Method for quickly obtaining firing data of precision guided aircraft based on BP neural network

Abstract

The application discloses a method for quickly obtaining firing data of a precision-guided aircraft based on a BP neural network, wherein initial data is obtained by giving aerodynamic parameters and step sampling, the initial data is substituted into a mathematical simulation model to obtain multiple groups of trajectory data, a neural network taking expected range and current meteorological information as input and taking firing data as output is obtained by training the neural network according to the trajectory data after removing redundant data through sensitivity analysis, and the neural network is filled into the aircraft, so that the aircraft can obtain the firing data by itself only by filling the expected range and the current meteorological information into the aircraft before flying, thereby simplifying an operation process and improving the hitting accuracy of the aircraft.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aircraft guidance control, in particular to a method for quickly obtaining firing data of a precision guided aircraft based on a BP neural network. BACKGROUND

[0002] Accurate data calculation is a prerequisite for ensuring the hit rate of a precision guided aircraft and forming combat capability, and is a necessary core function of a fire control computer and a command instrument. The data calculation of a precision guided aircraft refers to calculating a set of accurate firing data before the aircraft is launched, according to actual launch conditions, so that the aircraft can meet the use requirements and complete the precision strike task by loading the data. The firing data of a typical laser terminal guided aircraft mainly include charge number, angle of elevation (also referred to as sight vane in the present application), azimuth (also referred to as azimuth in the present application), mechanical gyro working time (also referred to as program loading in the present application), and laser irradiation start time (also referred to as delay in the present application).

[0003] The calculation of the data takes many factors into account, including charge temperature, Magnus effect, and meteorological conditions, among which the meteorological conditions include ground pressure, temperature, crosswind, and longitudinal wind. Many constraints are considered, including range, program loading and angle range, seeker field of view, overload capacity, and terminal guidance segment speed. The calculation output is also multiple, including sight vane, azimuth, program loading, and delay. Therefore, the preparation of a firing table is a parameter optimization problem with multiple inputs and multiple outputs, which requires complex optimization calculation with spatial and temporal coupling iteration. One way is to obtain the calculation result by trajectory iteration optimization, which involves numerical solution of six-degree-of-freedom nonlinear differential equations, resulting in slow calculation speed and long time consumption, which is extremely disadvantageous for the precision strike of time-sensitive targets in some tasks. Another way is to modify each factor according to the current meteorological conditions by querying the previously loaded firing table, which has poor calculation accuracy and reduces the combat effectiveness of the precision guided aircraft.

[0004] Due to the above reasons, the present inventors have conducted in-depth research on the existing data calculation method in order to design a method for quickly obtaining firing data that can solve the above problems. SUMMARY

[0005] In order to overcome the above problems, the present inventors have made intensive studies and designed a method for quickly obtaining firing data of an accurate guidance aircraft based on a BP neural network, in which initial data is obtained by giving aerodynamic parameters and step sampling, a plurality of trajectory data is obtained by substituting the initial data into a mathematical simulation model, a neural network with expected range and current meteorological information as input and firing data as output is obtained by removing redundant data through sensitivity analysis and training the neural network, and the neural network is loaded into the aircraft, so that the aircraft can obtain the firing data by itself only by loading the expected range and current meteorological information before flight, thereby simplifying the operation process, improving the hit accuracy of the aircraft, and thus completing the present application.

[0006] Specifically, the present application aims to provide a method for quickly obtaining firing data of an accurate guidance aircraft based on a BP neural network, which comprises the following steps:

[0007] Step 1: Given the aerodynamic parameters, establish a five-degree-of-freedom motion equation and write a five-degree-of-freedom mathematical simulation model;

[0008] Step 2: Small step sampling is used to sample the weather, sight vane, range, and delay in the predetermined interval, and different values obtained by sampling are used as the input of the five-degree-of-freedom mathematical simulation model to perform batch simulation and obtain corresponding trajectory data;

[0009] Step 3: Select neural network samples from the trajectory data through sensitivity analysis;

[0010] Step 4: Use Adam learning law to train the neural network offline;

[0011] Step 5: Load the trained neural network into the fire control computer of the aircraft as a data obtaining model;

[0012] Step 6: Input the expected range and current meteorological information into the data obtaining model to obtain the firing data.

