Aircraft guidance and control method for multiple target domain tasks

Abstract

The application discloses a kind of aircraft guidance control methods for multiple target domain tasks, in which trained domain adversarial neural network is pre-filled on aircraft, and the domain adversarial neural network needs to specify different target domain tasks during training, so that the adversarial neural network can give guidance bias for each group of target domain tasks after receiving the state vector of the aircraft, and the aircraft only needs to determine and input the desired target domain task before launch, so as to find the corresponding guidance bias, and then obtain the guidance instruction, and work with rudder, control the aircraft to hit the target accurately, so as to complete the application.

Description

TECHNICAL FIELD

[0001] The present application relates to a control method of an aircraft, in particular to a kind of aircraft guidance control method for multiple target domain tasks. BACKGROUND

[0002] Guidance control is the core influencing factor of whether aircraft is accurate or not, and the quality of guidance instruction will directly affect the accuracy of aircraft, and even determine the success or failure of the task.Currently widely used method is the mature analytical method and numerical iteration method and other traditional guidance algorithms, can guarantee good control effect and high precision within foreseeable range.But with the continuous development of aerospace field, the complexity of task is increasingly improved, and people's requirements for the precision, robustness and real-time of guidance control system are also higher and higher, and the limitations of traditional method gradually appear.

[0003] Deep learning guidance method as the main representative of data-based algorithm, in recent years, in the guidance law design field, it is concerned, it regards the guidance problem as a "black box", uses neural network to learn the mapping relationship between the input and output of airborne computer.However, for traditional deep learning guidance algorithm, with the change of guidance task, the guidance model trained in the original task environment cannot be directly applied to the new task environment, and the guidance model adapted to the current task condition needs a large amount of trajectory data for retraining, but it is very difficult to obtain new training data in strict task;

[0004] In addition, in the existing scheme, each aircraft corresponds to a single task target / environment, if the task target / environment of the aircraft is determined to be different from the preset before launching, it needs to be recharged with new task target / environment, which leads to the extension of the launch preparation time of the aircraft, which may cause the loss of the best launch opportunity;Based on this, how to shorten the launch preparation time and improve the adaptability of the aircraft to different task targets / environment is the main problem to be solved at present.

[0005] When two or more sets of control systems for different task targets / environment are filled in the aircraft at the same time, there are a series of problems such as information redundancy, low running efficiency, high production cost and so on.

[0006] Based on the above problems, the present application has made in-depth research on the aircraft guidance control method, and in view of the shortcomings of existing research institutes, a kind of aircraft guidance control method for multiple target domain tasks is proposed to solve the above problems. SUMMARY

[0007] In order to overcome the above problems, the present inventors have made intensive studies and designed an aircraft guidance control method for multiple target domain tasks, in which a trained domain adversarial neural network is pre-installed on the aircraft, the domain adversarial neural network needs to specify different target domain tasks during training, so that the adversarial neural network can give a guidance bias item for each group of target domain tasks after receiving the state vector of the aircraft, and the aircraft only needs to determine and input the desired target domain task before launch to find the corresponding guidance bias item, and then obtain the guidance instruction, and work the rudder to control the aircraft to accurately hit the target, thereby completing the present application.

[0008] Specifically, the present application aims to provide an aircraft guidance control method for multiple target domain tasks, in which a trained domain adversarial neural network is pre-installed in the aircraft;

[0009] The method comprises:

[0010] S1, during the flight of the aircraft, obtaining the state vector of the aircraft at a predetermined frequency, and outputting multiple guidance bias items each matched to different target domain tasks through the domain adversarial neural network,

[0011] S2, selecting the corresponding guidance bias item based on the preset target domain task,

[0012] S3, obtaining the guidance instruction in real time based on the guidance bias item,

[0013] S4, the aircraft controls the rudder based on the guidance instruction and finally hits the target.

[0014] Wherein, the flight trajectory discrete point data of the aircraft under different initial states obtained by GPOPS-II is collected to obtain a data set for training the domain adversarial neural network;

[0015] According to different constraint conditions and different optimization objectives, the data set is divided into source domain data and multiple target domain data;

[0016] Preferably, the target domain data includes target domain I data and target domain II data.

[0017] Wherein, the domain adversarial neural network comprises a feature extractor, a domain discriminator, a source domain predictor, a target domain I predictor and a target domain II predictor.

