Aircraft RCS fast prediction method loaded with distributed intelligent stealth skin

Abstract

The application provides a kind of fast estimation method of the RCS of aircraft loaded with distributed intelligent stealth skin.The method is mainly divided into two parts.One part is an electromagnetic response detection method of aircraft intelligent stealth skin based on ray tracing,which can quickly and accurately obtain the number and position of distributed intelligent stealth skin that regulates and controls the RCS of aircraft when threat radar wave irradiates to the aircraft.The other part is a multi-level multi-layer neural network model,which can quickly and accurately predict the RCS value of aircraft at different azimuth angles.The method can quickly estimate the RCS of aircraft loaded with distributed intelligent stealth skin.Compared with traditional RCS searching methods such as table lookup and interpolation,the speed increases by several orders of magnitude,and it can be easily deployed on PC and FPGA,with good compatibility and portability.

Description

Technical Field

[0001] This invention relates to the field of aircraft stealth, and more particularly to a method for rapid RCS prediction of aircraft equipped with distributed intelligent stealth skin. Background Technology

[0002] By distributing multiple smart stealth skins composed of tunable metasurfaces on the outer surface of an aircraft, and altering the scattering characteristics of these skins, the radar cross-section (RCS) of the aircraft can be effectively reduced. To facilitate the control of the smart stealth skins in practical engineering applications, theoretically, the RCS of the aircraft under different scattering states with the distributed smart stealth skins should be tested and stored in the aircraft's embedded device. For an aircraft with n distributed smart stealth skins (m scattering states), the amount of RCS test data is m. n The number of intelligent stealth skin layers increases exponentially. When the number of distributed intelligent stealth skin layers is large, it becomes impractical to test and obtain the entire RCS of the aircraft. Typically, in engineering practice, only the RCS of the aircraft under typical scattering conditions and within a certain angular range of the distributed intelligent stealth skin layers is tested, meaning that only partial RCS data of the aircraft can be obtained.

[0003] To address the aforementioned issues, current technical measures involve using test or calculation data under partial scattering conditions as a basis, and employing search methods such as table lookup and interpolation to predict the RCS of aircraft equipped with distributed intelligent stealth skin. However, such methods are inefficient and have limited accuracy.

[0004] Therefore, how to process the measured RCS data of some aircraft to quickly predict the RCS of aircraft with distributed intelligent stealth skin under other scattering states, and to establish a general rapid prediction method, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The technical problem solved by this invention is to provide a method for rapid prediction of the RCS of an aircraft with distributed intelligent stealth skin based on ray tracing and deep learning algorithms. This method addresses the problem that the amount of RCS data for an aircraft increases exponentially when the number of intelligent stealth skins is too large, making it difficult to obtain all of them through testing.

[0006] A method for rapid RCS prediction of an aircraft with distributed intelligent stealth skin includes the following steps:

[0007] S1. The aircraft is placed horizontally with its center at the origin of the spherical coordinate system. A transmitter capable of emitting a family of electromagnetic rays is set at a certain position P(r,θ,φ) in space. The aircraft has k identical smart stealth skins distributed on it, each with j scattering states. A receiver for receiving electromagnetic rays is set at the center of each smart stealth skin. The position coordinate θ of the transmitter corresponds to 90° minus the pitch angle of the aircraft, and the position coordinate φ corresponds to the azimuth angle of the aircraft.

[0008] S2. The transmitter emits an initial family of N rays toward the origin of the coordinate system. After the rays hit the aircraft, each ray in the family is tracked.

[0009] S3. Record the distance l from the ray to the center of the smart stealth skin of a receiver after it acts on the aircraft and is reflected, transmitted, and diffracted. Compare this distance with the set distance threshold L. When l ≤ L, the ray is counted in the receiver's ray count; otherwise, it is not counted.

[0010] S4. Count the number of electromagnetic rays received by the receiver at the center of each smart stealth skin panel on the aircraft and compare it with the set threshold number M; for the number of electromagnetic rays m received by the receiver at the i-th smart stealth skin panel... i When m i When M ≥ M, it is assumed that the i-th smart stealth skin will respond to the electromagnetic ray beam emitted from position P(r,θ,φ), that is, changing the scattering state of the i-th smart stealth skin will affect the RCS of the aircraft at pitch angle θ and azimuth angle φ; otherwise, it is assumed that changing the scattering state of the i-th smart stealth skin will not affect the RCS of the aircraft at pitch angle θ and azimuth angle φ.

