Stealth aircraft recognition method based on CNN-MSA model
By using a recognition method based on the CNN-MSA model, combined with CATIA modeling, FEKO simulation, and the improved Raccoon Optimization Algorithm (ICOA), the problem of high difficulty in recognizing stealth aircraft was solved, and accurate recognition of stealth aircraft was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENYANG AEROSPACE UNIVERSITY
- Filing Date
- 2024-03-21
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies are insufficient to effectively identify next-generation stealth aircraft, as they are difficult to identify and highly deceptive, resulting in insufficient identification accuracy.
A recognition method based on the CNN-MSA model is adopted, which combines CATIA modeling, FEKO simulation, multi-head self-attention mechanism and improved raccoon optimization algorithm ICOA to construct feature dataset and train model to improve recognition accuracy.
By combining feature extraction and discrete wavelet transform on the feature data of stealth aircraft, and introducing a multi-head self-attention mechanism and an improved raccoon optimization algorithm (ICOA), the recognition accuracy of stealth aircraft and the performance of the recognition model are significantly improved.
Smart Images
Figure CN118194016B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of aircraft radar target recognition, and specifically to a stealth aircraft recognition method based on the CNN-MSA model. Background Technology
[0002] Radar target identification technology refers to the technique of using radar to emit electromagnetic waves to illuminate a target and analyzing the obtained echoes to determine the target's type and model attributes. For next-generation stealth aircraft, due to their greater difficulty in detection and their strong deceptive capabilities, identification is even more challenging than for conventional aircraft. Therefore, proposing a novel target identification method to achieve accurate identification of stealth aircraft and promote the advancement of radar stealth technology has become an urgent problem to be solved. Summary of the Invention
[0003] In view of this, the present invention provides a stealth aircraft identification method based on the CNN-MSA model to achieve accurate identification of stealth aircraft.
[0004] The technical solution provided by this invention is as follows: A stealth aircraft identification method based on a CNN-MSA model, comprising the following steps:
[0005] Step 1: Use CATIA software to model the F-22, F-35 and FA-XX stealth aircraft;
[0006] Step 2: Use FEKO electromagnetic simulation software to solve the full-space static RCS database of the three types of stealth aircraft under both HH and VV polarization modes;
[0007] Step 3: Based on the static RCS database in Step 2, obtain the dynamic RCS sequences of the three types of stealth aircraft under steady circling tracks;
[0008] Step 4: Combine statistical feature extraction with discrete wavelet transform feature extraction to extract features from the dynamic RCS sequence and construct a feature dataset;
[0009] Step 5: Introduce a multi-head self-attention mechanism into the classic convolutional neural network to construct a CNN-MSA model, wherein the input of the CNN-MSA model is feature data and the output is stealth aircraft type data;
[0010] Step 6: Divide the feature datasets of the three types of stealth fighters under different polarization modes into training set and test set, and use the training set to train the CNN-MSA model and use the test set to test the CNN-MSA model to obtain the trained CNN-MSA model. During the training of the CNN-MSA model, the improved raccoon optimization algorithm ICOA is used to optimize the hyperparameters of the CNN-MSA model. The improved raccoon optimization algorithm ICOA uses a composite chaotic system to improve the population initialization stage of the raccoon optimization algorithm COA.
[0011] Step 7: Use radar to acquire dynamic RSC data of the stealth aircraft and extract features from the dynamic RSC data to obtain feature data. Then, input the feature data into the trained CNN-MSA model to obtain the recognition result of the stealth aircraft.
[0012] Preferably, in step 3, the method for obtaining the dynamic RCS sequence of the three types of stealth aircraft under steady circling tracks is as follows:
[0013] Establish the stealth aircraft's maneuvering trajectory, perform coordinate system transformation, and solve for the attitude angle of the radar line of sight in the aircraft's coordinate system;
[0014] The calculated attitude angles are combined with a full-space static RCS database to obtain the dynamic RCS sequences of the three types of aircraft.
[0015] Further optimization, in step 4, the feature parameters extracted by statistical feature extraction include location feature parameters, scattering feature parameters, and distribution feature parameters.
