An online identification method for drag coefficient of intercepted projectile based on artificial intelligence

Abstract

The application discloses an online identification method for drag coefficient of an interceptor based on artificial intelligence, which utilizes sensors on an aircraft to measure kinematic information of the aircraft and the interceptor in real time, inputs the measured information into an identification model, and obtains the drag coefficient of the interceptor. The online identification method for the drag coefficient of the interceptor based on artificial intelligence has faster identification speed and higher accuracy.

Description

Technical Field

[0001] This invention relates to an online identification method for the drag coefficient of interceptor missiles based on artificial intelligence, belonging to the field of guidance and control. Background Technology

[0002] Currently, the main strategies for aircraft to evade interception include differential game-based evasion strategies and stochastic maneuver evasion strategies. Regardless of the game strategy adopted, to better evade interception, it is necessary to use the relative motion information of our aircraft and the enemy intercepting aircraft detected by our aircraft to predict its trajectory. An important prerequisite for accurately predicting its trajectory is to identify its drag coefficient.

[0003] Traditional online identification methods are typically based on Kalman filtering. However, these methods have drawbacks: Kalman filtering is a prediction-correction method, and its computational process can be divided into two steps: state prediction and probability correction. To accurately predict enemy-friendly game systems, information about enemy interceptor missiles, such as their velocity and direction, must be accurately obtained. This is difficult to achieve in complex air combat environments. In Kalman filtering, to address this issue, the difference between the true and estimated values ​​is usually treated as significant process noise. However, this approach can easily lead to non-convergence of the identification model. Furthermore, Kalman filtering requires high continuity of information acquisition, significantly reducing its engineering application value and limiting the maneuverability of friendly aircraft.

[0004] Therefore, it is necessary to further study the identification of drag coefficients of traditional interceptor missiles in order to solve the above problems. Summary of the Invention

[0005] To overcome the above problems, the inventors conducted in-depth research and designed an online identification method for the drag coefficient of interceptor missiles based on artificial intelligence. The method uses sensors on the aircraft to measure the kinematic information of the aircraft and the interceptor missile in real time, and inputs the measured information into the identification model to obtain the drag coefficient of the interceptor missile.

[0006] In a preferred embodiment, the kinematic information of the aircraft and the interceptor missile includes:

[0007] Aircraft acceleration a T The aircraft's velocity tilt angle θ m The aircraft's velocity deflection angle ψ V ;

[0008] The relative distance r between the aircraft and the interceptor missile TM The interceptor missile's line-of-sight tilt angle q relative to the aircraft y TM The interceptor missile's line-of-sight angle q relative to the aircraft z TM;

[0009] The relative distance r between the aircraft and the interceptor missile TM rate of change, interceptor missile tilt angle relative to the aircraft's line of sight q y TM Rate of change, interceptor missile relative to the aircraft's line-of-sight angle q z TM The rate of change.

[0010] In a preferred embodiment, the identification model is obtained by the following method:

[0011] Step 1: Obtain kinematic information data of multiple sets of aircraft and interceptor missiles, use the drag coefficient data of the interceptor missiles in the kinematic information as type labels, and use the remaining information as input data to build a sample library;

[0012] Step 2: Train the neural network using a sample library to obtain the recognition model.

[0013] In a preferred embodiment, in step 1, multiple sets of kinematic information data of the aircraft and interceptor missile are obtained through a relative kinematics simulation model.

[0014] The relative kinematics simulation model is used to simulate the process of an interceptor missile attacking a flight vehicle.

[0015] In a preferred embodiment, the kinematic equations of the interceptor missile, the acceleration required for interception, and the velocity change of the interceptor missile caused by drag are set in the relative kinematics simulation model to obtain the motion trajectory of the aircraft and the interceptor missile, and the kinematic information data of the aircraft and the interceptor missile are obtained based on the motion trajectory.

