Inertial missile trajectory parameter measurement method based on Conv-LSTM and Kalman filter fusion

Abstract

The present invention relates to a method for measuring inertial ballistic parameters based on the fusion of Conv-LSTM and Kalman filtering, and belongs to the field of ammunition inertial measurement technology. The method of the present invention can achieve high-precision real-time trajectory measurement of artillery-launched intelligent ammunition in complex electromagnetic environments or when satellite positioning fails due to enemy interference or deception. The present invention is primarily aimed at a type of intelligent ammunition that relies on inertial / satellite combined signals for trajectory measurement. It can compensate for the problem of measurement errors accumulating over time that occur in traditional combined navigation algorithms when satellite positioning fails, and has high accuracy.

Description

Technical Field

[0001] The present invention relates to an inertial missile trajectory parameter measurement method based on Conv-LSTM and Kalman filtering fusion, belonging to the technical field of ammunition inertial measurement. Background Art

[0002] Accurate trajectory measurement is a crucial prerequisite for flight control and precision strikes of artillery-launched intelligent munitions. Satellite positioning systems are currently the mainstream method for trajectory measurement of intelligent munitions. However, satellite signals are susceptible to electromagnetic interference, severely reducing their availability under intense warfare conditions. Developing trajectory measurement methods for satellite-denied environments is crucial for improving the survivability of intelligent munitions. Three-axis geomagnetic sensors can be used to measure projectile attitude, but due to the weak geomagnetic field signal, they are highly susceptible to interference from strong magnetic environments. Radio beacon technology is widely used for aircraft navigation in the civilian sector. However, in actual battlefield environments, the number, range, and terrain of beacons are limited, often preventing them from providing complete, high-precision three-dimensional measurement information over a long period of time, resulting in poor observability when estimating projectile attitude. In recent years, the application of low-cost, compact, and overload-resistant micro-electromechanical inertial sensors in intelligent munitions has become increasingly feasible. However, errors in low-cost inertial navigation systems accumulate over time, often requiring reliance on other auxiliary navigation systems. Summary of the Invention

[0003] The technical problem solved by the present invention is: to overcome the shortcomings of the existing technology, and to propose an inertial ballistic parameter measurement method based on the fusion of Conv-LSTM and Kalman filter. The method is an artificial intelligence-based projectile trajectory estimation method in a satellite-denied environment that only uses micro-electromechanical inertial sensors and pre-flight parameters unique to ammunition. The method is intended to solve the problem of measurement errors accumulating over time in traditional inertial navigation systems in a satellite-denied environment, and can improve the accuracy of ballistic measurement in a satellite-denied environment.

[0004] The technical solution of the present invention is:

[0005] A method for measuring inertial ballistic parameters based on the fusion of Conv-LSTM and Kalman filtering, the method comprising the following steps:

[0006] The first step is to establish a training data set, which includes basic trajectory parameters and simulated noise output data of inertial sensors;

[0007] The basic trajectory parameters include the trajectory angle A and the trajectory initial position, trajectory initial velocity, trajectory initial posture, trajectory position at each moment, trajectory velocity at each moment, and trajectory posture at each moment obtained under the conditions of simulating the trajectory angle A;

[0008] The method for obtaining the simulated noise output data of the inertial sensor is as follows: inverting the basic parameters of the trajectory to obtain the noise-free output data of the inertial sensor, and then introducing noise uniformly distributed within a set range into the obtained noise-free output data of the inertial sensor to obtain the simulated noise output data of the inertial sensor;

[0009] The setting range is: [-f r ,f r ]、[-ω r ,ω r ], f r is the bias stability index of the accelerometer in the inertial sensor, ω r It is the bias stability index of the gyroscope in the inertial sensor;

[0010] The second step is to establish a test data set, which includes the basic parameters of the trajectory and the measured noise output data of the inertial sensor;

[0011] The basic trajectory parameters include the trajectory angle B, and the trajectory initial position, trajectory initial velocity, trajectory initial attitude, trajectory position at each moment, trajectory velocity at each moment, and trajectory attitude at each moment obtained under the conditions of simulating the trajectory angle B; the minimum value of the trajectory angle B value range is less than the minimum value of the trajectory angle A value range, and the maximum value of the trajectory angle B value range is greater than the maximum value of the trajectory angle A value range;

[0012] The method for obtaining the measured noise output data of the inertial sensor is as follows: inverting the basic parameters of the trajectory to obtain the noise-free output data of the inertial sensor, and then introducing the measured static noise data into the obtained noise-free output data of the inertial sensor to obtain the measured noise output data of the inertial sensor;