[0016] Further optimization, in step 5, the calculation formula for the multi-head self-attention mechanism is as follows:
[0017] MultiHead(Q,K,V)=Concat(head1,...,head i ...,head h W o
[0018] In the formula, h is the total number of heads, and W o This represents the weight matrix, Concat represents the vector concatenation operation, and head... i The head represents the characteristic of the i-th head. i The calculation method is as follows:
[0019] head i =Attention(QW i Q ,KW i K VW iV )
[0020] In the formula, W i Q W represents the weight matrix when the i-th head Q is used as input. i K W represents the weight matrix when the i-th head K is used as input. i V Let V represent the weight matrix when the i-th head V is used as input.
[0021] Further optimization involves dividing the training and test sets of the feature dataset in a 7:3 ratio during step 6.
[0022] Further preferred, in step 6, the mathematical formula of the complex chaotic system is as follows:
[0023]
[0024] In the formula, x(i) is the initial population, x(i+1) is the improved initial population, and r is the coefficient.
[0025] The stealth aircraft identification method based on the CNN-MSA model provided by this invention obtains clearer target features by combining statistical feature extraction and discrete wavelet transform feature extraction on the dynamic RCS sequence of the stealth aircraft. Introducing a multi-head self-attention mechanism into the CNN and constructing a CNN-MSA model can improve network performance. Introducing a Tent-Logistic-Cosine composite chaotic system to improve the population initialization stage of the CoA algorithm can enhance the algorithm's search ability. In the process of optimizing the hyperparameters of the CNN-MSA model, the optimal hyperparameters can be obtained more quickly. The stealth aircraft identification method based on the CNN-MSA model provided by this invention can achieve accurate identification of stealth aircraft. Attached Figure Description
[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.
[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0028] Figure 1 A flowchart of the stealth aircraft identification method based on the CNN-MSA model provided by the present invention;
[0029] Figure 2 The following are the dynamic RCS sequences of the F-22 aircraft in the VHF band under HH and VV polarization modes: (a) is the dynamic RCS sequence of the F-22 aircraft in the VHF band under HH polarization mode, and (b) is the dynamic RCS sequence of the F-22 aircraft in the VHF band under VV polarization mode.
[0030] Figure 3 The following are the dynamic RCS sequences of the FA-XX aircraft in the VHF band under HH and VV polarization modes: (a) is the dynamic RCS sequence of the FA-XX aircraft in the VHF band under HH polarization mode, and (b) is the dynamic RCS sequence of the FA-XX aircraft in the VHF band under VV polarization mode.
[0031] Figure 4 The following are the dynamic RCS sequences of the F-35 aircraft in the VHF band under HH and VV polarization modes: (a) is the dynamic RCS sequence of the F-35 aircraft in the VHF band under HH polarization mode, and (b) is the dynamic RCS sequence of the F-35 aircraft in the VHF band under VV polarization mode.
[0032] Figure 5 This is a wavelet transform decomposition diagram;
[0033] Figure 6 The diagram shows a comparison of the performance optimization of the test functions for the ICOA and COA algorithms. (a) corresponds to the single-modal test function, (b) corresponds to the multimodal test function, and (c) corresponds to the fixed-dimensional multimodal test function.
[0034] Figure 7 The image shows the optimal target recognition result obtained after using the classic CNN model to identify the test set under HH polarization conditions.
[0035] Figure 8 The image shows the optimal target recognition result obtained after using the classic CNN model to identify the test set under VV polarization conditions.
[0036] Figure 9 This is the optimal result image obtained after using the COA-CNN model to identify the test set under HH polarization conditions;
[0037] Figure 10 This is the optimal result image obtained after using the COA-CNN model to identify the test set under VV polarization conditions;
[0038] Figure 11 This is the optimal result image obtained after using the ICOA-CNN-MSA model to identify the test set under HH polarization conditions;
[0039] Figure 12This is the optimal result image obtained after using the ICOA-CNN-MSA model to identify the test set under VV polarization conditions;
[0040] Figure 13 A comparison chart of the target recognition accuracy of the three models on the HH polarization test set in 10 randomized experiments;
[0041] Figure 14 A comparison chart of the target recognition accuracy of the three models on the VV polarization test set in 10 randomized experiments;
[0042] Figure 15 This is a comparison chart showing the optimal recognition accuracy of the three models on two test sets.