[0016] In a preferred embodiment, the kinematic equations of the interceptor missile are expressed as:

[0017]

[0018] Where, x m y m z m Let θ be the coordinates of the aircraft in the inertial frame. m V is the velocity tilt angle of the aircraft. m Let ψ be the velocity of the aircraft. V Let be the velocity deflection angle, t represent time, g be the gravitational acceleration, m(t) be the interceptor mass, T(t) be the interceptor thrust, D be the drag coefficient of the interceptor, and n be the velocity angle. y For the pitch acceleration of the interceptor missile, n z This refers to the yaw acceleration of the interceptor missile.

[0019] In a preferred embodiment, in step 2, the neural network includes an input layer, hidden layers, an output layer, and a multi-model layer.

[0020] The input to the input layer is the kinematic information of the aircraft and the interceptor missile at different times;

[0021] The output of the input layer is represented as:

[0022] h t =tanh(w xh x t +b xh )

[0023] Among them, h t w is the information passed from the input layer to the hidden layer. xh Let x be the weight matrix. t Input information at time t, b xh tanh() is the bias passed from the input layer to the hidden layer, and tanh() is the activation function.

[0024] In a preferred embodiment, the hidden layer has multiple layers, preferably three layers, and each hidden layer has n neurons, wherein the neurons are GRU neurons.

[0025] In a preferred embodiment, the multi-model layer is connected to the output layer of the neural network and includes multiple models. Each model is connected to the output layer. Each model takes the output of the neural network as input, performs calculations, and outputs its own output. The outputs of multiple models are combined to obtain a regression result, which is the output of the multi-model layer. This regression result is the output of the identification model.

[0026] In a preferred embodiment, the output of the multi-model layer is:

[0027] O = Λ T G

[0028] Where Λ represents the state matrix of the model state values.

[0029] The beneficial effects of this invention include:

[0030] (1) According to the online identification method for the drag coefficient of interceptor missiles based on artificial intelligence provided by the present invention, a trained neural network is used as the identification model, which can improve the identification speed;

[0031] (2) According to the online identification method of interceptor drag coefficient based on artificial intelligence provided by the present invention, the identification accuracy and training speed are greatly improved by optimizing the traditional neural network, and the identification accuracy is ultimately higher. Attached Figure Description

[0032] Figure 1A schematic flowchart of an online identification method for the drag coefficient of an interceptor missile based on artificial intelligence, according to a preferred embodiment of the present invention, is shown.

[0033] Figure 2 The graph shows the change of the loss function during the training of the identification model in Example 1;

[0034] Figure 3 The graph shows the change in recognition accuracy during training in Example 1.

[0035] Figure 4 The diagram shows the changes in the loss function during training for Example 1 and Comparative Example 1. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Through these descriptions, the features and advantages of the present invention will become clearer and more apparent.

[0037] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments. Although various aspects of embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless specifically indicated otherwise.

[0038] According to the present invention, an online identification method for the drag coefficient of an interceptor missile based on artificial intelligence is provided, such as... Figure 1 As shown, the kinematic information of the aircraft and the interceptor missile is measured in real time using sensors on the aircraft. The measured information is then input into the identification model to obtain the drag coefficient of the interceptor missile.

[0039] In a preferred embodiment, the kinematic information of the aircraft and the interceptor missile includes:

[0040] Aircraft acceleration a T The aircraft's velocity tilt angle θ m The aircraft's velocity deflection angle ψ V ;

[0041] The relative distance r between the aircraft and the interceptor missile TM The interceptor missile's line-of-sight tilt angle q relative to the aircraft y TM The interceptor missile's line-of-sight angle q relative to the aircraft z TM ;

[0042] The relative distance r between the aircraft and the interceptor missile TM rate of change, interceptor missile tilt angle relative to the aircraft's line of sight q y TM Rate of change, interceptor missile relative to the aircraft's line-of-sight angle q z TMThe rate of change.