[0013] The third step is to use the min-max normalization method to normalize the training data set established in the first step to obtain a normalized training data set, and to use the min-max normalization method to normalize the test data set established in the second step to obtain a normalized test data set;

[0014] The fourth step is to establish a Conv-LSTM network. The established Conv-LSTM network includes a one-dimensional residual network, a dual LSTM network, and a parameter regression layer. The one-dimensional residual network includes eighteen basic modules, each of which includes two convolution units with a convolution kernel of 3, a Relu activation function, batch normalization, a dropout layer, and a residual calculation module. The one-dimensional residual network is used to learn the hidden variables of the ballistic equation.

[0015] The dual LSTM network consists of two standard LSTM networks, a ReLU activation function, and a dropout layer. Each standard LSTM network has 5-10 neurons. The dual LSTM network is used to fuse the current hidden state with the previous hidden state to estimate the optimal current hidden motion state.

[0016] The parameter regression layer includes three branches, namely branch one, branch two and branch three. Each branch is divided into two small branches. Among them, one small branch in branch one is used to regress the current position information, and the other small branch in branch one is used to regress the position covariance information at the current moment. One small branch in branch two is used to regress the current speed information, and the other small branch in branch two is used to regress the speed covariance information at the current moment. One small branch in branch three is used to regress the current attitude information, and the other small branch in branch one is used to regress the current attitude covariance information.

[0017] The loss function of the established Conv-LSTM network is designed to minimize the mean square error between the predicted and true values;

[0018] The optimizer of the established Conv-LSTM network is the SGD optimization model;

[0019] Step 5: Use the normalized training data set obtained in step 3 to train the Conv-LSTM network established in step 4 to obtain a trained Conv-LSTM network.

[0020] Step 6: The trained Conv-LSTM network obtained in step 5 is stored in the missile-borne computer. After the missile is launched, the measured output data of the inertial sensor is acquired and stored in real time.

[0021] In the seventh step, the measured output data of the inertial sensor obtained in real time in the sixth step or the normalized test data set obtained in the third step are input into the Conv-LSTM network trained in the fifth step to obtain the predicted trajectory posture parameters. The trajectory posture parameters include trajectory position, trajectory velocity, trajectory posture, trajectory position covariance value, trajectory velocity covariance value, and trajectory posture covariance value.

[0022] Step 8: Establish an extended Kalman filter model;

[0023] The observation quantities of the established extended Kalman filter model are the predicted trajectory position, trajectory velocity, and trajectory attitude obtained in the seventh step;

[0024] The observation variance value of the established extended Kalman filter model is the predicted trajectory position covariance value, trajectory velocity covariance value and trajectory attitude covariance value obtained in the seventh step;

[0025] The state variables of the established extended Kalman filter model are the current trajectory position, trajectory velocity, trajectory attitude, trajectory acceleration, and trajectory angular velocity;

[0026] The state equation of the established extended Kalman filter model is the inertial navigation solution model;

[0027] In the ninth step, the predicted trajectory position, trajectory velocity, and trajectory attitude obtained in the seventh step are used as the initial state values ​​of the extended Kalman filter model obtained in the eighth step. The extended Kalman filter model is updated to obtain the optimized trajectory position, trajectory velocity, and trajectory attitude data.

[0028] This paper discloses a pure inertial trajectory parameter measurement method based on the fusion of Conv-LSTM and Kalman filtering. Compared with the existing technology, it has the following advantages and beneficial effects:

[0029] 1. This invention realizes for the first time the use of Conv-LSTM and Kalman filter fusion method to measure the ballistic parameters of artillery-fired intelligent ammunition, reducing the physical model established based on the dynamics of intelligent ammunition.

[0030] 2. This invention realizes for the first time trajectory prediction based solely on low-cost micro-inertial sensors, reducing the increase in cost, power consumption, and weight caused by adding sensors.

[0031] 3. By establishing a Conv-LSTM network, the present invention can effectively predict the strapdown inertial navigation solution posture and correct the accumulated error based on the inertial sensor measurement data and pre-launch trajectory parameters over a period of time, thereby improving the accuracy of trajectory measurement.

[0032] 4. The present invention establishes a Kalman filter model and uses the output of the Conv-LSTM network as the observation quantity to further improve the measurement accuracy of the pure inertial navigation system, thereby accurately obtaining the trajectory posture parameters and accurately predicting the trajectory landing point in advance, and then solving the control instructions to enable the projectile to accurately hit the target, thereby improving the hit accuracy and damage effect of the intelligent ammunition.