[0018] Wherein, the training process of the domain adversarial neural network comprises:

[0019] Step 1, extracting the common features of the input source domain data, target domain I predictor and target domain II predictor through the feature extractor;

[0020] Step 2, forward propagate the obtained common features to the domain discriminator G d , obtain the domain classification loss of the domain discriminator and the classification loss is back propagated to the feature extractor in the form of gradient ascent through the gradient reversal layer;

[0021] Step 3, forward propagate the obtained common features to the source domain predictor G s , target domain I predictor target domain II predictor obtain the prediction value of the source domain predictor, the prediction value of the target domain I predictor, the prediction value of the target domain II predictor; further obtain the loss of the source domain predictor the loss of the target domain I predictor the loss of the target domain II predictor

[0022] Step 4, obtain the total objective function of the domain adversarial neural network based on minimizing the total objective function of the domain adversarial neural network feedback the trainable model parameters;

[0023] Step 5, stop training when the domain adversarial neural network reaches the maximum epoch.

[0024] wherein the total objective function of the domain adversarial neural network is obtained by the following formula (I):

[0025]

[0026] wherein, represents the classification loss of the source domain predictor;

[0027] represents the classification loss of the target domain I predictor;

[0028] represents the classification loss of the target domain II predictor;

[0029] represents the classification loss of the domain discriminator;

[0030] μ represents a hyperparameter for balancing the loss of the target domain I predictor;

[0031] ν represents a hyperparameter for balancing the loss of the target domain II predictor;

[0032] γ represents a hyperparameter for balancing the loss of the domain discriminator.

[0033] wherein the classification loss of the domain discriminator is obtained by the following formula (II):​

[0034]

[0035] wherein, when χ α = β, when χ α ≠ β,

[0036] p ic represents the output of the domain discriminator;

[0037] In the classification process of the domain discriminator, the source domain data is given a label χ α = 1, the target domain I data is given a label χ α = 2, and the target domain II data is given a label χ α = 3.

[0038] N represents the number of label categories, and takes a value of 3,

[0039] β represents different task labels, and takes a value of 1, 2, or 3.

[0040] wherein, the classification loss of the source domain predictor is obtained by the following formula (three):

[0041]

[0042] the classification loss of the target domain I predictor is obtained by the following formula (four):

[0043]

[0044] the classification loss of the target domain II predictor is obtained by the following formula (five):

[0045]

[0046] wherein, i represents the i-th data sample;

[0047] n represents the number of sample of the source domain data;

[0048] m1 represents the number of sample of the target domain I data;

[0049] m2 represents the number of sample of the target domain II data;

[0050] represents the predicted value of the source domain predictor for the i-th source domain data sample;

[0051] represents the label in the i-th source domain data sample;

[0052] represents a predicted value of the target domain I predictor for the i-th target domain I data sample;

[0053] represents a label in the i-th target domain I data sample.

[0054] represents a predicted value of the target domain II predictor for the i-th target domain II data sample;

[0055] represents a label in the i-th target domain II data sample.

[0056] where p ic is obtained by the following equation (six):

[0057] p ic = G d (G f (x i ; θ f ); θ d ) (six)

[0058] The is obtained by the following equation (seven):

[0059]

[0060] The is obtained by the following equation (eight):

[0061]

[0062] The is obtained by the following equation (nine):

[0063]

[0064] where G f (x i ; θ f ) represents a feature extractor;

[0065] x i represents the i-th sample input to the domain adversarial neural network;

[0066] θ f represents a model parameter of the feature extractor that can be trained;

[0067] θ d represents a model parameter of the domain discriminator that can be trained;

[0068] θ s represents a model parameter of the source domain predictor that can be trained;

[0069] denotes target domain I predictor trainable model parameters;

[0070] denotes target domain II predictor trainable model parameters.

[0071] wherein the value of the hyperparameter μ balancing the target domain I predictor loss is obtained by the following equation (X):

[0072]

[0073] the value of the hyperparameter v balancing the target domain II predictor loss is obtained by the following equation (XI):

[0074]

[0075] the value of the hyperparameter γ balancing the domain discriminator loss is obtained by the following equation (XII):

[0076]

[0077] wherein μ k denotes the value of the hyperparameter μ balancing the target domain I predictor loss in the kth training;

[0078] v k denotes the value of the hyperparameter v balancing the target domain II predictor loss in the kth training;

[0079] γ k denotes the value of the hyperparameter γ balancing the domain discriminator loss in the kth training;

[0080] denotes the classification loss of the domain discriminator in the kth training;

[0081] denotes the classification loss of the source domain predictor in the kth training;

[0082] denotes the classification loss of the target domain I predictor in the kth training;

[0083] denotes the classification loss of the target domain II predictor in the kth training;

[0084] ρ k denotes the training progress.

