A deep learning-based hidden control integrated control distribution method

Abstract

This invention discloses a deep learning-based integrated stealth and control allocation method, belonging to the field of aircraft stealth and control technology. By using a deep neural network, the three-axis torque commands generated by the flight control law are mapped to control surface deflection commands, addressing the issues of control surface nonlinearity and the impact of control surface deflection on stealth performance. Specifically, it includes: 1) fitting the nonlinear relationship between control surface deflection and torque using wind tunnel and CFD data; 2) constructing a deep autoencoder network, using a combination of virtual control error and control surface deflection as the loss function; 3) offline training of the network, including supervised and unsupervised training, and online deployment to the flight control system. This invention achieves high-precision nonlinear control allocation while optimizing control surface deflection to reduce stealth performance loss, offering advantages such as strong real-time performance and low data requirements.

Description

Technical Field

[0001] This invention belongs to the field of aircraft stealth and control technology, specifically relating to a stealth and control integrated control allocation method based on deep learning. Background Technology

[0002] Traditional conventional aircraft typically employ three independent control surfaces: ailerons, elevators, and rudders, which control roll, pitch, and yaw moments respectively. Since the relationship between control surfaces and moments is approximately linear, the control allocation problem can be directly solved analytically using methods such as the pseudo-inverse method. However, tailless aircraft (such as flying wing configurations), by eliminating the vertical and horizontal stabilizers, rely on multiple redundant control surfaces (such as split control surfaces and ailerons) to achieve three-axis control, resulting in strong nonlinear coupling characteristics between control surfaces and moments. Furthermore, to meet stealth requirements, the surface design of such aircraft is extremely sensitive to control surface deflection; even a small deflection can significantly increase the radar cross-section (RCS), reducing stealth performance.

[0003] Existing control allocation methods (such as pseudo-inverse methods and mathematical programming) rely on linear assumptions, making it difficult to handle nonlinear rudder effect problems. While traditional optimization algorithms can introduce stealth constraints, their high computational complexity makes it difficult to meet the real-time requirements of flight control. Furthermore, the relationship between stealth performance and rudder deflection angle usually requires costly wind tunnel testing or electromagnetic simulation, leading to data scarcity and further limiting the applicability of data-driven methods. In recent years, neural networks have shown potential in the field of control due to their powerful nonlinear fitting capabilities; however, existing research has largely focused on single control objectives without comprehensively considering stealth constraints and real-time requirements.

[0004] Therefore, there is an urgent need for a new control allocation method that can simultaneously meet the following requirements: 1) adapt to the nonlinear characteristics of rudder effect; 2) minimize the impact of rudder surface deflection on stealth performance; 3) have real-time computing capabilities; and 4) reduce dependence on high-cost labeled data. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a deep learning-based integrated control allocation method for stealth and control, which maps the three-axis control torque commands generated by the control law of a tailless aircraft into control surface deflection commands, thereby reducing the impact of control deflection on the aircraft's stealth performance while ensuring control effectiveness.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A deep learning-based method for integrated implicit control allocation, the method comprising:

[0008] Step 1: Establish a nonlinear relationship model between the control surface deflection angle and the control torque;

[0009] Step 2: Construct a deep autoencoder neural network, which includes a deep neural network and a nonlinear relational model. The output of the deep neural network is converted into a virtual control quantity and standardized, and then connected to the nonlinear relational model to obtain the deep autoencoder neural network.

[0010] Step 3: Perform supervised training on the deep autoencoder neural network based on pre-assigned data;

[0011] Step 4: Define the loss function as the weighted sum of the virtual control quantity reconstruction error and the rudder surface deflection penalty term, and perform unsupervised training on the deep autoencoder neural network.

[0012] Step 5: Extract the deep neural network from the trained deep autoencoder neural network and connect it to the flight control system to receive standardized three-axis torque commands in real time and output the optimal control surface deflection angle allocation scheme.

[0013] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned deep learning-based hidden control and allocation method.

[0014] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned deep learning-based hidden control and allocation method.

[0015] The beneficial effects of this invention are as follows:

[0016] Efficiently handles nonlinear control assignment problems: Through the powerful nonlinear mapping capability of deep autoencoder networks, this invention can accurately fit the complex coupling relationship between control surface deflection angle and three-axis control torque, overcoming the limitations of traditional pseudo-inverse methods and mathematical programming methods that rely on linear assumptions, and significantly improving the accuracy and adaptability of flight control.