[0033] 5. The present invention can achieve high-precision real-time trajectory measurement of gun-launched intelligent ammunition in complex electromagnetic environments or when satellite positioning fails due to enemy interference or deception, thereby improving subsequent flight control and hit accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0035] The specific embodiments of the present invention will be described below with reference to the accompanying drawings.

[0036] Example

[0037] like Figure 1 As shown, a method for measuring inertial ballistic parameters based on the fusion of Conv-LSTM and Kalman filter is provided, and the steps of the method include:

[0038] The first step is to establish a training data set, which includes basic trajectory parameters and simulated noise output data of inertial sensors;

[0039] The basic trajectory parameters include the trajectory angle [35°, 50°] (interval 1°) and the trajectory initial position, trajectory initial velocity, trajectory initial posture, trajectory position at each moment, trajectory velocity at each moment, and trajectory posture at each moment obtained under the conditions of simulating the trajectory angle;

[0040] The method for obtaining the simulated noise output data of the inertial sensor is as follows: inverting the basic parameters of the trajectory to obtain the noise-free output data of the inertial sensor, and then introducing noise uniformly distributed within a set range into the obtained noise-free output data of the inertial sensor to obtain the simulated noise output data of the inertial sensor;

[0041] The setting range is: [-2mg, 2mg], [-10° / h, 10° / h], f r is the bias stability index of the accelerometer in the inertial sensor, ω r It is the bias stability index of the gyroscope in the inertial sensor;

[0042] The second step is to establish a test data set, which includes the basic parameters of the trajectory and the measured noise output data of the inertial sensor;

[0043] The basic trajectory parameters include the trajectory angle [30°, 55°] (interval 1°) and the trajectory initial position, trajectory initial velocity, trajectory initial posture, trajectory position at each moment, trajectory velocity at each moment, and trajectory posture at each moment obtained under the conditions of simulating the trajectory angle;

[0044] The method for obtaining the measured noise output data of the inertial sensor is as follows: inverting the basic parameters of the trajectory to obtain the noise-free output data of the inertial sensor, and then introducing the measured static noise data into the obtained noise-free output data of the inertial sensor to obtain the measured noise output data of the inertial sensor;

[0045] The measured static noise of the three-axis accelerometer in the inertial sensor has a zero bias of 1.44 mg, 0.09 mg, and 0.99 mg, with standard deviations of 2.3388 mg, 2.8898 mg, and 2.0049 mg. The measured static noise of the three-axis gyroscope in the inertial sensor has a zero bias of 30.00° / h, 43.20° / h, and 2.31° / h, with standard deviations of 262.50° / h, 286.1771° / h, and 221.4064° / h.

[0046] In this data set, when the flight time is 100 seconds, the accuracy of the traditional strapdown inertial solution is as follows: the prediction errors of the pitch angle, roll angle and yaw angle are: 5.82°, 5.86° and 5.23°; the prediction errors of the east velocity, north velocity and celestial velocity are: 49.21m / s, 14.03m / s and 33.72m / s respectively; the prediction errors of the east position, north position and celestial position velocity are: 2852.5m, 5096.5m and 1901.9m respectively.

[0047] The third step is to use the min-max normalization method to normalize the training data set established in the first step to obtain a normalized training data set, and to use the min-max normalization method to normalize the test data set established in the second step to obtain a normalized test data set;

[0048] The fourth step is to build a Conv-LSTM network. The Conv-LSTM network includes a one-dimensional residual network, a dual LSTM network, and a parameter regression layer. The one-dimensional residual network includes eighteen basic modules, each of which includes two convolution units with a convolution kernel of 3, a Relu activation function, batch normalization, a dropout layer, and a residual calculation module. The one-dimensional residual network is used to learn the hidden variables of the ballistic equation. In this example, the one-dimensional residual network uses the Resnet18 model.

[0049] The dual LSTM network consists of two standard LSTM networks, a ReLU activation function, and a dropout layer. Each standard LSTM network has five neurons. The dual LSTM network is used to fuse the current hidden state with the previous hidden state to estimate the optimal current hidden motion state.

[0050] The parameter regression layer includes three branches, namely branch one, branch two and branch three. Each branch is divided into two small branches. Among them, one small branch in branch one is used to regress the current position information, and the other small branch in branch one is used to regress the position covariance information at the current moment. One small branch in branch two is used to regress the current speed information, and the other small branch in branch two is used to regress the speed covariance information at the current moment. One small branch in branch three is used to regress the current attitude information, and the other small branch in branch one is used to regress the current attitude covariance information.