[0043] According to the present invention, the identification model is obtained by the following method:

[0044] Step 1: Obtain kinematic information data of multiple sets of aircraft and interceptor missiles, use the drag coefficient data of the interceptor missiles in the kinematic information as type labels, and use the remaining information as input data to build a sample library;

[0045] Step 2: Train the neural network using a sample library to obtain the recognition model.

[0046] Preferably, in step 1, multiple sets of kinematic information data of the aircraft and interceptor missile are obtained through a relative kinematics simulation model.

[0047] The relative kinematics simulation model is used to simulate the process of an interceptor missile attacking a flight vehicle.

[0048] Furthermore, according to the present invention, the kinematic equations of the interceptor missile, the acceleration required for interception, and the velocity change of the interceptor missile caused by drag are set in the relative kinematic simulation model. Simulation is performed according to the above settings to obtain the motion trajectory of the aircraft and the interceptor missile, and the kinematic information data of the aircraft and the interceptor missile are obtained based on the motion trajectory.

[0049] In this invention, no special settings are made for the specific simulation platform; any known simulation platform, such as MATLAB, can be used.

[0050] More preferably, the kinematic equations of the interceptor missile are expressed as:

[0051]

[0052] The kinematic equations of the interceptor missile are based on an inertial coordinate system, where x m y m z m Let θ be the coordinates of the aircraft in the inertial frame. m For the velocity inclination angle, V m Let ψ be the velocity of the aircraft. V Let θ be the velocity deflection angle, t be time, g be the gravitational acceleration, m be the interceptor mass, T(t) be the interceptor thrust, D be the drag coefficient of the interceptor, and n be the velocity deflection angle. y For the pitch acceleration of the interceptor missile, n z Let n be the yaw acceleration of the interceptor missile; where n is the yaw acceleration of the interceptor missile. y n z It varies with the drag coefficient and guidance parameters.

[0053] The interception requires acceleration, which is expressed as follows:

[0054]

[0055] n PN n represents the required acceleration of the interceptor missile under the PN guidance law. APN This represents the required acceleration of the interceptor missile under the APN guidance law, where N is the guidance parameter and n is the acceleration required under the APN guidance law. T For the target acceleration, V r The relative velocity between the aircraft and the interceptor missile. The interceptor missile's line-of-sight angular rate relative to the aircraft;

[0056] In this invention, the acceleration required for interception includes not only the PN guidance law but also the APN guidance law. The PN drag coefficient is the most widely used drag coefficient currently. It requires that during the attack on the target, the overload required by the aircraft is proportional to the product of the relative velocity and the line-of-sight angle between the projectile and the target. The APN adds a target maneuver compensation term on the basis of PN, thereby reducing the normal overload required by the trajectory near the point of impact and increasing the accuracy of the simulation results.

[0057] The velocity change of the interceptor missile due to drag is expressed as:

[0058]

[0059] S M Let D be the reference area of ​​the interceptor missile, m be the drag coefficient of the interceptor missile, ρ be the mass of the interceptor missile, and V be the air density. m For speed

[0060] According to a preferred embodiment of the present invention, during the simulation process, by setting different drag coefficients D and guidance parameters N, multiple sets of flight trajectories of the aircraft and interceptor missile are obtained, and the kinematic information of the aircraft and interceptor missile is obtained based on the flight trajectories.

[0061] According to a preferred embodiment of the present invention, after obtaining the flight trajectories of the aircraft and the interceptor missile, the relative distance r between the aircraft and the interceptor missile is... TM The interceptor missile's line-of-sight tilt angle q relative to the aircraft y TM The interceptor missile's line-of-sight angle q relative to the aircraft z TM The data is obtained through the following formula

[0062]

[0063]

[0064]

[0065] Where, x r y r z rThis represents the components of the relative distance between the interceptor missile and the aircraft in the x, y, and z directions in the inertial coordinate system.

[0066] According to a preferred embodiment of the present invention, in step 2, the neural network includes an input layer, a hidden layer, an output layer, and a multi-model layer.

