Aircraft remaining flight time prediction method based on deep learning

Abstract

The application discloses a kind of aircraft residual flight time prediction methods based on deep learning, the method is for aircraft using specific guidance algorithm, using the current speed information of aircraft itself, the relative position information of aircraft and target, the residual flight time is predicted by deep neural network.The aircraft residual flight time prediction method based on deep learning provided by the application, the input state selected is current speed, current speed direction, current height, current lateral position, mapping relationship is reasonable, the feasibility of using deep learning to fit this mapping relationship is high;High prediction accuracy, less calculation, can be applied to online prediction scene, ensure that aircraft efficiently realizes cooperative penetration.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of aircraft residual flight time prediction, and particularly relates to an aircraft residual flight time prediction method based on deep learning. BACKGROUND

[0002] Aircrafts (such as missiles) are the main force to attack important strategic targets, but in modern wars, the enemy's defense countermeasures are various, especially the ground or shipborne platforms have long-range interception weapons and close-in defense weapons, which have caused great threat to aircrafts.

[0003] Multi-aircraft cooperative penetration can saturate the enemy's defense system and increase the probability of penetration success, and ensuring that multiple aircrafts can reach the target at the same time has become a key problem to realize cooperative penetration. Accurate residual flight time prediction is the basis of accurate arrival time control. If the residual flight time cannot be obtained directly or indirectly, the arrival time control problem can only be solved in an open loop, and the control accuracy of open loop control is generally less than that of closed loop control.

[0004] At present, the online prediction problem of aircraft residual flight time mainly adopts an analytical method, and is based on the assumption of constant flight speed, and the prediction accuracy is poor. Although the iterative calculation of the differential equation can improve the prediction accuracy, the calculation amount is large, and it is difficult to realize online prediction.

[0005] Therefore, it is necessary to provide a method for predicting the residual flight time of an aircraft, which has high prediction accuracy, small calculation amount and can be applied to online prediction scenarios. SUMMARY

[0006] In order to overcome the above problems, the present inventors have made intensive research and designed a method for predicting the residual flight time of an aircraft based on deep learning. The method uses the current speed information of the aircraft itself, the relative position information of the aircraft and the target, and predicts the residual flight time through a deep neural network for aircrafts using a specific guidance algorithm, has good prediction ability, high prediction accuracy, small calculation amount, and can be applied to online prediction scenarios, thereby completing the present application.

[0007] Specifically, the present application aims to provide the following aspects:

[0008] In a first aspect, a method for predicting the residual flight time of an aircraft based on deep learning is provided, characterized in that the method comprises the following steps:

[0009] Step 1: Collect flight data and train a neural network model;

[0010] Step 2: When the aircraft is flying, use the neural network model to predict the residual flight time in real time.

[0011] wherein step 1 comprises the following sub-steps:

[0012] Step 1-1, the aircraft runs a flight simulation program to collect training sample data during flight;

[0013] Step 1-2, a neural network is established, and the neural network is trained using the collected training sample data;

[0014] Step 1-3, the trained neural network is tested.

[0015] In step 1-1,

[0016] The obtaining of the training sample comprises the following steps:

[0017] Before each run of the simulation program, the initial state of the aircraft is randomly modified.

[0018] The initial state includes initial speed, initial speed direction, initial lateral position, and initial height.

[0019] In step 1-1, the collected training sample data includes the speed of the aircraft, the speed direction, the spatial position, and the remaining flight time.

[0020] In step 1-2, the established neural network is a deep neural network, which has multiple hidden layers of neural networks, and the function of the hidden layer is as follows:

[0021] C j = f(∑ i (w ij × R i + b ij )

[0022] Wherein, R represents the input of the hidden layer, C represents the output of the hidden layer, w represents the weight of the hidden layer, b represents the offset of the hidden layer, subscript j represents the number of hidden layers, and subscript i represents the i-th neuron in the hidden layer j.

[0023] In a second aspect, a computer readable storage medium is provided, which stores a deep learning-based aircraft remaining flight time prediction program, and the program, when executed by a processor, causes the processor to perform the steps of the deep learning-based aircraft remaining flight time prediction method.

[0024] In a third aspect, a computer device is provided, which includes a memory and a processor, and the memory stores a deep learning-based aircraft remaining flight time prediction program, and the program, when executed by the processor, causes the processor to perform the steps of the deep learning-based aircraft remaining flight time prediction method.

[0025] The beneficial effects of the present application include:

[0026] (1) The aircraft residual flight time prediction method based on deep learning provided by the present application selects the input state as the current speed, the current speed direction, the current height and the current lateral position, the mapping relationship is reasonable, and the feasibility of using deep learning to fit the mapping relationship is high.

