Midcourse guidance method based on integrated transfer learning

Abstract

The application discloses a guidance method (ETLS) based on integrated transfer learning, which can not only generate optimal guidance commands in real time, but also quickly adapt to a new working environment by fine-tuning with a small amount of new data after a scene changes, and the performance is almost as good as before; the method combines multiple traditional trained DNN neural networks with a meta-learner, simplifies the optimal control problem of the midcourse guidance of a new aircraft into the problem of finding an optimal weighting function and an optimal bias function, and the two functions can be quickly determined with a small amount of data, thereby avoiding the time-consuming and data shortage problems of retraining a new network, so that the control instructions of the midcourse guidance section meeting the terminal velocity and accuracy requirements can be given in a very short time for a new aircraft and a new application scene.

Description

TECHNICAL FIELD

[0001] The present application relates to a midcourse guidance method for a flight vehicle, and in particular to a midcourse guidance method based on integrated transfer learning. BACKGROUND

[0002] In the process of performing a task by a flight vehicle, the flight vehicle needs to go through three stages of launch, midcourse guidance and terminal guidance. Among them, the midcourse guidance process takes the longest time and is the most critical step in the guidance process of the flight vehicle.

[0003] The guidance system is the core of the high hit rate of the missile, and the quality of the guidance algorithm will directly affect the hit accuracy of the missile. The algorithms widely used at present are relatively mature analytical and numerical methods, etc., which are called traditional guidance algorithms, and can ensure high guidance accuracy within a predictable range. In recent years, scholars have introduced machine learning methods into the guidance field, thus giving birth to a series of emerging data-based guidance algorithms, the representatives of which are deep learning and reinforcement learning guidance algorithms. The characteristics of these methods are that a large amount of data and a large amount of time are needed to train the deep neural network (DNN). Once the training is completed, the deep neural network can quickly generate results with less calculation cost.

[0004] However, one inherent defect of this method is poor generalization ability. A well-trained deep neural network usually cannot provide satisfactory performance in a completely new task scenario, and even in most cases it cannot work normally. This means that when the application scenario changes, a new DNN needs to be retrained. However, due to the very time-consuming training process and the need for a large amount of labeled data, it is difficult for data-based guidance algorithms to be applied to those tasks that can only provide a small amount of data or have strict time limits.

[0005] Based on this, when a new flight vehicle is designed or a mature flight vehicle is applied to a new application scenario, the control system of the flight vehicle in the midcourse guidance stage often cannot meet the guidance requirements due to the lack of sufficient data flushing, the control instructions given by the control system cannot meet the guidance requirements, the flight vehicle cannot obtain the maximum terminal velocity in the midcourse guidance stage, and the final hit accuracy will also be affected.

[0006] Based on the above problems, the present inventors have made in-depth analysis on the data-based midcourse guidance method, in the hope of designing a guidance method based on integrated transfer learning (ETLS) which can solve the above problems. SUMMARY

[0007] In order to overcome the above problems, the inventors have made intensive research and designed an integrated transfer learning-based guidance method (ETLS), which can not only generate optimal guidance commands in real time, but also quickly adapt to new working environments through fine-tuning with very little new data after the scene changes, and the performance is almost as good as before; the method combines multiple traditional trained DNN neural networks with a meta-learner, simplifies the optimal control problem of the midcourse guidance of a new aircraft into finding optimal weighting functions and optimal bias functions, and these two functions can be quickly determined with a small amount of data, avoiding the time-consuming and data shortage problem of retraining a new network, so as to give control instructions of the midcourse guidance segment that meet the terminal velocity and accuracy requirements in a very short time for new aircraft and new application scenarios, thereby completing the present application.

[0008] Specifically, the purpose of the present application is to provide an integrated transfer learning-based midcourse guidance method, characterized in that in the method, in the midcourse guidance segment, the optimal control instruction a c new is obtained in real time c new The rudder of the aircraft is controlled to work, so that the aircraft flies according to the predetermined trajectory, thereby completing the midcourse guidance task of maximizing the terminal velocity.