[0085] wherein in S3, the guidance command is obtained by the following equation (XIII):

[0086]

[0087] wherein a crepresents a guidance command;

[0088] N represents a proportional coefficient;

[0089] V represents the speed of the aircraft;

[0090] represents a missile-target line-of-sight angle change rate;

[0091] a b represents a guidance bias term.

[0092] The present application has the beneficial effects including:

[0093] (1) The aircraft guidance control method for multiple target domain tasks provided by the present application can overcome the environment difference between pre-training and online control, and realize accurate guidance of the aircraft in different guidance task environments.

[0094] (2) The aircraft guidance control method for multiple target domain tasks provided by the present application can be trained based on multiple target domain tasks when training the adversarial neural network in the training domain, so that the specific application of the aircraft can adapt to multiple target domain tasks, and only needs to determine and input the target domain task to be executed before launching the aircraft to realize guidance control of the aircraft based on the target domain task, which can improve the universality of the aircraft and the launch efficiency of the aircraft.

[0095] (3) The aircraft guidance control method for multiple target domain tasks provided by the present application can give real-time control instructions for the target domain task during flight based on the target domain task given before launch. The control instructions are basically consistent with the theoretical optimal control instructions, and the flight trajectory and other related parameters obtained based on the control instructions are also basically consistent with the optimal parameters, so that the control method can provide optimal control instructions based on the preset target domain task to ensure that the aircraft finally hits the target. BRIEF DESCRIPTION OF DRAWINGS

[0096] Figure 1 Fig. 3 shows the aircraft trajectory diagram for the target domain task I in the embodiment;

[0097] Figure 2 Fig. 4 shows the aircraft speed-time change diagram for the target domain task I in the embodiment;

[0098] Figure 3 Fig. 5 shows the aircraft trajectory inclination-time change diagram for the target domain task I in the embodiment;

[0099] Figure 4 Fig. 6 shows the aircraft guidance bias-time change diagram for the target domain task I in the embodiment;

[0100] Figure 5 A schematic diagram showing the trajectory of the aircraft in the embodiment for the target domain task II;

[0101] Figure 6 A schematic diagram showing the velocity of the aircraft in the embodiment for the target domain task II over time;

[0102] Figure 7 A schematic diagram showing the ballistic angle of the aircraft in the embodiment for the target domain task II over time;

[0103] Figure 8 A schematic diagram showing the guidance bias term of the aircraft in the embodiment for the target domain task II over time. DETAILED DESCRIPTION

[0104] The application will be further described below in detail with the aid of the accompanying drawings and embodiments. The features and advantages of the application will become more apparent through these descriptions.

[0105] 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 disclosure is not limited to the specific embodiments illustrated in the drawings.

[0106] According to the aircraft guidance control method for multiple target domain tasks provided by the application, in the method, the current state of the aircraft is input into the pre-trained domain-adversarial neural network filled in the aircraft in the mid-guidance phase to obtain multiple guidance bias terms, each of which corresponds to a target domain task, and the desired target domain task is filled into the aircraft before the launch of the aircraft, so that the guidance bias term output by the domain-adversarial neural network can be called according to the filled target domain task, and the guidance instruction is obtained by the aircraft based on the guidance bias term in real time, and then the aircraft hits the target according to the target domain task under the control of the guidance instruction.

[0107] In the present application, the specific target domain task can be set according to specific requirements, and in the present application, the target domain task is preferably set to two; specifically:

[0108] The guidance task with the impact angle as the constraint condition and the minimum flight time as the optimization target is taken as the source domain task.

[0109] The guidance task with the impact angle as the constraint condition and the maximum terminal velocity as the optimization target is taken as the target domain task I.

[0110] The guidance task with the impact angle and the attack time as the constraint condition and the maximum terminal velocity as the optimization target is taken as the target domain task II; the impact angle is the ballistic angle at the terminal guidance segment.

[0111] Preferably, more target domain tasks can be set. For example, target domain task III can also be set: a guidance task with landing angle as a constraint and energy optimization as the optimization goal; the energy optimization means minimizing the energy consumption of the aircraft.

[0112] Preferably, the neural network is a pre-trained domain adversarial neural network.

[0113] In this application, the number of discrete points is adaptively increased or decreased based on the magnitude of change in the state and control variables. More discrete points are used where the variables change more significantly, and fewer discrete points are used where the changes are less drastic. This results in a more reasonable optimal data distribution, which is more conducive to the thorough training of the neural network.

[0114] Preferably, discrete point data of the flight trajectory obtained by GPOPS-II under different initial states of the aircraft are collected to obtain the dataset for training the domain adversarial neural network; the GPOPS-II is General-Purpose Pseudospectral Optimal Control Software (GPOPS-II) based on the hp-adaptive lasing pseudospectral method.