[0027] (2) The aircraft residual flight time prediction method based on deep learning provided by the present application has high prediction accuracy, less calculation amount, and can be applied to online prediction scenarios to ensure efficient implementation of aircraft cooperative penetration. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The deep neural network structure diagram according to a preferred embodiment of the present application is shown;

[0029] Figure 2 The comparison diagram of the predicted value and the true value of the deep neural network according to a preferred embodiment of the present application is shown;

[0030] Figure 3 The prediction error diagram of the deep neural network according to a preferred embodiment of the present application is shown;

[0031] Figures 4 to 7 The effect comparison diagram of the residual flight time predicted by the method of the present application and the existing analytical method in Embodiment 1 of the present application is shown, wherein, Figure 4 and 5 The comparison results of the predicted value and the true value of the method of the present application, the prediction error are shown respectively; Figure 6 and Figure 7 The comparison results of the predicted value and the true value of the existing analytical method, the prediction error are shown respectively. DETAILED DESCRIPTION

[0032] The present application will be further described in detail through preferred embodiments and examples. Through these descriptions, the characteristics and advantages of the present application will become clearer and more explicit.

[0033] Herein, the word "exemplary" means "serving as an example, embodiment, or illustration". Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments.

[0034] The present inventors have found through research that the residual flight time of an aircraft is determined by the residual flight path and the flight speed, and the flight parameters are as follows:

[0035]

[0036]

[0037]

[0038] where t go represents the remaining flight time, L go represents the remaining flight path length, v(t) represents the flight speed, x represents the lateral spatial position of the aircraft, θ represents the angle between the aircraft speed vector and the horizontal plane, the subscript 0 represents the initial value, and the subscript f represents the terminal value.

[0039] The present inventors found that the problem of predicting the remaining flight time of an aircraft can be converted into the problem of predicting the future speed magnitude and speed direction at t0. However, the speed magnitude and direction change over time and are affected by various factors such as air dynamics, aircraft weight, air density, etc., and the aircraft itself also exerts some control force to change its flight state.

[0040] The aerodynamic model of the aircraft is as follows:

[0041] q(t)=0.5ρv(t) 2 ;

[0042] F L =c L qS;

[0043] F D =c D qS;

[0044] F G =mg;

[0045] where q represents the dynamic pressure, F L represents the lift, F D represents the drag, F G represents the gravity, c L represents the lift coefficient, c D represents the drag coefficient, and the lift coefficient and the drag coefficient change with the change of the angle of attack, the Mach number, and the rudder deflection angle.

[0046] In the above aerodynamic model, when calculating the dynamic pressure q, the air density needs to be considered, but the atmospheric density at different altitudes is different, i.e. ρ=g(y0); the lift coefficient, the drag coefficient, and the air density are generally obtained by table lookup or Lagrange interpolation.

[0047] Taking the proportional guidance method as an example, the overload command is:

[0048]

[0049] Considering the time-varying nature of the parameters, the above formula can be rewritten as:

[0050]

[0051] Then the flight parameters are obtained as:

[0052] t go =f(v,θ,x0,y0)

[0053] When predicting the remaining flight time at the current position (x0, y0) of the aircraft, all the speeds on the remaining flight path need to be solved. However, the future speed cannot be directly expressed in an analytical form, and the calculation amount of solving the integral of the dynamic differential equation is too large to be used in real time. In order to circumvent this problem, the remaining flight time estimation algorithm is designed based on the constant speed assumption, but these algorithms cannot be practically applied to the aircraft.

[0054] Therefore, in the present application, the relationship between the current speed information of the aircraft, the distance between the aircraft and the target and the arrival time of the aircraft is preferably established by using a deep learning method.

[0055] The present application provides a deep learning-based aircraft remaining flight time prediction method, which comprises the following steps:

[0056] Step 1, collecting flight data to train a neural network model;

[0057] Step 2, when the aircraft is flying, the neural network model is used to predict the remaining flight time in real time.

[0058] The deep learning-based aircraft remaining flight time prediction method is further described as follows:

[0059] Step 1, collecting flight data to train a neural network model.

[0060] The present inventors have found that the deep neural network constructed by the deep learning technology can fit a complex mapping relationship through learning and training, and the trained network has a relatively small calculation amount and low complexity when actually used, and is suitable for the prediction of the remaining flight time of the aircraft.

[0061] In the present application, the aircraft can be a missile, a drone or the like, and is preferably a missile.