[0043] Figure 16 This is a comparison chart of the average recognition accuracy of the three models on two test sets. Detailed Implementation
[0044] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of systems consistent with some aspects of the invention as detailed in the appended claims.
[0045] In order to more effectively and accurately identify stealth aircraft, such as Figure 1 As shown, this invention provides a stealth aircraft identification method based on a CNN-MSA model, comprising the following steps:
[0046] Step 1: Use CATIA software to model the F-22, F-35 and FA-XX stealth aircraft;
[0047] For complex, electrically large targets like aircraft, their Rectangular Cross Section (RCS) largely depends on their external shape. During modeling, specialized modeling software can more accurately reflect the detailed structure of the target model, improving the accuracy of RCS simulation data. CATIA software is recognized in the industry for its powerful surface design capabilities; therefore, CATIA was used to build the target model. CATIA's freeform surface and generative shape design modules are particularly distinctive in surface modeling. The data for target modeling came from the internet and some publicly available literature. During modeling, the aircraft's three-view drawings were first imported into the freeform surface design module, and key feature lines for each part were drawn based on these drawings. Then, the generative shape design module was used to fill and sweep the feature lines, thus completing the 3D model of the aircraft.
[0048] Step 2: Use FEKO electromagnetic simulation software to solve the full-space static RCS database of the three types of stealth aircraft under both HH and VV polarization modes;
[0049] FEKO software is easier to operate, with built-in solvers applicable to electromagnetic simulation calculations across various frequency bands. FEKO also features script editing capabilities, allowing users to create scripts as needed. Furthermore, it supports CPU multi-core parallel processing and GPU acceleration, significantly improving computational efficiency. In FEKO, the signal source is set to a plane wave, the incident wave frequency to 300MHz, the elevation angle to 0–180°, the azimuth angle to 0–360°, and the step angle to 1°. The solver selected is the geometric optics method with ray tracing (RL-GO), to solve for the static RCS across the entire spatial domain.
[0050] Step 3: Based on the static RCS database in Step 2, obtain the dynamic RCS sequences of the three types of stealth aircraft under steady circling tracks;
[0051] Steady hovering tracks are commonly used to test the dynamic RCS of aircraft. First, a maneuvering track of the stealth aircraft is established, a coordinate system transformation is performed, and the attitude angles of the radar line of sight in the aircraft's coordinate system are calculated. In space, let any point P be the position of the radar, represented in Cartesian coordinates as (x...). p ,y p ,z p If the target is considered as a point mass, then the aircraft's trajectory coordinates in the radar coordinate system can be represented by rectangular coordinates as (x...). b (t),y b (t),z b (t)). Then (x1(t), y1(t), z1(t)) represents the coordinates of point P in the target body coordinate system. According to the principle of coordinate transformation, the radar's position coordinates (x1(t), y1(t), z1(t)) in the radar coordinate system are transformed. p (t),y p (t),z p Transform the radar coordinate system (t) into the body coordinate system. The transformed coordinates are represented as (x1(t), y1(t), z1(t)). The transformation matrix from radar coordinate system to body coordinate system is... As shown in the following formula;
[0052]
[0053] The position coordinates in the radar coordinate system are transformed to the position coordinates in the body coordinate system as follows:
[0054]
[0055] Where α, γ, and η are the yaw angle, pitch angle, and roll angle, respectively.
[0056] Based on the transformed real-time position coordinates of the target, the azimuth angle of the radar line of sight in the target's coordinate system as a function of time is calculated. The pitch angle θ(t) is calculated using the following formula:
[0057]
[0058] The calculated attitude angles are combined with a full-domain RCS database to obtain the dynamic RCS sequences of the three aircraft types, such as... Figure 2 , Figure 3 , Figure 4 As shown in the table below, the statistical characteristics of the dynamic RCS of the three types of aircraft are as follows.
[0059] Table 1. Statistical characteristics of dynamic RCS of three types of aircraft
[0060]
[0061] Step 4: Combine statistical feature extraction with discrete wavelet transform feature extraction to extract features from the dynamic RCS sequence and construct a feature dataset;
[0062] The feature parameters extracted by statistical feature extraction mainly include location feature parameters, scatter feature parameters, and distribution feature parameters. Discrete wavelet transform (DWT) is an important time-frequency analysis method in wavelet analysis, possessing excellent multi-scale decomposition capabilities for complex non-stationary signals. The original signal can be processed by DWT to obtain multi-scale approximation coefficients and detail coefficients. The Mallat algorithm, as a fast algorithm for discrete wavelet transform, decomposes the original signal f(t) as shown in the following equation.