[0115] Based on different constraints and optimization objectives, the dataset is divided into source domain data, target domain I data, and target domain II data, as shown in Table 1 below:

[0116] Table 1

[0117] Constraint Optimization objective Source domain task Impact angle constraint (-90°) Minimum flight time Target domain task I Impact angle constraint (-90°) Maximum terminal velocity Target domain task II Impact time constraint (70s) and impact angle constraint (-90°) Maximum terminal velocity

[0118] The initial state and landing angle constraints are shown in Table 2 below:

[0119] Table 2

[0120]

[0121] In this application, the current state includes the target distance R, the vehicle speed v, and the trajectory inclination angle. The target line-of-sight angle λ1, the initial launch angle Initial horizontal position x f The column vector formed It is also the input vector of the neural network. The initial launch angle is the angle between the aircraft's axis and the horizontal plane in the geodetic coordinate system when the aircraft is launched towards the target.

[0122] Preferably, in this application, the source domain data is divided into 20 intervals, and 10 ballistics are randomly selected from each interval to obtain 200 ballistic data points, totaling 120,816 sample points;

[0123] In the present application, the ballistic data is a set of discrete points of the ballistic trajectory solved by GPOPS-II; the sample point is a discrete point in the ballistic data, i.e., a vector composed of the current state of the vehicle and the guidance bias term is a sample point.

[0124] The target domain I data is divided into 3 intervals, and 10 trajectories are randomly selected in each interval to obtain 30 ballistic data, a total of 6018 sample points; the target domain II data is divided into 3 intervals, and 10 trajectories are randomly selected in each interval to obtain 30 ballistic data, a total of 4506 sample points. Therefore, the training set is composed of 200 source domain ballistic data, 30 target domain I ballistic data and 30 target domain II ballistic data.

[0125] In the present application, only a small amount of target domain data is added, and the guidance accuracy of the vehicle can be greatly improved.

[0126] The data selection method in the present application is based on Latin hypercube sampling: first, the data set is evenly divided into several non-overlapping intervals, so that each interval has the same length; second, samples are randomly selected in each interval. Such selection method can reduce the demand for data volume as much as possible, and can make the neural network fully learn the mapping relationship between the flight state of the missile and the guidance bias term in the ballistic data.

[0127] Preferably, each set of data in the data set in the present application includes the current state of the vehicle and the corresponding guidance bias term, and the guidance bias term is the label term.

[0128] The output of the source domain predictor, the output of the target domain I predictor and the output of the target domain II predictor in the domain adversarial neural network of the present application are all the outputs of the domain adversarial neural network, i.e., the domain adversarial neural network simultaneously outputs three groups of data, wherein the output of the target domain I predictor corresponds to the target domain task I, and the output of the target domain II predictor corresponds to the target domain task II.

[0129] In a preferred embodiment, the domain adversarial neural network comprises a feature extractor, a domain discriminator, a source domain predictor, a target domain I predictor and a target domain II predictor.

[0130] The domain adversarial neural network in the present application selects the adaptive moment estimation (Adam) algorithm for error back propagation. The adaptive moment estimation calculates the first moment estimation and the second moment estimation of the gradient to design independent adaptive learning rates for different parameters, which has the characteristics of faster model convergence.

[0131] The network parameter update formula is:

[0132]

[0133] wherein, γ is a learning rate, and is 0.001.

[0134] β n denotes the network parameter at the current moment, and β n+1 denotes the network parameter at the next moment, denotes the total loss of the network corresponding to the network parameter at the current moment, denotes the partial derivative of β.

[0135] The dual regression domain adversarial neural network in the application is composed of 1 input layer and 4 hidden layers, and the number of hidden layer neurons is 256; the domain discriminator is composed of 2 hidden layers and 1 output layer, and the number of hidden layer neurons is 256; the source domain predictor, the target domain I predictor and the target domain II predictor are all composed of 3 fully connected hidden layers and 1 output layer. A gradient reversal layer is inserted between the feature extractor and the domain discriminator.

[0136] The domain adversarial neural network selects the Rectified Linear Unit (Relu) as the activation function:

[0137]

[0138] The training process of the neural network comprises:

[0139] Step 1: extracting the common features of the input source domain data, the target domain I predictor and the target domain II predictor through the feature extractor;

[0140] Step 2: forward propagating the obtained common features to the domain discriminator G d to obtain the domain classification loss of the domain discriminator and reversely propagating the classification loss to the feature extractor in the form of gradient ascent through the gradient reversal layer;

[0141] The gradient reversal layer acts as an identity transformation in the gradient forward propagation, and multiplies the gradient by -λ2 in the reverse propagation, so that the gradient direction is automatically negated in the reverse propagation process; if the gradient reversal layer is regarded as a pseudo function then the related expression is as follows:

[0142]

[0143]

[0144] wherein, I represents a unit matrix, and λ2 represents a weight coefficient in the reverse propagation, and is preferably 1.