[0067] The input layer takes kinematic information about the aircraft and the interceptor missile at different times as input, and its input is size(input) = n. input ×n step ,

[0068] Where size(input) represents the size of the input matrix, n input This represents the input dimension, i.e., the kinematic information of the aircraft and the interceptor missile at the same moment, n. step This represents the input step size, i.e., different time points. The input data is in the form of a two-dimensional matrix, which, compared to a one-dimensional input BP network, can quickly identify and detect changes in different features at different time points;

[0069] The output of the input layer is represented as:

[0070] h t =tanh(w xh x t +b xh )

[0071] Among them, h t w is the information passed from the input layer to the hidden layer. xh Let x be the weight matrix. t Input information at time t, b xh tanh() is the bias passed from the input layer to the hidden layer, and tanh() is the activation function.

[0072] The hidden layer has multiple layers, preferably three layers, and each hidden layer has n neurons. More preferably, the neurons are GRU neurons.

[0073] According to the present invention, the output layer has multiple neurons, and the output G of the output layer is represented as:

[0074]

[0075] Where G is the output of the output layer. G i Let represent different neurons in the output layer, representing the probability of different resistance coefficients, where p represents the number of neurons in the output layer, and ... w ho Weights from hidden layer to output layer For the output of the last hidden layer at time t, b hoThe hidden layer is used to bias the output layer, and softmax() is the activation function.

[0076] According to the present invention, the multi-model layer is connected to the output layer of the neural network and includes multiple models. Each model is connected to the output layer. Each model takes the output result of the neural network as input, and outputs the output result of the model after calculation. The output results of multiple models are merged to obtain a regression result, which is the output result of the multi-model layer. This regression result is the output of the identification model.

[0077] The multi-model mechanism in this invention can achieve a smooth mapping from the unknown range to the known range, which can greatly accelerate the training speed and prediction accuracy of neural networks in regression problems, thereby improving the accuracy of identification.

[0078] Furthermore, the number of models in the multi-model layer is set before neural network training. Preferably, the number of models p is 2 to 10, and more preferably p = 5.

[0079] Furthermore, the output of the multi-model layer is:

[0080] O = Λ T G

[0081] Wherein, Λ represents the state matrix of the model state values, and preferably, its value is:

[0082] Λ = [0.0, 0.1, 0.2, 0.3, 0.4] T

[0083] The above values ​​are empirical values, which can greatly improve the accuracy of neural network recognition after training.

[0084] According to a preferred embodiment of the present invention, the Adam method is used to train the neural network.

[0085] The Adam method combines the Momentum algorithm with the RMSprop algorithm, incorporating momentum and enabling adaptive parameter training. Compared to the BP algorithm, the Adam method can effectively prevent the network from getting trapped in local optima, accelerate network convergence, and improve the robustness of the algorithm.

[0086] More preferably, the Adam method is updated as follows:

[0087]

[0088]

[0089] Where, θ t The parameters before the update; θ t+1The updated parameters are: η is the learning rate, which takes the value 0.001; ε is the smoothing term to prevent division by zero; m' t For the first-order moment estimate, v' t For second-order moment estimation; m t v is the first-order gradient matrix; t It is a second-order gradient matrix;

[0090] In a preferred embodiment, after establishing the GRU network, the data in the sample library is divided into training data and test data. Training data is extracted from the sample library in batches and input into the GRU network for training. The training set loss function value L is then calculated. train The network training refers to the process by which the network continuously adjusts its parameter values ​​to adapt to the input and output of the training samples; after each training iteration, test data is input into the network for testing, and the test set loss function value L is obtained. test Repeat the above process until the number of training iterations or the loss function reaches the preset value.

[0091] According to a preferred embodiment of the present invention, before inputting the measured information into the identification model and before constructing the sample library, the obtained data is further normalized, and the data is scaled to the [0,1] interval, as follows:

[0092]

[0093] Where ξ i For the processed data, ξ i ′ represents the data before processing, ξ i ′ ,max To process the maximum value of the first i-th dimension of the input data, ξ i ′ ,min This represents the minimum value of the i-th dimension of the input data before processing.