[0013] In step 1, the five-degree-of-freedom motion equation includes a center of mass motion dynamics equation, a center of mass rotation dynamics equation, a center of mass kinematics equation, a center of mass rotation kinematics equation, and a mass change equation;

[0014] Preferably, the center of mass motion dynamics equation is formula (1) as follows:

[0015]

[0016] Wherein, m represents the real-time mass of the aircraft, V mV represents the flight speed of the aircraft, P represents the engine thrust on the aircraft, X represents the aerodynamic drag received by the aircraft, Y represents the lift received by the aircraft, and Z represents the lateral force received by the aircraft; a represents the attack angle of the aircraft, b represents the sideslip angle of the aircraft, and 0 represents the ballistic angle of the aircraft, g represents the flight path angle of the aircraft, and y represents the yaw angle of the aircraft v represents the speed inclination angle of the aircraft v represents the ballistic angle of the aircraft

[0017] Preferably, the rotation dynamics equation around the center of mass is as follows (2):

[0018]

[0019] wherein J y represents the moment of inertia of the aircraft in the Y-axis of the missile body coordinate; J z represents the moment of inertia of the aircraft in the Z-axis of the missile body coordinate; M y represents the projection of the moment of force on the Y-axis generated by the external force on the center of mass of the aircraft; M z represents the projection of the moment of force on the Z-axis generated by the external force on the center of mass of the aircraft; w y represents the yaw angular velocity of the aircraft rotating around the center of mass; w z represents the roll angular velocity of the aircraft rotating around the center of mass

[0020] Preferably, the center of mass kinematics equation is as follows (3):

[0021]

[0022] wherein (x, y, z) represents the position coordinates of the center of mass of the aircraft in the inertial coordinate system

[0023] Preferably, the rotation kinematics equation around the center of mass is as follows (4):

[0024]

[0025] wherein, represents the pitch angle of the aircraft; y represents the yaw angle of the aircraft; and g represents the inclination angle of the aircraft

[0026] Preferably, the mass change equation is as follows (5):

[0027]

[0028] wherein m represents the real-time mass of the aircraft, m c represents the mass flow rate of the aircraft per second

[0029] In step two, the sampling initial value of the weather is that the initial value of the air pressure is 950 hPa, the sampling step is 10 hPa, the initial value of the temperature is 15°, and the sampling step is 1°; in the present application, since the influence of the ground air pressure is greater, for the convenience of training, only the ground air pressure and the temperature value are considered;

[0030] The sampling initial value of the sight is 200 mil, and the sampling step is 1 mil;

[0031] The sampling initial value of the program is 7 s, and the sampling step is 1 s;

[0032] The sampling initial value of the delay is 10 s, and the sampling step is 1 s.

[0033] In step three, for the parameters obtained by sampling, the control variable method is used to filter out the trajectory data which has little influence on the range, and the neural network sample is obtained;

[0034] Preferably, the neural network sample is in the form of, the expected range and the weather in the trajectory data are used as inputs, and the sight, the program, and the delay in the trajectory data are used as outputs.

[0035] In step four, the neural network is a BP neural network,

[0036] Preferably, the neural network includes 5 layers of hidden layers, and the number of neurons in each layer is 10.

[0037] In step four, 70% of the neural network samples obtained by screening in step three are used as a training set, 15% are used as a test set, and 15% are used as a verification set.

[0038] In step four, the following sub-steps are included:

[0039] Sub-step 1: a random number in an interval is assigned to each connection weight, and a Loss function is set,

[0040] Sub-step 2: an input sample and the corresponding expected output are randomly selected from the training set, and the input sample is input into the neural network;

[0041] Sub-step 3: the input and output of each neuron in the hidden layer are calculated;

[0042] Sub-step 4: the partial derivative of the error function with respect to each neuron in the output layer is calculated using the expected output in the sample and the actual output of the neural network;

[0043] Sub-step 5: the Adam learning law is used to correct the weight and bias term;

[0044] Sub-step 6: the rationality of the model is judged, if the error reaches the preset accuracy, the training is ended, if the preset accuracy is not reached, steps 2-6 are repeated.

[0045] The application also provides a BP neural network-based precise guided aircraft shooting parameter quick obtaining system, which comprises a parameter obtaining model loaded in the aircraft, and the shooting parameter is obtained through the parameter obtaining model before the aircraft is launched, and the aircraft is launched according to the shooting parameter.