[0017] Optimizing stealth performance: A control surface deflection weighting factor is introduced into the loss function, imposing a higher penalty on control surfaces that significantly impact stealth performance, thereby automatically suppressing their deflection amplitude during control allocation. This design effectively reduces the impact of control surface deflection on the aircraft's radar cross-section (RCS), balancing the dual requirements of flight control and stealth performance.

[0018] High real-time performance and high computational efficiency: By combining offline training with online deployment, complex nonlinear optimization problems are transformed into forward computation of neural networks, which greatly reduces the online computational burden and meets the stringent real-time requirements of flight control systems.

[0019] Reduced data dependence and good versatility: The method adopts a self-supervised learning mechanism, which only requires torque-deflection angle data generated by wind tunnel tests or CFD simulations to complete the training. It does not rely on a large amount of labeled measured data, which reduces the data acquisition cost and enhances the applicability of the method on different aircraft platforms.

[0020] Flexible and scalable: By adjusting the weighting factors in the loss function, control performance and stealth requirements can be flexibly balanced to adapt to different mission scenarios. Furthermore, the network structure can be expanded according to the number of control surfaces and aircraft characteristics, demonstrating broad engineering application potential. Attached Figure Description

[0021] Figure 1 This is a flowchart of a hidden control and allocation method based on deep learning according to the present invention.

[0022] Figure 2 This is a diagram of the neural network structure used in this invention;

[0023] Figure 3 This is a graph showing the error reduction during supervised training in an embodiment of the present invention;

[0024] Figure 4 This is a graph showing the error reduction during unsupervised training in an embodiment of the present invention;

[0025] Figure 5 This is a network allocation error diagram in an embodiment of the present invention;

[0026] Figure 6 This is a diagram illustrating the tracking of attitude signals by the aircraft in an embodiment of the present invention. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] This method consists of two main parts: offline network training and online network usage. First, data from wind tunnel and computational fluid dynamics tests are collected, and a polynomial fitting method is used to obtain the expression between the control surface deflection angle and the three-axis control torque. Then, a deep neural network is constructed. The training of the deep neural network is divided into two steps: the first step is supervised learning based on pre-assigned data, and the second step is unsupervised learning using the autoencoder principle. The unsupervised learning process ensures that the network's assignment results simultaneously consider the aircraft's control performance and stealth performance. This is the offline network training process. After completion, the constructed neural network is connected to the flight control system. The three-axis torque commands of the flight control program are standardized according to the test data standardization method and then input into the connected network. The network then provides the corresponding control surface deflection angle. This process occurs online, hence the name "online network usage." Specifically, the method includes:

[0029] Step 1: Establish a nonlinear relationship model between control surface deflection angle and control torque: Using wind tunnel and computational fluid dynamics test data, a polynomial fitting method is employed to obtain the nonlinear mapping relationship between control surface deflection angle and three-axis control torque. ;in, Let be a column vector consisting of the deflection values ​​of each control surface of the aircraft. If the aircraft has , The rudder surface, then , They are respectively Dimensionality and 3-dimensional real vector space.

[0030] Step 2: Construct a deep autoencoder neural network, which includes a deep neural network and a nonlinear relational model. The output of the deep neural network is converted into a virtual control variable and standardized, then connected to the nonlinear relational model to obtain the deep autoencoder neural network. For example... Figure 2 As shown, a deep neural network is first constructed, including 3-5 hidden layers. The input layer consists of 3 nodes corresponding to the three-axis normalized torque, and the number of neurons in the output layer is the number of aircraft control surfaces. The activation function of the hidden layers is the ReLU function, and the activation function of the output layer is a function modified according to the control surface deflection limit. The output of the deep neural network is utilized The virtual control quantity is converted into a virtual control quantity and standardized, then connected to a nonlinear relational model to form an autoencoder structure. The error function is a combination of the virtual control quantity error and the control surface deflection.

[0031] Step 3: Perform supervised learning on the deep autoencoder neural network based on pre-assigned data. The labeled dataset used for supervised learning... The pseudo-inverse method is used to solve the problem based on the linearized control surface efficiency model. The batch size is set to... Every 40 epochs of training, the output error of the neural network is evaluated using test data. The evaluation is completed when all three axes of error reach a certain threshold. Stop training when the intensity reaches a certain level.