[0011] S5. Set the smart stealth skin that does not respond to electromagnetic beams to a non-working state, and the smart stealth skin that responds to electromagnetic beams to a scattering state. Complete the changes of j scattering states, test and obtain the RCS values ​​of the aircraft at all pitch angles θ and some azimuth angles φ, and construct a dataset.

[0012] S6. Construct a multi-level, multi-layer perception neural network model based on the scattering state and pitch angle of the distributed intelligent stealth skin.

[0013] S7. Select the RCS data of the aircraft at some azimuth angles φ under different scattering states of the distributed intelligent stealth skin at the first pitch angle in the dataset, and divide 90% of it into the training dataset and 10% into the test dataset.

[0014] S8. Input the training dataset into the neural network and train the neural network through backpropagation;

[0015] S9. Input the test dataset into the neural network for cross-validation. When the accuracy is below 90%, execute step S8 to continue training the neural network. When it is above 90%, stop training the model.

[0016] S10. Repeat steps S7 to S9 until the multi-level, multi-layered sensing neural network model is trained, and the fitting relationship between the RCS and azimuth angle φ of the aircraft under different scattering states of the distributed intelligent stealth skin at all pitch angles is obtained.

[0017] Furthermore, in step S6:

[0018] For each different sub-model, the input of the model is the azimuth angle of the aircraft, and the output is the RCS value. A multilayer perceptron neural network model is selected as the basic model. The hyperparameters of the model are shown below, where there are 26 hidden layers. Each layer uses the ReLU function as the activation function. Batch normalization is enabled on some hidden layers, that is, the data distribution is normalized to the standard normal distribution before the activation function receives the input value.

[0019] x and a represent the input value of the previous layer and the hidden layer value; w represents the hidden layer weight value, and b represents the hidden layer bias value. Therefore, the neuron value is:

[0020] a = Relu(wx + b)

[0021] The ReLU activation function expression is:

[0022]

[0023] Furthermore, the characteristic is that in step S7:

[0024] The dataset was randomly divided into training and test sets in a 9:1 ratio.

[0025] Furthermore, the characteristic is that in step S8:

[0026] After building the model, a learning rate of 0.0001 is defined, and the convergence condition is set at 40,000 training iterations for each sub-model or a loss value below 1; the loss function is defined as:

[0027] Loss = MSE(S - S')

[0028] S is the true label value, i.e., the simulated RCS value; S' is the RCS value of the neural network output; and MSE is the mean squared error function, the expression of which is as follows:

[0029]

[0030] Backpropagation is performed according to the following formula:

[0031]

[0032] Where η is the learning rate, and backpropagation is ultimately completed by adjusting the intrinsic parameters of the neural network step by step.

[0033] Furthermore, in step S9:

[0034] During testing, the intrinsic parameters of the fixed neural network model remain unchanged. All test set data is input into the neural network, and error analysis is performed between the output values ​​and the true label values. The error analysis function is shown below:

[0035]

[0036] When the value of p is less than 0.2, the two are considered to be the same; when it is greater than 0.2, the output value of the neural network is considered to be inaccurate.

[0037] The beneficial effects of this invention are:

[0038] This invention provides a method for rapidly predicting the radar cross-section (RCS) of an aircraft with distributed intelligent stealth skin based on ray tracing and deep learning algorithms. The method consists of two main parts: one is a ray-tracing-based electromagnetic response detection method for the intelligent stealth skin, which can quickly and accurately determine the number and location of the distributed intelligent stealth skin components that influence the RCS of the aircraft when threatened radar waves strike it; the other part is based on a multi-level, multi-layer neural network model to quickly and accurately predict the RCS values ​​of the aircraft at different azimuth angles. This method can rapidly predict the RCS of aircraft with distributed intelligent stealth skin, offering a geometrically faster speed compared to traditional RCS search methods such as lookup tables and interpolation. Furthermore, it can be easily deployed on PCs and FPGAs, exhibiting good compatibility and portability. Attached Figure Description

[0039] Figure 1 The following is an overall flowchart of a method for predicting the RCS of a spacecraft segment based on ray tracing and a multi-level, multi-layer perceptual neural network algorithm, provided for an embodiment of the present invention.

[0040] Figure 2 A schematic diagram of a hexahedral metal section of an aircraft provided for an embodiment of the present invention.

[0041] Figure 3 This represents the response state of the intelligent stealth skin on the cabin section to electromagnetic ray beams emitted from different locations in an embodiment of the present invention.