[0046] Preferably, the parameter obtaining model is obtained by the following method:

[0047] Step one: given the aerodynamic parameters, establish the five-degree-of-freedom motion equation, and compile the five-degree-of-freedom mathematical simulation model;

[0048] Step two: use small steps to sample the weather, sight, range, delay in the predetermined interval, and use the different values obtained by sampling as the input of the five-degree-of-freedom mathematical simulation model to perform batch simulation and obtain the corresponding trajectory data;

[0049] Step three: filter the neural network samples from the trajectory data through sensitivity analysis;

[0050] Step four: use the Adam learning law to perform offline training on the neural network, thereby obtaining the parameter obtaining model.

[0051] The application also provides an aircraft, wherein the BP neural network-based precise guided aircraft shooting parameter quick obtaining system as described above is arranged in the aircraft.

[0052] The application has the following beneficial effects:

[0053] (1) The BP neural network-based precise guided aircraft shooting parameter quick obtaining method can be conveniently deployed in the fire control computer, and the tedious shooting table query is avoided;

[0054] (2) The training samples required by the BP neural network-based precise guided aircraft shooting parameter quick obtaining method are easy to generate, the neural network of the application can be fully trained, the accuracy of the parameter solving of the obtained model is high, the calculation amount is small, and the time consumption is short;

[0055] (3) The BP neural network-based precise guided aircraft shooting parameter quick obtaining method has good universality, and can be applied to a series of precise guided aircrafts such as 155mm or less mortar shells, rocket shells, and cannon shells to obtain the parameters at different altitudes. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1The figure shows the whole logic diagram of the method for quickly obtaining the firing data of the precision guided aircraft based on BP neural network according to a preferred embodiment of the present application;

[0057] Figure 2 The figure shows the case of hitting the target by the first aircraft according to an embodiment of the present application;

[0058] Figure 3 The figure shows the case of hitting the target by the second aircraft according to an embodiment of the present application;

[0059] Figure 4 The figure shows the case of hitting the target by the third aircraft according to an embodiment of the present application. DETAILED DESCRIPTION

[0060] The present application will be further described in details by the accompanying drawings and embodiments. The features and advantages of the present application will become more apparent through these descriptions.

[0061] The word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any implementation described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other implementations. Unless specifically stated otherwise, the drawings are not drawn to scale and the depiction of features in the drawings is intended to provide a conceptual understanding of the embodiments.

[0062] The method for quickly obtaining the firing data of the precision guided aircraft based on BP neural network provided by the present application comprises the following steps: Figure 1

[0063] Step one, given the aerodynamic parameters, establish the five-degree-of-freedom motion equation, and compile the five-degree-of-freedom mathematical simulation model;

[0064] The aerodynamic parameters described in the present application include the lift coefficient, the drag coefficient, and the side force coefficient, i.e. according to these aerodynamic parameters, the X, Y, and Z, i.e. the lift, the drag, and the side force can be obtained.

[0065] The five-degree-of-freedom mathematical simulation model includes the five-degree-of-freedom motion equation and the guidance and control model; wherein, at the beginning of the delay time in the terminal guidance section, the seeker starts to work. Through the guidance and control model, the rotation angular velocity of the missile-target line is measured, and then the control signal of the rudder is given according to the guidance law to correct the longitudinal and lateral deviations until the target is hit.

[0066] The five-degree-of-freedom motion equation includes the mass center motion dynamics equation, the rotation dynamics equation around the mass center, the mass center kinematics equation, the rotation kinematics equation around the mass center, and the mass change equation;

[0067] Preferably, the mass center motion dynamics equation is the following formula (1):

[0068] ​

[0069] wherein m represents the real-time mass of the aircraft, which is obtained in real time by formula (5) so as to obtain the continuously changing mass information in real time; g represents the gravity acceleration, which is 9.8;

[0070] V m represents the flight speed of the aircraft, and the value thereof is obtained by solving the differential equation by using the Runge-Kutta method;

[0071] P represents the engine thrust on the aircraft, the value of which changes with time and is obtained by solving the differential equation, i.e., solving formula (1) above, and when no engine is arranged on the aircraft, the value of the force is 0; preferably, the specific value of the thrust is set before launching, such as the relationship between the working time and the thrust given in the following table:

[0072] Time 0 0.15 2.1 3.53 8 Thrust 330 610 600 48 0

[0073] X represents the aerodynamic drag received by the aircraft, which is obtained according to the aerodynamic parameters;

[0074] Y represents the lift received by the aircraft;

[0075] Z represents the side force received by the aircraft;

[0076] α represents the attack angle of the aircraft, which can be obtained by the following formula:

[0077] β represents the side slip angle of the aircraft, which can be obtained by the following formula: β = arcsin [cos θ sin (ψ - ψ v )]

[0078] θ represents the ballistic angle of the aircraft, the value of which is obtained by formula (1);

[0079] γ v represents the speed inclination angle of the aircraft, which can be obtained by the following formula: γ v = arcsin (tan β tan θ)

[0080] ψ v represents the ballistic deflection angle of the aircraft.

[0081] Preferably, the rotation dynamics equation around the center of mass is formula (2) as follows:

[0082]

[0083] wherein J y represents the rotation inertia of the Y axis of the aircraft in the body coordinate;

[0084] J zIzz represents the moment of inertia of the aircraft in the Z axis in the missile body coordinate system;

[0085] M y Iy represents the projection of the moment of inertia of the aircraft in the Y axis in the missile body coordinate system;

[0086] M z Iz represents the projection of the moment of inertia of the aircraft in the Z axis in the missile body coordinate system;

[0087] w y ωy represents the yaw angular velocity of the aircraft rotating around the center of mass;

[0088] w z ωx represents the roll angular velocity of the aircraft rotating around the center of mass;

[0089] Preferably, the center of mass kinematics equation is as follows (3):

[0090]

[0091] where (x, y, z) represents the position coordinates of the center of mass of the aircraft in the inertial coordinate system.

[0092] Preferably, the rotation kinematics equation around the center of mass is as follows (4):

[0093]

[0094] where, θ represents the pitch angle of the aircraft;

[0095] ψ represents the yaw angle of the aircraft;

[0096] γ represents the tilt angle of the aircraft, which can be obtained by the following formula:

[0097] Preferably, the mass change equation is as follows (5):

[0098]

[0099] where m represents the real-time mass of the aircraft, m c m represents the mass flow rate of the aircraft per second, i.e., the value of the mass reduction per second.

[0100] Step two: in a predetermined interval, small steps are used to sample the weather, the sight, the program, and the delay, and different values are used as inputs of the five-degree-of-freedom mathematical simulation model to perform batch simulation to obtain corresponding trajectory data, which includes specific values of the firing elements.

[0101] Preferably, in the present application, the five-degree-of-freedom mathematical simulation model is trained separately, i.e. trained once for each charge number, so that the corresponding model can be selected according to the actual charge number; the charge number and the initial velocity are associated with each other, and the initial velocity is determined after the charge number is determined.

[0102] In step two, the weather mainly includes ground pressure and temperature, and in the sampling of the weather, the ground pressure is 950 hPa-1050 hPa, the initial value is 950 hPa, the temperature is 15°-30°, the initial value is 15°, the sampling step of the ground pressure is 10 hPa; the sampling step of the temperature is 1°

[0103] The sampling initial value of the sight is 200 mil, the sampling step is 1 mil, and the predetermined interval of the sampling is 200 mil-400 mil; wherein, mil represents one-thousandth of an inch, equal to 0.0254 mm;

[0104] The sampling initial value of the program is 7s, the sampling step is 1s, and the predetermined interval of the sampling is 7s-16s;

[0105] The sampling initial value of the delay is 10s, the sampling step is 1s, and the predetermined interval of the sampling is 10s-20s.

[0106] Preferably, in step two, the input of the five-degree-of-freedom mathematical simulation model includes the weather, the sight, the program, the delay obtained by small-step sampling, and the charge number which takes a fixed value, which can determine the initial velocity, and the specific values are respectively 6# charge: initial velocity 325 m / s; 5# charge: initial velocity 410 m / s; 4# charge: initial velocity 495 m / s; 3# charge: initial velocity 570 m / s; 1# charge: initial velocity 740 m / s.