[0063]
[0064] In the above formula, c j,k Φ represents the approximation coefficient. j,k (t) represents the scale space, d m,k ψ represents the detail coefficient. m,k (t) represents the wavelet space. The decomposition can be expressed as:
[0065]
[0066] In the above formula, h(·) and g(·) represent that the scale space and wavelet space have low-pass and high-pass filters, respectively.
[0067] The Mallat algorithm is a classic tower decomposition process, and its decomposition process is as follows: Figure 5 As shown. Figure 5 The original sequence f(t) is decomposed into two sub-band signals by high-pass and low-pass filters, and the approximation coefficients c are obtained through further calculation. 1,k and detail coefficient d 1,kAnd select the approximation coefficient c 1,k Proceed to the next level of decomposition.
[0068] The inverse process of the above decomposition is the reconstruction process of the discrete wavelet transform, i.e.
[0069]
[0070] This method uses the energy of the reconstructed approximate signal and detail signal at each scale of the target dynamic RCS time series as the training features of the classifier.
[0071] Step 5: Introduce the multi-head self-attention (MSA) mechanism into the classic convolutional neural network (CNN) to construct a CNN-MSA model, wherein the input of the CNN-MSA model is feature data and the output is stealth aircraft type data;
[0072] Among them, the self-attention mechanism can focus more on the key information in the input, and its definition is as follows:
[0073]
[0074] In the formula, Q and K are feature dimensions of d. k The input data, V, is a feature dimension of d. v The input data is given by Q and K. The dot product is obtained from Q and K, and the softmax function can be used to obtain the weights of each element in V. It can scale the dot product, improving learning efficiency.
[0075] Multi-head self-attention (MSA) is essentially a linear transformation resulting from the combined computations of multiple self-attention mechanisms. Its main advantage is that it allows the model to obtain different feature information at different locations, increasing feature diversity. The calculation formula for multi-head self-attention is as follows.
[0076] MultiHead(Q,K,V)=Concat(head1,...,head i ...,head h W o (8)
[0077] In the formula, h is the total number of heads, and W o This represents the weight matrix, Concat represents the vector concatenation operation, and head... i The feature representing the i-th head can be calculated according to the self-attention mechanism in equation (7), and the calculation method is as follows:
[0078] head i =Attention(QW iQ ,KW i K VW i V (9)
[0079] In the formula, W i Q W represents the weight matrix when the i-th head Q is used as input. i K W represents the weight matrix when the i-th head K is used as input. i V Let V represent the weight matrix when the i-th head V is used as input.
[0080] Adding the MSA mechanism to the CNN network structure can further improve the network performance of CNN.
[0081] Step 6: Divide the feature datasets of the three types of stealth fighters under different polarization modes into training set and test set. Use the training set to train the CNN-MSA model and use the test set to test the CNN-MSA model to obtain the trained CNN-MSA model. During the training of the CNN-MSA model, the improved raccoon optimization algorithm (Improve COA, ICOA) is used to optimize the hyperparameters of the CNN-MSA model to improve the convergence speed and search capability of the algorithm. The improved raccoon optimization algorithm (Improve COA, ICOA) uses a Tent-Logistic-Cosine composite chaotic system to improve the population initialization stage of the raccoon optimization algorithm (COA).
[0082] Preferably, the training set and test set of the feature dataset are divided in a 7:3 ratio.
[0083] The Co-O-A algorithm, a population-based metaheuristic algorithm, uses each raccoon as a member of the population. Its position in the search space determines the value of each decision variable. Therefore, before implementing the Co-O-A algorithm, the following formula is used to randomly initialize the positions of population members in the search space:
[0084] X i :x i,j =lb j +r·(ub j -lb j ),i=1,2,...,N,j=1,2,...m (10)
[0085] In the formula, X i Let x be the position of the i-th population in the search space. i,jLet j be the values of decision variables, N be the number of individuals in the population, m be the number of decision variables, and r be a random number in the interval [0,1]. j and lb j These represent the upper and lower bounds of the j-th decision variable, respectively. Since random initialization of the population leads to uneven distribution of population members, it significantly affects the convergence speed of the algorithm and can cause it to get trapped in local optima. To address this issue, this invention employs a Tent-Logistic-Cosine composite chaotic system to improve population initialization. The mathematical formula for the Tent-Logistic-Cosine composite chaotic system is as follows:
[0086]
[0087] In the formula, x(i) is the initial population, x(i+1) is the improved initial population, and r is the coefficient.