[0032] Step 4: Define the loss function as the weighted sum of the virtual control quantity reconstruction error and the control surface deflection penalty term, and perform unsupervised training on the deep autoencoder neural network: Define the loss function as the weighted sum of the virtual control quantity reconstruction error and the control surface deflection penalty term, where the control surface penalty weights are dynamically adjusted according to stealth sensitivity; set the batch size to... The training data is input into a deep autoencoder neural network in batches. The loss function for each batch is the average of the error functions of that batch of training data. The gradients of the parameters in the neural network are obtained by backpropagation using the loss function, and the parameter values ​​are updated using batch stochastic gradient descent. Every 40 epochs of training, the output error of the neural network is evaluated using test data. When the network evaluation error reaches a certain threshold... Stop training when the intensity reaches a certain level.

[0033] Step 5, Online Deployment: Extract the deep neural network from the trained deep autoencoder neural network and connect it to the flight control system. It receives standardized three-axis torque commands in real time and outputs the optimal control surface deflection allocation scheme, achieving real-time control allocation that balances control performance and stealth requirements. Extracting the deep neural network portion from the trained deep autoencoder neural network yields a neural network with Z-SCORE standardized three-axis torque as input and control surface deflection as output. The parameters used for Z-SCORE standardization are the same as those used in Step 2 for standardizing the training data.

[0034] The activation function mentioned in step 2 is as follows:

[0035] (1)

[0036] (2)

[0037] In the formula For the input of neurons, These represent the maximum and minimum deflection angles of the control surface corresponding to a certain output layer neuron, in radians.

[0038] In step 3, the label dataset The specific construction method is as follows: First, a random sampling method is used to sample the control surface deflection amount at the aircraft's control surface deflection limit, forming a control surface deflection amount set. ;use Map the set of rudder surface yaw values ​​to a set of three-axis torques. Using the least squares principle, we obtain linear fitting , The control efficiency matrix of the aircraft. The label dataset is a constant vector generated during the fitting process. The definition of is:

[0039] (3)

[0040] Will Triaxial torque in Standardization is performed, and the standardized three-axis torque constitutes the standardized aircraft torque, which can reach a total torque. Thus, the labeled dataset used for supervised training is obtained: ;

[0041] exist The method used is random sampling, extracting 70% of the data as the training set for supervised training. 15% of the data was selected as the test set. The remaining 15% of the data was used as the evaluation set. .

[0042] The loss function for supervised learning is:

[0043] (4)

[0044] In the formula This indicates that the neural network is in operation when the input is... Output at time The first step is to train the Bach Size.

[0045] The error evaluation function for supervised learning is:

[0046] (5)

[0047] In the formula for The number of elements in the middle.

[0048] In step 4, the dataset used for unsupervised learning is the triaxial torque set generated in step 3. ,Will The elements in the equation are standardized using the Z-SCORE method, and the standardized triaxial torque set is still denoted as . ,exist 70% of the data was randomly sampled as the training set. 15% of the data was selected as the test set. 15% of the data was extracted as the evaluation set. .

[0049] The loss function for unsupervised training is:

[0050] (6)

[0051] In the formula As a stealth weight factor, For the first The stealth weight of control surfaces is determined by the degree to which their deflection affects the aircraft's stealth performance. A greater stealth weight corresponds to a control surface whose deflection has a greater impact on the aircraft's stealth performance. By selecting appropriate weights, the impact of control surface deflection on the aircraft's stealth performance can be reduced. As the rudder deflection weighting factor, the balancing network reduces the allocation error and the degree of reduction in rudder deflection angle during the second training step. The larger the network, the more it tends to reduce the control surface deflection angle. This determines the batch size for the second training step.

[0052] The error evaluation function for unsupervised training is:

[0053] (7)

[0054] In the formula for The number of elements in the middle.

[0055] In step 5, the trained deep neural network is connected to the flight control system. The control law in the system generates three-axis control torque signals according to flight commands. The three-axis control signals generated by the control law are Z-SCORE standardized to obtain the standardized three-axis torques to be assigned. ;Will The deep neural network is input to obtain the allocation results.

[0056] Example:

[0057] This method is used for attitude control of a tailless aircraft with 10 control surfaces, including 4 elevons and 6 split-drag rudders. First, based on literature data and CFD simulation results, its nonlinear control surface efficiency function is obtained. , The function that generates roll and pitch moments due to the deflection of the control surfaces is a cubic polynomial. The function that generates the yaw moment is a quadratic polynomial.

[0058] A deep neural network was then constructed with an input dimension of 3 and an output dimension of 10, consisting of 5 hidden layers. The number of neurons in each hidden layer was 64, 128, 256, 128, and 64, respectively. The training set for both the first and second training steps consisted of 280,000 data points. The training errors for the first supervised training and the second unsupervised training steps are as follows: Figure 3 and Figure 4 As shown.