[0009] wherein the optimal control instruction a c new is obtained in real time by inputting the state vector S of the aircraft into the pre-trained network E.

[0010] The training process of the network E includes the following steps:

[0011] Step 1: Train to obtain at least 5 DNN neural networks to form a base learner;

[0012] Step 2: Connect the base learner with a meta-learner to obtain the network E, i.e. the output of the base learner is taken as the input of the meta-learner;

[0013] Step 3: Train the network E with a small amount of aircraft training data, and the trained network E is obtained.

[0014] In step 1, each DNN neural network in the at least 5 DNN neural networks corresponds to an application scenario, i.e. the application scenarios of the respective DNN neural networks are different.

[0015] In step 1, the DNN neural network is a deep feedforward neural network, which has 3 hidden layers, 20 neurons in each layer, and each hidden layer is fully connected with the neurons of the previous layer.

[0016] Wherein, in the step 1, the training process of the DNN neural network comprises:

[0017] Step a, normalizing and grouping the training data;

[0018] Step b, inputting the training set data into the DNN neural network, and comparing the predicted value with the standard value in the training set to obtain a loss;

[0019] Step c, error back propagation, parameter update;

[0020] Step d, when the neural network completes a training, inputting the validation set and test set data into the neural network and calculating the loss value of the network as a measurement index of the generalization ability of the neural network; when the loss value is reduced to a set value or reaches the maximum epoch, the training is stopped.

[0021] Wherein, the meta-learner is a single hidden layer feedforward neural network, and the input of the meta-learner is at least 5, that is, at least 5 outputs of the DNN neural network are the input of the meta-learner; the output of the meta-learner is the optimal control instruction a c new .

[0022] Wherein, the algorithm in the single hidden layer feedforward neural network is:

[0023]

[0024] Wherein, i represents the number of single hidden layer feedforward neural network input;

[0025] N represents the number of single hidden layer feedforward neural network input;

[0026] a ci represents the i-th input of the single hidden layer feedforward neural network;

[0027] C j represents a weighting function;

[0028] b j represents a bias function.

[0029] Wherein, in step 3, the small amount refers to less than 500 groups of data.

[0030] The present application has the beneficial effects, including:

[0031] (1) According to the midcourse guidance method based on integrated transfer learning provided by the application, several neural networks are trained for different aerodynamic models, referred to as base learners, and then a small feedforward neural network is used to learn the mapping relationship from the old optimal control to the optimal control in the new environment, so that the new environment with insufficient data can be quickly adapted;

[0032] (2) According to the midcourse guidance method based on integrated transfer learning provided by the application, the training data can be greatly reduced while ensuring the guidance performance;

[0033] (3) According to the midcourse guidance method based on integrated transfer learning provided by the application, the adaptation to the new environment can be completed in a few seconds, and it is suitable for use in time-critical situations;

[0034] (4) According to the midcourse guidance method based on integrated transfer learning provided by the application, it can be flexibly applied to other scenarios, such as minimum control energy guidance, minimum time guidance, etc. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The DNN neural network structure diagram according to a preferred embodiment of the application is shown;

[0036] Figure 2 The optimal guidance command under different aerodynamic parameters under the same initial condition is shown;

[0037] Figure 3 The meta-learner structure diagram according to a preferred embodiment of the application is shown;

[0038] Figure 4 The training loss and validation loss of the meta-learner under different neuron numbers in Example 1 are shown;

[0039] Figure 5 The position error comparison diagram in Example 2 is shown;

[0040] Figure 6 The velocity error comparison diagram in Example 2 is shown;

[0041] Figure 7 The terminal angle error comparison diagram in Example 2 is shown;

[0042] Figure 8 The time error comparison diagram in Example 2 is shown;

[0043] Figure 9 The position error comparison diagram in Example 3 is shown;

[0044] Figure 10 The velocity error comparison diagram in Example 3 is shown;

[0045] Figure 11 This diagram illustrates the comparison of terminal angle errors in Example 3.