[0094] Example

[0095] Example 1

[0096] The model experiment includes the following steps:

[0097] The identification model was obtained using the following method:

[0098] Step 1: Obtain kinematic information data of multiple sets of aircraft and interceptor missiles, use the drag coefficient data of the interceptor missiles in the kinematic information as type labels, and use the remaining information as input data to build a sample library;

[0099] Step 2: Train the neural network using a sample library to obtain the recognition model.

[0100] In step 1, the kinematic equations of the interceptor missile, the acceleration required for interception, and the velocity change of the interceptor missile caused by drag are set in the relative kinematics simulation model. The simulation is performed according to the above settings to obtain the motion trajectory of the aircraft and the interceptor missile, and the kinematic information data of the aircraft and the interceptor missile are obtained based on the motion trajectory.

[0101] The kinematic equations of the interceptor missile are expressed as follows:

[0102]

[0103] The interception requires acceleration, which is expressed as follows:

[0104]

[0105] The velocity change of the interceptor missile due to drag is expressed as:

[0106]

[0107] Furthermore, the simulation step size was set to 0.001s, the aircraft updated the kinematic information of the interceptor missile at a frequency of 50Hz, and finally 16,000 training samples were extracted. The drag coefficient of the interceptor missile was randomly set from [0.0, 0.4]. 10% of the extracted samples were randomly selected for testing to form a test set, and the rest were used for training.

[0108] Before constructing the sample library, the obtained data was normalized and scaled to the [0,1] interval, as follows:

[0109]

[0110] In step 2, the neural network comprises an input layer, hidden layers, an output layer, and a multi-model layer. The output of the input layer is represented as follows:

[0111] h t =tanh(w xh x t +b xh )

[0112] The hidden layer consists of three layers, each containing 96 GRU neural units;

[0113] The output G of the output layer is represented as:

[0114]

[0115] The multi-model layer contains 5 models, and its output is:

[0116] O = Λ T G

[0117] Where Λ = [0.0, 0.1, 0.2, 0.3, 0.4] T

[0118] The Adam method is used to train the neural network, and the Adam method update is as follows:

[0119]

[0120]

[0121] In the training parameters, the dropout failure rate is 5%, the batch size is 3000, the initial learning rate is 0.0015, and the learning rate decays by 0.98 every 100 iterations. The loss function changes during training as follows: Figure 2 As shown.

[0122] The trained identification model was used to conduct simulation experiments. Data from the test set was used as kinematic information measurements of the aircraft and the interceptor missile by sensors on the aircraft. These measurements were then input into the identification model to obtain the drag coefficient of the interceptor missile.

[0123] In the test set, data with an enemy interceptor missile drag coefficient of 0.2 were selected for verification, and the recognition accuracy changed during training as follows: Figure 3 As shown, it can be seen that during the identification process, the model can quickly converge to the vicinity of the true value (0.2), and remains stable in the vicinity in subsequent identifications, indicating that the method has extremely high identification accuracy.

[0124] Comparative Example 1

[0125] Using the sample library obtained in Example 1, and replacing the neural network in S2 with a traditional GRU neural network, the same experiment as in Example 1 was conducted. The loss function of the neural network during training was as follows: Figure 4 As shown.

[0126] As can be seen from the figure, the method in Example 1 can obtain a smaller loss function compared to Comparative Example 1, that is, it is more accurate. Moreover, the loss function of the method in Example 1 decreases more quickly, that is, it can obtain relatively accurate identification results even with low training samples.

[0127] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "front," and "rear," etc., indicate the orientation or positional relationship based on the orientation or positional relationship in the working state of this invention, and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. Furthermore, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0128] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0129] The present invention has been described above with reference to preferred embodiments; however, these embodiments are merely exemplary and illustrative. Various substitutions and modifications can be made to the present invention based on these embodiments, all of which fall within the scope of protection of the present invention.