[0107] In step two, different values of each parameter obtained by sampling are cross-combined with each other, each combination corresponds to the input of a set of five-degree-of-freedom mathematical simulation models, i.e. each set of data can obtain a corresponding set of trajectory data. In the present application, the process of obtaining the trajectory data includes inputting the sight, the program, the delay, and the weather into the five-degree-of-freedom mathematical simulation model to obtain the corresponding expected range, and the trajectory data includes the sight, the program, the delay, the weather, and the expected range.

[0108] In the training of the neural network, the weather and the expected range are taken as the input, and the sight, the program, and the delay are taken as the output.

[0109] Step three: selecting the neural network sample from the trajectory data through sensitivity analysis.

[0110] In the step three, the control variable method is adopted for the parameters obtained by sampling, i.e. the weather, the sight, the program, and the delay, to filter out the trajectory data which has less influence on the range, so as to obtain the neural network sample.

[0111] Specifically, for the parameters obtained by sampling, their influence on the range is compared one by one; for example, for the scale data, under the condition of fixing other parameters, two groups of trajectory data corresponding to two adjacent samplings are called, the range values in the two groups of trajectory data are compared, if the range difference reaches 10 meters, it is marked as a large influence, that is, a single variable of one step can cause the range to change more than 10 meters, otherwise it is marked as a small influence; then two other groups of trajectory data corresponding to two adjacent samplings are randomly called, the range values in the two groups of trajectory data are compared, the marks are obtained, the above steps are repeated, and 10 marks are obtained, if all the 10 marks are small influences, the sampling step of the scale data is enlarged, such as from 1s to 2s, that is, half of the trajectory data is deleted, otherwise no special treatment is performed.

[0112] Step four: using Adam learning law to perform offline training on the neural network.

[0113] In step four, the neural network is a BP (Back Propagation) neural network,

[0114] Preferably, the neural network includes 5 hidden layers, and the number of neurons in each layer is 10. The input data of the neuron must be processed by the activation function before it can be output, and the activation function applied in the patent is the ReLU function: f(x) = max(0, x)

[0115] In this application, by using the ReLU activation function, the problems of gradient explosion and disappearance can be avoided.

[0116] In a preferred embodiment, the learning process of the BP neural network consists of two processes of signal forward propagation and error back propagation. When forward propagating, the input sample is transmitted from the input layer to the output layer through the processing of each hidden layer. If the actual output of the output layer does not match the expected output, the error back propagation stage is entered. When back propagating, the output is transmitted to the input layer through the hidden layer, and the error is allocated to all units of each layer to obtain the error signal of each unit, and then the weight of each unit is corrected through the error signal.

[0117] In step four, 70% of the neural network samples obtained by screening in step three are used as a training set, 15% are used as a test set, and 15% are used as a validation set. The test set is used to test whether the model is accurate after training the training set, if the test set is accurate, then the validation set is used for verification, if both are accurate, the network model can be output, wherein the error of each parameter between the single verification result and the actual result is less than 2% to determine that the verification result is accurate, and the overall accuracy of the test set or the validation set is more than 90% to be called accurate.

[0118] Preferably, in the training process, for the forward propagation, the value of the bias term b is initialized as a random number between (-1, 1) according to the input sample, the given initial weight value w is a random number between (-1, 1), the final output value is calculated, and the loss value between the output value and the actual value is calculated. The calculation formula between the input and the output is as follows (6):

[0119]

[0120] Wherein, f(·) represents an activation function, y represents an output, x represents an input, and n represents a number of layers.

[0121] In a preferred embodiment, the step four comprises the following sub-steps:

[0122] Sub-step 1: Assign a random number in an interval to each connection weight, specifically a random number between (-1, 1), set a Loss function, and the Loss function is selected as a mean square error. In this application, the mean square error and the activation function are used to update the weight and the bias; the mean square error function is shown in the following formula (7):

[0123]

[0124] Wherein, n represents a number of layers, y represents an output, represents an expected output value;

[0125] Sub-step 2: Randomly select an input sample and the corresponding expected output from the training set, and input the input sample into the neural network;

[0126] Sub-step 3: Calculate the input and output of each neuron in the hidden layer;

[0127] Sub-step 4: Calculate the partial derivative of the error function with respect to each neuron in the output layer using the expected output in the sample and the actual output of the neural network;

[0128] Sub-step 5: Use the Adam learning law to correct the weight and the bias term;

