is Xk-1 , then the one-step prediction value X of the current state kk-1 for:

[0182] X kk-1 =Φ k|k-1 X k-1 (1.15)

[0183] in, is the state transfer matrix, and T represents the sampling time interval.

[0184] At the same time, the one-step prediction P of the state error covariance is calculated k|k-1 ,

[0185]

[0186] in, is the noise driving matrix, P k-1 is the state error covariance at the previous moment.

[0187] Step 3.2: Measurement update. After obtaining the measurement input of the GPS at the current moment (the output of the carrier's three-directional position and speed information), combined with the measurement equation, the measurement Z is obtained. k , and combined with X kk-1 and P kk-1 Estimate the current state.

[0188] X k =X kk-1 +K k δZ k (1.17)

[0189] P k =P kk-1 -K k H k P kk-1 (1.18)

[0190] Among them, X k ,P k is the estimated value of the state and error covariance at time k, K k is the Kalman filter gain matrix, δZ k is the new interest, which is calculated as follows:

[0191]

[0192]

[0193] In the formula, is the actual measured value, S k is the innovation covariance matrix,

[0194]

[0195] Step 3.3: Adaptive adjustment of noise covariance.

[0196] The performance of traditional Kalman filtering depends on the accuracy of noise covariance, but in a high dynamic environment, the vibration of the carrier itself and the drastic changes in the external environment will cause the prior information of the noise covariance to become inaccurate, thus affecting the navigation accuracy of the filter. Therefore, this application introduces an adaptive mechanism to dynamically adjust the noise covariance matrix. The main idea is to introduce multiple time-varying adaptive factors to adaptively change the gain matrix and the state error covariance matrix. The specific method is as follows:

[0197] When the system noise changes dramatically, the estimated error of the state should also increase. Based on this idea, an adaptive factor is introduced to adjust the state error covariance adaptively. We introduce the adaptive matrix M k

[0198]

[0199] In the formula, diag means constructing a diagonal matrix, If are multiple time-varying adaptive factors, the one-step prediction of the new adaptive error covariance is rewritten as:

[0200]

[0201] Since a single adaptive factor has a poor effect in multi-dimensional state estimation, M k The adaptive factors in are not equal. According to orthogonality, we have:

[0202]

[0203] in,

[0204]

[0205] In the formula, α is the forgetting parameter.

[0206] The adaptive factors designed in this application are selected as follows:

[0207]

[0208] Where, T k ,N k ,Θ k are the noise-related variables in the design process,

[0209] T k =H k N k Hk T (1.26)

[0210]

[0211] Θ k =P kk-1 -Q(1.28)

[0212] In order to make the state estimation parameters smoother, the adjustment factor can be used to reduce the interference of measurement noise, then equation (1.27) can be rewritten as:

[0213]

[0214] In the formula, κ≥1 is the adjustment factor.

[0215] S4. Transformer-LSTM learning framework

[0216] Although LSTM can effectively solve the problems of gradient vanishing and gradient exploding in traditional recurrent neural networks, a single LSTM model may not be able to guarantee prediction accuracy in a highly dynamic environment. Therefore, this patent introduces the Transformer-LSTM learning framework to enhance the capture of long-term dependencies and dynamically allocate weights to suppress error propagation.

[0217] Step 3.1: When GPS is available, use GPS data for network training. When GPS data is available, it is input into the Transformer-LSTM network for training while performing integrated navigation settlement.

[0218] The core part of LSTM is the memory unit c k And three gating units: input gate i k 、Forget gate k AND output gate o k . Three gate control units i k ,f k ,o k They are all output vectors of the neural network layer after the sigmoid activation function.

[0219]

[0220] Forget Gate k It is used to control the memory unit c k-1 Output, it determines the memory unit c at the previous moment k-1 What information has been forgotten?

[0221] f k =sigmoid(w fx x k+w fh h k-1 +b f )(1.31)

[0222] In the formula, w fx and w fh is the input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the forget gate, b f is the bias coefficient of the forget gate.

[0223] Input Gate i k It is used to control the memory unit update value The output of Which features can be used to update c k .

[0224] i k =sigmoid(w ix x k +w ih h k-1 +b i ) (1.32)

[0225]

[0226] In the formula, w ix and w ih is the input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the input gate, b i is the bias coefficient of the input gate, w cx and w ch is the input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the memory unit, b c is the bias coefficient of the memory unit. The memory unit update value c k is the output vector of the neural network layer after the tanh activation function, which is:

[0227] tanh(x)=(e x -e -x ) / (e x +e -x ) (1.34)

[0228] At the same time, the new memory unit is updated as:

[0229]

[0230] Output gate o k It is used to control the memory unit c kWhether the features are retained for the hidden layer output h k .