[0042] Figure 4 This is a diagram of a multi-level, multi-layer perceptual neural network architecture provided in an embodiment of the present invention.

[0043] Figure 5The loss curve is provided for an embodiment of the present invention.

[0044] Figure 6 This is a schematic diagram illustrating the results of an embodiment of the present invention. Detailed Implementation

[0045] The present invention will now be described in further detail with reference to the accompanying drawings.

[0046] In this invention, ray tracing algorithm is first used to identify which smart stealth skin coatings on the surface of the aircraft mainly affect the RCS value of the aircraft at different pitch and azimuth angles. For smart stealth skin coatings that do not affect the RCS value, their scattering states are no longer considered, simplifying the number of distributed smart stealth skin scattering states. Then, according to the simplified number of distributed smart stealth skin scattering states, the aircraft's RCS is tested at all pitch angles and some azimuth angles to form a dataset. Finally, a multi-level, multi-layer neural network model is constructed according to the distributed smart stealth skin scattering states and pitch angles, and the dataset is trained to obtain a high-precision fitting relationship between the aircraft's RCS and azimuth angle, thus completing the rapid and accurate prediction of the aircraft's RCS at unknown azimuth angles.

[0047] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0048] This invention provides a method for predicting the RCS of a vehicle segment loaded with distributed intelligent stealth skin based on ray tracing and deep learning algorithms. The overall flowchart is shown below. Figure 1 The specific operating steps are as follows:

[0049] S1, such as Figure 2 As shown, a hexahedral metal section of an aircraft is horizontally placed at the origin of a spherical coordinate system, with its center coinciding with the origin. The section is 4m long, 0.5m wide, and 0.5m high. At a certain location in space... A transmitter capable of emitting electromagnetic rays is installed at the site. Four identical rectangular smart stealth skin panels, each 1m long and 0.2m wide, are distributed across the module's four sides (top, bottom, left, and right), with the center of each skin coinciding with the center of the four sides. Each smart stealth skin panel has two scattering states, and a receiver for receiving electromagnetic rays is placed at the center of each panel. The transmitter's position coordinate θ corresponds to 90° minus the aircraft's pitch angle, which in this embodiment is between 60° and 90°, spaced 5° apart. The azimuth angle of the corresponding aircraft is an integer degree between 0° and 359° in this embodiment.

[0050] S2. Divide the hexahedral metal section into triangular faces (represented by the vertex coordinates (A,B,C)). The transmitter's initial position coordinates are (1000m,0,0). An initial ray family of 1000 rays is emitted towards the origin. A ray tracing algorithm is used to trace the reflection path of each ray through the triangular faces, obtaining all ray paths from the transmitter to the four receivers. Then, following step S1, the transmitter position coordinates θ and... The requirements are to adjust the position of the transmitter in sequence to obtain all the ray paths from the transmitter to the four receivers at different positions.

[0051] For the normal vector is The triangle faces (A1, B1, C1) and their normal vectors are: Whether there is a reflection path between the triangular faces (A2, B2, C2) should follow the criterion: when When X and Y are A, B, and C, a reflection path exists between two triangular faces; otherwise, no reflection path exists. Using common visibility analysis algorithms, this criterion can be used to quickly obtain the set of triangular faces with reflection paths, reducing the need to track non-existent reflection paths.

[0052] S3. Calculate and count all ray path lengths l obtained in step S2, and compare them with the path length threshold of 1005m set in this embodiment. When l ≤ 1005m, the ray is counted in the number of rays received by the receiver; otherwise, it is not counted.

[0053] S4, Statistics on the transmitter's different locations When electromagnetic rays are emitted, the number of electromagnetic rays received by the receiver at the center of the i-th smart stealth skin is: Where i = 1, 2, 3, 4 correspond to the intelligent stealth skin located on the four sides of the compartment (top, bottom, left, and right), and are compared with the threshold number 400 set in this embodiment. If It is assumed that the intelligent stealth skin responds to the emitted electromagnetic beam, meaning that changing the scattering state of the i-th intelligent stealth skin will affect the aircraft's pitch angle (90°-θ) and azimuth angle. The RCS at the i-th point is determined by the scattering state of the i-th smart stealth skin; conversely, it is assumed that changing the scattering state of the i-th smart stealth skin will not affect the aircraft's RCS at pitch angle 90°-θ and azimuth angle θ. RCS at the location. Figure 3 To statistically determine whether the four intelligent stealth skin panels on the hexahedral aircraft section in this embodiment respond to electromagnetic beams emitted by transmitters at different locations.