[0046] Figure 12 This diagram illustrates the comparison of time errors in Example 3.

[0047] Figure 13 This diagram illustrates the position error comparison in Example 4.

[0048] Figure 14 This diagram illustrates the speed error comparison in Example 4.

[0049] Figure 15 This diagram illustrates the comparison of terminal angle errors in Example 4.

[0050] Figure 16 This diagram illustrates the comparison of time errors in Example 4.

[0051] Figure 17 A schematic diagram of the aircraft trajectory in Example 5 is shown;

[0052] Figure 18 This diagram illustrates the change in velocity over time in Example 5.

[0053] Figure 19 This diagram illustrates the change of the ballistic inclination angle over time in Example 5.

[0054] Figure 20 The diagram shows the change of control commands over time in Embodiment 5. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Through these descriptions, the features and advantages of the present invention will become clearer and more apparent.

[0056] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments. Although various aspects of embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless specifically indicated otherwise.

[0057] According to the present invention, a mid-course guidance method based on ensemble transfer learning is provided, wherein the optimal control command a is obtained in real time during the mid-course guidance phase. c newThe frequency of obtaining the optimal control instruction in the present application is not lower than the calculation speed of the neural network. The neural network calculation time used in the present application is not more than 1 ms, about 0.3-0.6 ms, so the optimal control instruction can be obtained at a frequency of not less than 1000HZ. The frequency of obtaining the optimal control instruction in the present application is preferably set to 20HZ; obviously, higher frequency can obtain higher precision. Through the optimal control instruction a c new The rudder of the aircraft is controlled to make the aircraft fly according to the predetermined trajectory, thereby completing the midcourse guidance task of maximizing the terminal velocity. The terminal velocity maximization is the maximization of the velocity of the aircraft at the intersection point of the midcourse guidance segment and the terminal guidance segment when the aircraft flies according to the optimal trajectory. In the present application, the aircraft can be a newly designed new type of aircraft, or an aircraft applied to a new application scenario, and a small amount of flight trajectory data can be obtained through the optimal control instruction a c new to control the rudder of the aircraft.

[0058] Preferably, the optimal control instruction a c new is obtained in real time by inputting the state vector S of the aircraft into the pre-trained network E in real time. The state vector S is obtained in real time by the sensor carried on the aircraft. In the present application, the selection of the state vector is not fixed, and appropriate parameters need to be selected for different tasks. For example, the state vector S can include where x and y represent the horizontal and vertical coordinates of the aircraft, the ballistic inclination angle, V is the speed of the aircraft, x f ,y f , represent the terminal position coordinates and the ballistic inclination angle.

[0059] In a preferred embodiment, the training process of the network E includes the following steps:

[0060] Step 1, train to obtain at least 5 DNN neural networks to form a base learner;

[0061] Step 2, connect the base learner with a meta-learner to obtain the network E, i.e. the output of the base learner is taken as the input of the meta-learner; the transfer learning module in the present application is composed of the meta-learner;

[0062] Step 3, train the network E through a small amount of aircraft training data, so as to obtain the trained network E.

[0063] Preferably, in the step 1, each of the at least five DNN neural networks corresponds to an application scenario, i.e., the application scenarios to which the respective DNN neural networks are directed are different. The application scenarios include system dynamics model, operating environment and task target, and any one aspect of the two application scenarios is different, i.e., it is considered as two application scenarios.

[0064] For example, there are five types of aircraft, called aircraft 1 to aircraft 5. Due to different configurations, their aerodynamic coefficients are also different, as shown in Table 1 and Table 2. Table 1 is the baseline of the aerodynamic coefficients, and the coefficients in Table 2 are all changed on this basis. Aircraft 1 to aircraft 5 have sufficient flight data, so each has a well-trained DNN, numbered B1 to B5, called base learners, which can control the corresponding aircraft to complete the terminal angle constraint maximum terminal velocity midcourse guidance task. At this time, if a new aircraft is put into use, its aerodynamic coefficients are shown in Table 3, and it is difficult to train a new DNN for it due to lack of data support; therefore, the ETLS proposed in the present application is needed to solve this problem in such a data-scarce situation.