[0088] This improvement can further enhance the convergence speed of the COA algorithm and avoid the algorithm getting stuck in local optima, thereby improving the algorithm's search capability and enabling it to obtain the optimal hyperparameters more quickly during the process of optimizing the hyperparameters of CNN-MSA.
[0089] The algorithm's local and global search capabilities are tested using three widely used standard test functions (unimodal test function, multimodal test function, and fixed-dimensional multimodal test function). Details of the three standard test functions are shown in the table below.
[0090] Table 2 Information on Three Types of Standard Test Functions
[0091]
[0092]
[0093] To verify the superiority of the improved algorithm, the basic Raccoon Optimization Algorithm (COA) and the composite chaotic mapping improved Raccoon Optimization Algorithm (ICOA) used in this invention were selected for comparison. The comparative experiment used the controlled variable method, setting the population size of the two algorithms to 30, the maximum number of iterations to 1000, and the dimension n to 30. Figure 6 The graphs in the middle show the iterative optimization of the two algorithms under three standard test functions. The horizontal axis represents the number of iterations, and the vertical axis represents the fitness function value. Figure 6In Figures (a)-(c), the blue line represents the fitness curve of the basic raccoon optimization algorithm, and the green line represents the fitness curve of the raccoon optimization algorithm improved by the composite chaotic mapping. Comparative analysis reveals that, compared to the ordinary raccoon optimization algorithm, the composite chaotic mapping-improved raccoon optimization algorithm provided in this invention has a faster search speed and stronger local and global search capabilities. This demonstrates that the ICOA optimization algorithm outperforms the COA algorithm.
[0094] The superiority of the proposed CNN-MSA model is verified below by conducting 10 randomized experiments on each of the following models: a typical CNN model (denoted as CNN), a CNN model using the COA algorithm for hyperparameter optimization (denoted as COA-CNN), and the proposed CNN-MSA model using the ICOA algorithm for hyperparameter optimization (denoted as ICOA-COA-CNN). The optimal recognition results of the three models on the HH and VV polarization test sets in the 10 randomized experiments are as follows: Figure 7-12 As shown in the figure. The red solid circles represent the actual categories, and the blue triangles represent the categories predicted by the models. The recognition accuracy of the three models in 10 randomized experiments is shown in the figure. Figure 13 , 14 As shown.
[0095] The recognition results of the three models are shown in the table below.
[0096] Table 3 Recognition results of the three models
[0097]
[0098]
[0099] according to Figure 13 , 14As shown in Table 3, in 10 randomized experiments, the optimal recognition accuracy of a typical CNN network for the HH polarization dataset was 96.61%, and the average recognition accuracy was 95.68%; for the VV polarization dataset, the optimal recognition accuracy was 94.92%, and the average recognition accuracy was 93.98%. In the 10 randomized experiments, the CNN model using the COA algorithm for hyperparameter optimization achieved an optimal recognition accuracy of 97.46% for the HH polarization dataset, an improvement of 0.88% compared to the standard CNN model, and an average recognition accuracy of 96.53%, also an improvement of 0.88% compared to the standard CNN model; for the VV polarization dataset, the optimal recognition accuracy was 96.61%, an improvement of 1.69% compared to the standard CNN model, and an average recognition accuracy of 95.25%, an improvement of 1.27% compared to the standard CNN model. The ICOA-CNN-MSA model used in this invention achieved an optimal recognition accuracy of 99.15% on the HH polarization dataset in 10 random trials, representing an improvement of 2.54% and 1.69% compared to the CNN and COA-CNN models, respectively. Its average recognition accuracy was 98.48%, an improvement of 2.8% and 1.95% compared to the CNN and COA-CNN models, respectively. On the VV polarization dataset, the optimal recognition accuracy was 98.31%, an improvement of 4.11% and 1.7% compared to the CNN and COA-CNN models, respectively. Its average recognition accuracy was 97.77%, an improvement of 3.79% and 2.52% compared to the CNN and COA-CNN models, respectively. A comparison of the recognition accuracy of the three models on different datasets is shown in the figure below. Figure 15 , 16 As shown.