[0145] Step 3, forward-propagating the obtained common features to the source domain predictor G s and the target domain predictor G t , obtaining the prediction value of the source domain predictor and the prediction value of the target domain predictor; further obtaining the loss of the source domain predictor and the loss of the target domain predictor

[0146] Step 4, obtaining the total objective function of the domain adversarial neural network based on minimizing the total objective function of the domain adversarial neural network to feed back the trainable model parameters;

[0147] Step 5, stopping the training when the domain adversarial neural network reaches the maximum epoch.

[0148] Preferably, the total objective function of the domain adversarial neural network is obtained by the following formula (I):

[0149]

[0150] wherein, represents the loss of the source domain predictor;

[0151] represents the classification loss of the target domain I predictor;

[0152] represents the classification loss of the target domain II predictor;

[0153] represents the domain classification loss of the domain discriminator;

[0154] μ represents a hyperparameter for balancing the loss of the target domain I predictor;

[0155] ν represents a hyperparameter for balancing the loss of the target domain II predictor;

[0156] γ represents a weight coefficient in backpropagation, i.e. a hyperparameter for balancing the loss of the domain discriminator.

[0157] Preferably, the classification loss of the domain discriminator is obtained by the following formula (II):

[0158]

[0159] wherein, when χ α = β, when χ α ≠ β,

[0160] p icrepresents the output of the domain discriminator, which is defined as the probability of making a correct classification for the task class of the input sample;

[0161] In the classification process of the domain discriminator, the source domain data is given a label χ α = 1, the target domain I data is given a label χ α = 2, and the target domain II data is given a label χ α = 3.

[0162] N represents the number of label classes, and is 3,

[0163] β represents different task labels, and is 1, 2, or 3.

[0164] Preferably, the classification loss of the source domain predictor is obtained by the following formula (three):

[0165]

[0166] The classification loss of the target domain I predictor is obtained by the following formula (four):

[0167]

[0168] The classification loss of the target domain II predictor is obtained by the following formula (five):

[0169]

[0170] Wherein, i represents the i-th data sample;

[0171] n represents the number of source domain data samples;

[0172] m1 represents the number of target domain I data samples;

[0173] m2 represents the number of target domain II data samples;

[0174] represents the predicted value of the source domain predictor for the i-th source domain data sample;

[0175] represents the label in the i-th source domain data sample;

[0176] represents the predicted value of the target domain I predictor for the i-th target domain I data sample;

[0177] represents the label in the i-th target domain I data sample;

[0178] represents the prediction of the target domain I predictor for the i-th target domain I data sample;

[0179] represents the label in the i-th target domain II data sample.

[0180] Preferably, the output p of the domain discriminator is ic is obtained by the following equation (Six):

[0181] p ic = G d (G f (x i ; θ f ); θ d ) (Six)

[0182] The is obtained by the following equation (Seven):

[0183]

[0184] The is obtained by the following equation (Eight):

[0185]

[0186] The is obtained by the following equation (Nine):

[0187]

[0188] wherein G f (x i ; θ f ) represents the feature extractor;

[0189] x i represents the i-th sample of the domain adversarial neural network input;

[0190] θ f represents the feature extractor trainable model parameter;

[0191] θ d represents the domain discriminator trainable model parameter;

[0192] θ s represents the source domain predictor trainable model parameter;

[0193] represents the target domain I predictor trainable model parameter;

[0194] represents the target domain II predictor trainable model parameter.

[0195] Preferably, the value of the hyperparameter μ balancing the target domain I predictor loss is obtained by the following formula (X):

[0196]

[0197] The value of the hyperparameter v balancing the target domain II predictor loss is obtained by the following formula (XI):

[0198]

[0199] The value of the hyperparameter γ balancing the domain discriminator loss is obtained by the following formula (XII):

[0200]

[0201] wherein μ k represents the value of the hyperparameter μ balancing the target domain I predictor loss in the kth training;

[0202] v k represents the value of the hyperparameter v balancing the target domain II predictor loss in the kth training;

[0203] γ k represents the value of the hyperparameter γ balancing the domain discriminator loss in the kth training;

[0204] represents the classification loss of the domain discriminator in the kth training;

[0205] represents the classification loss of the source domain predictor in the kth training;

[0206] represents the classification loss of the target domain I predictor in the kth training;

[0207] represents the classification loss of the target domain II predictor in the kth training;

[0208] ρ k represents the training progress, ρ k =k / K, k represents the current training number; K represents the total training number, and is 500.