[0065] Table 1 Baseline of aerodynamic coefficients

[0066]

[0067] Table 2 Aerodynamic coefficients of aircraft 1 to 5

[0068]

[0069] Table 3 Aerodynamic coefficients of new aircraft

[0070]

[0071] Preferably, the base learner needs a data set generation method. The optimal control problem is highly nonlinear, so there is no analytical solution and it can only be solved by numerical methods; there are many existing numerical methods, but their convergence and calculation speed are uneven. The present application selects hp-FRPM as the solving method. This method is not sensitive to initial value guess, and the convergence speed is fast, and in the demonstration scenario in the present application, it only takes an average of 1 second to calculate the optimal solution of an optimal control problem. This algorithm can be directly called using the existing solver GPOPS-II. The biggest advantage of hp-FRPM is that it can adaptively increase or decrease the discrete points according to the change gradient of the state variables and control variables. More discrete points are arranged at positions with larger variable change gradients, and fewer discrete points are arranged at positions with less variable change. This makes the optimal data distribution solved more reasonable and more conducive to the full training of the neural network.

[0072] The present application randomly generates 15000 initial conditions, in which the random initial launch angle and the horizontal position of the target exist, so there are 15000 optimal control problems. The solution accuracy of GPOPS-II is set to 1 x 10 -8 . The solved state-optimal control pairs [s, a c ] are stored as a training set.

[0073] Preferably, in the step 1, the DNN neural network is a deep feedforward neural network, which has 3 hidden layers, each layer has 20 neurons, and the neurons of each hidden layer are fully connected with the neurons of the previous layer. The structure of the DNN neural network is as shown in Figure 1 .

[0074] The calculation method of each layer in the neural network is as follows:

[0075] L i+1 = σ(W i L i + b i )

[0076] Where W i is the weight matrix, b i is the bias matrix, L i+1 represents the output of the i+1 layer, and σ represents a nonlinear activation function. The nonlinear activation function is an indispensable part of a neural network, and different types of activation functions have different effects on the training results. The present patent selects to use function as the activation function, because this function is more suitable for fitting problems.

[0077] Preferably, in the step 1, the training process of the DNN neural network includes:

[0078] Step a, normalizing and grouping the training data; normalizing can increase the training speed and stability of the neural network. The present application uses to normalize;

[0079] When grouping, 70% of the data is used as the training set, 15% as the validation set, and 15% as the test set.

[0080] Step b, input the training set data to the DNN neural network, and compare the predicted value with the standard value in the training set to obtain a loss; the present application uses mean square error as the loss function:

[0081] Step c, error back propagation, parameter update; use the back propagation algorithm to propagate the loss value calculated by the loss function to the inside of the neural network. Then each parameter W i and b iThe updating is performed according to the loss value. The Levenberg-Marquardt algorithm is used for updating in the application. The updating formula is:

[0082] x k+1 = x k (J T J+μI) -1 J T e

[0083] Wherein, J represents the Jacobian matrix, μ is the radius of the trust region, the initial value is 0.001, when the loss of the neural network training decreases, let μ' = 0.1 μ, so as to accelerate the convergence speed; on the contrary, if the loss of the neural network increases, let μ' = 10 μ. The adaptive adjustment of μ can make the Levenberg-Marquardt algorithm obtain the fastest convergence speed when training a small neural network.

[0084] Step d, after the neural network completes one training, that is, after 1 epoch, the validation set and test set data are input into the neural network and the loss value of the network is calculated, which is used as a measurement index of the generalization ability of the neural network; when the loss value decreases to 1 x 10 －6 The following or the maximum epoch is reached to stop the training. Since the convergence speed of the L-M algorithm is extremely fast, the patent sets the number of training to 200 epochs.