[0231] o k =sigmoid(w ox x k +w oh h k-1 +b o ) (1.36)

[0232] h k =o k εtanh(c k ) (1.37)

[0233] In the formula, w ox and w oh is the input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the output gate, b o is the bias coefficient of the output gate.

[0234] At the same time, the network introduces a self-attention mechanism to capture global dependencies in the sequence. The self-attention mechanism adjusts the representation of different positions by calculating the similarity between Query, Key, and Value. The higher the similarity, the stronger the dependency between different positions, and the network will pay more attention to the relevant information of these positions. In this process, the attention weight of each position is first calculated, then normalized using softmax, and the corresponding values ​​are weighted summed as follows:

[0235]

[0236] In the formula, Q, K, and V represent Query, Key, Value, and k Indicates the dimension of the Key.

[0237] Step 3.2: When the GPS signal is lost, the network outputs prediction information for integrated navigation. When the GPS signal is lost, the trained Transformer-LSTM network predicts the GPS signal at the time of loss and outputs it to the integrated navigation filter as the measurement data when the GPS signal is lost, thereby ensuring that integrated navigation can still be performed when the GPS signal is lost and correcting the accumulated error of the INS.

[0238] The algorithm flow chart is as follows Figure 3 shown.

[0239] Mathematical simulation

[0240] To verify the effectiveness of this application, the following navigation scenario is constructed for simulation. It is assumed that the carrier is at 100° longitude, 50° latitude, and 5km altitude at the initial moment. The accelerometer has a constant drift of 1mg, a random drift of 25mg, a gyroscope has a constant drift of 20° / h, and a random drift of The simulation duration is 100s and the step length is 0.1s. The simulation scenario is as follows Figure 4 shown.

[0241] The position error of GPS navigation is 10m, the speed error is 2.5m / s, and the GPS signal is lost after 50s. The tracking simulation is performed using adaptive extended Kalman filter (AEKF), LSTM-based AEKF method and the algorithm proposed in this patent (The proposed filter).

[0242] Tracking results comparison Figure 5 and Figure 6 shown.

[0243] It can be seen that the traditional Kalman filter cannot handle the situation where the GPS signal is lost, so its estimation accuracy gradually diverges after the GPS signal is lost; although the LSTM-based Kalman filter can handle the situation where the GPS signal is lost, the Transformer-LSTM method proposed in this patent has a higher prediction accuracy for GPS signals due to the introduction of the self-attention mechanism, so its combined navigation accuracy is also higher. The method proposed in the present invention effectively improves the navigation accuracy of the carrier when the GPS signal is lost in a high dynamic environment.

[0244] It should be noted that the specific implementation is only an explanation and description of the technical solution of the present invention, and cannot be used to limit the scope of protection of the rights. Any partial changes made according to the claims and description of the present invention should still fall within the scope of protection of the present invention.

Claims

1. A high dynamic aircraft inertial / satellite combined navigation method, characterized in that The following steps are involved: Step 1: construct an INS / GPS integrated navigation model, wherein the INS / GPS integrated navigation model includes a state equation and a measurement equation; Step 2: Obtain the inertial information measured by INS, and use the inertial information and state equation to calculate the error state variable X k Update the time and get the predicted value X of the current state kk-1 and the state error covariance P k|k-1 ; Step 3: Determine whether GPS is available. If GPS is available, obtain the current GPS measurement, i.e., the three-directional position and speed information of the carrier. If GPS is not available, obtain the GPS measurement at the previous moment, and input the GPS measurement at the previous moment into the pre-trained network model to obtain the predicted measurement, and use the predicted measurement as the GPS measurement at the current moment. Step 4: Use the current GPS measurement and combine it with the measurement equation to get the measurement Z k , then based on the quantity measurement Z k , the predicted value of the current state X kk-1 and the state error covariance P k|k-1 Estimate the current state and get the estimated state value X k and the error covariance estimate P k , and then the state estimate X k and the error covariance estimate P k Input the inertial navigation solution model and obtain the output attitude, velocity and position information.

2. A high dynamic aircraft inertial / satellite integrated navigation method according to claim 1, characterized in that The state equation is expressed as: X k =[δα,Δp,Δv,Δb g ,Δb a ] T in, For X k The first derivative of k is the state quantity of the combined navigation filter, F is the system matrix, G is the noise input matrix, w is the zero-mean system noise, δα is the attitude error angle of the carrier, Δp is the position error of the carrier, Δv is the velocity error of the carrier, Δb g is the gyro constant drift error of the carrier, Δb a is the constant drift error of the carrier's accelerometer, Representation calculation The antisymmetric matrix of is the direction cosine matrix between the carrier system and the launching inertial system, is the specific force obtained after correction of the accelerometer, GM is the earth's gravitational constant, is the distance from the carrier to the center of the earth, x, y, z are the position coordinate components of the carrier in the inertial system, and R0 is the average radius of the earth.