[0209] In the present application, the selection of the hyperparameters μ, v and γ has a great influence on the network performance. In the classical domain adversarial neural network, the prediction loss and the classification loss of the domain discriminator are both binary cross-entropy functions, so they have the same order of magnitude. However, for the guidance instruction prediction task in the present application, the mean square error function is selected to measure the loss of the predictor, which is very different from the domain discriminator using cross-entropy as the loss function. At the beginning of training, the prediction loss and Classification loss one order of magnitude smaller. Therefore, by balancing the prediction loss and the classification loss with μ, v and γ, they are always kept at the same order of magnitude, and μ, v and γ are self-adaptively adjusted according to the training process.

[0210] Preferably, in step 5, when the classification loss value of the source domain predictor of the neural network, the classification loss value of the target domain I predictor, and the classification loss value of the target domain II predictor are all reduced to 1x10 -5 , it can be considered that the neural network has been trained to the optimal state, and the training of the neural network is completed. Since the convergence speed of the Adam algorithm is fast, the training is stopped when the number of training reaches 500 epochs.

[0211] In a preferred embodiment, the guidance command is obtained by the following formula (thirteen):

[0212]

[0213] wherein a c represents the guidance command;

[0214] N represents a proportional coefficient; preferably, the value is 3;

[0215] v represents the speed of the aircraft; it is obtained in real time by sensors on the aircraft, such as a satellite signal receiver;

[0216] represents the rate of change of the line-of-sight angle between the missile and the target; the rate of change of the line-of-sight angle between the missile and the target is obtained by the following formula:

[0217]

[0218] wherein θ represents the trajectory inclination angle, λ1 represents the line-of-sight angle between the missile and the target, and R represents the distance between the missile and the target; the specific parameters are obtained in real time by the seeker on the aircraft.

[0219] a b represents the guidance bias term.

[0220] In the present application, is a proportional guidance term, and by setting the proportional guidance term, the aircraft can be ensured to hit the target with zero control miss distance, and the guidance bias term a b satisfies the given angle constraint requirement.

[0221] Embodiment

[0222] The discrete point data of the flight trajectory of the aircraft obtained by GPOPS-II under different initial states is collected; according to the different constraints and different optimization objectives, the data set is divided into source domain data, target domain I and target domain II data, as shown in the following Table 1.

[0223] Table 1

[0224] Constraint Optimization objective Source domain task Impact angle constraint (-90°) Minimum flight time Target domain task I Impact angle constraint (-90°) Maximum terminal velocity Target domain task II Impact time constraint (70s) and impact angle constraint (-90°) Maximum terminal velocity

[0225] The source domain data is divided into 20 intervals, and 10 trajectories are randomly selected from each interval to obtain 200 trajectory data, a total of 120816 sample points;

[0226] The target domain I data is divided into 3 intervals, and 10 trajectories are randomly selected from each interval to obtain 30 trajectory data, a total of 6018 sample points; the target domain II data is divided into 3 intervals, and 10 trajectories are randomly selected from each interval to obtain 30 trajectory data, a total of 4506 sample points.

[0227] The domain adversarial neural network training process based on the data includes:

[0228] Step 1, extracting common features of the input source domain data, target domain I predictor, and target domain II predictor through a feature extractor;

[0229] Step 2, the obtained common features are forward propagated to the domain discriminator G d , to obtain the domain classification loss of the domain discriminator

[0230] Step 3, the obtained common features are forward propagated to the source domain predictor G s and the target domain predictor G t , to obtain the prediction value of the source domain predictor and the prediction value of the target domain predictor; further obtaining the loss of the source domain predictor and the loss of the target domain predictor

[0231] Step 4, obtaining the total objective function of the domain adversarial neural network Based on minimizing the total objective function of the domain adversarial neural network to feedback the trainable model parameters;

[0232] Step 5, stopping training when the domain adversarial neural network reaches the maximum epoch.

[0233] Wherein, the total objective function of the domain adversarial neural network is obtained by the following formula (I):

[0234]

[0235] Wherein, represents the loss of the source domain predictor;

[0236] represents the classification loss of the target domain I predictor;

[0237] the classification loss of the target domain I predictor;

[0238] the domain classification loss of the domain discriminator;

[0239] μ represents a hyper-parameter balancing the target domain I predictor loss;

[0240] ν represents a hyper-parameter balancing the target domain II predictor loss;

[0241] γ represents a hyper-parameter balancing the domain discriminator loss.