[0129] Wherein, the update formula of the parameter p (p = w, b) in the Adam learning law is as follows (8):

[0130]

[0131] Wherein, m t ' represents the gradient first moment estimation, v t ' represents the gradient second moment estimation, which is obtained by the following formula (9):

[0132]

[0133] wherein, p t represents the parameter before updating;

[0134] p t+1 represents the parameter after updating;

[0135] η represents a learning rate, and the value is 0.01;

[0136] ε represents a smoothing term, and the value is 10 to the power of-8, which prevents division by zero;

[0137] m t represents the first-order moment of the gradient,

[0138] v t represents the second-order moment of the gradient,

[0139] β1 and β2 represent hyperparameters for correcting the first-order and second-order moment estimates;

[0140] β1 t represents the momentum decay hyperparameter at the t-th time, and the value is 0.9;

[0141] β2 t represents the scaling decay hyperparameter at the t-th time, and the value is 0.999;

[0142] m t represents the first-order moment of the gradient;

[0143] v t represents the second-order moment of the gradient;

[0144] m t-1 represents the first-order moment of the gradient at the t-th time;

[0145] v t-1 represents the second-order moment of the gradient at the t-1-th time;

[0146] g t represents the gradient at the t-th time;

[0147] Substep 6: Determine the rationality of the model. If the error reaches the preset accuracy, the training is completed. If the preset accuracy is not reached, repeat steps 2-6.

[0148] The preset accuracy is 0.001.

[0149] Step five: load the trained neural network into the fire control computer of the aircraft as a model for obtaining elements;

[0150] Use the element calculation model to replace the traditional flight vehicle carried shooting table or ballistic iterative optimization module.

[0151] Step six: input the desired range, current weather information into the said data obtaining model, the firing data can be obtained, and the aircraft is launched according to the firing data, so that the aircraft hits the target finally.

[0152] The aircraft is launched according to the charge number, sight, and azimuth in the firing data, and during the flight of the aircraft, the range charge and delay start work in the firing data are also followed, the mechanical gyro is started according to the range charge time, and the fairing is thrown according to the delay time.

[0153] The application also provides a precise guidance aircraft firing data quick obtaining system based on a BP neural network, which comprises a data obtaining model loaded in an aircraft, and the firing data is obtained through the data obtaining model before the aircraft is launched, and the aircraft is launched according to the firing data.

[0154] Preferably, the data obtaining model is obtained by the following method:

[0155] Step one: given the aerodynamic parameters, establish the five-degree-of-freedom motion equation, and compile the five-degree-of-freedom mathematical simulation model;

[0156] Step two: small steps are adopted to sample the weather, sight, range charge, and delay in the predetermined interval, different values obtained by sampling are taken as the input of the five-degree-of-freedom mathematical simulation model, batch simulation is carried out, and corresponding trajectory data is obtained;

[0157] Step three: through sensitivity analysis, neural network samples are screened from the trajectory data;

[0158] Step four: the Adam learning law is adopted to carry out offline training on the neural network, so that the data obtaining model is obtained.

[0159] The application also provides an aircraft, which is provided with the precise guidance aircraft firing data quick obtaining system based on a BP neural network as described above.

[0160] Embodiment:

[0161] Generate training samples:

[0162] Step 1, given the aerodynamic parameters, establish the five-degree-of-freedom motion equation, and compile the five-degree-of-freedom mathematical simulation model; the aerodynamic parameters include the lift coefficient, the drag coefficient, and the lateral force coefficient, that is, the lift, the drag, and the lateral force can be obtained according to these data, the five-degree-of-freedom mathematical simulation model comprises the mass center motion dynamics equation, the rotation dynamics equation around the mass center, the mass center kinematics equation, the rotation kinematics equation around the mass center, the mass change equation, and the guidance and control model;

[0163] Step 2, in the predetermined interval, the weather, the sight, the range charge, and the delay are sampled with small steps, and different values are taken as the input of the five-degree-of-freedom mathematical simulation model to perform batch simulation to obtain corresponding trajectory data.