[0062] Preferably, step 1 comprises the following sub-steps:

[0063] Step 1-1, running an aircraft flight simulation program to collect training sample data during flight.

[0064] According to a preferred embodiment of the present application, the flight simulation program of the aircraft is obtained by measuring the aerodynamic parameters of the aircraft,

[0065] The aerodynamic parameters include the lift coefficient, the induced drag coefficient and the zero-lift drag coefficient.

[0066] In the present application, when the aircraft structure is determined, the aerodynamic parameters can also be determined. In actual flight, the aerodynamic parameters are generally related to the Mach number, the attack angle and the rudder deflection angle of the aircraft.

[0067] Preferably, the Mach number is related to the sound speed at the current height of the aircraft, and is obtained from the current speed information / sound speed of the aircraft;

[0068] Wherein, the program in the aircraft includes a navigation module, a guidance module and a control module, and the height information and the current speed information of the aircraft are obtained by the navigation module;

[0069] The attack angle represents the direction of the air flow, and is obtained by the navigation module of the aircraft;

[0070] The rudder deflection angle is obtained by the control module of the aircraft.

[0071] Wherein, the sound speed is obtained by interpolation of the pre-determined atmospheric data, and then the Mach number is obtained.

[0072] More preferably, the aerodynamic parameters corresponding to the current Mach number, attack angle and rudder deflection angle are obtained by wind tunnel experiment and interpolation calculation.

[0073] In the present application, the state of the aircraft at the next time is obtained according to the following differential equation of aerodynamics of the aircraft:

[0074]

[0075] Wherein, v represents the speed, θ represents the angle between the speed vector of the aircraft and the horizontal plane, x represents the lateral spatial position of the aircraft, y represents the longitudinal spatial position of the aircraft, m represents the weight of the aircraft, P represents the engine thrust, α represents the attack angle of the aircraft, X represents the drag, L represents the lift, m c represents the fuel consumption per unit time;

[0076] The relationship between the lift, the drag and the aerodynamic parameters is:

[0077] X=(c d0 +c d )qS

[0078] L=c L qS

[0079] Wherein, c d0 represents the zero-lift-drag coefficient, c d represents the induced drag coefficient, c L represents the lift coefficient, q is the dynamic pressure, and S is the reference area of the aircraft.

[0080] Preferably, the state includes position information, speed information and attitude information.

[0081] wherein the simulation program solves the differential equations by the fourth-order Runge-Kutta method, iteratively calculates the state of the aircraft using a fixed step size until the aircraft completes the mission.

[0082] According to a preferred embodiment of the present application, the obtaining of the training samples comprises the following steps:

[0083] Before each run of the simulation program, the initial state of the aircraft is randomly modified:

[0084] The initial state comprises an initial speed, an initial speed direction, an initial lateral position and an initial altitude.

[0085] Since the present application employs deep learning, which is implemented based on a large amount of data, a sufficient number of training samples need to be collected, and a large number of training samples can be obtained by running the flight simulation program of the aircraft multiple times. The present inventors have found that modifying the initial state of the aircraft before each run of the simulation program can increase the randomness of the samples and improve the training effect of deep learning.

[0086] In a further preferred embodiment, during the running of the simulation program, the simulation program iteratively updates according to a fixed step size, and after each iteration, a set of data of the aircraft is obtained,

[0087] The data comprises a current speed, a current speed direction, a current lateral position, a current altitude and a current time.

[0088] In a still further preferred embodiment, after the aircraft completes the mission, the total flight time is subtracted from the current time in each set of data to obtain the remaining flight time corresponding to each set of data.

[0089] Preferably, the collected training sample data comprises the speed size, speed direction, spatial position and remaining flight time of the aircraft.

[0090] The input of the training sample is the speed size, speed direction and spatial position, and the output is the remaining flight time.

[0091] Step 1-2, establish a neural network, and train the neural network using the collected training sample data.

[0092] According to a preferred embodiment of the present application, the established neural network is a deep neural network, which is a neural network having multiple hidden layers, and the function of the hidden layer is as follows:

[0093] C j = f(∑ i (w ij × R i + b ij ))

[0094] where R represents the input of the hidden layer, C represents the output of the hidden layer, w represents the weight of the hidden layer, b represents the offset of the hidden layer, subscript j represents the number of the hidden layer, and subscript i represents the i-th neuron in the hidden layer j.

[0095] In a further preferred embodiment, each neuron has a non-linear activation function f, which is shown as follows:

[0096] f(x) = max(0, x).

[0097] In a still further preferred embodiment, the established deep neural network has 3 hidden layers, each of which has 100 neurons, as shown in Figure 1 .