[0242] the classification loss of the domain discriminator obtained by the following formula (II):

[0243]

[0244] the classification loss of the source domain predictor obtained by the following formula (III):

[0245]

[0246] the classification loss of the target domain I predictor obtained by the following formula (IV):

[0247]

[0248] the classification loss of the target domain II predictor obtained by the following formula (V):

[0249]

[0250] wherein n represents the sample number of the source domain data, and takes a value of 120816;

[0251] m1 represents the sample number of the target domain I data, and takes a value of 6018;

[0252] m2 represents the sample number of the target domain II data, and takes a value of 4506;

[0253] θ f represents the trainable model parameter of the feature extractor;

[0254] θ d represents the trainable model parameter of the domain discriminator;

[0255] θ s represents the trainable model parameter of the source domain predictor;

[0256] represents the trainable model parameter of the target domain I predictor;

[0257] representing target domain II predictor trainable model parameters.

[0258] The trained neural network is directly loaded into the computer on the aircraft, and the time for the neural network to calculate a guidance bias term to indirectly form an optimal guidance command is 0.4 milliseconds;

[0259] The target domain task of the aircraft is set to target domain task I, that is, a guidance task with a landing angle as a constraint condition and a maximum terminal velocity as an optimization target as the target domain task;

[0260] The aircraft is tested using 500 times of Monte Carlo simulation, and a trajectory trajectory is as shown by a dashed line in Figure 1 ; a curve of a change of a velocity of the aircraft over time is as shown by a dashed line in Figure 2 ; a curve of a change of a trajectory inclination angle of the aircraft over time is as shown by a dashed line in Figure 3 ; and a curve of a change of a guidance bias term of the aircraft over time is as shown by a dashed line in Figure 4 .

[0261] The target domain task of the aircraft is set to target domain task II, that is, a guidance task with a landing angle and an attack time as constraint conditions and a maximum terminal velocity as an optimization target as the target domain task;

[0262] The aircraft is tested using 500 times of Monte Carlo simulation, and a trajectory trajectory is as shown by a dashed line in Figure 5 ; a curve of a change of a velocity of the aircraft over time is as shown by a dashed line in Figure 6 ; a curve of a change of a trajectory inclination angle of the aircraft over time is as shown by a dashed line in Figure 7 ; and a curve of a change of a guidance bias term of the aircraft over time is as shown by a dashed line in Figure 8 .

[0263] Figure 1 and the solid lines in Figure 5 are a theoretical optimal trajectory trajectory for the target domain task; the theoretical optimal trajectory trajectory is obtained by GPOPS-II and is a best trajectory trajectory that can be theoretically achieved for a target domain task. Correspondingly, Figure 2 and Figure 6 are a theoretical optimal velocity curve over time; Figure 3 and Figure 7 are a theoretical optimal trajectory inclination angle curve over time; Figure 4 and Figure 8 are a theoretical optimal guidance bias term curve over time. The terminal error statistical results of 500 times of Monte Carlo simulation in the embodiment are shown in Table 3 and Table 4.

[0264] Table 3 End error statistics of terminal in example target domain task I

[0265]

[0266] Table 4 End error statistics of terminal in example target domain task II

[0267]

[0268] It can be known from the examples that the aircraft guidance control method for multiple target domain tasks provided in the application can achieve high-precision control guidance under different target domain task conditions, that is, the specific aircraft state information such as the flight trajectory obtained according to the method is basically consistent with the theoretically best state information, which indicates that the method can provide a high-precision control guidance scheme for any one of the target domain tasks involved, thereby solving the problem that the aircraft is difficult to quickly adapt to multiple target domain tasks, and improving the launch efficiency of the aircraft.

[0269] The application is described above in combination with preferred embodiments, but these embodiments are only exemplary and serve only to illustrate. On this basis, various substitutions and improvements can be made to the application, and these all fall within the protection scope of the application.