[0100] The above analysis shows that the stealth aircraft identification method based on the CNN-MSA model proposed in this invention further improves the identification performance of the stealth aircraft identification method based on the classic CNN model, and can adapt to target identification under different polarization conditions.
[0101] Step 7: Use radar to acquire dynamic RSC data of the stealth aircraft and extract features from the dynamic RSC data to obtain feature data. Then, input the feature data into the trained CNN-MSA model to obtain the recognition result of the stealth aircraft.
[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.
Claims
1. A stealth aircraft identification method based on a CNN-MSA model, characterized in that, Includes the following steps: Step 1: Use CATIA software to model the F-22, F-35 and FA-XX stealth aircraft; Step 2: Use FEKO electromagnetic simulation software to solve the full-space static RCS database of the three types of stealth aircraft under both HH and VV polarization modes; Step 3: Based on the static RCS database in Step 2, obtain the dynamic RCS sequences of the three types of stealth aircraft under steady circling tracks; Step 4: Combine statistical feature extraction with discrete wavelet transform feature extraction to extract features from the dynamic RCS sequence and construct a feature dataset; Step 5: Introduce a multi-head self-attention mechanism into the classic convolutional neural network to construct a CNN-MSA model, wherein the input of the CNN-MSA model is feature data and the output is stealth aircraft type data; Step 6: Divide the feature datasets of the three types of stealth fighters under different polarization modes into training set and test set, and use the training set to train the CNN-MSA model and use the test set to test the CNN-MSA model to obtain the trained CNN-MSA model. During the training of the CNN-MSA model, the improved raccoon optimization algorithm ICOA is used to optimize the hyperparameters of the CNN-MSA model. The improved raccoon optimization algorithm ICOA uses a composite chaotic system to improve the population initialization stage of the raccoon optimization algorithm COA. Step 7: Use radar to acquire dynamic RSC data of the stealth aircraft and extract features from the dynamic RSC data to obtain feature data. Then, input the feature data into the trained CNN-MSA model to obtain the recognition result of the stealth aircraft.
2. The stealth aircraft identification method based on the CNN-MSA model according to claim 1, characterized in that: In step 3, the method for obtaining the dynamic RCS sequence of the three types of stealth aircraft under steady circling tracks is as follows: Establish the stealth aircraft's maneuvering trajectory, perform coordinate system transformation, and solve for the attitude angle of the radar line of sight in the aircraft's coordinate system; The calculated attitude angles are combined with a full-space static RCS database to obtain the dynamic RCS sequences of the three types of aircraft.
3. The stealth aircraft identification method based on the CNN-MSA model according to claim 1, characterized in that: In step 4, the feature parameters extracted by statistical feature extraction include location feature parameters, scatter feature parameters, and distribution feature parameters.
4. The stealth aircraft identification method based on the CNN-MSA model according to claim 1, characterized in that: In step 5, the calculation formula for the multi-head self-attention mechanism is as follows: MultiHead(Q,K,V)=Concat(head1,...,head i ...,head h )W o In the formula, h is the total number of heads, and W o This represents the weight matrix, Concat represents the vector concatenation operation, and head... i The head represents the characteristic of the i-th head. i The calculation method is as follows: head i =Attention(QW i Q ,KW i K ,VW i V ) In the formula, W i Q W represents the weight matrix when the i-th head Q is used as input. i K W represents the weight matrix when the i-th head K is used as input. i V Let V represent the weight matrix when the i-th head V is used as input.
5. The stealth aircraft identification method based on the CNN-MSA model according to claim 1, characterized in that: In step 6, the training set and test set of the feature dataset are divided in a 7:3 ratio.
6. The stealth aircraft identification method based on the CNN-MSA model according to claim 1, characterized in that: In step 6, the mathematical formula for the complex chaotic system is as follows: In the formula, x(i) is the initial population, x(i+1) is the improved initial population, and r is the coefficient.