[0164] wherein the initial value of the air pressure is 950 hPa, the initial value of the temperature is 15°, the initial value of the sight is 200 mil, the initial value of the range charge is 7 s, and the initial value of the delay is 10 s; the step length of the air pressure is 10 hPa, the step length of the temperature is 1°, the step length of the sight is 1 mil, the step length of the range charge is 1 s, and the step length of the delay is 1 s; the predetermined interval of the air pressure is 950 hPa-1050 hPa, the predetermined interval of the temperature is 15°-30°, the predetermined interval of the sight is 200 mil-400 mil, the predetermined interval of the range charge is 7 s-16 s, and the predetermined interval of the delay is 10 s-20 s; the number of individual sampling of the air pressure is 10, the number of individual sampling of the temperature is 15, the number of individual sampling of the sight is 200, the number of individual sampling of the range charge is 9, and the number of individual sampling of the delay is 10;

[0165] Step 3, sensitivity analysis is performed on the obtained trajectory data, and data that causes the range to be less than 10 meters is removed, and finally 11562 neural network samples are obtained;

[0166] Training the network model:

[0167] Step 4, the neural network is trained offline using the Adam learning law,

[0168] wherein 70% of the obtained neural network samples in step 3 are selected as the training set, 15% are selected as the test set, and 15% are selected as the verification set;

[0169] The update formula of the parameter ρ (ρ=w,b) in the Adam learning law is as follows (8):

[0170]

[0171] wherein m t ′ represents the first-order moment estimation of the gradient, v t ′ represents the second-order moment estimation of the gradient, and is obtained by the following formula (9):

[0172]

[0173] When the error reaches 0.001, the training is stopped;

[0174] Aircraft target shooting experiment loaded with the network model:

[0175] Step 5, load the trained neural network into the fire control computer of the three aircrafts respectively, wherein the expected range of the first aircraft is 5000m, the expected range of the second aircraft is 5500m, and the expected range of the third aircraft is 6000m;

[0176] Step 6, input the expected range and the current meteorological information into the data obtaining model to obtain the firing data, and the three aircrafts fire according to the obtained firing data, and the actual hitting effects of the three aircrafts are shown in Figure 2 、 Figure 3 and Figure 4 .

[0177] Wherein, the current meteorological information includes ground air pressure 993.6hPa, temperature 15.8℃;

[0178] The firing data of the three aircrafts are shown in Table 1:

[0179]

[0180] According to the range coverage of each charge number, the No.5 charge is selected in the three experiments, and the corresponding initial speed is 326m / s.

[0181] From Figure 2 、 Figure 3 and Figure 4 , it can be seen that the three aircrafts can accurately hit the target with small miss distance when the firing data obtained according to the data obtaining model installed on the aircraft is fired, which fully illustrates the correctness of the data obtaining model.

[0182] The above describes the present application in combination with the preferred embodiments, but these embodiments are only exemplary and are used for illustration only. On this basis, various substitutions and improvements can be made to the present application, and these all fall within the protection scope of the present application.

Claims

1. A method for quickly obtaining firing data of a precision-guided aircraft based on a BP neural network, characterized in that, The method comprises the following steps: Step one: given the aerodynamic parameters, establish the five-degree-of-freedom motion equation, and compile the five-degree-of-freedom mathematical simulation model; Step two: using small steps, sampling the weather, the sight, the range charge, and the delay in the predetermined interval, using the different values obtained by sampling as the input of the five-degree-of-freedom mathematical simulation model, performing batch simulation to obtain the corresponding trajectory data, the weather being the ground pressure and temperature values; Step three: through sensitivity analysis, filtering out the trajectory data with less influence on the range from the trajectory data to obtain the neural network sample, the sensitivity analysis being performed by using the control variable method; Step four: using the Adam learning law, performing offline training on the neural network; Step five: loading the trained neural network into the fire control computer of the aircraft as the data obtaining model; Step six: inputting the expected range and the current weather information into the data obtaining model to obtain the firing data, and launching the aircraft according to the firing data.