[0051] The loss function of the established Conv-LSTM network is designed to minimize the mean square error between the predicted and true values;

[0052] The optimizer of the established Conv-LSTM network is the SGD optimization model;

[0053] Step 5: Use the normalized training data set obtained in step 3 to train the Conv-LSTM network established in step 4 to obtain a trained Conv-LSTM network.

[0054] In the sixth step, the trained Conv-LSTM network obtained in the fifth step is stored in the missile-borne computer. After the missile is launched, the measured output data of the inertial sensor is acquired and stored in real time. In this embodiment, the normalized test data set obtained in the third step is used as the measured output data of the inertial sensor.

[0055] In the seventh step, the normalized test data set obtained in the third step is input into the Conv-LSTM network trained in the fifth step to obtain the predicted trajectory posture parameters, which include trajectory position, trajectory velocity, trajectory attitude, trajectory position covariance value, trajectory velocity covariance value, and trajectory attitude covariance value. After passing through the Conv-Lstm network model, the prediction errors of the pitch angle, roll angle, and yaw angle are: 5.10°, 0.59°, and 0.29°; the prediction errors of the east velocity, north velocity, and celestial velocity are: 2.02m / s, 1.56m / s, and 1.74m / s, respectively; the prediction errors of the east position, north position, and celestial position velocity are: 400.41m, 68.11m, and 108.45m, respectively.

[0056] Step 8: Establish an extended Kalman filter model;

[0057] The observation quantities of the established extended Kalman filter model are the predicted trajectory position, trajectory velocity, and trajectory attitude obtained in the seventh step;

[0058] The observation variance value of the established extended Kalman filter model is the predicted trajectory position covariance value, trajectory velocity covariance value and trajectory attitude covariance value obtained in the seventh step;

[0059] The state variables of the established extended Kalman filter model are the current trajectory position, trajectory velocity, trajectory attitude, trajectory acceleration, and trajectory angular velocity;

[0060] The state equation of the established extended Kalman filter model is the inertial navigation solution model;

[0061] In step nine, the predicted trajectory position, velocity, and attitude obtained in step seven are used as the initial state values ​​for the extended Kalman filter model obtained in step eight. The extended Kalman filter model is then updated to obtain optimized trajectory position, velocity, and attitude data. After Kalman filtering, the predicted accuracies for pitch, roll, and yaw angles are 0.001°, 0.005°, and 0.02°, respectively, for a flight time of 100 seconds. The predicted errors for east, north, and celestial velocities are 0.70 m / s, 0.55 m / s, and 0.07 m / s, respectively. The predicted errors for east, north, and celestial position velocity are 131.34 m, 28.49 m, and 59.47 m, respectively.

[0062] It should be understood that the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for measuring inertial ballistic parameters based on the fusion of Conv-LSTM and Kalman filter, characterized in that The steps of the method include: The first step is to create a training data set; The second step is to create a test data set; The third step is to normalize the training data set established in the first step to obtain a normalized training data set, and to normalize the test data set established in the second step to obtain a normalized test data set; Step 4: Establish a Conv-LSTM network; Step 5: Use the normalized training data set obtained in step 3 to train the Conv-LSTM network established in step 4 to obtain a trained Conv-LSTM network. Step 6: The trained Conv-LSTM network obtained in step 5 is stored in the missile-borne computer. After the missile is launched, the measured output data of the inertial sensor is acquired and stored in real time. Step 7: Input the measured output data of the inertial sensor obtained in real time in step 6 into the Conv-LSTM network trained in step 5 to obtain the predicted trajectory pose parameters; Step 8: Establish an extended Kalman filter model; In the ninth step, the trajectory posture parameters obtained in the seventh step are used as the initial state values ​​of the extended Kalman filter model obtained in the eighth step. The extended Kalman filter model is updated to obtain the optimized trajectory position, trajectory velocity and trajectory posture data.

2. The inertial ballistic parameter measurement method based on Conv-LSTM and Kalman filter fusion according to claim 1 is characterized in that: In the first step, the training data set established includes basic trajectory parameters and simulated noise output data of the inertial sensor; The basic trajectory parameters include the trajectory angle A and the trajectory initial position, trajectory initial velocity, trajectory initial posture, trajectory position at each moment, trajectory velocity at each moment, and trajectory posture at each moment obtained under the conditions of simulating the trajectory angle A; The method for obtaining the simulated noise output data of the inertial sensor is as follows: inverting the basic parameters of the trajectory to obtain the noise-free output data of the inertial sensor, and then introducing noise uniformly distributed within a set range into the obtained noise-free output data of the inertial sensor to obtain the simulated noise output data of the inertial sensor; The setting range is: [-f r ,f r ]、[-ω r ,ω r ], f r is the bias stability index of the accelerometer in the inertial sensor, ω r It is an indicator of the bias stability of the gyroscope in the inertial sensor.