[0085] In a preferred embodiment, the meta-learner is a single hidden layer feedforward neural network, the input of the meta-learner has at least 5, that is, the output of at least 5 DNN neural networks is the input of the meta-learner; the output of the meta-learner is the optimal control instruction a c new .

[0086] Preferably, Figure 2 It is shown in FIG. 2 that under different aerodynamic parameters, the optimal guidance instruction profile (expressed by lateral acceleration) required for the aircraft to complete the optimal midcourse guidance of terminal velocity is different under the same initial condition; Figure 2 In FIG. 2, the legend Aero i Corresponding to Table 2, it refers to the result of the ith aerodynamic coefficient; from Figure 2 It can be known from FIG. 2 that the optimal solution of a specific aerodynamic model is different from that of another model; however, from Figure 2 It can also be clearly observed from FIG. 2 that there is also strong similarity between different solutions within the same time interval. This means that we can calculate the new acceleration by using the weighted combination of the existing optimal acceleration, and based on this, the algorithm in the single hidden layer feedforward neural network is:

[0087]

[0088] wherein i represents the number of the single-hidden-layer feedforward neural network input;

[0089] N represents the number of single-hidden-layer feedforward neural network inputs;

[0090] a ci represents the i-th input of the single-hidden-layer feedforward neural network;

[0091] C j represents a weighting function;

[0092] b j represents a bias function.

[0093] C j and b j are parameters to be determined; thus, the midcourse guidance optimal control problem of a new vehicle is simplified to finding the optimal weighting function C i and the optimal bias function b i ; if these two functions can be quickly determined with a small amount of data, the time-consuming and data-insufficient problems of retraining a new network can be avoided.

[0094] In the present application, a meta-learner essentially a single-hidden-layer feedforward neural network is set as shown in Figure 3 , the input of the meta-learner is the optimal guidance command a ci generated by the well-trained base learner, and the output is the optimal control instruction a c new required by the new missile.

[0095] In a preferred embodiment, in step 3, the training method of the meta-learner is also similar to that of the base learner; first, a small amount of data [s, a c new ] is prepared, and the state vector S of the vehicle is first input into the five existing base learners to obtain five outputs a c1 to a c5 . The five outputs are input into the meta-learner to obtain the predicted output of the network, the predicted value is compared with the true value in the training set, the loss is calculated and backpropagation is performed until the loss is small enough, wherein the specific training termination condition is that the loss is less than 1x10 －6 below. Finally, the trained meta-learner is obtained, and thus the network E is obtained.

[0096] Preferably, in step 3, the small amount refers to less than 500 groups of data, and the vehicle training data is vehicle trajectory data. That is, each group of vehicle training data contains all the data in a trajectory of the vehicle.

[0097] Embodiment 1

[0098] The single-hidden-layer feedforward neural networks with different neuron numbers are selected as the meta-learners, and the neuron numbers are 10 to 90, and there are 9 meta-learners in total; the training input data of the meta-learners is the output data of the 5 trained DNN neural networks, and the training output data of the meta-learners is the optimal control instruction;

[0099] The algorithm in the single-hidden-layer feedforward neural network is as follows:

[0100]

[0101] Wherein, i represents the number of the single-hidden-layer feedforward neural network input;

[0102] N represents the number of the single-hidden-layer feedforward neural network input;

[0103] a ci represents the i-th input of the single-hidden-layer feedforward neural network;

[0104] C j represents the weighting function;

[0105] b j represents the bias function.

[0106] The 9 meta-learners are trained and verified, and the training loss and verification loss are recorded, and the results are shown in Figure 4 , and it can be known from Figure 4 that the performance loss is minimum when the number of neurons is 40, and therefore the numerical simulation of the present application is performed by using the meta-learner with 40 neurons.