[0054] S5. The intelligent stealth skin that does not respond to electromagnetic beams is placed in a non-operating state, while the intelligent stealth skin that responds to electromagnetic beams is placed in a scattering state. This process of changing between two scattering states is repeated. The RCS values ​​of the module at elevation angles of 0°–30° (5° intervals) and azimuth angles of 0°–92° are then measured, and a dataset is constructed. Table 1 shows the first set of data in the constructed dataset, namely the RCS test data of the module at an elevation angle of 0°, with the right skin in scattering state 1 and the other three skin panels in a non-operating state.

[0055] The first set of data in Table 1

[0056]

[0057] S6. Construct a multi-level, multi-layered sensing neural network model based on the distributed intelligent stealth skin scattering state and pitch angle, such as... Figure 4 As shown in Table 2, for each different sub-model, considering that the model input is the RCS value of the aircraft at some azimuth angles and the output is the RCS prediction at other azimuth angles, a multilayer perceptron neural network model is selected as the basic model. The hyperparameters of the model are shown in Table 2, where there are 26 hidden layers, each with a ReLU function as the activation function. Batch normalization (BN) layers are enabled on some hidden layers, that is, the data distribution is normalized to the standard normal distribution before the activation function receives the input value.

[0058] x and a represent the input value (RCS value) of the previous layer and the hidden layer value, respectively. w represents the hidden layer weight value, which is 0.1 in this embodiment, and b represents the hidden layer bias value, which is 0.05 in this embodiment. Therefore, the neuron value is:

[0059] a = Relu(wx + b)

[0060] The ReLU activation function is expressed as follows:

[0061]

[0062] Table 2 Hyperparameters of the Neural Network

[0063]

[0064]

[0065] S7. Select the first set of RCS data in the dataset (i.e., Table 1), divide the RCS data with azimuth angles between 0° and 83° into the training dataset, and divide the RCS data with azimuth angles between 84° and 92° into the test dataset.

[0066] S8. Input the training dataset into the neural network and train the neural network through backpropagation. After building the model, define the learning rate η as 0.0001, and use 40,000 training iterations for each sub-model or a loss value below 1 as the convergence condition. The loss function is defined as:

[0067] Loss = MSE(S - S')

[0068] S is the true label value, i.e., the RCS value; S' is the RCS value of the neural network output; and MSE is the mean squared error function, the expression of which is as follows:

[0069]

[0070] Backpropagation is performed according to the following formula:

[0071]

[0072] Backpropagation is ultimately completed by adjusting the intrinsic parameters of the neural network step by step. For example... Figure 5 As shown, as the number of training iterations increases, the model's loss value continuously decreases, eventually reaching a convergent state.

[0073] S9. Input the test dataset into the neural network for cross-validation. If the accuracy is below 90%, execute step S8 to continue training the neural network; if it is above 90%, stop training the model. When entering the test phase, keep the intrinsic parameters of the neural network model fixed and do not change them. Input all the test set data into the neural network and perform error analysis between the output values ​​and the true label values. The error analysis function is shown below:

[0074]

[0075] When the value of p is less than 0.2, the two are considered to be the same; when it is greater than 0.2, the neural network output value is considered inaccurate. For example... Figure 6 The figure shows a comparison between the RCS prediction and the measured RCS values ​​obtained by the neural network algorithm for the section in the azimuth range of 0° to 180° when the pitch angle is 0°, the right skin is in scattering state 1 and the other three skins are in non-working state. The RCS prediction values ​​have a good fit.

[0076] S10. Repeat steps S7 to S9 until all groups of RCS data in the dataset have been trained, obtaining the RCS and azimuth of the aircraft under different scattering states with distributed intelligent stealth skin at all pitch angles. The fitting relationship between them.

[0077] The above-described specific embodiments are limited to explaining and illustrating the technical solutions of the present invention, but do not constitute a limitation on the scope of protection of the claims. Those skilled in the art should understand that any new technical solutions obtained by making simple modifications or substitutions based on the technical solutions of the present invention fall within the scope of protection of the present invention.