3. A high dynamic aircraft inertial / satellite integrated navigation method according to claim 2, characterized in that The measurement equation is expressed as: Z k =H k X k +V GPS Among them, Z k For quantity measurement, H k is the intermediate variable, V GPS is the measurement noise of the GPS system.

4. A high dynamic aircraft inertial / satellite integrated navigation method according to claim 3, characterized in that The predicted value X of the current state kk-1 It is expressed as: X kk-1 =Φ k|k-1 X k-1 Among them, Φ k|k-1 is the state transfer matrix, T represents the sampling time interval, and I represents the identity matrix.

5. A high dynamic aircraft inertial / satellite integrated navigation method according to claim 4, characterized in that The state error covariance P k|k-1 It is expressed as: M k =diag[θ1,θ2,K,θ n ] Θ k =P kk-1 -Q Among them, M k is the intermediate variable, P k-1 is the state error covariance of the previous moment, Γ k|k-1 is the noise driving matrix, Q is the intermediate variable, diag means to construct a diagonal matrix, θ n are multiple time-varying adaptive factors, κ≥1 is the adjustment factor, T k ,N k ,Θ k are the noise-related variables in the design process, are the random drift noise variances of the gyroscope and accelerometer, respectively, and C k is the intermediate variable, δZ k is the new information, α is the forgetting parameter, and k is the kth moment.

6. A high dynamic aircraft inertial / satellite integrated navigation method according to claim 5, characterized in that The state estimate X k It is expressed as: X k =X kk-1 +K k δZ k Among them, X k is the estimated value of the state at time k, K k is the Kalman filter gain matrix, is the actual measured value, S k is the innovation covariance matrix, R k is the measurement noise variance.

7. A high dynamic aircraft inertial / satellite integrated navigation method according to claim 6, characterized in that The error covariance estimate P k It is expressed as: P k =P kk-1 -K k H k P kk-1 Among them, P k is the error covariance estimate at time k.

8. A high dynamic aircraft inertial / satellite integrated navigation method according to claim 7, characterized in that The inertial navigation solution model includes an update equation for the attitude in the launch inertial system, an update equation for the velocity in the launch inertial system, and an update equation for the position of the velocity in the launch inertial system; The update equation of the attitude in the launch inertial system is expressed as: <h2 style=";text-align:left;direction:ltr">q=[q1 q2 q3 q4]<h2 style=";text-align:left;direction:ltr"> T in, is the first-order derivative of q, q is an intermediate variable, is the carrier angular velocity measured by the gyroscope, q1, q2, q3 and q4 are the carrier attitude quaternions; The update equation of the velocity in the launch inertial system is expressed as: in, is the first-order derivative of V, V is the velocity vector of the carrier in the inertial system, V = [v x v y v z ] T , v x 、v y and v z are the x, y, and z axis components of the velocity in the launching inertial system, g is the gravitational acceleration, and f b is the specific force measured by the accelerometer; The position update equation of the velocity in the launch inertial system is expressed as: in, is the first-order derivative of x, is the first-order derivative of y, is the first-order derivative of z, v x 、v y and v z is the position vector of the carrier in the inertia system.

9. A high dynamic aircraft inertial / satellite integrated navigation method according to claim 1, characterized in that The pre-trained network model is Transformer-LSTM.

10. The high dynamic aircraft inertial / satellite integrated navigation method according to claim 1, characterized in that The pre-trained network model is an improved Transformer-LSTM, and the improved Transformer-LSTM includes Transformer-LSTM and a self-attention mechanism; The hidden layer state output h of the Transformer-LSTM k It is expressed as: h k =o k εtanh(c k ) o k =sigmoid(w ox x k +w oh h k-1 +b o ) f k =sigmoid(w fx x k +w fh h k-1 +b f ) i k =sigmoid(w ix x k +w ih h k-1 +b i ) Among them, c k is the memory unit, i k is the input gate, f k For the forget gate, k is the output gate, c k-1 is the memory unit of the previous moment, w fx and w fh The input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the forget gate, b f is the bias coefficient of the forget gate, Update the value of the memory cell, w ix and w ih The input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the input gate, b i is the bias coefficient of the input gate, w cx and w ch The input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the memory unit, b c is the bias coefficient of the memory unit, w ox and w oh The input vector x k and the previous hidden layer state h k-1 Relative to the weight coefficient of the output gate, b o is the bias coefficient of the output gate; Output the hidden layer state h k As the input of the self-attention mechanism, the prediction result is obtained; The self-attention mechanism is expressed as: Among them, Q, K, and V represent Query, Key, Value, and d respectively. k Indicates the dimension of the Key.

Images