[0059] Figure 5 This demonstrates the torque distribution error of the neural network. Figure 6 The study demonstrated the aircraft's tracking of attitude command signals when using a neural network as the control allocation mechanism. Evaluation of the neural network showed that the three-axis torque error was small. For the vast majority of the torques to be allocated (over 92%), the torque allocation error of the neural network was below 2%. Furthermore, when using the neural network as the control allocation mechanism, the aircraft tracked attitude commands well.

[0060] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned deep learning-based hidden control and allocation method.

[0061] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned deep learning-based hidden control and allocation method.

[0062] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based hidden control integrated control distribution method, characterized in that, The method comprises: Step 1, establishing a nonlinear relationship model of rudder deflection angle and control moment; Step 2, constructing a deep auto-encoding neural network, including a deep neural network and a nonlinear relationship model, wherein the output of the deep neural network is converted into a virtual control variable and standardized, and then connected with the nonlinear relationship model to obtain the deep auto-encoding neural network; Step 3, performing supervised training on the deep auto-encoding neural network based on pre-allocated data; Step 4, defining a loss function as the weighted sum of virtual control variable reconstruction error and rudder deflection penalty term, and performing unsupervised training on the deep auto-encoding neural network; Step 5, taking out the deep neural network from the trained deep auto-encoding neural network and connecting it to the flight control system, and real-time receiving standardized three-axis moment instructions and outputting the optimal rudder deflection angle allocation scheme; The step 1 comprises: based on wind tunnel test and fluid mechanics test data, adopting a polynomial fitting method to obtain a nonlinear mapping relationship between the rudder deflection angle and the three-axis control moment ; wherein, is a column vector composed of deflection values of each rudder surface of the aircraft, represents common block rudder surfaces of the aircraft, respectively, dimension and 3-dimensional real vector space; The step 2 comprises that the deep neural network comprises 3-5 hidden layers, the input layer is 3 nodes corresponding to the three-axis normalized moment, and the number of neuron nodes of the output layer is the number of aircraft control surfaces; the activation function of the hidden layer is a ReLU function, and the activation function of the output layer is a modified The output of the deep neural network is changed into a virtual control variable by using a nonlinear mapping relationship and normalized, and then connected with a nonlinear model to jointly constitute a self-encoder structure; wherein the error function is a combination of the virtual control variable error and the control surface deflection.

2. The method of claim 1, wherein the method is based on deep learning. The step 3 includes supervising a label data set used for training Based on the linearization of the control surface efficiency model, the pseudo-inverse method is used to solve, is the normalized reachable set of the aircraft's torque.

3. The method of claim 2, wherein the method further comprises: The step 3 specifically comprises: Aircraft control surface deflection vector set is constituted by using random sampling method to sample control surface deflection vector in control surface deflection limit, ; Nonlinear mapping relationship is used Control surface deflection vector set is mapped into three-axis moment set ; Least square principle is adopted to obtain linear fitting , Aircraft control efficiency matrix, Constant vector generated in fitting process, then the definition of label data set is:​ （3） The triaxial moment in the triaxial moment is normalized, and the normalized triaxial moment constitutes a normalized aircraft torque reachable set , so as to obtain a label data set for supervised training: .

4. The method of claim 3, wherein the method further comprises: In step 3, the The training set for supervised training is divided by the method of random sampling The test set The evaluation set The loss function of supervised training is: （4） In the formula represents the output of the neural network when the input is , bach Size is the size of the supervised training The error evaluation function of the supervised training is: （5） In the formula is the number of middle elements.

5. The method of claim 4, wherein the method further comprises: The training set used in the unsupervised training in step 4 is the triaxial moment set generated in step 3 , the elements in are standardized by Z-SCORE method, and the standardized triaxial moment set is denoted as , the elements in are divided into training set , test set , and evaluation set by random sampling.

6. The method of claim 5, wherein the method further comprises: In the step 4, the loss function of the unsupervised training is: （6） wherein is the stealth weight factor, is the stealth weight for the rudder surface, is the rudder deflection weight factor, is the size of the batch size for unsupervised training; The error evaluation function of the unsupervised training is: （7） In the formula is the number of middle elements.

7. An electronic device, comprising: Comprise: One or more processors; Memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the deep learning-based hidden control integrated control allocation method of any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, Executable instructions are stored thereon, which can make the processor implement the deep learning-based hidden control integrated control allocation method of any one of claims 1-6 when executed by the processor.

Images