2. The method according to claim 1, wherein in step one, the five-degree-of-freedom motion equation comprises a center-of-mass motion dynamics equation, a center-of-mass rotation dynamics equation, a center-of-mass kinematics equation, a center-of-mass rotation kinematics equation, and a mass change equation; the center-of-mass motion dynamics equation is as follows (1): the center-of-mass rotation dynamics equation is as follows (2): wherein m represents the real-time mass of the aircraft, V m represents the flight speed of the aircraft, P represents the engine thrust on the aircraft, X represents the aerodynamic resistance received by the aircraft, Y represents the lift received by the aircraft, and Z represents the lateral force received by the aircraft; a represents the angle of attack of the aircraft, b represents the sideslip angle of the aircraft, q represents the ballistic angle of inclination of the aircraft, g v represents the speed inclination angle of the aircraft, and y v represents the ballistic deflection angle of the aircraft. the center-of-mass kinematics equation is as follows (3): where J y represents the moment of inertia of the aircraft about the Y axis in body coordinates; J z represents the moment of inertia of the aircraft about the Z axis in body coordinates; M y represents the projection of the moment of the external forces on the mass center of the aircraft on the Y axis; M z represents the projection of the moment of the external forces on the mass center of the aircraft on the Z axis; w y represents the yaw angular velocity of the rotation of the aircraft about the mass center; w z represents the roll angular velocity of the rotation of the aircraft about the mass center; where (x, y, z) represents the position coordinates of the center of mass of the aircraft in the inertial coordinate system; the center-of-mass rotation kinematics equation is as follows (4): the mass change equation is as follows (5): wherein denotes the pitch angle of the aircraft; denotes the yaw angle of the aircraft; denotes the roll angle of the aircraft; 3. The method according to claim 1, wherein in step two, the initial value of the pressure in the weather is 950 hPa, the sampling step is 10 hPa, the initial value of the temperature is 15°, and the sampling step is 1°; where m represents the real-time mass of the aircraft, m c represents the mass flow rate of the aircraft. the initial value of the sampling of the sight is 200 mil, and the sampling step is 1 mil; the initial value of the sampling of the range charge is 7 s, and the sampling step is 1 s; the initial value of the sampling of the delay is 10 s, and the sampling step is 1 s.

4. The method according to claim 1, wherein in step three, the control variable method is used to filter out the trajectory data with less influence on the range from the parameters obtained by sampling to obtain the neural network sample.

5. The method according to claim 4, wherein in the neural network sample, the expected range and the weather in the trajectory data are used as the input, and the sight, the range charge, and the delay in the trajectory data are used as the output.

6. The method according to claim 1, wherein in step four, the neural network is a BP neural network, the neural network comprises 5 layers of hidden layers, and the number of neurons in each layer is 10. ​ ​ ​ ​ ​ 7. The method of claim 1, wherein, in step four, 70% of the neural network samples obtained in step three are taken as a training set, 15% are taken as a test set, and 15% are taken as a validation set.

8. The method of claim 1, wherein, the step four comprises the following sub-steps: sub-step 1, assigning a random number in an interval to each connection weight, setting a Loss function, sub-step 2, randomly selecting an input sample and a corresponding expected output from the training set, and inputting the input sample into the neural network; sub-step 3, calculating the input and output of each neuron in the hidden layer; sub-step 4, calculating the partial derivative of the error function with respect to each neuron in the output layer using the expected output in the sample and the actual output of the neural network; sub-step 5, correcting the weight and bias using the Adam learning law; sub-step 6, judging the rationality of the model, if the error reaches the preset accuracy, the training is ended, if the preset accuracy is not reached, repeating steps 2-6.

9. A system for quickly obtaining firing data of a precision-guided aircraft based on a BP neural network, characterized in that, The system comprises a firing data obtaining model loaded in the aircraft, before the aircraft is launched, the firing data is obtained through the firing data obtaining model, and the aircraft is launched according to the firing data; the firing data obtaining model is obtained by the following method: step one: given the aerodynamic parameters, establish a five-degree-of-freedom motion equation, and compile a five-degree-of-freedom mathematical simulation model; step two: using small steps, sampling the weather, sight, range, delay in the predetermined interval, taking the different values obtained by sampling as the input of the five-degree-of-freedom mathematical simulation model, performing batch simulation to obtain corresponding trajectory data, the weather is the ground pressure and temperature value; step three: through sensitivity analysis, filter out the trajectory data with less influence on the range from the trajectory data to obtain the neural network sample, the sensitivity analysis is to filter out the trajectory data with less influence on the range by using the control variable method; step four: using the Adam learning law, the neural network is trained offline, thereby obtaining the firing data obtaining model.

10. An aircraft, characterized in that The aircraft is provided with the firing data obtaining system of claim 9.

Images