[0107] The training time of the 9 meta-learners is shown in Table 4, and the time required for retraining the DNN is 2 hours, and it can be known from Table 4 that the meta-learner can quickly complete the learning of the control instruction under the new aerodynamic parameters.

[0108] Table 4 Training time of the meta-learner under different neuron numbers

[0109]

[0110] The network E composed of the trained base learner and the meta-learner can be directly loaded into the onboard computer. There are only a small amount of operations in the network E, and therefore the time for calculating the optimal control instruction is very short, and it is only about 0.3-0.6 milliseconds on a notebook computer, and it will be faster on a customized onboard computer. After the launch of the vehicle, the onboard computer calculates a group of state vectors S of the vehicle every 0.05 seconds and inputs them into the network E, and then the network E can quickly give the optimal control instruction a c new required at this time, and the vehicle only needs to follow the instruction of the network E to achieve the maximum terminal velocity under the terminal angle constraint.

[0111] Embodiment 2

[0112] The five trained DNN neural networks corresponding to the five aircrafts are called, B1, B2, B3, B4, B5, and a base learner is further formed by the five DNN neural networks, and the base learner is connected with a meta-learner to obtain a network E, wherein the network E trained by 100 new trajectories is E(5, 100), and the network E trained by 500 new trajectories is E(5, 500).

[0113] The error performances of E(5, 100), E(5, 500) and the five DNNs corresponding to the five aircrafts are counted, and the position error is shown in Figure 5 , the speed error is shown in Figure 6 , the terminal angle error is shown in Figure 7 , and the time error is shown in Figure 8 .

[0114] It can be seen from Figures 4 to 8 that the error of the network E obtained by using only 500 or less new data is very close to the error of the retrained DNN, which shows the effectiveness of the guidance method based on integrated transfer learning in the present application.

[0115] Embodiment 3

[0116] The five trained DNN neural networks corresponding to the five aircrafts are called, and two, three, four or five of them are used to form a base learner, and the base learner is connected with a meta-learner to obtain a network E, wherein the network E trained by 100 new trajectories is E(2, 100), E(3, 100), E(4, 100) or E(5, 100).

[0117] The error performances are counted, and the position error is shown in Figure 9 , the speed error is shown in Figure 10 , the terminal angle error is shown in Figure 11 , and the time error is shown in Figure 12 .

[0118] It can be seen from Figures 9 to 12 that when the number of DNN neural networks in the base learner is reduced to two, the guidance performance will be greatly reduced. However, when the number of DNN neural networks in the base learner is continuously increased, the improvement of the guidance performance is also limited, which shows that the number of DNN neural networks is preferably greater than two, but should not exceed five.

[0119] Embodiment 4

[0120] The five trained DNN neural networks corresponding to the aircraft are called, and a base learner is formed, the base learner is connected with a meta-learner to obtain a network E, wherein the network E is trained using different numbers of new trajectories, and the network E is E(5, 6), E(5, 12), E(5, 25), E(5, 50), E(5, 100), E(5, 200), and E(5, 500).

[0121] The error performance is counted, and the position error is counted as shown in Figure 13 , the speed error is counted as shown in Figure 14 , the terminal angle error is counted as shown in Figure 15 , and the time error is counted as shown in Figure 16 .

[0122] It can be seen from Figures 13 to 16 that when the training data is reduced to 6 trajectories, the performance of the guidance method based on integrated transfer learning ETLS method deteriorates rapidly, and the increase of data to 500 trajectories does not bring much improvement; the experimental results fully illustrate that the ETLS method can complete the learning of the optimal control under new aerodynamic parameters with a small amount of data.