Claims

1. An online identification method for the drag coefficient of interceptor missiles based on artificial intelligence, characterized in that, The kinematic information of the aircraft and the interceptor missile is measured in real time using sensors on the aircraft. The measured information is then input into the identification model to obtain the drag coefficient of the interceptor missile. The kinematic information of the aircraft and the interceptor missile includes: Aircraft acceleration aircraft speed tilt angle Aircraft speed deflection angle ; Relative distance between the aircraft and the interceptor missile The angle of view of the interceptor missile relative to the aircraft's line of sight The angle of deflection of the interceptor missile relative to the aircraft's line of sight ; Relative distance between the aircraft and the interceptor missile Rate of change, interceptor missile's line-of-sight tilt relative to the aircraft Rate of change, interceptor missile deflection relative to the aircraft's line of sight The rate of change; The identification model was obtained through the following method: Step 1: Obtain kinematic information data of multiple sets of aircraft and interceptor missiles, use the drag coefficient data of the interceptor missiles in the kinematic information as type labels, and use the remaining information as input data to build a sample library; Step 2: Train the neural network using a sample database to obtain the recognition model; In step 2, the neural network includes an input layer, hidden layers, an output layer, and a multi-model layer. The output layer has multiple neurons, and the output of the output layer... Represented as: in, For the output of the output layer, , These represent the probabilities of different neurons in the output layer, characterizing different resistance coefficients. This indicates the number of neurons in the output layer, and has... , Weights from hidden layer to output layer The output of the last hidden layer at time t. The hidden layer is used to bias the output layer, and softmax() is the activation function. The multi-model layer is connected to the output layer of the neural network and includes multiple models. Each model is connected to the output layer. Each model takes the output of the neural network as input, and outputs its own output after processing. The outputs of multiple models are combined to obtain a regression result, which is the output of the multi-model layer. This regression result is the output of the identification model.

2. The online identification method for the drag coefficient of an interceptor missile based on artificial intelligence according to claim 1, characterized in that, In step 1, multiple sets of kinematic information data of the aircraft and interceptor missile are obtained through a relative kinematics simulation model. The relative kinematics simulation model is used to simulate the process of an interceptor missile attacking a flight vehicle.

3. The online identification method for the drag coefficient of an interceptor missile based on artificial intelligence according to claim 2, characterized in that, In the relative kinematics simulation model, the kinematic equations of the interceptor missile, the acceleration required for interception, and the velocity change of the interceptor missile caused by drag are set to obtain the motion trajectory of the aircraft and the interceptor missile. Based on the motion trajectory, the kinematic information data of the aircraft and the interceptor missile are obtained.

4. The online identification method for the drag coefficient of an interceptor missile based on artificial intelligence according to claim 3, characterized in that, The kinematic equations of the interceptor missile are expressed as follows: in, , , The coordinates of the aircraft in the inertial frame are... For the aircraft's speed tilt angle, For the speed of the aircraft, For velocity deflection angle, Indicates time, It is the acceleration due to gravity. For the interceptor missile mass, To intercept the missile's thrust, The drag coefficient experienced by the interceptor missile. For the pitch acceleration of the interceptor missile, This refers to the yaw acceleration of the interceptor missile.

5. The online identification method for the drag coefficient of an interceptor missile based on artificial intelligence according to claim 1, characterized in that, The input to the input layer is the kinematic information of the aircraft and the interceptor missile at different times; The output of the input layer is represented as: in, Information passed from the input layer to the hidden layer. This is the weight matrix. Input information at time t. tanh() is the bias passed from the input layer to the hidden layer, and tanh() is the activation function.

6. The online identification method for the drag coefficient of an interceptor missile based on artificial intelligence according to claim 5, characterized in that, The hidden layer has multiple layers, and each hidden layer has n neurons, which are GRU neurons.

7. The online identification method for the drag coefficient of an interceptor missile based on artificial intelligence according to claim 1, characterized in that, The output of the multi-model layer is: in, The state matrix represents the state values ​​of the model.

Images