3. The inertial ballistic parameter measurement method based on Conv-LSTM and Kalman filter fusion according to claim 2 is characterized in that: In the second step, the test data set established includes basic trajectory parameters and measured noise output data of the inertial sensor; The basic trajectory parameters include the trajectory angle B, and the trajectory initial position, trajectory initial velocity, trajectory initial attitude, trajectory position at each moment, trajectory velocity at each moment, and trajectory attitude at each moment obtained under the conditions of simulating the trajectory angle B; the minimum value of the trajectory angle B value range is less than the minimum value of the trajectory angle A value range, and the maximum value of the trajectory angle B value range is greater than the maximum value of the trajectory angle A value range; The method for obtaining the measured noise output data of the inertial sensor is as follows: inverting the basic parameters of the trajectory to obtain the noise-free output data of the inertial sensor, and then introducing the measured static noise data into the obtained noise-free output data of the inertial sensor to obtain the measured noise output data of the inertial sensor.

4. A method for measuring inertial ballistic parameters based on Conv-LSTM and Kalman filtering fusion according to any one of claims 1 to 3, characterized in that: In the third step, the min-max normalization method is used for normalization.

5. The method for measuring inertial ballistic parameters based on the fusion of Conv-LSTM and Kalman filter according to claim 4, characterized in that: In the fourth step, the established Conv-LSTM network includes a one-dimensional residual network, a dual LSTM network and a parameter regression layer.

6. The method for measuring inertial ballistic parameters based on the fusion of Conv-LSTM and Kalman filter according to claim 5, characterized in that: The one-dimensional residual network consists of eighteen basic modules, each of which includes two convolutional units with a convolution kernel of 3, a ReLU activation function, batch normalization, a dropout layer, and a residual calculation module. The one-dimensional residual network is used to learn the hidden variables of the ballistic equation. The dual LSTM network consists of two standard LSTM networks, a ReLU activation function, and a dropout layer. Each standard LSTM network includes 5-10 neurons. The dual LSTM network is used to fuse the current hidden state with the previous hidden state to estimate the optimal current hidden motion state.

7. The method for measuring inertial ballistic parameters based on Conv-LSTM and Kalman filter fusion according to claim 5 or 6, characterized in that: The parameter regression layer includes three branches, namely branch one, branch two and branch three. Each branch is divided into two small branches. Among them, one small branch in branch one is used to regress the current position information, and the other small branch in branch one is used to regress the position covariance information at the current moment. One small branch in branch two is used to regress the current speed information, and the other small branch in branch two is used to regress the speed covariance information at the current moment. One small branch in branch three is used to regress the current attitude information, and the other small branch in branch three is used to regress the current attitude covariance information. The loss function of the established Conv-LSTM network is designed to minimize the mean square error between the predicted and true values; The optimizer of the established Conv-LSTM network is the SGD optimization model.

8. The method for measuring inertial ballistic parameters based on the fusion of Conv-LSTM and Kalman filtering according to claim 7, characterized in that: In the seventh step, the trajectory posture parameters include trajectory position, trajectory velocity, trajectory posture, trajectory position covariance value, trajectory velocity covariance value and trajectory posture covariance value.

9. The method for measuring inertial ballistic parameters based on the fusion of Conv-LSTM and Kalman filter according to claim 8, characterized in that: In the eighth step, the observation quantities of the established extended Kalman filter model are the predicted trajectory position, trajectory velocity, and trajectory attitude; The observation variance values ​​of the established extended Kalman filter model are the predicted trajectory position covariance values, trajectory velocity covariance values ​​and trajectory attitude covariance values; The state variables of the established extended Kalman filter model are the current trajectory position, trajectory velocity, trajectory attitude, trajectory acceleration, and trajectory angular velocity; The state equation of the established extended Kalman filter model is the inertial navigation solution model.

10. The method for measuring inertial ballistic parameters based on Conv-LSTM and Kalman filter fusion according to claim 9, characterized in that: In the ninth step, the predicted trajectory position, trajectory velocity, and trajectory attitude are used as the initial state values ​​of the extended Kalman filter model.

Images