[0123] Embodiment 5

[0124] The network E(5, 12) in embodiment 4, the DNN neural network B5 in embodiment 2, and a traditional trajectory shaping guidance law (TSG) are called, the traditional trajectory shaping guidance law is described in Zarchan, P., Tactical and strategic missile guidance, Vol. 239, American Institute of Aeronautics and Astronautics, Inc., 2012. The same aircraft is simulated by using the three guidance control schemes, and the obtained aircraft trajectory is as shown in Figure 17 , the speed-time curve is as shown in Figure 18 , the trajectory angle-time curve is as shown in Figure 19 , and the control command-time curve is as shown in Figure 20 .

[0125] It can be seen from Figures 17 to 20 that the performance of the integrated transfer learning midcourse guidance method, i.e., the network E(5, 12), is very close to the optimal solution, which proves the effectiveness of the method, and the integrated transfer learning midcourse guidance method can control the aircraft to fly according to the optimal trajectory and complete the midcourse guidance task of maximizing the terminal speed; the traditional TSG method can meet the terminal angle constraint, but cannot optimize the terminal speed; in addition, the old neural network B5 cannot work normally under the new aerodynamic parameters, resulting in a failed task.

[0126] The application has been described above with reference to preferred embodiments. However, these embodiments are merely exemplary and are intended to be illustrative only. Various substitutions and alterations are possible in view of the disclosure of this application without departing from the spirit and scope of the application.

Claims

1. A mid-course guidance method based on ensemble transfer learning, characterized in that, In this method, the optimal control command a is obtained in real time during the mid-guidance phase. c new ; through this optimal control command a c new The servo motors control the aircraft to steer, enabling the aircraft to fly along a predetermined trajectory and thus achieve mid-course guidance with maximum terminal velocity. The optimal control command a is obtained in real time by inputting the aircraft's state vector S into a pre-trained network E. c new ; The training process of network E includes the following steps: Step 1: Train at least 5 DNN neural networks to form a base learner; Step 2: Connect the base learner and the meta learner to obtain network E, that is, use the output of the base learner as the input of the meta learner; Step 3: Train network E using a small amount of aircraft training data to obtain the trained network E; In step 1, each of the at least 5 DNN neural networks corresponds to an application scenario, meaning that each DNN neural network targets a different application scenario. The application scenarios include: system dynamics model, operating environment, and task objective. If any aspect of two application scenarios is different, they can be considered as two different application scenarios. In step 1, the DNN neural network is a deep feedforward neural network with 3 hidden layers, each with 20 neurons, and each neuron in the hidden layer is fully connected to the neurons in the previous layer. The calculation method for each layer in a neural network is as follows: L i+1 =σ(W i L i +b i ) Among them, W i Let b be the weight matrix. i Let L be the bias matrix. i+1 L represents the output of the (i+1)th layer, and σ represents the non-linear activation function. i This represents the output of the i-th layer; The meta-learner is a single-hidden-layer feedforward neural network. The meta-learner has at least five inputs, meaning the outputs of at least five DNN neural networks are the inputs to the meta-learner. The output of the meta-learner is the optimal control command a. c new .

2. The mid-course guidance method based on ensemble transfer learning according to claim 1, characterized in that, In step 1, the training process of the DNN neural network includes: Step a: Normalize and group the training data; Step b: Input the training set data into the DNN neural network and compare the predicted values ​​with the standard values ​​in the training set to obtain a loss. Step c: Backpropagation of error, parameter update; Step d: After the neural network completes one training cycle, the validation set and test set data are input into the neural network and the network's loss value is calculated as a measurement indicator of the neural network's generalization ability; training stops when the loss value decreases to a set value or reaches the maximum epoch.

3. The mid-course guidance method based on ensemble transfer learning according to claim 1, characterized in that, The algorithm in the single hidden layer feedforward neural network is as follows: Where i represents the number of the input to a single hidden layer feedforward neural network; N represents the number of inputs to a single hidden layer feedforward neural network; a ci This represents the i-th input to a single hidden layer feedforward neural network. C j Represents the weighting function; b j This represents the bias function.

4. The mid-course guidance method based on ensemble transfer learning according to claim 1, characterized in that, In step 3, "small number" refers to less than 500 sets of data.

Images