[0098] Preferably, the loss function of the deep neural network is set to the minimum mean square error, i.e. the mean square error of the estimated value and the actual value, as shown in the following formula:

[0099]

[0100] where L is the loss function, N is the number of samples, and i is the sample number.

[0101] First, the gradient g t of each parameter to the objective function is calculated:

[0102]

[0103] where β is the parameter to be trained, including the weight w and the offset b, and the initial learning rate r is set before training, and the parameter update is performed according to the following formula:

[0104] β i+1 = β i - rg t

[0105] In the present application, the training samples are preferably divided into two parts, 80% of which are used as the training set to train the deep neural network, and 20% of which are used as the test set to test the performance of the deep neural network.

[0106] According to a preferred embodiment of the present application, in each training cycle, a certain number of samples are randomly selected from the training set to train the neural network.

[0107] Preferably, 1000 samples are selected in each training cycle, and the training cycle is set to 100,000 times.

[0108] The inventors find that, due to a large amount of training sample data, directly training using all the training samples will result in a large amount of calculation, therefore, preferably, in each training period, a certain number of samples are randomly selected from the training set to train the neural network.

[0109] The trained neural network model converges.

[0110] Step 1-3, testing the trained neural network.

[0111] The trained neural network is tested by using the test set to determine whether the prediction error meets the standard.

[0112] According to a preferred embodiment of the present application, the threshold of the prediction error is set according to the application scenario of the prediction error, and the influence of the prediction error on the guidance control accuracy of the aircraft is fully considered,

[0113] Preferably, it is set to be less than 1s in the experimental stage.

[0114] The prediction error is the maximum error of the complete coincidence of the predicted value and the true value, when it is greater than the threshold, it is determined that it does not meet the standard, and when it is less than or equal to the threshold, it is determined that it meets the standard.

[0115] Step 2, when the aircraft is flying, the neural network model is used to predict the remaining flight time in real time.

[0116] In the present application, the trained deep neural network is saved in two parts of model structure and model parameters, and in actual use, the model structure is first loaded on the embedded computer of the aircraft, and then the model parameters are loaded in the model structure, and the model loading is completed.

[0117] According to a preferred embodiment of the present application, when the aircraft is flying, according to the real-time flight speed, flight speed direction, height of the aircraft, and the lateral relative distance between the aircraft and the target, the trained deep neural network model is used to obtain the predicted value of the remaining flight time in real time.

[0118] The aircraft remaining flight time prediction method based on deep learning provided by the present application selects the input state of speed, speed direction and spatial position, the designed mapping relationship is reasonable, and it is feasible to use deep learning to fit this mapping relationship. Moreover, the method has high prediction accuracy and less calculation amount, and can be applied to online prediction scenarios.

[0119] The present application also provides a computer readable storage medium storing a program for predicting the remaining flight time of an aircraft based on deep learning, wherein the program is executed by a processor to make the processor execute the steps of the aircraft remaining flight time prediction method based on deep learning.

[0120] The deep learning-based aircraft residual flight time prediction method described in the present application can be realized by means of software and a necessary general hardware platform, the software being stored in a computer readable storage medium (including ROM / RAM, a magnetic disk, an optical disk), and including a plurality of instructions for causing a terminal device (which can be a mobile phone, a computer, a server, a network device, etc.) to execute the method described in the present application.

[0121] The present application also provides a computer device comprising a memory and a processor, the memory storing a deep learning-based aircraft residual flight time prediction program, the program being executed by the processor to cause the processor to execute the steps of the deep learning-based aircraft residual flight time prediction method.

[0122] Embodiments

[0123] The present application is further described below through specific examples, but these examples are merely exemplary and do not constitute any limitation on the scope of protection of the present application.

[0124] Example 1

[0125] In the present embodiment, the selected aircraft is a missile, and its aerodynamic parameters are shown in Table 1:

[0126] Table 1

[0127]

[0128]

[0129] The aerodynamic differential equation is:

[0130]

[0131] The fixed step size used by the simulation program is 0.01s;

[0132] The simulation program runs 1000 times, and each run generates about 2000 groups of data, a total of about 2 million groups of data.

[0133] The deep neural network constructed uses 3 hidden layers, each with 100 neurons, and the activation function of each neuron is: f(x) = max(0, x);

[0134] The loss function is:

[0135] During the training process, 1000 samples are selected for each training cycle, and the training cycle is set to 100,000 times;

[0136] After the training is completed, the converged neural network is tested, and the results are as follows Figure 2and 3 As shown in Figure 2 and 3 It can be seen that the predicted value of the residual flight time of the neural network trained in the embodiment is completely consistent with the true value, and the maximum error is not greater than 0.6s, indicating that the deep neural network can well fit the mapping relationship between the flight state of the missile and the residual flight time.