Claims

1. A method for rapid RCS prediction of an aircraft with distributed intelligent stealth skin, characterized in that, Includes the following steps: S1. The aircraft is placed horizontally with its center at the origin of the spherical coordinate system. A transmitter capable of emitting a family of electromagnetic rays is set at a certain position P(r,θ,φ) in space. The aircraft has k identical smart stealth skins distributed on it, each with j scattering states. A receiver for receiving electromagnetic rays is set at the center of each smart stealth skin. The position coordinate θ of the transmitter corresponds to 90° minus the pitch angle of the aircraft, and the position coordinate φ corresponds to the azimuth angle of the aircraft. S2. The transmitter emits an initial family of N rays toward the origin of the coordinate system. After the rays hit the aircraft, each ray in the family is tracked. S3. Record the distance l from the ray to the center of the smart stealth skin of a receiver after it acts on the aircraft and is reflected, transmitted, and diffracted. Compare this distance with the set distance threshold L. When l ≤ L, the ray is counted in the receiver's ray count; otherwise, it is not counted. S4. Count the number of electromagnetic rays received by the receiver at the center of each smart stealth skin panel on the aircraft and compare it with the set threshold number M; for the number of electromagnetic rays m received by the receiver at the i-th smart stealth skin panel... i When m i When M ≥ M, it is assumed that the i-th smart stealth skin will respond to the electromagnetic ray beam emitted from position P(r,θ,φ), that is, changing the scattering state of the i-th smart stealth skin will affect the RCS of the aircraft at pitch angle θ and azimuth angle φ; otherwise, it is assumed that changing the scattering state of the i-th smart stealth skin will not affect the RCS of the aircraft at pitch angle θ and azimuth angle φ. S5. Set the smart stealth skin that does not respond to electromagnetic beams to a non-working state, and the smart stealth skin that responds to electromagnetic beams to a scattering state. Complete the changes of j scattering states, test and obtain the RCS values ​​of the aircraft at all pitch angles θ and some azimuth angles φ, and construct a dataset. S6. Construct a multi-level, multi-layer perception neural network model based on the scattering state and pitch angle of the distributed intelligent stealth skin. S7. Select the RCS data of the aircraft at some azimuth angles φ under different scattering states of the distributed intelligent stealth skin at the first pitch angle in the dataset, and divide 90% of it into the training dataset and 10% into the test dataset. S8. Input the training dataset into the neural network and train the neural network through backpropagation; S9. Input the test dataset into the neural network for cross-validation. When the accuracy is below 90%, execute step S8 to continue training the neural network. When it is above 90%, stop training the model. S10. Repeat steps S7 to S9 until the multi-level, multi-layered sensing neural network model is trained, and the fitting relationship between the RCS and azimuth angle φ of the aircraft under different scattering states of the distributed intelligent stealth skin at all pitch angles is obtained.

2. The method for rapid RCS prediction of an aircraft with distributed intelligent stealth skin according to claim 1, characterized in that, In step S6: For each different sub-model, the input of the model is the azimuth angle of the aircraft, and the output is the RCS value. A multilayer perceptron neural network model is selected as the basic model. The hyperparameters of the model are shown below, where there are 26 hidden layers. Each layer uses the ReLU function as the activation function. Batch normalization is enabled on some hidden layers, that is, the data distribution is normalized to the standard normal distribution before the activation function receives the input value. x and a represent the input value of the previous layer and the hidden layer value; w represents the hidden layer weight value, and b represents the hidden layer bias value. Therefore, the neuron value is: a = Relu(wx + b) The ReLU activation function expression is:

3. The method for rapid RCS prediction of an aircraft with distributed intelligent stealth skin according to claim 1, characterized in that, In step S7: The dataset was randomly divided into training and test sets in a 9:1 ratio.

4. The method for rapid RCS prediction of an aircraft with distributed intelligent stealth skin according to claim 1, characterized in that, In step S8: After building the model, a learning rate of 0.0001 is defined, and the convergence condition is set at 40,000 training iterations for each sub-model or a loss value below 1; the loss function is defined as: Loss = MSE(S - S') S is the true label value, i.e., the simulated RCS value; S' is the RCS value of the neural network output; and MSE is the mean squared error function, the expression of which is as follows: Backpropagation is performed according to the following formula: Where η is the learning rate, and backpropagation is ultimately completed by adjusting the intrinsic parameters of the neural network step by step.

5. The method for rapid RCS prediction of an aircraft with distributed intelligent stealth skin according to claim 1, characterized in that, In step S9: During testing, the intrinsic parameters of the fixed neural network model remain unchanged. All test set data is input into the neural network, and error analysis is performed between the output values ​​and the true label values. The error analysis function is shown below: When the value of p is less than 0.2, the two are considered to be the same; when it is greater than 0.2, the output value of the neural network is considered to be inaccurate.

Images