Claims

1. A method for aircraft guidance and control for multiple target domain tasks, characterized in that, The method includes: S1, obtaining a state vector of the aircraft in flight according to a predetermined frequency, and outputting a plurality of guidance bias items corresponding to different target domain tasks through the domain adversarial neural network, S2, selecting a corresponding guidance bias item based on a preset target domain task, S3, obtaining a guidance instruction in real time based on the guidance bias item, S4, the aircraft controls the rudder based on the guidance instruction, and finally hits the target; The training process of the domain adversarial neural network includes: Step 1, extracting common features of the input source domain data, target domain I predictor and target domain II predictor through a feature extractor; Step 5, stopping training when the domain adversarial neural network reaches the maximum epoch; Step 2: Propagate the obtained common features forward to the domain discriminator G. d Obtain the classification loss of the domain discriminator. The classification loss of the domain discriminator is then converted using a gradient inversion layer. Backpropagation is performed to the feature extractor using gradient ascent; Step 3, forward propagating the obtained common features to the source domain predictor G s , target domain I predictor target domain II predictor obtaining the prediction value of the source domain predictor, the prediction value of the target domain I predictor, and the prediction value of the target domain II predictor; further obtaining the classification loss of the source domain predictor the classification loss of the target domain I predictor the classification loss of the target domain II predictor Step 4, obtaining the total objective function of the domain adversarial neural network based on the total objective function of the domain adversarial neural network feedback trainable model parameters; μ represents a hyperparameter for balancing the classification loss of the target domain I predictor; a total objective function of the domain adversarial neural network is obtained by the following equation (one) wherein, denotes the classification loss of the source domain predictor; denotes the classification loss of the target domain I predictor; denotes the classification loss of the target domain II predictor; The classification loss of the domain discriminant; ν represents a hyperparameter for balancing the classification loss of the target domain II predictor; γ represents a hyperparameter for balancing the classification loss of the domain discriminator; The value of the hyperparameter μ for balancing the classification loss of the target domain I predictor is obtained by the following formula (ten): The value of the hyperparameter ν for balancing the classification loss of the target domain II predictor is obtained by the following formula (eleven): The value of the hyperparameter γ for balancing the classification loss of the domain discriminator is obtained by the following formula (twelve): Where k represents the kth training; 2. The aircraft guidance control method for multiple target domain tasks according to claim 1, wherein ρ k indicates the training progress. Collecting flight trajectory discrete point data of the aircraft in different initial states obtained by GPOPS-II to obtain a data set for training the domain adversarial neural network; According to different constraint conditions and different optimization targets, the data set is divided into source domain data and multiple target domain data; The target domain data includes target domain I data and target domain II data.

3. The aircraft guidance control method for multiple target domain tasks according to claim 1, wherein The domain adversarial neural network includes a feature extractor, a domain discriminator, a source domain predictor, a target domain I predictor and a target domain II predictor.

4. The aircraft guidance control method for multiple target domain tasks according to claim 1, wherein N represents the number of label categories, and takes a value of 3, Classification loss of the domain discriminator is obtained by the following equation (two): wherein when χ α = β, when χ α ≠ β, p ic represents the output of the domain discriminator; In the classification process of the domain discriminator, the source domain data is given a label χ α = 1, the target domain I data is given a label χ α = 2, the target domain II data is given a label χ α = 3, β represents different task labels, and takes a value of 1, 2 or 3.

5. The aircraft guidance control method for multiple target domain tasks according to claim 1, wherein Where i represents the ith data sample; Classification loss of the source domain predictor is obtained by the following equation (three) Classification loss of target domain I predictor is obtained by the following equation (four) Classification loss of target domain ii predictor is obtained by the following equation (five) n represents the number of source domain data samples; m1 represents the number of target domain I data samples; m2 represents the number of target domain II data samples; 6. The aircraft guidance control method for multiple target domain tasks according to claim 5, wherein denotes the predicted value of the source domain predictor for the i-th source domain data sample; represents a label in the i-th source domain data sample; represents the predicted value of the target domain I predictor for the i-th target domain I data sample; represents the label in the i-th target domain I data sample; represents the predicted value of the target domain II predictor for the i-th target domain II data sample; represents the label in the i-th target domain II data sample.

7. The aircraft guidance control method for multiple target domain tasks according to claim 1, wherein Output p of the domain discriminator ic is obtained by the following equation (six) p ic = G d (G f (x i ; θ f ) ; θ d )(six) The By the following formula (seven): The By the following formula (eight): The By the following formula (IX) wherein G f (x i ; θ f ) denotes a feature extractor; x i represents the i-th sample of the domain-adversarial neural network input; θ f representing that the feature extractor trainable model parameters; θ d denote the domain discriminator trainable model parameters; θ s denote source domain predictor trainable model parameters; representing target domain I predictor trainable model parameters; represent target domain II predictor trainable model parameters. In S3, the guidance instruction is obtained by the following formula (thirteen): N represents a proportionality coefficient; wherein a c represents a guidance command; V represents the speed of the aircraft. ​ represents the rate of change of the line of sight angle; a b represents a guidance bias term.

Images