[0137] Further, the simulation effect of the method of the present application and the existing residual flight time prediction method (approximate residual flight time is calculated by an analytical form, specifically as described in the document "In-Soo Jeon, Jin-Ik Lee, and Min-Jea Tahk. Impact-time-control guidance law for anti-ship missiles. IEEE Transactions on Control Systems Technology 2006, 14(2): 260-266.") is compared, and the result is as shown in Figures 4 to 7 .

[0138] As shown in Figures 4 to 7 It can be seen that the prediction method of the present application has good prediction effect, the predicted value is basically consistent with the actual value, and the prediction error is less than 0.3s; the existing analytical method has a large prediction error when the residual flight time is large, and the maximum prediction error can reach 60s, and the prediction effect is poor.

[0139] The present application has been described in detail above with reference to specific embodiments and exemplary examples, but these descriptions cannot be understood as limiting the present application. Those skilled in the art understand that the technical solutions and embodiments of the present application can be variously replaced, modified or improved without deviating from the spirit and scope of the present application, and these all fall within the scope of the present application.

Claims

1. A deep learning-based aircraft remaining flight time prediction method, characterized by, The method comprises the following steps: Step 1, collecting flight data, training a neural network model; Step 2, using the neural network model to predict the remaining flight time in real time when the aircraft is flying; Step 1 comprises the following sub-steps: Step 1-1, running an aircraft flight simulation program to collect training sample data during flight; Step 1-2, establishing a neural network and training the neural network using the collected training sample data; Step 1-3, testing the trained neural network; In step 1-1, the aircraft flight simulation program is obtained by measuring the aerodynamic parameters of the aircraft, which are related to the Mach number, angle of attack and rudder deflection angle of the aircraft, and the Mach number is obtained from the current speed information / sound speed of the aircraft; The height information and current speed information of the aircraft are obtained from the navigation module, and the sound speed is obtained by interpolating the pre-measured atmospheric data; The aerodynamic parameters corresponding to the current Mach number, angle of attack and rudder deflection angle are obtained by wind tunnel experiment and interpolation calculation; According to the following differential equation of the aerodynamics of the aircraft, the state of the aircraft at the next time is obtained, which includes position information, speed information and attitude information: where v represents the speed magnitude, Θ represents the angle between the aircraft speed vector and the horizontal plane, x represents the lateral spatial position of the aircraft, y represents the longitudinal spatial position of the aircraft, m represents the weight of the aircraft, P represents the engine thrust, a represents the angle of attack of the aircraft, X represents the drag, L represents the lift, m c represents the fuel consumption per unit of time; The relationship between lift, drag and aerodynamic parameters is: X = (c d0 + c d )qS L = c L qS where c d0 represents the zero-lift drag coefficient, c d represents the induced drag coefficient, c L represents the lift coefficient, q is the dynamic pressure, and S is the reference area of the aircraft; In step 1-1, The obtaining of the training sample comprises the following steps: Before running the simulation program each time, the initial state of the aircraft is randomly modified; The initial state includes initial speed, initial speed direction, initial lateral position and initial height; In step 1-1, the collected training sample data includes the speed, speed direction, spatial position and remaining flight time of the aircraft; In step 2, the remaining flight time is predicted in real time by the neural network model according to the real-time flight speed, flight speed direction, height of the aircraft, and the lateral relative distance between the aircraft and the target when the aircraft is flying.

2. The deep learning-based aircraft time-to-fatigue prediction method of claim 1, wherein, In step 1-2, the established neural network is a deep neural network, which has multiple hidden layers, and the function of the hidden layer is as follows: Wherein, R represents the input of the hidden layer, C represents the output of the hidden layer, w represents the weight of the hidden layer, b represents the offset of the hidden layer, subscript j represents the number of hidden layers, and subscript i represents the i th neuron in the hidden layer j.

3. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a deep learning based aircraft remaining flight time prediction program, and the program is executed by the processor to make the processor execute the steps of the deep learning based aircraft remaining flight time prediction method of claim 1 or 2.

4. A computer device, comprising: The computer device comprises a memory and a processor, and the memory stores a deep learning based aircraft remaining flight time prediction program, and the program is executed by the processor to make the processor execute the steps of the deep learning based aircraft remaining flight time prediction method of claim 1 or 2.

Images